Tilburg University Describing, explaining and - Research portal

Loading...

Tilburg University

Describing, explaining and predicting health care expenditures with statistical methods Wong, A.

Document version: Publisher's PDF, also known as Version of record

Publication date: 2012 Link to publication

Citation for published version (APA): Wong, A. (2012). Describing, explaining and predicting health care expenditures with statistical methods Enschede: Gildeprint

General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. - Users may download and print one copy of any publication from the public portal for the purpose of private study or research - You may not further distribute the material or use it for any profit-making activity or commercial gain - You may freely distribute the URL identifying the publication in the public portal Take down policy If you believe that this document breaches copyright, please contact us providing details, and we will remove access to the work immediately and investigate your claim.

Download date: 07. Jan. 2018

Describing, explaining and predicting health care expenditures with statistical methods

The research as described in this thesis was carried out at the National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands, and at Tilburg University, Tilburg, the Netherlands. The results could not have been obtained without the financial support of the National Institute for Public Health and the Environment, the Dutch Ministry of Health, Welfare and Sport (Ministerie van VWS), and the Netherlands Organization for Health Research and Development (ZonMW).

ISBN/EAN: 9789461082633 Copyright © 2012, Albert Wong Cover design: Hiuli Wong Artwork: Hiuli Wong Printing: Gildeprint Drukkerijen, Enschede, The Netherlands All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior written permission of the author.

Describing, explaining and predicting health care expenditures with statistical methods

PROEFSCHRIFT ter verkrijging van de graad van doctor aan Tilburg University op gezag van de rector magnificus, prof. dr. Ph. Eijlander, in het openbaar te verdedigen ten overstaan van een door het college voor promoties aangewezen commissie in de aula van de Universiteit op vrijdag 16 maart 2012 om 14:15 uur door

Albert Wong geboren op 21 september 1980 te Enschede

Promotiecommissie Promotores prof. dr. J.J. Polder prof. dr. H.C. Boshuizen

Overige leden prof. dr. S. Felder prof. dr. R.T.J.M. Janssen dr. M.A. Koopmanschap prof. dr. P. Kooreman

Contents Chapter 1

General Introduction

Part I Chapter 2

The Red Herring Phenomenon Exploring the Influence of Proximity to Death on DiseaseSpecific Hospital Expenditures: a Carpaccio of Red Herrings Standardizing the Inclusion of Indirect Medical Costs in Economic Evaluations Time to Death and the Forecasting of Macro-Level Health Care Expenditures: Some Further Considerations

Chapter 3 Chapter 4 Part II Chapter 5 Chapter 6 Chapter 7 Chapter 8 Part III

The Relationship between the Presence of One or More Diseases and Health Care Expenditures Predictors of Long-Term Care Utilization by Dutch Hospital Patients Aged 65+ The Disabling Effect of Diseases: A Study on Trends in Diseases, Activity Limitations, and Their Interrelationships Longitudinal Administrative Data Can Be Used to Examine Multimorbidity Provided False Discoveries Are Controlled for Comorbidity and Hospital Care Expenditures: Does One Disease Affect the Expenditures for Another?

1

11 37 55

87 105 123 139

Chapter 11

The Interpersonal and –Generational Distribution of Health Care Expenditures Medical Innovation and Age-Specific Trends in Health Care Utilization: Findings and Implications Modeling the Distribution of Lifetime Health Care Expenditures: a Nearest Neighbor Resampling Approach Interpersonal Variation in Lifetime Health Care Expenditures

185 213

Chapter 12

General Discussion

235

Bibliography Summary Samenvatting

247 265 273

Acknowledgements Curriculum Vitae

281 285

Chapter 9 Chapter 10

165

Chapter 1 General Introduction

1.1 Background In the past centuries, the worldwide economy has changed considerably. In the 18th century, the economy was transformed forever with the first Industrial Revolution. During this period great progress was made in textile manufacturing, mining, metallurgy and transport, thanks to the introduction of steam power and coal as energy sources. As a result, socioeconomic conditions improved considerably. In the 19th century, innovations such as the lightbulb, railways and steamboats further boosted the economy. And in the 20th century, the innovation rate was even more rapid. In recent decades however, the economy has seen a substantial shift from the primary sector (agriculture, fishing and mining) and secondary sector (manufacturing) to the tertiary sector (services). The service industry is the largest sector in Western countries, and includes many forms of services. Health care is seen as one of the most prominent, and is only expected to become more important in the future. As Robert Fogel, Nobel prize laureate in economics, once said (Fogel, 2004): “…Just as electricity and manufacturing were the industries that stimulated the growth of the rest of the economy at the beginning of the twentieth century, health care is the growth industry of the twenty‐first century. It is a leading sector, which means that expenditures on health care will pull forward a wide array of other industries including manufacturing, education, financial services, communications, and construction…” Aside from stimulating other industries, the benefits of health care are, without a doubt, enormous for society. In the Netherlands, life expectancy has increased from around 70 [73] years in 1950 to 78 [83] years in 2010 for men [women]

CHAPTER 1

(Statistics Netherlands, 2010). Where heart diseases and cancer led to almost immediate death in the past, health care now ensures much longer survival. Clearly, health care is a necessity on the individual level. However, it is not necessarily the driving factor behind health care expenditures on an aggregate level. Getzen (2000) argued that macro-level expenditures is not simply the sum of all individual health care demand, but that the expenditures are often restricted in the sense that government or health care insurer sets a budget for health care spending. Since the budget is related to how much a nation can spend (in terms of national income), health care may be considered as a luxury good on a national level, rather than a necessity. Figure 1.1 shows how health care expenditures in the Netherlands have evolved through time. The total health care expenditures have increased thirteen-fold over the period 1972-2009. Of course, these expenditures include the growth in population size. The per capita expenditures increased eleven-fold during this period, and when keeping the price fixed at the 2000 level, the increase is threefold. Not only do the health care expenditures grow with national income (Gross Domestic Product), they also grow faster than the GDP: the share of health care expenditures in the Dutch GDP has risen from near 7% to almost 12% in the span 38 years. This trend is expected to continue in the coming decades. Fogel (2004) noted the following on this: “…The increasing share of global income spent on healthcare expenditures is not a calamity; it is a sign of the remarkable economic and social progress of our age…” While health care is seen as something worth investing in, policy makers are worried about the financing of health care. Therefore, many policy makers are not only interested in cost containment in health care, but also in alternative health care financing systems. To make well-informed decisions in these areas, a bigger understanding of the drivers and dynamics of health care expenditures is needed. Current research can be divided into macro-level and micro-level studies. Koopmanschap et al. (2011) and Kommer et al. (2010) provide a comprehensive review of all currently studied determinants in health care expenditures. In this chapter, we will outline the determinants and some properties of health expenditures that bear relevance to the remainder of this thesis. Aging and the last year of life Population aging can occur both in a temporal and structural way. Birth cohorts may differ in size through time. For example, there was a large amount of births in 2

GENERAL INTRODUCTION

the period following World War II (commonly referred to as ‘baby boomers’). Given the large share of this cohort in the total population, the population –often expressed in terms of median age– will age as a result. As this cohort ages the demand for health care will inevitably rise. Structural aging is caused by decreasing fertility rates and increasing longevity (or simply, ‘aging’). The latter has been extensively debated in research. Since health care expenditures rise steeply with age (Figure 1.2), it is suggested that increasing longevity leads to a great growth in health care expenditures. Yet, this line of thought can be nuanced. As Zweifel et al. (1999) found, the health care expenditures are much higher in the last year of life. For the Netherlands, the expenditures in the last year of life are 13.5 times higher than the expenditures in other years (Polder et al., 2006). As such, using the crosssectional age-distribution of health care costs in Figure 1.2 may not lead to a representative assessment of the growth in health care expenditures as result of aging. The finding suggests that with increasing longevity, a large part of the expenditures will be simply postponed, as they fall in the last year of life. Thus, increasing longevity may not lead to as a large as increase in expenditures as was once thought. The role of aging in the growth of health care expenditures has therefore been commonly referred to as a ‘Red Herring’, as it diverts the attention from other factors, such as medical technology and institutional factors. (Co-)Morbidity and disability On the individual level, health plays an important part in the demand for health care. Health is a fairly abstract concept, which can be operationalised in many ways. Morbidity, co-morbidity (the presence of two or more diseases), and disability are determine health care demand on an individual level. While age might be correlated, age alone does not approximate these processes well. Morbidity and co-morbidity have been associated with higher health care expenditures (see Gijsen et al., 2001). The exact relationship is not yet understood however, as each of these are difficult to disentangle from one another. Particularly, comorbidity is a complicating factor, as there exist a multitude of diseases, of which –at least in theory– any combination can co-exist. While the presence of multiple diseases will likely lead to higher health care expenditures, it is not clear whether the total expenditures is more or less than the sum of average expenditures for each disease alone (i.e., the expenditures in case when only one disease is present). Lifetime dynamics in health care expenditures Most studies on individual health care expenditures focus on estimating averages over age. However, patterns in these averages do not reflect individual patterns very well. De Nardi et al. (2010) found that both the level and volatility of health care expenditure increase sharply with age. Health care expenditures tend to fluctuate strongly within a lifetime, as well as between individuals. 3

CHAPTER 1

1980

1990

2000

12 7

8

9

10

11

Share HCE in GDP (percent)

50000 30000 10000

Total HCE (Euros)

70000

Figure 1.1: Health care expenditures over 1972-2010 (OECD, 2010).

2010

1980

1990

1990

2000

2000

2010

1000

2000

3000

Average HCE (2000 price level)

3000 2000 1000

Average HCE (Euros)

1980

2010

Year

4000

Year

2000

2010

1980

1990

Year

Year

Figure 1.2: Average health care expenditures by age and gender in the Netherlands in 2007 (Slobbe et al., 2011). Men

Women

55000 50000 45000 40000 35000 30000 25000 20000 15000 10000 5000

4

95+

90-94

85-89

80-84

75-79

70-74

65-69

60-64

55-59

50-54

45-49

40-44

35-39

30-34

25-29

20-24

15-19

10-14

5-9

1-4

0

0

GENERAL INTRODUCTION

Few studies have dealt with these dynamics. French and Jones (2004) found that 0.1% of US households suffer a health shock that leads to lifetime costs over $125,000. Long-term care, as the name suggests, features dynamics that are different from those in acute health care expenditures. Brown and Finkelstein (2008) found that only 40% of current 65-year old males and 54% of 65-year old females will use long-term care at some point in time, suggesting long-term care is highly uneven distributed amongst individuals. Advances in medical technology In all Western countries, the growth in health care expenditures is greater than the rate of population aging. Newhouse (1992) found that after accounting for per capita income, insurance coverage level and population aging, a large part of the growth was still unexplained. He attributed the remaining growth for the greater part to ‘the enhanced capabilities of medicine’, which includes a wide range of technologies: new drugs, medical devices, implants, nanotechnology, genetic and biochemical techniques, imaging techniques and bioinformatics (Knecht & Oomen, 2006). Innovations in these areas are likely to spur the demand for health care, and thus, determine the growth in health care expenditures. While a lot of research has already been done on health care expenditures, it is clear that still a lot is not yet understood. Given the importance for the future of health care, further research on this topic is strongly desired.

1.2 Aim/Motivation The main goal of this thesis is to identify several aspects of health care expenditures that have been given little attention in the literature, and investigate how statistical methods can be used to explore, and generate additional knowledge on, these facets. The knowledge may contribute to further health economics research, or facilitate decision making with regards to the future of health care and health care financing. As outlined in the previous section, the relationship between health care expenditures and aging is clearly multifaceted. Within this thesis, we will focus on the following three themes that deserve further exposition: I. The ‘Red Herring’ phenomenon, which is used by researchers to understand the role of aging, and to make projections of health care expenditures based on demographic changes;

5

CHAPTER 1

II. The relationship between the presence of one or more diseases, and the amount of health care utilization, often expressed in terms of health care expenditures; III. The interpersonal and -generational distribution of health care expenditures. Each of these themes can be divided into subproblems. Following Harrell (2001), Shmueli (2010) and Shmueli and Koppius (2011), we can distinguish between three classes of modeling to tackle these subproblems:  Descriptive modeling, which is aimed at summarizing or representing the data structure in a compact manner. The focus is on discovering and capturing associations between variables, without necessarily relying on a causal framework;  Explanatory modeling, which is used to test causal hypotheses. This class is driven by an underlying theoretical framework, which allows the structure of the model to be more defined a priori. Testing a hypothesis will not necessarily prove the causality (as this requires experiments in which factors are changed by the experimenter), but rather provide empirical support for the causality (which follows from the theoretical framework);  Predictive modeling, where new observations are predicted using a set of explanatory variables. A special subclass of predictive modeling is forecasting, where only future observations are predicted. These are similar to explanatory modeling, with the exception that the model structure is largely data driven. The choice for modeling class strongly depends on the nature of the problem. Typically, with a more open-ended nature of the problem, a descriptive approach is needed, while explanatory modeling is suited for a more close-ended nature. If predictions are of interest, rather than associations or causal relations, predictive modeling is required. This thesis will show how all three approaches are useful in gaining insight into the following:  The determinants of individual health care expenditures, and how they may, or may not, affect aggregate health care expenditures;  The distribution of health care expenditures over individuals, and over groups of individuals (e.g. gender, age, and those who are in their last year of life);  The dynamics of health care expenditures over a lifetime.

6

GENERAL INTRODUCTION

1.3 Outline of the thesis The thesis is outlined as follows. There are three themes in the context of health care expenditures which are explored: the Red Herring (Chapters 2-4), the relationship between (co-)morbidity and health care expenditures (chapters 5-8), and interpersonal and -generational health care expenditures (chapters 9-11). The thesis ends with a general discussion (chapter 12). Chapter 2 further examines the concept of ‘the health care expenditures in the last year of life’ by examining disease-specific hospital care expenditures. An explanatory modeling approach is proposed, to determine for which diseases an association exists with the last year of life. Results from Chapter 2 are used in Chapter 3 to provide a standardized way of predicting unrelated medical care expenditures in cost-effectiveness analyses. In contrast to all other chapters in this thesis, a deterministic approach is chosen here. Chapter 4 gives an exposition on how current researchers use ‘time-to-death’ in the forecasting of macro-level health care expenditures. The ‘time-to-death’ effect is tested on macro-level data. Some methodological issues of forecasting with timeto-death are being raised here, and some propositions are made to improve on these issues. In Chapter 5 a combination of explanatory and descriptive modeling is used to find which factors predict long-term care utilization after a hospital discharge. A theoretical framework is used to determine many of the explanatory variables in the model, with the exception of diseases, which are selected based on the data at hand. In Chapter 6 the hypothesis is tested that diseases may have become less disabling over time. Using an explanatory modeling approach, the existence of trends in morbidity, comorbidity and disability are explored, as well as the association between (co-)morbidity and disability in the Netherlands. Given that there may be a trend in comorbidity, Chapter 7 explores which specific types of comorbidity may occur more frequently than expected due to chance, amongst Dutch hospital patients. This is a purely descriptive approach, with no underlying etiological framework.

7

CHAPTER 1

This specific comorbidity approach is taken further in Chapter 8 by examining the relationship with hospital care expenditures. An overview is given of all comorbidities that are associated with a higher than additive level of expenditures. Chapter 9 considers a theoretical framework to explain hospital care utilization. The hypothesis being tested here, is that advances in medical technology will benefit the elderly more. Chapter 10 adapts an existing method to forecast synthetic individual lifecycles in health care expenditures. Using these lifecycles, a predicted lifetime distribution in health care expenditures can be obtained. The method from Chapter 10 is further modified in Chapter 11 to forecast synthetic lifecycles for health care expenditures in each specific health care sector. This thesis ends with Chapter 12, where all findings are briefly summarized and a general discussion is provided.

8

10

Chapter 2 Exploring the Influence of Proximity to Death on DiseaseSpecific Hospital Expenditures: A Carpaccio of Red Herrings ‡

It has been demonstrated repeatedly that time to death is a much better predictor of health care expenditures than age. This is known as the ‘red herring’ hypothesis. In this article, we investigate whether this is also the case regarding disease-specific hospital expenditures. Longitudinal data samples from the Dutch hospital register (n = 11,253,455) were used to estimate 94 disease-specific two-part models. Based on these models, Monte Carlo simulations were used to assess the predictive value of proximity to death and age on disease-specific expenditures. Results revealed that there was a clear effect of proximity of death on health care expenditures. This effect was present for most diseases and was strongest for most cancers. However, even for some less fatal diseases, proximity to death was found to be an important predictor of expenditures. Controlling for proximity to death, age was found to be a significant predictor of expenditures for most diseases. However, its impact is modest when compared to proximity to death. Considering the large variation in the degree to which proximity to death and age matter for each specific disease, we may speak not only of age as a ‘red herring’ but also of a ‘carpaccio of red herrings’. 2.1 Introduction In recent years, a lot of research has been done on the impact of ageing of the population on health care expenditures (HCE). Arguably the most influential paper This chapter is based on: Wong A, van Baal PHM, Boshuizen HC, Polder JJ. 2011. Health Economics 20(4): 379-400. This article has been reproduced with permission from John Wiley & Sons, Ltd. ‡

CHAPTER 2

in this area was published by Zweifel et al. (1999). They analyzed the relationship between age and HCE, using longitudinal Swiss sick fund data, and found that the magnitude of HCE is explained to a greater extent by proximity to death than by age. As such, population ageing might have a more limited impact on HCE growth than generally believed. The authors, therefore, suggested that ageing of the population was a ‘red herring’ that diverts attention from the real causes of HCE growth, such as government regulations in the health care sector and advances in medical technology. Excluding time to death in estimates of future HCE will result in an overestimation of total HCE (in future projections based on demographic trends), as several papers have found to varying degrees (Wickstrøm et al., 2002; Stearns and Norton, 2004; Polder et al., 2006). The study by Zweifel et al. received a lot of attention, not only due to its strong conclusions about the relatively mild effect of ageing on HCE but also because of methodological issues. These issues include the endogeneity of closeness to death (Salas and Raftery, 2001) and the use of the Heckit model over a two-part model (Dow and Norton, 2003; Seshamani and Gray, 2004a). Zweifel et al. (2004) subsequently addressed most of these methodological concerns analyzing both old and new Swiss sick fund data and found that the claim made in the seminal paper (Zweifel et al., 1999) was still valid. Other studies have also supported this claim in recent years, sometimes using different data sets and employing different methodologies. Yang et al. (2003) confirmed it for a longitudinal survey of Medicaid beneficiaries. Seshamani and Gray (2004b) performed a random effects analysis using an English data set of hospital admissions spanning a period of 29 years and found that approaching death affects HCE up to 15 years before death, and that the increase in HCE in the years before death overshadows the increase in HCE associated with age. Dormont et al. (2006) found that the rise in HCE due to ageing is small, especially when compared to the rise caused by changes in medical practices. Häkkinen et al. (2008) also confirmed the limited role of ageing in HCE from Finnish data. The red herring claim has also been explored for different health care components (Werblow et al., 2007). Proximity to death was found to be a good predictor for ambulatory care, use of drugs, hospital inpatient, and outpatient care. Werblow et al. concluded that there is a ‘school of red herrings’. Long-term care might be the sole exception, where both age and proximity to death play a major role. Weaver et al. (2008) found that proximity to death is one of the main drivers of long-term care, but that changes in the availability of informal care might diminish its importance. In this article, we make a contribution to the literature by investigating the red herring claim for disease-specific hospital HCE. Data from a nationwide hospital register were used to estimate disease-specific two-part models for 94 disease categories (referred to as ‘diseases’ hereafter). Our aim was to assess the relationship between proximity to death and hospital HCE for each disease and to 12

A CARPACCIO OF RED HERRINGS

examine the extent to which age influences disease-specific hospital HCE while controlling for proximity to death. The influence of proximity to death on diseasespecific hospital HCE is evaluated for each disease by means of estimating the (disease-specific) ratio of hospital HCE of those who died in a particular year (the ‘deceased’ hereafter) to those of survivors. For determining the importance of age as a predictor compared to proximity to death, the successive ratios of current hospital HCE to hospital HCE at an age 5 years younger were estimated for each disease. We provide detailed analyses for eight diseases and summarize the results for the other disease categories.

2.2 Methods Previous HCE studies (Seshamani and Gray, 2004b; Werblow et al., 2007) used the so-called two-part models (Mullahy, 1998) to estimate HCE and to examine the association with age and time to death. In the case of hospital HCE, the first part of the two-part model estimates the proportion of individuals being hospitalized and the second part estimates the HCE conditional on being hospitalized. This model is preferred over the Heckit model in analyses of determinants of actual HCE in situations where sample selectivity is not a problem (Dow and Norton, 2003). The main difference with previous studies is that we modeled the hospital HCE for a given disease, rather than estimating the total hospital HCE. Essentially this means that we ran multiple two-part models, one for each disease. The disease-specific HCE encompass all hospital expenditures belonging to those admissions that have that disease coded as their primary hospital diagnosis. In other words, we split up total hospital HCE by primary hospital diagnosis, and modeled each disease separately, while the sum of all disease-specific HCE is equal to the total HCE. Furthermore, instead of combining all data into one model with death and time to death as explanatory variables (Werblow et al., 2007), we constructed separate models for deceased and the survivors, in order to deal with the different age patterns between both groups. This is explained further in the later sections. In the literature dealing with the issue, there are two common ways to test the ‘red herring hypothesis’. The first is to look at the statistical significance of age (Zweifel et al., 1999). In our case, this would have been an inappropriate choice, because we modeled age using ‘splines’, which precludes testing for statistical significance (explained further below). The second way is to calculate age-specific HCE predictions, and then to examine absolute differences over age intervals, particularly if age is statistically significant (Zweifel et al., 2004; Werblow et al., 2007). In some studies that have used this approach, these differences are tested for significance (Seshamani and Gray, 2004b). In this study, we tested ratios for 13

CHAPTER 2

statistical significance 1. On the basis of expenditure estimates calculated with the model, disease-specific ratios were estimated of deceased and survivor HCE, as well as disease-specific ratios of successive ages among survivors. Data Data on hospital inpatient care utilization collected through the Dutch Hospital Discharge Register (LMR) were obtained from the Prismant health care services institute (Prismant, 2008). All university and general hospitals and most specialized hospitals agreed to participate in this register for the period 1995–2004. As a result, the LMR provides a nearly complete coverage of all hospital inpatient admission in the Netherlands. It includes administrative patient data such as date of admission and discharge, and extensive diagnosis (on ICD-9 level) and treatment information (including about 10 000 medical procedures). In this article, we focus on inpatient care including all clinical procedures and day cases, comprising 60% of total hospital HCE, or about 16.1% 2 of the total HCE in the Netherlands (Slobbe et al., 2006). Costs per admission consisted of two parts: intervention costs and all other costs associated with hospital stay. Since all interventions were registered in the LMR, intervention costs per patient could be calculated using the detailed remuneration schemes of the Dutch hospital payment system, which provided for each intervention all relevant doctor fees and the hospital’s reimbursement for associated costs of, among other things, equipment, materials, and personnel. All other costs of hospital stay such as nursing and accommodation costs were calculated on a daily basis, using average costs per day. Costs were aggregated per admission. The resulting average costs per patient were validated using health insurance data on average hospital HCE by age and gender. The LMR data as such were not suited for longitudinal analysis as patients could not be identified over longer periods of time. To deal with this, Statistics Netherlands linked the LMR data set to the Dutch Person Register, a nationwide register of all Dutch individuals. It includes variables such as date of birth, date of death, gender, living situation, and residence, and was available for the period 1995–2005. The success rate of this linkage has been found to be satisfactory (for the LMR, 87% of the yearly admissions were linked successfully (Bruin et al., 2004)). Just like absolute differences, ratios have the advantage of being easier to interpret than coefficients from a multitude of models. However, they were preferred over absolute differences as they also allow for easier comparison between and within diseases. 2 This percentage might seem low in comparison to other countries. However, the definition of total health care expenditures used in Dutch statistical records includes many forms of long-term care, such as homes for the elderly, and disabled care. As the line between health care and social care is drawn differently in different countries, international comparison is difficult (Heijink et al., 2008). 1

14

A CARPACCIO OF RED HERRINGS

A number of steps were needed to prepare the data sets for the statistical analysis. These include reformatting the data into a panel structure, where each individual had a yearly observation. A life year of a deceased individual was based on the date of death, whereas for survivors it was based on the date of birth. As the time to death effect has been found for up to 15 years before death (Seshamani and Gray, 2004b), restricting the deceased HCE to the last year of life was deemed insufficient. Definitions were chosen for deceased and survivors in such a way that a balance was struck between having a reasonable number of yearly observations per individual (six), and having a reasonable number of last years of life for the deceased (five). Thus, the deceased were defined as those persons who died during the study period or within 5 years of the last year of the study period. Thus, as dates of death were available until 2005, only observations of individuals from the period 1995–2000 were used. The remaining individuals were considered survivors: they survived for at least 5 years after the end of the study period. Admissions were linked to a particular year based on the hospital discharge data. Information on demographics, admission(s), diagnosis, and costs were coded in such a way that all variables correspond to the period to which an observation relates. Diagnoses were limited to those that were coded as the principal diagnosis. Diagnoses were originally coded in ICD-9 format, but recoded to ISHMT format (WHO, 2010), as ICD-9 provides a categorization of diseases that we found to be too detailed for our purposes. The ISHMT format leaves us with 130 disease categories. From these we discarded the diseases that are restricted to younger age groups (i.e., those related to pregnancy and childbirth, perinatal conditions, and congenital malformations), or are caused by external factors (injury and poisoning), or belong to the classification ‘unknown’ (ISHMT disease chapters 18 and 21), leaving a total of 94 diseases. Individuals who could not be linked to the Dutch Person Register throughout the period were excluded from the data set. The resulting data set contains approximately 11.25 million individuals. The data set was divided into two subsets: one set of individuals who were admitted to a hospital at some point through 1995–2000 (roughly 39%) and another set of individuals without any form of hospital inpatient care during this period (61%). Characteristics of the population are listed in Table 2.1. Sampling Procedure Modeling all 11.25 million individuals at once was not possible due to hardware and software restrictions, and so a sampling procedure was used. First, the subset with admitted individuals was split into 94 smaller subsets, one for each disease (category) as the principal diagnosis. This is possible because each admission has exactly one principal diagnosis. Each of these datasets contains all the individuals who were admitted at least once during 1995–2000 for that particular disease.

15

CHAPTER 2 Table 2.1: Characteristics of the main dataset (individuals in the period 1995–2000). Percentage (n=11253455)

Variable Admissions throughout period 0 1 or more

61% 39%

Time to death (years) 1-5 6 or more

11% 89%

Age 0 1-24 25-44 45-64 65+

8% 24% 27% 25% 16%

Sex Male Female

49% 51%

Figure 2.1: Probability sampling scheme for a given disease x . Population (N=11,253,455)

[1] Population admitted for disease x (N1)

Sampling probability pi for each stratum i such that

pn i

i

i

 min(80,000; N 1 )

[2] Population admitted for diseases other than x (N2)

[3] Population without any admission (N3)

Sampling probability qj for each stratum j such that

q n j

j

 50,000

j

Sampling probability rk for each stratum k such that

r n k

k

 50,000

k

Sample for disease x (n  min(180,000;N1+100,000))

Consequently, individuals who have been admitted for more than one principal disease are found in at least two subsets. For example, if an individual was admitted once for stroke and once for diabetes, the expenditures due to the stroke admissions were added to the stroke data set, while the expenditures due to the diabetes admissions were allocated to the diabetes data set. Each of these subsets 16

A CARPACCIO OF RED HERRINGS

was used to create data sets for the two-part statistical model, each with its own response variables, selection of individuals, and sampling procedure. Figure 2.1 gives an overview of the sampling procedure. The first part of the model estimates the proportion of individuals with hospital HCE for a specific disease and, therefore, requires data on individuals who have hospital HCE for that disease as well as individuals without hospital HCE for that disease. The latter can be (conceptually) subdivided into a group without any hospital HCE, and a group with hospital HCE for other diseases. Because the data already had a similar group structure, it was decided to keep this group structure intact to simplify the complex linkage and sampling process for each disease. The individuals with hospital HCE for a given disease are sampled from one of the 94 subsets. The probability of sampling such an individual was chosen such that the resulting sample size for individuals with hospital HCE for that disease would not exceed N1 ≈ 80,000 (deceased and survivors together, as portrayed by box [1] in Figure 2.1). For most diseases, the original sample size fell short of this number, and so they were sampled with a probability of one. For instance, this applies to individuals with epilepsy (5,610 deceased and 15,810 survivors), while those with stroke were much more numerous (56,667 and 50,621, respectively). Individuals with epilepsy therefore were sampled with probability one, whereas those with stroke were sampled with a probability such that the total number of survivors and deceased would roughly amount to 80,000. The individuals with no hospital HCE for that disease were sampled from two subgroups: First, the group of individuals who had no admissions at all ( N 2 ≈ 50,000, box [2] in Figure 2.1). Second, a group consisting of the remainder of the 94 disease subsets ( N 3 ≈ 50000, box [3] in Figure 2.1). As a result, the total sample size was either N1 +100,000 or 180,000, depending on whether the population in box [1] was sampled with probability of one or not, respectively. Larger sample sizes led to negligible differences in estimates. Part two of the model uses the cost observations from all the individuals with at least one admission in a year and thus required no sampling. Random sampling within each box was deemed inappropriate, as this resulted in very small numbers in specific groups (such as the survivors age 90 or above, or those deceased at age 50 or below). Oversampling of such groups was used to resolve this. This was done as follows. First, the population was stratified by both age and survivor status. Each individual from stratum i in box [1] was sampled with a probability pi . Similarly, individuals from strata j (box [2]) and k (box [3]) were sampled with a probability q j and rk , respectively. Smaller groups, like those mentioned above, were sampled at a higher rate, such that the resulting samples from each box had approximately the same amount of individuals as mentioned above ( N1 ≈ min(80,000; N1 ), N 2 ≈50,000 and N 3 ≈50,000). To account for the 17

CHAPTER 2

fact that the resulting data would no longer be representative of the distribution of the population in terms of age and survivor status (Table 2.1), weighting was implemented by giving each individual a weight equalling the inverse of the sampling probability. The Huber–White estimator was used to give robust variance estimates (Rogers, 1993). Model specification In the case of hospital HCE, each individual has multiple observations, and as a result, these observations are correlated. Using a cross-sectional model with such panel data would result in overestimated standard errors of time-varying covariates and underestimated standard errors of time-invariant covariates (Fitzmaurice et al., 1993; Dunlop, 1994; Hu et al., 1998). To account for these correlations, we used generalized estimating equations (GEE), as proposed by Liang and Zeger (1986). These are an extension to the generalized linear model. GEE were chosen over random effect models, as we were more interested in population-averaged expenditures than in expenditures of the average individual. These estimates are not the same when the model uses a distribution other than the normal distribution (Molenberghs and Verbeke, 2000). Population-averaged expenditures can be used directly to estimate total HCE, by simply multiplying by the number of individuals. This is not as straightforward with estimates of the average individual. One limitation of GEE, however, is that they are estimated not with a likelihood function, but with a quasi-likelihood function, which only specifies the first two moments (mean  and variance V (  ) ). This means that most likelihood-based approaches for goodness-of-fit testing cannot be used. In recent research, alternatives for GEE model selection have appeared, however. Thus, for model selection, we used the QIC value, which is a modified version of the Akaike Information Criterion for GEE (Pan, 2001) and has been implemented in Stata (Cui, 2007). Part one models the proportion 3 of individuals with hospital HCE for a given disease, given a set of covariates x . It is common to use the logit or probit link in conjunction with the binomial family, and so we opted for the logit link here. Let  be the error term, i the index for the individual, and t the calendar time. If we denote P[ HCE  0 | x] , the proportion of individuals with non-zero HCE, by p , then the regression equation for part one becomes In the literature pertaining to time to death analyses (Zweifel et al., 1999, 2004; Seshamani and Gray 2004a, b, c; Werblow et al., 2007) the dependent variable of ‘part one’ is often taken to represent the probability of having HCE. We believe this is not quite appropriate, as the dependent variable is regressed on the time to a future event (death). It should, in our view, rather be seen as a retrospective analysis where the proportion of individuals with expenditures in the last year of life is compared to other years. This description also suggests an analysis of associations, rather than one of causality. By using this definition, we steer clear of the ‘reversed causality’ nature of the model (the level of expenditures can affect the chance of survival, yet expenditures is regressed on time to death).

3

18

A CARPACCIO OF RED HERRINGS

f ( pi ,t )  logit(pi,t )  log[pi,t /(1 - p i,t )]  xiT,t    i ,t

(2.1)

In part two, the hospital HCE for the given disease in a particular year, conditional on the hospital HCE (for that same disease) being nonzero and a set of covariates x , is estimated. In this case, there is not an obvious candidate for the family and link function. We used the algorithm proposed by Manning and Mullahy (2001) to determine an appropriate family and link selection. Some studies have suggested that the gamma family and log link function are most suitable (Blough and Ramsey, 2000), thus our initial choice fell on this combination. Inspection of our data showed that running the model under the gamma family and log link yielded logscaled residuals that had a kurtosis value of less than 3. In addition, the raw-scale variance was nearly quadratic in the raw-scale prediction, and so according to Manning’s algorithm the gamma model provided a good fit. This resulted in the following regression equation for part two: g ( i ,t )  log(i ,t )  xiT,t    i ,t

(2.2)

where  =E[HCE | HCE>0, x ] and the mean-variance relationship is characterized by the two-parameter gamma distribution. See also McCullagh and Nelder (1989) for the generalized linear model framework. The average HCE can then be calculated by multiplying the two components: E[HCE] = P[HCE>0 | x]  E[HCE | HCE>0, x]

(2.3)

The variance-covariance matrix of each part is: Vi (  ,  ,  )  Ai (  )1/ 2 Ri ( ) Ai (  )1/ 2

(2.4)

with Ai (  ) being t  t diagonal variance matrices for regression parameters  , Ri ( ) being the working correlation matrix that is characterized by the t  1 vector  and the dispersion parameter  (Molenberghs & Verbeke, 2000). The working correlation matrix Ri ( ) needs to be specified by the modeller. For part one, we used the unstructured correlation matrix, because it imposes the fewest number of restrictions on the correlations. Models with the unstructured correlation matrix provided the lowest QIC value. Analysis was performed in Stata 9 using the xtgee command. The exchangeable structure was used for part two, as the unstructured matrix did not lead to convergence. In some cases, both the unstructured and exchangeable structures resulted in no convergence. We suspect this may have

19

CHAPTER 2

been the result of few cost observations in part one and short panels in part two (most individuals had only one year with hospital inpatient HCE). For these cases, we used the independent structure matrix. This structure is identical to running the GLM variant, except for the computation of robust variances with the Sandwich estimator (Williams, 2000). In HCE studies, it is common to add polynomial terms to model the nonlinear relationship between age and the response variables. Initial models were tested using age squared and age cubed for a small selection of diseases. Model fit was assessed by overlaying the fitted curve with the empirical means (based on the 11.25 million individuals). This proved to give poor results in some cases. The first issue was that the shape of the empirical curves showed little resemblance to what might possibly be described by second or third degree age terms. The second issue was that the empirical curves differed between the deceased and survivors, which could not be modeled properly with higher order age interactions. To address these issues, the following measures were taken. First, the data set for each disease was split into two, one for deceased and one for survivors. For each group, we then used cubic B-Splines, which are special functions defined piecewise by third degree polynomials in a variable x . The shape of the curve is dependent on the position and number of the knots, which are the points that tie the piecewise functions together. If we denote the k user-defined knots by s1 ,..., sk , then the scaleinvariant B-splines in Stata (see Newson, 2000, for the full mathematical description) are defined as B( x; s1 ,..., s k  2 )  s k  2

1

   s1    s h  s j  Pn ( x; s j ) j 1 1 h  k  2 , h  j  k 2

(2.5)

where Pn ( x; s j ) is the so-called nth power plus function at knot s j (n =3 in case of cubic splines): Pn ( x; s j ) = ( x  s j ) n

=0

if x  s j if x  s j

(2.6)

These knots s1 ,..., sk are usually best chosen at points where the curve slope (as seen in the data) is particularly steep. However, given our aim of calculating HCE for as many diseases as possible, and the fact that each disease has different age HCE patterns, picking the knots manually is a rather cumbersome task. We therefore opted to pick the knots at uniformly distributed intervals for each disease. The knot positions differ for each disease, as each disease has a different age range. The knot range s1 , s k  was set at the 1st and the 99th percentiles of age, 20

A CARPACCIO OF RED HERRINGS

to avoid problems with the small number of cases near the minimum and maximum age values. For larger ranges, more knots were selected, resulting in interval widths of 15–20 years. The splines were then regressed alongside other variables, for deceased and survivors separately. Using smaller intervals led to overfitting near the age range borders. Based on graphical checks and the QIC criterion, the splines were favored over age squared and cubic terms. Variable selection Next to the age splines, time to death (in years) and time to death squared, sex, and calendar year were included in the model. Time to death squared was included because the observed HCE in the last year of life were often considerably higher than in the other years preceding death. Splines were not used for time to death, as the time to death could only assume values from 1 to 10 years, which is a much smaller interval than the one by age. Since we model deceased and survivors separately, a specific dummy for being deceased (Werblow et al., 2007) is not needed. Calendar year was included to correct for any potential cohort effects or autonomous trends in HCE. The QIC value was lowest after having included all aforementioned variables in one model. Ratio predictions and their confidence intervals Our next step was to use these models to calculate several ratios of HCE. This study focuses on two ratios: (1) HCE during the last year of life divided by HCE of survivors (hereafter the ‘deceased/survivor ratio’) and (2) ‘ratios of successive ages’ for survivors, defined as HCE at a certain age divided by HCE incurred at an age five years younger (controlled for calendar time). Ratio (2) can be interpreted as an age gradient, giving an indication of how much HCE grows as an individual ages. The ratios are simple measures that allow a direct comparison between diseases. Moreover, they also provide a way to compare effects due to age, and effects that are due to the proximity to death. If ratio (1) is higher than ratio (2), this implies that high hospital HCE have a stronger association with the process of dying than with age itself. The more values differ from one, the stronger the relation between disease-specific hospital HCE and time to death and age, respectively. Ratios and their confidence intervals were estimated using the software package R. Confidence intervals were estimated by performing Monte Carlo simulation for each disease separately, with 10000 runs per simulation (more runs led to negligible change in confidence intervals). Within each run, the first step was to randomly draw values for the regression coefficients from a multivariate normal distribution, using the mvrnorm function in R. This distribution has means equal to the regression coefficients ˆ , and a variance-covariance matrix Vi (  ,  ,  ) that follows directly from the regression (Molenberghs & Verbeke, 2000):

21

CHAPTER 2

^   ~ MVN   ,Vi (  , , )   

(2.7)

The assumption that underlies this analysis is that all maximum likelihood-based regressions, like GEE, share the property of having multivariate normally distributed parameter estimates under asymptotic conditions. The random drawing was done for each submodel separately (part one and two, for both deceased and survivors). The second step involved making predictions in each sub-model for all possible combinations of age and sex, using the drawn coefficients from the first step. For each prediction, the calendar year was set at 2000 4 and for deceased the time to death was set at one year. In the third step, these predictions were used to calculate hospital HCE. Finally, two gender and age-specific ratios were calculated: (1) by dividing the last year of life HCE by the predicted value for survivors and (2) by dividing HCE of survivors at the end and at the beginning of (successive) 5-year time intervals. After all runs were performed, confidence intervals for (1) and (2) were determined by taking the (α/2)th and (1-α/2)th percentiles of the ratios.

2.3 Results In Table 2.2 regression results are shown for two diseases, one considered as lethal (lung cancer) and the other one as nonlethal (gonarthrosis). Both models performed significantly better than the null model according to the Wald test (p<0.0001). The proportions with HCE and conditional HCE were significantly smaller for females in the case of lung cancer, but this was not the case with gonarthrosis. No specific calendar year trend effects were found, with the exception of a decreasing trend in conditional HCE, for both diseases. The decrease in conditional HCE over time is most likely due to a decrease in the average length of stay in the hospital, which is in line with the government policy with regard to increasing the efficiency in hospital care (Statistics Netherlands, 2010). Time to death and its square were highly significant variables for lung cancer in part one, but not in part two. Gonarthrosis shows similar results, although the statistical significance of both terms in part one is not as strong. The coefficients of the regression models, in particular those for the age splines (see Table 2.A1 in Appendix 2.A), are difficult to interpret with regard to the response variables. In addition, the distinction between deceased and survivors for each response variable is not clear. Therefore, we plotted the expected values for

4

The most recent calendar year in the two-part model.

22

A CARPACCIO OF RED HERRINGS Table 2.2: Regression results for malignant neoplasms of trachea, bronchus and lung, and gonarthrosis. Malignant neoplasms of trachea, bronchus and lung Part One Part Two Variable Deceased Spline 1 Spline 2 Spline 3 Spline 4 Spline 5 Spline 6 TTD TTD^2 Female 1996 1997 1998 1999 2000

Beta

-4.413 -1.682 -1.411 -1.247 -4.121 -5.938 -0.723 -0.043 -0.807 0.220 0.197 0.163 0.123 0.158

S.E. Sign.

1.379 0.249 0.134 0.132 0.246 1.487 0.074 0.016 0.037 0.049 0.049 0.051 0.051 0.051

** ** ** ** ** ** ** ** ** ** ** ** * **

Beta

8.416 9.344 9.174 9.433 9.193 9.189 -0.034 0.008 0.133 0.006 -0.048 -0.082 -0.139 -0.150

S.E. Sign.

0.530 0.082 0.040 0.036 0.061 0.383 0.019 0.004 0.011 0.015 0.016 0.016 0.016 0.016

** ** ** ** ** **

** ** ** ** **

Gonarthrosis Part One Beta

Part Two

S.E. Sign.

-16.880 16.353 -12.565 2.089 -8.973 0.707 -9.064 0.617 -7.878 0.565 -19.547 4.094 0.824 0.270 -0.100 0.044 0.762 0.172 0.416 0.262 0.342 0.254 0.079 0.146 -0.053 0.126 0.174 0.124

** ** ** ** ** ** * **

Beta

9.397 8.524 8.654 9.339 9.416 9.190 -0.015 0.000 0.147 -0.069 -0.077 -0.122 -0.183 -0.304

S.E. Sign.

3.530 0.471 0.158 0.111 0.139 0.614 0.047 0.007 0.024 0.040 0.039 0.040 0.039 0.038

** ** ** ** ** **

**

** ** **

Survivors Spline 1 -24.169 7.393 ** 9.399 1.091 ** -13.524 4.117 ** 8.348 0.424 ** Spline 2 -9.154 0.899 ** 9.347 0.159 ** -9.597 0.536 ** 7.887 0.061 ** Spline 3 -8.675 0.323 ** 9.402 0.073 ** -8.236 0.241 ** 8.215 0.027 ** Spline 4 -6.665 0.364 ** 9.542 0.070 ** -7.068 0.199 ** 9.098 0.024 ** Spline 5 -8.949 0.965 ** 9.444 0.168 ** -5.777 0.359 ** 9.064 0.042 ** Spline 6 -15.211 7.203 * 6.854 1.607 ** -17.036 2.016 ** 10.085 0.279 ** Female -0.801 0.166 ** -0.042 0.022 * 0.697 0.073 ** 0.205 0.008 ** 1996 0.316 0.209 -0.034 0.034 0.339 0.088 ** -0.029 0.012 * 1997 0.006 0.087 -0.086 0.034 * 0.482 0.093 ** -0.076 0.012 ** 1998 0.040 0.179 -0.081 0.034 * 0.339 0.081 ** -0.132 0.012 ** 1999 0.201 0.178 -0.158 0.034 ** 0.451 0.083 ** -0.154 0.012 ** 2000 -0.208 0.084 * -0.203 0.035 ** 0.608 0.086 ** -0.229 0.012 ** Number of observations [groups] for Neoplasms of lung (Deceased part one and part two, survivors part one and part two respectively): 116350 [54560], 40606 [34564], 205507 [54576], 4937 [4635]. Number of observations [groups] for Gonarthrosis (Deceased part one and part two, survivors part one and part two respectively): 80296 [30191], 4045 [3805], 276159 [80381], 42607 [37468]. Wald’s test for Neoplasms of lung (Deceased part one and part two, survivors part one and part two respectively):

 2 (13) =46932;  2 (11) =56725;  2 (13) =3864507;  2 (11) =838540 (p<0.0001 for all). Wald’s test for Gonarthrosis (Deceased part one and part two, survivors part one and part two respectively): 2  2 (13) =26046;  2 (11) =75590;  (13) =685418;  2 (11) =5529697 (p<0.0001 for all).

Abbreviations: S.E., standard error; Sign., significance; TTD, Time to death in years. Key: *, p<0.05; **, p<0.01.

each component, for both diseases (Figure 2.2). The values were estimated for a female in the year 2000. The following things can be concluded from these graphs:  As is the case in the two-part model for total hospital HCE (Seshamani and Gray, 2004a,b,c; Zweifel et al., 2004), the proportion of individuals with HCE determines the curve of the expected HCE. Part two, the conditional HCE seems to be less influential.  The hospital HCE in the last year of life for lung cancer are clearly higher than those for gonarthrosis, for all ages. This can be explained with lung

23

CHAPTER 2 Figure 2.2: Predicted two-part model values for neoplasm of trachea, bronchus, and lung, and gonarthrosis. Dashed lines represent the deceased in their last year of life, and the solid lines represent the survivors. All values refer to females. Expenditures are in euros.

60

65

70

75

0.005 50

55

60

65

70

Malignant Neoplasm of Lung

Gonarthrosis

65

70

75

80

50

55

60

65

70

Age

Malignant Neoplasm of Lung

Gonarthrosis

65 Age

70

75

80

75

80

30 20 10

Average expenditures

400 200

60

80

40

Age

55

75

7000

Conditional expenditures 60

80

4000

10000

55

75

10000

Age

11000

Age

0 50

0.003

80

600

50

0.001

Proportion with expenditures

0.06 0.04 0.02

55

9000

Conditional expenditure

50

Average expenditures

Gonarthrosis

0.00

Proportion with expenditures

Malignant Neoplasm of Lung

50

55

60

65

70

Age

cancer being a much more common cause of death, which has a large influence on the values in part one.  Expected lung cancer HCE of deceased are higher than those of survivors, whereas the HCE for gonarthrosis show a reversed image: the HCE for survivors are higher due to a higher proportion of individuals with HCE. This proportion under survivors is very small for lung cancer (in the range between 10 -6 and 10 -4).  The conditional HCE for gonarthrosis in the last year of life are higher than those for survivors.

24

A CARPACCIO OF RED HERRINGS Figure 2.3: Predicted ratio of deceased in their last year of life and survivor HCE (left), predicted ratio of successive ages (right), and corresponding 95% confidence intervals, for malignant neoplasms of trachea, bronchus and lung (top), and gonarthrosis (bottom).

1.5 1.0 0.5

55

60

65

70

75

80

50

55

60

65

70

Age

Age

Gonarthrosis

Gonarthrosis

75

80

75

80

50

55

60

65 Age

70

75

80

1.8 1.6 1.4 1.2 1.0

0.2

0.4

0.6

0.8

Successive Age Ratio

1.0

2.0

50

Deceased/Survivor Ratio

2.0

2.5

Malignant Neoplasm of Lung

Successive Age Ratio

1500 1000 500

Deceased/Survivor Ratio

Malignant Neoplasm of Lung

50

55

60

65

70

Age

The age patterns not only clearly differ between part one and part two but also between diseases. For gonarthrosis the highest HCE are found in the age group 75–80, whereas the peak for lung cancer occurs somewhere between 60 and 65 years. Lung cancer is a disease that can also occur at younger ages, whereas gonarthrosis, like most forms of arthrosis, is a chronic disease that is especially prevalent amongst the elderly. Figure 2.3 shows estimations of the ratios obtained by Monte Carlo simulation for two diseases 5. It is evident that proximity to death is not a good predictor of high hospital HCE for all diseases, as the results for gonarthrosis show: for most ages the deceased/survivor ratio is significantly smaller than one. On the other

5

Gender differences are minimal, and so only results for females are shown here.

25

CHAPTER 2 Figure 2.4: Deceased/survivor and successive age ratios, and corresponding 95% confidence intervals, for a few other diseases. Abbreviations: TIA, transient cerebral ischemic attacks.

65

70

75

65

70

75

60

65

80

75

80

70

75

80

1.6 1.2

8 10

Age

75

70

2.0

TIA Successive Age Ratio

TIA

70

80

1.4

55

Age

65

75

1.2 50

Age

6

60

70

1.0

80

4

55

65

Diabetes

2 50

60

Diabetes 40

60

55

Age

30

55

1.5 50

Age

20 50

1.3

80

Successive Age Ratio

60

1.1

Successive Age Ratio

150 100

55

10

Deceased/Survivor Ratio

50

Deceased/Survivor Ratio

Septicaemia

50

Deceased/Survivor Ratio

Septicaemia

50

55

60

65 Age

hand, we see a very large deceased/survivor ratio for lung cancer. By contrast, the successive age ratios are higher for gonarthrosis, while being lower for lung cancer. Note that the uncertainty in the estimates for both diseases increases as age decreases; this is a natural consequence of the smaller number of deceased among younger individuals. These regression models and Monte Carlo simulations were repeated for the remainder of the selected diseases. Figures 2.4 and 2.5 show some results for other diseases. It is clear that the deceased/ survivor and successive age ratios differ strongly between diseases. Potentially lethal diseases such as septicaemia and renal failure are associated with the highest deceased/survivor ratios, whereas less lethal diseases such as transient cerebral ischemic attacks show a much smaller ratio.

26

A CARPACCIO OF RED HERRINGS Figure 2.5: Deceased/survivor and successive age ratios, and corresponding 95% confidence intervals, for some more diseases. Abbreviations: COPD, chronic obstructive pulmonary disease.

65

70

75

60

1.7

65

COPD

65

70

75

70

75

80

70

75

80

75

80

0.8

1.2

1.6

COPD

80

50

55

60

65 Age

Renal Failure

Renal Failure

65 Age

70

75

80

1.1

Successive Age Ratio

70

60

1.3

Age

50

55

60

Age

30 50

55

0.9

55

1.5 50

Age

90

50

1.3

80

Successive Age Ratio

60

1.1

Successive Age Ratio

30 20

55

10 20 30 40 50

Deceased/Survivor Ratio

50

Deceased/Survivor Ratio

Cerebrovascular

10

Deceased/Survivor Ratio

Cerebrovascular

50

55

60

65

70

Age

This becomes more evident when looking at the other diseases (Table 2.36). Shown are the ratios with corresponding confidence intervals for females (unless the disease is male-specific) at three ages 50, 65, and 807, to show the spread in ratios during life time. One can see at first glance that most ratios are significantly greater than one, for most age points. This is not only true for those diseases considered as very lethal, such as the cancers, but also for seemingly nonlethal diseases such as asthma and cholelithiasis. Thus, lethality does not seem to be a necessary condition for the ratio to be greater than one. However, lethality does One disease model, for chronic diseases of tonsils and adenoids (disease group 52), failed to converge in the regression, probably due to the small amount of cost observations among the deceased, and therefore this disease was subsequently left out from the results. 7 The age range shown here was deliberately picked as narrow, as many diseases usually fall within a limited age range. In addition, the small amount of individuals over 80 and older caused the uncertainty to be so large that estimates might have little value. Even with this narrow range some diseases showed extreme wide confidence intervals at either age 50 or 80, and these values were left out the analysis. 6

27

CHAPTER 2

seem to go hand in hand with the statistical significance and magnitude of the ratio. Common lethal diseases show a ratio greater than one, and the ones that are considered particularly lethal, such as cancers, septicaemia and renal failure, have the highest ratios. Cancers that particularly stand out are lung cancer and ovary cancer. Diseases of the circulatory system can be considered as potentially lethal, yet, after survival of the acute event, they may manifest as chronic diseases. This might explain why the ratios, although much larger than one, but much smaller than the values for cancer. Among the diseases of the circulatory system heart failure has the highest ratios, which is to be expected, since heart failure is one of the deadliest amongst heart diseases. In the Netherlands about 30% of all patients die within one year after their first admission for heart failure (Statistics Netherlands, 2010). Of the 93 diseases investigated, 17 diseases showed ratios that were not significantly greater than one for at least two age points. All of these diseases are not associated with high mortality rates. These diseases can be characterised as nonlife threatening or curable illnesses, requiring treatment (such as intervertebral disc disorders, cataract, internal derangement of knee, and benign neoplasms), or as diseases with a chronic nature (gonarthrosis and coxarthrosis), and so it is plausible that the HCE for deceased are not (significantly) higher than those for survivors, or in some cases, even significantly less. Most diseases have the highest ratios at age 50. Exceptions to this rule are diseases found only among the elderly, such as dementia and Alzheimer. Generally speaking, this coincides with the two-part models for total HCE. At high age, hospital HCE for deceased are relatively low, which results in a relatively smaller ratio. The lower HCE at advanced age might be due to the tendency to treat the elderly less intensively than would be done at lower ages in similar conditions (Long and Marshall, 2000). Alternatively, it could be due to a substitution of hospital care by long term care (McGrail et al., 2000; Spillman and Lubitz, 2000; Hogan et al., 2001), or to the simple fact that the elderly are more frail and succumb quicker to a serious disease. Having controlled for proximity to death, we can examine the ‘pure’ influence of age (Table 2.4 8). Successive age ratios for surviving females are presented here. They are evaluated by comparing HCE at ages 70,75,80 and dividing them by HCE at ages 65,70,75, respectively. We present the results for more advanced ages, as these are most informative in this context (i.e., to study the role of ageing). In contrast to the findings of studies that have focused on total hospital HCE (Seshamani and Gray, 2004a,b,c), we find that for many separate diseases the HCE

In line with previous tables, model predictions for some diseases at age 80 have extreme wide intervals and are not shown in the table.

8

28

A CARPACCIO OF RED HERRINGS Table 2.3: Estimated disease-specific deceased/survivor ratios for females at ages 50, 65, and 80. Ratio at age No

Disease

1

Intestinal infectious diseases except diarrhoea

2

50

65

80

5.94

*

4.19

**

2.68

Diarrhoea and gastroenteritis of presumed infectious origin

17.69

**

7.23

**

3.07

** *

3

Tuberculosis

15.32

**

7.65

**

4.11

**

4

Septicaemia

129.22

**

42.28

**

15.82

**

5

Human immunodeficiency virus [HIV] disease

100.68

**

44.94

**

–––

6

Other infectious and parasitic diseases

7

Malignant neoplasm of colon, rectum and anus

8

Malignant neoplasms of trachea, bronchus and lung

9 10

24.68

**

11.05

**

5.43

197.38

**

45.53

**

13.06

**

1028.03

**

295.33

**

146.35

**

Malignant neoplasms of skin

28.50

**

10.16

**

3.30

**

Malignant neoplasm of breast

28.59

**

11.94

**

4.00

**

11

Malignant neoplasm of uterus

62.22

**

21.66

**

7.12

**

12

Malignant neoplasm of ovary

198.41

**

97.26

**

55.08

**

13

Malignant neoplasm of prostate

39.21

**

6.69

**

14.06

**

14

Malignant neoplasm of bladder

126.44

**

40.68

**

15.77

**

15

Other malignant neoplasms

512.35

**

202.69

**

66.50

**

16

Carcinoma in situ

1.10

1.03

0.62

#

17

Benign neoplasm of colon, rectum and anus

1.74

1.58

1.24

18

0.51

2.86

22

Leiomyoma of uterus Other benign neoplasms and neoplasms of uncertain or unknown behaviour Anaemias Other diseases of the blood and bloodforming organs and certain disorders involving the immune mechanism Diabetes mellitus

23

Other endocrine, nutritional and metabolic diseases

24

Dementia

–––

25

17.72 9.72

27

Mental and behavioural disorders due to alcohol Mental and behavioural disorders due to use of other psychoactive subst. Schizophrenia, schizotypal and delusional disorders

28

Mood [affective] disorders

29 30

19 20 21

26

**

–––

13.12

**

10.10

**

6.91

**

73.13

**

28.05

**

9.47

**

76.01

**

39.77

**

12.19

**

36.10

**

20.63

**

10.26

**

19.08

**

14.88

**

9.45

**

19.40

**

10.18

**

**

8.04

**

–––

**

8.12

**

3.51

**

5.68

**

5.06

**

5.92

**

5.28

**

3.08

**

2.24

**

Other mental and behavioural disorders

10.25

**

Alzheimer's disease

–––

31

Multiple sclerosis

10.64

32

Epilepsy

53.22

33

Transient cerebral ischaemic attacks and related syndromes

34

Other diseases of the nervous system

35 36 37

Diseases of the ear and mastoid process

38

Hypertensive diseases

39

Angina pectoris

40

Acute myocardial infarction

41

Other ischaemic heart disease

42 43 44

Heart failure

45 46 47

Varicose veins of lower extremities

48

Other diseases of the circulatory system

13.55

**

9.49

**

16.23

**

77.84

**

**

6.39

**

22.09

**

22.07

**

10.84

**

6.84

**

5.79

**

2.58

**

12.61

**

11.13

**

5.61

**

Cataract

1.68

*

0.67

##

0.39

##

Other diseases of the eye and adnexa

1.28

0.89

0.88

0.81

–––

0.69

14.84

**

13.25

**

7.45

**

3.69

**

2.63

**

2.02

**

19.90

**

12.65

**

8.55

**

6.97

**

5.37

**

2.51

**

Pulmonary heart disease & diseases of pulmonary circulation

30.38

**

14.54

**

5.86

**

Conduction disorders and cardiac arrhythmias

13.28

**

7.19

**

3.62

**

134.39

**

47.71

**

20.59

**

Cerebrovascular diseases

27.96

**

18.97

**

11.98

**

Atherosclerosis

17.64

**

12.35

**

6.96

**

1.14 18.84

–––

1.49 **

10.29

**

5.67

**

29

CHAPTER 2 Table 2.3 (continued). Ratio at age No

Disease

49

Acute upper respiratory infections and influenza

31.33

50 **

13.68

65 **

5.88

80 **

50

Pneumonia

36.08

**

22.80

**

9.80

**

51

Other acute lower respiratory infections

39.22

**

20.04

**

8.80

**

52

Chronic diseases of tonsils and adenoids

–––

–––

53

Other diseases of upper respiratory tract

1.53

2.26

**

––– –––

54

Chronic obstructive pulmonary disease and bronchiectasis

55

Asthma

39.78

**

24.25

**

12.21

8.09

**

8.16

**

4.82

56

Other diseases of the respiratory system

**

54.33

**

30.50

**

13.22

**

57

Disorders of teeth and supporting structures

58

Other diseases of oral cavity, salivary glands and jaws

11.68

**

6.34

**

3.21

**

59

Diseases of oesophagus

30.74

**

12.53

**

5.12

**

60

Peptic ulcer

46.92

**

19.37

**

9.81

**

61

Dyspepsia and other diseases of stomach and duodenum

29.05

**

14.15

**

7.99

**

1.03

**

–––

1.12

62

Diseases of appendix

1.97

63

Inguinal hernia

0.43

##

1.35 0.61

##

2.21 0.55

64

Other abdominal hernia

3.78

**

2.70

**

1.73

65

Crohn's disease and ulcerative colitis

8.35

**

7.30

**

5.34

**

66

Other noninfective gastroenteritis and colitis

39.33

**

21.02

**

9.01

**

67

Paralytic ileus and intestinal obstruction without hernia

46.27

**

18.45

**

7.09

**

68

Diverticular disease of intestine

7.08

**

7.35

**

4.50

**

69

Diseases of anus and rectum

6.20

**

7.08

**

4.59

**

70

Other diseases of intestine

30.40

**

15.85

**

6.51

**

##

71

Alcoholic liver disease

185.88

**

90.65

**

17.60

**

72

Other diseases of liver

91.04

**

47.13

**

17.92

**

73

Cholelithiasis

1.84

**

1.85

**

1.50

*

74

Other diseases of gall bladder and biliary tract

22.45

**

11.09

**

7.36

**

75

Diseases of pancreas

44.85

**

23.86

**

9.11

**

76

Other diseases of the digestive system

34.27

**

19.91

**

9.77

**

77

Infections of the skin and subcutaneous tissue

9.72

**

9.51

**

4.40

**

78

Dermatitis, eczema and papulosquamous disorders

2.88

**

2.42

**

1.27

79

Other diseases of the skin and subcutaneous tissue

11.88

**

9.12

**

5.07

**

80

Coxarthrosis [arthrosis of hip]

0.53

0.37

##

0.28

##

81

Gonarthrosis [arthrosis of knee]

0.54

##

0.33

##

0.20

##

82

Internal derangement of knee

0.28

##

0.26

##

–––

83

Other arthropathies

2.80

**

3.61

**

1.74

**

27.63

**

12.70

**

5.47

**

84

Systemic connective tissue disorders

85

Deforming dorsopathies and spondylopathies

2.37

86

Intervertebral disc disorders

1.23

87

Dorsalgia

5.58

**

5.93

**

3.71

**

88

12.30

**

5.84

**

5.70

**

5.64

**

6.74

**

3.67

**

90

Soft tissue disorders Other disorders of the musculoskeletal system and connective tissue Glomerular and renal tubulo-interstitial diseases

20.28

**

14.77

**

6.38

**

91

Renal failure

71.10

**

56.18

**

34.77

**

92

Urolithiasis

2.20

**

2.90

**

1.70

93

Other diseases of the urinary system

7.30

**

6.00

**

89

94 Hyperplasia of prostate 0.24 Key: *, ratio greater than one with p<0.05; **, ratio greater than one with p<0.01, #, ratio smaller than one with p<0.05; ##, ratio smaller than one with p<0.01.

30

1.96

0.60

–––

1.05

0.76

5.00 0.68

**

A CARPACCIO OF RED HERRINGS Table 2.4: Estimated disease-specific successive age ratios for females evaluated at ages 70, 75 and 80. Ratio at age No

Disease

1

Intestinal infectious diseases except diarrhoea

1.28

**

1.43

**

1.34

2

Diarrhoea and gastroenteritis of presumed infectious origin

1.29

**

1.58

**

1.31

*

3

Tuberculosis

1.21

**

1.34

*

1.35

**

4

Septicaemia

1.44

**

1.32

**

1.20

**

5

Human immunodeficiency virus [HIV] disease

0.30

6

Other infectious and parasitic diseases

1.18

**

1.22

**

1.22

7

Malignant neoplasm of colon, rectum and anus

1.33

**

1.35

**

1.29

**

8

Malignant neoplasms of trachea, bronchus and lung

1.20

**

0.78

0.47

##

9

Malignant neoplasms of skin

1.59

**

1.49

**

1.27

**

10

Malignant neoplasm of breast

1.13

**

1.15

*

0.98

11

Malignant neoplasm of uterus

1.00

1.03

12

Malignant neoplasm of ovary

0.96

0.71

##

0.60

##

13

Malignant neoplasm of prostate

1.00

0.78

##

0.80

##

14

Malignant neoplasm of bladder

1.42

**

1.21

**

1.02

15

Other malignant neoplasms

1.15

**

1.08

16

Carcinoma in situ

1.26

**

1.15

*

0.90

17

Benign neoplasm of colon, rectum and anus

1.38

**

1.30

**

1.03

18

1.25

7.79

**

–––

22

Leiomyoma of uterus Other benign neoplasms and neoplasms of uncertain or unknown behaviour Anaemias Other diseases of the blood and bloodforming organs and certain disorders involving the immune mechanism Diabetes mellitus

23

Other endocrine, nutritional and metabolic diseases

24

Dementia

25

1.29

27

Mental and behavioural disorders due to alcohol Mental and behavioural disorders due to use of other psychoactive subst. Schizophrenia, schizotypal and delusional disorders

28

Mood [affective] disorders

0.87

29

Other mental and behavioural disorders

30

Alzheimer's disease

31

Multiple sclerosis

0.55

##

0.50

#

0.50

32

Epilepsy

1.34

**

1.33

*

1.13

33

Transient cerebral ischaemic attacks and related syndromes

1.45

**

1.65

**

1.53

**

34

Other diseases of the nervous system

1.23

**

1.31

**

1.08

*

35

Cataract

1.97

**

1.75

**

1.41

**

36

Other diseases of the eye and adnexa

1.23

**

1.21

**

1.07

**

37

Diseases of the ear and mastoid process

0.94

38

Hypertensive diseases

1.13

1.26

1.02

39

Angina pectoris

1.23

**

1.06

0.83

40

Acute myocardial infarction

1.24

**

1.15

41

Other ischaemic heart disease

1.24

**

1.04

42

Pulmonary heart disease & diseases of pulmonary circulation

1.33

**

1.30

**

1.20

43

Conduction disorders and cardiac arrhythmias

1.44

**

1.38

**

1.25

**

44

Heart failure

1.56

**

1.48

**

1.43

**

45

Cerebrovascular diseases

1.54

**

1.40

**

1.14

**

46

Atherosclerosis

1.24

**

1.22

**

1.04

19 20 21

26

70

75

80 **

–––

0.37

**

1.07

0.97 #

1.11

**

1.15

**

0.97

1.69

**

1.66

**

1.54

1.10

*

1.03

1.38

**

1.26

**

1.25

**

1.33

**

1.36

**

–––

2.08

**

1.91

**

0.88

0.57

##

–––

*

1.37

**

1.35

1.08

*

1.34

**

##

0.84

#

0.82

##

0.93

1.25

**

1.38

**

–––

0.83

0.92

**

0.94 1.10

0.78

–––

0.89

**

##

1.03 0.73

47

Varicose veins of lower extremities

0.80

##

0.86

#

–––

48

Other diseases of the circulatory system

1.33

**

1.15

**

0.92

##

##

31

CHAPTER 2 Table 2.4 (continued). Ratio at age No

Disease

49

Acute upper respiratory infections and influenza

1.60

70 **

75

50

Pneumonia

1.57

51

Other acute lower respiratory infections

1.32

80

1.34

**

**

1.46

**

1.27

**

1.21

**

1.08

–––

1.13

52

Chronic diseases of tonsils and adenoids

–––

53

Other diseases of upper respiratory tract

0.89

#

0.78

##

––– –––

54

Chronic obstructive pulmonary disease and bronchiectasis

1.29

**

1.07

*

0.90

55

Asthma

0.90

56

Other diseases of the respiratory system

1.18

** ##

1.06

**

##

1.33

1.16

**

1.11

0.66

#

–––

57

Disorders of teeth and supporting structures

0.68

58

Other diseases of oral cavity, salivary glands and jaws

0.99

59

Diseases of oesophagus

1.32

**

1.43

**

1.47

60

Peptic ulcer

1.38

**

1.40

**

1.45

**

61

Dyspepsia and other diseases of stomach and duodenum

1.29

**

1.36

**

1.39

**

62

Diseases of appendix

1.33

**

1.24

**

0.75

63

Inguinal hernia

1.30

**

1.28

**

64

Other abdominal hernia

1.19

**

1.11

1.01

1.03

1.13

**

**

1.05

65

Crohn's disease and ulcerative colitis

1.08

*

1.13

*

0.93

66

Other noninfective gastroenteritis and colitis

1.56

**

1.42

**

1.25

**

67

Paralytic ileus and intestinal obstruction without hernia

1.53

**

1.41

**

1.24

*

68

Diverticular disease of intestine

1.34

**

1.27

**

1.19

**

69

Diseases of anus and rectum

1.06

1.26

**

70

Other diseases of intestine

1.26

1.48

**

71

Alcoholic liver disease

0.79

0.79

0.83

72

Other diseases of liver

0.98

1.04

1.10

73

Cholelithiasis

1.12

**

1.19

**

1.16

74

Other diseases of gall bladder and biliary tract

1.26

**

1.27

**

1.30

**

75

Diseases of pancreas

1.05

1.17

**

1.24

**

76

Other diseases of the digestive system

1.31

1.31

**

1.34

**

77

Infections of the skin and subcutaneous tissue

1.07

1.23

**

78

Dermatitis, eczema and papulosquamous disorders

1.14

**

1.24

**

79

Other diseases of the skin and subcutaneous tissue

1.37

**

1.37

**

1.38

80

Coxarthrosis [arthrosis of hip]

1.46

**

1.20

**

0.97

81

Gonarthrosis [arthrosis of knee]

1.59

**

1.35

**

0.92

82

Internal derangement of knee

0.80

##

0.93

–––

83

Other arthropathies

1.09

*

1.07

0.96

84

Systemic connective tissue disorders

1.34

**

1.29

**

1.13

**

85

Deforming dorsopathies and spondylopathies

1.40

**

1.47

**

1.24

*

86

Intervertebral disc disorders

0.94

87

Dorsalgia

1.07

88

1.06 1.07

**

1.23

90

Soft tissue disorders Other disorders of the musculoskeletal system and connective tissue Glomerular and renal tubulo-interstitial diseases

1.18

**

1.08

0.96

91

Renal failure

1.10

**

1.07

1.06

92

Urolithiasis

1.08

*

0.99

93

Other diseases of the urinary system

1.38

**

1.34

**

1.28

94 Hyperplasia of prostate 1.57 Note: Ratios are relative to baseline age 65. Expenditures for gender-specific diseases are averages within that specific gender. Key: *, ratio greater than one with p<0.05; **, ratio greater than one with p<0.01,

**

1.28

**

1.06

89

#, ratio smaller than one with p<0.05; ##, ratio smaller than one with p<0.01.

32

1.11 **

**

1.45

**

1.14

1.13

1.08 ** #

–––

0.94 *

**

*

0.92 **

1.18

**

0.86

#

1.24

**

0.91 **

A CARPACCIO OF RED HERRINGS

increase significantly with age in itself. Thus, the successive age ratios are often greater than one. For successive age ratios evaluated at ages 70 and 75 years, this was the case for more than 60 out of 93 diseases, while for age 80 there were 40 out of 83 diseases with a ratio higher than one. It should be noted that while many of these diseases show considerable increases in HCE over age, the increases are relative to the survivor HCE at 5 years before that age, and for most diseases, these absolute HCE are rather modest. On the other hand, 6, 8, and 11 diseases, respectively, have successive age ratios significantly smaller than one. Seven diseases have a successive age ratio significantly smaller than one for at least two ages (neoplasms of ovary and of prostate, mood disorders, multiple sclerosis, varicose veins of lower extremities, other disease of upper respiratory tract, and disorders of teeth and supporting structures). When comparing these ratios to the deceased/survivors ratios, we can conclude that the proximity to death ratios are much higher, and that proximity to death, therefore, is a much better predictor, on aggregate level, of high HCE.

2.4 Discussion This article builds on the red herring debate. The main novelty of this paper is that it investigates the red herring claim for disease-specific hospital HCE. The first major conclusion is that proximity to death is an important predictor of high HCE for most, but not all diseases. The majority of the diseases do have a ratio of deceased/survivor HCE significantly greater than one, of which many are, perhaps surprisingly, not immediately associated with high mortality risk. Strong lethality is thus not a prerequisite for a positive influence of proximity to death on hospital HCE, but it is associated with a higher ratio. The greatest ratios were found for the most lethal diseases such as lung cancer, septicaemia, heart, and renal failure. The diseases where proximity to death was not a good predictor of high HCE had, when adequately treated, a nonlife threatening nature, and were often either chronic or only required planned nonurgent inpatient treatment. Secondly, in contrast to common views about total hospital inpatient HCE, for most diseases age did significantly influence HCE of survivors. However, in terms of the consequences of ageing, each of these diseases has a limited impact because they are associated with a modest level of average HCE. The influence of age also seems to be modest in comparison to proximity to death. This study allows us to interpret time to death analyses in a different manner. Although time to death is a much better predictor of HCE than age, it must be realized that just like age, time to death approximates underlying processes (Gray, 2005). These processes, of course, are mainly due to the presence of disease, 33

CHAPTER 2

particularly in the case of hospital HCE. The simple mechanism of the presence of diseases lowering chances of one’s survival plays a role. It can be argued that time to death is a rather crude approximation of this decline in health. This perception sheds a different light on the traditional two-part red herring model. Part one can be interpreted as the proportion of individuals having one or more disease(s), and also utilizing health care for the disease(s). This proportion is higher in the last years of life for most diseases, as we see in this paper. Moreover, conditional on having a disease and utilizing care for it, as the severity of diseases is greater toward the end of life, treatment is in most cases more intensive in the last years of life. The severity is described by part two of the two-part model. The results of this article may also have implications outside the red herring context. While time to death is a better predictor than age, a lot of variation is found in HCE between diseases (Polder et al., 2006). Disease-specific HCE estimates may be used to investigate the consequences of changes in epidemiology, such as the shift from heart diseases to cancer as the major cause of death. Dormont et al. (2006) estimated HCE for a small selection of diseases and found that changes in morbidity led to savings in HCE which offset the increase in spending due to ageing. The results may also be used to investigate the HCE of healthy ageing (Lubitz et al., 2003). A hypothesis would be that as people age healthier, the onset of diseases is postponed, thus skewing lifetime HCE even more towards the last years of life (compression of morbidity). Our results indicate this proposition might hold true for most diseases. Health economists might also be interested in disease-specific HCE for cost effectiveness analyses or for analyses concerning HCE in life years gained as a result of preventive measures (Gandjour and Lauterbach, 2005; van Baal et al., 2007). Finally, this study shows that a hospital register provides insights different from what insurance data might contribute, as it provides extensive diagnosis information and also has the luxury of having a large sample size. This study does not come without limitations. First of all, the data deals only with hospital inpatient care. Other health care sectors, such as long-term care and general practitioner care, may show different patterns for some diseases. The diseases in this study clearly have a more acute nature. Because this study only addresses a small proportion of all HCE, results may not be generalized to total HCE including all health care sectors. Secondly, confounding variables such as frailty, disability, and comorbidity (Schwartz et al., 1996; Fried et al., 2004) were not included in this study. Frailty and disability data were not readily available for the whole population, and while extensive diagnostic information was available, modeling comorbidity is an extensive and complicated task that falls outside the scope of this article. In addition, the results are heavily dependent on the diagnosis coding used. We used the ISHMT format, which is used by Eurostat, to improve comparability between European states. Clinicians might prefer the ICD-10 34

A CARPACCIO OF RED HERRINGS

format, however, which generally features smaller groupings, and as a result, the proportion of individuals having such a disease and utilizing care for it is smaller. This means the average HCE will differ for a lot of diseases when calculated on ICD-10 level. Finally, there is the issue of endogeneity of proximity to death (Salas and Raftery, 2001; Terza et al., 2008). To our knowledge, the severity of this issue has of now yet been unresolved in HCE analyses.

2.5 Conclusion Proximity to death is a significant predictor of high hospital HCE for most, but not all, diseases. Age, while significantly influencing disease-specific hospital HCE, has a much more limited impact in most cases. Exceptions are diseases which are commonly viewed as nonlethal and prevalent among elderly, such as arthrosis of hip and knee. Considering the large variation in the degree to which proximity to death and age matter, we support the notion that time to death and age are crude estimators of high HCE, as they approximate processes that really matter, namely health status and particularly in the case of hospital HCE, the presence of diseases. We therefore dub this the ‘carpaccio of red herrings’.

35

CHAPTER 2

Appendix 2.A: Additional regression output Table 2.A1: Spline values for malignant neoplasms of lung and gonarthrosis. Malignant neoplasms of lung

Gonarthrosis

Spline Age

1

2

3

Spline 4

5

6

1

2

3

4

5

6

50 0.005 0.345 0.593 0.057 0.000 0.000 0.000 0.167 0.667 0.167 0.000 0.000 51 0.003 0.303 0.620 0.074 0.000 0.000 0.000 0.140 0.664 0.196 0.000 0.000 52 0.001 0.262 0.642 0.094 0.000 0.000 0.000 0.117 0.655 0.228 0.000 0.000 53 0.000 0.223 0.659 0.118 0.000 0.000 0.000 0.097 0.641 0.262 0.001 0.000 54 0.000 0.187 0.668 0.145 0.000 0.000 0.000 0.078 0.623 0.297 0.002 0.000 55 0.000 0.154 0.670 0.176 0.000 0.000 0.000 0.063 0.600 0.333 0.004 0.000 56 0.000 0.126 0.664 0.211 0.000 0.000 0.000 0.049 0.574 0.370 0.006 0.000 57 0.000 0.101 0.650 0.249 0.000 0.000 0.000 0.038 0.545 0.407 0.010 0.000 58 0.000 0.079 0.630 0.289 0.001 0.000 0.000 0.029 0.513 0.444 0.015 0.000 59 0.000 0.061 0.605 0.332 0.003 0.000 0.000 0.021 0.479 0.479 0.021 0.000 60 0.000 0.046 0.574 0.375 0.006 0.000 0.000 0.015 0.444 0.513 0.029 0.000 61 0.000 0.033 0.539 0.418 0.010 0.000 0.000 0.010 0.407 0.545 0.038 0.000 62 0.000 0.023 0.501 0.460 0.016 0.000 0.000 0.006 0.370 0.574 0.049 0.000 63 0.000 0.016 0.460 0.501 0.023 0.000 0.000 0.004 0.333 0.600 0.063 0.000 64 0.000 0.010 0.418 0.539 0.033 0.000 0.000 0.002 0.297 0.623 0.078 0.000 65 0.000 0.006 0.375 0.574 0.046 0.000 0.000 0.001 0.262 0.641 0.097 0.000 66 0.000 0.003 0.332 0.605 0.061 0.000 0.000 0.000 0.228 0.655 0.117 0.000 67 0.000 0.001 0.289 0.630 0.079 0.000 0.000 0.000 0.196 0.664 0.140 0.000 68 0.000 0.000 0.249 0.650 0.101 0.000 0.000 0.000 0.167 0.667 0.167 0.000 69 0.000 0.000 0.211 0.664 0.126 0.000 0.000 0.000 0.140 0.664 0.196 0.000 70 0.000 0.000 0.176 0.670 0.154 0.000 0.000 0.000 0.117 0.655 0.228 0.000 71 0.000 0.000 0.145 0.668 0.187 0.000 0.000 0.000 0.097 0.641 0.262 0.001 72 0.000 0.000 0.118 0.659 0.223 0.000 0.000 0.000 0.078 0.623 0.297 0.002 73 0.000 0.000 0.094 0.642 0.262 0.001 0.000 0.000 0.063 0.600 0.333 0.004 74 0.000 0.000 0.074 0.620 0.303 0.003 0.000 0.000 0.049 0.574 0.370 0.006 75 0.000 0.000 0.057 0.593 0.345 0.005 0.000 0.000 0.038 0.545 0.407 0.010 76 0.000 0.000 0.043 0.562 0.387 0.009 0.000 0.000 0.029 0.513 0.444 0.015 77 0.000 0.000 0.031 0.526 0.429 0.014 0.000 0.000 0.021 0.479 0.479 0.021 78 0.000 0.000 0.022 0.488 0.469 0.021 0.000 0.000 0.015 0.444 0.513 0.029 79 0.000 0.000 0.015 0.448 0.508 0.030 0.000 0.000 0.010 0.407 0.545 0.038 80 0.000 0.000 0.009 0.406 0.544 0.041 0.000 0.000 0.006 0.370 0.574 0.049 Note: List of B-Spline values for malignant neoplasms of lung, and gonarthrosis. These covariate values can be used in conjunction with the regression coefficients to make model predictions. Values are equivalent for part one and part two of the model. Splines with a lower number correspond with lower ages, and splines with a higher number correspond with a higher age interval.

36

Chapter 3 Standardizing the Inclusion of Indirect Medical Costs in Economic Evaluations ‡

A shortcoming of many economic evaluations is that they do not include all medical costs in life-years gained (also termed indirect medical costs). One of the reasons for this is the practical difficulties in the estimation of these costs. While some methods have been proposed to estimate indirect medical costs in a standardized manner, these methods fail to take into account that not all costs in life-years gained can be estimated in such a way. Costs in life years gained caused by diseases related to the intervention are difficult to estimate in a standardized manner and should always be explicitly modeled. However, costs of all other (unrelated) diseases in life-years gained can be estimated in such a way. We propose a conceptual model of how to estimate costs of unrelated diseases in life-years gained in a standardized manner. Furthermore, we describe how we estimated the parameters of this conceptual model using various data sources and studies conducted in the Netherlands. Results of the estimates are embedded in a software package called ‘Practical Application to Include future Disease costs’ (PAID 1.0). PAID 1.0 is available as a Microsoft® Excel tool and enables researchers to ‘switch off’ those disease categories that were already included in their own analysis and to estimate future healthcare costs of all other diseases for incorporation in their economic evaluations. We assumed that total healthcare expenditure can be explained by age, sex and time to death, while the relationship between costs and these three variables differs per disease. To estimate values for age- and sex-specific per capita health expenditure per disease and healthcare provider stratified by time to death we used ‡

This chapter is based on: Van Baal PHM, Wong A, Slobbe LCJ, Polder JJ, Brouwer WBF, de Wit GA. 2011. Pharmacoeconomics 29(3), 175-187. This article has been reproduced with permission of Adis Data Information BV. © Adis Data Information BV, 2011. All rights reserved.

CHAPTER 3

Dutch cost-of-illness (COI) data for the year 2005 as a backbone. The COI data consisted of age- and sex-specific per capita health expenditure uniquely attributed to 107 disease categories and eight healthcare provider categories. Since the Dutch COI figures do not distinguish between costs of those who die at a certain age (decedents) and those who survive that age (survivors), we decomposed average per capita expenditure into parts that are attributable to decedents and survivors, respectively, using other data sources.

3.1 Introduction Life-saving (or death-postponing) interventions induce medical consumption in socalled life-years gained (LYG). This medical consumption in LYG has also been labeled as ‘indirect’ medical costs and in the theoretical economic literature a further distinction has been made between related and unrelated medical costs in LYG (Garber and Phelps, 1997). Subsequently, there has been discussion as to whether all of this medical consumption in LYG (related and unrelated) should be included in economic evaluations (Gerber and Phelps, 1997; Melzer et al., 2000; Nyman, 2004; van Baal et al., 2007; Lee, 2008; Melzer, 2008; Feenstra et al., 2008; Rappange et al., 2008) In practice, as prescribed in (pharmacoeconomic) guidelines (National Institute for Health and Clinical Excellence, 2008; College voor Zorgverzekeringen, 2006), many economic evaluations do take into account those costs in LYG that are related to the intervention under evaluation, while ignoring other medical costs altogether. However, the costs that are termed ‘related’ and therefore included, in practice do not necessarily adhere to the definitions of ‘related’ and ‘unrelated’ employed in the theoretical literature (van Baal et al., 2011a). In the practice of economic evaluations, related costs are usually defined on the level of diseases, and only the costs occurring in LYG of diseases to which the intervention is targeted are taken into account. For instance, in an economic evaluation of statins for the prevention of cardiovascular disease, usually all costs of future cardiovascular disease are included and costs of all other diseases in LYG are excluded. In an evaluation of a colorectal cancer screening program, only the future (averted) costs of colorectal cancer are included. However, if these interventions result in gains in life expectancy, it is likely that costs for other diseases, besides the diseases to which the intervention is targeted, will occur, so that the cost effectiveness might change (Van Baal et al., 2007; Manns et al., 2003; de Kok et al., 2009; van den Berg et al., 2008; van Baal et al., 2008). Theoretically, the distinction between related and unrelated has nothing to do with diseases, and costs are only unrelated if they, conditional on reaching a certain age, are independent of the intervention (Garber and Phelps, 1997). In the above38

STANDARDIZING INDIRECT MEDICAL COSTS

mentioned examples on cardiovascular disease and colorectal cancer, some of these disease-specific costs may be theoretically called related while others are not. Furthermore, costs of other diseases which are not included in the economic evaluation may also be partly related. Besides the lack of consensus regarding the theoretical appropriateness of including future medical costs, we think that an important reason why many guidelines still do not advocate the inclusion of all future medical costs is the lack of practical tools to facilitate their inclusion. Since economic evaluations of lifesaving (preventive and curative) interventions are conducted in a variety of settings, including those of new (and often expensive) drugs, a standardized way to account for medical costs in LYG is of great importance. However, the question then becomes ‘how can we standardize the inclusion of indirect medical costs?’. The simplest way to include indirect medical costs in a standardized way is to multiply age-specific per capita medical consumption with the LYG in an economic evaluation. For example, if an intervention causes a person to die at his/her 80th birthday instead of his/her 79th, the indirect medical costs are then estimated by simply taking the average per capita health expenditure of a 79-yearold person. However, adding age-specific average per capita health consumption has been shown to result in biased estimates of the apparent costs of aging. Zweifel et al. (1999) were the first to conclude, using longitudinal Swiss sick fund data, that healthcare expenditure depend on time to death, rather than time since birth (age). Higher average healthcare costs at a higher age are caused mainly by the fact that elderly people die, with associated high healthcare utilization in the period just before dying. The role of proximity to death (also known as the ‘red-herring’ hypothesis) has been confirmed in other studies (Polder et al., 2006; Seshamani and Gray, 2004a; Häkkinen et al., 2008; O’Neill et al., 2000). Further research revealed that the strength of the proximity to death effect differed starkly between healthcare providers (O’Neill et al., 2000; Werblow et al., 2007) and between different diseases. Wong et al. (2011a) found that the time-to-death effect was strongest for the most lethal diseases such as lung cancer, septicaemia, heart failure and renal failure. The diseases where the time-to-death effect could not be found had a non-life-threatening nature, and were either chronic or only had planned non-urgent inpatient treatment. Gandjour and Lauterbach (2005) were the first to link the ‘red herring’ literature to the practice of economic evaluations. By modeling total per capita health expenditure conditional on age and proximity to death they demonstrated that cost effectiveness analyses overestimate the incremental cost-effectiveness ratio (ICER) of preventive interventions if they do not explicitly model the high costs of the last year of life, as these costs are only postponed by prevention. Although Gandjour and Lauterbach (2005) showed that adding age-specific per capita costs without 39

CHAPTER 3

accounting for the high expenditure near death results in overestimates of medical costs in LYG. Their approach also has its limitations since it cannot be combined with most economic evaluations in practice. This holds because costs of related diseases are already included in most economic evaluations, and therefore it is incorrect to add all medical costs in LYG to the ICER, even when corrected for the costs in the last year of life (van Baal et al., 2011). Clearly, only the costs of all other (unrelated) diseases should be included. One solution to this problem may be to not model the costs of related diseases in cases of life-saving interventions. Then, the approach proposed by Gandjour and Lauterbach (2005) could be considered appropriate. However, simply adding per capita costs stratified by age and proximity to death ignores cost differences between diseases as well as the fact that some of these per capita costs will indeed significantly change as a result of the intervention. A successful weight loss intervention will change the per capita expenditure on diabetes and cardiovascular disease. However, it will probably not alter the expenditure on dementia, for instance. A colorectal cancer screening prevention will probably influence future spending on colorectal cancer (and possibly also other types of cancer) but will not influence future spending on all other diseases. Therefore, while it is impossible to standardize the inclusion of indirect medical costs for all diseases and for all interventions, it might be possible to standardize inclusion of the indirect medical costs of all other diseases besides the diseases to which the intervention is targeted. This would make ICER estimates more precise as well as improve the comparability of results of different economic evaluations. We describe a methodology that can be used to include costs of unrelated diseases during LYG in a standardized manner in economic evaluations. This methodology has been implemented in a toolkit designed to facilitate inclusion of indirect medical costs in economic evaluations in practice in the Netherlands: the ‘Practical Application to Include future Disease costs’ (PAID 1.0) [available as Supplemental Digital Content 1, http://links.adisonline.com/PCZ/A95. Important: be sure to enable the macros embedded in PAID 1.0. If these are disabled, PAID 1.0 does not function properly; please also download the manual (Supplemental Digital Content 2) from http://links.adisonline.com/PCZ/A96]. This tool enables researchers to incorporate indirect medical costs in their economic evaluation in a tailor-made fashion. De-pending on the diseases for which costs are already included in the basic economic evaluation, future costs of all other diseases can be added using PAID 1.0, in combination with the survivor curves from the basic economic evaluation. This article highlights the methodology underlying PAID 1.0. In the following section, we explain the conceptual model and the data sources and methodology used to estimate the parameters of the conceptual model behind PAID 1.0. In the results section, we present the estimated model parameters embedded in PAID 1.0. 40

STANDARDIZING INDIRECT MEDICAL COSTS

3.2 Methodology underlying the PAID model Suppose someone conducts an economic evaluation of a stroke care intervention resulting in a substantial increase in life expectancy. The costs of stroke have already been estimated in this study as they are expected to change due to the intervention. An important question in that context obviously is how the costs of all other diseases should then be estimated? Conceptual framework If the goal is to develop a general framework to estimate the costs of all diseases not directly related to an intervention, it is convenient to start by breaking total healthcare expenditure down by diseases. Conceptually, lifetime healthcare costs are then the sum of disease-specific expenditure one incurs throughout life. Since disease-specific expenditure is strongly determined by age and time to death (Wong et al., 2011a), individual lifetime healthcare costs can be estimated using equation 3.1: n 1

lhc( g )    sci (a, g )   dci (n, g ) a

i

(3.1)

i

where lhc(g) = lifetime healthcare costs for an individual of sex g; a=age in years; n=age at death; dc=decedent costs (per capita healthcare costs in the last year of life); sc=survivor costs (per capita healthcare costs in all other years); and i = index for diseases. Equation 1 simply states that individual healthcare expenditure are the sum of per capita disease-specific expenditure in the last year of life and ‘other’ years, and can be thought of as lifetime health expenditure if the current health expenditure pattern would remain constant. Now suppose an intervention that increases life expectancy influences the health expenditure for Z, a set of related diseases. The costs of all other diseases can then be simply estimated by summing over the remaining disease categories (equation 3.2): n 1

  sc (a, g )   dc (n, g ) i

a

iZ

i

(3.2)

iZ

where Z=the set of related diseases. By first breaking down lifetime healthcare expenditure into disease components, it is simple to exclude costs of certain diseases to avoid double counting of costs and to model the costs of those diseases for which treatment patterns are expected to change separately. 41

CHAPTER 3

The toolkit PAID 1.0 contains estimates of age- and sex-specific costs for a range of diseases stratified by last year of life and other years as in equation 3.1. PAID 1.0 is available as a Microsoft® Excel tool (see the Supplemental Digital Content 1) and enables researchers to select the diseases whose costs are already modeled and therefore should be excluded to calculate per capita costs for all other diseases as in equation 3.2. The costs of all other diseases as estimated with PAID 1.0 can then be combined with the survivor curves of the intervention and comparator under study to estimate differences in costs of unrelated diseases. The number of survivors in the scenarios can be multiplied with survivor costs of unrelated diseases estimated by PAID 1.0 and the number of deaths in both scenarios can be multiplied by the decedent costs of unrelated diseases estimated by PAID 1.0 (see the PAID 1.0 user manual available as Supplemental Digital Content 2 for more details on how to use PAID 1.0). Estimating the input of PAID 1.0: Disease-specific per capita health expenditure stratified by last year of life and ‘other years’ To produce consistent estimates of disease-specific per capita costs for decedents (costs per capita in the last year of life) and survivors (costs per capita in all other years) as in equation 3.1, we combined information from several data sources. As the backbone we used cost-of-illness (COI) data for the Netherlands in 2005 (Poos et al., 2008). In that study, the 2005 total direct healthcare costs in different healthcare settings in the Netherlands were uniquely attributed to 107 disease categories (including the remainder category ‘not disease related’) and eight healthcare provider categories, specified by sex and 21 age classes. Appendix A in the Supplemental Digital Content 3 (http://links. adisonline.com/PCZ/A97) displays tables of the health providers (Table A1) and diseases (Table A2) distinguished in the 2005 COI study. This was a sequel to earlier 1999 and 2003 Dutch COI studies (Polder et al., 1998; Meerding et al., 1998; Slobbe et al., 2006) and COI estimates were made using the healthcare cost definitions of the System of Health Accounts (SHA, see Orosz and Morgan (2004)) for international comparability. To translate the age categories from the COI data into age-yearspecific per capita health expenditure, we interpolated the 21 age classes using cubic splines. Since the Dutch COI figures do not distinguish between costs of survivors and decedents, the most important step in the estimation of equation 3.1 was the decomposition of average per capita expenditure into a part that is attributable to those who die at a certain age and a part that is attributable to those who survive that age. This decomposition was accomplished by assuming that average costs in a single year at a particular age is the weighted average of those surviving that year and those dying that particular year (note that all input parameters and model

42

STANDARDIZING INDIRECT MEDICAL COSTS

calculations are age and sex specific, but that for notational purposes age and sex indices were omitted) [equation 3.3]: aci  (1  m)  sci  m  dci

(3.3)

where aci =average per capita healthcare expenditure for disease i; and m=mortality rate. Per capita healthcare expenditure for survivors and decedents for a particular disease can then be calculated if we know the mortality rate and the ratio ri between healthcare costs for those dying at a particular age and those surviving that age (equation 3.4): dci  ri  sci aci  sci  (ri  1)  m  sci sci 

(3.4)

aci 1  (ri  1)  m

To divide the average per capita costs per disease according to the above-specified relationships, we used additional data sources. Mortality rates for 2005 from Statistics Netherlands were employed (Statistics Netherlands, 2010). Given mortality rates, the only additional input needed is disease-specific cost ratios of decedents and survivors. However, these were only available for hospital expenditure (Wong et al., 2011a). Since, the effect of proximity to death on healthcare expenditure differs strongly per healthcare provider (Werblow et al., 2007), we could not use these ratios directly to decompose all disease-specific per capita health expenditure. Therefore, we used these ratios only to decompose hospital expenditure (equation 3.5): sci , j 1 

aci , j 1 1  (ri , j 1  1)  m

(3.5)

with index j denoting the healthcare provider; j = 1 refers to the hospital sector. Wong et al. (2011a) estimated disease-specific ratios for 75 diseases categorized using the International Shortlist for Hospital Morbidity Tabulation (ISHMT) format, which is highly compatible with the COI categories, resulting in 71 matches of 107 disease categories, which amounts to 60% of total hospital expenditure in 2005 (Table A2 in the Supplemental Digital Content 3 displays the matches of COI categories to the ISHMT categories). For the non-disease-related expenditure (11.7% of total expenditure), we assumed the ratios to equal one and 43

CHAPTER 3

thus, conditional on age and sex, the costs to be equal for survivors and decedents. For the remaining disease categories, we used the age- and sex-specific mode of the 71 matched disease ratios. The mode was estimated by kernel density estimates using average costs per disease as weights. For other healthcare providers besides hospitals, no empirically estimated disease-specific ratios were available. However, for some major health providers (providers of ambulatory healthcare, drugs and appliances, nursing and residential care) we had access to decedent/survivor ratios for total expenditure in 1999 estimated in previous research using data from insurance claims (Polder et al., 2006). To estimate disease-specific ratios for these three healthcare providers (ambulatory healthcare, drugs and appliances, nursing and residential care), we exponentiated all disease-specific hospital ratios by a constant (equation 3.6): ri , j 1  ri , j 1

x j 1

(3.6)

where the j index denotes the healthcare provider; j= 1 implies hospital care; ri,j>1 is the ratio ([costs decedents]/[costs survivors]) for disease-specific health expenditure of disease i for healthcare provider j other than hospital care; xj>1 is a scaling constant for healthcare provider j other than hospital care. Thus, if, for example, the disease-specific hospital ratios for diseases a, b and c equal 4, 9 and 16, respectively, and the scaling factor x for long-term care equals 0.5, the disease-specific ratios for this healthcare provider would equal 2, 3 and 4, respectively. Equation 6 implies that, for each healthcare provider, the age- and sex-specific distributions of ratios [(disease costs decedents)/ (disease costs survivors)] are proportional on the log scale. Suppose we use equation 3.6 for a given baseline disease (denoted by i=1), then this can be rewritten as shown in equation 3.7: log(ri 1, j 1 )  log(ri 1, j 1

x j 1

)  x j 1 

ri 1, j 1 ri 1, j 1

(3.7)

Since we assume xj>1 to be equal for all diseases, we can similarly state that (equation 3.8): x j 1 

ri 1, j 1 ri 1, j 1

Thus (equation 3.9):

44

(3.8)

STANDARDIZING INDIRECT MEDICAL COSTS

x j 1 

log(ri 1, j 1 ) log(ri 1, j 1 )



log(ri 1, j 1 ) log(ri 1, j 1 )

, i

(3.9)

Equation 3.9 describes how the effect of proximity to death on healthcare expenditure differs between healthcare sectors. A value of x higher than one implies that, for that health provider, the relationship between time to death and healthcare expenditure is stronger for all diseases than in the hospital sector. A value for x lower than one implies that the relationship is less strong. An alternative way of scaling the ratios would be to multiply all hospital ratios by a constant. However, since some ratios were smaller than one, we chose to scale the hospital ratios on a log scale. This way, we ensured that the relationship between time to death and healthcare expenditure did not change from negative (ratio smaller than one) to positive (ratio greater than one) or vice versa. Equation 3.9 can be rearranged (equation 3.10) to describe how the proximity to death relationship differs between diseases: log(ri 1, j 1 ) log(ri 1, j 1 )

log(ri 1, j 1 )



log(ri 1, j 1 )

, i

(3.10)

In the example mentioned above, it is easy to check that log(16)/log(4) = log(4)/log(2) = 2. To ensure that the sum of disease-specific costs of decedents and survivors match those of total costs in such a way that the ratio for total expenditure in that healthcare sector equals the empirically estimated ratios (denoted rtot;j>1), we exponentiated all disease-specific hospital ratios by the constant x such that the following assumption is not violated (equation 3.11):

rtot , j 1

 dc   sc i

i , j 1

i

 r  sc   sc x

i , j 1

i , j 1

i

i , j 1

(3.11)

i , j 1

i

Combining equation 3.11 with equation 3.5, we can calculate total survivor expenditure for healthcare providers other than hospital care using the estimated ratio for total expenditure as a function of mortality rates, average costs per disease for that healthcare provider, disease-specific hospital ratios and the scaling constant (equation 3.12):

45

CHAPTER 3

sctot , j 1  sci , j 1 

 ac

i , j 1

i

1  (rtot , j 1  1)  m aci , j 1

1  (ri , j 1

x j 1

(3.12)

 1)  m

sctot , j 1   sci , j 1   i

i

aci , j 1 1  (ri , j 1

x j 1

 1)  m

Equation 3.12 now contains only one unknown variable: the scaling factor x . Age-, sex- and healthcare provider-specific values for x were found by numerically minimizing the error, as defined by the distance between total survivor costs calculated using the empirically estimated ratios for total expenditure (denoted sctot;j>1 for these three healthcare providers (ambulatory healthcare, drugs and appliances, nursing and residential care) and the total survivor costs calculated as the sum of the disease-specific survivors costs (equation 3.13): sctot , j 1   i

aci , j 1 x

1  (ri , j 1  1)  m

(3.13)

For the remaining provider categories (mainly being overhead type healthcare costs), it is assumed that costs are equal for decedents and survivors and that costs depend solely on age and sex.

3.3 Estimated model parameters in PAID 1.0 To show the effect of the decomposition of average per capita health expenditure by costs related to those dying and those surviving, Figure 3.1 displays average per capita costs in the last year of life and other years specified by sex and age, stacked for the different healthcare providers (in this figure we have omitted costs for all other healthcare providers as these depend on age solely and not on time to death). Figures 3.1a and 3.1b display the average per capita health expenditure resulting from interpolating the COI study (summed over all 107 disease categories). Figures 3.1c–f display the estimates summed over all 107 disease categories that are the result of the decomposition of the COI data into costs of decedents and costs of survivors. Please note that the y-axis of different panels have different scales. From Figure 3.1 it can be concluded that costs in the last year of life are very high at a very young age and decrease sharply thereafter.

46

STANDARDIZING INDIRECT MEDICAL COSTS Figure 3.1: Average annual healthcare expenditure per capita (h, year 2005 values) in the Netherlands by age and sex, for four healthcare providers and stratified by last year of life and other years: (a) men average; (b) women average; (c) men survivors; (d) women survivors; (e) men decedents; and (f) women decedents. GP= providers of ambulatory healthcare; HC= hospitals; LTC = nursing and residential care facilities; Med = retail sale and other providers of medical goods.

15000 0

20

40

60

80

20

40

60

80

Age

Age

C: per capita costs men survivors

D: per capita costs women survivors

5000

15000

GP Med HC LTC

0

5000

Annual costs in €

GP Med HC LTC

15000

0

0

20

40

60

80

0

20

40

60

80

Age

Age

E: per capita costs men decedents

F: per capita costs women decedents

Med

HC

LTC

60

80

GP

Med

HC

LTC

60

80

0

0

10000

30000

Annual costs in €

GP

30000

0

10000

Annual costs in €

0

Annual costs in €

GP Med HC LTC

5000

Annual costs in €

5000

15000

B: per capita costs women average

0

Annual costs in €

A: per capita costs men average GP Med HC LTC

0

20

40 Age

0

20

40 Age

The major cause for this decrease is that mortality in the first year of life is often preceded by a period of intensive hospital care, whereas mortality among children, adolescents and especially young adults is mostly caused by (traffic) accidents (Statistics Netherlands, 2010). At middle age, costs in the last year of life increase again. Total costs of survivors increase exponentially at old age mainly due to frailty, disability, co-morbidity and subsequent needs for nursing and residential care. Survivor costs in hospital, for GPs and for medicines, do not depend strongly on age, so the age-related increase in total healthcare expenditure is produced mainly in the long-term care sector. At older age, the share in long-term care costs increases sharply in survivors. Interestingly, absolute cost levels are somewhat

47

CHAPTER 3 Table 3.1: Ratios of (decedent costs)/(survivor costs) for men aged 75 years. Healthcare provider

Ratio for

Scaling

total costs

factor x

Lung cancer ratio

Stroke

Depression

ratio

ratio

Hospitals

8.5

-

121.1

6.3

2.8

Nursing and residential care facilities

7.6

0.90

73.7

5.2

2.5

Providers of ambulatory healthcare

2.2

0.42

7.6

2.2

1.6

Retail sale and other providers of medical goods

2.3

0.46

8.9

2.3

1.6

Table 3.2: Estimated lifetime healthcare costs (€, year 2005 values) stratified by last year of life and all other years and healthcare provider. Healthcare provider

Men

Women

last year

other

of life

years

Total

30017

207979

Hospitals

% costs

last year

other

of life

years

13 (4)

34766

279059

in last year of lifea

% costs in last year of lifea 11 (4)

15571

80389

16 (3)

13017

91452

12 (2)

Nursing and residential care facilities

8046

20897

28 (10)

15373

45633

25 (11)

Providers of ambulatory healthcare

3011

44596

6 (3)

3020

65980

4 (3)

2767

34872

7 (4)

2664

45323

6 (3)

622

27225

2 (2)

691

30671

2 (2)

Retail sale and other providers of medical goods Other healthcare providers a

The numbers between brackets are the percentage of costs in the last year of life if the relationship between proximity to

death and healthcare expenditures is ignored and only the cost of illness data are used.

higher in women than in men, especially at a higher age. This may be explained to a certain extent by the fact that the nursing and residential care population mainly consists of women (Poos et al., 2008). Table 3.1 displays estimates of some disease-specific ratios for different healthcare providers for men aged 75 years. It can be seen that the relationship between time to death and healthcare expenditure is strongest for the hospital care providers. As a result, the scaling factors estimated needed to calculate diseasespecific decedent/ survivor ratios are all below one. Furthermore, the diseasespecific ratio is highest for lung cancer and lowest for depression. Table 3.2 displays estimates of lifetime health expenditure broken down by healthcare provider. Lifetime healthcare expenditure was estimated by calculating the expected value of equation 3.1 using mortality rates for 2005 summed over all diseases. To show the importance of including time to death, we compared the percentage of healthcare expenditure consumed in the last year of life with a naive estimate in which we did not make a distinction between costs in the last year of life and other years as derived from the original COI data.

48

STANDARDIZING INDIRECT MEDICAL COSTS Figure 3.2: Average annual healthcare expenditure per capita (h, year 2005 values) in the Netherlands by age, stratified by last year of life and other years, for two different disease categories: (a) average circulatory system; (b) average neoplasms; (c) survivors’ circulatory system; (d) survivors’ neoplasms; (e) decedents’ circulatory system; and (f) decedents’ neoplasms. GP= providers of ambulatory healthcare; HC= hospitals; LTC = nursing and residential care facilities; Med= retail sale and other providers of medical goods. 2500 1500 0 20

40

60

80

20

40

60 Age

D: survivors neoplasms 1500

Age

C: survivors circulatory system

80

500

1000

GP Med HC LTC

0

500

Annual costs in €

GP Med HC LTC

1000

0

0

20

40

60

80

40

60

80

Age

F: decedents neoplasms

Annual costs in € 20

40

60 Age

80

10000

Age

0 2000

6000

20

E: decedents circulatory system GP Med HC LTC

0

0

GP Med HC LTC

6000

10000

0

0 2000

Annual costs in €

1500

0

Annual costs in €

B: average neoplasms GP Med HC LTC

500

Annual costs in €

1500 500 0

Annual costs in €

2500

A: average circulatory system GP Med HC LTC

0

20

40

60

80

Age

Table 3.2 makes clear that the share of hospital costs is much higher in the last year of life than in other years. However, also for long-term care, a large share of lifetime healthcare expenditure is realized in the last year of life (28% for men and 25% for women). It should be noted that although this seems very large, the share that is expected to be spent on long-term care in the last year of life would also be substantial if costs in the last year of life (conditional on age) are the same as in other years (10% for men and 11% for women). Seen this way, it can be concluded that the effect of including proximity to death effect is most pronounced for hospital expenditure.

49

CHAPTER 3 Table 3.3: Estimated lifetime healthcare costs (€, year 2005 values) stratified by last year of life and other years specified by disease category. Disease category

Men

Women %

% last year of life

other years

costs

last

in last

year

year of

of life

life Total Infectious and parasitic disease Neoplasms Endocrine, nutritional and metabolic diseases Diseases of the blood and the blood-forming organs

costs

other years

in last year of

a

life

a

30017

207979

13 (4)

34766

279059

11 (4)

535

7019

7 (3)

454

7454

6 (2)

5091

8976

36 (5)

3744

11562

24 (3)

805

5190

13 (4)

1117

7261

13 (5)

331

859

28 (6)

340

1081

24 (6)

Mental and behavioural disorders

6136

29522

17 (6)

10080

47875

17 (7)

Diseases of the nervous system

1727

16094

10 (4)

1532

20152

7 (3)

Diseases of the circulatory system

4822

27450

15 (5)

5882

25456

19 (6)

Diseases of the respiratory system

2372

10634

18 (5)

1499

10835

12 (4)

Diseases of the digestive system

1294

19537

6 (2)

1183

22806

5 (2)

Diseases of the genitourinary system

1209

6096

17 (6)

1050

10540

9 (4)

15

669

2 (0)

47

12480

0 (0)

315

3855

8 (3)

348

4422

7 (3)

1189

13506

8 (3)

2112

23292

8 (3)

Congenital malformations

34

1839

2 (0)

28

1728

2 (0)

Certain conditions originating in the perinatal period

65

2781

2 (0)

54

2484

2 (0)

2166

20914

9 (3)

1840

27201

6 (3)

1061

6900

13 (5)

2063

10304

17 (7)

850

26136

3 (3)

1396

32123

4 (4)

Pregnancy, childbirth and the puerperium Diseases of the skin and subcutaneous tissue Diseases of the musculoskeletal system and connective tissue

Symptoms, signs and abnormal clinical and laboratory findings, not elsewhere classified Injury, poison and certain other consequences of external causes Not allocated/not disease related a

The numbers between brackets are the percentage of costs in the last year of life if the relationship between proximity to

death and healthcare expenditures is ignored.

To focus on the differences in healthcare expenditure patterns between diseases, Figure 3.2 displays a similar graph as Figure 3.1, but now for two different disease categories instead of total expenditure: neoplasms and diseases of the circulatory system. Please note that the y-axes have different scales. Figure 3.2 clearly illustrates the differences between disease categories. Per capita expenditure for neoplasms is, on average, lower than for circulatory diseases. However, in the last year of life, per capita health expenditure is much higher for neoplasms. More specifically, the average healthcare expenditure for neoplasms is largely determined by hospital expenditure in the last year of life. Table 3.3 presents estimates of lifetime healthcare costs broken down into costs in the last year of life and other years for different disease categories. Table 3.3 demonstrates, for example, that cancer is a major cost component in the last year 50

STANDARDIZING INDIRECT MEDICAL COSTS

of life, but hardly in other years. The same goes for diseases of the blood and blood-forming organs. 3.4 Discussion Since economic evaluations of life-saving (preventive) interventions are conducted in a variety of settings, including those of new (and often expensive) drugs, a standardized way to account for indirect medical costs is of great importance. While some methods have been proposed to estimate medical costs in LYG (Gandjour and Lauterbach, 2005), these methods do not take into account that, in most economic evaluations, a part of these indirect medical costs have already been covered; namely, those costs related to the disease or intervention that was evaluated. These future costs that are expected to change as a result of an intervention should always be explicitly modeled and, hence, this is common practice in the majority of economic evaluations. Subsequently, simply adding per capita health expenditure stratified by age and proximity to death will result in double counting of the costs of related diseases (van Baal et al., 2011a). We have proposed a methodology to estimate the costs of all other diseases in LYG in a standardized manner that avoids this double counting. Starting from the framework developed by Gandjour and Lauterbach (2005) we present a methodology to adjust per capita health expenditure stratified by age and proximity to death for the costs of the diseases already included in the main economic evaluation. In this conceptual model, it is assumed that total healthcare expenditure can be explained by age, sex and time to death, while the relationship between costs and these three variables differs per disease. We present estimates of our conceptual model, which are embedded in a toolkit called PAID 1.0, tailored for economic evaluations in the Netherlands. Disease-specific average per capita expenditure from the Dutch COI 2005 study (Poos et al., 2008) were decomposed into a part that is attributable to those who die at a certain age and a part that is attributable to those who survive that age. To accomplish this we used estimates of ratios of decedent/survivor costs for disease-specific hospital expenditure and total expenditure for healthcare providers other than hospitals. Our results on the effect of the last year of life with respect to total health expenditure calculated as the sum of disease-specific health expenditure for all healthcare providers are in line with previous research conducted in the Netherlands (van Baal et al., 2011a). In our methodology, we accounted for the fact that the relationship between age, sex and proximity to death and per capita costs differs between diseases (Wong et al., 2011a). This allows the relationship between time to death and healthcare costs to be altered if the costs of related diseases are excluded (van Baal et al., 2011a). Consequently, for our methodology it is essential to know what the 51

CHAPTER 3

role of age and proximity to death on disease-specific per capita health expenditure is for each disease. Our conceptual model is similar to the concept of ‘other cause’/‘background’ mortality, which is often used in simulation models to decompose total mortality rates into a part related to the intervention and a part unrelated to the intervention. To decompose average per capita health expenditure into costs in the last year of life and all other years, we had to make several assumptions. Most importantly, we assumed that the disease-specific ratios estimated in Wong et al. (2011a) based on 60% of hospital expenditure, could be generalized to total hospital expenditure and that the disease-specific ratios could be used to decompose disease-specific costs for some other healthcare providers under some constraints. Furthermore, the validity of PAID 1.0 crucially depends on the validity of the COI study and the allocation of the health expenditure to disease categories in that study. Another limitation is that we dichotomized proximity to death into two categories. A further version of PAID could be improved by stratifying into more periods. This becomes more important if we consider the timing of health expenditure and the role of discounting therein. We used total healthcare expenditure in the Netherlands from 2005 as a starting point. This implies that agespecific cross-sectional data are interpreted in a longitudinal fashion, as is done when constructing life tables and also in many Markov models. The implicit assumption is that current observed patterns of health expenditure remain constant. Of course, the longer the period modeled, the more problematic this assumption becomes. We do not claim that the parameters included in our conceptual model are the only ones that are needed to estimate indirect medical costs. Technological progress, innovation, changes in morbidity patterns, developments in the labor market and institutional changes may have an impact on future healthcare costs, but have not been included. Although we recognize these limitations, we are convinced that it is better to provide an estimate using all current, albeit imperfect, knowledge than no estimate at all. While the former estimate may be imprecise, the latter is surely wrong. In economic evaluations, modeling techniques are applied frequently to estimate the effects of life-prolonging interventions on health and healthcare costs. Usually, in these models, an intermediate effect such as blood pressure, newly detected cases through screening, or short-term survival (as estimated using observational data, an RCT or meta-analyses) is connected to causally related events (most importantly, death) that could not be observed within the trial period of the intervention because the follow-up period is too short. Thus, models are used to reach beyond the time horizon of trials. As a result, costs and effects beyond the observed period have to be estimated from other data sources. In cost effectiveness studies that capture both health effects and costs of related diseases during added LYG, PAID 1.0 allows the estimation of future healthcare costs, 52

STANDARDIZING INDIRECT MEDICAL COSTS

correcting for costs of diseases already included in the basic evaluation, taking into account that the relationship between healthcare expenditure and proximity to death differs per disease and healthcare provider. If costs are included for only a limited follow-up period while, at the same time, health effects are modeled for the whole course of life after the follow-up period, PAID 1.0 can be used for the inclusion of age-specific costs of survivors and decedents. PAID 1.0 is populated with country-specific (Dutch) data and not immediately transferable for use in other countries. The most important ingredient to construct similar tools for other countries is top-down COI studies, covering all healthcare expenditure, which have already been conducted in a variety of countries (Heijink et al., 2008). The relationship between healthcare costs and proximity to death is less well researched for different countries, but we expect this variable to be less susceptible to variation between countries than healthcare expenditure itself (Payne et al., 2007). Therefore, if country-specific information on the influence of time to death on health expenditure is lacking, an option might be to ‘borrow’ data from other countries, e.g. those presented in this article.

3.5 Conclusions We think that the use of PAID 1.0 improves comparability between economic evaluations in the Netherlands and we hope that our proposed methodology may inspire researchers from other countries to further refine and improve standardized estimation of indirect medical costs.

53

CHAPTER 3

54

Chapter 4 Time to Death and the Forecasting of Macro-Level Health Care Expenditures: Some Further Considerations ‡

Although the effect of time to death (TTD) on health care expenditures (HCE) has been investigated using individual level data, the most profound implications of TTD have been for the forecasting of macro-level HCE as the TTD relation at the micro-level suggests a positive relation between changes in mortality rates and changes in per capita HCE. In this paper, we estimate the TTD model using macro-level data consisting of mortality rates and age-specific per capita health expenditures for the years 1981-2007. Forecasts for the years 2008-2020 of this macro-level TTD model are compared to forecasts that exclude any link to mortality. Results revealed that the effect of TTD on HCE in our macro model was similar as found in micro-econometric studies. However, as the inclusion of changes in mortality rates pushed downwards the estimates of the growth rates due to unidentified causes, forecasts of HCE for the period 2008-2020 were similar for the two models. We conclude that including TTD will not improve forecasts of macro-level HCE. 4.1 Introduction Although the rising number of elderly certainly increases the group of individuals who need health care, the extent to which population aging actually drives HCE upward has been debated (Gerdtham and Jönsson, 2000; Reinhardt, 2003; Gray, 2005; Bech et al., 2011). One topic that has received particular attention is the question whether longevity gains increase health care spending and if so, by how This chapter is based on: Van Baal PHM, Wong A. 2011. Time to death and the forecasting of macro-level health care expenditures: some further considerations (submitted). ‡

CHAPTER 4

much (Payne et al., 2007). In an important paper in this area Zweifel and colleagues (Zweifel et al., 1999) argued that differences in HCE between individuals can be better explained by time to death (TTD) than time since birth (age). They referred to the role of aging in HCE growth as a ‘red herring’, as the issue of aging diverts attention from the real causes of HCE growth, such as government regulations in the health care sector and advances in medical technology. It is not too surprising that time to death better explains health care utilization than age if we consider that both age and TTD are proxy variables for morbidity and disability, which are the real drivers of individual health care demand. Although the prevalences of morbidity and disability generally increase with age, these conditions also increase mortality risk (Carstensen et al., 2008; Feigin et al., 2003; Majer et al., 2011; Engelfriet et al., 2011). As a result, TTD might capture more variation in HCE than age, as it not only incorporates age but also mortality risk. Subsequently, some studies have confirmed that TTD loses its explanatory power once adjustments are made for morbidity and disability (Shang and Goldman, 2008; de Meijer et al., 2011). With respect to the question whether HCE increases with longevity Zweifel et al. in their 1999 paper stated “per capita HCE is not necessarily affected by the ageing of the population due to an increase in life expectancy. Rather, an increase in the elderly’s share of population seems to shift the bulk of HCE to higher age, leaving per capita HCE unchanged” (Zweifel et al., 1999). Although the effect of time to death (TTD) on HCE has been investigated using individual level data, including TTD in analyses has most profound implications in the forecasting of macro-level HCE. Seshamani and Gray (2004a) noted that “Incorporation of proximity to death into health expenditure projection models will prove critical to more accurate estimation of future health expenditure trends, which will better meet the challenge of population ageing”. Assuming that individual health care consumption is concentrated in a relatively brief period when death is approaching implies (ceteris paribus) that there should be a positive relation between changes in mortality rates and per capita HCE. It is this relation between mortality rates and HCE at the macro-level that has been exploited in what we will call TTD prediction models. Since life expectancy has been rising in most western countries and is expected to rise even further (Oeppen and Vaupel, 2002), it has been argued that forecasts of HCE should include TTD (Stearns and Norton, 2004). In multiple studies the argument has been brought forth that excluding time to death in estimates of future HCE results in overestimation of total health expenditure. This has been confirmed to varying degrees in several papers (Stearns and Norton, 2004; Madsen et al., 2002; Polder et al., 2006; Breyer and Felder, 2006). Madsen et al. (2002) modeled HCE in Denmark over a 25-year old period, and found that the overestimation to be equal to 3.4%. Stearns and Norton (2004) found a difference of 15% for Medicare projections up until 2020. Using a model pertaining to the period 1999-2020 in the Netherlands, Polder et al. (2006) found the difference to 56

TIME TO DEATH AND FORECASTING OF EXPENDITURES

be in the order of 9%. Breyer and Felder (2006) estimated that projections of per capita HCE in 2050 would be 18.5% lower if the TTD effect is taken into account. Given the increased pressure of HCE on public expenditures these more optimistic forecasts of HCE were received as good news by policy makers, and subsequently national and international governmental bodies adjusted their forecasts of HCE downwards to incorporate the effect of increased longevity (DG-ECFIN, 2009; OECD, 2006; Besseling and Shestalova, 2010) Given the widespread use of TTD models in forecasting HCE, surprisingly little attention has been paid to the validity of these models. We believe three issues need to be addressed before it can be concluded that including TTD improves forecasts of macro-level HCE. First of all, the relation between mortality rates and per capita HCE as assumed in these models has never been estimated using macro-level data. This is remarkable given the known difference between determinants of individual health care needs and the national HCE (Getzen, 2000; Getzen, 2001). Macro HCE is not simply the sum of individual health care needs since the budget for health care spending is often set by an insurer and/or government. This budget determines which technologies are supplied to the patients. As a consequence, even though in a given period persons closer to death consume more care than those further away from death, this does not automatically imply that a decline in the mortality risk of the entire population leads to a slower growth of HCE. These theoretical differences between individual and aggregate level determinants are also reflected in the empirical literature. Microeconometric studies usually focus on factors associated with health care demand (Payne et al., 2007), while macro-level studies traditionally have focused more on supply side factors such as GDP and technological change (Gerdtham and Jönsson, 2000). Secondly, future life expectancy is an unknown and its values need to be forecast. As such, they cannot be considered as a given and are surrounded by considerable uncertainty. This has been recognized in the actuarial and demographic community (Pitacco et al., 2008; Tabeau, 2001; De Waegenaere et al., 2010). However, in all of the studies using TTD to forecast macro-level HCE it was implicitly assumed that mortality rates are deterministic (Stearns and Norton, 2004; Madsen et al., 2002; Polder et al., 2006; Breyer and Felder, 2006) and uncertainty surrounding forecasts of mortality rates was not included in the projections of HCE. Forecasts of mortality rates in these studies were either taken from sources like census bureaus (Madsen et al., 2002; Polder et al., 2006; Breyer and Felder, 2006) or were obtained by simple extrapolation of trends in mortality rates without including uncertainty surrounding the trends (Stearns and Norton, 2004). Compared to simpler forecasting models, forecasts of HCE using TTD add additional complexity as forecasts of HCE require forecasts of mortality rates as input. 57

CHAPTER 4

Thirdly, in all of the studies that used TTD to forecast macro HCE the degree of overestimation due to the neglect of the TTD effect was assessed by comparing the predictions of the TTD models with predictions of so-called naïve models in which the influence of mortality risk was not included (Stearns and Norton, 2004; Madsen et al., 2002; Polder et al., 2006; Breyer and Felder, 2006). However, in these comparisons the influence of many potentially relevant factors that might influence HCE was ignored. In some studies the focus was on the impact of aging on HCE, while the only additional source that might affect macro HCE was forecast population size (Polder et al., 2006). In these projections age- and genderspecific per capita HCE in the naïve models were assumed constant and changes in HCE were purely the result of changes in the age and gender distribution of the population. As the population in most western countries ages, health care costs in these models were projected to rise. In contrast, in the TTD models in these studies age- and gender-specific per capita HCE were allowed to change as a function of mortality rates (Stearns and Norton, 2004; Madsen et al., 2002; Polder et al., 2006; Breyer and Felder, 2006). Not surprisingly, a decrease in mortality rates led to lower forecasts of HCE than in the naïve models. Although it was recognized in these studies that other factors are important in predicting HCE, the TTD effect was assumed to operate independently. For instance, Stearns and Norton note that “Future expenditures will also be affected by a number of other factors including health status trends and technological change…Despite these considerations, improvements in predictions that are easily achieved have intrinsic value, and the predicted differences between the models are sufficient to justify reassessment of the value of inclusion of time to death in models for predicting HCE”. Implicit in this notion is the assumption that the effect of other determinants of HCE are not affected by in- or exclusion of TTD. Some other studies did include growth rates in per capita HCE due to unspecified causes (these rates could refer to anything from technological change to changes in clinical practice). In these studies age- and gender-specific per capita HCE were assumed to grow with a yearly constant percentage (Madsen et al., 2002; Breyer and Felder, 2006). For instance Breyer and Felder assumed that age- and gender-specific per capita HCE grow at a rate of 1% yearly due to unspecified causes: “To forecast the total development of HCE (at constant prices), it is necessary to account for the growth factor “technological change” in medicine, which has increased per-capita expenditures by 1% per annum in the period 1970–1995, holding income and age structure constant”. They assumed that per capita HCE at all ages increase yearly with 1% in both the TTD and the naïve model. The degree of overestimation of future HCE by ignoring TTD was similar as in studies assuming no growth due to other causes. Summarizing, in all forecasting studies using TTD it was implicitly assumed that the influence of other factors besides TTD on macro HCE was insensitive to the inclusion of TTD in the prediction model. Furthermore, the magnitude of TTD was assessed by applying the ceteris paribus assumption. 58

TIME TO DEATH AND FORECASTING OF EXPENDITURES

In this paper we will investigate the three above mentioned issues in order to assess the relevance of TTD in forecasting macro HCE. Given that the most profound implications of the red herring hypothesis pertain to the forecasting of macro HCE we tested the TTD hypothesis using macro-level data. Furthermore, since the effect of TTD depends crucially on forecasts of future mortality rates we incorporated the uncertainty to which these are subject when forecasting HCE with TTD. Finally, and most importantly, we investigated to what extent including TTD in a forecasting model influences the strength of the growth rate due to other unspecified causes. The structure of this paper is as follows. We will first describe how TTD models are currently used to forecast macro HCE. From this structure we derive a theoretical model that, under assumptions as made in the TTD context, allows us to relate changes in age-specific mortality rates to changes in age-specific per capita spending. Then, we estimate the parameters of this model and compare its parameters to a model without TTD by using aggregate level panel data. These data were age- and sex-specific per capita hospital expenditures and mortality rates for the years 1981-2007. Finally, we will forecast hospital expenditures for the period 2008-2020 and compare predictions for different versions of the TTD and the naïve model. Actuarial methods are used to forecast mortality rates that serve as input for the TTD models which allows the quantification of uncertainty surrounding the forecasts of mortality rates.

4.2 Theoretical model In this section we will explain two models that are currently used to forecast HCE. We start by explaining the so-called naïve model because in all TTD studies this model was used as a baseline model to quantify the influence of including TTD. Then we will explain the TTD prediction model that forecasts HCE incorporating the mechanism that individuals consume most health care in the last year of their lives. The naïve prediction model Before TTD was used in forecast models, predicting HCE t * years ahead of known time point t 0 was done by estimating:





E H (t 0  t * )   (1   a , g ) t  C a , g (t 0 )  N a , g (t 0  t * ) *

(4.1)

a A gG

where a is an index for age, g an index for gender, H (t 0  t * ) is the total HCE at time t  (t 0  t * ) , N a , g (t 0  t * ) the age- and gender-specific population size at 59

CHAPTER 4

time t , C a , g (t 0 ) the age- and gender-specific per capita HCE at t 0 , and  a , g a constant exogenous growth rate. In the Red Herring literature, this model is referred to as the ‘naïve model’. In this naïve model forecasts of HCE crucially depend (I) on forecasts of N a1 , g1 (t 0  t * ),..., N a p , g 2 (t 0  t * ) , which are usually





taken from sources like national census offices, and (II) on estimates of  a, g , which are usually based on the growth in total HCE and are therefore considered equal for all age and gender classes. These changes in age- and gender-specific per capita HCE can be stated in log form as follows: d log(C a , g ) dt

(4.2)

 log(1   a , g )

In this naïve model uncertainty in the prediction of total HCE stems mainly from uncertainty surrounding the future population size and the growth rates. However, in practice both uncertainties are not taken into account, often out of convenience. The TTD prediction model The ‘Red Herring’ model was conceived by Zweifel et al. (1999) to account for the fact that the individual HCE are, on average, higher in the last year of life. By including TTD as an additional covariate in the estimation of average individual HCE, the effect of age on HCE is greatly diminished. The newfound estimates can then be combined with projections of demographics to predict future aggregate HCE. Estimates of H (t 0  t * ) using a TTD model are obtained by substituting the following in equation (4.1) (Stearns and Norton, 2004; Madsen et al., 2002; Polder et al., 2006; Breyer and Felder, 2006): C a , g (t 0 )  M a , g (t 0  t * )  C a , g ,d (t 0 )  [1  M a , g (t 0  t * ))]  C a , g ,u (t 0 ) (4.3)

where M a , g (t 0 ) denotes the mortality rate, C a , g ,d denotes per capita HCE in the last year of life and Ca , g ,u denotes per capita HCE for survivors (i.e., all ‘other’ years of life besides the last year of life), for a given age group a and gender g. Essentially, under the ‘Red Herring’ model the per capita HCE C a , g is decomposed into a deceased and survivor part. Equation (4.3) implies that C a , g is a weighted cost average over deceased (persons dying at that age) and survivors (people who survive that age), with the weights being determined by the mortality rates in that age and gender group. While in the naïve model changes in mortality rates may influence forecasts of HCE indirectly through its effect on forecasts of N a1 , g1 (t 0  t * ),..., N a p , g 2 (t 0  t * ) , in the TTD model they influence forecast HCE



60



TIME TO DEATH AND FORECASTING OF EXPENDITURES

directly by influencing per capita HCE. If mortality rates decline over time, C a , g ,d will have a smaller weight, while C a , g ,u will have a larger weight. Thus, it seems natural that per capita spending for a given age and gender group will decrease as a result of lower mortality rates within that same group, as C a , g ,d  C a , g ,u . If we denote the strength of the relation between TTD and per capita HCE by the ratio C a , g ,d R , the derivative of the log of per capita HCE with respect to time can be C a , g ,u rewritten as: d log(C a , g ) dt

  M a , g (t )  R  C a , g ,u (t )  [1  M a , g (t ))]  C a , g ,u (t )   log  M (t  1)  R  C (t  1)  [1  M (t  1))]  C (t  1)  a , g ,u a,g a , g ,u   a,g  ( R  1)  M a , g (t )  1 C a , g ,u (t )    log  ( R  1)  M (t  1)  1 C (t  1)  a,g a , g ,u     ( R  1)  M a , g (t )  1  1     log   ( R  1)  M (t  1)  1 a,g    ( R  1)  M a , g (t  1)  M a , g ,t   1    log(1   )  log   ( R  1)  M a , g (t  1)  1     ( R  1)  M a , g ,t  (4.4)  log(1   )  log1   ( R  1)  M (t  1)  1  a , g  

From equation (4.4) we can see the changes in C a , g depend on some estimated growth rate  and a function, which has the level of mortality at a previous point of time, the growth in mortality in the previous time unit and the expenditures ratio as arguments. We can easily observe two boundary cases. On one hand, no matter what value mortality assumes, if costs in the last year equal those in other years (R=1) then changes in per capita HCE only depend on  and thus equal those of the naïve model (    ). On the other hand, for any constant R  1 , if dM a , g  0 ), we also see that    . mortality rates do not change over time ( dt However, if we assume that R  1 , a decrease in mortality rates then leads to a decrease in per capita HCE. Note that this effect is stronger for higher values of R. Furthermore, this effect also depends on the level of mortality at t  1 . For R  1 , a decrease in mortality will have a stronger effect when the level of mortality is higher (i.e., closer to one). Also note that this model does not actually use TTD as a variable, but rather, a slight reformulation of it in the form of R. R>1 61

CHAPTER 4

corresponds with a significant TTD. This approach has been used before (Wong et al., 2011a). Equation (4.4) implies that decreasing mortality rates will dampen the growth of per capita HCE. The Red Herring model was proposed as a better way to extrapolate trends: the trend in H (t 0  t ) was thought to be forecasted more accurately when it is decomposed in mortality and cost trends. However, when using equation (4.3) to forecast HCE instead of equation (4.1) one needs to forecast mortality rates. In all TTD studies so far forecasts thereof are considered deterministic. Furthermore, although it is a not a necessary implication of the Red Herring theory, all studies so far have used identical growth rates when comparing the predictions of (4.1) and (4.3), i.e.,    . An important point worth mentioning is that equation (4.2) is a simplified case of the ‘Red Herring’ model, where only a distinction is being made between the last year of life and the other years of life. Seshamani & Gray (2004a/b/c), amongst others, have shown that increased HCE are found for the last years of life. In Appendix 4.A we will generalize the relation between TTD and per capita HCE to the continuous case. We show that under specific conditions it is likely that, ceteris paribus, declining mortality rates at a given age will lead to a decrease in per capita HCE at that same age if HCE are related to time to death, similar to in equation (4.4). Empirical model To assess whether the inclusion of TTD in macro-economic models is also supported in an empirical context, we estimated two models to predict aggregate hospital care HCE during the period 1981-2007: A first differenced model is used to estimate changes in log-scaled HCE as a function of changes in mortality, and age-, gender- and period-specific growth rates similar as in equation (4.4), and a naïve first differenced model, that excludes any link to changes in mortality as specified in equation (4.2). In the model including mortality rates, we will take into account uncertainty in the predictions of mortality rates. The latter will be quantified using the Lee-Carter model which is a popular method amongst actuaries to forecast mortality rates and life expectancy (Lee and Carter, 1992; Lee, 2000). Forecasts of mortality rates using the Lee-Carter model serve as input for the model that forecasts HCE as a function of mortality rates. All analyses were done in R (www.r-project.org). We used age- and gender-specific mortality rates and per capita HCE for the period 1981-2007 in the Netherlands. Data To construct age- and gender-specific time series of per capita HCE we used as a starting point data from National Health Care Accounts that are published

62

TIME TO DEATH AND FORECASTING OF EXPENDITURES Table 4.1: Summary of the dataset. Men

Women %growth

Per Age

capita HCE 2007*

Number of deaths per 100,000

per capita HCE 19812007

%growth mortality rate 19812007

Per capita HCE 2007*

Number of deaths per 100,000

%growth HCE 19812007

%growth mortality rate 19812007

0-1

16810

428

128.8

-47.6

14954

310

123.2

-49.8

2-5

783

29

-9

-66.4

620

22

-2.5

-71

4-9

355

10

-30.5

-68.5

301

9

-16

-52.8

10-14

382

10

-5.7

-63.7

379

11

14.1

-43

15-19

534

32

-6

-48.2

698

17

19

-39.7

20-24

612

48

0.2

-45.9

1157

25

10.1

-22.3

25-29

606

46

3.2

-41.9

1866

23

26.2

-43.7

30-34

669

60

7.2

-34.2

2288

37

85.2

-29.5

35-39

815

85

14.3

-33

1667

54

44.1

-33.5

40-44

1050

118

18.1

-39.5

1372

93

6.4

-28.5

45-49

1431

204

23.6

-40.4

1649

174

17

-18.2

50-54

1977

353

29.5

-45.1

2041

284

26.7

-21.2

55-59

2883

599

44.8

-45.2

2527

424

40.7

-15.4

60-64

3997

1019

52.9

-44.7

3276

655

48.1

-18.2

65-69

5686

1674

65.5

-46.1

4595

961

55.1

-28.4

70-74

8507

2863

87.3

-43.4

6374

1603

63.8

-28.5

75-79

11548

5216

114.4

-35.4

8893

2885

77.4

-31.2

80-84

13812

9145

132.3

-25.3

11250

5528

94.2

-26.9

85-89

15763

16174

172.9

-13

12721

10897

118.8

-21.2

90+

15131

31496

234.8

8.3

11955

24310

161.1

-4.4

*expressed in euro price level 2007

annually starting from 1972 to 2007 (Statistics Netherlands, 2010). In that study, multiple data sources are employed to comprehensively estimate the national expenditures related to health care and general well-being. From that study we took annual HCE for the Netherlands for the period 1981 to 2007. These annual HCE include the costs of all hospitals and ambulatory care but do not include the costs of long-term care. These costs were decomposed into gender and 20 age classes (01 year, 2-5, 6-10, etc… 90+) by using information on the number of clinical admissions and the average length of stay after clinical admission. Values for these variables were available for each age and gender class for the period 1981-2007 and were taken from the Dutch Hospital Discharge Register (LMR). All university and general hospitals and most specialized hospitals have agreed to participate in this register. As a result the LMR provides a nearly complete coverage of all hospital inpatient admissions in the Netherlands. The age- and gender-specific per capita

63

CHAPTER 4

C a , g (t ) in a particular year t can then be calculated as the product of the average

per diem HCE and the average number of hospital days per individual: C a , g (t ) 

K a , g (t ) La , g (t ) H (t )  K a, g (t ) La, g (t ) N a, g (t ) a

(4.5)

g

where K a , g (t ) is the number of clinical admissions and La , g (t ) the average length of stay per clinical admission. After decomposing total annual HCE using equation (4.7) this resulted in 20 × 2 × 27 = 1080 observations (20 age classes, 2 genders, 27 years). Table 4.1 displays a summary of the created dataset. From Table 4.1 the cross-sectional relation between age, HCE and mortality becomes obvious as the age gradient for mortality rates and HCE is pretty similar. If we look at the changes in per capita HCE and mortality rates between 1981 and 2007 we can clearly see that while mortality rates have been declining for most ages, HCE has been increasing for most ages and that the increase has been higher at older ages. This has resulted in a shift of the peak of per capita expenditures to higher ages. Modeling health care expenditures Since our main research question is whether changes in mortality rates can be related to changes in age-specific per capita HCE, we differenced the data on mortality rates and per capita HCE with respect to time t. Using differenced data has the advantage that it removes (the majority of) the serial correlation in the error terms. After differencing the age- and gender-specific log of mortality rates and per capita HCE the number of observations equaled 20 × 2 × 26 = 1040 observations (20 age classes, 2 genders, 26 years). Based on the prediction models described in the previous section and plots of the marginal growth rates by gender and age we estimated the following naïve model:  log(C a , g ,t )   0  i 1 i  X i   j 1982  j  Z j   1;a , g ,t 39

2006

(4.6)

α0 indicates the average growth rate over all gender and age groups and all years. αi and β2 are vectors of parameters to be estimated, Xi is a vector of dummies indicating each combination of gender and each age category , and Zj is a vector of dummies indicating each year. The two vectors dummy variables were both coded using so-called effects coding so that the intercept α0 equals the average growth rate over all the years and age and gender categories. We estimated the growth rates directly since growth rates are small and thus we can reasonably assume log(1   i )   i .

64

TIME TO DEATH AND FORECASTING OF EXPENDITURES

In addition to all the parameters in the naïve model, the TTD model (as in equation 4.7) includes parameters related to the changes in mortality rates (as derived in equation 4.4):  log(C a , g ,t )   0  i 1  i  X i   j 1982  j  Z j 39

2006

  ( R1  R2  age  R3  baby  1)  M a , g ,t  (4.7)  log1   ( R  R  age  R  baby  1)  M (t  1)  1  2;a , g ,t 1 2 3 a,g  

Our hypothesis was that increases in mortality rates are negatively related to HCE and therefore that the estimate of R1 is positive. Given results of previous studies that have showed that the ratio of decedents and survivors is lower at higher ages (80+) compared to 65+ to 80 we expect R2 to be negative (Payne et al., 2007). A dummy variable for newborns was included as the mortality rate and HCE for newborns is substantially different from adjacent ages. Estimating the values for R1, R2 and R3 allowed us to back calculate the costs for the last year of life (decedent costs) and in other years (survivor costs). The  in equation (4.7) can be interpreted as the age- and gender-specific net growth rates due to all causes not included in the model. In all forecasting studies so far it has been assumed that  0   0 and that there is no age gradient in the growth rate of health expenditures. Since our data pointed towards age differences in terms of growth rates we not only compared the average growth over all age categories but also the age- and gender-specific growth rates of both models. Since the second term in equation (4.7) contains a non-linear term we could not estimate this model using OLS. To estimate equation (4.7) we minimized the sum of squares using an algorithm developed by Nelder-Mead (Nelder and Mead, 1965) as implemented in the function optim() in R. Starting values for this algorithm were found by estimating the parameters of a linear approximation for equation (4.7):  log(C a , g ,t )

  0  i 1  i  X i  i 1982  j  Z j   1  ( M a , g ,t ) 39

2006

  2  age  ( M a , g ,t )   3  baby  ( M a , g ,t )   3;a , g ,t

(4.8)

It was assumed that errors (  1;a , g ,t ,  2;a , g ,t and  3;a , g ,t ) were normally distributed. For all time series in each model (i.e., 20 age categories times 2 gender categories for each model) we ran Breusch-Godfrey and Durbin-Watson tests in order to test for any remaining first order serial correlation in the errors. In all cases, the null hypothesis of no serial correlation could not be rejected. The Breusch-Godfrey test rejected the null hypothesis of homoscedasticity (p<0.001) in both linear models, so we used the White’s heteroscedasticity-consistent estimator for 65

CHAPTER 4

estimation of standard errors. Even though the Breusch-Godfrey test could not be applied to the non-linear model we also used White’s heteroscedasticity-consistent estimator for estimation of standard errors to allow a more straightforward comparison with the linear models. Modeling mortality rates To forecast mortality rates needed to forecast HCE for the TTD model we used the so-called Lee-Carter model. The Lee-Carter model as applied with respect to mortality postulates that the mortality rate for a given age and calendar year could be modeled as a function of three sets of parameters: age-specific constants, a time-varying index and interaction terms between time and age (Lee and Carter, 1992; Lee, 2000). In our analyses we used the following version of the Lee-Carter model: logM a , g ,t    a , g   a , g   t   4;a , g ,t

(4.9)

Since the model is underdetermined, the following constraints have to be put on the parameters so that a unique solution will be possible to find:  a , g  0 , a,g

  t   1. Normalising the parameters this way does not alter the model fit, in t

addition these constraints make the parameters interpretable.  a , g indicates the time-averaged log-mortality rate for each age and gender class.  t  refers to an age-independent latent mortality index, which describes a general level of mortality at time t ; that is  t  quantifies the evolution of mortality rates over time.  a , g can be interpreted as the interaction between each age/gender category with mortality. In other words, it tells us at which age and gender class mortality rates decline rapidly and which rates decline slowly in response to changes in  t  . The white noise error term  4;a , g ,t , which reflects particular age- and gender specific historical influences not captured by the model, has mean 0 and variance  2 . Lee and Carter showed how Singular Value Decomposition (SVD) can be used to find the least square solution. SVD is applied to the logarithm of the death rates after the age-specific averages of the log-mortality rates have been subtracted. After fitting the model, one retains a series of time-dependent  t  values for each calendar year, which can be treated as a time series and forecasts of mortality can be made by forecasting  t  .

66

TIME TO DEATH AND FORECASTING OF EXPENDITURES

Forecasting mortality rates and per capita health care expenditures We forecasted HCE up until 2020 using the naïve and TTD models. When making predictions we did not attempt to forecast year-specific fluctuations but only used the average growth rate observed over the period 1981-2007. For the naïve model forecasts were made by employing estimates of equation (4.7) in the following way: E H (t 0  1)   exp[(ˆ 0  i 1ˆ i  X i   1;a , g ,t )  log(C a , g (t 0 ))]  na , g (4.10) 39

aA gG

where na , g  N a , g (t 2007 ) /  N a , g (t 2007 ) . a A gG

Forecasts of per capita HCE for the years 2008-2020 were made using the 2007 population. This has the advantage that we ruled out the effect of changes in population size. To forecast age-specific health expenditures with the TTD model whilst assuming mortality rates are estimates, rather than deterministic values, we proceeded in two steps. First, we forecasted mortality rates by applying the LeeCarter method. Forecasts of the Lee-Carter served as input for the TTD model. Lee and Carter proposed that specifying a time series model of random walk with drift parameter describes  t  . If there is a stable linear declining trend in  t  then the following model is sensible:  t   d   t  1  u t  , where d is the drift term, and u t  ~ N 0,  2  .Then predicting the future log-mortality rates can be done by substituting predicted values of  t  into the systematic part of the LeeCarter model. Although Lee & Carter have suggested alternate specifications time series, one of the main insights of applying the Lee & Carter method was that the time parameter for time series in many data sets could be modeled simply as a random walk with drift. This resulted in the following prediction formula for mortality rates:  log( Mˆ a , g )  (dˆ  u (t ))  ˆ a , g   4;a , g ,t

(4.11)

These estimated values of the mortality rates were used in conjunction with equation (4.7) to obtain forecasts based on the TTD prediction model: E H (t 0  1)   ( Rˆ1  Rˆ 2  age  Rˆ 3  baby  1)  Mˆ a , g ,t0 1  39 ˆ ˆ   exp[( 0  i 1  i  X i  log 1   ˆ ˆ ˆ   ( 1 ) ( ) 1 R R age R baby M t        aA gG 1 2 3 a,g 0   (4.12)  ˆ2 ;a , g ,t0 1 )  log(C a , g (t 0 ))]  na , g

67

CHAPTER 4

To quantify prediction intervals for all forecasts we used residual bootstrapping techniques (also known as model-based resampling). This means that after fitting the models random samples from the residuals are taken and added to the fitted values of the response variable to construct a new data set from which the model can be re-estimated. The steps outlined above were repeated 10,000 times and confidence intervals of the predictions were calculated by taking the 0.025 and 0.975 percentile interval of these predictions. We chose for model-based resampling over moving-blocks bootstrapping, which have also been employed in time-series based analyses, as the number of periods was too small to conduct moving blocks bootstrapping. It has also been applied in conjunction with the Lee-Carter model (Härdle et al., 2003; Koissi et al., 2006). Furthermore, using model-based bootstrapping allowed us to take into account the correlation between the Lee-Carter and TTD prediction in a simple manner. After estimating equations (4.7) and (4.9) we calculated the model predictions and residuals for the two models and randomly drew with replacement from a vector identifying year, age and gender. Then we added residuals for both the Lee-Carter and TTD model using the same vector identifying years, ages and gender to account for the correlation between the different models. Besides comparing the forecasts of the estimated naïve and TTD models for the years 2008-2020, we will also forecast with the TTD model by using the growth rates as estimated by the naïve model. This reflects the current practice of comparing the TTD model with the naïve model as the growth rates in the TTD model are not adjusted for the influence of changing mortality rates. To compare our results with previous studies we also forecasted HCE for both the naïve and TTD model by using the average growth rate in per capita HCE for the period 1981-2007 (which equaled 2.3%) and forecasted HCE for the TTD model assuming a zero growth rate. Given that we used a standardized population a zero growth implies that HCE do not change according to the naïve model.

4.3 Results Regression models for health care expenditures Table 4.2 displays the results from the regression analyses for the three different regression models. It shows that in the naïve model the average annual growth rate of per capita HCE for all ages (as captured by the intercept) equals 1.36 % in the years 1981-2007 and that for the TTD model this rate is pushed upwards to 1.52% in the linear approximation and 1.54% in the non-linear model. To illustrate the age and gender gradient in growth rates the average annual growth rates by gender and age as estimated by the three different models for the period

68

TIME TO DEATH AND FORECASTING OF EXPENDITURES Table 4.2: Results from regression analyses. Robust standard errors (SE) are given. Coefficients for the dummy variables are shown in Table 4.B1 in Appendix 4.B. Naïve Model

TTD Model Approximation

Variable

Beta

SE

0

0.014

0.001

Variable *

Beta

TTD Model

SE

Variable

Beta

SE

0

0.015

0.001

*

0

0.015

0.001

*

γ1

16.382

2.711

*

R1

28.830

2.713

*

γ2

-0.338

0.063

*

R2

-0.578

0.064

*

γ3

-27.815

9.486

*

R3

-50.966

9.511

*

*significant at p<0.05

Figure 4.1: Estimated growth rates by gender and age for the three different regression models. Men TTD

-0.01

Growth rate HCE 0.01 0.03

Naive TTD approximation

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

65-69

80-84

Women Growth rate HCE 0.00 0.01 0.02 0.03

Naive TTD approximation

0-1

5-9

20-24

TTD

35-39 50-54 Age

1981-2007 is displayed in Figure 4.1. It shows that the assumption of constant growth rates for all ages is not a realistic one and that the growth rates are highest for newborns and the elderly – for the latter the level of costs are also higher. Furthermore, the influence of mortality rates is most relevant at older ages and decreases at the oldest age group. Noteworthy is that including mortality in the model as in TTD models increases the average growth rate. Ignoring the effect of mortality rates results in a so-called omitted variable bias as the growth rate

69

CHAPTER 4 Figure 4.2: Back calculated values for costs in the last year of life (decedent costs) and costs in ‘other years’ (survivor costs) for the year 2007. € 1,000 40 60 80 100

Per capita HCE men Average

0

20

Last year of life Other years

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

€ 1,000 40 60 80 100

Per capita HCE women Average

0

20

Last year of life Other years

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

absorbs the dampening effects of decreasing mortality over time. At most ages the growth rate is higher for the TTD model, except for 90+ males: this can be explained by the fact that mortality has increased rather than decreased for this group. The estimates of the growth rates between the non-linear and linear TTD model are almost identical. The explained variance is only a bit higher for the TTD models but F-tests indicated that both TTD models are preferred over the naïve model. Adding mortality, as is done in the TTD model, confirms our hypothesis: decreasing mortality rates will dampen the growth of per capita HCE. Let R  ( R1  R2  age  R3  baby  1) then for the estimates values of R1, R2 and R3, R will be positive for all values of age and indicate a strong age gradient in the TTD effect: for newborns costs in the last year of life are 6 times higher than in other years, at age 65 costs are about 20 times higher while at age 90 this will be about 3 times. Figure 4.2 displays the back calculated costs in the last year of life and other years using the estimates for R1, R2 and R3. The pattern we see here is very similar to micro-econometric studies. Costs are much higher in the last year of life but this difference decreases at older ages.

70

TIME TO DEATH AND FORECASTING OF EXPENDITURES Figure 4.3: Parameter estimates of the Lee-Carter model. Lee Carter model parameters Women

-8

V coefficients -6 -4 -2

Men

0

20

40

60

80

-10

K coefficients -5 0 5

Age

1990

1995 Year

0.06

1985

2005

Women

0.00

X coefficients 0.02 0.04

Men

2000

0

20

40

60

80

Age

Lee-Carter model for mortality Figure 4.3 displays the estimates of the parameters of the Lee-Carter model. The top panel are estimates of the age- and gender-specific v coefficients which represent the time-averaged log mortality rates. The second panel displays the general time trend in mortality rates which have been decreasing over time. Finally, the lowest panel indicates the age- and gender-specific interactions with the general time trend. From this panel it can be seen that in the highest category mortality rates have remained almost constant as indicated by the coefficients being close to zero. Forecasts Figure 4.4 displays the estimates of the mortality rates for 2020 forecasted using the Lee-Carter model together with the mortality rates for the years 1981 and 2007. From this figure it can be seen that mortality rates are expected to decrease for most ages except for the 90+.

71

CHAPTER 4 Figure 4.4: Mortality rates in 1981,2007 and 2020 as forecasted by the Lee-Carter model. Mortality rates men

1

Rate * 10,000 5 50 500

1981 2007 2020

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

Mortality rates women

5e-01

Rate * 10,000 1e+01 5e+02

1981 2007 2020

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

Figure 4.5 displays estimates of per capita HCE for the years 1981, 2007 and 2020 as estimated using the naïve model and TTD model of equation (4.7). Note that forecasts using the linear approximation were almost identical. From this figure it can be seen that the higher growth rates for the elderly will cause the age at which the expenditures peak to shift towards the higher ages for both men and women. Furthermore, the forecasts of the naïve and TTD model are remarkably similar. Only if growth rates are not adjusted for the decreasing trends in mortality rates (which is usually done in TTD forecasting models) we see that the forecasts of the naïve and TTD model differ. For the oldest age category (90+) in which mortality rates hardly change the TTD model with unadjusted growth rates results in similar forecasts as other models. Table 4.3 displays forecasts of per capita HCE for the two different models as well as the forecasts of life expectancy using the Lee-Carter model that serve as input for the forecasts of the TTD model. In the second row we see that per capita HCE will decrease (1) if mortality rates decrease, (2) if there are no other variables that influence forecasts, and (3) if the growth rate due to other causes is assumed to equal zero. In the following two rows we see that if the growth rate equals 0.023 for all ages in both the naïve and TTD model, per capita HCE in the naïve model

72

TIME TO DEATH AND FORECASTING OF EXPENDITURES Figure 4.5: Age- and gender-specific per capita health care expenditures for the years 1981, 2001 and 2020. Per capita HCE Men

0

10000



20000

1981 Data 2007 Data 2020 Naive model 2020 TTD model estimate 2020 TTD model unadjusted growth rates

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

Per capita HCE Women

0

5000



15000

1981 Data 2007 Data 2020 Naive model estimate 2020 TTD model estimate 2020 TTD model with unadjusted growth rates

0-1

5-9

20-24

35-39 50-54 Age

65-69

80-84

Table 4.3: Forecasts of life expectancy (in years) and growth in per capita HCE (in percentages). 95% prediction intervals are between brackets. Model type

2010

2015

2020

80.4 (79.6 – 81.2)

81.0 (79.7 – 82.2)

81.6 (80.0 – 83.1)

-0.6

-1.5

-2.4

7

19.8

34.1

6.4

18.0

31.0

Naïve model growth in per capita HCE [Growth rates model age specific as in equation (4.6)]

6.0 (4.3 – 7.9)

17.2 (13.6 – 20.8)

29.9 (24.6 – 35.4)

TTD model growth in per capita HCE [Growth rates model age specific as in equation (4.6)]

5.3

15.1

26.2

TTD model growth in per capita HCE [Growth rates model age specific as in equation (4.7)]

6.1 (3.7 – 8.4)

17.6 (13.2 – 22.3)

30.8 (24.4 – 37.9)

Predicted life expectancy using Lee-Carter model TTD model growth in per capita HCE [Zero growth rate] Naïve model growth in per capita HCE [0.023 growth rate for all ages] TTD model growth in per capita HCE [0.023 growth rate for all ages]

73

CHAPTER 4

is expected to grow with 34.1 percent and in the TTD model HCE is expected to grow with only 31.0 percent. If we now include the age and gender gradient in the growth rate as estimated using the naïve model [equation (4.5)], we see that the forecasts of per capita HCE are lower for both the naïve and TTD models but that the difference between these two forecasts remain approximately the same to the situation where growth rates are assumed equal for ages. However, if we account for the fact that estimates of growth rates are altered when the influence of mortality rate changes is included, we see that the big difference in predictions between the two models disappears and that the TTD model forecasts are even somewhat higher on average. However, since forecasts of HCE using the TTD model include uncertainty surrounding future developments of life expectancy, prediction intervals are a bit wider. 4.4 Discussion In this paper we have investigated the relevance of TTD with respect to the forecasting of macro-level HCE. Using age- and gender-specific time-series data on macro HCE our analyses showed that changes in mortality rates could indeed be linked to age-specific per capita HCE, thereby confirming the red herring hypothesis using macro-level data. Our model allowed us to back calculate values the ratio between HCE in the last year of life and other years. These values were similar to those found in micro-econometric studies. In this regard our results seem to justify the use of TTD in prediction models. However, we showed that forecasts for the period 2008-2020 based on the TTD model resulted in a growth in per capita HCE that was nearly identical to the predicted growth under the naïve model; though it should be noted that the TTD model forecasts had slightly larger confidence intervals as result of the uncertainty associated with the predicted mortality rates. The forecasts were similar because the estimate of the growth rates due to unidentified causes is altered if TTD is included in the model. In previous studies the size of this ‘residual’ growth rate was not adjusted for the influence of mortality rates and therefore assumed to be equal for the naïve and TTD model. When making this latter assumption, our findings were more in line with those of previous studies. While time-to-death models were initially conceived to show the 'true' effect of aging –if there even is such a thing– they are also used to make macro-health predictions. We found that, if time-to-death is indeed used to forecast HCE, any 'benefit' from including time-to-death in forecast models is likely to be negated by a higher growth rate in HCE. In a sense, our results are not surprising since we have witnessed a strong increase in HCE accompanied with a decrease in mortality rates in the last decades in many western countries. In all studies different time 74

TIME TO DEATH AND FORECASTING OF EXPENDITURES

windows and populations were used to study trends in mortality compared to the data used to study the relation between TTD and HCE. In our study we used the same population and time frame to study the relation between TTD and HCE. If mortality rates are assumed to continue to decrease at roughly the same pace, there is no reason to assume that the trend in HCE would suddenly level off because of this. Previous studies have compared forecasts based on TTD models with naïve models, as well as comparing the share of TTD in HCE growth to other determinants. However, these studies have all used the same growth rate estimates, which is ill-advised based on our findings. Since HCE in the last year of life are higher than in other years, declining mortality rates act as a dampening factor on HCE growth. The actual growth due to unidentified causes was therefore estimated to be higher when mortality rates were included in the prediction model. The question then arises as to whether it is not more efficient to use the naïve model without have to bother with forecasting mortality rates. Furthermore, in contrast with TTD prediction studies but in line with previous empirical studies, we find that growth in per capita HCE may very well be age-specific (Polder et al., 2002; Meara et al., 2004; Wong et al., 2012). This growth is greatest in the highest age groups, both in absolute as well as relative terms. We also found that the age at which the per capita HCE is highest increased over time. This growth might be explained by changes in practice, as well as medical technology, which in turn might be spurred on by the aging of the population. The higher growth for the elderly will cause the age profile to be steeper in the future. In this sense, both the age profile of per capita HCE and mortality rates have become more rectangular over time with the only difference that the level of per capita HCE has shifted upwards enormously for most age groups. In contrast to previous studies using macro-level data on health expenditure a unique feature of our study was that we used macro-level data stratified by age and gender. This allowed us to test the implications of the TTD prediction models. We created this panel data set of age- and gender-specific per capita HCE by decomposing total health expenditures into age- and gender-specific using data on clinical hospital admissions and average length of stay. This method has its drawbacks as it does not include day admissions which more and more form a large portion of total hospital care. The reason for this was that only had data on day admissions for from 1993 onwards and we wanted to create a longer time series. However, when we estimated the naïve and TTD models from equations (4.5) and (4.6) using a shorter time series but including day admissions to decompose total hospital HCE we found similar results: changes in mortality rates could still be linked to changes in per capita HCE and the growth rates were higher in the TTD model than in the naïve model. Furthermore, in our study we only looked at how annual changes in mortality rates affected per capita HCE in that same year. It has been argued that increased 75

CHAPTER 4

HCE are found for the last years of life. While this would complicate the empirical estimation using macro-level data (as it requires lagged influences) it would complicate forecast even more. For instance, if HCE are already higher q years before death, one needs to forecast mortality rates q years ahead in order to predict HCE one year ahead. It is worth mentioning that, although we have argued that the uncertainty of future mortality rates should be accounted for when using the TTD model, the Lee-Carter model is known for its rather small prediction intervals (De Waegenaere et al., 2010). Consequently, using other methods would have probably increased prediction intervals for HCE even more. We think our results are robust, and when we shortened the time window we got similar results for the regression models and for the forecasts. Also, different functional forms (e.g. polynomials) did not influence our main results. The inclusion of time-to-death in models has led to strong debates about endogeneity. Felder et al. (2010) have shown that while endogeneity is present, it is a small issue: time-to-death influences costs more than the other way around. Here, the difference between micro-level and macro-level analyses is important (Getzen, 2000). On an individual level health care is a necessity. And since being close to death is an approximation for the presence of severe morbidity and disability, it is even more so a necessity in the last years of life. However, since health care is a luxury on macro-level, decreases in age-specific mortality rates do not necessarily lead to lower age-specific HCE. In the TTD literature life expectancy is seen as a determinant of HCE. However, life expectancy has been increasing partly because of new medical technologies and the diffusion thereof (Cutler et al., 2006a; Cutler et al., 2006b). Further aging due to increasing life expectancy is probably not for free, as it probably is a result from progress in medical technology (Hall and Jones, 2007). The endogeneity debate focuses at individual differences in remaining life expectancy within a given period while for forecasting macro HCE it is relevant how the relation between life expectancy and HCE changes over time. It may be just as important to focus on what investments are needed to increase life expectancy rather than to focus just on the postponement of the death-related costs. In concordance with the TTD literature we did not attempt to try to explain the growth rates due to other causes. We hypothesize that investments needed to increase life expectancy are much more important than the postponement of health care demand due to increases in life expectancy unrelated to health care (i.e., lifestyle, nutrition and welfare). This is not too surprising if you look at results of economic evaluations in which initial investments are usually much larger than possible savings that might accrue due to less complications, avoidance or postponement of disease.

76

TIME TO DEATH AND FORECASTING OF EXPENDITURES

4.5 Conclusion We conclude that including TTD will not improve forecasts of HCE. More importantly, including TTD unnecessarily complicates forecast of HCE as it requires forecasts of mortality rates. Many ‘Red Herring’ studies have pointed out that ‘age has diverted attention from factors that really matter such as medical technology’, but the irony of this statement is that the fixation on proximity to death implicitly assumes that effects of changes in medical care and other determinants of growth in health expenditures can be estimated independently from TTD effects. Even though it may be that life expectancy is easier to predict than real determinants like disease and disability we still think it seriously complicates forecasts of HCE. The only relevance of TTD with respect to forecasting might be for forecasting of groups that differ in mortality risk. This might the case in economic evaluations of life prolonging interventions in which results from TTD models can be linked to models used in economic evaluations (Gandjour and Lauterbach, 2005; van Baal et al., 2011b). Time-to-death should mainly be used to investigate how health care spending is distributed over agegroups, i.e., to find the effect of aging. However, we find there to be too many caveats (inability to take into account time-to-death dependent growth, uncertainty and possible endogeneity on macro-level) that makes projections based on time-todeath a lot less attractive.

77

CHAPTER 4

Appendix 4.A: Generalization of the relation between mortality rates and health care expenditures Let S a () be the continuous average survival time for an individual aged a (for notational purposes, we omit any mention of time t and gender g). Furthermore, let Ca ( S a ) be a monotonous non-increasing average cost function of S a for age a . Assuming that [1] C a ( S a ) is completely governed by how far each individual aged a is away from their death and [2] mortality rates are not affected by the level of spending in the years leading up to the individuals reaching age a , it can be shown that, under specific parameter conditions of S a 1 , that C a ( S a ) will decrease if S a increases: ~

~

S a  S a  C a (S a )  Ca (S a )

(4.A1)

Proof of theorem Assume that C a ( S a ) is, for each a , non-increasing 2, right-continuous and finite ( 0  C a ( S a )   ). Also assume that the distribution function of S a , Fa (s ) , is continuous at 0, so that Fa (0)  0 . The expectation of the average cost function C a ( S a ) can be obtained by integrating over the product of survival time dependent cost function C a (s ) and f a (s ) : 



C a ( S a )   C a ( s ) f a ( s )ds   C a ( s )dFa ( s ) 0

(4.A2)

0

Integration of (4.A3) by parts gives: 

Ca (S a )

 C a () Fa ()  C a (0) Fa (0) 

 F (s)dC a

a

( s)

0



 C a ( ) 

 F (s)dC a

a

( s)

(4.A3)

0

We can rewrite the last-term of (4.A4): 1

We assume that lower mortality rates go hand in hand with higher survival. In other words, we can only speak of lower mortality rates if the following applies for Fa(s), the distribution function of Sa : F*a(s)< Fa(s) for all s, with the asterisk denoting the ‘new’ distribution. However, Fa(s) is characterized by a set of parameters, and changes in survival are governed by changes in these parameters. Under specific values for these parameters, the inequality cannot be found for all s, as the distribution function might intersect at some s. See the end of this section for more information. 2 The empirical results from the Red Herring literature all indicate that the HCE are non-increasing with increasing survival, so this assumption is not likely to be a stretch at all.

78

TIME TO DEATH AND FORECASTING OF EXPENDITURES







 Fa ( s )dC a ( s )  0

 F (s)d 1  C a

a

( s )

(4.A4)

0

Equation (4.A5) can be better understood when considering an approximation of the integral, based on Riemann sums, and rewriting the approximation:   Fa ( s )dC a ( s )    Fa ( s * )C a ( s i 1 )  C a ( si )  i

 F (s )1  C (s )  1  C (s )   F ( s )1  C ( s )   1  C ( s )  

*

a

a

i

i 1

a

i

*

a



i 1

a

i

 F (s)d 1  C a

a

a

i

( s )

If we normalize C a (s ) by dividing by a constant C a (0) , we get for (4.A5): 





 Fa (s)d 1  C a (s)  C a (0) Fa (s)d 1  0

 0



C a (s)   C a (0) 

(4.A5)

Substituting (4.A6) into (4.A4) gives: 

Ca (S a )





 C a ()  C a (0) Fa ( s ) d 1  0



 C a ()  C a (0) E Fa ( a ) ,

C a ( s)   C a ( 0) 

(4.A6)

where  a is a random variable with cumulative distribution function (CDF) C (s) . Ga (s ) fulfills the requirements for a CDF, because [1] Having Ga ( s)  1  a C a (0) normalized in step (4.A2), we ensured that 0  Ga ( s )  1 and [2] C a (s ) is a nonincreasing function, which means Ga (s ) is a non-decreasing function. Assuming that C a () and C a (0) are not affected by changing survival rates (which was one of the theorem conditions), changes in C a ( S a ) are completely determined by E Fa ( a ). Now assume that if the mortality decreases, then the CDF of  a will face a shift to the right:

79

CHAPTER 4 ~ ~ ~ S a  S a s a  F a ( a )  Fa ( a ) a  E  F a ( a )  EFa ( a )  a  

Following from (4.A7) C a ( S a ) will then decrease: ~ ~      E  F a ( a )  E Fa ( a )  E C a ( S a )  E C a ( S a ) .    

(4.A7)

Let F ( s; k ,  ) be the CDF that of the survival time s . We show that the theorem holds for three frequently used survival functions (Weibull, log-logistic and Gompertz-Makeham), under specific conditions. Note that the following examples apply to survival time s at age 0, but can easily be generalized to a survival time s * for all ages by adding a location parameter age in each CDF, i.e., by substituting s  s *  age . Weibull: F ( s; k ,  )  1  e

s ( ) k



(Case a: changes in the shape parameter k ) Let Fa ( s )  1  e 

s ( ) k

~

~

, F a (s)  1  e



s ( ) k



. Then

~

~

~

F a ( s)

k ~

0s

0k k

0s

0k k

~

C a ( s)

~

F a ( s )  Fa ( s ) ~

F a ( s )  Fa ( s )

~

C a ( s)  C a ( s) ~

C a ( s)  C a ( s)

(Case b: changes in the scale parameter  ) Let Fa ( s )  1  e k 0k

s ( )k



~

, F a (s)  1  e

s ( ~ ) k



~

~



F a ( s) ~

0 

Log-logistic: F ( s; k ,  )  1 

~

F a ( s )  Fa ( s ) 1

1  s 

(Case a: changes in the shape parameter k )

80

. Then

k

~

C a ( s) ~

C a ( s)  C a ( s)

TIME TO DEATH AND FORECASTING OF EXPENDITURES

Let Fa ( s)  1 

1  s 

k

1 s

1  s

1  s 

. Then

~ k

~

~

F a ( s)

k

0 

1

~

, F a (s)  1 

~



0

1

C a ( s)

~

~

~

F a ( s )  Fa ( s )

0k k

C a ( s)  C a ( s)

~

~

~

F a ( s )  Fa ( s )

0k k

C a ( s)  C a ( s)

(Case b: changes in the scale parameter  ) Let Fa ( s)  1  ~

1

1  s 

k

~

 ~

~ 1    s   

k

.Then

~

F a ( s)

0 

1

~

, F a (s)  1 

C a (s)

~

F a ( s )  Fa ( s )

~

C a ( s)  C a (s)

Gompertz-Makeham: F ( s;  ,  ,  )  1  e

 s 



 s e 1 



(Case a: changes in the scale parameter  ) ~   ~   s  e s 1  s  e s 1  , F a ( s)  1  e  . Then Let Fa ( s )  1  e ~

~



~

F a ( s)

C a ( s)

~

~

~

F a ( s )  Fa ( s )

0 

C a ( s)  C a ( s)

(Case b: changes in the baseline mortality parameter  ) Let Fa ( s )  1  e

 s 



 s e 1 



~

~

, F a ( s)  1  e

~

~

 ~



 s e 1 



. Then

~

F a ( s)

0 

 s 

C a ( s)

~

F a ( s )  Fa ( s )

~

C a ( s)  Ca ( s)

(Case c: changes in the senescent component  ) Let Fa ( s )  1  e

 s 



 s e 1 



~

, F a (s)  1  e

~    s   e  s 1  ~   

.

~

Then no solution exists such that F a ( s )  Fa ( s ) s . 81

CHAPTER 4

Appendix 4.B: Additional regression output Table 4.B1: Regression coefficients with robust standard errors (SE). Naïve Model Variable

Beta

TTD Model Approximation SE

Variable

Beta

TTD Model

SE

Variable

Beta

SE



0.018

0.004

*

β1

0.017

0.005

*

β1

0.017

0.005



-0.017

0.007

*

β2

-0.018

0.007

*

β2

-0.018

0.007

*



-0.028

0.011

*

β3

-0.029

0.011

*

β3

-0.029

0.011

*



-0.016

0.006

*

β4

-0.017

0.006

*

β4

-0.017

0.006

*



-0.016

0.006

*

β5

-0.017

0.006

*

β5

-0.017

0.006

*



-0.014

0.004

*

β6

-0.015

0.004

*

β6

-0.015

0.004

*



-0.012

0.004

*

β7

-0.014

0.004

*

β7

-0.014

0.004

*



-0.011

0.005

*

β8

-0.012

0.005

*

β8

-0.012

0.005

*



-0.009

0.004

*

β9

-0.010

0.004

*

β9

-0.010

0.004

*

10

-0.007

0.004

*

β 10

-0.008

0.003

*

β 10

-0.008

0.003

*



-0.006

0.003

β 11

-0.006

0.003

β 11

-0.006

0.003

12

-0.004

0.004

β 12

-0.004

0.004

β 12

-0.004

0.004

13

0.001

0.003

β 13

0.002

0.003

β 13

0.002

0.003

14

0.003

0.003

β 14

0.005

0.003

β 14

0.005

0.003

15

0.006

0.004

β 15

0.010

0.004

*

β 15

0.010

0.004

*

16

0.011

0.003

*

β 16

0.016

0.003

*

β 16

0.016

0.003

*

17

0.016

0.004

*

β 17

0.022

0.004

*

β 17

0.022

0.003

*

18

0.019

0.004

*

β 18

0.023

0.004

*

β 18

0.023

0.004

*

19

0.025

0.006

*

β 19

0.027

0.005

*

β 19

0.027

0.005

*

20

0.033

0.008

*

β 20

0.030

0.008

*

β 20

0.030

0.008

*

*

21

0.017

0.005



-0.015

0.008

23

-0.020

0.008

24

-0.009

0.005

25

-0.007

0.004

26

-0.010

0.004

27

-0.005

0.004

28

0.010

0.002

29

0.000

0.004

30

-0.011

0.004

31

-0.008

0.004

*

β 21

0.016

0.005

*

β 21

0.016

0.005

*

β 22

-0.016

0.008

*

β 22

-0.016

0.008

*

β 23

-0.022

0.008

*

β 23

-0.022

0.008

*

β 24

-0.010

0.005

β 24

-0.010

0.005

β 25

-0.008

0.004

β 25

-0.009

0.004

β 26

-0.011

0.004

β 26

-0.012

0.004

β 27

-0.006

0.004

β 27

-0.006

0.004

β 28

0.009

0.002

β 28

0.009

0.002

β 29

-0.001

0.004

β 29

-0.001

0.004

*

β 30

-0.013

0.004

*

β 30

-0.012

0.004

*

*

β 31

-0.009

0.003

*

β 31

-0.009

0.003

*

*

*

*

* *

* *

32

-0.005

0.003

β 32

-0.006

0.002

β 32

-0.006

0.002

3

-0.001

0.002

β 33

-0.002

0.002

β 33

-0.002

0.002

* *

4

0.002

0.004

β 34

0.001

0.003

β 34

0.001

0.003

35

0.003

0.003

β 35

0.003

0.003

β 35

0.003

0.003

36

0.005

0.004

β 36

0.006

0.004

β 36

0.006

0.004

37

0.008

0.004

*

β 37

0.010

0.004

*

β 37

0.010

0.004

*

38

0.012

0.004

*

β 38

0.014

0.004

*

β 38

0.015

0.004

*

39

0.017

0.006

*

β 39

0.019

0.005

*

β 39

0.019

0.005

*

82

TIME TO DEATH AND FORECASTING OF EXPENDITURES Table 4.B1 (continued). Naïve Model Variable

TTD Model Approximation Variable

TTD Model

Beta

SE

Beta

SE

Beta

SE

λ1982

-0.025

0.004

*

λ1982

-0.026

0.003

*

λ1982

-0.026

0.003

*

λ1983

-0.053

0.004

*

λ1983

-0.054

0.004

*

λ1983

-0.054

0.004

*

λ1984

-0.049

0.004

*

λ1984

-0.050

0.004

*

λ1984

-0.050

0.004

*

λ1985

-0.033

0.004

*

λ1985

-0.036

0.004

*

λ1985

-0.036

0.004

*

λ1986

-0.016

0.004

*

λ1986

-0.018

0.003

*

λ1986

-0.018

0.003

*

λ1987

-0.001

0.004

λ1987

0.004

0.005

λ1987

0.004

0.005

λ1988

0.008

0.003

λ1989

0.002

0.003

*

λ1988

0.007

0.003

λ1989

-0.003

0.003

Variable

*

λ1988

0.007

0.003

λ1989

-0.003

0.003

*

λ1990

0.030

0.003

*

λ1990

0.031

0.003

*

λ1990

0.031

0.003

*

λ1991

0.025

0.003

*

λ1991

0.025

0.003

*

λ1991

0.025

0.003

*

λ1992

0.010

0.004

*

λ1992

0.011

0.004

*

λ1992

0.011

0.004

*

λ1993

-0.023

0.003

*

λ1993

-0.031

0.003

*

λ1993

-0.031

0.003

*

λ1994

-0.020

0.004

*

λ1994

-0.015

0.004

*

λ1994

-0.015

0.004

*

λ1995

-0.013

0.004

*

λ1995

-0.015

0.004

*

λ1995

-0.015

0.004

*

λ1996

-0.011

0.003

*

λ1996

-0.012

0.003

*

λ1996

-0.012

0.003

*

λ1997

-0.008

0.003

*

λ1997

-0.006

0.004

λ1997

-0.005

0.003

λ1998

0.003

0.003

λ1998

0.003

0.003

λ1999

0.016

0.004

λ1999

0.016

0.004

λ2000

0.009

0.005

λ2000

0.009

0.005

λ1998

0.004

0.003

λ1999

0.018

0.004

λ2000

0.009

0.005

λ2001

0.063

0.004

*

λ2001

0.064

0.004

*

λ2001

0.064

0.004

*

λ2002

0.047

0.005

*

λ2002

0.045

0.005

*

λ2002

0.045

0.005

*

λ2003

0.019

0.003

*

λ2003

0.020

0.003

*

λ2003

0.020

0.003

*

λ2004

-0.001

0.006

λ2004

0.004

0.005

λ2004

0.004

0.005

λ2005

0.002

0.004

λ2005

0.003

0.004

λ2005

0.003

0.004

λ2006

0.004

0.006

λ2006

0.007

0.006

λ2006

0.007

0.006

*

*

*

*significant at p<0.05

83

CHAPTER 4

84

86

Chapter 5 Predictors of Long-Term Care Utilization by Dutch Hospital Patients Aged 65+ ‡

Long-term care is often associated with high health care expenditures. In the Netherlands, an ageing population will probably increase the demand for long-term care within the near future. The development of risk profiles will not only be useful for projecting future demand, but also for providing clues that may prevent or delay long-term care utilization. Here, we report our identification of predictors of long-term care utilization in a cohort of hospital patients aged 65+ following their discharge from hospital discharge and who, prior to hospital admission, were living at home. The data were obtained from three national databases in the Netherlands: the national hospital discharge register, the long-term care expenses register and the population register. Multinomial logistic regression was applied to determine which variables were the best predictors of long-term care utilization. The model included demographic characteristics and several medical diagnoses. The outcome variables were discharge to home with no formal care (reference category), discharge to home with home care, admission to a nursing home and admission to a home for the elderly. The study cohort consisted of 262,439 hospitalized patients. A higher age, longer stay in the hospital and absence of a spouse were found to be associated with a higher risk of all three types of longterm care. Individuals with a child had a lower risk of requiring residential care. Cerebrovascular diseases [relative risk ratio (RRR) = 11.5] were the strongest disease predictor of nursing home admission, and fractures of the ankle or lower leg (RRR = 6.1) were strong determinants of admission to a home for the elderly. Lung cancer (RRR = 4.9) was the strongest determinant of discharge to the home with home care. These results emphasize the impact of age, absence/presence of a This chapter is based on: Wong A, Elderkamp-de Groot R, Polder JJ, Van Exel NJA. 2010. BMC Health Services Research 6(10). ‡

CHAPTER 5

spouse and disease on long-term care utilization. In an era of demographic and epidemiological changes, not only will hospital use change, but also the need for long-term care following hospital discharge. The results of this study can be used by policy-makers for planning health care utilization services and anticipating future health care needs.

5.1 Introduction In countries all over the world, the health sector faces the challenge of an ageing population. It is expected that the prevalence of chronic diseases will rise, and thus the number of people in need of long-term care. This global development will have a significant effect on health care services in terms of capacity, planning and costs. The World Health Organization (WHO) defines long-term care as ‘the system of activities undertaken for persons that are not fully capable of self-care on a longterm basis, by informal caregivers (family and friends) or by formal caregivers (professionals). It encompasses a broad array of services, delivered in homes, in communities or in institutional settings’ (WHO, 2002a). The Netherlands provides a wide range of long-term care services and facilities, most of which are completely or almost completely covered by national health insurance plan (Exceptional Medical Expenses Act (Dutch Ministry of Health, Welfare and Sport, 2009). These include home care and support services, varying from the very basic to more intensive forms, and two types of residential facilities – homes for the elderly and nursing homes (Statistics Netherlands, 2010; Portrait et al., 2000; Poos et al., 2008; also see Appendix 5.A for a summary). Nursing homes differ from homes for the elderly in that the care and treatments provided in the former are much more intensive than those provided in the latter, with up to 24-h care being provided for the terminally ill in nursing homes. In 2005, about 18.5% of the total health care budget in the Netherlands was spent on long-term care provided by nursing homes, homes for the elderly and home care, largely to individuals aged 65+ (Poos et al., 2008). In the Netherlands, life expectancy has risen from 70.4 years for men and 72.7 years for women in 1950 to 78.0 and 82.3 years, respectively, in 2007 (National Public Health Compass, 2008). Due to the ageing of the post-war ‘baby-boomers’ and a further predicted increase in life expectancy, a substantial further increase in the proportion of elderly in the general population is anticipated from 2010 onwards. Given the high costs associated with long-term care and the anticipated growth in the demand for such care in the next two decades, there has been an increasing interest among health professionals and policy-makers for less costly care alternatives. This has resulted in the development of risk profiles of future health

88

PREDICTORS OF LONG-TERM CARE UTILIZATION

care utilization that can be used to provide information on the predictors and probabilities of the types of health care. These profiles are not only useful for projecting future demand for different types of health care, but they also provide some degree of insight into possibilities for preventing and delaying long-term care use in general and the costly admission to residential care facilities in particular (Tomiak et al., 2000). To date, studies on the 65+ population have mainly focussed on determinants of admission to a nursing home. Age has been found to be a strong and consistent predictor of admission to a nursing home (Ahmed et al., 2003; Cai et al., 2009; Campbell et al., 2004; Greene and Ondrich, 1990; Miller and Weissert, 2000; Mustard et al., 1999; Ohwaki et al., 2005; Wolinsky et al., 1992), as has the need for care, with lower physical and mental health status associated with an increased risk of institutionalization (Tomiak et al., 2000; Cai et al., 2009; Campbell et al., 2004; Ohwaki et al., 2005; Gaugler et al., 2007; Hancock et al., 2002). Depression has also been determined to be an important predictor of admission to a nursing home (Ahmed et al., 2007; Harris, 2007; Harris et al., 2006; Nuyen et al., 2006; Onder et al., 2007). On the basis of an epidemiological study, Bharucha et al. (2004) reported that dementia was the strongest disease predictor of admission to a nursing home, with dementia patients having a nearly fivefold higher risk of requiring nursing home facilities. A number of other studies have reported an association with living alone prior to hospitalization (Cai et al., 2009; Hancock et al., 2002; de Pablo et al., 2004; Freedman, 1996; Glader et al., 2003; Rundek et al., 2000). Although this latter finding is not consistent across studies, it emphasizes the protective role often played by spouses and other relatives and the potential role of these informal caregivers in preventing or delaying the institutionalization of elderly patients. A few studies have found that women are more likely to be admitted to a nursing home than men (Ohwaki et al., 2005; Breeze et al., 1999; Grundy and Glaser, 1997). However, this effect may in part be related to two of the above-mentioned determinants, age and living alone, as women have a higher life expectancy and tend to outlive their partner. The traditional division of roles in the household has also been put forward as a possible explanation: according to this argument, women are both more used to and willing to perform household and caring tasks, making it more likely that their male partners will be discharged home (rather than vice versa). Studies addressing ethnicity have reported that being non-white decreases the likelihood of institutionalization (Cai et al., 2009; Miller and Weissert, 2000; Ottenbacher et al., 2008; Salive et al., 1993; Stansbury et al., 2005). The relationship between socio-economic status and admission to an institution has been found to be rather ambiguous, but some studies have reported a positive association between admission to a care facility and a higher income or level of education (Miller and Weissert, 2000; Mustard et al., 1999; de Pablo et al., 2004).

89

CHAPTER 5

The majority of studies in this field have only considered admission to a nursing home. In addition, these studies have largely focussed on groups of patients with a specific disease, mostly dementia or stroke. Less is known about predictors of discharge from hospital to other types of long-term care services/facilities, particularly in the context of different diseases. In the study reported here, we have investigated and compared predictors for the hospital discharge of Dutch patients aged 65+ to alternative types of long-term care – i.e., discharge to home without any formal care, discharge to home with home care, discharge to a nursing home and discharge to a home for the elderly. The health care system in the Netherlands is very comprehensive and therefore allows for a clear view of the whole range of health care needs of hospital patients. Using a target group of patients aged 65+ who were living in their own homes prior to hospital admission, we have applied a regression model to predict the long-term care needs of a new patient with a specific combination of diseases, length of hospital stay and social–demographic characteristics. In times of rapid demographic and epidemiological changes, the results of these analyses can help policy-makers to address future health care needs.

5.2 Methods Data This study had a retrospective design based on individual patient-level data obtained from the database of Statistics Netherlands for 2005. Consent for the data was given by Statistics Netherlands. The dataset comprises three registers, i.e., the national hospital discharge register, the exceptional health care expenses register and the Dutch population register. The hospital discharge register contains administrative patient data on hospital admissions, covering all general and academic hospitals and most specialized hospitals in the Netherlands. It includes date of admission and discharge and medical data, such as diagnoses and treatment; outpatient and ambulatory care are not included. The Dutch population register contains demographic data of all registered inhabitants in the Netherlands, including gender, date of birth, zip code, family relations and date of death. The hospital discharge register records were deterministically linked to population register records by Statistics Netherlands using date of birth, gender and zip code as primary linkage keys. The linkage was of good quality as 87.6 % of the records were successfully linked. There is a slight bias towards elderly from a non-Western origin but, overall, the data are considered to be representative (de Bruin et al., 2004). Data on long-term care use were retrieved from the register of exceptional health care expenses, which was linked by Statistics Netherlands to the population register records using the same primary linkage keys as mentioned above. The 90

PREDICTORS OF LONG-TERM CARE UTILIZATION

long-term care register contains information on the starting date and the type and amount of long-term care used. In cases where individuals required more than one type of long-term health care, only the type consumed directly after discharge was considered. The combined dataset therefore comprised basic demographic characteristics and hospital history (e.g. date of admission and discharge, diagnosis) of all individuals discharged from hospital in the Netherlands in 2005 and, when applicable, data on long-term care consumption after hospital discharge. Study population The target group of our study comprised long-term care users aged 65+ who were living at home prior to their admission to hospital, and not utilizing any kind of formal care at that time. This study population was derived as follows. The main source for the population were the hospitalized patients in 2004 (n=1,414,142). All individuals in this dataset aged 65 years and older were selected (n=434,142 individuals). Because we focussed on risk factors for long term care utilization, other than end-of-life care, we also excluded hospitalized patients who died in 2004 or 2005. Within the remaining individuals (n=323,923), hospitalized patients were identified who lived at home prior to hospitalization and did not use any type of formal care. This group was derived by linking the remaining sample to longterm care data for 2004, and excluding individuals who had consumed long-term care in 2004. The final study group therefore consisted of 262,439 persons (11.5 % of the population aged 65+ in the Netherlands). For each individual, exactly one year of hospitalization data was used to determine the number of hospital days and diagnoses. More specifically, for individuals that utilized long-term care in 2005 the study-year was defined as the 365 days preceding the first contact with long-term care in 2005. Since the hospital register also contained admissions from 2005, as well as from 2004, all admissions within this year could be retraced. For individuals that did not utilize long-term care in 2005, the study year was defined by picking a random date in 2005, and taking exactly one year preceding this random date. Dependent variables The response variables were the four possible destinations after discharge from hospital: (1) a nursing home, (2) a home for the elderly, (3) home with home care or (4) or home without formal care. The dependent variables to be modelled were the probabilities of each outcome. Since the patients who died during hospitalization or in the year following hospital admission were excluded from our analysis, the probabilities calculated are conditional on the individual being alive at least 1 year after being admitted to hospital.

91

CHAPTER 5

Independent variables from the population register Several demographic variables were included in the various databases that were linked in this study and used to explain the probability of long-term care utilization. Age and age squared were included as continuous variables, with age squared included because of the anticipated non-linear relationship between age and longterm care utilization (Gaugler et al., 2007; Himes et al., 2000). Further examination of the data confirmed that these relationships were not of linear but parabolic. Cubic splines for age confirmed this parabolic relationship, but the age and age squared terms were retained as the splines offered nearly identical model predictions over age. Other demographic variables were gender and the presence of a spouse or a child in the patient’s household. Interactions between spouse and gender and between spouse and child were also examined. Independent variables from the hospital register Length of stay (LOS) in the hospital was defined as the total number of hospital days, accumulated over exactly one year (see above paragraph ‘Study population’), with longer hospital stays regarded as a possible indicator of more severe diseases and complications. As binning the LOS and calculating average probabilities for each outcome suggested a parabolic relationship between LOS and each outcome, length of stay squared was also included in the analysis. We also considered LOS smoothing splines, but this resulted in a nearly identical fit. LOS and LOS squared terms were preferred because of their relatively simpler nature. Medical diagnoses were included, based on the International Shortlist for Hospital Morbidity Tabulation (ISHMT, see WHO (2010)). Only primary diagnoses were considered. This shortlist was compiled by the Hospital Data Project (HDP) of the European Union Health Monitoring Programme and is aimed at maximizing the statistical comparability of hospital care. In 2005, Eurostat, the Organization for Economic Co-operation and Development (OECD) and the Family of International Classifications (WHO-FIC) adopted this shortlist for data collection and presentation. This shortlist is categorized in 130 groups and 20 chapters. Diagnoses related to pregnancy and childbirth, perinatal conditions and congenital malformations were not relevant to our study population and therefore excluded from the analyses. Groups that included unclassified symptoms and factors influencing health status were also considered to be too broad and also excluded. Statistical analysis A regression model was used to explore the predictors of long-term care utilization after hospital discharge. Since the outcome of interest, the discharge destination, has multiple categories, a multinomial logistic regression model was used (Hosmer and Lemeshow, 2000). Discharge to home without any formal care was chosen as the reference category. For the demographic variables, male gender, no spouse 92

PREDICTORS OF LONG-TERM CARE UTILIZATION

present and no child present were chosen as reference categories. For the disease variables, the reference category was the group not having that specific disease. A backwards stepwise regression (including demographic variables) was used to determine which diagnoses needed to be included to simplify the model. The use of all remaining diagnoses (110) simultaneously did not lead to convergence (which suggests the model is overspecified), so it was opted to perform a backwards regression per disease chapter as specified in the ISHMT tabulation. Those diseases with a significant coefficient (α ≤ 0.10) for at least one outcome were included in the final model. A total of 23 diagnoses were used in the final model (referred to as ‘core diseases’ hereafter). This model was run in Stata 9 with the – mlogit– command. The results are reported as relative risk ratios (RRR). Interpretation of the RRR is similar to that of the odds ratio, with the exception that it involves a specific outcome as a reference group instead of a group with a negative outcome. In terms of absolute effect of the probabilities, the RRR is hard to interpret. Similar to the simple logistic regression model, the model has a non-linear character that does not allow the researcher to associate a RRR in itself with changes in absolute probabilities. The multinomial logistic model also has the added difficulty of multiple categories, all of which can be considered as competing risks. In some cases this means that while a covariate dummy may have a positive RRR for a certain outcome, the probability of the outcome for the dummy taking on a value of 1 may actually be lower than that when the dummy is zero. This occurs when the probability of another outcome category increases even more, such that the probability of the outcome of interest would fall relative to that other outcome (Statacorp, 2007). For this reason, a straightforward interpretation of the RRR is complicated. A marginal (partial) effect can be computed, but this is also not easy to interpret, as this effect depends on the chosen set of values for the covariates in terms of effect sign and effect size. We therefore opted to make multiple model predictions, changing the value of one covariate while holding all other covariates constant (Statacorp, 2007). By plotting these model predictions, we were able to perform more comprehensive comparisons.

5.3 Results Descriptive analysis The demographic and clinical characteristics of the study cohort are shown in Table 5.1. The patient population comprised 262,439 individuals with a mean age of 74.2 years [standard deviation (SD) = 6.4] and an almost equal distribution of males and females. Most of the patients lived only with their spouse. The median

93

CHAPTER 5 Table 5.1: Summary statistics and discharge destination after hospital discharge according to the characteristics of the study population. n

Share of total sample

149,750

57.1%

75-84

99,242

37.8%

75.1%

20.2%

1.6%

3.2%

84+

13,447

5.1%

64.7%

23.6%

4.5%

7.2%

Male

132,326

50.4%

82.8%

14.2%

0.8%

2.2%

Female

130,113

49.6%

77.9%

18.2%

1.5%

2.5%

Categories

No formal Home Care (n = care 42,418) (n = 210,972)

Home for the elderly (n = 2,918)

Nursing home (n = 6,131)

0.5%

1.4%

Age, years (mean = 74.23, SD = 6.39) 65-74

85.3%

12.9%

Sex

Family situation Living Alone Spouse Child Hospitalization duration, days (median = 4, SD = 13.79) 01-10 11-20 20+ Diagnoses Intestinal, stomach and rectum cancer Lung cancer Uterus cancer Ovary cancer Prostate cancer Bladder cancer Diabetes mellitus Dementia Schizophrenia Alzheimer’s disease Epilepsy Heart Failure Cerebrovascular diseases COPD Alcoholic liver disease Infections of skin Coxarthrosis Gonarthrosis Glomerular disorders Intracranial injury Fracture of elbow and forearm Fracture of femur Fracture of ankle or lower leg Strictly other diagnoses

70,596

26.9%

73.4%

20.4%

2.6%

3.6%

186,674 18,604

71.1% 7.1%

83.1% 81.2%

14.4% 0.6%

1.8% 15.8%

0.7% 2.5%

190,094 40,824 31,521

72.4% 15.6% 12.0%

87.3% 66.1% 57.1%

11.2% 27.7% 31.1%

0.6% 2.2% 2.5%

0.8% 4.1% 9.3%

5,744

2.2%

49.8%

46.1%

1.6%

2.6%

2,957 662 414 3,351 2,538 10,720 1,064 640 240 1,087 8,688 9,411 9,107 153 915 9,220 6,910 1,329 1,557 1,188 4,327 1,141 190,010

1.1% 0.3% 0.2% 1.3% 1.0% 4.1% 0.4% 0.2% 0.1% 0.4% 3.3% 3.6% 3.5% 0.1% 0.4% 3.5% 2.6% 0.5% 0.6% 0.5% 1.7% 0.4% 72.4%

45.8% 65.1% 48.1% 77.1% 74.2% 68.1% 48.9% 52.2% 57.1% 67.3% 65.0% 61.5% 68.3% 54.3% 67.2% 62.3% 70.3% 68.9% 72.9% 67.9% 46.2% 57.6% 86.6%

50.1% 32.0% 47.3% 20.2% 23.9% 26.3% 26.5% 28.1% 23.8% 22.6% 29.4% 17.9% 27.6% 34.6% 28.4% 29.6% 23.1% 29.0% 17.4% 24.4% 29.9% 26.7% 11.8%

1.3% 1.7% 1.9% 0.8% 0.6% 1.5% 4.6% 3.8% 2.9% 2.2% 2.3% 1.4% 1.5% 2.6% 1.2% 3.4% 2.1% 1.0% 2.2% 2.5% 5.5% 4.8% 0.7%

2.9% 1.2% 2.7% 2.0% 1.3% 4.1% 20.0% 15.9% 16.3% 7.8% 3.3% 19.2% 2.6% 8.5% 3.2% 4.7% 4.5% 1.1% 7.5% 5.1% 18.4% 10.9% 0.9%

Abbreviations: SD, Standard Deviation; COPD, Chronic Obstructive Pulmonary Disorder. Sample size n = 262,439. With the exception of the ‘n’ column, which provides absolute numbers, all other values are given as percentages. The percentages in each column for diagnoses do not add up to 100% because patients can have more than one diagnosis. The percentages given for discharge destination do add up to 100%.

94

PREDICTORS OF LONG-TERM CARE UTILIZATION

length of hospital stay was 4 days. The five most prevalent diseases among the 23 diagnoses included in the model were diabetes mellitus, cerebrovascular diseases, coxarthrosis, chronic obstructive pulmonary disease (COPD) and heart failure. About 72% of the patients included in the study had none of these 23 diseases, but they did have at least one of the remaining 87 diseases (79% of all diseases that were considered for model inclusion). Averaged over the study population, most patients (80.3 %) returned home without home care after being discharged from hospital. The others returned home with home care (16.1%) or were admitted to a nursing home (2.5%) or a home for the elderly (1.1%). This distribution of outcomes was found to be distinctly different for those patients with the core diseases, as they had a higher probability of formal care. The characteristics of the study population in terms of discharge destinations are summarized in Table 5.1. We identified a positive association between increasing age and the use of long-term care. The proportion of older adults who were institutionalized was greatest among the group of 85+. Males had a higher probability of being discharged to their homes after hospital admission than females. People living alone or with their children only were more likely to be discharged to a nursing home or homes for the elderly than those who lived with a spouse or with a spouse and children. Patients with prostate cancer had the highest probability of going home without home care (77%). Patients with dementia had the highest probability of a discharge to a nursing home (20%), while patients with a fracture of femur had the highest probability of being discharged to a home for the elderly (5.5%). The latter also had the highest probability of receiving formal care after their hospital discharge (100 – 46.18 = 53.82%). Demographics predictors of discharge destination The results of the multinomial regression model are shown in Table 5.2. In describing these results, we made a distinction between demographic predictors, hospital care utilization predictors and disease predictors. The presence of a spouse can be seen to lower the probability of all three types of long-term care, with a particularly strong effect observed for residential care (homes for the elderly and nursing homes). Although women were found to be more likely than men to use home care services or go to a home for the elderly after discharge from the hospital, there was no significant difference between the sexes in terms of nursing home care. The sex of the spouse also did not significantly affect the probability of nursing home care. The presence of a child was found to be significant only for residential care, while having both a spouse and a child had no significant effect on the probability of long-term care. Age and age squared terms were both significant, but these were hard to interpret, as mentioned above. To further clarify the results, we

95

CHAPTER 5 Table 5.2: Relationship between background/disease predictors and discharge destination (relative risk ratios, standard errors and statistical significance). Home care

Home for the elderly

Nursing home

Variable

RRR

SE

Sgn

RRR

SE

Sgn

RRR

SE

Sgn

Demographics Age (in years) squared Age (in years) Female Presence of a spouse Female spouse Presence of a child Female spouse and child

1.00 1.37 1.22 0.65 1.25 0.97 1.03

0.00 0.02 0.03 0.01 0.03 0.04 0.05

*** *** *** *** ***

1.00 1.55 1.14 0.26 1.20 0.64 1.10

0.00 0.08 0.07 0.02 0.10 0.07 0.21

*** *** * *** * ***

1.00 1.34 0.97 0.48 1.12 1.17 0.85

0.00 0.05 0.05 0.02 0.07 0.09 0.09

*** ***

Hospital care utilization Hospital days squared Hospital days

1.00 1.08

0.00 0.00

*** ***

1.00 1.10

0.00 0.00

*** ***

1.00 1.12

0.00 0.00

*** ***

3.36

0.10

***

1.72

0.19

***

1.25

0.11

*

4.89 2.05 2.88 1.60 1.43 1.32 2.08 1.62 1.20 1.42 1.26 1.17 1.29 2.63 1.61 2.54 1.69 1.31 1.10 1.86 2.03 2.21

0.20 0.18 0.31 0.08 0.07 0.03 0.18 0.16 0.22 0.11 0.03 0.04 0.03 0.49 0.13 0.06 0.05 0.09 0.08 0.14 0.08 0.16

*** *** *** *** *** *** *** ***

2.61 1.65 1.93 1.14 0.61 1.08 4.09 2.23 1.08 2.25 1.14 1.32 1.06 3.84 0.95 5.16 2.57 0.73 1.53 2.27 3.83 6.12

0.44 0.52 0.71 0.24 0.16 0.09 0.69 0.51 0.46 0.49 0.09 0.12 0.10 2.03 0.30 0.34 0.23 0.21 0.28 0.45 0.30 0.92

***

2.22 0.95 1.45 1.15 0.51 1.00 7.50 3.89 2.14 1.33 0.65 11.55 0.73 4.03 1.00 4.93 3.89 0.32 2.21 2.41 9.30 8.18

0.27 0.35 0.47 0.16 0.10 0.06 0.83 0.55 0.51 0.18 0.05 0.43 0.05 1.38 0.21 0.28 0.26 0.09 0.27 0.38 0.48 0.93

***

Hospital care diagnoses Intestinal, stomach and rectum cancer Lung cancer Uterus cancer Ovary cancer Prostate cancer Bladder cancer Diabetes mellitus Dementia Schizophrenia Alzheimer’s disease Epilepsy Heart Failure Cerebrovascular diseases COPD Alcoholic liver disease Infections of skin Coxarthrosis Gonarthrosis Glomerular disorders Intracranial injury Fracture of elbow and forearm Fracture of femur Fracture of ankle or lower leg

*** *** *** *** *** *** *** *** *** *** *** ***

*** *** *** ** * *** *** * *** *** ***

*** *

*** *** *** ** * *** *** *** *** *** *** *** *** *** *** ***

Log likelihood = -135296.99. Likelihood Ratio (96) = 48248.56 (p < 0.0001). Pseudo R² = 0.1513. Abbreviations: RRR, Relative Risk Ratio; SE, Standard Error; Sgn, Significance; COPD, Chronic Obstructive Pulmonary Disorder. Key to significance values: *, p<0.05; **, p<0.01; ***, p<0.001.

made several model predictions and plotted the results. Figure 5.1 shows the probability of each type of long-term care for different combinations of gender, presence of a spouse and presence of a child in the household of the patient, plotted against age. In making these predictions, we assumed that the patients had none of the 23 core diseases and had a median length of stay in the hospital. The probability of home care can be seen to be higher than that for either type of residential care. The figure also shows the age–outcome relationship. For residential care, the probability increased with age, with the slope of the line also

96

PREDICTORS OF LONG-TERM CARE UTILIZATION

0.20

0.08

0.06

Figure 5.1: Predicted probability of discharge for different combinations of gender, presence of spouse and presence of child. Discharge destinations are home with home care, home for the elderly and a nursing home.

2 4

0.02

6 5 8

90

Age

100

0.05 0.04 0.03

8 6 5

0.00

0.00

0.00

80

7

0.02

0.04

3

2

0.01

4

Probability of nursing home care

1

7

70

4

0.06

1 Probability of elderly home care

0.10

5 7

0.05

Probability of home care

0.15

6 8 1 3

3

2

70

80

90

100

70

Age 1 Male , no spouse, no child 2 Female, no spouse, no child 3 Male , no spouse, child 4 Female, no spouse, child

80

90

100

Age 5 Male , spouse, no child 6 Female, spouse, no child 7 Male , spouse, child 8 Female, spouse, child

increasing with age. However, there was a decline in the probability of home care after an age of 90. This figure clearly shows that both men and women with a spouse (lines 5-8) had lower probabilities of requiring long-term care than those without a spouse (lines 1-4). Gender differences were minimal. The presence of a child seems to have had an effect on the use of residential care. Interestingly, those with a child had a higher probability of going home with home care. Hospital care utilization Both hospital days RRR terms were significant, but as is the case with age, it was difficult to interpret these coefficients. Figure 5.2 shows the relationship between the length of stay in the hospital and the distribution of outcomes, based on model predictions for females aged 74 years (mean age) without a spouse or a child in their household and without a model disease. The probability of long-term care

97

CHAPTER 5

1.0

Figure 5.2: Predicted cumulative probability of discharge for different lengths of hospitalization. Discharge destinations are home with no formal care, home with home care, home for the elderly and a nursing home.

0.4

0.6

Home of the elderly care Nursing home care

0.0

0.2

Probability

0.8

No formal care Home care

0

10

20

30

40

50

60

Hospital days

increased with increasing stay in the hospital for the first 62 days. This effect was relatively stronger for residential care. After a hospitalization of 62 days, the overall probability of long-term care decreased, while the probability of nursing home and home elderly care combined continued to increase until an age of 87 years. Note that about 99% of all patients had a hospital stay of 62 days or less, which means that model predictions for more than 62+ days have relatively wide confidence intervals (not depicted here). Disease predictors of discharge destination Of all 23 core diseases, all but two had a significant RRR for the probability of going back home with home care. Only five diseases had a non-significant RRR for nursing homes as the discharge destination, and ten diseases had a non-significant RRR for homes for the elderly as discharge destination. The diagnoses with the highest RRR for home care utilization were mainly cancers (lung cancer, intestinal, stomach and rectum cancer and ovary cancer), with alcoholic liver disease and coxarthrosis completing the top five. Of all the cancers examined, only lung cancer and intestinal, stomach and rectum cancer showed a significantly positive RRR for both types of residential care. High RRR for residential care were also found for dementia and alcoholic liver disease and for diseases related to physical functioning (coxarthrosis, gonarthrosis, fracture of femur and fracture of ankle or lower leg).

98

PREDICTORS OF LONG-TERM CARE UTILIZATION

0.6

1 2 3 4 5 0

0.0 0.2 0.4

Probability of home care

Figure 5.3: Five diseases associated with the largest predicted discharge probability for each discharge destination. Discharge destinations are home with home care, home for the elderly, and a nursing home.

70

80

90

1 Lung cancer 2 Intestinal cancer 3 Ovary cancer 4 Alcoholic liver dis. 5 Coxarthrosis 0 Baseline

100

0.00 0.10 0.20 0.30

Probability of elderly home care

Age

1 2 34 5 0

70

80

90

1 Fracture of low er leg 2 Coxarthrosis 3 Dementia 4 Alcoholic liver dis. 5 Fracture of femur 0 Baseline

100

0.4

1 0.2

2 34 5 0

0.0

Probability of nursing home care

Age

70

80

90

1 Cerebrovascular dis. 2 Fracture of femur 3 Fracture of low er leg 4 Dementia 5 Coxarthrosis 0 Baseline

100

Age

One crucial difference between the forms of residential care was found for cerebrovascular diseases: the RRR was the highest of all diseases for nursing home care, but it was ranked 15th for homes for the elderly. Figure 5.3 shows the five diseases with the highest probabilities for each outcome. Model predictions were based on female gender, absence of a spouse and child and median LOS. The diagnoses clearly had a larger effect than demographic variables on the probability of formal care. The top five diseases for each outcome category can be seen to correspond with the diseases with the highest RRRs. The differences within these top five diseases are more pronounced for predicting hospital discharge to a nursing home than discharge to a home for the elderly or to the patient’s home with home care, with cerebrovascular diseases clearly the most important predictor of nursing home care.

99

CHAPTER 5 Figure 5.4: Predicted cumulative probabilities for discharge destinations associated with nine distinct diseases. Discharge destinations are home with no formal care, home with home care, home for the elderly and a nursing home.

75

85

85

95

65

75

85

95

Epilepsy

Heart Failure

Cerebrovascular dis.

85

95

0.0 65

75

85

95

65

85

95

Age

COPD

Alcoholic liver dis.

Fracture of low er leg

85

95

Age

0.0 65

75

85

95

Age

No formal care Home care

0.4

Probability

0.0

0.4

Probability

0.8

Age

0.4

75

75

Age

0.8

75

0.4

Probability

0.0

0.4

Probability

0.4

0.8

Age

0.8

Age

0.0 65

0.8 0.0

75

Age

0.8

65

0.4

Probability

0.8 65

0.0

Probability

0.4

95

0.8

65

Probability

Dem entia

0.0

0.4

Probability

0.8

Diabetes m ellitus

0.0

Probability

Lung cancer

65

75

85

95

Age

Home of the elderly care Nursing home care

Figure 5.4 shows the distribution of outcomes plotted against age for nine diseases. These predictions were based on female gender, absence of a spouse and child and median LOS. All cancers (not depicted) had a similar distribution to that of lung cancer, with a high probability of discharge to the home with home care, while the risk of residential care was comparable to the baseline value (none of the core diseases). On the other end of the spectrum, fracture of the lower leg was associated with a high probability of residential care. Other mobility-related diagnoses (not depicted here) also had comparable patterns. Dementia and cerebrovascular diseases were also good predictors of residential care. Interestingly, heart failure did not lead to nearly as much formal care as cerebrovascular diseases.

100

PREDICTORS OF LONG-TERM CARE UTILIZATION

5.4 Discussion The aim of this study was to assess and compare predictors of long-term care utilization by hospital patients aged 65+. By linking data from three national registers, we have been able to show that the large majority of the study population (80.3%) did not use any form of formal long-term care directly after discharge from the hospital. Our results are in agreement with those reported in previous studies (Ahmed et al., 2003; Cai et al., 2009; Campbell et al., 2004; Greene and Ondrich, 1990; Miller and Weissert, 2000; Mustard et al., 1999; Ohwaki et al., 2005; Wolinsky et al., 1992; Hancock et al., 2002; de Pablo et al., 2004; Freedman, 1996; Glader et al., 2003; Rundek et al., 2000) in showing that higher age, living alone and length of stay in hospital are associated with an increased probability of longterm care utilization. In general, higher age was associated with an increased risk of long-term care utilization, but above the age of 90 years, the probability of the patient returning home with home care only declined and the probability of residential care increased. This result is very plausible and emphasizes the relevance of using a dataset that includes different types of long-term care. A longer stay in hospital, within the constraints of the first 62 days, was also associated with a higher risk of formal long-term care. Living with a spouse or living in the household of a child had a considerable impact on long-term care utilization. The presence of a spouse reduced the risk of all types of long-term care, probably because the spouse was able to provide informal care. Living with a child reduced the risk of admission to a home for the elderly but increased the risk of admission to a nursing home. One possible explanation for this latter effect is that of selection bias. The personal circumstances of an elderly parent(s) who is sharing a household with a child before admission to hospital may differ structurally from those of a parent living alone or with a spouse, such as in terms of frailty and capacity to live independently. Earlier changes in these latter two factors may, in fact, have provided the motivation for the parent to move in with the child in the first place. Although children in such situations may be able to provide informal care, they are likely to have other tasks and responsibilities, such as work and parenting activities. Consequently, there is a critical threshold in their capacity to provide the necessary level of care to the parent: below that level, middle-aged children are able to combine care for their parent with their normal daily activities; beyond that threshold, the burden of care becomes too high and the care of the parent has to be transferred to a facility. The dynamics of maintaining the patient at home versus institutionalization differ depending on the individuals providing the informal care. Partners tend to be more sensitive to the patient’s desire to live at home as long as possible and to 101

CHAPTER 5

feel more responsible (or obliged) to provide the care the patient needs, leading to a longer perseverance time, many times at the expense of their own health (Schulz and Beach, 1999; van Exel, 2005). We were unable to distinguish between sons and daughters in terms of providing informal care. In contrast, Freedman (1996) showed that having at least one daughter reduced the risk of admission to a home, while the presence of a son did not have any positive effect in this context. Our results are a valuable contribution to existing scientific literature in the field as they provide information on the effect of major diseases in the 65+ population on long-term care utilization following hospital discharge. Based on our data, hospital diagnoses do indeed play a major role in determining the care utilization of the patient following his/her release from hospital. In our model, different diagnoses were the most important predictors for the three types of formal longterm care, with cancers being important predictors of home care, cerebrovascular diseases being important predictors of admission to a nursing home and disabilityrelated diseases, such as fracture of the ankle or of the lower leg, and mental health problems, such as dementia and schizophrenia, being important predictors of admission to a home for the elderly. Changing patterns in hospital morbidity may, therefore, induce a shift in long-term care needs. Policy-makers and health care planners can make use of this information to anticipate on future health care needs. These findings underline a number of the strengths of this study. Due to the availability of national registers, we were able to include a large number of subjects and a variety of diagnoses and discharge destinations in the analysis. Consequently, we were able to investigate long-term care utilization following discharge from the hospital in a setting with a whole range of long-term care facilities as well as a large variety of disease variables. However, our study also has a number of limitations. First, although the data included information on a large variety of diagnoses, no information was available on the severity of the associated functional and cognitive limitations which, ultimately, can be expected to be stronger predictors of longterm care use. Secondly, this study is largely based on data for a single year (2005). As a consequence we could effectively only observe patients’ long-term care utilization up to one year. As such, our conclusions are valid for this single year, and strictly do not regard a subsequent extended period of long term care use. However, given the structure of the Dutch long-term care sector, with comprehensive residential facilities for elderly people, it is common that such longterm decisions are made for the remaining lifetime of the elderly person. Therefore, we expect that the group of elderly with short-term need for extra services is relatively small. We expect that only home care results might be affected by this, and not the residential care. Data covering a longer period would provide more insight into time trends and patterns of long-term care utilization. Regrettably these data were not available. Thirdly, in this study we do not only look 102

PREDICTORS OF LONG-TERM CARE UTILIZATION

at hospitalized patients that utilize long-term care directly after being discharge, but also patients that use long-term care after a certain period following hospitalization. The wider this timeframe is, the weaker the causal link between hospitalization and long-term care becomes. However, the predictors were relatively stable within the observed period of one year. Many of the core diseases are chronic and are therefore unlikely to have become more or less severe in the window. Finally, due to the nature of the data sources used, this study did not take into account a number of other factors that may influence discharge destination, such as supply-side restrictions (e.g. waiting lists), patient preferences and the burden and positive aspects of providing informal care (Brouwer et al., 1999; Brouwer et al., 2005; Nieboer and Koolman, 2006; van Exel et al., 2007; van Exel et al., 2008).

5.5 Conclusions The results of this study may be relevant to health professionals and policy-makers. Firstly, our data on the positive influence of the presence of a spouse in terms of home care could have a considerable effect on planning future long-term care use as they provide further evidence that informal care may considerably reduce the need for formal care. In addition, however, this observation is particularly relevant for countries such as the Netherlands in which the gap in life expectancy between men and women is narrowing. It is expected that in the upcoming decades a higher proportion of women will have a spouse, with the result that the need for residential long-term care will show less of an increase relative to conservative estimates. For policy-makers this trend may provide an important incentive to invest in alternative approaches to providing informal care and to enhance and underpin policies that provide support to spousal caregivers. Secondly, we found that several diagnoses, including cerebrovascular diseases, fractures of the ankle or lower leg, cancers and dementia, are not only important predictors of long-term care utilization but also of specific discharge destinations. Any future change in the prevalence of these diseases will therefore result in changing health care needs, not only within hospitals but also in terms of providing long-term care after discharge. By linking our findings to epidemiological scenarios for the prevalence of these diagnoses, scenarios of future demand for long-term care can be developed and used for capacity and manpower planning of home care, nursing homes and homes for the elderly. Our figures can also be used for the development of cost-containment measures to limit the financial burden of an ageing population on national budgets. They also highlight the importance of focussing on disease prevention and early

103

CHAPTER 5

intervention as both activities may have a real impact on the number of individuals requiring long-term care. In conclusion, the ageing of Western societies is a multifaceted phenomenon in which health care needs change much more than has been suggested in traditional scenarios. The need for long-term care following hospital discharge depends heavily on demographic and epidemiological trends. Policy-makers would be advised to anticipate these developments.

Appendix 5.A: Long-term care in the Netherlands The Dutch health system comprises three major types of formal care dedicated to the health care needs of the elderly: home care, homes for the elderly and nursing homes. In recent years home care has been restructured into two different categories, but here we describe the situation as it was in 2005, the year of our data collection. Of all Dutch individuals above the age of 65 years in 2005, about 20% received formal care at home and close to 7% lived in either a nursing home (65,000 people) or a home for the elderly (100,000 people). Home care largely consisted of personal care and help with daily household activities. The Netherlands has a strong primary care system, and each citizen – elderly or not – can contact his general practitioner in case of sickness or physical/mental complaints. This is also the case for residents of homes for the elderly, which in essence are residential facilities that provide living assistance. Nursing homes provide both personal care and living assistance, with geriatrists and gerontologists responsible for providing integrated medical care. As a general rule, the health status of elderly residents of nursing homes is much poorer than that of the elderly in residential homes. It is important to note that the availability of formal long-term care services is highly regulated. The service capacity within a specific region is directly related to the budget received by local authorities. During recent decades, the national government has promoted an increase in the capacity of community care services and a decrease of the number of beds in institutional care, but compared to many other countries the capacity of residential care in the Netherlands (i.e., homes for the elderly and nursing homes) is still relatively high. Access to formal long-term care services is also strongly regulated. Individuals in need of home care or admission to either a nursing home or a home for the elderly have to apply to a Municipal Committee for a Need Assessment. The applicant may be rejected or, due to a shortage of placements, put on a waiting list. If the request is granted, a (large) share of the costs incurred for formal care utilization is covered by national insurance systems (the Exceptional Medical Expenses Act).

104

Chapter 6 The Disabling Effect of Diseases: A Study on Trends in Diseases, Activity Limitations, and Their Interrelationships ‡

Time trend data from the Netherlands suggest an increase in the prevalence of chronic diseases and a stable prevalence of disability. We studied how the association between chronic diseases and activity limitations in the Netherlands has changed in the past 15 to 20 years. Five Dutch surveys among non-institutionalized older persons aged 55 to 84 years (n=54,847) provided self-reported data on chronic diseases (diabetes, heart disease, peripheral arterial disease, stroke, lung disease, joint disease, back problems, and cancer) and activity limitations (Organisation for Economic Co-operation and Development [OECD] long-term disability questionnaire or 36-item Short Form Health Survey [SF-36]) over the period 1990 to 2008. Prevalence rates of chronic diseases increased over time, whereas prevalence rates of activity limitations were stable (OECD) or slightly decreasing (SF-36). The associations between chronic diseases and activity limitations were also stable (OECD) or slightly decreasing (SF-36). We saw opposing trends between surveys. The hypothesis that diseases became less disabling over the period 1990 to 2008 was only supported by results based on activity limitation data as assessed with the SF-36. Further research on how diseases and disability are associated over time is needed.



This chapter is based on: Hoeymans N, Wong A, Van Gool CH, Deeg DJ, Nusselder WJ, De Klerk MM, Van Boxtel MP, Picavet SH. 2012. American Journal of Public Health 102(1): 163-170. This article has been reproduced with the permission of the American Public Health Association.

CHAPTER 6

6.1 Introduction As populations in many Western countries are aging, a major public health issue is how the health of the population is changing. Insight into the trends in health status of the older population may help in estimating future health care needs and in setting priorities for improving the health of the aging population. Important indicators for health at old age are the prevalence of chronic diseases and disability. Earlier studies in the Netherlands showed that the prevalence of chronic diseases increased in the past decades (Uijen and van de Lisdonk, 2008). This increase was not only attributable to aging of the population; the age group– specific prevalence rates also increased (Jagger et al., 2007). In contrast, the ageadjusted prevalence of activity limitations and disabilities show varying trends in the Netherlands, including remaining constant, increasing, or even decreasing over the past decades as can be inferred from several data sources(Picavet and Hoeymans, 2002; Hoogendijk et al., 2002; Puts et al., 2008; van Gool et al., 2011). Decreasing disability trends have also been observed in several other western countries, suggesting a postponement of limitations and disabilities, despite increases in chronic diseases in many countries (Christensen et al., 2009). Because diseases are important causes of activity limitations and disabilities (Stuck et al., 1999; Tas et al., 2007), these opposing trends warrant clarification. Several possible explanations for these contradictory trends have been suggested. First, trends in disability rates do not necessarily follow trends in chronic diseases, because increasing use of aids and devices facilitates greater independence in people with or without diseases (Murabito et al., 2008; Jeune and Brønnum-Hansen, 2008). Also changes in the environment can prevent disabilities for people with a disease (Schoeni et al., 2008). Safer and better sidewalks can, for example, make it easier for people with joint disease to move around. Another explanation is that improved medical knowledge and health services utilization lead to more detection of disease, while the actual prevalence remains the same. For some diseases, such as type 2 diabetes, hypertension, and some cancers, people get diagnosed earlier and receive better treatment than before (Murabito et al., 2008). This progress leads to longer periods of known morbidity (and, thus, higher estimates of the prevalence of diseases), but with an improved functional status. This would result in a stable or decreasing prevalence of activity limitations and disabilities (Christensen et al., 2009; Schoeni et al., 2008; Parker and Thorslund, 2007). In that case, the conclusion will be that diseases have become less disabling over time. For different diseases, different underlying developments are possible, resulting in different trends in the disabling impact of diseases (Puts et al., 2008). We sought to investigate what can be said about the time trend in the disabling impact of chronic diseases in the Netherlands. We examined the time trends in the prevalence of diseases, in activity limitations, and in the strength of the associations 106

THE DISABLING EFFECT OF DISEASES

between diseases and activity limitations. To give the best estimate of this trend we combined all available data sources.

6.2 Methods In the Netherlands, several surveys give information on diseases and functional limitations, and no single survey can be regarded as the “best.” Therefore, we selected all available Dutch population-based surveys relevant to the subject. Inclusion criteria were that the studies should (1) contain data on self-reported activity limitations and chronic diseases, (2) have at least 3 data collection moments, (3) represent both genders, (4) cover a minimum period of 10 years, and (5) concern non-institutionalized respondents aged 55 to 84 years. We harmonized original data from 5 surveys: the Amenities and Services Utilization Survey (Dutch abbreviation: AVO, see Stoop (2000)), the Netherlands Health Interview Survey (in 1997 renamed Permanent Life Situation Survey [Dutch abbreviation: POLS], see Statistics Netherlands (1996)), the Longitudinal Aging Study Amsterdam (LASA, see Deeg et al. (2002)), the Maastricht Aging Study (MAAS, see Jolles et al., 1995)), and the Doetinchem Cohort Study (DCS, see Verschuren et al. (2008)). The surveys differ in whether proxy measurements were used, frequency of measurement, whether they cover a region or the whole country, whether it is a repeated cross-sectional survey or a time series survey, and several other aspects. Table 6.1 gives information on these and other general characteristics of the 5 surveys. Many of these characteristics cannot be adjusted for in the models. We therefore analyzed all surveys separately to combine the results in a meta analysis (see statistical analyses). Respondents who turned 55 years in between cycles of data collection were added to the data set in the subsequent cycle. Also, in 2003 a new cohort aged 55 to 65 years was added to the surviving LASA cohort. These “new” respondents miss self-reports in the years before their 55th birthday or before 2003, respectively. Measurement of Diseases and Activity Limitations On the basis of the condition that at least 3 surveys had to collect information on these self-reported diseases in a comparable way (see Table 6.2), we selected 8 diseases for analysis: diabetes, heart disease, peripheral arterial disease, stroke, lung disease, joint diseases, back problems, and cancer. We constructed 1 summary measure to distinguish people with morbidity (no disease vs 1 or more diseases) and another to distinguish people with multimorbidity (no disease vs 1 disease vs 2 or more diseases).

107

CHAPTER 6 Table 6.1: General characteristics and prevalence rates of diseases and activity limitations of the selected surveys, the Netherlands. Characteristic Commissioning organization Geographical representation Type of survey Frequency of measurement Duration of 1 cycle of measurement Period Interview and/or questionnaire Use of proxies (Baseline) Age range (Baseline) Age range of selected resp. Response range No. of selected c respondents d Diseases, % Diabetes Heart disease Peripheral arterial Stroke Lung disease Joint disease Back problems Cancer Morbidity summary Exactly 1 dis. Two or more dis. b Disability, % At least moderate OECD Severe OECD At least moderate SF-36 Severe SF-36

AVO Netherlands Institute for Social Research National

POLS Statistics Netherlands

LASA VU University Amsterdam

MAAS Maastricht University

National

Multiregional

Provincial

DCS National Institute for Public Health and the Environment Municipal

Repeated crosssectional Every 4 year

Repeated crosssectional Yearly

Longitudinal

Longitudinal

Longitudinal

Every 3 year

Every 3 year

Every 5 year

1 year

1 year

2 year

3 year

5 year

1991–2007 Both

1990–2008 Both

1995–2006 Both

1993–2007 Questionnaire

1995–2007 Questionnaire

Allowed 6 and older 55–84

Allowed 0 and older 55–84

Allowed 55–84 55–84

Not allowed 24–81 55–81

Not allowed 20–59 55–59

43% to 70% of households 3411

60%–65%

62%–82%

54%–75%

62%–80%

2190

1638

934

2037

8.8 9.6 8.0 3.3 10.5 25.0 15.6 3.2

8.3 9.2 5.3 3.4 9.8 26.5 13.6 6.6

9.1 23.8 10.1 5.8 13.8 47.6 NA 13.4

7.4 16.2 6.2 NA 11.3 17.4 NA 7.7

4.8 4.2 24.1 2.5 3.6 NA 59.4 7.1

24.1 20.7

29.4 23.9

37.3 34.3

27.3 14.8

43.3 27.8

24.0

32.2

40.3

NA

NA

4.8 NA

7.8 NA

13.8 NA

NA 52.1

NA 38.0

NA

NA

NA

16.8

9.4

a a

b

Notes. AVO=Amenities and Services Utilization Survey; DCS=Doetinchem Cohort Study; LASA=Longitudinal Aging Study Amsterdam; MAAS=Maastricht Aging Study; NA=not available; OECD=Organisation for Economic Co-operation and Development long-term disability questionnaire; POLS=Permanent Life Situation Survey; SF-36=36-item Short Form Health Survey; resp=respondents; dis=diseases. a Except for the period 2001 to 2003 where the old cohort was interviewed in 2001 to 2002 and a new cohort was interviewed in b 2002 to 2003. In 2005 to 2006 both cohorts were interviewed at the same time. Waves over 1993 and 1994 did not include c d 36-item Short Form Health Survey data and were excluded from the analyses. Averaged over all years of measurement. Crude prevalences; mean over all years of measurement.

The assessment of activity limitations was based on 3 items that the 5 surveys had in common: stair climbing, walking outside, and getting dressed (van Gool et al., 2011). AVO, POLS, and LASA used variants of items from the OECD long-term disability questionnaire and MAAS and DCS used items based on the Medical Outcomes Study (MOS) 36-item Short Form Health Survey (SF-36) to assess activity limitations in these domains (McWhinnie, 1981; Ware and Sherbourne, 1992; Ware et al., 1995). In our analyses, we differentiated between activity limitations based on these 2 instruments because of differences in wording between the OECD questionnaire and the SF-36 (difficulty in performing actions 108

Narrowing of veins in the abdomen g or legs (no varicosis)

Peripheral arterial disease Stroke

Arthritis of knees, hips; rheumatoid m arthritis

Severe or long-lasting back o problems

Cancer or malignant condition

Joint disease

Back problems

Cancer d,r

Did you ever have any kind of cancer (malignant condition)?

Severe or long-lasting back p problems

Arthritis of knees, hips, or n hands; rheumatoid arthritis

POLS I am going to mention a number of diseases and conditions. Please indicate whether you currently have any of these, or have had in the b past 12 mo. c,d Do you have diabetes? Did you ever have a heart attack? Did you have another severe heart condition in the d,f past 12 mo? Narrowing of veins in the abdomen or legs (no h varicosis) Did you ever have a stroke, brain hemorrhage, or cerebral d,j infarction? Asthma, chronic bronchitis, l emphysema, or CNSLD

Do you have tumor formation or cancer, or did you have it in the past?

Do you have arthritis of knees, hips, or hands? Do you (also) have rheumatoid arthritis? NA

Do you have CNSLD (asthma, chronic bronchitis, emphysema)?

Did you ever have a stroke or a brain attack? (also brain hemorrhage)

Do you have abnormalities or diseases of arteries or veins in abdomen or legs?

LASA Now, I’d like to ask about a number of diseases. It’s about diseases and symptoms that last for at least 3 mo, or for which people need to be treated by a physician. Please indicate whether you currently have any of the diseases that I list. Do you have diabetes? Do you have heart disease, or did you have a heart attack?

Cancer or malignancy

NA

CNSLD, chronic bronchitis, or emphysema Rheumatic condition or joint disease

NA

In the past 12 mo, did you have problems (pain, discomfort) in the low back? In the past 12 mo, did you have problems (pain, discomfort) in the upper back? Do you have any kind of cancer, or did you have it in the past?

NA

Did you ever have asthma?

Did you ever have a stroke?

Do you experience pain in your legs when walking?

Do you have diabetes? Did you ever have a heart attack?

Diabetes (IDD or NIDD) Heart failure or cardiac insufficiency; heart attack

Peripheral arterial diseases or claudication

DCS NA

MAAS Have you been told by a physician within the past 3 y that you suffer from any of the following conditions?

Notes. CNSLD=chronic non-specific lung disease; IDD=insulin-dependent diabetes; NIDD= non–insulin-dependent diabetes; NA=not available. a b Wording used in 1995 differed from the wording in this column, no disease information available before 1995; Interview wording used in the period 1990 to 2000 differed from the wording used in c d e the questionnaire (in this column); Before 2001, this question read “diabetes.”; These questions precede the introductory text.; Two separate questions—in 1995 this question read “Severe f g h i heart condition or heart attack.”; Two separate questions—before 2001 this question read “Severe heart condition or heart attack.” ; Not available before 1999; Not available before 2001; In j k l 1995 this question read “(Consequences of a) stroke.”; Before 2001 this question read “(Consequences of a) stroke.”; In 1995, this question read “Asthma, chronic bronchitis, or CNSLD.”; Before m n 2001, this question read “Asthma.”; Two separate questions—in 1995 the first question read “Arthritis of knees, hips, or hands.”; Two separate questions—before 2001 these questions (3) read o “Arthritis of knees, hips, or hands; rheumatoid arthritis; other chronic rheumatoid condition lasting for 3 mo or more.”; In 1995, this question read “Persistent back problems, lasting longer than 3 p q r mo, or spinal hernia.”; Before 2001, this question read “Persistent back problems, lasting longer than 3 mo.”; In 1995, the order of cancer and malignant neoplasm was reversed. Before 2001, this question read “Malignant condition or cancer.”

q

Asthma, chronic bronchitis, k emphysema, or CNSLD

Lung disease

(Consequences of) stroke, brain i hemorrhage, or cerebral infarction

Diabetes Severe heart condition (such as heart failure or angina), e (consequences of a) heart attack

AVO The following is a list of diseases and conditions. Please indicate whether you currently have any of these diseases or conditions, or a have had in the past 12 mo.

Diabetes Heart disease

Introductory text

Table 6.2: Questionnaire and interview wording in the selected surveys regarding selected diseases, the Netherlands, 1990 to 2008.

CHAPTER 6

vs health-limiting effects in the performance of activities, respectively), and because earlier studies suggested differences in criterion validity between the 2 questionnaires (van Gool et al., 2011; Reijneveld et al., 2007; Reuben et al., 1995). We combined the 3 items to create 1 summary measure for activity limitations based on the OECD and 1 based on the SF-36. For both, we distinguished 3 levels: (1) not limited at all, (2) moderately limited in at least 1 activity, and (3) severely limited in at least 1 activity. We further dichotomized these categories into 2 variables indicating either severe or at least moderate (including severe) limitations. Statistical Analyses We used descriptive analyses to present the prevalence of the selected diseases and activity limitations in the separate surveys averaged over all observation years. Next, we used regression models to estimate time trends for (1) each disease and morbidity summary measure, (2) each activity limitation, and (3) the association between diseases or morbidity summary measures and activity limitations. In models a and b, we used year of observation as the independent variable. In model c, the disease or morbidity summary variable, year of observation, and the interactions between these 2 served as independent variables. We carried out the regression analyses on each of the 5 data sets and adjusted all parameters for age and sex. For the prospective studies, we fitted a generalized estimation equations model (Liang and Zeger, 1986); for the cross-sectional studies, a generalized linear model. For more details on the regression models, we refer to Appendix 6.A. To reduce loss of information because of missing values in the data on diseases and activity limitations, we imputed missing values by using the multiple imputations by chained equations algorithm (van Buuren and GroothuisOudshoorn, 2011). Most variables missing varied from 1% to 20%, with 2 exceptions: 56% missing in peripheral arterial disease in POLS, because this disease was not measured before 2001, and 43% missing in SF-36 in MAAS. We used 10 imputations per survey including all data on diseases, activity limitations, and demographic characteristics (age, gender, marital status, and household size). Per survey, we pooled the regression parameters of the 10 imputations to gain surveyspecific estimates by using Rubin’s rules (Rubin, 1987). To get overall estimates we performed meta-analyses on all survey-specific estimates. Because there might be unexplained heterogeneity between surveys that could be attributed to survey-specific aspects, such as data sampling and interviewing methods, we used a random effects model based on the DerSimonian-Laird method (DerSimonian and Laird, 1986). The random effects model assumes each study has its own mean and (within-study) variance for its effect, and the mean of each study comes from a hyper-distribution of means. That distribution is assumed to be Gaussian, with an overall mean, and an overall 110

THE DISABLING EFFECT OF DISEASES

variance (between-study variance, which quantifies the amount of heterogeneity). Estimating the overall mean, while taking into account the heterogeneity, allows for inference over all studies (Hu et al., 1998; Marquardt, 1970). Tests for heterogeneity justified the choice for the random effects model. As a result, this will generally yield wider confidence intervals (than fixed effect models), resulting in a less likely rejection of the null hypothesis (i.e., results are less likely to be significant and, thus, are more conservative).

6.3 Results General characteristics of the 5 surveys are summarized in Table 6.1. Two of the 5 surveys had collected nationally representative data with a repeated cross-sectional design. The other 3 were longitudinal studies performed in 1 specific region or in several regions. Response rates varied from 43% to 82%. Mean age varied between 60 and 70 years. Three surveys showed that almost 50% (42% to 54%) of the respondents had at least 1 disease (Table 6.1). In the other 2 surveys, this rate was a little higher than 70%. Prevalence rates of stroke, diabetes, and lung disease were fairly consistent over the surveys, with the DCS study reporting the lowest prevalences. These observed differences might be explained by differences in mean age, because the DCS study had the youngest study population. The differences in prevalence rates of the other diseases were greater. The way the presence of these diseases was ascertained might explain these differences (Table 6.2). For example, the prevalence of peripheral arterial disease was much higher in the DCS study than in the other surveys. This is most likely attributable to the nonspecific way this disease was presented to respondents. The prevalence of at least moderate activity limitations varied between 24% and 52%, and that of severe limitations between 5% and 17%. Trends in Chronic Diseases and Disabilities In the period 1990 to 2008, the odds of reporting diabetes increased by 5% per year (odds ratio [OR]=1.05; 95% confidence interval [CI]=[1.04, 1.07]; Table 6.3). In addition, the odds of reporting stroke increased 5% per year (OR=1.05; 95% CI=[1.01, 1.08) and the odds of reporting cancer increased 6% per year (OR=1.06; 95% CI=[1.01, 1.12]). We found stable prevalence rates for heart disease, peripheral arterial disease, lung disease, joint disease, and back problems. Overall, the odds of reporting 1 or more of the selected diseases increased by 3% per year (OR=1.03; 95% CI=[1.00, 1.05]), and the odds of reporting 2 or more diseases increased by 3% per year (OR=1.03; 95% CI=[1.01, 1.04]). For most diseases the

111

CHAPTER 6 Table 6.3: Time trends over the period 1990 to 2007 in the prevalence of chronic diseases and activity limitations across different surveys and overall, persons aged 55 to 84 years, the Netherlands. AVO, OR (95% CI)

POLS, OR (95% CI)

LASA, OR(95% CI)

MAAS, OR (95% CI)

DCS, OR (95% CI)

Overall, OR (95% CI)

1.07* (1.04, 1.09) 1.04* (1.02, 1.06) 0.94* (0.90, 0.99) 1.05* (1.02, 1.08) 1.06* (1.04, 1.08) 0.95* (0.94, 0.97) 1.01 (1.00, 1.03) 1.07* (1.03, 1.10)

1.05* (1.04, 1.05) 1.03* (1.03, 1.04) 1.01 (0.98, 1.04) 1.08* (1.07, 1.09) 1.01* (1.00, 1.01) 1.01* (1.01, 1.02) 1.01* (1.00, 1.02) 1.14* (1.12, 1.15)

1.07* (1.05, 1.09) 1.02* (1.01, 1.04) 0.99 (0.97, 1.01) 1.03* (1.01, 1.06) 1.02* (1.00, 1.04) 1.0*3 (1.01, 1.04) NA

1.01 (0.97, 1.04) 0.99 (0.96, 1.01) 1.00 (0.97, 1.04) NA 0.99 (0.96, 1.01) 1.05* (1.03, 1.08) NA

1.06* (1.01, 1.10) 0.96 (0.91, 1.01) 0.95* (0.93, 0.97) 1.00 (0.93, 1.07) 0.99 (0.94, 1.04) NA

1.04* (1.02, 1.06)

1.02 (0.99, 1.05)

1.09* (1.07, 1.11) 1.05* (1.01, 1.09)

1.05* (1.04, 1.07) 1.02 (1.00, 1.04) 0.98 (0.95, 1.00) 1.05* (1.01, 1.08) 1.01 (0.99, 1.04) 1.01 (0.98, 1.04) 1.04 (1.00, 1.08) 1.06* (1.01, 1.12)

0.98* (0.97, 1.00)

1.03* (1.02, 1.03)

1.04* (1.02, 1.05)

1.00 (0.98, 1.02)

1.09* (1.07, 1.11)

1.03* (1.00, 1.05)

0.96* (0.94, 0.97) 1.02* (1.00, 1.04)

1.00 (0.99, 1.02) 1.04* (1.04, 1.05)

1.00 (0.98, 1.01) 1.04* (1.03, 1.06)

0.98 (0.96, 1.03) 1.03* (1.00, 1.05)

1.08* (1.06, 1.11) 0.99 (0.97, 1.01)

1.00 (0.97, 1.03) 1.03* (1.01, 1.04)

0.99* (0.99, 1.00) 0.99* (0.98, 0.99) NA

1.05* 1.04, 1.06) 1.02 (1.00, 1.04) NA

NA

NA

NA

NA

ALM SF-36

1.02* (1.01, 1.03) 1.03* (1.01, 1.05) NA

Severe SF-36

NA

NA

NA

0.97 (0.95, 1.00) 0.98* (0.95, 0.99)

0.95* (0.94, 0.97) 0.97* (0.94, 0.99)

1.02 (0.99, 1.05) 1.01 (0.98, 1.04) 0.96* (0.94, 0.98) 0.97* (0.95, 0.99)

Disease-specific Diabetes Heart disease Peripheral arterial disease Stroke Lung disease Joint disease Back problems Cancer Morbidity smry I 1 or more dis. Morbidity smry II Exactly 1 dis. Multimorbidity: 2 or more dis. Activity limitations ALM OECD Severe OECD

Notes: Asterisks indicate statistical significance of P<0.05. AVO=Amenities and Services Utilization Survey; CI=confidence interval; DCS=Doetinchem Cohort Study; LASA=Longitudinal Aging Study Amsterdam; MAAS=Maastricht Aging Study; NA=not available; OECD=Organisation for Economic Co-operation and Development long-term disability questionnaire; OR=odds ratio; POLS=Permanent Life Situation Survey; SF-36=36-item Short Form Health Survey; smry=summary; dis=disease; ALM=At least moderate.

results were fairly consistent between studies. The only exception was heart disease: 3 studies found an increasing trend whereas the other 2 found this trend to be declining (although not significantly). In the period 1990 to 2008, the odds of reporting at least moderate SF-36 limitations decreased by 4% per year (OR=0.96; 95% CI=[0.94, 0.98]) and the odds of reporting severe SF-36 limitations decreased by 3% per year (OR=0.97; 95% CI=[0.95, 0.99]; Table 6.3). In the same time period, overall trends in activity limitations as assessed with the OECD instrument were stable. Results were, however, not consistent among surveys. POLS, MAAS, and DCS showed decreasing trends in activity limitations, whereas AVO and LASA showed increasing trends. 112

1.03 (1.00, 1.06) 0.99 (0.96, 1.02)

1.02* (1.00, 1.04)

0.94 (0.88, 1.00) 1.11* (1.03, 1.19) 1.03 (0.95, 1.11) 1.01 (0.91, 1.12) 1.09* (1.03, 1.16) 1.03 (0.99, 1.08) 1.11* (1.06, 1.16) 0.93 (0.84, 1.03)

1.07* (1.01, 1.15) 0.99 (0.95, 1.03)

1.03 (1.00, 1.06)

0.98 (0.89, 1.09) 1.08 (0.97, 1.21) 1.09 (0.96, 1.24) 0.95 (0.84, 1.07) 1.21* (1.10, 1.33) 1.01 (0.94, 1.08) 1.07 (0.99, 1.15) 0.98 (0.83, 1.15)

0.97* (0.97, 0.98) 0.96* (0.95, 0.97)

0.98* (0.97, 0.98)

0.97* (0.96, 0.99) 0.96* (0.94, 0.98) 0.99 (0.96, 1.02) 0.92* (0.90, 0.95) 0.99 (0.97, 1.01) 0.97* (0.96, 0.98) 0.98* (0.96, 0.99) 0.96* (0.94, 0.99)

0.96* (0.95, 0.98) 0.96* (0.95, 0.97)

0.97* (0.97, 0.98)

0.98 (0.95, 1.01) 0.95* (0.92, 0.97) 0.99 (0.96, 1.02) 0.93* (0.90, 0.97) 0.98 (0.96, 1.01) 0.96* (0.94, 0.97) 0.99 (0.97, 1.02) 0.96* (0.93, 1.00)

1.04* (1.01, 1.06) 1.05* (1.03, 1.07)

1.05* (1.03, 1.06)

1.03 (0.98, 1.08)

1.06* (1.01, 1.11) 1.03 (0.99, 1.06) 1.08* (1.02, 1.13) 0.98 (0.93, 1.05) 1.03 (0.99, 1.08) 1.05* (1.03, 1.08) NA

LASA At Least Moderate

1.01 (0.98, 1.05) 1.01 (0.98, 1.04)

1.02 (0.99, 1.04)

0.99 (0.94, 1.06)

1.01 (0.95, 1.09) 0.99 (0.94, 1.04) 1.01 (0.96, 1.08) 0.96 (0.90, 1.03) 1.02 (0.97, 1.08) 1.02 (0.98, 1.06) NA

Severe

1.01 (0.96, 1.06) 1.00 (0.94, 1.06)

1.02 (0.97, 1.07)

0.99 (0.93, 1.05) 1.02 (0.95, 1.10) 1.03 (0.97, 1.09) 0.96 (0.91, 1.02) 1.03 (0.98, 1.09) 1.02 (0.96, 1.08) 1.04 (0.92, 1.17) 0.98 (0.93, 1.03)

Overall At Least Moderate

1.01 (0.95, 1.07) 0.99 (0.95, 1.02)

1.00 (0.97, 1.04)

0.98 (0.96, 1.01) 0.99 (0.93, 1.04) 1.00 (0.97, 1.04) 0.94* (0.91, 0.97) 1.05 (0.96, 1.15) 0.99 (0.94, 1.04) 1.02 (0.95, 1.09) 0.97 (0.94, 1.00)

Severe

0.95* (0.91, 1.00) 0.99 (0.95, 1.04)

0.97 (0.94, 1.00)

0.97 (0.92, 1.02) 0.99 (0.94, 1.04)

0.99 (0.95, 1.02)

0.97 (0.86, 1.08)

0.95 (0.88, 1.03) 1.00 (0.93, 1.06) NA

0.93 (0.87, 1.00) 0.98 (0.93, 1.03) NA 0.97 (0.87, 1.07)

0.98 (0.90, 1.08) 0.97 (0.89, 1.04) 0.91 (0.82, 1.02) NA

0.97 (0.90, 1.04) 0.98 (0.93, 1.04) 0.92 (0.84, 1.01) NA

0.94* (0.92, 0.97) 0.90* (0.87, 0.93)

0.92* (0.90, 0.94)

0.95* (0.92, 0.97) 1.02 (0.94, 1.10)

0.99 (0.90, 1.09) 0.96 (0.87, 1.06) 0.96 (0.92, 1.01) 0.97 (0.85, 1.11) 1.03 (0.94, 1.13) NA

0.94* (0.90, 0.99) 0.95* (0.91, 0.99)

0.94* (0.91, 0.96)

0.97 (0.92, 1.03) 1.10 (0.96, 1.25)

1.07 (0.93, 1.23) 1.07 (0.94, 1.21) 1.04 (0.97, 1.11) 0.98 (0.82, 1.16) 1.04 (0.91, 1.19) NA

SF-36 Activity Limitations, OR (95% CI) MAAS DCS At Least At Least Severe Severe Moderate Moderate

0.94* (0.92, 0.97) 0.94 (0.85, 1.04)

0.94* (0.89, 1.00)

1.00 (0.94, 1.06)

NA

0.98 (0.89, 1.08) NA

0.98 (0.92, 1.03) 0.98 (0.93, 1.03) 0.95* (0.91, 1.00) NA

Overall At Least Moderate

0.95* (0.92, 0.99) 0.97 (0.93, 1.01)

0.96 (0.91, 1.01)

1.03 (0.90, 1.16)

NA

0.98 (0.90, 1.07) NA

1.01 (0.93, 1.09) 1.00 (0.91, 1.10) 0.98 (0.86, 1.11) NA

Severe

Notes. Asterisks indicate statistical significance of P<.05. AVO=Amenities and Services Utilization Survey; CI=confidence interval; DCS=Doetinchem Cohort Study; LASA=Longitudinal Aging Study Amsterdam; MAAS=Maastricht Aging Study; NA=not available; OECD=Organisation for Economic Co-operation and Development long-term disability questionnaire; OR=odds ratio; POLS=Permanent Life Situation Survey; SF-36=36-item Short Form Health Survey.

Multimorbidity (2 or more diseases)

Morbidity summary II Exactly 1 disease

1 or more diseases

Morbidity summary I

Cancer

Back problems

Joint disease

Lung disease

Stroke

Peripheral arterial disease

Heart disease

Disease-specific Diabetes

OECD Activity Limitations, OR (95% CI) AVO POLS At Least At Least Severe Severe Moderate Moderate

Table 6.4: Time trends over the period 1990 to 2007 in the relation between selected diseases and activity limitations across different surveys and overall among persons aged 55 to 84 years, the Netherlands.

CHAPTER 6

Trends in Associations Between Diseases and Activity Limitations There was a strong association between diseases and activity limitations and the different surveys were fairly comparable with consistently higher associations for joint disease, stroke, and heart disease (see Table 6.A1 in Appendix 6.A). For most diseases, we observed no statistically significant changes in the strength of the associations with activity limitations in the period 1990 to 2008 (Table 6.4). However, the association between stroke and severe OECD limitations decreased over time by 6% per year (OR=0.94; 95% CI=[0.91, 0.97]). Also, the association between peripheral arterial disease and at least moderate SF-36 limitations decreased by 5% per year (OR=0.95; 95% CI=[0.91, 1.00]). For all diseases taken together, the association with activity limitations as assessed with the OECD instrument was stable over time and the association with at least moderate SF-36 activity limitations decreased, by 6% per year (OR=0.94; 95% CI=[0.89, 1.00]; Table 6.4). Also, the association of having exactly 1 disease with at least moderate or severe SF-36 limitations decreased by 6% and 5%, respectively (OR=0.94; 95% CI=[0.92, 0.97]; and OR=0.95; 95% CI=[0.92, 0.99]; respectively). There were, however, remarkable differences between surveys: POLS—a large, repeated crosssectional survey that assessed limitations with the OECD—showed a significantly weakening association between diseases and limitations. This weakening association was seen for almost all diseases and both for at least moderate and severe limitations. Both DCS and MAAS also found the associations between diseases and limitations becoming less strong over time, although not as convincingly as the POLS study. On the other hand, LASA and AVO showed the disabling effect of some diseases becoming stronger over time.

6.4 Discussion The main question of our study was whether diseases became less disabling over the period 1990 to 2008. Based on meta-analyses with 5 large-scale Dutch surveys we found (1) increasing prevalence rates of 3 out of 8 chronic diseases and of morbidity measures, (2) stable (OECD items) or decreasing (SF-36 items) activity limitation prevalence rates, and (3) stable (OECD) or decreasing (SF-36) associations between chronic diseases and activity limitations. The hypothesis that diseases became less disabling over the period 1990 to 2008 was therefore only supported by results based on activity limitation as assessed with the SF-36. Explanation of the Trends in Diseases and Activity Limitations Our finding of an increasing prevalence of chronic diseases was based on selfreports. Better knowledge of diseases or a changing willingness to report diseases might have caused a trend in self-reported disease while the actual prevalence did 114

THE DISABLING EFFECT OF DISEASES

not change. However, registrations in general practice has also shown an increase in the prevalence of chronic diseases (Puts et al., 2008). Furthermore, studies in most other countries have reported evidence for an increasing prevalence of chronic diseases (Jagger et al., 2007; Crimmins, 2004; Parker et al., 2005; Martin et al., 2009; Martin et al., 2010; Crimmins et al., 2011). The major explanations for this trend are changes in risk factor distributions, earlier diagnosis, and better survival among those with a chronic disease (Parker and Thorslund, 2007). For activity limitations, we found a stable prevalence without a clear, or at most a mild, trend. This was discussed in our earlier paper on trends in activity limitations (van Gool et al., 2011), where we studied trends in the prevalence of separate limitations in 3 activities: walking, climbing stairs, and dressing and undressing. These earlier results showed stable prevalence rates of these separate limitations as assessed with the SF-36, and a slight increase in the prevalence of at least moderate activity limitations in stair climbing and in getting dressed with OECD items. In the current paper, where activity limitations were defined on the basis of the 3 items combined, we found a stable prevalence for OECD limitations, and a slight decrease in the prevalence for SF-36 limitations. Are Diseases Becoming Less Disabling? The main result of our study concerns the change in the association between diseases and activity limitations over time. Our hypothesis that the strength of this association would grow weaker over time could only partly be confirmed. We found the disabling effect of having a disease to be relatively constant based on activity limitation data as assessed with the OECD instrument, but we found weakening associations between diseases and activity limitation as assessed with the SF-36. Results were, however, not consistent over all surveys. Data from POLS, DCS, and MAAS suggested that diseases became less disabling over the period 1990 to 2008 and data from AVO and LASA suggested the contrary. Put together, the overall result shows a fairly stable association. Other studies on trends in the disabling impact of chronic diseases are few. A previous Dutch study showed different results for different diseases, in the sense that the disabling impact of the more fatal diseases decreased, whereas the impact of nonfatal diseases increased (Puts et al., 2008). We did not reproduce this finding in our current meta-analysis. Other recent studies have shown that the change in the relation between disease and functional status differs across age groups (Martin et al., 2009, 2010; Crimmins et al., 2011), indicating functioning in younger age groups to be affected more by the increase in chronic diseases than in older age groups. Earlier, Freedman and Martin (2000) and Freedman et al. (2007) showed that for older age groups many conditions became less debilitating and that the prevalence of disability among those with a particular condition declined. Christensen et al. (2009) also concluded in their overview on challenges for aging 115

CHAPTER 6

populations that the link between diseases and activity limitations or disability is loosening. The definition of disability might be a possible explanation for the fact that our findings regarding the trend in the associations with diseases were different from the findings of these others (Field and Jette, 2007). Disability is an umbrella term for different dimensions of functioning. The International Classification of Functioning, Disability, and Health model of the World Health Organization distinguishes body functions and structures, activities, and participation (WHO, 2011), and the disablement process described by Verbrugge and Jette (1994) comprises pathology, impairments, functional limitations, and disability. Using aids and assistive devices does not alter activity limitations or impairments, but does change the way these limitations affect people in performing their activities and in participating. For example, people with osteoarthritis have difficulties bending their knees (problem in body function and structure), and walking (activity limitation), but because they use a walking stick or other device, they still move around and do their shopping (participation). To what extent an increasing prevalence in diseases is reflected in an increasing prevalence in limitations is dependent on what dimension of disability is assessed. It might well be that the improvements in health care and environmental factors affect participation, but not activity limitations. The instruments in our study assessed disability mainly in the sense of activity limitations. This might explain why we did not find a clear change in the disabling impact of diseases. Future studies in which a distinction between these dimensions in disability can be made are needed to shed further light on this issue. Another explanation for our finding that diseases hardly became less disabling might be that the more objective functioning of the older population actually did improve, but we were not able to measure this change by our self-reports. Disability is a socially defined construct, which not only depends on a person’s activity limitations and the demands of the environment, but also on that person’s expectations of daily life (Schoeni et al., 2008). If these expectations changed in the past decades, perhaps because older persons are less inclined to accept deterioration in their functioning (Martin et al., 2009), this might have concealed a true decrease in activity limitations. Measuring functional status more objectively with performance tests should contribute to the understanding of the changes in the relationship among diseases, objective functioning, and self-rated disability. Methodological Issues The major strength of our study is that we combined a number of large-scale surveys in the Netherlands, giving us the opportunity of studying a long period based on repeated cross-sectional as well as on longitudinal surveys using the same method. Moreover, questionnaire items within surveys hardly changed over time giving the possibility of trend analysis. Because of the heterogeneity among 116

THE DISABLING EFFECT OF DISEASES

surveys, using meta-analyses to combine the results from the different surveys to calculate best estimates of the time trends in activity limitations seemed most appropriate. The Netherlands is a small country with a relatively homogeneous population, so it is unlikely that differences in coverage of regions had affected the results. Our study lacks information on the institutional population. The proportion of older people aged 55 to 85 years living in institutions decreased from 2.3% in 1995 to 1.2% in 2008 (Statistics Netherlands, 2010). If this is attributable to higher thresholds for institutionalizing people, this would imply that more people with diseases and activity limitations can be found in the community and our figures would mask a decline in disability prevalence. However, these effects are expected to be relatively small. To what extent the exclusion of the institutionalized affected our results concerning the association between diseases and activity limitations cannot be said. There were large differences in the proportion of missing values between the different surveys. Although any potential bias that may have been introduced by these missing values has been minimized by using multiple imputation, the effect of the imputation is that the concerning estimates are less precise. This might explain why the decreasing time trends in at least moderate activity limitations and in the association between having 1 or more diseases and limitations were not statistically significant in MAAS. The overall results in the trend for activity limitations differed for the 2 instruments: OECD limitations were stable whereas SF-36 limitations were declining. Systematic differences between the OECD long-term disability indicator and the SF-36 might explain these differences in the findings (Van Gool et al., 2011; Reijneveld et al., 2007; Reuben et al., 1995). Whereas OECD items ask about difficulty in the ability to perform actions, the SF-36 asks about limitations in executing activities because of impaired health. However, this cannot be the whole explanation because 1 of the surveys that used OECD (POLS) also found the prevalence of activity limitations to be declining. Impact of the Findings Our data suggest an overall increase in the prevalence of diseases in the past 15 to 20 years and, in the same time period, no trend or slightly decreasing trend in activity limitations. One of the explanations is that diseases have become less disabling, but, according to our results, this is only a small part of the explanation. It is also possible that with the currently available data we were not able to fully pick up the changing relationship between diseases and their disabling consequences. Further research using performance tests to measure disability or measuring disability at the level of participation might shed more light on these trends. 117

CHAPTER 6

Appendix 6.A: Supplementary material to methods and results Statistical analyses details of the regression models used Survey- and imputation-specific regression models were used to estimate: A. the time trend in diseases; B. the time trend in activity limitations, C. the relationship between diseases or morbidity summary measures and activity limitations. D. the time trend in the relationship between diseases or morbidity summary measures and activity limitations A. After imputation a series of regression models were estimated. Let the subscript i denote individual i . For each imputation k within a survey s , the prevalence pi ,t ( z j ) of disease z j ( j  1,2,...,8 ) can be regressed on a time t (in calendar years), while adjusting for confounders age ai ,t and gender g i ,t . Suppressing the s , k , i and t indices, we have the following model to estimate imputation-specific trends in prevalence rates of diseases and activity limitations (model type I): logit  p( z j )    0, j  1, j t   1, j a   2, j g

(6.A1)

Year t was defined as the year in which the cycle of data collection of the survey took place. If this data collection of a specific cycle took more than one year, the midyear was used. Age a and year t were considered as linear continuous variables, as higher order polynomials negligibly improved fit (for all our regression models). Alternatively, we fitted a second class of models, where morbidity summary measures replaced the disease-specific dependent variables. Two measures were considered, an indicator variable L (with none of the selected diseases, and one or more of the selected diseases as its categories) and a categorical variable M with three categories (none of the selected diseases, exactly one of the selected diseases, and two or more of the selected diseases). To our knowledge, the polytomous outcome M cannot be fitted while taking into account the correlation between observations (i.e. in a GEE context), so we opted to use two separate binary logistic regressions: logit  p (m1 )    0,m1   1,m1t   1,m1 a   2,m1 g logit  p (m2 )    0,m 2   1,m 2 t   1,m 2 a   2,m 2 g

(6.A2) (6.A3)

where M 1 and M 2 denote the event of having exactly one disease, and two or more diseases, respectively. 118

THE DISABLING EFFECT OF DISEASES

B. The model for prevalence rates in activity limitations had a similar form. This model was fitted as a Generalized Estimating Equations (GEE) model for panel surveys, using a logit link and a binomial distribution, to adjust for the correlation in observations within individuals (Liang and Zeger, 1986). If this correlation is not taken into account, the standard errors might be overestimated for time-varying covariates such as the time trend t , which might lead to potentially wrong inferences (Hu et al., 1998). An exchangeable working correlation matrix was used here. For the cross-sectional surveys, since the observations are uncorrelated, a Generalized Linear Model was used. Furthermore, age- and gender-based inverse probability weights were constructed by dividing five-year age and gender strata proportions in the sample by proportions of the same strata in the Dutch population, and were subsequently used as weights in all the regression models, to account for survey-specific sampling designs. C and D. Let D be the binary outcome for a given activity limitation, then we use the following model to estimate the associations between each disease or morbidity summary measure and disability (see table A1 below), as well as the trend in these associations (model type I): logit  p (d )    0, j   1, j t    j * z j    j * z j * t   1, j a   2, j g (6.A4) j

j

Here,  j signifies the effect of disease j on the activity limitation in the reference year. By simultaneously including all diseases, these effects have been adjusted for each other. The  j represent the interaction effect between disease j and time t , and can be interpreted as the extent to which each  j changed over time, compared to the reference trend  1 . The actual trend in the effect of disease j can then be calculated by taking  1   j  as the point estimate. The corresponding standard error can be derived using the well-known variance property var( 1   j )  var( 1 )  var( j )  2 cov( 1 ,  j ) .

(6.A5)

One problem that may arise from using so many model terms is multicollinearity. The Variance Inflation Factor (VIF) was used to determine the extent to which multicollinearity is an issue. Marquardt indicates that a VIF greater than 10 might indicate serious multicollinearity (Marquardt, 1970). We initially found VIF values over 30. However, after centering the time variable t (i.e., subtracting the mean of t from t ), VIF values were reduced to just under 10, suggesting that the

119

CHAPTER 6

multicollinearity might be inconsequential here. After centering, the middle year of each study (between 1997-2001, depending on the study) becomes the reference year. We also fitted alternative versions using the morbidity variables L and M (model type II): logit  p (d | l )    0, j  1, j t   * l   * l * t   1, j a   2, j g logit  p (d | m)    0, j   1, j t   * m   * m * t   1, j a   2, j g

120

(6.A6) (6.A7)

6.29* (6.11, 6.77) 21.59* (17.44, 26.71)

11.37* (9.38, 13.78)

1.44* (1.12, 1.87) 2.04* (1.55, 2.68) 3.23* (2.30, 4.54) 4.78* (3.12, 7.31) 2.37* (1.85, 3.04) 4.87* (4.16, 5.70) 2.90* (2.37, 3.54) 2.66* (1.70, 4.16)

3.40* (3.13, 4.37) 11.59* (7.73, 17.39)

7.10* (4.78, 10.55)

1.41 (0.95, 2.10) 1.58* (1.04, 2.38) 1.34 (0.74, 2.41) 6.12* (3.89, 9.66) 1.27 (0.85, 1.91) 3.03* (2.29, 4.01) 2.09* (1.56, 2.81) 1.69 (0.88, 3.24)

3.50* (3.27, 3.74) 11.05* (10.24, 11.92)

5.54* (5.21, 5.91)

1.79* (1.63, 1.97) 2.43* (2.20, 2.69) 3.83* (3.29, 4.46) 3.68* (3.09, 4.38) 2.63* (2.42, 2.87) 3.35* (3.16, 3.56) 2.98* (2.77, 3.21) 1.46* (1.28, 1.66)

2.86* (2.52, 3.24) 8.01* (6.93, 9.27)

4.84* (4.27, 5.50)

1.86* (1.59, 2.18 1.81* (1.57, 2.09 3.19* (2.59, 3.92 3.31* (2.71, 4.05 2.19* (1.93, 2.47 1.90* (1.74, 2.06 2.09* (1.88, 2.32 1.08 (0.89, 1.32

2.40* (2.04, 2.81) 5.35* (4.47, 6.40)

3.18* (2.74, 3.71)

1.24* (1.00, 1.53)

1.67* (1.30, 2.13) 1.70* (1.46, 1.98) 2.22* (1.72, 2.85) 2.18* (1.67, 2.84) 2.22* (1.82, 2.71) 2.44* (2.14, 2.79) NA

LASA At Least Moderate

Notes: Asterisks indicate statistical significance of P<0.05. NA=not available.

Multimorbidity (2 or more diseases)

Morbidity summary II Exactly 1 disease

Morbidity summary I 1 or more diseases

Cancer

Back problems

Joint disease

Lung disease

Stroke

Peripheral arterial disease

Heart disease

Disease-specific Diabetes

OECD Activity Limitations, OR (95% CI) AVO POLS At Least At Least Severe Severe Moderate Moderate

2.30* (2.07, 2.69) 5.00* (3.78, 6.61)

3.23* (2.49, 4.19)

1.10 (0.84, 1.43)

1.57* (1.17, 2.09) 1.90* (1.57, 2.31) 1.62* (1.24, 2.10) 2.61* (1.95, 3.47) 2.17* (1.70, 2.77) 1.67* (1.41, 1.98) NA

Severe

3.73* (3.47, 4.06) 10.82* (5.83, 20.07)

5.83* (3.38, 10.07)

1.71* (1.53, 1.90) 2.05* (1.58, 2.64) 3.04* (2.12, 4.34) 3.31* (2.19, 5.00) 2.50* (2.25, 2.78) 3.41* (2.51, 4.62) 2.97* (2.77, 3.19) 1.56* (1.17, 2.08)

Overall At Least Moderate

2.76* (2.37, 3.22) 7.59* (5.12, 11.26)

4.69* (3.29, 6.68)

1.72* (1.49, 1.99) 1.82* (1.63, 2.03) 1.97* (1.13, 3.45) 3.59* (2.45, 5.26) 1.94* (1.52, 2.49) 2.06* (1.61, 2.63) 2.09* (1.89, 2.31) 1.12 (0.96, 1.30)

Severe

1.82* (1.46, 2.27) 3.58* (2.71, 4.74)

2.29* (1.86, 2.81)

1.21 (0.86, 1.70)

1.92* (1.41, 2.61) 1.91* (1.44, 2.53) NA

1.37 (0.97, 1.93) 1.81* (1.36, 2.41) 2.15* (1.31, 3.52) NA

1.81* (1.36, 2.42) 3.96* (2.88, 5.46)

2.54* (1.95, 3.32)

1.10 (0.71, 1.72)

1.99* (1.32, 2.99) 1.84* (1.25, 2.70) NA

2.25* (1.54, 3.27) 1.42* (1.07, 1.89) 2.10* (1.20, 3.67) NA

1.66* (1.40, 1.96) 4.32* (3.59, 5.20)

2.39* (2.05, 2.78)

1.31* (1.14, 1.50) 1.37* (1.01, 1.86)

1.50* (1.02, 2.21) 2.28* (1.57, 3.30) 4.38* (3.67, 5.22) 1.36 (0.84, 2.20) 1.43 (0.97, 2.10) NA

2.64* (1.82, 3.84) 9.29* (6.47, 13.34)

4.78* (3.47, 6.58)

1.81* (1.39, 2.36) 1.04 (0.67, 1.65)

1.44 (0.89, 2.32) 1.89* (1.18, 3.02) 5.50* (4.26, 7.10) 1.74 (0.95, 3.19) 1.87* (1.13, 3.11) NA

SF-36 Activity Limitations, OR (95% CI) MAAS DCS At Least At Least Severe Severe Moderate Moderate

1.71* (1.50, 1.96) 4.06* (3.41, 4.83)

2.35* (2.08, 2.66)

1.30* (1.03, 1.63)

NA

1.69* (1.27, 2.26) NA

1.43* (1.10, 1.85) 1.97* (1.57, 2.47) 3.19* (1.59, 6.38) NA

Overall At Least Moderate

2.15* (1.49, 3.10) 6.04* (2.62, 13.93)

3.46* (1.87, 6.42)

1.08 (0.78, 1.48)

NA

1.94* (1.41, 2.67) NA

1.84* (1.19, 2.85) 1.54* (1.20, 1.98) 3.51* (1.37, 9.01) NA

Severe

Table 6.A1: Associations between selected chronic diseases and activity limitations across different surveys and overall, based on the median year of measurement in the surveys, persons aged 55 to 84 years, the Netherlands.

CHAPTER 6

122

Chapter 7 Longitudinal Administrative Data Can Be Used to Examine Multimorbidity, Provided False Discoveries are Controlled for ‡

This article presents methods for using administrative data to study multimorbidity in hospitalized individuals and indicates how the findings can be used to gain a deeper understanding of hospital multimorbidity. A Dutch nationwide hospital register (n = 4,521,856) was used to calculate age- and sexstandardized observed/expected ratios of disease-pairing prevalences with corresponding confidence intervals. The strongest association was found for the combination between alcoholic liver and mental disorders due to alcohol abuse (observed/expected = 39.2). Septicemia was found to cluster most frequently with other diseases. The consistency of the ratios over time depended on the number of observed cases. Furthermore, the ratios also depend on the length of the time frame considered. Using observed/expected ratios calculated from the administrative data set, we were able to (1) better quantify known morbidity pairings while also revealing hitherto unnoticed associations, (2) find out which pairings cluster most strongly, and (3) gain insight into which diseases cluster frequently with other diseases. Caveats with this method are finding spurious associations on the basis of too few observed cases and the dependency of the ratio magnitude on the length of the time frame observed.

This chapter is based on: Wong A, Boshuizen HC, Schellevis FG, Kommer GJ, Polder JJ. 2011. Journal of Clinical Epidemiology 64(10), 1109-1117. This article has been reproduced with permission of Elsevier. ‡

CHAPTER 7

7.1 Introduction The increasing occurrence of multimorbidity –the simultaneous presence of more than one disease in an individual– poses several difficulties for society. Not only can multimorbidity be an obstacle for patients in leading a normal and productive life, it also alters the quantity and type of social, medical, and health care services that are needed to support them. Studies (Gijsen et al., 2001) have shown that multimorbidity is strongly associated with a higher mortality risk (Charlson et al., 1987; Deyo et al., 1992; Romano et al., 1993; Southern et al., 2004; Sundararajan et al., 2004) and in many cases, multimorbidity is associated with more health care utilization (Shwartz et al., 1996; Roe et al., 1998; Librero et al., 1999; Westert et al., 2001; Struijs et al., 2006). Specific disease combinations have been found to affect functional status and quality of life (Fuchs et al., 1998; Xuan et al., 1999; Fried et al., 1999; Rijken et al., 2005). The importance of knowledge in this area is widely acknowledged. Although multimorbidity is often merely used as an explanatory variable in research to adjust for outcomes, the view that it is an object of study is increasingly common. Most studies that focus on multimorbidity fall into one of the following categories. The first type deals with identifying and investigating the clinical relevance of specific disease combinations, usually an index disease and other diseases (Nuyen et al., 2006; John et al., 2003). The second focuses on describing the epidemiology of disease combinations that are found among patients, especially in the general population (Van den Akker et al., 1998; 2001). Although in its approach it is purely descriptive, this study aims to contribute to the first category. In this article, we examine multimorbidity in the Dutch hospitalized population using data on hospital admissions over a 10-year period from a nationwide Dutch hospital register. Our primary goal is to show how a national hospital register can be used to gain insight into hospital multimorbidity, by examining the associations between all possible disease pairs. We also show how these associations can be calculated, summarized, and presented to form a comprehensive overview of hospital multimorbidity and what pitfalls an analyst may encounter when working with these results. In particular, the role of pure chance has to be addressed. Multiple testing could lead to many accidental associations. Many approaches have been suggested to deal with this. The Benjamini-Hochberg procedure is generally considered as one of the most powerful (Benjamini and Hochberg, 1995), but it still can be very conservative. Based on our findings, we propose an alternative way to filter out significant associations that probably follow from chance.

124

EXAMINING MULTIMORBIDITY

7.2 Methods Data For this study, the Dutch Hospital Discharge Register (LMR, see Dutch Hospital Data (2009)) was used. Data are available for the period 1995-2004. All general and academic hospitals and most specialized hospitals in the Netherlands participated in the register during this period. This resulted in a data set that captures nearly all hospital admissions in the Netherlands throughout that period. For each admission, demographic information (date of birth, sex, and postal code) and details regarding the admission (date of admission and discharge, principal and secondary diagnoses, and discharge destination) were recorded. The LMR database was linked to the national Population Register (GBA) by Statistics Netherlands using the date of birth, sex, and postal code that were registered at the time of admission. The linkage allowed for longitudinal analysis of hospital care utilization on an individual level and also for obtaining additional patient information, such as the date of decease. Although the aforementioned linkage keys did not allow for a complete identification of individuals, most admissions were successfully linked to individuals (more than 87% of all admissions; see de Bruin et al., 2004). The demographic composition of the resulting data set may differ slightly from that of the Dutch population (in particular, elderly from nonwestern origins more often have a date of birth, that is, not registered correctly and thus are relatively underrepresented) but is nonetheless considered sufficiently representative by Statistics Netherlands. Diagnoses were recorded in International Classification of Diseases, Ninth Revision (ICD-9) format. They were only recorded by the specialist when relevant to the specific admission in question (i.e., either related to the reason of admission or affecting the length of stay). Definition of multimorbidity Multimorbidity is defined here as the co-occurrence of two or more diseases (discharge diagnoses) within one person, in a specific period of time. This is in line with the common definition (van den Akker et al., 1996). For our classification of diseases, we used the International Shortlist for Hospital Morbidity Tabulation (ISHMT) format, which is compiled by the Hospital Data Project group and adopted by Eurostat, Organisation for Economic Co-Operation and Development (OECD), and the World Health Organisation-Family of International Classifications (WHO-FIC) network. The ISHMT format features a higher aggregation level than ICD-9 and International Classification of Diseases, Tenth Revision (ICD-10). It uses 138 groups of diseases in total, eight of which are related to external causes (which can never be the principal diagnosis of an admission). ICD-9 codes in our database were converted to ISHMT format using 125

CHAPTER 7

the official ICD to ISHMT conversion table (WHO, 2010). For our purposes, we considered the ICD-9 format too detailed, as it would feature many theoretical disease pairings that rarely occur in practice. The ISHMT format is developed with the hospital setting in mind. The disease grouping in the ISHMT format includes both chronic and non-chronic nature diseases. For the analysis, we used different time periods: 1, 3, 5, or 10 years. No distinction between principal and secondary diagnoses is made. Population The data set contained hospital discharges from the period 1995-2004. To study multimorbidity over different periods of time, we selected several time intervals. For a 1-year period, we took 2004 as the baseline year. Findings from admissions in 1998 and 2001 were compared with 2004, to test for temporal consistency of the results, and also as means of validation. On top of that, 3-year periods (1996-1998, 1999-2001, and the baseline period 2002-2004), 5-year periods (1995-1999 and the baseline period 2000-2004), and a 10-year period (1995-2004) are selected to examine the effects of scaling of the observation period. The restriction was imposed that patients were uniquely identifiable throughout the whole period and that they were alive during the whole of the particular period. This was done to obtain a more homogenous population with similar person-at-risk times, such that the selected patients had equal “opportunity” to consume care. Disease-tracking level Multimorbidity was analyzed on a patient level. Disease combinations were restricted to pairs of diseases. This means that for a patient with three diseases (A, B, and C) within the observed period, three pair combinations were found and counted (AB, AC, and BC). Because the ISHMT format uses 130 diagnosis groups (excluding external causes; see above), counts were tracked for 8,385 combinations in total. In this article, we mainly show combinations of diseases that are of relevance for the elderly, and we excluded some classifications that do not refer to specific diseases, such as admissions related to pregnancy (chapter 15), conditions originating in the perinatal period (chapter 16), congenital malformations (chapter 17), symptoms (chapter 18), injury (chapter 19), and contacts with health services (chapter 21). Observed/expected ratio Considering absolute number of persons with a certain disease combination has limited value. The absolute numbers are to a large extent determined by the prevalence rates of each disease in the combination. For instance, heart diseases are part of the most common disease combinations, as the prevalence of these diseases is very high. Therefore, it is more informative to view these prevalences in 126

EXAMINING MULTIMORBIDITY

perspective and compare observed with expected numbers. The latter are the counts that are calculated under the assumption of statistical independence of the occurrence of the respective diseases and are found by multiplying the population size with the prevalence rates (van den Akker et al., 2001). These probabilities vary according to the composition of the study population, depending in particular, on age. In order to control for age and sex, a stratified approach was used. Four age groups, further divided according to sex, were chosen with the middle of the observation period as reference (0-24, 25-44, 45-64, and 65+). Thus, the study population was divided into eight strata, and the probabilities Pk in each stratum k were weighted with the size N k of the stratum. The standardized observed/expected ratio is then as follows: nk ( AB) k N k N k k Ratio    k Pk ( A) Pk ( B) N k  nk ( A) nk ( B) N k Nk Nk k

 Pk ( AB) N k

n( AB) n ( A)n ( B) k k N k k

(7.1)

where nk ( A) , nk ( B) and nk ( AB) denote the number of patients with disease A, B and both (i.e. the pair AB), respectively, in stratum k. A ratio greater than 1 implies a positive association, and a ratio smaller than 1 implies a negative association between the two diseases. To assess the statistical significance of these ratios, 95% lower and upper confidence limits were calculated. The observed counts were assumed to be Poisson distributed, and the Poisson confidence limits were calculated by using Byar’s approximations (Breslow and Day, 1987). Summarizing the ratios The ratios were not only used to confirm known disease pairings and find new ones. In this article, we show listings of which diseases cluster most strongly and which diseases cluster most frequently with other diseases. Validation of the ratios The data set spans a number of years. We make use of this aspect by checking consistency of ratios over time. Furthermore, the longitudinal aspect of the data set allowed us to count disease pairs in individuals over different lengths of time. In most studies, a 1-year observation period is taken (e.g. Van den Akker et al., 1998; 2001) often because researchers are interested in short-term clustering or otherwise because of a lack of long-term data. However, the observed degree of disease clustering within individuals might depend on the observation period. We therefore experimented with different lengths of observation periods to see 127

CHAPTER 7

whether certain clusters are only observed over longer periods of time, or whether the time period has little effect on the clustering.

7.3 Results Characteristics of the hospital study population are found in Table 7.1, for a few different periods of time. Also shown in Table 7.1 is the number of primary diagnoses per individual. These numbers are similar over 1-, 3-, and 5-year periods, respectively. As the observation period is longer, the average number of diagnoses per patient increases, and a shift in the distribution is clearly visible. Over 10 years, however, 37% of the patients still had only one diagnosis. In Table 7.2, the 15 disease pairs most frequently occurring in 2004 are shown. Most pairs involve diabetes mellitus, heart diseases (hypertensive diseases, heart failure, and angina pectoris), and respiratory diseases (chronic obstructive pulmonary disease [COPD] and pneumonia). Some differences may be observed when comparing the 1-year periods, even when adjusted for age and sex (shown between parentheses). The combination diabetes mellitus with heart failure, and combinations of COPD with heart failure, conduction disorders, and diabetes mellitus occurred less often in 2004 than in previous 1-year periods. Differences between periods with similar lengths were found to be smaller when longer observation periods were chosen as basis of measurement (not depicted here). Conversely, diabetes mellitus with hypertensive diseases and cataract with conduction disorders occurred more often in 2004. These combinations are highly prevalent because of high occurrence rates of each separate disease. Note that the percentages increase as the observation period increases. This shows, as is to be expected, that pairs of diseases become more prevalent over a longer time period. Age- and sex-standardized observed/expected ratios were calculated for all pairs. Outcomes were initially classified as positively associated (ratio significantly greater than 1), negatively associated (significantly smaller than 1), and nonassociated (not significantly different from 1). Table 7.3 gives the 20 strongest associations found in our data, for the year 2004. The top 20 in 2004 seems to be dominated by pairs that correspond to similar diagnosis or body system where the symptoms appear, such as acute lower respiratory infections and COPD. The top 20 in other years are not an identical match, however. This is caused by observed counts that change over time (not depicted). Some combinations show an upward trend in the observed counts, such as the pairings of cholelithiasis with other diseases of gall bladder (from 1,654 in 1998 to 2,240 in 2004) and pancreas (from 745 in 1998 to 1,030 in 2004). These varying ratios raise the question whether such approach using an administrative database is sufficiently reliable over time. 128

EXAMINING MULTIMORBIDITY Table 7.1: Characteristics of the hospital population, over different observation periods. 2004

2002-2004

2000-2004

1995-2004

Characteristics

(n=1,414,142)

(n=2,880,410)

(n=4,138,529)

(n=4,521,856)

Age/Sex (%) 0-24, Male 0-24, Female 25-44, Male 25-44, Female 45-64, Male 45-64, Female ≥65, Male ≥65, Female

9.0 7.9 7.6 16.5 13.5 15.0 13.6 17.1

9.3 8.3 8.3 17.5 13.8 15.0 12.0 15.8

10.6 9.2 8.9 17.8 13.8 15.2 10.4 14.1

9.1 8.4 9.8 17.8 15.3 16.6 9.5 13.5

Diagnoses per individual within period (%) 1 2 3 4 ≥5

56.7 23.6 10.4 4.4 5.0

49.1 24.5 12.5 5.9 8.0

43.0 24.8 14.0 7.3 10.9

37.0 23.3 15.0 8.9 15.8

Mean diagnoses 1.84 2.12 2.38 2.76 Notes: Shares of age/sex and diagnoses per individual strata in the hospital population are given in crude percentages, while mean diagnoses are expressed in terms of number of diagnoses per individual. Due to space considerations numbers are not shown for all studied periods of time.

Table 7.2: Crude prevalences of disease pairs and indirectly standardized prevalences (with respect to age and sex distribution in 2004) between parentheses, for different periods of time. Listed pairs refer to the 15 most common pairs in 2004. Group

Group

2004

2002-2004

2000-2004

1995-2004

(n=1,414,142)

(n=2,880,410)

(n=4,138,529)

(n=4,521,856)

Ear Tonsils 1.27 (-) 1.65 (1.58) 1.98 (1.68) 1.66 (1.60) Diabetes Hypertensive 0.40 (-) 0.46 (0.49) 0.47 (0.54) 0.56 (0.64) Diabetes CDO 0.29 (-) 0.35 (0.38) 0.37 (0.45) 0.44 (0.54) Diabetes Heart failure 0.24 (-) 0.28 (0.30) 0.28 (0.33) 0.31 (0.38) HF COPD 0.21 (-) 0.24 (0.26) 0.24 (0.29) 0.28 (0.35) CDO COPD 0.20 (-) 0.26 (0.29) 0.28 (0.34) 0.35 (0.44) Diabetes COPD 0.20 (-) 0.21 (0.23) 0.21 (0.24) 0.23 (0.28) Anaemias Diabetes 0.18 (-) 0.21 (0.23) 0.21 (0.25) 0.24 (0.28) Anaemias Heart failure 0.17 (-) 0.20 (0.22) 0.20 (0.24) 0.23 (0.29) Anaemias CDO 0.17 (-) 0.23 (0.25) 0.25 (0.30) 0.32 (0.40) Diabetes CVD 0.16 (-) 0.21 (0.22) 0.22 (0.26) 0.27 (0.32) Diabetes AP 0.15 (-) 0.21 (0.22) 0.26 (0.30) 0.36 (0.41) CDO Pneumonia 0.14 (-) 0.20 (0.21) 0.21 (0.26) 0.27 (0.34) LM uterus MMP 0.14 (-) 0.21 (0.21) 0.31 (0.30) 0.49 (0.45) Diabetes Cataract 0.14 (-) 0.31 (0.33) 0.43 (0.51) 0.61 (0.73) Notes: “Other” disease groups and pairs within one disease chapter are not shown here. Due to space considerations numbers are not shown for all studied periods of time. Abbreviations: CDO, Conduction Disorders; HF, Heart Failure; COPD, Chronic Obstructive Pulmonary Diseases; CVD, Cerebrovascular Diseases; AP, Angina Pectoris; LM, Leiomyoma; MMP, menstrual and menopausal diseases.

We tried to answer that question by verifying all results within 1-, 3-, and 5-year periods, respectively. Disease pairs were considered as consistent if they either had ratios significantly greater than 1, ratios significantly smaller than 1, or ratios not significantly different from 1, in all periods of similar lengths examined (most of

129

CHAPTER 7 Table 7.3: Disease pairs with the highest O/E ratio in 2004, compared to two other years (1998 and 2001). Disease 1 Mental disorders (Alcohol) Alcoholic liver disease Mood [affective] disorders Other acute lower respiratory infections Diseases of gall bladder HIV Mental disorders (Alcohol) Cholelithiasis Diseases of oesophagus Cholelithiasis Glomerular/renal tubulo-interstitial dis. Peptic ulcer Mental disorders (Alcohol) Malignant neoplasm of ovary Diseases of oesophagus Infectious and parasitic diseases Systemic connective tissue disorders

Disease 2 Alcoholic liver disease Other diseases of liver Other mental and behavioural COPD Diseases of pancreas Other infectious and parasitic diseases Mental disorders (Psych. subst.) Other diseases of gall bladder Dyspepsia Diseases of pancreas Renal failure Dyspepsia Diseases of pancreas Other malignant neoplasms Abdominal hernia Acute lower respiratory infections Glomerular/renal tubulointerstitial dis. Urolithiasis

1998 O/E (95% CI) 47.9 (43.1-53.1) 30.2 (26.1-34.7) 27.3 (26.0-28.6) 15.7 (15.0-16.4)

2004 O/E (95% CI) 39.2 (35.2-43.5) 34.3 (30.2-38.8) 27.4 (26.0-28.9) 22.9 (21.8-24.0)

21.4 (19.1-23.9) 23.1 (19.9-26.6)

22.4 (20.3-24.6) 22.3 (19.0-26.0)

20.4 (18.5-22.6) 18.7 (17.9-19.7) 17.6 (16.3-19.0) 17.1 (15.9-18.3) 21.9 (20.2-23.7)

21.5 (19.8-23.3) 21.1 (20.3-22.0) 19.8 (18.5-21.2) 19.5 (18.3-20.7) 18.9 (17.5-20.3)

14.1 (12.9-15.4) 16.7 (14.9-18.7) 15.7 (14.5-17.0) 17.9 (17.0-18.9) 16.0 (15.2-16.9) 11.4 (09.6-13.4)

18.1 (16.5-19.8) 17.8 (16.1-19.7) 17.5 (16.3-18.8) 17.2 (16.3-18.1) 16.8 (15.9-17.8) 15.1 (13.1-17.4)

Glomerular/renal tubulo-interstitial 13.7 (12.5-14.9) dis. Diseases of oesophagus Peptic ulcer 10.8 (09.7-11.9) HIV Pneumonia 12.6 (09.6-16.1) Notes: Selection of disease pairs limited to pairs from different ISHMT disease chapters (excluding the symptoms, injury, external causes and contacts with health services), with at least 50 observed pairs, space considerations numbers are not shown for all studied periods of time. Abbreviations: E, Expected; O, Observed; dis., diseases; Psych.Subst, Psychoactive substances

14.6 (13.5-15.7) 13.9 (12.6-15.3) 13.7 (10.4-17.6) chapters concerning shown here. Due to

which are found in Table 7.4). About 80.2% of all outcomes were consistent for three 1-year periods (1998, 2001, and 2004). About 19.8% of those outcomes were found to be discordant in at least one of the years. Furthermore, the impact of observed count size and the magnitude of the ratio on the consistency were assessed. Of the inconsistent pairs in the 1-year periods, 80.8% had less than 100 observed pairs in all 3 years. About 71.3% of these cases, in which the number of observations was less than 100, had a ratio smaller than 0.5 or larger than 2 in at least one of the years. It shows that even with a large ratio, results can be very unstable because of a small number of observed cases. For those inconsistent pairs observed at least 100 times, the majority (97.8%) had a ratio between 0.5 and 2 in all 3 years. Similar findings can be found for 3- and 5-year periods, with the difference that small observed counts became increasingly rare. Among these consistent pairs, combinations can be found that are not obvious at first glance and that may draw attention to medically interesting associations. For instance, the groups “diseases of the esophagus” and “paralytic ileus and intestinal obstruction without hernia” were found to cluster (ratio of 2.34 in 2004, with a corresponding

130

EXAMINING MULTIMORBIDITY Table 7.4: Consistency of disease pair associations within an observation period length. Total Consistent Pairs Inconsistent Pairs Consistent Pairs Total positive associations O/E ≥ 2 Remainder Total negative associations O/E ≤ 0.5 Remainder No associations

1 year (3 periods) Count (%) 8,385 (100%) 6,723 (80.2%) 1,662 (19.8%)

5 years (2 periods) Count (%) 8,385 (100%) 7,310 (87.2%) 1,075 (12.8%)

10 years (1 period) Count (%) 8,385 (100%) 8,385 (100%) -

6,723 (100%) 1,586 (23.6%) 866 (12.9%) 720 (10.7%) 4,250 (63.2%) 3,310 (49.2%) 940 (14.0%) 887 (13.2%)

7,310 (100%) 2,883 (39.4%) 1,338 (18.3%) 1,545 (21.1%) 3,618 (49.5%) 1,331 (18.2%) 2,287 (31.3%) 809 (11.1%)

8,385 (100%) 3,801 (45.3%) 1,653 (19.7%) 2,148 (25.6%) 3,284 (39.2%) 751 (09.0%) 2,533 (30.2%) 1,300 (15.5%)

1

Inconsistent Pairs 1,662 (100%) 1,075 (100%) O <100 , 0.5≤ O/E ≤2 386 (23.2%) 202 (18.8%) O <100 , Remainder 957 (57.6%) 293 (27.3%) O ≥100 , 0.5≤ O/E ≤2 312 (18.8%) 530 (49.3%) O ≥100 , Remainder 7 (00.4%) 50 (04.7%) Notes: A disease pair is only considered to be consistent when the outcome (association, negative association or no association) is equal in all periods, for the given observation period length. Likewise, an (dis)association is only defined as strong if the ratio is at least 2 (at most 0.5) and statistically significant in all periods. Consistency check could only be done for two five year periods, an no check could be made for 10 years. Due to space considerations numbers are not shown for a period length of 3 years. Abbreviations: O, Observed; E, Expected.

confidence interval of [1.89, 2.79]). Ileus is a well-known complication of gastrointestinal surgery. It is likely that the combination occurs with the ileus being a complication of surgery for an esophagus disorder, but this invites further research. Other interesting pairings include anaemia and schizophrenia (2.33 [1.99, 2.71]), anemia and dementia (2.85 [2.56, 3.16]), “glomerular and renal tubulointerstitial diseases” and “other diseases of the liver” (3.64 [3.01, 4.36]), and dermatitis and renal failure (4.60 [3.77, 5.56]). Table 7.5 shows the 30 diseases with the greatest number of strong associations, where “strong” means a ratio smaller than 0.5 or greater than 2.0, for 1-, 3-, 5-, and 10-year periods. Septicemia, anemia, and renal failure are often found to be strongly associated to other diseases. Other diseases that are of interest are dermatitis, eczema, and diabetes mellitus. These diseases are remarkable in that most associations are positive. The length of the observed time period was found to have a substantial effect on the association patterns over which multimorbidity is observed. With longer periods of observation, the numbers of “strong associations” were seen to increase, an increase for the ‘strong associations’ and the ‘remainder’ category is visible with a larger time period. With a longer period, it is

1

Because no check could be made on the 10-year period, the distribution of consistent and inconsistent pairs was not known. For indication purposes only, all pairs over 10 year were considered to be consistent, and described accordingly in the table.

131

CHAPTER 7 Table 7.5: Top 30 diseases with the highest number of strong statistical associations (O/E≥2) in 1 year, shown together with strong negative associations (O/E≤0.5) and a remainder category, containing weak (dis)associations and non-associations. Numbers are also shown for 5 and 10 years. 1 year: O/E 5 year: O/E 10 year: O/E Disease ≥2.0 ≤0.5 R ≥2.0 ≤0.5 R ≥2.0 ≤0.5 Septicaemia 54 13 62 70 5 54 75 1 Anaemias 48 12 69 61 3 65 67 1 Other endocrine/nutritional/metabolic dis. 45 23 61 54 2 73 61 0 Other infectious and parasitic dis. 43 13 73 54 2 73 64 0 Other dis. of liver 41 16 72 49 5 75 56 1 Other dis. of the blood 38 13 78 54 3 72 57 1 Renal failure 37 20 72 60 10 59 66 11 Other dis. of the urinary system 37 17 75 47 5 77 54 1 Dyspepsia & other dis. of stomach 33 27 69 47 7 75 60 2 Mental d.o. due to psychoactive subst. 32 27 70 48 13 68 53 10 Pneumonia 32 31 66 40 8 81 46 1 Dermatitis, eczema, papulosquamous d.o. 31 16 82 47 3 79 53 1 Other noninfective gastroenteritis and colitis 31 33 65 42 7 80 51 0 Other dis. of the respiratory system 29 25 75 46 8 75 56 2 Dis. of oesophagus 29 29 71 46 4 79 54 2 Other dis. of the digestive system 29 27 73 48 7 74 58 2 Diabetes mellitus 29 13 87 39 11 79 45 3 Glomerular and renal tubulo-interstitial dis. 28 27 74 46 6 77 51 2 Alcoholic liver disease 28 29 72 44 21 64 50 17 Other mental and behavioural d.o. 28 24 77 41 7 81 57 3 Paralytic ileus, intestinal obstruction 27 28 74 43 8 78 55 2 Human immunodeficiency virus 26 24 79 45 17 67 53 16 Other dis. of intestine 26 25 78 42 9 78 52 1 Systemic connective tissue d.o. 26 24 79 38 8 83 53 4 Heart failure 25 30 74 41 9 79 48 8 Mental and behavioural d.o. due to alcohol 25 47 57 36 25 68 40 17 Other unspecified effects of external causes 24 13 92 49 5 75 58 0 Dementia 24 25 80 28 15 86 34 14 Other symptoms, signs, abnormal findings 23 24 82 39 3 87 46 0 Intestinal infectious dis. except diarrhoea 23 28 78 37 10 82 45 4 Notes: Due to space considerations numbers are not shown for a period length of 3 years. Abbreviations: E, Expected; excl, excluding; O, Observed; R., Remainder; dis., diseases; d.o., disorders; subst; substances.

R 53 61 68 65 72 71 52 74 67 66 82 75 78 71 73 69 81 76 62 69 72 60 76 72 73 72 71 81 83 80

more likely that the individuals will suffer from more than one disease. Thus, over a longer period, more clustering of diseases is found. This is virtually the case for all combinations. For most ratios, this meant a slight increase in magnitude. Pairs for which a change in outcome was found as the length of the observation period increased include atherosclerosis and dyspepsia (ratio 2004: 1.04 [0.77, 1.37], ratio 1995-2004: 2.19 [2.03, 2.26]), infections of the skin and glomerular diseases (1.04 [0.76, 1.40] and 2.23 [2.06, 2.41], respectively), and dyspepsia and dorsalgia (1.13 [0.80, 1.55] and 2.26 [2.08, 2.45]). Dyspepsia, dorsalgia, diarrhea, and gastroenteritis are conditions that frequently formed part of these “temporally inconsistent” pairs. Some combinations share a common risk factor (alcoholic liver disease and intracranial injury, 1.08 [0.47, 2.14] and 3.23 [2.73, 3.78]). The strongest associations for the top 12 diseases are found in Table 7.6. Aside from these diseases having many strong associations, the associations often exhibited a high ratio.

132

12.5 (11.3-13.8) 12.5 (10.9-14.1) 11.0 (10.4-11.6) 10.9 (10.1-11.6) 10.8 (9.50-12.3) 10.2 (8.84-11.7) 10.1 (8.86-11.6) 7.38 (6.76-8.04)

10.2 (8.84-11.7) 9.02 (8.49-9.57) 8.81 (7.85-9.86) 8.35 (6.87-10.1) 7.45 (6.47-8.52) 7.01 (6.45-7.61) 6.92 (6.53-7.33) 5.48 (4.77-6.28)

21.5 (19.8-23.3) 11.0 (9.14-13.2) 9.29 (7.93-10.8) 8.74 (7.85-9.71) 7.99 (7.38-8.63) 6.56 (6.06-7.09) 6.05 (5.65-6.48) 4.76 (4.26-5.31)

Septicaemia Renal failure Glomerular dis. Oth. urinary dis. Oth. infectious dis. Oth. gall bladder Oth. blood dis. Oth. liver dis. Pneumonia

Oth. Blood dis. Septicaemia Anaemias Oth. liver dis. SCT Malign. neopl. lung Oth. respiratory dis. Oth. malign. neopl. Renal failure

Ment. DO (Ps.subst) Ment.DO (Alcohol) Dementia Mood disorders Oth. Ment. DO. COPD Hypertensive Oth. endocrine dis. AMI

Pneumonia Septicaemia Asthma COPD Oth. respiratory dis. Oth. acute resp. inf. Pulm. heart dis. Heart failure Oth. dis. of liver

Renal Failure Glomerular dis. Septicaemia SCT Anaemias Oth. endocrine dis. Oth. urinary dis. Heart failure Oth. gastroenteritis

Anaemia Oth. blood dis. Alcoholic liver dis. Peptic ulcer Renal failure Dyspepsia Oth. digestive dis. Septicaemia Oesophagus

Disease 1 Disease 2

7.38 (6.76-8.04) 7.17 (6.58-7.79) 7.08 (6.85-7.31) 6.75 (6.45-7.07) 5.96 (5.37-6.59) 4.59 (4.19-5.02) 4.10 (3.92-4.29) 3.96 (3.59-4.37)

18.9 (17.5-20.3) 12.5 (11.3-13.8) 7.71 (6.44-9.16) 7.42 (7.04-7.82) 6.65 (6.35-6.96) 6.29 (5.97-6.62) 6.22 (5.88-6.57) 5.94 (5.31-6.62)

9.02 (8.49-9.57) 7.92 (6.66-9.36) 7.42 (6.92-7.95) 7.42 (7.04-7.82) 7.15 (6.73-7.59) 5.87 (5.57-6.19) 4.97 (4.53-5.44) 4.95 (4.61-5.31)

O/E (95% CI)

Dermatitis Oth. dis. of liver Asthma Renal failure Oth. infectious dis. Infections of the skin Anaemias Oth. skin dis. Oth. respiratory dis.

Oth. urinary dis. Septicaemia Glomerular dis. Oth. infectious dis. Renal failure Hyperplasia prostate Urolithiasis Oth. dis. of liver Malign. neopl. bladder

Oth. Endocrine dis. Diarrhoea/gastroenteritis Intestinal inf. dis. Oth. gastroenteritis Renal failure Hypertensive dis. Ment. DO (ps.subst.) Alcoholic liver disease Diabetes mellitus

Disease 1 Disease 2

6.06 (5.06-7.21) 5.51 (4.60-6.56) 4.60 (3.77-5.56) 4.47 (4.03-4.95) 3.84 (3.16-4.62) 3.32 (2.92-3.76) 3.27 (2.81-3.77) 2.75 (2.31-3.24)

11.0 (10.4-11.6) 8.70 (8.25-9.16) 7.67 (7.46-7.89) 6.29 (5.97-6.62) 4.26 (4.06-4.47) 4.22 (3.96-4.50) 4.19 (3.90-4.50) 3.87 (3.59-4.17)

12.5 (10.4-15.0) 11.5 (10.8-12.1) 7.43 (7.15-7.72) 6.65 (6.35-6.96) 6.10 (5.95-6.25) 6.05 (5.65-6.48) 4.82 (4.14-5.58) 4.68 (4.57-4.80)

O/E (95% CI)

Oth. Gastroenteritis Oth. endocrine dis. Crohn's disease Renal failure Paralytic ileus Oth. digestive dis. Intestinal inf. dis. Septicaemia Oth. liver dis.

Dyspepsia Oesophagus Peptic ulcer Oth. digestive dis. Oth. abd. hernia Oth. infectious dis. Anaemias Oth. liver dis. Dis. of pancreas

Oth. infectious dis. HIV Oth.acute resp. inf. Septicaemia Dyspepsia Oth.. urinary dis. Glomerular dis. Acute resp. inf. Oth. liver dis.

Disease 1 Disease 2

7.43 (7.15-7.72) 7.17 (6.40-8.01) 5.94 (5.31-6.62) 4.62 (4.07-5.22) 4.54 (4.07-5.04) 4.40 (3.81-5.04) 3.93 (3.25-4.71) 3.86 (3.32-4.45)

19.8 (18.5-21.2) 18.1 (16.5-19.8) 9.79 (9.01-10.6) 9.04 (8.38-9.74) 8.00 (7.52-8.49) 7.15 (6.73-7.59) 5.77 (5.04-6.58) 5.39 (4.41-6.52)

22.3 (19.0-26.0) 16.8 (15.9-17.8) 10.9 (10.1-11.6) 8.00 (7.52-8.49) 7.67 (7.46-7.89) 6.73 (6.30-7.18) 6.37 (5.97-6.79) 6.24 (5.81-6.70)

O/E (95% CI)

Notes: Selection of disease pairs limited to pairs from different disease chapters (excluding the chapters concerning symptoms, injury, external causes and contacts with health services), with at least 50 observed pairs, shown here. Associations apply to combinations that are consistent and have a ratio of at least 2 in all one-year periods. Abbreviations: COPD, Chronic Obstructive Pulmonary Diseases; dis, diseases; E, Expected; HIV, Human immunodeficiency virus; Inf, Infections; Malign neopl, Malignant neosplasms; Ment. DO, Mental disorders; O, Observed; Oth, Other; Ps subst, Psychoactive substances; Pulm, Pulmonary; Resp, respiratory; SCT, Systemic Connective Tissue disorders; TIA, Transient Ischaemic Attack.

O/E (95% CI)

Disease 1 Disease 2

Table 7.6: Strongest associations in 2004 for a selection of diseases (O≥50).

CHAPTER 7

7.4 Discussion Summary of results This article sought to give an overview of methods that may be used to study multimorbidity when a hospital administrative data set is available. Using standardized observed/expected ratios, the strengths of the pairwise associations between diseases, allowing the pairs to be classified as positively associated, negatively associated, or not associated. This information can be used to confirm known occurrence relations between diseases or search for hitherto unknown associations. If ratios are estimated for all possible pairings, additional information can be derived. Listings can be made of the pairings with the greatest ratios, that is, the strongest associations. Furthermore, it can be identified which diseases cluster most frequently with other diseases. However, a problem with these ratios is that results become unreliable when observed numbers are smaller than 100 and when ratios are close to 1. The magnitude of the ratios can also be sensitive to the length of the time interval over which they are estimated. Strengths and weaknesses of this study As far as we know, no other study has examined multimorbidity among hospital patients in such detail. Multimorbidity has been examined in a similar fashion in the Dutch general practice population (Van den Akker et al., 1998; 2001). Also cluster analysis of US primary care data has been used to assess multimorbidity (Cornell et al., 2008). Specific hospital multimorbidity has been studied by analyzing statistical associations in UK hospital data (Goldacre et al., 2000). We made calculations for all diseases, classified according to the ISHMT format. But these are easily replicable using other formats such as ICD-9 and ICD-10 if further differentiating is needed. The major strength is the use of a data set that encompasses nearly all Dutch hospital admissions. The scope of this data set allows an accurate description of multimorbidity of hospital patients. A novelty of this study is its approach for analyzing multimorbidity over a long term. The same method of comparing observed with expected frequencies was applied using different time intervals of observation. Although this adds little value from an etiological point of view, it yields new insights into how multimorbidity evolves over time. Also, we performed the analysis for several calendar years to further validate our findings. However, this data set did not come without some limitations. First, we had to make a choice on how to deal with individuals who were not alive during the whole period (births and deaths). Our choice to exclude these to obtain a more homogenous study population also introduced a source of bias. Particularly those diseases with a high mortality rate (such as the cancers, acute myocardial infarction, and other heart diseases) were slightly underrepresented in this study. However, a 134

EXAMINING MULTIMORBIDITY

sensitivity analysis showed that these selection criteria did not affect the results in a drastic way. The combinations that were sensitive to these criteria were mainly those that had small observed counts in the first place. Second, the validity of this study depends strongly on how accurate diagnoses were recorded in the administrative process. Also, discharge hospital diagnoses were only recorded when they were relevant to the specific admission in question (i.e., they either were related to the reason of admission or affected the length of stay). This would probably lead to underreporting of specific diseases (Schram et al., 2008). Furthermore, we cannot estimate the frequency and impact of misdiagnoses. We also observed time trends in some diagnostic pairings, without being able to determine their source. The pairings might truly increase or decrease in frequency over time or might simply be a reflection in improvements in diagnostics or a change in classification of the disease. However, no financial incentives are in place for registering specific diagnoses, and it is unlikely that the results of register itself have affected the registration procedure over time (e.g. for purposes of quality control). The Dutch Hospital Register as used in this study is a register funded by public money and was set up with the goal of promoting research in medicine and related disciplines. Hospitals did not get compensated for the number of diagnoses as recorded in this register, but they were compensated by global budgets as set by the Dutch government. Finally, some further potential sources of bias should be mentioned. Thus, the so-called surveillance bias could result if admittance to the hospital for a disease A increases the probability of being diagnosed with disease B. However, if this bias would have been involved, we would expect it to be equally present for all disease combinations. As our data set was restricted to the hospital setting, referral bias certainly also played a role because diseases differ in the likelihood of hospital admission and are treated more exclusively than in other care providers. Just as is the case with surveillance bias, it is difficult to quantify the influence it may have had on our findings. Implications for future research on multimorbidity and health policy This study confirms the potential of using large administrative data sets such as the Dutch Hospital Discharge Register for research. Using an automated approach, statistical associations were deducted for each possible pairs of diseases. This approach is an effective way to confirm and further quantify the prevalence of disease combinations that are known in the literature. But above all, it allows us to detect hitherto unknown associations between diseases. Although the nature and clinical significance of these associations cannot be deducted from administrative data, the finding that there is a statistical association is revealing in itself, prompting further research on the relationship of the respective diseases. In addition, an administrative data set allows for a comprehensive listing of the most 135

CHAPTER 7

strongly associated pairings and the most frequently associated diseases. In combination with the absolute numbers, these can give researchers additional clues as to which pairings may be given priority in further research and formulating health care policies. However, in considering observed/expected ratios estimated in the manner described in this study, researchers should keep the following in mind. Patterns of multimorbidity vary over time. Although our approach was purely statistical, allowing us no insight into the causal mechanisms behind the observed associations, it did become clear that for many diseases a longer period of observations is required before clustering becomes manifest. Thus, clinical and health care researchers interested in assessing etiological relationships should consider this time dependency. The second issue is the variability over time of these ratios. Our analyses show that there is reason for caution when interpreting observed/expected ratios. Although from a strictly statistical point of view, any ratio that differs significantly from 1 can be interpreted as meaning that there is an association, our data show these requirements are not sufficient for practical purposes. The outcomes were sensitive to the observed count sizes and the relative magnitude of the ratio. It reaffirmed the notion that consistency needs to be formally checked for all findings. These caveats are especially relevant in a study as this one, which relies heavily on administrative data, and does not provide any biomedical support for the nature of the associations (Taubes, 1995). One main contribution of this article is that we propose a way to deal with these inconsistent findings, which point toward chance, as an alternative to multiple testing corrections such as Benjamini-Hochberg (1995). Although these corrections are useful when the researcher has no external data to test for validity (which is the case in many studies), they are conservative in terms of the false discovery rate because they are purely based on the probability theory. In our case, we have multiple years of data to check the findings against. It is highly unlikely that a chance finding is consistently reproduced over multiple years (barring any systematic bias because of diagnostic procedures), so information on consistency can be used to filter out chance findings. Suppose the false discovery rate is a in a given year. Assuming independence, the rate of finding a pairing that has seen a significant ratio by chance in n years is an. In our case, the false discovery rate is 0.053 = 0.000125 or 1 of 8,000. Second, if this is not possible (e.g. because no longitudinal data are available), we propose to use our general rule of thumb of only considering disease pairings with an observed/expected ratio outside the (0.5, 2) interval and with an observed count of at least 100. Information on specific disease pairings derived from administrative databases can be used by health care providers (in this case, hospital managements) to improve the quality of care and optimize the efficiency of health care processes. Care should not only be directed at specific diseases but at specific combinations 136

EXAMINING MULTIMORBIDITY

of diseases as well. In particular, the complicating diseases are of great interest to these providers. Governments should consider the complexity of multimorbidity and acknowledge that improving quality and efficiency of care, and having detailed insight into health care costs, requires a deeper understanding of multimorbidity. To achieve that goal, extensive data sets such as ours are needed.

137

CHAPTER 7

138

Chapter 8 Comorbidity and Hospital Care Expenditures: Does One Disease Affect the Expenditures for Another? ‡

Comorbidity is often associated with high health expenditures, although its precise impact remains unclear. In this study we test the hypothesis that hospital expenditures for a principal disease is increased by the presence of a given comorbid disease. We test this for a broad range of diseases. Using extensive treatment and diagnosis information from the Dutch hospital register, a three-part model is proposed. Results showed that hospital expenditures for a specific disease A are not increased by all co-occurring diseases, but only by roughly 7.5% of all disease pairings. For a specific disease pair A and B, it is entirely possible that B might lead to higher expenditures of A, but not vice versa. The implication here is that general trends in comorbidity might lead to additional per capita hospital care utilization, if and only if the diseases that were identified in this study are underlying the trend.

8.1 Introduction Worldwide health care expenditures (HCE) have steadily increased in recent years. In the Netherlands, the HCE have increased from 13 billion to 59 billion euros during 1981-2008 (OECD, 2010). This trend is expected to continue in the coming decades, with the ageing of the population. The growth in HCE has imposed serious concerns among policy makers on the sustainability of current health care systems. The ability to distinguish and understand determinants of health care,

This chapter is based on: Wong A, Boshuizen HC, Polder JJ. 2011. Comorbidity and hospital care expenditures: does one disease affect the expenditures for another? Submitted paper. ‡

CHAPTER 8

both on a macro- and a micro-level, is imperative for solid decision making with regards to the future of health care financing. Koopmanschap et al. (2010) summarized these determinants as follows. They specified four categories (predisposing, enabling, need and societal), while each category was subdivided into two aggregation levels (micro- and macro-level). Predisposing determinants reflect the individual’s propensity towards use; examples are age, gender, household composition and socio-economic status. Enabling determinants refer to resources available to satisfy a need for health care (e.g. informal care supply and income). Need determinants regard health status related variables, such as morbidity and disability. Finally, examples of societal determinants are medical technology and characteristics of the health care system. In the literature proximity to death is found to be an important predictor of high HCE (Zweifel et al., 1999). It can be seen as an approximation of underlying processes (Gray, 2005). More specifically, it can be seen as a proxy for morbidity (Wong et al., 2011a) for hospital care, while for long-term care it can be seen as a proxy for disability (De Meijer et al., 2011). While a higher level of disability is associated with higher HCE, the relationship between morbidity and HCE is not as straightforward. Comorbidity is a particularly complicating factor. Since there exist a multitude of diseases, of which –at least in theory– any combination can coexist, isolating the influence of each disease on total HCE is difficult. In the literature most studies have therefore simplified comorbidity by either studying a small selection of diseases and/or used a summary score. An oft used score is the Charlson’s index (Charlson et al., 1987), which was developed to predict one year mortality. The score is based on the presence on any of 22 diseases, each of which have an own importance weight. Given the higher expenditures in the last year of life it is not surprising that studies have found that a higher Charlson’s score is associated with higher expenditures for a specific main disease (e.g. Beddhu et al., 2000; Charlson et al., 2008; Kuo et al., 2008). Other studies have used other self-developed scores or used the number of diseases as a proxy (e.g. Shwartz et al., 1996; Fleischman & Cohen, 2010; Kuo & Lai, 2010; Zekry et al., 2010). However, to the best of our knowledge, none of these HCE studies have looked into specific combinations of diseases. Combinations of diseases are typically studied in the field of etiology, i.e. the study of the causes of diseases. Etiological research is often based on empirical associations between diseases, which follow from studies that study a particular disease from a biological perspective, or from studies that focus on extracting patterns from large-scale administrative datasets (Van den Akker et al., 1998; Wong et al., 2011b). The main aim of this paper is to contribute to existing HCE literature by exploring the relationship between comorbidity and hospital HCE, for a broad range of diseases. We do this in the context of the ‘Red Herring’, i.e. we distinguish between death-related and -unrelated HCE. The main question that we will answer 140

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

is closely related to the expression ‘The whole is greater than the sum of its parts’. More specifically, in the context of HCE the question arises whether the total HCE for a combination of diseases within an individual is greater than the sum of HCE for each disease separately (i.e. the HCE for two individuals, each with exactly one of the diseases). This is true if the presence of one comorbid disease affects the HCE for another disease. We examine the latter by using a nationwide Dutch Hospital register with extensive diagnosis information and exploring disease-specific hospital HCE. We propose a three-part model to deal with comorbidity, which will be further explained below.

8.2 Methods Many researchers have examined the relationship between age, proximity to death and health care expenditure (Zweifel et al., 1999; Seshamani & Gray, 2004a/b/c; Werblow et al., 2007). The two-part model was the model of choice in most of these studies. It models the proportion of persons with expenditures, and expenditure conditional on having expenditures greater than zero separately. In our attempt to estimate the effects of comorbidity we introduce a disease-specific three-part model. In line with Wong et al. (2011a), it deals with disease-specific hospital inpatient expenditures. In other words, it estimates all expenditures that are related to a disease as a principal diagnosis. The first part of the model is similar to the two-part model, in which the proportion of the population with expenditures for the principal disease is estimated. The second part deals with the number of admissions for the principal disease conditional on having expenditures for that disease greater than zero, and the third part estimates the costs per admission for the principal disease. The three-part model differs from the traditional two-part model in the sense that it is able to deal with the different diagnoses per admission within an individual. Similar to Wong et al. (2011a), separate models are estimated for deceased and survivors. Description of the data The Dutch Hospital Discharge Register (LMR) for the period 1995-2004 was used for this study. This is a nationwide register, to which nearly all hospitals in the Netherlands have contributed their admission records. Each record contains detailed information on patients, from demographics (age and sex) and general admission characters (data of admission, discharge, and urgency of admission), to clinical information (principal and secondary diagnoses on ICD-9 level) and treatment information. Costs per admission were not available, but were calculated by using estimates for hospital day rates and the expenses of medical procedures by using the Dutch costs of illnesses study (Slobbe et al., 2006). The LMR was linked 141

CHAPTER 8

to the Dutch Municipal Population Registration (GBA) and the Cause of Death register, which, together, contain information such as date of birth, gender, living situation, and if applicable, date of death and cause of death. This linkage was considered to be successful and representative for the Dutch population (for the LMR, 87% of the yearly admissions were linked successfully. See Bruin et al., 2004). Definition of comorbidity and hypothesis testing An important discussion point in epidemiological literature is the definition of comorbidity. It is often seen as the co-existence of a chronic condition alongside a chronic index disease. Another oft-used alternative is multimorbidity, which is defined as the existence of two chronic conditions without any hierarchical structure. See Gijsen et al. (2001) for an overview of definitions. For our purposes of estimating HCE of diseases, we have chosen a definition that follows the data structure as closely as possible. In the LMR a distinction is made between principal and secondary diagnoses. Principal diagnoses are diagnoses which are, based on medical grounds, seen as the actual cause for a hospital admission. The secondary diagnoses concern the other diseases that were reported by medical doctors whenever deemed clinically relevant. To preserve this hierarchical structure we have adopted a definition, where the principal diagnosis is seen as the index disease and secondary diagnoses as comorbid diseases. Because hospital HCE are not just restricted to chronic conditions, but also apply to many short-term conditions with an acute nature, we were able to estimate the effect of comorbidity for acute as well as for chronic diseases. Therefore we decided to analyze a broad range of diseases. In terms of the actual impact of each disease on the length of stay and the treatment involved, no information was available, however. This means that the expenditure cannot be a-priori distributed over the principal and secondary diagnoses. We therefore attributed the HCE to the principal diagnosis, with the idea that these costs are dependent on the secondary diagnoses involved. The assumption that underlies this is that the non-admitted population does not have a severe form of the disease(s), because they were not admitted and thus were not diagnosed. Suppose we are interested the HCE of two specific diseases. Let Z1 and Z 2 denote the indicator functions for each disease, i.e. 1Z : Z j  {0,1} . Furthermore, let Y1 and Y2 denote the per capita HCE for each respective principal diagnosis. Suppose the HCE Y1 would be independent of Z 2 , and vice versa: E[Y1 | Z 2 ]  E[Y1 ] E[Y2 | Z1 ]  E[Y2 ]

142

(8.1) (8.2)

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

Now, we are interested in whether this is the case. More specifically, we are interested in whether: E[Y1 | Z 2  1]  E[Y1 | Z 2  0] E[Y2 | Z1  1]  E[Y2 | Z1  0]

(8.3) (8.4)

Note that E[Yi | Z j  1]  E[Yi | Z j  0] is possible as well, but we will not consider these combinations in the remainder of this article. Note that (8.3) does not necessarily imply (8.4), and vice versa. Now, since individuals may have more than one comorbidity, we need to adjust the effect of each diseases for one another. Let  be the link function in the Generalized Linear Model context, X be a covariate matrix, Z be a matrix of disease indicators,  x and  z be their respective regression parameters, then we estimate the following model for each principal disease Yi : E[Yi | X  x, Z  z ]   1  X x  Z z 

(8.5)

Under assumption of additivity on the link scale, we can then find all Z j for which {Z j  Z |  j  0} holds. Suppose we have 94 diseases in total. Using this strategy, we ‘only’ have to fit 94 models1, one for each Yi . Since we have a three-part model of which each part has a different aggregation level, the interpretation of Z j also differs. The first two parts, the proportion with at least one hospital admission and the conditional number of admissions, are defined on an annual level, and thus Z j is also defined on an annual level. This means Z j can either be a comorbid disease for the principal disease within the same admission, or Z j can be a principal or comorbid disease (for a different principal disease) within a different admission. For the third part, the conditional HCE per admission, Z j is defined on admission level, and thus it can only be interpreted as a comorbid disease within the same admission. The average HCE for a specific principal disease conditional on Z j , which follows from the multiplication of all three parts, can then be interpreted as the average HCE for the principal disease given that Z j has been present within the patient all year.

If we for instance used an alternative definition of estimating total HCE given a pair of diseases, the number of models needed would increase quickly with the number of diseases considered. Given 94 diseases, 94!/(92!·2!)=4371 separate models would be needed. Using our definition, we can keep the number of models feasible.

1

143

CHAPTER 8

Note that Z i is only observed in the hospitalized population, such that E[Yi | Z j ] has a specific interpretation: it is not the population-average HCE conditional on the disease, but rather, the average HCE conditional on a severe form of the disease. Given the mandatory health insurance in the Netherlands, and no financial incentives to avoid health care utilization, this a decent approximation of population-average HCE for many severe acute diseases, as these diseases will most likely lead to hospitalization. Data preparation The preparation of data for the analysis largely follows the description in Wong et al. (2011a). The following steps were taken (i) the data was divided by deceased and survivors. In order to strike a balance between amount of yearly observations per individual and the last years of life for deceased, the following definitions were chosen. Deceased individuals were defined as those who died in the study period, or died within five years of the last year of the study period. Since the date of death was available until 2005, only observations of individuals from the period 19952000 were used for the study period. The remaining individuals were considered survivors: these stayed alive throughout the study period, and survived for at least five years after the study period. (ii) The data was reformatted into a panel structure, where, depending on the model, each individual either had an annual (part one and two) or admission level structure (part three). The annual structure was based on the date of death for the deceased, and on the date of birth for the survivors. Furthermore, admissions were added to the observation years based on the discharge data. Information on social-demographics, admission(s), diagnosis and costs were coded in such a way that all variables correspond to the aggregation level of the observation. (iii) Diagnoses were originally coded in ICD-9 format, which we found too detailed, and were subsequently recoded to ISHMT format (WHO, 2010). Furthermore, we made a selection of diagnoses that we found interesting. This leaves us with 94 disease categories (which we will to refer to as ‘diseases’ from hereon). (iv) Individuals that could not be linked to the Dutch Person Register throughout the period were excluded from the data. The resulting dataset had approximately 11.25 million individuals. (v) Since computer hardware restrictions did not allow us to model the full dataset at once, a disproportional sampling procedure was used. Note that, this was only necessary for part one; part two and three were sampled with probability one. First, for each deceased and survivor subset, the data was further split into smaller subsets by using the following subsequent steps: [1] distinguishing between individuals admitted at least once, and not admitted at all during 1995-2000. [2] records for admitted individuals were split by principal diagnosis (each admission has exactly one principal diagnosis). Consequently, individuals that have been admitted for n principal

144

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

diseases are found in n subsets. [3] Stratification by age. Since random sampling within each subgroup resulted in very small numbers in specific groups (e.g. the survivor group at age 90 and higher), smaller groups were sampled at a higher rate, such that the resulting samples had approximately the same amount of individuals. Datasets for part one, given a specific principal diagnosis, were constructed by adding individuals that were admitted for a principal diagnose, individuals that were admitted for other principal diseases, and individuals that were not admitted at all. See Wong et al. (2011a) for a more detailed description and a schematic overview of this procedure. Since the resulting data are not representative of the real population, inverse probability weighting was implemented in further analysis. All models used the Huber-White estimator (Williams, 2000), which yields robust variance estimates. Model specification The distribution of health care expenditure often exhibits a large proportion of zeroes, and is heavily skewed to the right. Common ways to deal with the large proportion of zeroes are using finite mixtures such as a zero-inflated distribution or zero-inflated negative-binomial distribution (Dev & Trivedi, 1997; Deb & Holmes, 2000), or two-part models (Mullahy, 1998). The latter has been particularly popular amongst health economics researchers, because of its relatively simple estimation procedure. For our purposes this two-part model is not appropriate. More specifically, we want to deal with the fact that diagnoses can differ per admission, for a patient in a given year. We therefore extend the two-part model to a three-part model. The first component models the proportion of zeroes p=P[costs>0 | x] in the population, and is identical to the first component of the two-part model: f ( pi ,t )  logit(p i, t )  xiT,t    i ,t

(8.6)

where xiT,t is the transposed vector of covariates for individual i at time t,  is a vector of parameters, and  i,t the residual error for individual i at time t. The second part is different to the traditional two-part expenditure model. The conditional expenditure C* in a specific year can be decomposed into two further parts:

E[C * | C *  0, X ]  E[ A | A  0, X ]  E[C | A  0, X ]

(8.7)

where A is the number of admissions per patient during a calendar year, and C the expenditures per admission.

145

CHAPTER 8

We then proceed to define the second part as the conditional number of admissions per individual as follows: g (ai ,t )  log (ai ,t )  xiT,t    i ,t

(8.8)

where  is a vector of parameters, and  i,t the residual error for individual i at time t. We assume the dependent variable to follow the zero-truncated Poisson distribution or the zero-truncated negative binomial distribution (Cameron & Trivedi, 1998). The main difference between these two distributions lies within the mean-variance relationship. The zero-truncated Poisson family assumes the mean to be equal to the variance: Var[  i ]   i

(8.9)

On the other hand, the negative binomial family is used to deal with overdispersion. For a given overdispersion factor k, the mean-variance relationship can be defined as: Var[  i ]  (1  k ) * E[  i ]

(8.10)

where k>0. Thus depending on whether the data exhibits signs of overdispersion, the Poisson family or negative binomial family needs to be chosen. The likelihoodratio test can be used the significance of k. If k is significant, the negative binomial model is preferred over the Poisson model. In our case, k was not significant. A mean-variance plot (Blough & Ramsey, 2000) for several covariate groups showed that there was, in fact, slight underdispersion. To our knowledge, this is not seen much in real life applications, and accordingly, there are not many standard ways to deal with underdispersion. Therefore we decided to use the zero-truncated Poisson distribution, as its mean-variance relationship is clearly more suitable. The conditional mean of the zero-truncated Poisson distribution is then: P( A  a)  P( A  a)  a P(a  0) a 1 1  P(a  0) a 1 a 1 a   exp( ) 1 1 exp( )a  a =a        a ! 1 exp( ) 1 exp( ) a! a 1 a 1 a  E[ A  a] 1 exp( ) a  = (8.11)  a! 1  exp( ) a 0 1  exp( ) 



E[ A  a | a  0] =  aP( A  a | a  0)   a

146

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

Finally, we define the third part of the model as the costs per admission. For this part we used the algorithm by Manning and Mullahy (2001). The modified Park test suggested that the gamma model with log link was most suitable. The regression model then becomes: h(ci ,t )  log(ci ,t )  xiT,t    i ,t

(8.12)

where  is a vector of parameters, and  i,t the residual error for individual i at time t. As the relationship between age and the outcome in each part could differ substantially between deceased and survivors, we opted to estimate separate models for deceased and survivors. This allows us to not having to deal with age interactions for each disease separately. Generalized Estimating Equations (GEE) were used to account for correlation within individuals (Liang & Zeger, 1986). GEE were chosen above random effect models because the resulting model predictions can be directly interpreted as population averages, as opposed a random effects model which only allows for population averages after taking the average of simulated subject-specific profiles (Molenberghs & Verbeke, 2000). However, as no GEE extension for the zerotruncated Poisson family was available, we used the generalized linear model version with robust standard errors instead. This probably led to a relative small loss of precision, as the number of admission for a given disease, per individual in a given year, is in most cases equal to one. In other words, the correlation between admissions will have a relatively small impact on variance estimates. For the working correlation structure we used the unstructured correlation for part one, and mainly the exchangeable structure in part three, as the unstructured matrix did not lead to convergence in part three. In a few select diseases convergence was not attained under either structure, which is probably caused by short panels. In these cases the independent working correlation matrix was used. Model variations and variable selection We considered two model variants: 1. A three-part model for expenditure of a given principal disease without disease variables: logit( pi,t )  xiT,t   i ,t

(8.13)

log(ai ,t )  xiT,t    i ,t

(8.14)

log(ci ,t )  x    i ,t

(8.15)

T i ,t

147

CHAPTER 8

2. A three-part model for expenditure of a given principal disease with a set of covariates that includes disease variables (that follow our definition of comorbidity) Z next to the variables included in model 1: logit( p*,i,t )  x*,T i ,t *  ziT,t z   *,i ,t

(8.16)

log(a*,i ,t )  x*,T i ,t  *  ziT,t  z   *,i ,t

(8.17)

log(c*,i ,t )  x*,T i ,t  *  ziT,t  z   *,i ,t

(8.18)

Essentially, model 1 estimates average expenditure given average comorbidity, whereas model 2 estimates average expenditure given specific comorbid diseases. Furthermore, if we set the dummies for the disease variables at zero in model 2, i.e. no diseases are present, we obtain predictions for the baseline (no comorbidity). This allows us to compare the effects of specific diseases on the expenditures for the principal disease, compared to expenditure of the principal diseases with average comorbidity, as well as no comorbidity. Both models are run separately for deceased and survivors (Wong et al., 2011a). The models for deceased include cause of death, time to death, time to death squared, calendar year and age as covariates, whereas the models for survivors exclude the variables related to death. The dummy for cause of death discerns between those individuals that die of the principal disease, and those that die of another cause. It is included as an indication of the severity of the principal disease. For some principal diseases, the cause of death dummy was omitted because of small numbers for the disease as cause of death, such as cataract. The time to death variables were used to control for expenditure differences between the last years of life, and calendar year was used to correct for autonomous trends in health expenditure. To estimate the non-linear relationship between age and each outcome, cubic B-Splines were used. Cubic B-Splines are 3rd degree polynomial functions defined piecewise. A mathematical description is given by Newson (2004). The user has to define knots which separate the x range into smaller intervals. The knot picking strategy follows that which is described in earlier work (Wong et al., 2011a). Model predictions were based on 50 to 80 year old females in the year 2000. The secondary diseases that were considered as a covariate were selected from the first 94 disease groups from the ISHMT system. For each principal disease, that same disease was left out as comorbid condition (i.e. a disease can not be comorbid with itself). Since including all disease variables to the model was unfeasible (it led to problems in convergence), we employed a manual backwards selection procedure from each disease chapter. We set the inclusion criterion on p<0.05. Each disease chapter yields a number of significant diseases, and these were put together into a new group. The variables in this new group were then, 148

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

reiteratively, subjected to a backwards selection procedure (with criterion p<0.05), until no disease variables had p>0.05. This whole procedure was only assessed for part one of the model, as we are most interested in the combinations of diseases that occur relatively frequently. In some cases, diseases were left out of part 2 and/or 3 because they were highly insignificant in those parts. Furthermore, for ten diseases, such as leiomyoma of uterus, diseases of anus and rectum, and internal derangement of knee, we had troubles fitting the model, presumably because of small numbers of relevant comorbid diseases. Therefore these diseases were defined to have zero HCE-increasing comorbidities, in the remainder of this article.

8.3 Results In Table 8.1 regression results (model type I) are given for stroke as the principal disease, as an example of the 94 principal diseases we modeled. According to the Wald test, the models for deceased and survivors performed significantly better than their respective null model (p<0.001). Cause of death was significantly associated with a higher proportion of admissions for stroke (p<0.001), a higher number of admissions (p<0.05) and, interestingly, a lower cost per admission (p<0.05). Similarly, a shorter time to death was positively associated with the first two parts (p<0.001), but negatively associated with the third part (p<0.001). Also shown in Table 8.1 are the comorbid diseases that were positively associated with a higher proportion of HCE. Nervous disorders, hypertension, epilepsy, conduction disorders, and urinary disorders were found for both deceased and survivors. On the other hand, Alzheimer’s disease, ear disorders and intestinal infections were comorbidities that were only relevant among deceased, while substance abuse disorders, transient ischaemic attacks and atherosclerosis were only found among survivors. Many of these diseases –but not all– were also positively associated with a higher number of admissions and HCE per admission. However, given the non-linearity of the models, as well as the inclusion of age splines into the models, the coefficients are difficult to interpret. Figure 8.1 plots predictions of each part against age, for two model types (estimation of HCE for average comorbidity and baseline, i.e. no comorbidity). The graphs show that the HCE for stroke were highest within the deceased with stroke as cause of death, followed by deceased with another disease as cause of death, and then survivors. For all three parts, and thus the average HCE, the predictions for average comorbidity were higher than for baseline, although the differences were greatest for the proportion with HCE (part one). The differences were particularly small for the conditional number of admissions per survivor. This was the case for many of the 94 principal diseases. 149

CHAPTER 8 Table 8.1 Three-part regression results for stroke, for deceased and survivors. Part Two

Part One Variable

Beta

S.E.

Deceased Age spline 1 Age spline 2 Age spline 3 Age spline 4 Age spline 5 Age spline 6 Age spline 7 Cause of death Time to death Time to death^2 Sex 1996 1997 1998 1999 2000 Alzheimer's Other nervous disorders Hypertension Epilepsy Ear disorders Conduction disorders Intestinal infections Other urinary disorders

-0.716 -3.323 -2.773 -2.282 -2.272 -2.264 -5.498 3.864 -1.970 0.259 -0.021 0.010 -0.015 0.059 -0.040 0.121 1.943 1.758 1.327 1.218 1.196 0.997 0.940 0.789

3.445 0.486 0.196 0.141 0.131 0.212 1.143 0.040 0.069 0.012 0.038 0.057 0.053 0.057 0.057 0.063 0.488 0.076 0.093 0.215 0.319 0.065 0.385 0.294

Survivors Age spline 1 Age spline 2 Age spline 3 Age spline 4 Age spline 5 Age spline 6 Age spline 7 Sex 1996 1997 1998 1999 2000 Substance disorders Other nervous disorders Hypertension Transient Ischaemic Attacks Epilepsy Conduction disorders Atherosclerosis Other urinary disorders

-9.189 -9.092 -7.269 -6.505 -5.308 -6.402 -1.479 -0.309 0.070 -0.022 0.168 0.075 0.051 3.341 2.942 2.930 2.425 2.287 1.345 1.276 0.978

1.841 0.365 0.174 0.122 0.141 0.321 2.509 0.045 0.056 0.060 0.071 0.072 0.059 0.241 0.091 0.097 0.280 0.316 0.108 0.259 0.214

Sign.

*** *** *** *** *** *** *** *** ***

*** *** *** *** *** *** * **

*** *** *** *** *** *** ***

*

*** *** *** *** *** *** *** ***

Beta

S.E.

-4.579 -1.943 -1.989 -2.233 -2.670 -3.775 -2.414 0.076 0.459 -0.059 -0.217 0.012 0.028 -0.037 0.031 0.110 0.541 0.609 0.292 0.384 0.088 0.147 -0.131 0.219

3.410 0.448 0.172 0.122 0.123 0.217 1.455 0.035 0.060 0.010 0.034 0.055 0.054 0.055 0.054 0.054 0.372 0.037 0.047 0.075 0.190 0.046 0.420 0.234

-2.153 -0.674 -0.958 -1.025 -1.402 -2.994 5.451 -0.117 -0.024 -0.077 -0.145 -0.180 -0.145 -0.174 0.297 -0.021 0.475 0.346 -0.283 0.444 0.118

1.306 0.200 0.087 0.067 0.088 0.250 2.458 0.025 0.039 0.040 0.041 0.042 0.041 0.120 0.034 0.035 0.051 0.072 0.054 0.091 0.066

Part Three Sign.

*** *** *** *** *** * *** *** ***

* *** *** *** **

** *** *** *** *** * ***

*** *** *** *** *** *** *** ***

Beta

S.E.

Sign.

6.173 7.642 7.556 7.981 8.230 8.530 7.541 -0.087 0.982 -0.148 0.125 0.020 0.040 0.001 -0.011 -0.085 0.573 0.345 0.047 0.325 0.498 0.127 0.429 0.604

1.482 0.203 0.073 0.045 0.042 0.066 0.393 0.012 0.021 0.004 0.011 0.019 0.019 0.019 0.019 0.019 0.142 0.017 0.021 0.040 0.089 0.017 0.276 0.118

*** *** *** *** *** *** *** *** *** *** ***

8.301 9.098 9.007 9.055 9.224 9.700 7.867 0.201 -0.017 -0.045 -0.124 -0.134 -0.213 -0.095 0.444 0.039 -0.147 0.211 0.153 -0.019 0.730

0.641 0.094 0.038 0.028 0.033 0.083 0.765 0.010 0.016 0.017 0.017 0.017 0.017 0.049 0.017 0.015 0.048 0.045 0.020 0.067 0.034

*

*** *** *** * *** *** *** ***

*** *** *** *** *** *** *** *** ** *** *** *** *** ** ** *** *** ***

Number of observations [groups] for deceased part one, part two, part three, survivors part one, part two, and part three respectively): 114664[55378], 265554[85106], 59067[N/A], 53338[N/A], 63552[56667], 60914[50621]. Wald’s test for each model: χ2(24)=39548.87 (p<0.001), χ2(20)=103595.71 (p<0.001), χ2(25)=13261.43, χ2(21)= 11306.40, χ2(25)=3003839, χ2(21)= 3676876. Abbreviations: SE, standard error; Sign., significance . *, p<0.05; **, p<0.01, ***, p<0.001.

150

COMORBIDITY AND HOSPITAL CARE EXPENDITURES Figure 8.1: Three-part predictions for stroke for average comorbidity (solid black lines) versus baseline (red dashed lines), distinguished by deceased (cause of death, left column), deceased (other cause of death, center column) and survivors (right column).

0.0005 0.0025

Survivors Proportion

0.025 0.015

0.40

0.50

Proportion

Other cause

Age

Age

Cause of death

Other cause

Survivors

1.08

1.08 1.04

1.08 1.04

50 55 60 65 70 75 80

1.14

Age

Admissions

50 55 60 65 70 75 80

Admissions

50 55 60 65 70 75 80

Age

Cause of death

Other cause

Survivors

9000

5000

12000

Age

Conditional HCE

Age

8000 11000

50 55 60 65 70 75 80

Conditional HCE

50 55 60 65 70 75 80

8000

50 55 60 65 70 75 80

5000

50 55 60 65 70 75 80

Age

Age

Age

Cause of death

Other cause

Survivors 30 10

Average HCE

50

50 55 60 65 70 75 80

100 200 300

50 55 60 65 70 75 80

Average HCE

2000 4000 6000

Average HCE

Conditional HCE

Admissions

Proportion

Cause of death

50 55 60 65 70 75 80

50 55 60 65 70 75 80

50 55 60 65 70 75 80

Age

Age

Age

Estimated values for the number of admission were typically close to one, with the exception of admission for cancers, which is understandable, given the recurrent nature of the treatment for cancers (i.e. chemotherapy sessions). Interestingly, in contrast to the proportion of HCE and the costs per admission, the number of admissions decreased with age. This might be a sign that the hospital care provided was substituted by long-term care for higher ages. To illustrate the effects of specific comorbid conditions on hospital HCE for stroke, HCE predictions for the regressions results from Table 8.1 are plotted in Figure 8.2. It becomes clear that the predicted HCE varied strongly even amongst

151

CHAPTER 8 Figure 8.2: Hospital care expenditures for stroke (Model II). 9

10000

7 6 5 3

6000

Average HCE

14000

Deceased (Cause of death)

8 4

1

Baseline Conduction disorders Hypertension Epilepsy Intestinal infections Other urinary Other nervous Ear disorders Alzheimer's

1: 2: 3: 4: 5: 6: 7: 8: 9:

Baseline Atherosclerosis Conduction disorders Other urinary TIA Epilepsy Hypertension Substance disorders Other nervous

2000

2

1: 2: 3: 4: 5: 6: 7: 8: 9:

50

55

60

65

70

75

80

Age

Survivors

600 400

7

5

0

200

Average HCE

800

9 8

4 2 1

6 3 50

55

60

65

70

75

80

Age

the HCE-increasing comorbid diseases. Particularly, the ratio between the highest HCE and the lowest HCE was high for survivors (around 8), compared to deceased (4). The ratios between HCE given average comorbidity and HCE given no relevant comorbidity are plotted for eight diseases in Figure 8.3. These also varied strongly between principal diseases, ranging from smaller than 1.7 (stroke) to up to 6 (lung cancer). They were strongest for the deceased (other cause of death) group, followed by deceased (cause of death) and by survivors. This might be explained by a higher prevalence of comorbidity amongst the deceased. An exception was stroke, where survivors had a much larger ratio than deceased (cause of death), which might be explained by the sudden and fatal character of stroke among

152

COMORBIDITY AND HOSPITAL CARE EXPENDITURES Figure 8.3: Ratio between HCE with average comorbidity and HCE with no comorbidity for eight common diseases, for deceased with the disease as cause of death (black dots), deceased with another disease as cause of death (red dashes) and survivors (blue solid line). 3.5 1.5 60

65

70

75

80

50

55

60

65 Age

Pneumonia

COPD

70

75

80

70

75

80

75

80

75

80

1.2

Ratio

1.5

1.8

Age

65

70

75

80

50

55

60

65

Age

Age

Diabetes

Renal failure Ratio

1.6

1.5

2.4

60

2.0

55

2.5

50

55

60

65

70

75

80

50

55

60

65

70

Age

Age

Colon cancer

Lung cancer

1.5

2

3.0

Ratio

6

4.5

50

4

Ratio

55

1.2 1.6 2.0

Ratio

50

Ratio

Heart failure

2.5

Ratio

1.5 1.3

Ratio

1.7

Cerebrovascular diseases

50

55

60

65 Age

70

75

80

50

55

60

65

70

Age

deceased. Of all the specific combinations 2 between a principal and secondary disease, we found that 7.5% of them were associated with higher HCE. About 25% of all disease combinations had at least one specific combination (between a principal and secondary disease) 3 that is HCE-increasing. To depict all disease pairings that were associated with a HCE increase, we labeled all diseases with their ISHMT number (see Appendix 8.A), and summarized these pairings in Figure 8.4. In the top panels, the exact combinations were marked with a black dot. For instance, a black dot was found at the cell (5,24). This

2

There are exactly (94-1)·94=8742 combinations of a principal and secondary disease. For a given disease pair A and B, either A or B can be a principal disease. The number of possible pairings is therefore exactly half of all specific combinations, or 4371 combinations.

3

153

CHAPTER 8 Figure 8.4: Patterns of comorbidity for deceased (left) and survivors (right), on disease category level (top) and on disease chapter level (bottom). Disease categories and chapters codings can be found in Appendix 8.A. Black dots in the top panels indicate specific disease combinations for which the comorbid disease increases HCE for the principal disease. Darker areas in the bottom panels indicate a higher number of comorbid diseases that were associated with higher HCE for the principal disease.

60 20 0

40

60

80

0

20

40

60

Principal disease

Principal disease

Deceased

Survivors

80

12 10 8 6 4 2 0

0

2

4

6

8

10

Comorbid disease chapter

12

14

20

14

0

Comorbid disease chapter

40

Comorbid disease

60 40 0

20

Comorbid disease

80

Survivors

80

Deceased

0

2

4

6

8

10

Principal disease chapter

12

14

0

2

4

6

8

10

12

14

Principal disease chapter

indicates that the expenditures for principal disease 5 (Septicaemia) was increased by the presence of comorbid disease 24 (diseases of the blood, excluding anaemia). These plots can be used to find overall patterns of comorbidity. The following can be deduced from these graphs. First, the patterns were similar, but not identical between deceased and survivors, suggesting that the effect of comorbidity on HCE among deceased might be slightly different. Secondly, we saw a stronger concentration of black dots along the diagonal for both deceased and survivors. This indicates that many comorbid diseases that increase HCE of the principal disease involve comorbid diseases that are similar to the principal disease (e.g. a heart disease as a principal disease will often have another HCE-increasing heart disease in its presence). On the other hand, about two thirds of the dots were still 154

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

found outside the diagonal, implying that the HCE-increasing comorbidities can also be completely different from the principal diseases. Thirdly, the patterns were not symmetrical in the diagonal, suggesting that a comorbid disease B might lead to higher HCE for principal disease A, but not vice versa. Finally, some ‘white areas’ were present; these areas indicate that no pairings were found that were associated with additional HCE. This is more obvious when considering the bottom panels in Figure 8.4. Essentially, these panels are an ‘aggregated’ version of the top panels. Each cell represents a potential combination between a principal disease chapter (according to the ISHMT) and a comorbid disease chapter. The cell is given a color based on the number of specific disease pairings found in each other chapter. In other words, they are based on the number of black dots from top panels within each disease chapter. Darker colors indicate a higher number of disease pairings found. Here, the white areas are more accentuated. For chapters 7 and 8 (diseases of eye and ear respectively; see Appendix 8.A) a white horizontal bar is found, suggesting that these chapters are never found as cost-increasing comorbidities. On the other hand, a horizontal and vertical dark line is found for chapter 11 (diseases of the digestive system). This indicates that, as a principal disease, these type of diseases are susceptible to comorbidities (i.e., a lot of other diseases are associated with an increase in HCE for these principal diseases), and at the same time, it often acts as a comorbidity with many other principal diseases. Some of the ‘most comorbid’ and ‘most susceptible’ diseases are plotted in Figure 8.5 (we have omitted the less specific disease categories here, such as “other urinary disorders”; see Appendix 8.A). Here it becomes clear that the ranks are different between deceased and survivors (left and right panels), as well as between comorbid and principal diseases (top and bottom panels). For top 10 most comorbid diseases, only anaemias, substance abuse disorders and dyspepsia were found among both survivors and deceased, whilst for the top 10 most susceptible diseases the only common diseases were chronic obstructive pulmonary disorder, tuberculosis and septicaemia. To further illustrate the frequency of HCE-increasing cases amongst diseases that are likely to be influential for population health and the demand for health care, Table 8.2 gives an overview of some highly prevalent diseases (van Oostrom et al., 2011) and whether these are HCE-increasing combinations. Within this selection disease pairings, this is also clearly not the case. HCE-increasing cases were only found for COPD and asthma (deceased), COPD and dermatitis/eczema (survivors), and partially for diabetes and heart failure, diabetes and COPD, ischaemic heart diseases (excluding acute myocardial infarct) and heart failure, and COPD and asthma (survivors). Thus, even though comorbidity might lead to higher health care demand, and thus, become a bigger strain on the health care system, this can be nuanced by bearing in mind that comorbidity only leads to additional HCE in a few cases. 155

CHAPTER 8 Figure 8.5: Most comorbid (top) and most susceptible diseases (bottom), for deceased (left) and survivors (right). Most comorbid diseases refer to those diseases that were found most often as a cost-increasing comorbidity (y-axis), while most susceptible diseases refer to those diseases that were found most often with a cost-increasing comorbidity (y-axis).

25 20 15 5 0 An ae m De i Su rm as bs a ta D nc ia titis e be di te so s rd e Tu C rs Al be O P co rc D ho u lic Dys los liv p e is er ps di ia se as Hy e pe SC r te TD ns io n

10

5

M

15 10 5 0

Di Pn a be eu te m s on Tu C i a be O P Se rcu D p lo Co tica sis lo e m n ia Ca can Su c rc er in bs ta De om nc m a e d en O iso tia es rd op er ha s gu s

Al

co ho li c

liv

C er O P d Tu i s D be ea O rcu se es lo o s Pe ph is Cr p ag oh tic us n' Ul s di cer s Pa eas Co nc e xa re Re r th a s n ro Se al fa sis pt ilu ica re em ia

0

5

10

15

20

25

Survivors: Most susceptible diseases 20

25

Deceased: Most susceptible diseases Number of associations

Survivors: Most comorbid diseases

0 ul An tip a le em S u E S c ia bs ar ler s ta di os nc so is e di rde s r Al ord s zh e Le e im rs io e r m 's y De om m a en To tia n A s il s Dy sth sp ma ep si a

10

15

20

Number of associations

25

Deceased: Most comorbid diseases

8.4 Discussion The aim of this paper was to analyze comorbidity and its effect on hospital HCE, in the context of death- and survivor-related HCE. Its contribution to existing literature is two-fold. First, it examines HCE for a large range of diseases, where most studies focus on a small selection of diseases. But more importantly, it addresses the extent to which comorbidities affect disease-specific hospital HCE, which has, to our knowledge, not been studied on a large scale. This broad approach has allowed for a bird eye’s view on the relationship between

156

COMORBIDITY AND HOSPITAL CARE EXPENDITURES Table 8.2: Some examples of disease pairs, and whether they are associated with additional hospital care expenditures or not. Deceased HCE higher for

Survivor HCE higher for

Disease 1 (D1)

Disease 2 (D2)

D1

D2

D1

D2

Diabetes mellitus Diabetes mellitus Diabetes mellitus Diabetes mellitus Diabetes mellitus Diabetes mellitus Diabetes mellitus Cataract Cataract Cataract Cataract Ischaemic heart disease Ischaemic heart disease Ischaemic heart disease Ischaemic heart disease Ischaemic heart disease Heart failure Heart failure Cerebrovascular diseases Cerebrovascular diseases COPD COPD

Cataract Other ischaemic heart disease Heart failure Cerebrovascular diseases COPD Dermatitis and eczema Dorsalgia Other ischaemic heart disease Cerebrovascular diseases COPD Dermatitis and eczema Heart failure Cerebrovascular diseases COPD Asthma Dermatitis and eczema Cerebrovascular diseases COPD COPD Dermatitis and eczema Asthma Dermatitis and eczema

No No No No No No No No No No No No No No No No No No No No Yes No

No No Yes No No No No No No No No Yes No No No No No No No No Yes No

No No No No No No No No No No No No No No No No No No No No No Yes

No No Yes No Yes No No No No No No No No No No No No No No No Yes Yes

hospital HCE and comorbidity. Out of the 8742 specific combinations between a principal and secondary disease considered, about 7.5% were found to be positively associated with higher hospital HCE. About one third of these associated pairings were diseases with a similar classification, i.e., diseases from the same disease chapter (e.g. one heart disease with another), meaning a large fraction (two thirds) comprise of diseases with different classifications. About 25% of all disease combinations had at least one case where a secondary disease was HCEincreasing for a primary disease. Furthermore, it was found that comorbidity patterns, while similar, were not completely the same for deceased and survivors. Finally, if a disease B led to higher HCE for disease A, the converse may not necessarily be true. The implications of this finding pertain to aggregate HCE spending in the future. In several Dutch health surveys (Hoeymans et al., 2012), a trend in the prevalence of comorbidity in the Dutch population was found for the period 19902007, even after correcting for demographic composition. While in itself comorbidity might lead to more spending, additional HCE due to the presence of two diseases is highly dependent on the specific comorbidity involved. In the end, trends in specific diseases and disease pairings will probably determine the per capita need for health care. A particularly heavily debated topic in health economics is ‘healthy ageing’. Fries (Fries, 1980; Vita et al., 1998) found using

157

CHAPTER 8

Medicare data that elderly become healthier over time, i.e., a 65-year old is healthier nowadays than a 65-year old would be decades ago. Therefore, he proposed the concept of ‘compression of morbidity’, where morbidity would be postponed towards the last years of life. Our results suggest that, if this concept indeed holds true, the effects of compression of morbidity for lifetime HCE depend highly on what type of (co-)morbidity is involved. If the diseases that are involved in the compression also belong to one of our established disease combinations, then it might be possible that the HCE in the last years of life increase so much that it overshadows the HCE reductions in the life years beforehand. If the diseases do not belong them, then a reduction in lifetime HCE will probably follow from a compression of morbidity. This study has a few limitations. First of all, only hospital inpatient care HCE (including day cases) were considered because of our data availability. Other health care sectors, such as primary, ambulatory and long-term care, may yield different disease pairings. Secondly, diagnoses were based on a hospital register. The hospital HCE estimates described in this paper are conditional on having a comorbid disease that is severe enough for the patient to be registered by the medical specialist or to lead to another hospital admission in the same year as the hospital admission for the principal disease. For severe diseases such as acute myocardial infarct and cancers, this might not be a large problem. Diseases like hypertension however, do not always lead to hospitalization. Thus, the costs of hypertension are likely to be overestimated in our study, because the less severe cases of hypertension –which could only be known by the general practitioner– were not registered as hypertensive patients in our analyses. Finally, the results are heavily dependent on the diagnosis coding used. We used the ISHMT format, so the results might differ when calculated on a different aggregation level, e.g. ICD10.

8.5 Conclusion Changing health status patterns in the population may lead to shifts in aggregate health care spending. In this paper we studied the relationship between HCE and one crucial dimension of health, namely comorbidity. More specifically, we tested whether comorbidity always leads to additional HCE. We found that it only holds true for specific disease pairings, and as a consequence, trends in comorbidity may only lead to higher additional expenditures when these trends are caused by those disease pairings.

158

COMORBIDITY AND HOSPITAL CARE EXPENDITURES

Appendix 8.A: ISHMT classification of diseases Table 8.A1: Overview of disease categories in this study, with their numbering. Also given are the disease chapters under which the disease categories fall, highlighted in gray. Disease chapters are numbered using roman numerals. No.

Full Description

(I)

Certain infectious and parasitic diseases

1

Intestinal infectious diseases except diarrhoea

2

Diarrhoea and gastroenteritis of presumed infectious origin

3

Tuberculosis

4

Septicaemia

5

Human immunodeficiency virus [HIV] disease

6

Other infectious and parasitic diseases

(II)

Neoplasms

7

Malignant neoplasm of colon, rectum and anus

8

Malignant neoplasms of trachea, bronchus and lung

9

Malignant neoplasms of skin

10

Malignant neoplasm of breast

11

Malignant neoplasm of uterus

12

Malignant neoplasm of ovary

13

Malignant neoplasm of prostate

14

Malignant neoplasm of bladder

15

Other malignant neoplasms

16

Carcinoma in situ

17

Benign neoplasm of colon, rectum and anus

18

Leiomyoma of uterus

19

Other benign neoplasms and neoplasms of uncertain or unknown behaviour

(III)

Diseases of the blood and bloodforming organs and certain disorders involving the immune mechanism

20

Anaemias

21

Other diseases of the blood and bloodforming organs and certain disorders involving the immune mechanism

(IV)

Endocrine, nutritional and metabolic diseases

22

Diabetes mellitus

23

Other endocrine, nutritional and metabolic diseases

(V)

Mental and behavioural disorders

24

Dementia

25

Mental and behavioural disorders due to alcohol

26

Mental and behavioural disorders due to use of other psychoactive subst.

27

Schizophrenia, schizotypal and delusional disorders

28

Mood [affective] disorders

29

Other mental and behavioural disorders

(VI)

Diseases of the nervous system

30

Alzheimer's disease

31

Multiple sclerosis

32

Epilepsy

33

Transient cerebral ischaemic attacks and related syndromes

34

Other diseases of the nervous system

(VII)

Diseases of the eye and adnexa

35

Cataract

36 37 / (VIII)

Other diseases of the eye and adnexa Diseases of the ear and mastoid process

159

CHAPTER 8 Table 8.A1 (continued). No.

Full Description

(IX)

Diseases of the circulatory system

38

Hypertensive diseases

39

Angina pectoris

40

Acute myocardial infarction

41

Other ischaemic heart disease

42

Pulmonary heart disease & diseases of pulmonary circulation

43

Conduction disorders and cardiac arrhythmias

44

Heart failure

45

Cerebrovascular diseases

46

Atherosclerosis

47

Varicose veins of lower extremities

48

Other diseases of the circulatory system

(X)

Diseases of the respiratory system

49

Acute upper respiratory infections and influenza

50

Pneumonia

51

Other acute lower respiratory infections

52

Chronic diseases of tonsils and adenoids

53

Other diseases of upper respiratory tract

54

Chronic obstructive pulmonary disease and bronchiectasis

55

Asthma

56

Other diseases of the respiratory system

(XI)

Diseases of the digestive system

57

Disorders of teeth and supporting structures

58

Other diseases of oral cavity, salivary glands and jaws

59

Diseases of oesophagus

60

Peptic ulcer

61

Dyspepsia and other diseases of stomach and duodenum

62

Diseases of appendix

63

Inguinal hernia

64

Other abdominal hernia

65

Crohn's disease and ulcerative colitis

66

Other noninfective gastroenteritis and colitis

67

Paralytic ileus and intestinal obstruction without hernia

68

Diverticular disease of intestine

69

Diseases of anus and rectum

70

Other diseases of intestine

71

Alcoholic liver disease

72

Other diseases of liver

73

Cholelithiasis

74

Other diseases of gall bladder and biliary tract

75

Diseases of pancreas

76

Other diseases of the digestive system

(XII)

Diseases of the skin and subcutaneous tissue

77

Infections of the skin and subcutaneous tissue

78

Dermatitis, eczema and papulosquamous disorders

160

COMORBIDITY AND HOSPITAL CARE EXPENDITURES Table 8.A1 (continued). No.

Full Description

(XIII)

Diseases of the musculoskeletal system and connective tissue

80

Coxarthrosis [arthrosis of hip]

81

Gonarthrosis [arthrosis of knee]

82

Internal derangement of knee

83

Other arthropathies

84

Systemic connective tissue disorders

85

Deforming dorsopathies and spondylopathies

86

Intervertebral disc disorders

87

Dorsalgia

88

Soft tissue disorders

89

Other disorders of the musculoskeletal system and connective tissue

(XIV)

Diseases of the genitourinary system

90

Glomerular and renal tubulo-interstitial diseases

91

Renal failure

92

Urolithiasis

93

Other diseases of the urinary system

94

Hyperplasia of prostate

161

CHAPTER 8

162

164

Chapter 9 Medical Innovation and Age-Specific Trends in Health Care Utilization: Findings and Implications ‡

Health care utilization is expected to rise in the coming decades. Not only will the aggregate need for health care grow by changing demographics, but will per capita utilization increase as well. It has been suggested that trends in health care utilization may be age-specific. In this paper, age-specific trends in health care utilization are presented for different health care sectors in the Netherlands, for the period 1981-2009. For the hospital sector we also explore the link between these trends and the state of medical technology. Using aggregated data from a Dutch health survey and a nationwide hospital register, regressions were used to examine age-specific trends in the probability of utilizing health care. To determine the influence of medical technology, the growth in age-specific probabilities of hospital care was regressed on the number of medical patents while adjusting for confounders related to demographics, health status, supply and institutional factors. The findings suggest that for most health care sectors, the trend in the probability of health care utilization is highest for ages 65 and up. Larger advances in medical technology are found to be significantly associated with a higher growth of hospitalization probability, particularly for the higher ages. Age-specific trends will raise questions on the sustainability of intergenerational solidarity in health care, as solidarity will not only be strained by the ageing population, but also might find itself under additional pressure as the gap in health care utilization between elderly and non-elderly grows over time. For hospital care utilization, this process might well be accelerated by advances in medical technology.



This chapter is based on: Wong A, Wouterse B, Slobbe LCJ, Boshuizen HC, Polder JJ. 2012. Medical innovation and age-specific trends in health care utilization: findings and implications. Social Science and Medicine 74(2): 263-272. This article has been reproduced with permission of Elsevier.

CHAPTER 9

9.1 Introduction Health care utilization in the Netherlands and most Western countries has risen considerably in recent decades, and as a result, the health care expenditures (HCE) have experienced a large growth as well. Over the period 1981 to 2008, the share of HCE in the Gross Domestic Product has risen from 7.6% to 9.9% in the Netherlands. HCE have increased from 13 billion to 59 billion euros (current prices), which corresponds with an increase of over 450% (OECD, 2010). In the Dutch definition of health care, which includes many long-term care provisions, this increase was even higher (Statistics Netherlands, 2010). The growth in HCE is expected to continue in the following decades, as a consequence of the ageing population, which will put considerable pressure on health care financing. Additionally, it requires a greater amount of intergenerational solidarity in countries with mandatory health insurance or social insurance systems, such as the Netherlands (Ministry of Welfare, Health and Sport, 2010). Since there are more elderly in absolute and relative terms, the non-elderly (working population) need to contribute higher premiums or tax payments while receiving a smaller proportional amount of care in return. In other words, the net payments of non-elderly increase with the ageing of the population. Ageing has garnered a lot of attention in health economics. Increasing life expectancy leads to more health care utilization over the total life span, which will lead to further gains in life expectancy, causing this cycle to start all over again. Zweifel et al. (2005) likened this to Sisyphus’ work, and labeled this the ‘Sisyphus Syndrome’ in health care. However, in recent years many researchers have argued that ageing itself has a smaller impact on HCE than previously thought, as proximity to death was found to be a much better predictor of high HCE (Zweifel et al., 1999; Seshamani & Gray, 2004; Werblow et al., 2007). This has also been confirmed in the Netherlands (Polder et al., 2006; Wong et. al, 2011a). The conclusion by Zweifel et al., in particular, is that ageing has diverted attention from other factors that may drive the health care expenditures. Much like time-to-death caused the age profiles in HCE to steepen, these factors may influence the shape of the age profiles, and ultimately, the net money transfers between generations. Studies on these age profiles over time have led to mixed conclusions. Polder et al. (2002) found that the relative growth in acute care HCE was greater for elderly during 1988-1994 in the Netherlands, while for long-term care HCE the exact opposite was found (Note that a relative greater growth also implies a absolute greater growth, because of the higher costs within elderly). Meara et al. (2004) analyzed US per person health care spending, and found that spending amongst the elderly grew relatively more rapidly than amongst the non-elderly from 1963 to 1987, but that this trend was reversed during 1988-2000, as a result of health care reforms. Buchner & Wasem (2006) found that growth in HCE was greatest, in 166

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

absolute terms, for the elderly as well in Germany during 1979 to 1996, and coined this phenomenon ‘age steepening’. Felder & Werblow (2008) could not find signs of the steepening in the age-HCE relationship for Swiss health insurance data from 1997-2006. When Zweifel et al. (1999) referred to ‘other factors’ that may influence HCE, they mainly alluded to the role of medical technology. Although Reinhardt (2003) and Getzen (2000, 2006) have argued that health care spending might depend more on what we can afford to spend based on national income, health care spending has an increasing share in the national income for all western countries, suggesting that the willingness to spend for health care becomes greater as well. The combination of population ageing and advances in medical technology might be a rationale for this willingness. Jones (2002) and Dormont et al. (2006) have suggested medical technology may be oriented towards the elderly. Consequently, medical technology may cause age profiles to steepen. To our knowledge, no study has tried to explore these effects. Rather, many studies have tried to isolate the influence of medical technology on aggregate HCE spending. These studies fall into two categories, the first of which is based on Newhouse (1992). In this seminal paper, the effect of medical technology on HCE was estimated by first accounting for relevant, quantifiable, factors such as per capita income, insurance coverage level and population ageing, and then attributing the remaining growth for the greater part to ‘the enhanced capabilities of medicine’. Other studies took a similar approach, assuming the residual growth as a proxy measure for the progress in technology (Peden & Freeland 1998; Buntin et al., 2004). The main issue with these models is that the residual effects can be caused by a wide range of phenomena, not limited to just medical technology. A second class of studies used variables that can be measured in a fairly consistent and objective manner. Okunade & Murthy (2002) used R&D spending as a proxy for the state in medical technology, and found that they were significantly associated with higher HCE. Other studies used the diffusion of specific technologies or procedures as a proxy for medical technology. Gerdtham et al. (1998), Oh et al. (2005) and Christiansen et al. (2006) used the number of renal dialyses and the number of tomography scanners per million inhabitants to approximate the diffusion level of technology, and found both to be significantly associated with higher HCE. Although these proxies are fairly objective, they may not accurately reflect advances in medical technology. R&D investments only refer to input, rather than actual technology output, and numbers of a specific piece of technology may not reflect the growth of technology as a whole. The goal of this paper was (1) to address age-specific trends in health care utilization for the Netherlands, and estimate these trends in the Netherlands for different health care sectors, and (2) to investigate whether medical innovation

167

CHAPTER 9

might influence the difference in trends between age groups for the hospital inpatient care sector. More specifically, the following hypotheses are tested. First, we reassess agespecific health care trends in the Netherlands for a longer period than Polder et al. (2002), and for specific health care sectors. The hypothesis that health care utilization increases faster for elderly is being examined for different health care sectors. Secondly, we examine whether the growth in the probability of health care utilization is associated with the state in medical technology. Thirdly, we test the hypothesis that this association is different between age groups. To do this, the probability of health care utilization is examined. A longrunning Dutch survey (1981-2009) is used to determine age-specific trends in the probability of health care utilization for different health care sectors, and to test whether the trend for elderly is steeper than for non-elderly. Secondly, using the Dutch Hospital Register, a regression model is used to explore the relationship between the state in medical technology and the age-distribution of hospital care utilization, whilst controlling for other variables such as policy changes and demographic changes. The state in medical technology is approximated by the number of granted patents in the area of medical technology. Patents are assumed to be a better proxy than R&D, because patents refer to actual output whereas R&D only regards input. Similarly, we prefer this over the diffusion of a single piece of technological equipment (e.g. tomography scanners), as patents reflect a wide range of technological progress.

9.2 Methods Data Five different aggregate data sources were used for this study, two of which are maintained by Statistics Netherlands. They include the POLS health survey from 1981 to 2009, which is a cross-sectional survey of approximately 10,000 individuals per year. This survey is considered to be representative for the Netherlands by Statistics Netherlands. Participants are asked about several aspects pertaining health and health care utilization. The latter is differentiated by sectors: e.g. general practitioner care, alternative medicine practitioner (similar to general practioner, with the difference that non-conventional medicine is applied), prescribed and non-prescribed drugs, dental care, physiotherapy, and medical specialist care. We used aggregated statistics (n=29; 29 time points for each health care type, except for alternative medicine practioner, where points were missing for 1983/1984; thus, having n=27), freely downloadable at Statistics Netherlands’s website (Statistics Netherlands, 2010). An alternative source was used for hospital inpatient care. The Dutch Hospital Register (LMR) is a nationwide hospital register, in 168

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

which nearly all hospitals participate, with the exception of a few specialized hospitals. Thus, it yields a nearly complete picture of inpatient hospital care in the Netherlands. Age- and gender-specific LMR statistics for the period 1981-2008 were also available from Statistics Netherlands. For 1981-2008 the annual hospitalization rates were available on an aggregated level, while for 1995-2008 annual individual probabilities of admission were also available. Individual hospitalization probabilities differ from hospitalization rates in the sense that individual probabilities take into account that individuals might be admitted more than once in a given year, thus giving an actual account of what proportion of the population utilized hospital care. No missing values within these periods were present, and the rates and probabilities were available by age and gender (n=28 for each age- and gender-specific hospitalization rate, and n=14 for each age- and gender-specific hospitalization probability). LMR hospitalization statistics are more accurate than those given by the POLS health survey. Finally, age- and genderspecific mortality rates from Statistics Netherlands were also used in our analysis. The rates are based on the Dutch municipal register, making it representative for the Netherlands. These were available for the whole period (1995-2008; n=14 for each age- and gender-class). Additional data were collected from the Dutch Health Care Authority (NZa) and the Dutch Patent Office. The NZa is the Dutch health care authority that, amongst others, determines annual hospital budgets, and by doing so, indirectly determines how much capacity in terms of materials and personnel hospitals can offer. Finally, we use filed patent numbers from the Dutch Patent Office, which represents the Netherlands in international patent organizations like the World Intellectual Property Organization and the European Patent Office. It keeps track of all worldwide patents, not limited to just Dutch patents. For this study, only patents concerning medical technology were selected. These patents fall into the following areas: drugs, medical devices, implants, nanotechnology, genetic and biochemical techniques, imaging techniques and bioinformatics (Knecht & Oomen, 2006). We took the total of these patents, with a time lag to approximate the moment of grant, and assumed that these lagged totals are a proxy for the state of medical technology advancement in a given year. No ethical approval was sought for the datasets, because they involve aggregated statistics from which no personal information can deduced. Estimating age-specific time trends from POLS data For the first part of our analyses, we analyzed health care utilization on an individual level. Age-specific probabilities of health care utilization were modeled with a regression model that allows us to determine in which age groups the probability of health care utilization has evolved most over time. In essence, we are revisiting previous literature on age-specific trends (Meara et al., 2004; Buchner & 169

CHAPTER 9

Wasem, 2006; Felder & Werblow, 2008) for different health care sectors; which gave mixed results as to whether elderly have seen the biggest growth in health care utilization. Let p s ,a ,t be the probability of utilizing health care sector s , for age group a j (j=0,1,2,3) at continuous time t , where j corresponds with age groups 65+, 0-19, 20-44, and 45-64 respectively. Let i be the subscript denoting the ith observation. Suppressing subscripts s, a and t to keep the notation clear, we can write the model as follows: 3

3

j 1

j 1

pi   0    j ai , j  t i    j ai , j t i   i ,

(9.1)

where  and  j are our parameters of interest.  describes the time trend for reference age group 65+, and  j the time trend deviation for age group a j . Particularly, we test whether  j is significant for each j, i.e., whether age-specific time trends in the probability health care utilization exist. Furthermore, if they are indeed different, we will examine the sign and coefficient magnitude of the  j for all j, to determine which age group has seen the greatest time trend. Because we use aggregated data (i.e., observed probabilities, rather than individual data), this model cannot be fitted with a straightforward logistic regression model. A reasonable alternative is to assume a normal distribution, truncated at zero and one (because 0  pi  1 ; see Greene, 2008). This was fitted with the software package GAMLSS in R (Stasinopoulos & Rigby, 2007). Estimating age-specific time trends from LMR data We used the POLS data to examine most health care sectors, but used the LMR dataset to examine time trends in hospital inpatient care. More specifically, we used hospitalization rates (average number of admissions per 100 inhabitants) from (non-day care) admissions 1981-2008. Since these rates could not be converted to individual probabilities, the comparison with the age-specific trends from POLS data was not fully compatible, but they did give a reasonable indication of individual-specific probabilities. The aggregate LMR statistics are more refined than the health survey statistics. The hospitalization rates are differentiated by 20 age groups, each of which has a width of five years, and by gender. Here, we fitted the following model for each gender g (we suppress indices a, g and t): 19

19

j 1

j 1

pi   0    j  ai , j    ti    j  ai , j ti   i ,

170

(9.2)

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

which only differs from the POLS model in the sense that (a) it is gender-specific and (b) it uses more age categories. The reference age-category is now age 95 and over. Age dummies were preferred over polynomials and cubic splines, because they resulted in a better fit, and allow for a more straightforward interpretation. Similar, to the POLS model, we are interested in the significance, sign and coefficient size of each  j , as this allows us to see for which age group hospital inpatient care has seen the biggest rise. Determining the relationship between the age- and gender-specific probability of hospitalization and the state of medical technology Once we have established (potential) differences in age-specific health care utilization probability trends, the question then remains (a) whether the trends in health care utilization probability can be linked to advances in the state of medical technology and (b) whether this relationship is different by age, i.e., whether technology will affect certain age groups more than others. We do this for one health care sector, hospital inpatient care, as only for this sector the most relevant confounders were available. To determine the association between the age- and gender-specific probabilities of hospitalization and utilization of medical technology, we used data from 1995 to 2008 only, as (1) individual probabilities were available from 1995 onwards and (2) a major confounding variable, hospital global budget, was only available from 1995 onwards as well. Initially, the following model form was used: pi ,t   0  1  Genderi   2  Agei    X i ,t   i ,t ,

(9.3)

where age and gender signify the age and gender strata respectively, X i ,t a set of time-dependent covariates,  i,t the residual error term, and pi the hospitalization probability in the population, given specific values for age, gender and X i ,t . However, strong serial correlation was found in the dependent, as well as the independent variables. Partial autocorrelation functions revealed a lag of order 1 in the dependent and independent variables. Breusch-Godfrey and Durbin-Watson tests rejected the null hypothesis of no serial correlated errors up to order 1 (p<0.001). Therefore the violation of the OLS assumption of independent errors could result in wrong standard errors and wrong inferences (Wooldridge, 2002). To avoid these problems, we adopted the strategy of Pischke and Velling (1997), who estimated differences in unemployment rates as a function of differences in immigration rates in Germany. Here, we differenced the response variable, and also the covariates, except for the age and gender terms, with respect to time t:

171

CHAPTER 9

pi ,t   0  1  Genderi   2  Agei    X i ,t   i ,t ,

(9.4)

where pi ,t denotes the first difference between subsequent years within an age and gender class, i.e., page, gender ,t  page, gender ,t 1 . Similarly, the interpretation here is that the growth (note that, from hereon, when we speak of growth, absolute growth is implied) in the proportion of hospital patients over time depends on the growth of several covariates, as well as on the age and gender class of the population. Once again, we fitted the model using a truncated normal distribution, this time with truncation at -1 and 1 (  1  pi ,t  1 ). The Breusch-Godfrey and Durbin-Watson tests no longer rejected the null hypothesis of no serial correlated errors. The errors were approximately normally distributed with mean 0 and variance  2 . The Breusch-Pagan test rejected the null hypothesis of homoscedasticity (p<0.001), so we opted to use the White’s heteroscedasticityconsistent estimator for estimation of standard errors. While this estimator is conservative, it requires no assumption on the heteroscedasticity function, in contrast to Weighted Least Squares (Wooldridge, 2002). The selection of covariates followed from various studies in the literature (see Koopmanschap et al., 2010, for an overview). Our selection incorporates aspects from the following determinants: predisposing determinants (e.g. age, gender, household composition), need (e.g. mortality, disability morbidity), enabling determinants (e.g. hospital care supply, individual income, national income), and societal determinants (e.g. technological innovation, institutional factors and wages). This selection was based on what was available, and what was considered most important. We also kept our model as parsimonious as possible and reduce the risk of overfitting. This meant that many demand side variables were not included. Getzen (2000) argues that “With insurance, it is the average income of the group, and the fraction that the group is collectively willing to spend on medical care, which determines the health care budget, not the income of the particular patient being treated.” Thus, individual income was not included as the Netherlands feature a mandatory health care insurance with very little copayments. Predisposing determinants in need were expressed in terms of age and gender. Age is essential, since our interest in the age-specific relationship between age and technology. Furthermore, gender was included since the data showed gender differences in hospital admission probabilities. In many macro-level HCE studies (Gerdtham et al., 1998; Christiansen et al., 2006), need was approximated by age, or fraction of elderly in the population. Health status was accounted for here by including age- and gender-specific mortality. It is widely known that health care utilization in the last years of life is much greater than in other years; particularly, the probability of hospital care admission is much higher (Seshamani & Gray, 2004). As was argued by Wong et al. (2011a), mortality rates approximate 172

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

morbidity, which is the underlying reason for hospitalizations. Getzen (2000) argued that health care expenditures, and therefore utilization, is mainly determined by how much a nation can spend. Thus, enabling factors in the form of supply factors need to be implemented in the model. Rather than using the number of specialists, the number of hospital beds, or national income as a proxy for supply (e.g. Gerdtham et al., 1998; Christiansen et al., 2006), we used the global allocated hospital budgets as an enabling determinant. These budgets can be seen as more precise indicators of what the government is willing to spend on hospital care, than national income for instance. The budget financing system was implemented by the Dutch government to control HCE growth, as well as to stimulate efficiency as hospitals had to maintain production whilst adhering to their respective hospital budgets. After the turn of the century, it became apparent that this system was consistently underestimating hospital care demand, as the patient waiting lists became longer (Crom, 2005). In 2001, the government attempted to reduce these waiting lists in two ways. First, they increased hospital budgets with a considerable amount (958 million euro in 2001 and an additional 1082 million euro in 2002). This hospital budget was included as a continuous variable. Secondly, they altered the way in which the lump-sum wages of medical specialists were established. During 1995-2000 lump-sum wages for specialists were based on a fixed pre-allocated budget. In 2001, this system was no longer in use. The lumpsum budgets were based on actual production in hospitals, giving financial incentives for increased production (Folmer et al., 2006; Folmer and Westerhout, 2008). The first system of lump-sum wages led to a clear lack of growth for the period 1998-2000, which suggests a lagged effect. Therefore, a dummy for lumpsum wages was added to our model for effects of the institutional change during 1998-2000. Note that this dummy does not reflect changes, but is rather a ‘level’ variable (i.e., it is not differenced in contrast to other covariates), because the first differenced data, and not the original data, showed different patterns for the period 1998-2000. Aside from the lump-sum wages, another societal determinant was used in the form of patent numbers, which was assumed to approximate medical innovation. Of course, medical innovation is widely thought to be a driver of health care utilization (Zweifel, 1999). The patent data referred to the moment of patent filing. It was assumed that the diffusion of medical technologies into hospitals does not start until patents had been granted. Typically, there is a fiveyear lag (Winnink, 2006) between patent application and grant. Thus, we chose to lag the patent numbers with five years. With the exception of supply factors and technology capacity, all variables were specified on age and gender-specific level. In contrast to the model in section (2.3), the age groups here consisted of 19 classes (18 five-year classes and one class for ages 90 and up), because this was how the data was defined by Statistics Netherlands. After differencing our data, 19  2  13=494 observations remained. 173

CHAPTER 9

Age was initially coded as dummy variables. A model with age and gender dummies was fitted, and showed that the relationship between pt and age was of a quadratic nature. To preserve degrees of freedom, we used age squared and age terms. Since age was originally defined in five-year classes, as an approximation, we used the center of each five-year class as a data point for the age terms. To assess the influence of technological innovation on the amount of number patients within specific age and gender strata, we used age term interactions with patents. We also used the age term interactions with gender, and with the dummy ‘lumpsumwages’ to adjust for additional variation. The final model was as follows: pi ,t   0   1  Genderi   2  Agei   3  Agei   4  Mortality i ,t   5  Patentsi ,t 5   6  Budget i ,t   7  Lumpsumwagesi ,t 2

  8  Agei  Genderi   9  Agei  Genderi 2

  10  Agei  Lumpsumwagesi ,t   11  Agei  Lumpsumwagesi ,t 2

  12  Agei  Patentsi ,t 5   13  Agei  Patentsi ,t 5 2

(9.5)

All continuous variables were centered to reduce the potential influence of multicollinearity. After centering, the intercept  0 reflects the autonomous growth in hospital probability (i.e., growth not captured by the fixed effects in this model) of the reference age group (47.5 years old), while the coefficients  2 and  3 for the age terms indicate how this growth may deviate for other ages. In this model, the interest mainly lies in the sign, magnitude and significance of the coefficients  5 (for the differenced patents), 12 (for the interaction between age squared and differenced patents) and 13 (for the interaction between age and differenced patents). Their interpretation allows us to infer whether a higher patent growth leads to higher growth in hospitalization probabilities, and whether this relation is different for age. Although the growth in medical technology may be influenced by the increasing demand thanks to the ageing population, medical technology was considered as an exogenous process here. This is more plausible when considering that the patents reflect worldwide technological process, which is unlikely to be affected much by the admission rates in the Netherlands, as the Netherlands are a small health care consumer, when compared to other countries. We also assumed hospital budget is a weak endogenous process (a weak endogenous process implies that the process, as a covariate, might also depend on the response variable, but that this dependence is not strong, and thus, should not lead to too different results from the case where the process was completely exogenous). Since patents were lagged,

174

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

1980

1985

1990

1995

2000

2005

2010

0.06

Alternative Practitioner

0.02

Proportion utilizing care

0.75

General Practitioner

0.60

Proportion utilizing care

Figure 9.1: Age-specific time trends in the population proportion utilizing health care, differentiated by health care type. Values are missing for 1983/1984, for the alternative practitioner care.

1980

1985

1990

1985

1990

1995

2000

2005

2010

1980

1985

1990

1995

2000

2005

2010

2005

2010

2005

2010

0.20

Physiotherapy

1980

1985

1990

Year

1995

2000

Year

Specialist 0.5

Age Category Age Category Age Category Age Category

0.3

Proportion utilizing care

2000

0.05

Proportion utilizing care

0.6 0.2

Proportion utilizing care

1990

1995 Year

Dental

1985

2010

Non-Prescribed Drugs

Year

1980

2005

0.10 0.25 0.40

Proportion utilizing drugs

0.8

Prescribed Drugs

0.5 1980

2000

Year

0.2

Proportion utilizing drugs

Year

1995

1980

1985

1990

1995

2000

2005

0-19 20-44 45-64 65+

2010

Year

Table 9.1: Regression results of time trend in proportion by age group, for seven different types of health care (n=29 for each type of health care). Time Beta

S.E. Sign.

General practitioner care

0.209 0.043

Alternative practitioner care

0.019 0.014

***

Time  Age[0-19]

Time  Age[20-44]

Time  Age[45-64]

Beta

Beta

Beta

S.E. Sign.

0.030 0.058

S.E. Sign.

S.E. Sign.

-0.311 0.058

***

-0.159 0.058

**

0.107 0.020

***

0.066 0.020

**

0.086 0.020

***

***

-0.378 0.044

***

-0.228 0.044

***

Prescribed drugs

0.707 0.031

***

-0.588 0.044

Non-prescribed drugs

0.902 0.040

***

0.097 0.054

Dental care

1.324 0.053

***

-1.239 0.073

***

-0.958 0.073

***

0.251 0.073

***

Physiotherapy

0.499 0.026

***

-0.300 0.036

***

-0.168 0.036

***

-0.090 0.036

*

0.095 0.054

-0.079 0.054

Specialist care 0.552 0.036 *** -0.598 0.047 *** -0.560 0.047 *** -0.515 0.047 *** Results are multiplied with 100, and are thus interpreted in the context of percentages, rather than fractions. The “Time” Beta coefficients report the trend for the reference group (Age 65 and over), while the Beta coefficients for the interaction terms between time and age give estimates for how much the trend in the respective age groups differs from age 65 and over. The coefficients for the intercept, and age group fixed effects are not shown here. Bolded beta terms signify the age group with the highest time trend within that particular health care sector. Respective AIC values for each health care model: -578.6516, -780.563, -607.2533, -595.3448, -525.1751, -634.3717, -625.9468. Legend: * p<0.05, ** p<0.01, *** p<0.001.

175

CHAPTER 9

and hospital budgets were allocated before the year starts, problems with endogeneity were minimalized. Furthermore, endogeneity is important if we are interested in the causal effects of technology. But here we are mainly interested in the association between the growth in technology, and the growth in specific growth trends in utilization.

9.3 Results Figure 9.1 gives the age-specific levels and trends in health care utilization probabilities, for seven health care sectors, based on the health survey data. For general practitioner care, prescribed drugs and specialist care the age group 65+ had the highest proportion of utilization during the whole period. By contrast, this group had the lowest proportion for dental care. With the exception of some trends for general practitioner, alternative medicine practitioner, and dental care, all trends seem to be positive. Table 9.1 gives the estimated time trends for the reference group 65+, and the interactions between time and the age groups. All time trends coefficients for 65+ are positive and significant, except for alternative medicine practitioner care, which is in line with what Figure 9.1 suggests. The interaction coefficients can be interpreted as the way in which the trends for the other age groups differ from the trend for 65+. For prescribed drug use, physiotherapy and specialist care, the time trends in other groups are all significantly lower than for 65+. For general practitioner care age groups 20-44 and 45-64 had significantly lower trends than 65+, although ages 0-19 did not have a significant different trend. For dental care only the age group 45-64 had a higher growth in utilization. Interestingly, 65+ had the lowest trend for alternative medicine practitioner care and the next-to-lowest for non-prescribed drugs. The evolution of the age distribution for hospital inpatient care is plotted in Figure 9.2. The distributions for both men and women are characterized by local maximums found at ages 0-4 (newborn) and ages 80-84, although women have an additional peak around 30-34 years, which can be explained by admissions related to pregnancy and childbirth. For both men and women, the figure clearly demonstrates that the peak at age 80-84 becomes steeper over time. By contrast, the level seems to remain relatively constant between 1995 and 2008 at other ages, though we observe some shift between ages 25-34 in females. In Table 9.2 regression results are presented for the time trend in hospitalization rate. The reference group here is ages 95 and over, which has a significant time trend for both men and women (0.546 and 0.330 respectively, both of which have significance p<0.001). Similar to most results from the POLS health

176

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION Figure 9.2: Evolution of the age-specific hospitalization rate, differentiated by gender (men on the left, women on the right).

Table 9.2: Regression results of time trend by age group, for the hospital inpatient care admission rates (n=400 for men and women each). Men Variable Time

Time  Age[0-4]

Time  Age[5-9]

Time  Age[10-14] Time  Age[14-19] Time  Age[20-24] Time  Age[24-29] Time  Age[30-34] Time  Age[34-39] Time  Age[40-44] Time  Age[44-49] Time  Age[50-54] Time  Age[54-59] Time  Age[60-64] Time  Age[64-69] Time  Age[70-74] Time  Age[74-79] Time  Age[80-84] Time  Age[84-89]

Women

Beta

S.E.

Sign.

Beta

S.E.

0.546

0.030

***

0.330

0.025

Sign. ***

-0.648

0.042

***

-0.366

0.035

***

-0.717

0.042

***

-0.446

0.035

***

-0.620

0.042

***

-0.382

0.035

***

-0.623

0.042

***

-0.388

0.035

***

-0.647

0.042

***

-0.471

0.035

***

-0.643

0.042

***

-0.473

0.035

***

-0.642

0.042

***

-0.276

0.035

***

-0.636

0.042

***

-0.375

0.035

***

-0.640

0.042

***

-0.487

0.035

***

-0.652

0.042

***

-0.480

0.035

***

-0.659

0.042

***

-0.436

0.035

***

-0.654

0.042

***

-0.382

0.035

***

-0.628

0.042

***

-0.342

0.035

***

-0.571

0.042

***

-0.307

0.035

***

-0.462

0.042

***

-0.238

0.035

***

-0.357

0.042

***

-0.201

0.035

***

-0.257

0.042

***

-0.162

0.035

***

-0.135

0.042

**

-0.110

0.035

**

-0.046 0.042 -0.044 0.035 Time  Age[90-94] Results are in the context of percentages. The “Time” Beta coefficients report the trend for the reference group (Age 95 and over), while the Beta coefficients for the interaction terms between time and age give estimates for how much the trend in the respective age groups differs from age 95 and over. The coefficients for the intercept, and age group fixed effects are not shown here. Bolded beta terms signify the age group with the highest time trend. Respective AIC values for men and women: 1931.581, 1733.160. Legend: * p<0.05, ** p<0.01, *** p<0.001.

177

CHAPTER 9

survey, we find that other age groups have a significant lower trend. As can be seen by the coefficients of the interactions between time and age groups, younger ages have a lower trend in hospitalization rate. Interestingly, the lowest trend for men is found for ages 5-9 (Beta=-0.717, p<0.001), while for women the trend is lowest for 40-44 (Beta=-0.487, p<0.001). Also, in accordance with Figure 9.2, we find a relatively high trend for ages 30-34. Another difference between men and women is that the variation in trend slopes is much greater within men, which suggests that supply factors (including medical technology) might have a stronger effect on demand under men. In Figure 9.3 observed levels, as well as first differences, are plotted for two relevant variables in our model that we used to explain the relationship between the growth in medical technology and the growth in hospitalization rate. For the level and the new number of filed patents, the time series are increasing for nearly the whole period observed. Between 1996 and 2004, the global hospital budget increased with over 50%, but stagnated after 2004. In fact, after 2006 a decline in budget is found. Partial autocorrelation function plots (not depicted here) showed that serial correlation was present in the level of filed patents and hospital budget, but that this serial correlation disappeared when differencing both series. Table 9.3 gives the regression results for the relationship between growth in medical technology and the growth in hospitalization probability, for the period 1996-2008. Of the variables that are not involved in interaction terms, the main effect of mortality rate, Δ(mortality rate), has an expected sign. A higher growth in mortality rate was associated with a higher growth in hospitalization probability (p<0.01). For the remaining variables that are involved in interaction terms, the main effects are interpreted in a different manner. For instance, the main effect for budget, Δ(Budget centered) is applicable for the situation when age centered assumes a value of zero (which corresponds with age of 47) and Δ(Patents centered)=0. Δ(Budget centered) is positive and significant, which is line with our expectations: a higher growth in budget is associated with a higher growth in hospitalization probability. Similarly, the main effects of lump-sum wages, gender and Δ(Patents centered) can be interpreted as the effects when age assumes a value of 47. Thus, for 47 year olds, lump-sum wages, females, and a larger growth in patents was associated with a larger growth in hospitalization probability (p<0.001). Most of the interactions between age and lump-sum wages, age and gender, as well between age and Δ(Patents centered), are significant (ranging between p<0.01 to p<0.001). This means that the growth in hospitalization probability depends on combinations of age and gender, as well as age and Δ(Patents centered). Since it is hard to interpret the regression coefficients due to the presence of age squared, the relationship of interest, between age and the growth in

178

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION Figure 9.3: Observed trends for the number of filed patents (lagged with five years; top), and for the hospital budget (in millions of euros; bottom).

2000

2004

35000 1996

2000

2004

2008

Year

Year

Hospital Budget

Hospital Budget (First Difference) 1000 500 0 -1000

-500

(Hospital Budget)

10000 9000 8000

Hospital Budget

25000

2008

12000

1996

15000

200000

300000

(Number of filed patents)

400000

Filed Patents (First Difference)

100000

Number of filed patents

Filed Patents

1996

2000 Year

2004

2008

1996

2000

2004

2008

Year

Table 9.3: Regression results of the hospital admission probability growth (n=494). Variable

Beta

S.E.

Sign.

Intercept

2.01E-03

1.93E-04

**

Female

1.46E-03

2.78E-04

***

Age centered

9.61E-05

9.05E-06

***

Age centered squared

1.63E-06

3.60E-07

***

Δ(Patents centered)

1.45E-07

1.85E-08

***

Δ(Budget centered)

1.46E-06

2.47E-07

***

Δ(Mortality rate)

2.48E-01

8.36E-02

**

Female  Age centered

-1.99E-03

3.24E-04

***

-1.43E-05

1.04E-05

Lump-sump wages  Age centered

-1.01E-06

4.37E-07

*

-6.77E-05

1.35E-05

***

-1.52E-06

5.65E-07

**

2.18E-09

6.29E-10

***

5.30E-11

2.71E-11

Lump-sum wages

Female  Age centered squared

Lump-sump wages  Age centered squared

 Age centered Δ(Patents centered)  Age centered squared

Δ(Patents centered)

Legend: * p<0.05, ** p<0.01, *** p<0.001. AIC -4508.9.

hospitalization probability, is depicted for different levels of Δ(Patents centered) in Figure 9.4. It shows that for higher levels of Δ(Patents centered), the age gradient steepens. Differences between men and women come from the steeper age profile

179

CHAPTER 9 Figure 9.4: Predicted growth in hospitalization probability against age, for different growth levels of medical patents, with differentiation to gender (men left, women right).

0.015 0.010 0.000

0.005

(Hospitalization Probability)

0.015 0.010 0.005 0.000

(Hospitalization Probability)

0.020

Women

0.020

Men

0

20

40

60

80

100

0

Age

20

40

60

80

100

Age (Patents) = 30,000 (Patents) = 35,000 (Patents) = 40,000

within men, which suggests elderly men have experienced a larger growth in hospital care utilization over time.

9.4 Discussion The aim of this paper was to (1) revisit the topic of age-specific trends in health care utilization, and estimate these trends for different health care sectors in the Netherlands and (2) investigate whether medical innovation might influence the difference in trends between age groups for the hospital inpatient care sector. We find that for prescribed drugs, physiotherapy, specialist and hospital inpatient care the trend in the probability of health care utilization is highest for ages 65 and up. For dental care, ages 65 and up have the second highest trend, behind ages 45-64. This is in line with previous Dutch findings (Polder et al., 2002) and US findings (Meara et al., 2004). However, for other health care sectors, like alternative medicine practitioner care and non-prescribed drugs, the trend was 180

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

actually lowest for ages 65+. This might be an indication that trends in health care utilization are not only driven by the demand side. Explanations could be sought from the supply side. For instance, as time passed, expenses for more and more drugs could be reimbursed under the Dutch health insurance system. The extent to which drugs are reimbursed will probably influence the utilization of these drugs. Furthermore, new drugs, as well as medical devices, might create additional health care demand. Since new drugs usually fall under prescribed drugs, the trends for non-prescribed drugs would be much less affected by new advancements. More indications that the supply side drives differences in age-specific trends can be found for the dental care. One plausible explanation for the highest trends for ages 45 and up would be the worldwide availability of fluoride toothpaste in the 1970s, which has resulted in a strong decrease of edentate patients within each generation, and subsequently, an increase in the demand for dental care. We have found first evidence for this line of thought, using a model that estimates age- and gender-specific absolute growth in the probability of hospital care utilization as a function of medical innovation and other covariates. These variables include (1) changes in allocated hospital budget (2) changes in mortality rate (3) health care organizational change in the form of lump-sum wages and (4) changes in medical innovation, which was approximated by granted medical patents. For variables (1)-(3), which are viewed as a confounder, we find plausible results. Not surprisingly, a larger growth in hospital budget was associated with a higher hospital care utilization, which means that higher spending is associated with more patients. This is in agreement with studies that find a strong association between national income and health care expenditures. Getzen (2000; 2006) has argued that health care spending is a necessity on individual level, but a luxury on macro-level, implying that spending is determined by how much a nation is capable and/or willing to spend. Mortality rate was positively associated as well, which follows previous findings that patients are more likely to be admitted in the last year of their life, then in other life years (Zweifel et al., 1999 Seshamani & Gray, 2004; Wong et al., 2011a). A fixed income for medical specialists, resulting from lump-sum financing, was negatively associated. This follows the findings of McGuire & Pauly (1991) and Rizzo & Blumenthal (1994), who gave evidence that no income incentives will lead to behavior of physicians that maximize their profits (i.e., keep production at a certain level). But the main contribution of this paper was the exploration of the relationship between changes in medical innovation and changes in the probability of age-specific hospital care utilization. We found that growth in medical innovation was indeed correlated to growth in the probability, but perhaps more interestingly, we found that this relationship was age-dependent. A larger growth in innovation led to steepening in the age profile for the probability of hospital care. This suggests that elderly benefit most from medical

181

CHAPTER 9

innovation. While this was speculated by Jones et al. (2002) and Dormont et al. (2006), we have been able to quantify this effect. The implications of this study are fivefold. First, it has implications for intergenerational differences in terms of benefits from technology. This study shows that the probability of utilizing care increases over time for all ages, but even more so for elderly. Furthermore, it shows that this gap between elderly and nonelderly becomes larger if as technology advances more quickly. Elderly benefit more from technology as more elderly are in position to be admitted in hospital to prolong their life expectancy (in good health). Thus, changes in average life expectancy can at least be partially attributed to more elderly benefiting from technology, and not just elderly benefiting from better technology. Secondly, implications of this study might also extend to the domain of solidarity in health care. It is widely known that ageing of the population could endanger the solidarity in health care. Because of the increase in elderly, large pressure will be exerted on the working non-elderly, who have to contribute more, in terms of net payments (i.e., payments minus the value of care received), to finance the rise in health care utilization that follows. The degree of medical innovation might influence the skewness in the distribution of net payments. Acceleration of medical innovation might induce higher trends for elderly, and ultimately, to greater net transfer payments between generations. Thirdly, if we –on the other hand– define annual solidarity between individuals as the extent to which net payments are distributed between individuals (irrespective of age), then this solidarity might become smaller, as the innovation might lead to health care utilization for more (new) individuals. Fourthly, medical technology ‘sells’, which means that the finding that medical technology influences health care utilization will, in itself, encourage the development of new technologies. Weisbrod (1991) argued that growth in health care expenditures might be a by-product of the complicated interaction between technology (in his case, expressed by R&D) and the insurance system. Not only is R&D of technology influenced by expected utilization, which is largely determined by the insurance system, the opposite is also true: the demand for health care insurance depends on the state of technology. Related to this is our fifth and final point. The intergenerational solidarity will probably depend on the extent to which each health care type is covered by health insurance. We found that age-specific trends differ between health care sectors. For instance, we particularly found that the elderly had higher trends than non-elderly in the probability of health care utilization for specialist and hospital inpatient care. In line with Weisbrod’s arguments, one might expect the insurance coverage as well as technology development to be the largest/fastest for these sectors. Since the costs conditional on utilization are especially high here as well, it is likely that these will influence the net payments between generations the most out of all health care sectors considered. Countries with health care insurance that are concerned about 182

MEDICAL INNOVATION AND HEALTH CARE UTILIZATION

intergenerational payments might therefore consider alternative funding principles for (parts of) hospital care, for instance based on health savings accounts. This study comes with some limitations. First, we used the probability of health care utilization as a proxy for total health care utilization. The per capita health care utilization can be seen as the product of the aforementioned probability of health utilization and the conditional amount of hospital care utilization (i.e., the amount for those who had at least some health care utilization). We decided to not model the relationship between technology and the conditional amount of hospital care utilization, as important covariates were not available for this study. For instance, changes in the treatment intensity through time, due to changes in hospital logistics and hospital practices, imply that a hospital day cannot be compared over time, thus making length of stay an imperfect indicator of utilization. The latter could be mitigated when taking into account the costs per hospital day. However, no hospital day and treatment rates were available for a length period of time, so we opted to not pursue analysis of the conditional amount of health care utilization. However, the conclusion that technology mainly leads to higher absolute growth within elderly is unlikely to be affected. According to the Dutch Costs of Illnesses study (2008), per capita hospital care costs have increased for the age groups 0-65 and 65+, from €692 to €783 and from €2799 to €2906 respectively (fixed prices 2005), over the period 2003-2005. The group 65+ actually have a bigger increase (€107 for 65+ and €91 for 0-65), in absolute terms. Since the age-specific per capita costs are equal to the product of the probability of costs and the conditional costs, and the per capita costs increase stronger over time for elderly, then it can be deduced that, if the growth in conditional costs was greater for the non-elderly, then this effect was nullified by the stronger rise in probability of utilization for the elderly. Secondly, the technology study only spanned the period 1995-2008. To investigate long-term influence of medical technology, as well as trends in this effect, a longer period might be required. Thirdly, there is a need for more detailed indicators of medical innovation over time. Although we believe that patents are a better indication of innovation than other indicators, such as R&D, we acknowledge that future research will be needed to find a better approximation of medical technology. Often though, researchers are limited by the availability of data. Fourthly, for the analysis of hospitalization rates, we left out day-care admissions, which were only registered after 1993. These day-care admissions in themselves reflect a change in practice – longer admissions are replaced by shorter admissions – which in turn might be made possible by medical advancements. However, these limitations are unlikely to change the main argument of this paper, as (a) day-care admissions show similar age-specific trend patterns; for instance, the day care admission rate for males aged 30-35 has risen from 287 to 717 admissions per 10,000 inhabitants, while for males aged 65-70 the rate has gone up from 415 to 2111 (Statistics Netherlands, 2010) and (b) this substitution probably is of a 183

CHAPTER 9

small scale (van de Vijsel, 2009; van de Vijsel et al., 2011). Finally, the medical technology results have been demonstrated for hospital care utilization, and can not be generalized for other health care sectors; particularly, for long-term care.

9.5 Conclusion In recent years a consensus has developed that not age, but medical technology will be the main driver of future HCE growth. However, medical technology might have more far-reaching consequences. The benefits of technology increase with age, and the age curve of HCE consequently becomes steeper over time. This could exert additional pressure on the solidarity between generations. Even if the age distribution as such is less relevant for aggregate health care spending, it can therefore still be of great importance for policy. Age might be a red herring in the context of future HCE, but keeping it in the center of attention for the public debate on solidarity is unmistakably ‘getting it right on the money’ in health care.

184

Chapter 10 Modeling the Distribution of Lifetime Health Care Expenditures: A Nearest Neighbor Resampling Approach ‡

To assess interpersonal payments in health care financing, often only crosssectional distributions of individual health care expenditures are available. However, for a proper understanding of solidarity issues, policy makers are often much more interested in lifetime differences because health care expenditures tend to fluctuate strongly each year within a lifetime. Actual lifetime data, however, do not exist, and even will not be available at all in the coming decades. A modeling approach is required to achieve some insight in lifetime health expenditures and their distribution. Using a longitudinal dataset, we propose to use nearest neighbor resampling to extrapolate observed panels into full life cycles. We validate this method and demonstrate how this approach can be used to obtain the distribution of lifetime health care expenditures.

10.1 Introduction Over the past few decades health care expenditures (HCE) have risen dramatically. Over the period 1981–2008, HCE have increased from 13 billion to 59 billion Euros in the Netherlands, which corresponds to an increase of more than 450% (OECD, 2010). The share of HCE in the Gross Domestic Product also increased from 7.6% to 9.9%. This trend is only expected to continue in the coming decades, given the ageing of the population. At the micro-level, the age-specific per capita HCE has steadily increased over time as well, which many believe is due to advances in medical technology (Newhouse, 1992; Zweifel et al., 1999). These This chapter is based on: Wong A, Boshuizen HC, Polder JJ. 2011. Modeling the distribution of lifetime health care expenditures: a nearest neighbor resampling approach (submitted). ‡

CHAPTER 10

trends have caused concern amongst policy makers regarding the sustainability of current health care systems. The call for an alternative health care system that not only reduces the growth in HCE, but also takes into account patterns in individual HCE, is persistent. In particular, in a recent report (Jeurissen, 2005) the Council for Public Health and Health Care advises the Dutch government to consider a form of health savings accounts, next to the mandatory health insurance system that is currently in vigour in the Netherlands. Health savings accounts are a form of health care financing, where individuals have a special account that is used to pay for health care costs. Many forms of health savings accounts already exist in the United States and Singapore. In most forms, the accounts are funded by some income premium and are exempt from taxes. The balance cannot be accessed by the individual saver until a pre-specified age (65 in the United States), i.e., it cannot be used for purposes other than health care before that age. In order to assess the feasibility of such a system, knowledge of health expenditures along an individual’s lifetime is required. Particularly, interest lies in assessing whether all individuals have a substantial amount of HCE throughout the course of their lives and comparing the spendings of the largest users of health care with those of average users. More generally, the distribution of lifetime HCE is of great interest. Interestingly, most studies in this area focus on estimating averages of individual HCE over age (e.g. Zweifel et al., 1999; Werblow et al., 2007; Wong et al., 2011). It is widely known that HCE vary immensely between individuals, as well as within individuals. Thus, average HCE patterns alone are unlikely to characterize individual patterns sufficiently well. Yet, few studies examined individual life cycles in HCE. Alemayehu and Warner (2004) used life table methods to simulate a lifetime distribution using cross-sectional data. Their methods probably underestimate lifetime variation in HCE between individuals because they do not take correlation within individuals into account. French and Jones (2004) find that the stochastic process for log health costs is well modeled as the sum of a white noise process and a highly persistent AR(1) process, which can be rewritten as an ARMA(1,1) process. Aside from an autoregressive component and a moving average component, the model accounts for heterogeneity between individuals by modeling a person-specific component. Furthermore, they control for variables such as age, gender and marital status. They use this model to estimate the effects of health cost shocks, as well as to obtain upper quantiles for the lifetime health costs. However, they do not account for the fact that life expectancy depends on the amount of health care utilization in the past (these tend to be negatively correlated; see Zweifel & Breyer, 1997). Scholz et al. (2006) use a similar AR(1) model, with a different set of covariates and without the person-specific component. This model was further expanded by De Nardi et al. (2009), who included transitions in health and allowed health care costs and survival to depend on the health status as well as on age, gender and income. Forget et al. (2008) 186

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

divided HCE into categories and used a first-order Markov model to model the transitions between categories of costs. They did not pursue higher order models, and did not check whether categorizing HCE results in a loss of precision, which could be especially large in the tails of the lifetime distribution. Finally, Ameriks et al. (2010) used a first-order Markov model for transitions between health states. Their approach captures the causal process of HCE. However, they associate each health state with a fixed amount of costs, neglecting the variance in HCE between individuals within a health state, which is likely to be large. Despite the relevance of lifetime HCE, studies on this topic are relatively scarce. An explanation for this state of things can be sought in the lack of reliable data in general, in terms of coverage (e.g. health care sectors and population groups). Furthermore, the studies that do address the topic are often bound by the limitations of their longitudinal data, as most studies only have a few years available. And that is why many of these studies are ‘parameter-driven’. For instance, French and Jones (2004) used parametric methods to estimate the tails of the lifetime costs distribution, as non-parametric approaches are generally less suited in the case of small sample sizes. The aim of this paper is to show that a non-parametric method based on nearest neighbor resampling is a suitable alternative for generating life cycles in HCE and, ultimately, for estimating the distribution of lifetime HCE. Nearest neighbor resampling can be described as a resampling method for dependent data (e.g. time series) that uses the well-known k -nearest neighbor algorithm. This method is attractive because it hardly requires assumptions regarding the marginal or joint distribution of the HCE, as well as regarding the (non-)linearity of the relationship between the response and covariates. This paper is structured as follows. In the next section we introduce the k nearest neighbor resampling algorithm for time series in general, and then a modified version of the algorithm for application to life cycles of HCE. Our k nearest neighbor approach needs several ‘tuning parameters’ that need to be chosen heuristically. We do this in section 10.3 by comparing statistics of the simulated and original data. In section 10.4 we simulate life cycles using the chosen parameters, and describe some of their characteristics. We end with a discussion of the results.

10.2 Methods and Data The k -nearest neighbor algorithm for resampling time series The nearest neighbor principle has been used for many purposes, including prediction, conditional density estimation, and the resampling of time series. Farmer and Sidorowich (1987) were one of the first to implement it for the 187

CHAPTER 10

forecasting of time series. Their algorithm has since then been used in physics (prediction of chaotic system behavior: Casdagli, 1989; Casdagli, 1991; Casdagli, 1992), hydrology (prediction of daily temperatures and rainfall: Lall and Sharma, 1996; Rajagopalan and Lall, 1999; Buishand and Brandsma, 2001), and finance (dynamics of stock markets: Hseih, 1991). The general idea behind a nearest neighbor algorithm can be described as follows. Suppose we have a realization of a long stationary or Markovian time series Y  {Yt : t  1,..., q} and we are interested in simulating realizations Z  {Z t : t  1,..., r} of Y , or, more precisely, in generating realizations which follow approximately the probabilistic law of Y . The purpose of generating such realizations could be, for example, the extrapolation or completion of individual ‘histories’, or the prediction of the population average of a functional of individual histories describing the long-term behavior of Y . For simplicity we shall consider the situation where Z corresponds to a number of future values of Y . Regard Y as consisting of overlapping ‘blocks’ of p successive observations (namely the q  p  1 blocks (Yt  p 1 ,..., Yt ) , t  p, , q  1 ), where p , the ‘history length’, is much smaller than q and has been chosen beforehand. Take the last block of Y as the first p observations of Z . To generate Z p 1 assume that k , a parameter called the number of nearest neighbours, is fixed. From all the blocks of Y preceding the last block find the k blocks that are ‘most similar’ (see below for details) to the most recent history of Z , namely ( Z 1 , , Z p ) ; each of these k most similar blocks is called a k -nearest neighbor of the most recent history of Z . Pick one (see below) of the k -nearest neighbors and assign its successor in the Y series (i.e., the observation that follows the block in the Y series corresponding to the selected neighbor) to Z p 1 . To generate Z p  2 , Z p 3 , we proceed similarly: each time we consider the most recent history of length p of the Z simulated so far, look for its k -nearest neighbors in the Y series, pick one of these nearest neighbors, and set the next value of Z equal to its successor in the Y series. (In this procedure Z is “in the making” and its most recent history of length p is constantly being updated.) Suppose the procedure stops with the simulation of Z r (r  p) . Then the first p observations of Z coincide with the last p observations of Y , and Z p 1 , Z p  2 ,  , Z r provide a simulated ‘continuation’ of Y . Although many other resampling algorithms based on the nearest neighbor principle are possible (depending on the problem and data at hand), the above description contains its essential elements and provides the basis for the presentation of our own algorithm given later on. In particular, all such algorithms require the specification of the history length or ‘lag’ p , the number of nearest neighbors k , the measure of similarity or distance between pairs of blocks which

188

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

determines the neighbors of a given block, and the method of selecting one amongst the neighbors whose successor will provide the next value of the series being simulated. To determine which blocks of Y are the nearest neighbors of a given block of Z , the Euclidean distance is often used: if y and z denote two blocks, regarded as p -vectors, the distance between them is d (y , z )  (y  z ) T (y  z ) . Sometimes, in order to account for the correlation between the components of y and to neutralize eventual differences in scale between the components of y , the Mahalanobis distance between y and z is used instead. This distance is defined

as d (y , z )  (y  z ) T  1 (y  z ) , where  is an estimate of the covariance matrix of (Y1 ,..., Y p ) (  is the sample covariance of the data matrix whose rows are the q  p  1 overlapping blocks). Now suppose that the distances between a given block of Z and its k nearest neighbors in Y have been computed. (In case there are ties in the distances, causing the number of nearest neighbors to exceed k , neighbors may be sampled randomly from among candidates at the same distance from the block.) Let z denote the block in question and let d z  (d z ,1 ,  , d z ,k ) be the k -vector of distances. To choose the neighbor whose successor is to be the successor of z we can, for example, pick the neighbor uniformly at random from among the k neighbors (each neighbor having probability 1 / k of being selected). Alternatively, we might think it more appropriate to give greater probability of being drawn to those neighbors which are closer to z , drawing the neighbor with label L according to the probabilities P ( L  l )  wl

k

w l  1

l

,

(10.1)

where wl  1 / d z ,l . The drawing of the nearest neighbor according to these probabilities will be referred to as the ‘nearest neighbor weighting’ from here on. Other weightings may of course be considered, such as the distance ranking based weighting system (e.g. Rajagopalan and Lall, 1999). The generation of the observations of Z can be seen as the successive generation of observations from the distribution of Yt conditional on the values of Yt 1 ,..., Yt  p . From this point of view, the random selection of one among k nearest neighbors is similar to the simulation of an observation from a estimate of the conditional density f Yt |Yt 1 ,...,Yt  p (based on all the blocks of Y ), which essentially describes the Markov Bootstrap (see Politis, 2003). This analogy provides some justification for the use of nearest neighbor resampling methods (in the sense that 189

CHAPTER 10

they should yield blocks of simulated time series with approximately the required pattern), since many consistency results have been proved for nearest neighbor density estimators based on stationary or Markovian time series (e.g. Yakowitz, 1993). However, the nearest neighbor draw differs from that of the density estimate in the sense that no smoothing of the distribution is involved. The value of k determines how strictly the similarity between two blocks is measured by the distance function, as it sets the volume around the p -vector being conditioned upon. Going back to the analogy with conditional density estimation, a large k provides a ‘liberal’ measure of distance and corresponds to using a very smooth and somewhat biased estimate of the conditional density to simulate from, whereas a small k imposes a more stringent measure of distance and corresponds to using a practically unbiased but somewhat ‘noisy’ density estimate. If Y is known to be Markov of a relatively small order then the correct choice of p is more or less obvious; if not, then any feasible value of p will always yield simulated series Z that are somewhat biased versions of Y . In the latter case, the bias decreases with p , but with increasing p the k nearest neighbors become more rare and heterogeneous and hence contribute more noise. Unlike the situation with the estimation of conditional densities, there are no standard methods based on theoretical (e.g. asymptotic) arguments to help us choose the number of neighbors k , the history length p , the distance function, and the probabilities (1). As far as we know, the choice of these ‘tuning parameters’ has been mostly based on heuristic criteria. Several such criteria, ranging from Generalized Cross Validation and AIC (Lall and Sharma, 1996) to marginal moments and autocorrelation (Rajagopalan and Lall, 1999), have been considered in literature. We will come to these later on when discussing the choice of our tuning parameters. Data In this study we use health insurance data provided by Vektis, an institution set up by Zorgverzekeraars Nederland, the union of all Dutch health care insurance companies. The data contain annual HCE of all those Dutch individuals who were, at some point, publicly insured during 1997-2005. These expenditures include all health care sectors (hospitals, general practitioners, drugs, and so on), with the exception of long-term care costs. In this paper we will only use the total health expenditures over all the health care sectors represented in the dataset to illustrate the application of the method. The dataset was linked to the Dutch Municipality Register by Statistics Netherlands in order to obtain additional information, such as date of death. Within the linked datasets, we made a further selection of (a) individuals who were

190

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

alive until at least the last year of the period, and (b) individuals who were fully publicly insured during the period. The resulting dataset, which we shall refer to as ‘database’ from hereon, contained 1.37 million men and 1.70 million women. A nearest neighbor resampling algorithm to simulate HCE We adapt the nearest neighbor resampling algorithm to simulate life cycles of HCE. Each of these life cycles consists of an age series Z i  {Z i ,a : a  0,..., Ai } (a series indexed by age a in years) containing the expenditures of an individual i at the individual’s age zero, one, two, etc., until death at age Ai . The database available to us, denoted by Y , is not a single, very long time series but a very large set of relatively short series of the form (Y1( j ) ,..., Y p( j )1 ) , where j is the label of individual, Yt( j ) is a vector containing the individual’s HCE (as well as other relevant information, such as age) in the t  th year of the period where our HCE data are available, and p  1 is the length of that period. In the particular application to be presented here, p  8 , the year p  1 corresponds to 2005, 1 to 1997, 2 to 1998, etc. Since the length of the Z i to be generated depends on the health of individual i over the course of his life, and since less healthy individuals generally have higher HCE and a higher mortality risk, the death of an individual must be simulated in conjunction with his HCE. Thus, besides the HCE, Z i must include an indicator of death and accordingly we write Z i ,a  ( X i ,a , Ci ,a ) , where X i ,a denotes the HCE and Ci ,a the indicator of death of individual i at age a . To simulate Z i,a conditional on the history z i  (Z i ,a 1 ,, Z i ,a  p ) we must have Ci ,a 1  0 ; otherwise the simulated history Z i has terminated at age a  1 . Assuming Ci ,a 1  0 , we determine the k nearest neighbors of z i in the subset of coevals of the individual i , namely in the subset of the database Y which contains all individuals who are a years old at year p  1 ; call this subset Y (a ) . The nearest neighbors are determined in terms of HCE alone, namely by computing the distances between x i  ( X i ,a 1 ,, X i ,a  p ) , the expenditure components of z i , and the expenditure components of all the series in Y (a ) . Next, we pick one from the k neighbors and then assign to Z i,a its expenditure and death status (since HCE are measured annually it is possible for an individual to have positive HCE at age a even if he dies at that age). Note that every Z i,a simulated in this way is drawn from the year p  1 , since each time z i  (Z i ,a 1 ,, Z i ,a  p ) is matched to some neighbor (Y1( j ) ,..., Y p( j ) ) in Y whose successor is Y p( j )1 . This implies that the simulated life cycles describe the

evolution of health care expenditures over an individual’s age under a fixed 191

CHAPTER 10

calendar time. In our application this fixed time is the year 2005. Therefore the simulated life cycles reflect how the HCE of the newborns in 2005 should evolve over the course of their lives if health care price and volume levels and mortality rates remained fixed at the levels of 2005. Given the observed trends in health care expenditures and mortality, this ‘stationary reality’ is probably not realistic. Trends in health care price and volume levels can be applied post-hoc to the simulated life cycles, however. In the remainder of this article we will restrict ourselves to simulation of life cycles and not discuss the trend scenarios. We will return to possible advantages and disadvantages of this aspect of our algorithm in the discussion. So far, we have defined the generation of Z i,a conditional on z i  (Z i ,a 1 ,, Z i ,a  p ) ; it remains to define the generation of the first p observations of Z i . To simulate the initial state of individual i , Z i , 0 , we draw an individual randomly from Y ( 0 ) , the subset of Y consisting of the newborns at year p  1 , and the expenditure and death status of the sampled individual are assigned to Z i , 0 . If the individual dies at age 0, its history terminates. Otherwise, we simulate Z i ,1 by conditioning on Z i , 0 , namely by drawing k nearest neighbors from Y (1) (the one-year olds at year p  1 in Y ), computing distances in terms of the expenditure component X i , 0 , and assigning to Z i ,1 the expenditure and death status of a chosen neighbor. And so on: the generation of Z i ,1 ,, Z i , p follows the method used to simulate Z i ,a from (Z i ,a 1 ,, Z i ,a  p ) for a  p , except that the conditioning is on Z i ,1 , (Z i , 2 , Z i ,1 ) , …, (Z i , p 1 ,, Z i , 2 , Z i ,1 ) , respectively. Note that we may also initiate a simulated life cycle by using the complete history of a p -year old randomly sampled from the year p  1 (an eight-year old from 2005 in our application) as initial states and then simulating the rest of the cycle iteratively by conditioning on the previous p states, as explained above. This will probably increase the realism of the simulated life cycles. However, in order to protect the identity of the individuals as much as possible, in our application we have not used the complete history available. That is why it was necessary to define the generation of the first p states of Z i in a special way. Our application has two other particularities. Because there are important differences in longevity and in age/HCE patterns (e.g. related to pregnancy or breast cancer) between the sexes, we generate histories separately for men and women. And because the data on ages 95 and over are sparse, individuals in this age group are pooled together and treated as if they had exactly the same age. The history length used in our application is p  8 . Since our dataset spans the period 1997-2005 and all successors are sampled from 2005, p could have been chosen between 1 and 8 . In our validation study (which is described later on), the 192

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

bigger value turned out to give more realistic results. However, this validation was based on predicting one successor (i.e., one year ahead), and does not provide insight into how well the nearest neighbor algorithm performs when making successive predictions that represent a longer period. Nor could our eventual simulated life cycles be useful for validation, as there are no actual life cycle data to compare with. To provide some justification for the use of our algorithm, we set up a simple simulation study. In this study a dataset was simulated by drawing 10,000 observations randomly from a multivariate normal distribution with a 3  3 covariance matrix that was chosen beforehand (Table 10.A1 in Appendix 10.A). The resulting 10000  3 dataset has a similar structure to our real database, with each row representing an individual, and each individual having three observations over time. The nearest neighbor resampling algorithm was then used to simulate a continuation for each individual in this dataset. Since each individual has three observations, of which the last is used to predict the successor, we use a history length p  2 . Table 10.A2 shows the correlation matrix of the simulated continuation for six time points ahead. We found that the correlation structure as used to generate the dataset is preserved in the simulated continuation. Furthermore, the marginal distribution at each time follows the standard normal distribution very closely (Figure 10.A1 in Appendix 10.A). Thus, in this simple example, nearest neighbor resampling performs well over a long period (even for a relatively short history length), which does not give us reason to believe that our method would perform poorly in a different context. Our algorithm differs from those presented in previous studies (e.g. Lall and Sharma, 1996; Rajagopalan and Lall, 1999) in some important aspects: (1) rather than simulating one large series, our algorithm simulates multiple (relatively short) series, one for each individual; (2) the length of the simulated series is regarded as a random variable which depends on the values within a series (death depends on past HCE); (3) since HCE generally increase with age, the algorithm generates a non-stationary ‘age series’. The last aspect is particularly important since previous studies have been mainly concerned with the generation of stationary series. On the other hand, HCE data features both age and time (or period) effects, and we did transform the original dataset in order to make it more stationary regarding trends in prices and volumes before applying the algorithm to it, as explained in the next section. Detrending the data As explained in the previous section, the database from which the life cycles can be simulated is tied to the 1997-2005 period, and the simulated life cycles simulated by our algorithm will always consist of observations from 2005. If the algorithm is

193

CHAPTER 10

applied to the original data the simulated life cycles will necessarily be constrained by the price and volume levels of 2005 and by eventual trends in HCE during 1997-2005. Price inflation, advances in medical technology and changes in insurance coverage have probably contributed to create such trends, and neither the trends nor the context of 2005 can be expected to apply throughout an individual’s life time. An amount of HCE incurred in 1997 may represent a level of health care utilization (and potentially a different state of health) that is different to the situation when that same amount is incurred in 2005. This raises the question of what exactly the life cycles simulated by our algorithm from the original data would represent. Instead of trying to answer this question we shall try to make the dataset stationary with respect to HCE by removing their annual trend. Essentially, we use age-specific HCE increments over time to convert the past expenditures to the level of 2005. Let W j ,a ,t* be the expenditures of an a -year old individual j at a past calendar year t * (1997  t *  2004 in our database), and let Wa ,t * be the mean over all HCE in the year t * . Then we define the increment of an a -year old individual j at a past year t * as the relative increase in the average HCE from t * to 2005, i.e., Wa , 2005 / Wa ,t * . We then scale the expenditures of

individual j at t * to the level of 2005 by multiplying his expenditures with the estimated increment. Since the annual trend in HCE is positive, the increments decrease with the calendar year t * . The scaling is applied separately for men and women. However, the above description ignores the fact that we are searching for nearest neighbors based on histories of expenditures up until age a . This implies that both simulated individuals and individuals in our database remained alive throughout this history. For the estimation of Wa ,t * and Wa , 2005 we must therefore restrict ourselves to individuals that remain alive at t * and 2005 respectively. Note that we already excluded individuals who die at a past year t * in a previous selection; see the section on the data. An additional exclusion of the individuals that pass away in 2005 is required to compute Wa , 2005 , however. This correction is necessary in our opinion as health care expenditures are much higher in the last year of life (e.g. Zweifel et al., 1999).

10.3 Validation of the nearest neighbor resampling application Setting up the validation study Prior to simulating life cycles, we validated the k -nearest neighbor algorithm using the database. In essence, we made a prediction for the HCE and death status of 194

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

every individual in the database, by searching for a nearest neighbor within that same database. This prediction was made for the final year in our database (2005) and was then compared to the observed HCE and death status for each individual, using several statistics. By changing k , p , the method of weighting of neighbors and the choice of distance function, we can observe how well the nearest neighbor algorithm performs under a variety of conditions, allowing us to make a choice for each tuning parameter that leads to the best results under the investigated conditions. The validation was done separately for each age and gender stratum. First, prior to the validation we chose a value for k , a value for p , the method of weighting of neighbors and the type of distance function. Define w i  (Wi , 20051 ,,Wi , 2005 p ) as the vector of past expenditures for an individual i that were incurred at calendar year 2004,...,2005  p respectively. For instance, if p  4 , the vector only contains expenditures of the years 2001-2004. For any given individual i within a stratum, we searched for k nearest neighbors based on his expenditures history w i , drew one of the nearest neighbors (who is found at row j , with j  i ) , and then assigned the expenditures X j , 2005 and the death indicator C j , 2005 to individual i . This prediction process was repeated for all individuals in the dataset. Note that no detrending of data was needed here. While the past expenditures during the period [2005  p,2005  1] exhibit a calendar year trend, this trend was present in the history of expenditures for both individual i and nearest neighbor j . As mentioned previously, other studies have compared the observed and predicted values in various ways. No golden standard exists here, but we provide some intuitive argumentation. The Generalized Cross Validation criterion and Root Mean Squared Error are mainly used in a cross validation context, where the parameter selection is based on the optimal tradeoff between bias and variance (Hastie et al., 2009). However, some have also used it in a nearest neighbor resampling context (e.g. Lall and Sharma, 1996). The key thing to realize is that these statistics are used in a mean prediction error setting. Here, our objective is not to predict means but rather individual values and ultimately, the lifetime distribution of individual values. Thus, a large prediction error may be somewhat misleading. Consider the following simplistic example. Suppose we have two time series, Y : {Yt | t  1,... p  1} and Z : {Z t | t  1,... p  1} . Furthermore, suppose they are mutual nearest neighbors, based on their respective histories from time 1 to p . Even though these neighbors may be very similar, Y p 1 and Z p 1 can differ greatly,

if the serial correlation in Y and Z is relatively weak. This is exactly the case with HCE, as we will see later on in the validation. There is a large variance in annual HCE, which can be explained by the presence of health shocks: an individual might become ill at any point of time, even if the recent history of health suggests

195

CHAPTER 10

it is unlikely. For HCE the prediction error would be high and perhaps suggest a poor performance. Yet, the distribution of summed (‘lifetime’) costs is similar, provided that the distance between these two neighbors is small. Since we are interested in producing realistic age series of expenditures, it makes sense to look at statistics that describe some key properties of the original series. For HCE, we consider two criterions. The first criterion is the degree of similarity between the marginal distribution of the predicted and observed expenditures in 2005. If the distribution of predicted expenditures deviates strongly from the actual distribution, then we might expect this error to propagate and amplify by the iterative application of the nearest neighbor algorithm. The distribution comparison can be done by Q-Q plots, or alternatively, by goodness-of-fit statistics. We used the two-sample Cramér-Von Mises (Anderson, 1962) and Anderson-Darling (Pettitt, 1976) statistics to assess the similarity between two cumulative distribution functions G and F . These statistics are of the form 

d (G, F ) 

 w(G  F ) dF 2

(10.2)



where w is a positive weight function. If w  1 , this statistic is the Cramér-Von 1 Mises criterion; if w  F 1  F  it is the Anderson-Darling statistic. The Anderson-Darling statistic places more weight on the tails of the distribution, which are of specific interest given the skewed and heavy-tailed distribution of HCE. Note that we do not carry out these statistics to test whether G and F are equal or not but simply to determine which tuning parameter choices lead to the largest degree of similarity between G and F . In line with the Anderson-Darling and Cramér-Von Mises statistics, some statistics related to the first four moments of the distribution can be used for comparison (i.e., mean, variance, skewness and kurtosis). The second criterion involves the correlation of subsequent observations, which should determine how the series behave over a longer period. For death status, we investigated the marginal mean of the death status (i.e., whether there are differences between the observed and predicted death rate in 2005). Furthermore, we also explored the correlation between the death status in 2005 and past expenditures. Choosing the distance measure We found that under many combinations of distance measure, k , weighting of nearest neighbors, and p , there was little difference between the results obtained with the Euclidean distance and those obtained with the Mahalanobis distance (results not shown here), as elements of the feature vector have a similar scale. For this reason we used the simplest (and fastest) measure, the Euclidean distance. 196

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

Choosing k and neighbor weighting Choosing k in the nearest neighbor setting indirectly determines the cutoff value of the maximum distance at which an object will be considered as a nearest neighbor. If k is large, a bias may occur because less-similar nearest neighbors are also considered. It can be expected that k  1 yields the best results in mimicking the original data, as it picks the neighbor that is closest to the individual of interest. Table 10.1 shows some simulation results for one stratum (men aged 60) that confirms this. We compared several statistics for different values of k . For k  1 we found that nearest neighbor weighting, as in equation (1), yielded better results than uniform weighting (results not shown here). Table 10.1 shows that k  1 leads to (1) a serial correlation that closely follows the original data and (2) a predicted marginal distribution that deviates as little as possible from the actual marginal distribution, based on the Anderson-Darling and Cramér-Von Mises statistics. Q-Q plots further confirmed this (see Figure 10.1). For some age groups, the upper tails (98th quantiles and above) were not as well reproduced for any k , which is likely explained by the lack of similar neighbors to the observations in the upper tails. No large differences were found for the serial correlations. The tradeoff for the good performance of k  1 is that it will yield a deterministic series in the sense that, for a given initial state in the age series, the next state will always be the same, if there is a unique nearest neighbor. To avoid this, we picked the next-best performing value for k ( k  2 ). Choosing the lag order In our view, there is no real argument to a priori assume that a lag order of one works sufficiently well. Quite the contrary, there is ample evidence that some individuals are chronically ill and require a continuous use of health care. This group is probably underrepresented by choosing a lag of one: the probability of returning to low expenditures, given high expenditures in each previous year, is larger when we consider a shorter history. Ideally, we would use as much long-term information as possible, which means the lag order should be chosen as large as possible, i.e., as large as the data permits. To add weight to this argument, consider the following extreme example. Suppose we have a very large dataset where only a few observations in the last year of life are missing. It would then be very inefficient to use a lag order of one. Because the dataset is so large and contains a wide range of permutations in individual histories, it seems much more realistic to take the complete histories and use these to predict the few missing observations for each individual. Of course, we must take into account that in reality the number of observations is limited, and does not necessarily cover all permutations well. This can be intuitively explained as follows. When taking a large feature vector (i.e., including many covariates in the distance measure), the number of possible realizations increases exponentially. 197

CHAPTER 10 Table 10.1: Number of nearest neighbors validation for men aged 60 (n=15,015), using the Euclidean distance, nearest neighbor weighting and lag order p  8 . Expenditures in Euros. Variable Type

Number of nearest neighbors k Data

k=1

k=2

k=3

k=4

k=5

Moment-related statistics Mean Standard Deviation Skewness Kurtosis

2532 7578 10.63 176.68

2549 7798 10.77 179.07

2468 7364 10.64 182.26

2536 7390 9.75 151.44

2421 7098 10.25 168.6

2383 6791 10.27 177.08

Quantiles 0.1 0.5 1 5 10 25 50 75 90 95 99 99.5 99.9

1 3 4 7 22 118 636 1.891 5576 11209 31760 42592 108091

1 3 4 7 22 121 641 1845 5521 11237 33257 42749 107134

1 4 4 7 22 121 645 1825 5316 10685 32760 42387 105647

0 4 4 7 19 120 649 1854 5483 11426 33478 46010 103665

1 4 4 7 21 122 635 1810 5235 10493 31177 42308 91012

1 4 4 7 23 122 637 1825 5205 10524 30282 41095 90045

-

1.761 0.070

2.291 0.120

2.436 0.083

2.935 0.169

2.341 0.146

0.363 0.268 0.228 0.197 0.164 0.188 0.126 0.153

0.371 0.297 0.216 0.220 0.173 0.191 0.153 0.170

0.365 0.293 0.236 0.215 0.185 0.225 0.143 0.150

0.341 0.265 0.205 0.182 0.146 0.162 0.115 0.104

0.298 0.238 0.20 0.196 0.135 0.161 0.093 0.102

0.301 0.252 0.165 0.168 0.123 0.152 0.095 0.112

Moment-related statistics Mean

0.014

0.014

0.014

0.014

0.014

0.014

Correlation with lagged HCE 1 2 3 4 5 6 7 8

0.214 0.135 0.058 0.062 0.067 0.035 0.060 0.037

0.188 0.106 0.049 0.045 0.038 0.036 0.045 0.018

0.195 0.108 0.069 0.040 0.036 0.054 0.049 0.027

0.197 0.129 0.058 0.057 0.060 0.053 0.045 0.032

0.224 0.145 0.057 0.099 0.082 0.052 0.052 0.045

0.240 0.144 0.061 0.060 0.078 0.070 0.066 0.044

Measure Health Care Expenditures

Goodness-of-Fit Anderson-Darling Cramér-von Mises Serial Correlation at lag 1 2 3 4 5 6 7 8 Mortality

198

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES Figure 10.1: Q-Q plot for the simulated (predicted) versus observed data of 60-year old men, as part of the validation study. The predictions pertain to parameter choices of k  2 and p  8 .

For instance, if we approximate the expenditures with v distinct categories, then the number of possible permutations would be v p , where p is the lag order. That means that the probability of finding a meaningful neighbor, i.e., finding a neighbor with a reasonably small distance, becomes increasingly smaller with p . This can even happen in dimensions as low as 10-15 (Beyer et al., 1999). By matching an object with neighbors that are in fact not similar can yield a bias. Thus, we have to find a balance between keeping this bias minimized and retaining as much longterm information as possible. Table 10.2 shows the validation results for varying values of p , for one age group. Although in absolute terms p  1 seems to perform the best in reproducing the distribution (based on the lowest values for the Cramér-von Mises and Anderson-Darling test), the differences are fairly small. Even if we do not formally test here, note that the Cramér-Von Mises and Anderson-Darling critical values for   0 .01 and n   are 0.743 and 3.857, respectively. Under all lag choices, both statistics are well below these critical values, suggesting a great degree of similarity between the predicted and actual marginal distribution. This was further confirmed by inspecting the Q-Q plots. Furthermore, we observed that the serial correlation for larger lags are better reproduced for a larger order, though the difference after p  2 seems to be small. In line with our previous argument, we favor the largest lag order ( p  8 ) to retain as much long-term properties as possible, without causing considerable bias in the short-term. Other age and gender groups showed similar results.

199

CHAPTER 10 Table 10.2: Lag order validation for men aged 60 (n=15,015), using the Euclidean distance, nearest neighbor weighting and k  2 . Expenditures in Euros. Variable Type

Nearest Neighbors with lag p Data

p=1

p=2

p=4

p=6

p=8

Moment-related statistics Mean Standard Deviation Skewness Kurtosis

2532 7578 10.63 176.68

2508 7140 9.62 148.41

2549 7816 11.36 199.75

2558 7460 8.99 118.71

2570 7567 10.28 169.35

2468 7364 10.64 182.26

Quantiles 0.1 0.5 1 5 10 25 50 75 90 95 99 99.5 99.9

1 3 4 7 22 118 636 1891 5576 11209 31760 42592 108091

1 4 4 7 23 120 627 1876 5625 11358 31197 41897 105622

0 4 4 7 22 116 645 1888 5662 11205 31016 44497 110019

1 4 4 7 23 123 632 1896 5545 11388 32353 46520 105622

1 4 4 7 21 123 640 1879 5602 11682 32031 43196 107113

1 4 4 7 22 121 645 1825 5316 10685 32760 42387 105647

-

1.792 0.045

1.769 0.050

1.873 0.051

2.245 0.107

2.291 0.120

0.363 0.268 0.228 0.197 0.164 0.188 0.126 0.153

0.356 0.173 0.114 0.109 0.099 0.113 0.077 0.091

0.354 0.259 0.208 0.186 0.161 0.191 0.132 0.130

0.368 0.288 0.253 0.242 0.196 0.178 0.153 0.146

0.346 0.286 0.234 0.230 0.191 0.222 0.138 0.143

0.365 0.293 0.236 0.215 0.185 0.225 0.143 0.150

Moment-related statistics Mean

0.014

0.014

0.013

0.013

0.014

0.014

Correlation with HCE at lag 1 2 3 4 5 6 7 8

0.214 0.135 0.058 0.062 0.067 0.035 0.060 0.037

0.228 0.091 0.057 0.062 0.055 0.045 0.058 0.058

0.232 0.115 0.050 0.054 0.041 0.027 0.039 0.034

0.168 0.101 0.056 0.051 0.017 0.016 0.015 0.034

0.183 0.122 0.066 0.060 0.039 0.050 0.030 0.039

0.195 0.108 0.069 0.040 0.036 0.054 0.049 0.027

Measure Health Care Expenditures

Goodness-of-Fit Anderson-Darling Cramér-von Mises Serial Correlation at lag 1 2 3 4 5 6 7 8 Mortality

200

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

10.4 Simulation of life cycles After detrending the data, we simulated 10,000 life cycles for men and women separately. As mentioned before, we present the life cycles describing the evolution of health care expenditures over age assuming the price and volume level stays fixed at the level of 2005, and do not adjust these HCE for trends. Figure 10.2 portrays some simulated life cycles for men. These life cycles are characterized by many years with low expenditures, and a few years with high expenditures. The age at which these health shocks occur can differ strongly between individuals. Furthermore, the level and persistence of these health shocks are also different for each individual. Table 10.3 gives a comparison between cross-sectional and simulated life cycle statistics for men. In 2005 the cross-sectional mean HCE for men was around 1,726 Euros (irrespective of age). The other moment-related statistics and quantiles indicate that the expenditures have a large variance, are heavily skewed to right and are heavy-tailed. About 1% of the insured population incurs more than 53,000 Euros in 2005. Given the lack of available data, policy makers often consider this cross-sectional distribution to be representative of the distribution of lifetime HCE, but our results provide a more nuanced view. We find that the mean lifetime acute HCE for men amount to nearly €99,000. Although the lifetime distribution is still highly skewed to the right, we see that the skewness is 2.55 here, compared to 12.52 for the cross-sectional distribution. About 1% of the simulated individuals incur more than 350,000 Euros during their lifetime for acute health care. On the other hand, 1% of the individuals have no more expenditures than 7,200 Euros. When we look at specific age intervals, we find that the average expenditures more or less follow a U-shape (expenditures are high for newborns, are followed by lower HCE for the middle ages, and are high again for the highest ages). As the Coefficient of Variation shows, the variation is relatively largest for the 45-64 age group. Skewness and kurtosis was found to decrease with age. The Gini coefficient is lowest for 65+, which implies that the inequality in HCE is lowest in this age group. It is even lower for lifetime expenditures, which is likely explained by the high probability of a transition from low acute health care utilization to high acute health care utilization at some point during lifetime. When we look at the average expenditures per life year (i.e., the total lifetime HCE divided by the life span of the individual; referred to as ‘until death’ in Table 10.3), the estimated Gini coefficient is smaller than the coefficient for lifetime, but the magnitude of this difference is small. The Gini coefficient for the average expenditures per life year can be interpreted as the inequality in health care expenditures that is adjusted for differences in longevity between individuals. In other words, differences in life expectancy seem to have a relatively small influence on the inequality in lifetime HCE. 201

CHAPTER 10

20000 40000 0

20 40 60 80

0 0

20000 40000

0

20000 40000

0 0

20000 40000

0

20000 40000

0 20000 40000 0 20000 40000 0

20 40 60 80

20 40 60 80

0

20 40 60 80

0

20 40 60 80

20000 40000

20 40 60 80

0

20000 40000 0

0

0

0

20 40 60 80

20 40 60 80

20 40 60 80

20 40 60 80

20000 40000

0

0

0

0

0

20 40 60 80

20 40 60 80

20 40 60 80

20000 40000

0

0

0

0

20 40 60 80

20 40 60 80

20000 40000

0

0

20000 40000

20 40 60 80

20000 40000

0

0

20000 40000 0

0

20000 40000

Figure 10.2: Sixteen examples of simulated life cycles for men: in each panel acute health care expenditures (in Euros) are plotted against age.

These findings seem to be in line with Lubitz et al. (1995) who find that while individuals with a longer life span spend more on health care as seen over a lifetime, the costs associated with longevity gains are not as high as one would expect. Finally, the simulated life span has similar moment-related statistics to those of the deceased individuals in 2005. However, we found that our simulated individuals lived slightly shorter on average. Figure 10.3 shows the histograms for the simulated lifetime expenditures for men and women. The distribution for women is similar to the distribution for men, but the quantiles for women are generally higher. Furthermore, the distribution is less skewed for women.

202

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES Table 10.3: Comparison between simulated lifetime statistics (n=10,000), and observed cross-sectional statistics (2005) for men. Simulated statistics are presented for the case that price and volume levels remain fixed at the level of 2005. Expenditures are presented in Euros. Variable Type

Data

Simulated Lifetime Statistics Age Age 25-44 45-64

2005

Until death

Age 0-24

Age 65+

Per life year

1.37M

10000

10000

9936

9701

8129

10000

Moment-related statistics Mean Stand. Dev. Skewness Kurtosis

1726 4784 12.52 364.95

98853 70919 2.55 13.95

16251 22294 9.35 155.46

11423 18918 6.36 69.85

29480 41339 4.39 35.54

52470 43240 4.46 54.82

1327 1056 6.16 111.24

Quantiles 0.1 0.5 1 5 10 25 50 75 90 95 99 99.5 99.9

0 1 4 10 25 90 387 1398 4178 7693 20292 27657 53722

4024 7152 10051 21185 30859 52868 83753 126619 179283 222886 350705 412892 613931

1456 1864 2164 3426 4359 6634 11021 18315 29661 41776 101643 142308 253952

148 358 466 859 1228 2337 5450 13059 26859 40235 83785 126721 223119

36 254 455 1243 2082 5723 15962 37113 70383 100992 202922 256192 374978

15 114 318 5579 11205 24424 44464 71319 99787 120292 184405 215418 397675

104 153 184 337 457 715 1101 1649 2356 2988 4950 6138 11105

Variation (between and within) Coefficient of Variation Gini

2.772 0.816

0.717 0.352

1.372 0.463

1.656 0.612

1.402 0.574

0.824 0.396

0.796 0.350

-

0.126 0.316 0.438 0.146 0.206 0.218

-

-

-

-

-

Moment-related statistics Mean 74.73 73.07 Stand. Dev. 13.17 13.03 Skewness -1.10 -1.45 Kurtosis 2.041 2.96 Abbreviations: M., million; Stand. Dev., Standard Deviation.

-

-

-

-

-

Measure Sample Size N

Health Care Expenditures

Share of q last life years Mean for q=1 Mean for q=5 Mean for q=10 St. Dev. for q=1 St. Dev. for q=5 St. Dev. for q=10

Life Span

203

CHAPTER 10 Figure 10.3: Histograms of simulated lifetime acute health care expenditures (in Euros) for men and women, under the assumption that price and volume levels remain fixed at the level of 2005.

30 0 10

Frequency

50

Men

0e+00

1e+05

2e+05

3e+05

4e+05

5e+05

6e+05

5e+05

6e+05

Simulated lifetime costs

30 10 0

Frequency

50

Women

0e+00

1e+05

2e+05

3e+05

4e+05

Simulated lifetime costs

In Figure 10.4 a smoothed density of acute HCE is shown for various cutoffs (25, 45, 65-years old and death). For both men and women, the distribution of HCE becomes less skewed with the length of the age span. Nevertheless, even though the lifetime distribution may appear to be more evenly distributed than when considering a smaller time frame, it is still right-skewed and heavy-tailed.

10.5 Discussion In this paper the nearest neighbor resampling algorithm was adapted and modified to extrapolate short panels into full life cycles in terms of acute HCE. In our application the resulting life cycles represent how the acute HCE of a cohort of newborns should evolve over lifetime, assuming that the price, volume and mortality rates remain fixed at the levels of the year in which the infants were born. 204

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES Figure 10.4: Kernel density estimates of the distribution of acute health care expenditures over lifetime until various age cut off points (age 25, age 44, age 65 and death), for men (solid lines) and women (dashed lines). Expenditures are presented in Euros.

6e-05

Until age 45

2e+05

4e+05

4e-05 2e-05

6e+05

0e+00

2e+05

4e+05

Simulated lifetime costs

Until age 65

Until death 6e-05

Simulated lifetime costs

4e-05

men women

0e+00

0e+00

2e-05

Density

4e-05

men women

6e+05

2e-05

6e-05

0e+00

Density

men women

0e+00

2e-05

Density

4e-05

men women

0e+00

Density

6e-05

Until age 25

0e+00

2e+05

4e+05

Simulated lifetime costs

6e+05

0e+00

2e+05

4e+05

6e+05

Simulated lifetime costs

This method assumes that the current health care expenditures and death status depend on health care expenditures that have been incurred in the past. In the simulation of life cycles the following aspects of the algorithm are important: (1) Making the data stationary in terms of time trends such that the HCE reflect the year in which the infants were born; (2) choosing the ‘tuning parameters’: the number of nearest neighbors that are considered for prediction [k ] , the type of distance function which is used to determine which observations are a nearest neighbor, the length of the history of expenditures as used in the distance function [ p ] , and the nearest neighbor weighting method which determines the probability of drawing each nearest neighbor. In our application, we found that k  2 , p  8 , the Euclidean distance and inverse distance based weighting gave good results; (3)

205

CHAPTER 10

initiating a life cycle by drawing randomly an observation from newborns and assigning the HCE and a death status of this observation; (4) for subsequent ages: employing a nearest neighbor search and drawing one of the nearest neighbors under the choices for the tuning parameters; (5) predicting the successor values (i.e., HCE and death at the next year of life) by assigning the HCE and death status of the nearest neighbor; (6) terminating a life cycle once a nearest neighbor with ‘death’ is drawn. We validated this algorithm by means of a study in which we predicted the HCE and death status for each individual in a given year and compared these predictions to the values that were observed in the original data. We found that the method was able to reproduce both the marginal distribution in this year and the serial correlation with previous years from the original data rather well, suggesting that the method at least performs well in the short-term. Although we could not test the long-term performance –the extrapolated life cycles cannot be used to assess this performance, as we have no full life cycles to compare with– we did investigate this by setting up an additional simulation study, in which a similarstructured dataset (i.e., multiple short series) was generated according to a known process, and nearest neighbor resampling was used to make a continuation for each of these series. In this example we found that the long-term performance was equally satisfying, which provides some justification for our method. Using our simulated life cycles a distribution of lifetime acute HCE can be obtained which may help policy makers in their decisions regarding the financing of health care. The life cycles reflect the price and volume level of 2005. Although not presented in this paper, a time trend on these HCE can be applied to obtain a more realistic distribution (as it is unlikely that HCE remain stationary). As an alternative we could have included the trends in the simulation process. As opposed to removing the trend in the database by scaling the observations to the level of one fixed year (2005 in our application), the observations could have also been scaled to the year at which an individual is being simulated; this implies that a separate scaling is applied for each life year. This will mainly affect the nearest neighbor search process. Although it is difficult to assess which method is better, we believe that it is more practical to keep the simulated life cycles stationary at one fixed year. In this case we can easily apply a post-hoc trend correction and examine several trend scenarios, as opposed to the situation where a separate (time-consuming) simulation is required for each trend scenario. This is particularly important as trend estimates, especially when based on a short period, can be divergent in nature. Furthermore, trend scenarios are also complicated by the existence of age-specific trends. We found in our data that trends in HCE were higher for elderly. An explanation for this would be that elderly benefit more from innovations in medical technology (Wong et al., 2012). However, note that there is no trivial way to account for trends in life expectancy in a post-hoc correction, as 206

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

well as in the simulation process itself. This remains a limitation of our method and is a topic for future research. Our method could also be applied to assess the introduction of a new health system and its effects on the spending on health care. For instance, our data covered the period 1997-2005, in which there were two types of insurance (of which only the public insurance is reflected by our data). In 2006 a new health insurance system was introduced. By simulating the continuation of life cycles under the old system, this ‘parallel history’ could be compared with the actual history of HCE in that period (assuming that data for this period was also available). Our nearest neighbor resampling algorithm makes a contribution to existing versions of the algorithm, as it differs in a few regards. First, we simulated multiple short ‘age series’ –one for each individual– rather than one long series. Secondly, the simulated ‘age series’ were not stationary in the sense that acute HCE evolve over age (though our series were stationary in terms of time trends). But perhaps most importantly, the length of each series was regarded as a random variable that depends on previously incurred health care expenditures, as empirical studies have pointed out that the level of health care expenditures and death are positively correlated. It also contributes to current literature on the dynamics of lifetime health care expenditures. We have proposed our method as an attractive alternative to some more common approaches. The latter mainly involve (I) parametric time series models where an autoregressive model is fitted and then used to simulate time series (e.g. French and Jones, 2004; de Nardi et al., 2010), (II) Markov-chain based models, where the current state of HCE depends on the previous state(s) (e.g. Forget et al., 2008; Ameriks et al., 2010). Nearest neighbor resampling falls under the category of non-parametric resampling methods. Many resampling methods exist, such as residual based bootstrapping, wild and local bootstrapping (see Buhlmann, 2002; Politis, 2003; Davison et al., 2003; for an overview of methods to generate realizations for dependent data). A special case of the local bootstrap is the Markov Bootstrap, where a so-called transition density is estimated, which assumes that the entire distribution of successor values is a smooth function of the last p values. The smoothing is often accomplished by means of kernel density estimation. Nearest neighbor resampling is similar to the Markov Bootstrap, but it does not explicitly estimate a density conditional on the last p values, but rather simulates the process of drawing an observation from this transition density. Furthermore, it does not use any smoothing of sorts, but draws directly from the original data. One advantage in this regard over the Markov Bootstrap is its simplicity, as the step of explicitly estimating a transition density is skipped. Nearest neighbor resampling also is characterized by its highly intuitive appeal. The Markov Bootstrap is mostly used for resampling one particular long time series. In 207

CHAPTER 10

our application we only have multiple short time series (one for each individual) to our disposal. Nearest neighbor resampling provides an intuitive way to ‘borrow’ information from other time series to predict a realistic continuation for each of these short series. The main strengths of the nearest neighbors resampling algorithm is that it is essentially model-free and therefore involves fewer assumptions and restrictions than some other approaches. These include: (1) no response distribution assumption. HCE are highly skewed, and require possibly a more complicated distribution in parametric settings (e.g. Generalized Gamma distribution, see Manning et al., 2005); (2) no homogeneity assumption for the distribution. The variance, skewness and kurtosis of HCE differed with age in our data; (3) no assumption on the nature of dependence between the response and covariates. In parametric models, the response variable often depends on covariates linearly, but this may lead to poor results in some cases; (4) no restrictions on the lag order/history length. Autoregressive models are often restricted to first order serial correlation to reduce the model complexity, but the nearest neighbor resampling algorithm easily allows for higher orders, as long as the data permits it. In our experience, HCE requires at least a second order simulation for some realistic results; (5) less restrictions for multivariate prediction. In our case, death and expenditures were simultaneously predicted based on past expenditures. In this way, the correlation between death and expenditures (see Liu, 2009) can be implemented easily. In model-based approaches complexity considerations may act as a limiting factor for multivariate prediction. The nearest neighbors resampling method also has its weaknesses. First, it requires a lot of data to be effective. For sparse data it may be less effective, as it requires the nearest neighbors to be similar enough. In our experience this can be a particular issue for the upper tails of the HCE distribution, because there is a lack of observations in this part. This has also been documented by Buishand (2007). Secondly, it is computationally intensive. Simulating time series can consume a lot of time, as the nearest neighbors can only be found by calculating distances for each observation in the data (although algorithms have been proposed to make the search for neighbors faster). Finally, in the nearest neighbor resampling application any bias that may occur in the short-term may evolve into a bigger bias over the long-term. Note that this may be true for a lot of iterative methods, and that it is not just limited to nearest neighbor resampling. However, in our experiences algorithm performed rather well when it comes to predictions in the short-term.

208

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES

10.6 Conclusion In this paper we modeled the distribution in lifetime acute health care expenditures using nearest neighbor resampling, a method for resampling time-series data. The distribution of lifetime expenditures is far less skewed than the distribution of annual expenditures, which puts interpersonal transfers in social health insurance systems –in terms of differences between the premiums paid and the amount of health care received in return– in a different light. While these transfers are great annually because of a highly skewed distribution (as evidenced by a relatively large difference between the median and the mean: assuming the price and volume level of the year 2005, 50% of the publicly insured population spend less than €400 on acute health care in one year, when the mean in a year in our sample is around €1,700), lifetime transfers are relatively much smaller (assuming the price and volume level during a lifetime remain fixed at the level of 2005), 50% of men spend less than €83,000 on health care during lifetime, which is closer to the lifetime mean of €99,000). Our results suggest that a large group of individuals will incur a high amount of HCE at some point of their lives, something which is not apparent when observing cross-sectional annual distributions. Nevertheless, the lifetime distribution is still highly skewed. About 5% of all men spend more than €222,000 Euros on acute health care, while 5% spend less than €22,000 Euros during lifetime. For women, we find a similar degree of inequality. In the Netherlands, the Council for Public Health and Health Care advised the government to consider alternate funding forms like health care savings accounts (Jeurissen, 2005). This is a form of funding which involves personal tax-free accounts, on which money is set aside by deducing a monthly premium off personal income. The balance on the account can be used to pay for health care, but cannot be withdrawn until a pre-specified age, or in some versions, until death (in that case, the balance will go to relatives). It is clear that individual savings accounts only make sense when there is something worth saving for, for most individuals involved. In other words, if there is a substantial fixed amount of HCE which most individuals have through the course of their lives. The problem is of course, that we cannot observe this amount, given only the relatively recent surge in data collection and computing technology. Although further research will be needed, the life cycles as generated with the nearest neighbor resampling algorithm allows us to make well-educated guesses as to what these baseline expenditures amount to, based on a large yet limited (in terms of time span) dataset.

209

CHAPTER 10

Appendix 10.A: Validation of nearest neighbor resampling for a known process Table 10.A1: Covariance matrix as used in our simulation to generate the dataset.

1.000 0.500 0.250

0.500 0.250  1.000 0.500  0.500 1.000  

Table 10.A2: Covariance matrix of the data matrix containing a simulated continuation on each row. Results are shown for the simulation of six time points ahead.

 1.000  0.496  0.236  0.114  0.065  0.003 

210

0.496 0.236 0.114 0.065  0.003 1.000 0.492 0.230 0.136 0.052  0.492 1.000 0.501 0.246 0.118  0.230 0.501 1.000 0.506 0.136 0.246 0.506 1.000 0.052 0.118 0.235 0.504

0.235 0.504 1.000

   

MODELING THE LIFETIME DISTRIBUTION OF HEALTH CARE EXPENDITURES Figure 10.B1: Q-Q plot for the simulated marginal distribution at each time point ahead versus the reference distribution (standard normal distribution).

211

CHAPTER 10

212

Chapter 11 Interpersonal Variation in Lifetime Health Care Expenditures ‡

While it is known that the interpersonal variation in health care expenditures is large on an annual level, variation as seen over a lifetime is still not understood well. Furthermore, the lifetime variation may also differ by health care sector. In this study we apply, and validate, the nearest neighbor resampling method on a large insurance dataset to generate individual life cycles in health care expenditures, which can be used to investigate the lifetime interpersonal variation. Six health care sectors are considered (hospital care, general practitioner care, dental care, paramedic care, drug treatments, medical devices utilization). We find that the inequality in health care expenditures differs strongly between health care sectors. Gini coefficients as estimated from the synthetic life cycles suggest that general practitioner, hospital and dental care exhibited a moderate degree of inequality. On the other hand, the interpersonal variation in medication, medical devices and paramedic care is large. These results may bear relevance to the implementation of health care financing schemes, such as health care savings accounts.

11.1 Introduction Rising health care expenditures (HCE) have exerted strong pressure on many governments in Western countries to reconsider their health care financing systems. With no end in sight for the growth in HCE, the call for a solution is ubiquitous, amongst policy makers and economists. Yet, the debate rages on, with no easy ‘cure all’ solution on the horizon. Particularly, many question the This chapter is based on: Wong A, Boshuizen HC, Polder JJ. 2011. Interpersonal variation in lifetime health care expenditures (submitted). ‡

CHAPTER 11

sustainability of current health care systems. At the heart of this debate are, amongst others, policies to control costs. Cost control measures include increasing out-of-pocket payments, increasing co-payments and exclusion of medical services and goods from the health insurance coverage. These measures imply a certain responsibility on the individual level. An important question is how much an individual has to pay for his or her own health care, and how much will the public services need to contribute? All of these issues are central in the concept of solidarity. In a social insurance program (e.g. Medicare and Medicaid in the US) or a compulsory health insurance scheme (e.g. in the Netherlands, Switzerland), all individuals contribute an amount (possibly income-dependent) such that health care is completely financed. These systems exhibit a large degree of solidarity, as many individuals make a financial contribution to keep health care accessible for those who are in need, but may otherwise not be able to finance it with their own income or savings. On the other hand, health care financing systems that make use of Medical Savings Accounts (MSA) are focused on a large individual responsibility. Medical savings accounts (MSA) involve individual tax-free saving accounts that are used solely for purposes of health or medical care spending. Although MSAs can come in many variations, they are often characterized by funding based on income, and by a withdrawal restriction that is released after a pre-specified age. MSAs mainly are introduced “to encourage savings for the expected high costs of medical care in the future, to enlist health care consumers in controlling costs, and/or to mobilize additional funds for health systems” (WHO, 2002b). MSAs are seen as potentially cost-controlling, as they might reduce moral hazard in health care (see Rasmusen, 1987). The efficiency of MSA depends on several assumptions. Within a compulsory health insurance scheme, payments are made on an annual and cross-sectional basis. Thus, little knowledge of (distant) future HCE is needed to make this financing work. MSAs, however, require a good idea of what the future holds. How much does an individual need to save, given his health at some point? Does everyone need to save the same amount? How many people save too much and unnecessarily lose purchasing power? How do time trends, as caused by changes in medical technology and policy, affect these savings? In case MSAs are coupled with a health insurance (as is the case in the US and Singapore), how should the funding of medical care be allocated over each? All of these questions bear great relevance to MSAs. However, in order to answer them, the lifetime mechanics behind individual HCE need to be understood better. Surprisingly, little has been done in this regard. French and Jones (2004) found, using an AR(1) model on data from the US Health and Retirement Study, that 0.1% of the households suffer a health shock that leads to lifetime costs over $125,000. De Nardi et al. (2009) used a similar model on AHEAD data to estimate HCE over lifetime, and find that both the level and volatility of health care 214

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

expenditure increase sharply with age. Brown and Finkelstein (2008) use an actuarial model with transition probabilities to find that 40% of current 65-year old men and 54% of 65-year old women will use long-term care (nursing homes, assisted living, and home health care) at some point in time. Forget et al. (2008) use a Markov transition model on Canadian data to find that men and women will have, on average, 7.8 and 11 high cost years during their life, respectively. Finally, Wong et al. (2011c) used a nearest neighbor resampling scheme on Dutch insurance data to estimate the lifetime distribution of acute HCE 1. However, none of these studies have looked into sector-specific HCE. The aim of this paper is to simulate life cycles in terms of sector-specific HCE, which can then be used to examine individual health care spending patterns over the course of a life. The method as proposed by Wong et al. (2011c) will be adapted and modified accordingly. We will highlight some of these spending patterns, which may be relevant in assessing the feasibility of MSA, as seen from an individual consumption perspective. This paper is structured as follows. First, we give an introduction to nearest neighbor resampling as used by Wong et al. (2011c), describe our data, consider some variations of the method to make life cycles in sector-specific HCE, and validate these methods. Then we describe the life cycles that were generated with these methods, and end with a discussion and conclusion.

11.2 Methods An introduction to nearest neighbor resampling Nearest neighbor resampling is a non-parametric method to resample dependent data. A particular instance of its application may be in the context of time series. Suppose a researcher has a time series to his disposal, but the end of the time series is missing. Using nearest neighbor resampling, he may be able to ‘complete’ this time series, therefore allowing him to make inferences on the full time series. Alternatively, nearest neighbor resampling enables a researcher to examine the long-term behavior of a time series which can only be observed for a relatively short period. Applications are found in many disciplines, including physics (Farmer and Sidorowich, 1987; Casdagli, 1989; Casdagli, 1991; Casdagli, 1992), hydrology (Lall and Sharma, 1996; Rajagopalan and Lall, 1999; Buishand and Brandsma, 2001; Buishand, 2007), and finance (Hseih, 1991). Wong et al. (2011c) adapted this method to construct multiple synthetic life cycles of HCE, one for each simulated individual. Figure 11.1 depicts a schematic overview of their method.

1

From this point we will regularly refer to expenditures in the acute health care sector as ‘HCE’.

215

CHAPTER 11 Figure 11.1: A schematic overview of the nearest neighbor resampling algorithm from Wong et al. (2011c). Nearest neighbor for age a+2 History matching (length p)

Nearest neighbor for age a+1 History matching (length p)

Nearest neighbor for age a History matching (length p)

Synthetic lifecycle at age a-1

length p

a-p

a-1

a

a+1

a+2

The general idea can be described as follows. A synthetic life cycle is constructed by repeatedly sampling actual observations at subsequent ages, each time taking the previously sampled observations for that life cycle into account. A synthetic life cycle is initiated by sampling randomly an observation from zero-year olds. Supposing this synthetic individual survives at age zero, he then ages one year and also incurs HCE at age one. To impute the expenditures at age one, we consider the synthetic history of expenditures up to this point (in this case, the expenditures for the synthetic individual at age zero). Here, the so-called k nearest neighbor principle is applied: we search the population of one-year olds for k individuals that are most similar to the synthetic individual, in terms of the history of expenditures. One of the nearest neighbors then will be randomly drawn (either based on the uniform distribution, or on a distribution which postulates that the probability of sampling a neighbor is directly related to the degree of similarity between the neighbor and the synthetic individual), and then the expenditures and death status of this drawn neighbor will be assigned to the synthetic individual. With death the synthetic life cycle also ends. With no death the process repeats itself for subsequent ages, each using an updated history of HCE. A more probabilistic interpretation can be given as follows. Suppose Z t and Ct are random variables representing the expenditures and the death indicator respectively at time t for the synthetic individual. Further suppose that the lag

216

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

order p determines how we far back are able to, or willing to, observe the history from t in our data, and is fixed for all ages (with the exception of lower ages, where only a smaller lag is possible). Drawing one of the k nearest neighbors is similar to the process of estimating a conditional density f Z t ,Ct |Z t 1 , Z t  2 ,..., Z t  p (often non-parametrically, see Li and Racine (2003) for instance) and then drawing one observation from this estimated density. The value for k is chosen beforehand, and determines indirectly how similar the neighbor is to the synthetic individual. For a larger k more heterogenous neighbors are also considered for imputation, possibly leading to more bias. To determine which observations are nearest neighbors a so-called distance function can be used. In this context, this function measures the degree of similarity between the potential neighbor and the synthetic time series of interest based on their respective histories Y  Yt 1 , Yt  2 ,..., Yt  p  and Z  Z t 1 , Z t  2 ,..., Z t  p  . We will discuss some choices for the distance function later on. Data In this study two datasets are used: Vektis Insurance data, and the Dutch Municipality Register (GBA). Vektis is an institution operating under Zorgverzekeraars Nederland, a union of all Dutch health care insurance companies. Amongst others, they collect and analyse insurance data. Our dataset from Vektis contains HCE of all those Dutch individuals who were, at some point, publicly insured during 1997-2005 (the so-called “Ziekenfondsverzekerden”). During this time, two types of insurances existed in the Netherlands: public and private insurance. On the one hand individuals with an annual income less than €33,000 (including family members that live in the same household as a publicly insured individual, but are not employed) could become publicly insured for a small premium (a monthly fixed fee of €35, as well as an income rate of 1.5%). Non-eligible for public insurance were mainly those who earned annually more than €33,000, public servants, and those who were unemployed and had no access to unemployment pay. Therefore, the Vektis dataset mainly contains those who belong to the bottom income deciles, but at the same time, it misses some individuals with zero income (i.e., those who did not live with a publicly insured individual, and/or had no income while having no access to unemployment pay). The Vektis dataset covers HCE in the acute health care: hospital in- and outpatient care (including birth and pregnancy related care, as well as ambulant transportation), general practitioner care, dental care, drugs, medical aid devices, and paramedic aid. The Vektis dataset was linked to the Dutch Municipality Register by Statistics Netherlands to obtain additional information such as date of death.

217

CHAPTER 11 Figure 11.2: Mean (solid lines) health care expenditures by age, as well as 5% and 95% quantiles (dashed lines), for different health care sectors in 2005, based on our sample.

40

60

80

20

60

40 Age

0

20

60

80

40

60

80

Age

300

Paramedic care

0

100 0

0

20

1000

80

GP care

Expenditures

100 200 300 400

40 Age

Dental care

0

500

Expenditures 0

Age

Expenditures 200 400 600 800

20

0

0 0

Expenditures

Medical devices

1500

2500

Medication

500

Expenditures

15000 5000 0

Expenditures

Hospital care

0

20

40 Age

60

80

0

20

40

60

80

Age

We made an additional selection of (a) individuals that were at least alive until the last year (i.e., 2005), and (b) individuals that were fully publicly insured during 1997-2005. In other words, we made a selection of those individuals who had a complete history of HCE during this period available. This selection probably led to exclusion of some self-employed individuals, whose income fluctuates around the €33,000 mark (causing them to gain and lose eligibility for the public insurance system). The resulting database, which we shall refer to as ‘main dataset’ from hereon, contained 1.3 million men and 1.7 million women. The dataset contains more women than men because women have, on average, lower incomes, often out of part-time employment. Furthermore, it contains housewives that are publicly insured because they are living together with a publicly insured partner. Figure 11.2 shows the mean, 5% and 95% quantiles by age for each sector. It shows that mean and variance increase steeply with age for all sectors, except for dental care (which involves many out-of-pocket payments for adults). Two versions of the nearest neighbor resampling algorithm In this paper we consider two versions of nearest neighbor resampling to estimate sector-specific lifetime HCE. Other versions have also been considered for analysis, but were mainly found to be performing less well and have been left out of this article for reasons of clarity and brevity. 218

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

1. Nearest neighbor search based on total expenditures, with post-hoc distribution Here, we used the Euclidean distance vector: d (y, z )  (y  z ) T (y  z ) ,

(11.1)

where y  Yt 1 , Yt 2 ,..., Yt  p  and z  Z t 1 , Z t 2 ,..., Z t  p  denote the vectors of previous total expenditures for the potential nearest neighbor and the synthetic time series, respectively. This is an identical nearest neighbor search to Wong et al. (2011c); although this time we modify the algorithm slightly to make a post-hoc distribution of HCE to each health care sector. Once we have found and drawn a nearest neighbor whose last value is Yt , we not only impute the expenditures Yt (and the death status), but also store the distribution  t  ( 1,t ,...,  s ,t ) , where s is a subsector index, and  s,t denotes the share of that subsector in the total HCE Yt . For the Vektis data, we can choose the lag order p between 1 and 8. 2. Nearest neighbor search based on sector-specific expenditures * * Let y t  Yt ;1 , Yt ; 2 ,..., Yt ;s  and z t  Z t ;1 , Z t ; 2 ,..., Z t ;s  for s  S , where S is the collection of all sectors within acute health care ( S  {1,2,..,5,6} ). Furthermore,



* (1)

*( n )



represents all potential neighbors in the data. Under suppose that y t ,..., y t this version, we used the Mahalanobis distance vector: d (y * , z * )  (y *  z * ) T  1 (y *  z * ) ,



*

y *  y t 1 ,..., y t p

where

*



(11.2)



*

and z *  z t 1 ,..., z t p

*



denote the vectors of

previous sector-specific expenditures for the potential nearest neighbor and the synthetic time series respectively, and  is an estimate of the covariance matrix of * (1) *( n ) y t ,..., y t .  is used to account for the differences in scale between health care sector. For instance, hospital expenditures might vary anywhere between zero and hundreds of thousands Euros, while GP expenditures fall between zero and several hundreds of Euros. The Mahalanobis distance differs from the Euclidean distance in the weighting of each sector. The Euclidean distance gives most weight to the hospital expenditures, because the scale (and thus, the variance) of hospital expenditures is much larger. On the other hand, the Mahalanobis distance rescales the variance of each sector such that they have more similar weights in the distance. This version simulates the process of estimating the conditional density f Z * ,C |Z * ,...,Z * and drawing one observation from this estimate.





t

t

t 1

t p

219

CHAPTER 11

Setting up a validation study A validation study was set up to estimate which tuning parameter values for the nearest neighbors algorithm will yield the best results, and to assess the performance of both versions of nearest neighbor resampling. The general idea behind the validation involved predicting the HCE for the most recent year t * available in the dataset, and comparing these to the actual observed expenditures in t * (Figure 11.3). First, each dataset was divided into strata, based on gender and age (in the most recent year of the dataset). Within each stratum, we predict HCE and death status in t * for each individual, by searching for the nearest neighbor(s) based on the history of HCE leading up to t * , with a pre-chosen lag order p , and then using the observed value of (one of) the nearest neighbor(s) as the predicted value. Although many criteria can be used for comparison, we follow argumentation and strategy as given by Wong et al. (2011c). The nearest neighbor resampling was assumed to work well if the validation study showed that some properties of the original data are retained. The two main criteria we considered were the marginal distribution of the HCE at t * , and the correlation between HCE at t * and at previous years. The similarity between the predicted and observed marginal distribution assesses the short-term performance of the nearest neighbor resampling algorithm. Since the algorithm is an iterative scheme, we can expect biases that occur in the short-term to become increasingly worse over the longterm. The similarity in predicted and observed correlations is a direct measure for long-term performance; it is possible that the marginal distribution might suggest a good short-term performance, but that the biases only become visible over a longer period of time. Choosing the nearest neighbor parameters and validation of the resampling algorithms 1. Nearest neighbor search based on total expenditures, with post-hoc distribution Since the nearest neighbor search procedure for the HCE is identical to Wong et al. (2011c) and was validated there, we chose k  1 and lag order p  8 , as they yielded the best results in terms of short-term and long-term performance of total HCE. For life cycle extrapolation purposes however, we used k  2 (with distance weighted neighbors), to prevent identical paths (see Wong et al., 2011c). Upon examination of the post-hoc distribution of HCE, it was found that the short-term performance was similarly well-performing. The Q-Q plots in Figure 11.A1 depict the similarity between the predicted and observed marginal distribution in each acute health care sector. However, when using total expenditures to search for nearest neighbors, the hospital care sector

220

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES Figure 11.3: Validation scheme for the nearest neighbor resampling algorithm.

Nearest neighbor search

History of length p

Last year

Prediction for last year

Individual of interest

Observed value

Prediction

Imputation

(Data)

Observed value

Nearest Neighbor

Correlations between observed last year and previous years

Correlations between predicted last year and previous years

Correlation comparison

Marginal density comparison Observed Predicted Density

Correlation

Observed Predicted

1

2

3

4

5

Lag

6

7

8

Expenditures

has the greatest weight in determining the distance, because of its large share in total HCE. While this resulted in good long-term performance of this sector, health care sectors with smaller mean and variance in expenditures such as GP and dental care exhibited a bias. Table 11.A1 in Appendix 11.A shows the predicted and observed correlations between the last year and previous years, for several lags, for GP and hospital care (other acute health care sectors not shown due to space restrictions). The correlations for the predicted values are similar to those of the observed values for hospital care, but for GP they deviate at all lags. The observed correlation for GP at lag 1 is around 0.271, but the predicted correlation is much smaller (0.046). 2. Nearest neighbor search based on sector-specific expenditures Here, we found k  1 to give the best short- and long-term performance (not shown here). A plausible explanation is that k  1 will yield only the nearest 221

CHAPTER 11

neighbor, which will probably lead to the smallest bias. Though similar to the nearest neighbor search, k  2 (with distance weighted neighbors) was chosen for the life cycle extrapolation, as it gave the next best results, while avoiding the ‘deterministic’ nature of k  1 . The choice of the lag order was less straightforward. Table 11.A2 gives the correlation between the prediction for the last year and observed values in previous years in acute health care for different lag orders, for sixty-year old men. Choosing p  1 led to a worse long-term performance, as correlations for lag of two and over were found to be smaller than the correlations as found in the data. Values of p between two and four generally led to best reproduction of the correlations for all ages and genders, although correlations in the GP sector were still remarkably smaller. Note that the optimal lag choice is smaller than the lag choice for the nearest neighbor search based on total expenditures ( p  8 ). The nearest neighbor search is based on computation of a distance, which becomes increasingly less meaningful with the dimension size of the feature vector (Beyer et al., 1999). In other words, for a large dimension the likelihood of finding a neighbor which is similar (i.e., with a small distance) is smaller. For the nearest neighbor search based on sector-specific expenditures, there is a larger risk of a non-meaningful distance, as we have a feature vector that may amount up to 48 features in the case of acute HCE (6 sectors, and with each sector a lag order of 8). This was apparent when we compared the predicted and observed marginal density. For higher orders the predicted quantiles deviate more strongly. For the lag order we eventually chose p  2 , as this gave a reasonable long-term performance, while keeping the differences between the predicted and observed marginal distribution acceptable (Figure 11.A2 in Appendix 11.A). Detrending the data The nearest neighbor algorithm involves matching past expenditures of a potential neighbor with a synthetic history. The key thing to realize is that the past expenditures from the potential neighbor always are extracted from the dataset, where a time trend is present. On the other hand, the synthetic life cycle is always updated with expenditures from the most recent year of the nearest neighbor (i.e., 2005 for the Vektis data), which leads to the absence of a trend in the synthetic life cycle. Applying the algorithm as such will complicate the interpretation. We therefore made all data stationary, such that the simulated life cycles represent a “stationary reality” in which the HCE and mortality rates remain fixed at the level of 2005. Trends in HCE can be applied in a post-hoc way, but this has not been done in this paper. For acute HCE we found a positive trend through 1997-2005, which could be explained by price inflation, advances in medical technology, and changes in insurance coverage. The HCE were made stationary by removing the time trend

222

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

separately for each stratum of gender and age. This was done as follows. Since we are comparing histories, we only consider expenditures of survivors (i.e., those who are not in the last year of their lives). Suppose we have an 80-year old individual from the insurance data, with a history of eight years (i.e., the expenditures incurred from ages 72 to 79, at years 1997 to 2004 respectively). The expenditures incurred for this individual at 72 are rescaled to 2005 level, by multiplying with the increase in average expenditures of a 72-year old from 1997 to 2005. Similarly, the expenditures at age 73 are rescaled by taking the increase in average expenditures of a 73-year old from 1998 to 2005, and so on. For the nearest neighbor search based on total HCE, we used increments based on average HCE, while for the nearest neighbor search based on sector-specific expenditures we used sectorspecific increments.

11.3 Results We simulated 10,000 life cycles for each version of nearest neighbour resampling. In this section we will first focus on the version that uses sector-specific HCE to search for nearest neighbors, as that is the version that performs reasonably well for all sectors. Table 11.1 shows descriptive statistics for the simulated life cycles of men, for each health care sector. Hospital expenditures have the largest share in all lifetime acute HCE (76%), followed by consumption of drugs (21%). The remainder (3%) is divided over medical devices, dental care, GP care and paramedic care. All health care sectors also differ strongly in their moment-related statistics. When considering the relative variance, we find that hospital care and dental care have a similar variance coefficient (between 0.7 and 0.9). GP care has a Coefficient of Variation considerably smaller than one, while the variance is relatively large for drugs, medical devices and paramedic care. All sectors are highly skewed and kurtotic, with the exception of GP care (with a skewness of 0.58 and kurtosis of 0.62). These differences are also reflected in the Gini-coefficients. GP care features the lowest Gini coefficient (0.359), while medical devices and paramedic care have the highest Gini coefficient (0.570 and 0.553 respectively). This is further illustrated in Figure 11.4, which shows the Lorenz curve for each health care sector. GP care is reasonably close to perfect equality line (which runs from the origin under an angle of 45 degrees; in other words, along the hypotenuse of the light gray triangle), while the curve for paramedic care and medical devices is closer to the perfect inequality line (which runs along the legs of the light gray triangle). For women, we find that that hospital care is closest to the perfect equality line. This is probably explained by the fact that hospital care expenditures

223

CHAPTER 11 Table 11.1: Descriptive statistics of the simulated lifetime health care expenditures for the nearest neighbor search based on sector-specific expenditures (in Euros). Type

Simulated Lifetime Statistics Measure

Total

Hospital

Drugs

Devices

Dental

GP

Paramedic

92056 64585 2.84 18.36

69810 51586 3.63 31.26

14519 22034 7.54 88.14

3179 4940 7.7 113.4

1985 1690 2.49 7.83

1496 537 0.58 0.62

1063 1251 2.86 13.97

9251 19517 27676 49030 80652 120700 163293

5486 13513 19988 36316 60686 91859 126622

413 913 1500 3986 9728 18043 28411

0 0 70 738 1981 3969 6694

349 601 739 1015 1418 2161 4312

462 701 846 1110 1450 1825 2197

0 0 0 222 698 1473 2471

Variation between individuals Coefficient of Variation Gini

0.702 0.346

0.739 0.356

1.518 0.530

1.554 0.570

0.851 0.389

0.359 0.200

1.177 0.553

Share q last years in lifetime Mean for q=1 Mean for q=5 Mean for q=10 St. Dev. for q=1 St. Dev. for q=5 St. Dev. for q=10

0.122 0.273 0.383 0.145 0.203 0.220

0.142 0.293 0.396 0.172 0.230 0.243

0.067 0.252 0.403 0.117 0.213 0.255

0.078 0.256 0.420 0.181 0.300 0.344

0.002 0.012 0.032 0.032 0.080 0.138

0.068 0.123 0.178 0.096 0.115 0.127

0.020 0.071 0.129 0.111 0.204 0.270

Share age 65 and over Mean St. Dev.

0.336 0.269

0.342 0.282

0.378 0.320

0.398 0.365

0.040 0.155

0.159 0.152

0.140 0.280

Moment-related statistics Mean St. Dev. Skewness Kurtosis Quantiles 5 10 25 50 75 90 95

include expenditures related to pregnancy and birth. GP care, drugs and dental care have similar degrees of inequality, while the distributions of medical devices and paramedic care are highly skewed. It is important to note however, that while the lifetime expenditures are highly skewed, they are less skewed than the crosssectional distribution of lifetime expenditures (Figure 11.5). The latter is often used as a point of departure for discussions on solidarity in insurance-based financing schemes, but is clearly not representative for discussions on lifetime solidarity. In terms of how the expenditures are distributed over age during lifetime, we could distinguish two groups: sectors in which the last years of life have a relatively large share (hospital care, consumption of drugs, and medical devices), and sectors in which most expenditures fall before the last years of life, which largely corresponds with ages 65 and under (dental care, GP care, and paramedic care). Figure 11.6 gives the number of years with catastrophic HCE. The definition of catastrophic HCE varies in the literature, but they are often defined as some fixed percentage of household income (Xu et al., 2003). Here, we consider three

224

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

1.0

Figure 11.4: Sector-specific Lorenz curves for the sector-specific nearest neighbor method (top: men, bottom: women).

0.6 0.4 0.0

0.2

Cumulative Fraction HCE

0.8

Hospital Care Paramedic Care Medical Devices Dental Care GP Care Medication

0.0

0.2

0.4

0.6

0.8

1.0

0.8

1.0

1.0

Cumulative Fraction Population

0.6 0.4 0.2 0.0

Cumulative Fraction HCE

0.8

Hospital Care Paramedic Care Medical Devices Dental Care GP Care Medication

0.0

0.2

0.4

0.6

Cumulative Fraction Population

225

CHAPTER 11 Figure 11.5: Lorenz curves for the nearest neighbor search based on sector-specific costs – difference between one year and lifetime.

1.0

Women, Acute Health Care

0.4

0.6

0.8

One year Lif etime

0.0

0.2

Cumulative Fraction HCE

0.2

0.4

0.6

0.8

One year Li fetime

0.0

Cumulative Fraction HCE

1.0

Men, Acute Health Care

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

Cumulative Fraction Population

Cumulative Fraction Population

Figure 11.6: Number of years with catastrophic health care expenditures during lifetime, for different thresholds of household income.

0.2

0.4

0.6

Threshold 10% Threshold 20% Threshold 40%

0.0

Fraction of population

0.2

0.4

0.6

Threshold 10% Threshold 20% Threshold 40%

0.8

Women, Acute health care

0.0

Fraction of population

0.8

Men, Acute health care

0

5

10

15

Number of catastrophic years

20

0

5

10

15

20

Number of catastrophic years

thresholds: 10%, 20% and 40% of household income. To give an indication of how common catastrophic HCE are during lifetime, we assume all households income are equal to the average household income of €20,600 (Statistics Netherlands, 2010) for this purpose. Even when considering a threshold of 40%, a non-negligible fraction has one to five catastrophic years during lifetime. Table 11.A3 shows the descriptive statistics for the simulated life cycles of men for each health care sector based on the nearest neighbor search with posthoc distribution. The main differences of these simulated results with that described above were: (1) almost all sectors have a higher mean lifetime HCE, (2) almost all sectors have a lower Coefficient of Variation of lifetime HCE, with the exception of hospital care, (3) almost all sectors have a lower Gini coefficient, with the 226

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

biggest differences observed for drugs, dental and paramedic care (4) most sectors feature a larger share of expenditures in the last years of life, and (5) the 95th quantiles (and above) of HCE are higher, particularly for hospital care.

11.4 Discussion In this paper we used the nearest neighbor resampling technique to simulate individual life cycles for HCE for six different health care sectors: hospital care, drugs, paramedic care, medical devices, dental care, and general practitioner care. These life cycles were then used to describe the interpersonal variation in HCE. Some of the key results and their implications are as follows. (1) The Gini coefficient of inequality over one year was roughly 0.8 for acute health care, but over the entire lifetime the Gini coefficient was around 0.3. Thus, the lifetime distribution in HCE, for each health care sector, is considerably less skewed than the cross-sectional (annual) distribution of HCE. This means that policy makers can be strongly misled by only considering annual distributions. (2) Although the lifetime distribution is less skewed, the inequality is still considerable. Similar to the findings of French and Jones (2003) and Forget et al. (2008), the interpersonal variance is still very high, which implies that it is not straightforward to set the target amount that individuals must save. Saving a higher amount means there will be more surpluses amongst individuals, causing a loss of welfare for these individuals, but saving a smaller amount means less potential cost reductions based on moral hazard (assuming savings will occur after introduction of MSAs, though evidence for this is still lacking). The substantial variation between individuals can be seen as an argument against MSAs. (3) The inequality in HCE differs fairly strong between health care sectors. Hospital, GP and dental feature a mild degree of inequality, while, on the other hand, paramedic care and medical devices featured a stronger degree of inequality. The large differences between the sectors may suggest that some sectors have more potential for MSA than others. For instance, it may be conceived that hospital, GP and dental care are financed by an individual savings-based approach, while sectors with a large degree of inequality fall under an insurance scheme instead. (4) The large variance between individuals, coupled with the high number of years with catastrophic expenditures amongst individuals, suggest that financing HCE with MSAs alone might be infeasible for some individuals, but that it must be combined with some kind of health insurance. This has been the case in the US and Singapore (WHO, 2002b). (5) In line with Forget et al. (2008) and de Nardi et al. (2009), the variation in age-specific HCE is large. As a result, this complicates making insurance coverage agedependent, or any other age-related policies, as any individual spending within an age group will be far from predictable. 227

CHAPTER 11

There is still a considerable amount of uncertainty associated with lifetime HCE predictions. While our main findings remain fairly robust, the quantile estimates of lifetime HCE differed between methods, particularly in the upper tails of the distribution. By contrast, an insurance type of scheme only requires short-term predictions of aggregate HCE. This large disparity in available information must be taken into account when considering MSA as a financing scheme. The variation in HCE also has relevance in other contexts. Ameriks et al. (2011), De Nardi et al. (2011) and Scholz et al. (2006) have looked into reasons for the so-called annuity puzzle: in the US, the elderly tend to keep large amounts of savings until very late of life, and have little demand for annuities. They found that, by saving, the elderly anticipate the high out-of-pocket expenditures associated with the health care at higher ages. Similar to these studies, we find a large amount of uncertainty in HCE. On average one third of all acute lifetime expenditures are incurred at ages 65 and over, but this share varies strongly between individuals. As the optimal annuitization depends on the health care cost risk early in retirement (Peijnenburg, 2011), the results in this study can help in determining the variety of optimal annuity demand required. We encountered some limitations of our methods. Wong et al. (2011c) already outlined the strengths and weaknesses of the nearest neighbor resampling algorithm. Its main weakness is that the algorithm has an iterative nature, where any short-term bias may evolve into a larger long-term bias. Of course, this is true for any iterative method (e.g. a Markov-based scheme). However, we have validated these results as extensively as possible, and have found the short-term bias to be reasonable. We did find different results between the nearest neighbor search versions. In this paper, we examined nearest neighbor search versions based on (1) total expenditures, and (2) based on sector-specific expenditures. The first version is more accurate in estimating hospital care and, thus, total acute HCE (as hospital care makes up for most of acute HCE). The second version is more effective with estimating the HCE for smaller sectors, such as GP care, but is less precise for hospital (and thus, total) expenditures. In a sense, the second version suffers more from the curse of dimensionality, which can occur in contexts other than nearest neighbor search as well (e.g. regression). Simultaneous prediction of six response variables (or seven, when taking into account the death indicator) is without question a more demanding task than the prediction of just one response variable. Clearly, there is no ‘one size fits all’ solution. The choice for a method depends on whether the goal is to estimate total expenditures, or the relative differences between health care sectors. In this paper it is the latter, but it is good to realize that the total expenditures are probably not the best estimates. The strengths of the method outweigh its weaknesses, however (see Wong et al., 2011c). They include no distributional or homogeneity assumptions, and the possibility for multivariate prediction (here the death status and sector-specific 228

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

expenditures are simultaneously predicted). As far as we know, no study has looked into the sector-specific inequality of HCE before. The method requires the data to be stationary. Of course, there is an upwards trend in HCE over time. We detrended the data by scaling the expenditures of each person by a proportional amount deduced from survivors of similar an age group. By doing so, we account for both cohort and calendar year effects (which are difficult to distinguish, as is apparent in literature. We have therefore not pursued age-period-cohort models). However, we realize that this is the simplest way to detrend the data, and that real trends in HCE are much more complex. For instance, the introduction of a new drug might lead to much higher health care utilization of specific patient groups, but may not affect other groups at all. And if other drugs and other factors come into play, then the end results might be a complicated mixture of each individual effect. However, we feel it is reasonable to assume that the average of a large number of individual effects will behave in a reasonably regular manner. Furthermore, causality may also provide additional complications. If a drug may effectively cure a disease, then the introduction of such a drug may lead to less health care expenditure in the future (which complicates the interpretation of subsequently observed expenditures). This leads to many ‘what if’ scenarios that cannot be understood based on this data alone. While other ways to detrend can be devised (note that model-based approaches also most often detrend data based on the mean), it cannot be validated extensively which detrending method is better. Recall our validation setup, where expenditures are predicted based on matching histories that comprised original data (i.e., including a time trend), and where a transformation is thus not needed. Furthermore, we can not use the synthetic life cycles to perform heuristics. The life cycles do not only include any potential biases as caused by the detrend method, but also by the nearest neighbor method itself (e.g. the slight short-term bias we observed in the Mahalanobis-based method). At best, the distribution of the simulated expenditures by age can be compared to the distribution of the original data. Indeed, we found that these two differed somewhat (not shown). While it seems evident that cohort differences play a part – the initial states, obtained by sampling expenditures of zero-year olds in the last year, may not reflect the ‘true’ initial states for the elderly in that same year – long-term errors of the nearest neighbor extrapolation cannot be ruled out (note that this is a common theme for any extrapolation), although Wong et al. (2011c) have shown that for a simple known process, the long-term properties of the nearest neighbor method are, in theory, good. Finally, our data also had some strengths and weaknesses. The insurance dataset had a large sample size, and allowed us to look at sector-specific HCE. It had as main limitation that it only contained publicly insured individuals, which often fall into the lower income group. These individuals tend to have poorer health, which 229

CHAPTER 11

was reflected in our data by slightly higher mortality rates than in the whole Dutch population. This means simulated life cycles based on the insurance dataset have a slightly shorter life expectancy. Thus, results from this study cannot be used for direct inferences on the entire Dutch population. However, they are likely to give a good indication. Finally, long-term care was not available in this dataset, and still remains a topic for further research.

11.5 Conclusion In this paper we have used the nearest neighbor resampling algorithm to simulate individual life cycles in health care expenditures for six different health care sectors. To the extent that medical savings accounts are feasible, our results show that not each type of health care expenditures may be suitable candidates for medical savings accounts. Of all acute health sectors, only hospital care, GP and dental care exhibited a moderate degree of inequality. The differences between sectors, both in terms of lifetime mean and variance, may warrant a sector-specific approach when considering medical savings accounts as a health care financing form. These differences may only appear after taking a life cycle approach such as the one presented in this paper.

230

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES

Appendix 11.A: Supplementary material to the validation study Figure 11.A1: Q-Q plots between the simulated and observed marginal distributions for the nearest neighbor search based on total acute health care expenditures. Results pertain to 60-year old men (under parameter choices k=1 and lag order p=8).

Table 11.A1: Comparison between the predicted and observed serial correlation for the nearest neighbor search based on total acute health care expenditures. Results pertain to 60-year old men (under parameter choices k=1 and lag order p=8). Data Measure Serial Correlation at lag 1 2 3 4 5 6 7 8

Predicted

Hospital

GP

Hospital

GP

0.285 0.184 0.155 0.113 0.087 0.111 0.050 0.076

0.271 0.209 0.221 0.168 0.084 0.082 0.075 0.013

0.276 0.182 0.117 0.119 0.068 0.087 0.043 0.044

0.046 0.027 0.036 0.005 0.002 -0.002 -0.001 0.003

231

CHAPTER 11 Figure 11.A2: Q-Q plots between the simulated and observed marginal distributions for the nearest neighbor search based on sector-specific acute health care expenditures. Plots pertain to 60-year old men (under parameter choices k=1 and lag order p=2).

Table 11.A2: Comparison between the predicted and observed serial correlation for the nearest neighbor search based on sector-specific acute health care expenditures. Results pertain to 60-year old men (under parameter choices k=1). Nearest Neighbor on hospital HCE with lag p Measure

Nearest Neighbors on GP HCE with lag p

Data

p =1

p =2

p =4

p =8

Data

p =1

p =2

p =4

p =8

0.285 0.184 0.155 0.113 0.087 0.111 0.050 0.076

0.285 0.124 0.081 0.077 0.061 0.093 0.044 0.051

0.303 0.213 0.164 0.095 0.093 0.121 0.061 0.047

0.230 0.142 0.113 0.116 0.064 0.076 0.031 0.031

0.215 0.144 0.057 0.089 0.063 0.078 0.046 0.047

0.271 0.209 0.221 0.168 0.084 0.082 0.075 0.013

0.360 0.205 0.127 0.098 0.026 0.032 0.031 -0.007

0.303 0.285 0.169 0.092 0.030 0.042 0.017 0.019

0.224 0.193 0.158 0.074 0.028 0.015 0.019 -0.003

0.205 0.188 0.148 0.026 0.014 0.025 0.023 0.016

Serial Correlation at lag 1 2 3 4 5 6 7 8

232

INTERPERSONAL VARIATION IN LIFETIME EXPENDITURES Table 11.A3: Descriptive statistics of the simulated lifetime health care expenditures for the nearest neighbor search based on total expenditures (men; in Euros). Type Measure

Simulated Lifetime Statistics Drugs Devices Dental

Total

Hospital

98853 70919 2.55 13.95

72753 55265 3.05 21.72

16451 15310 3.49 29.11

4120 5285 5.16 54.37

10051 21185 30859 52868 83753 126619 179283

5750 13971 21347 37601 60863 93658 134379

636 1709 2955 6462 12788 21928 33633

Variation between individuals Coefficient of Variation Gini

0.717 0.355

0.760 0.368

Share of q last life years in lifetime Mean for q=1 Mean for q=5 Mean for q=10 St. Dev. for q=1 St. Dev. for q=5 St. Dev. for q=10

0.126 0.316 0.438 0.146 0.206 0.218

Share age 65 and over Mean St. Dev

0.402 0.277

Moments Mean St. Dev. Skewness Kurtosis Quantiles 5 10 25 50 75 90 95

GP

Paramedic

1737 1722 2.01 6.15

1754 563 0.74 1.99

2040 1422 2.59 9.67

0 69 315 1170 2719 5182 8956

0 33 159 514 1251 2444 3901

637 928 1096 1373 1701 2080 2465

582 815 947 1214 1585 2263 3936

0.931 0.440

1.283 0.533

0.991 0.497

0.321 0.177

0.697 0.322

0.147 0.340 0.456 0.177 0.237 0.245

0.067 0.276 0.439 0.111 0.191 0.226

0.091 0.316 0.483 0.185 0.297 0.320

0.021 0.109 0.178 0.102 0.229 0.283

0.065 0.132 0.194 0.086 0.105 0.116

0.002 0.026 0.051 0.030 0.119 0.165

0.412 0.293

0.422 0.303

0.464 0.360

0.179 0.289

0.184 0.145

0.057 0.174

233

CHAPTER 11

234

Chapter 12 General Discussion

The aim of this thesis was to apply statistical methods to generate additional knowledge on the determinants, distribution and dynamics of health care expenditures. We developed knowledge on the following three themes: I. The ‘Red Herring’ phenomenon (Chapters 2-4) The link between individual health care expenditures and the last year of life has been used by researchers to estimate the effect of aging in the growth of health care expenditures, as well as to forecast aggregate health care expenditures. This relationship was studied on a disease-specific level (i.e., health care expenditures are separated for each disease), which helps us to understand both the nature of the link, and potential changes in this link over time. Furthermore, the link was also explored on an aggregate level (i.e., the link between aggregate health care expenditures and mortality rates). This brought up some issues with forecast models that are based on the link (i.e., on individual level), that require careful consideration by researchers. Finally, we also used the link in a cost-effectiveness context, which allows us to better anticipate additional future medical costs in the life years gained as a result of a health intervention. II. The relationship between the presence of one or more diseases and health care expenditures (Chapters 5-8) Although the presence of one or more diseases (the latter is commonly referred to as comorbidity) has been associated with high health care expenditures, the exact relationship is still not particularly well-understood, as comorbidity encompasses a wide range of disease combinations. Four specific problems were examined. First, the relationship between morbidity and long-term care utilization was explored. This provides a hold on which specific diseases are relevant for preventing the need for long-term care. Second, it is not clear whether the population, as a whole,

CHAPTER 12

becomes less healthy – and thus, whether the need for health care increases over time. Using self-reported data, trends in morbidity and comorbidity were estimated that give an indication of how population health has evolved. Third, comorbidity is an umbrella term for any co-occurrence of two or more diseases: in theory, any combination of disease can co-exist. A study was set up to find which diseases were often simultaneously present within patients, as well as which disease combinations occurred more often than is expected based on chance alone. Finally, the relationship between comorbidity and health care expenditures was further explored. More specifically, it was studied whether one disease affects the expenditures for another disease (i.e., are the health care expenditures of two diseases additive in nature). III. Interpersonal and –generational distribution of health care expenditures (Chapters 9-11) A good understanding of the interpersonal and intergenerational distribution of health care expenditures is required to make well-informed decisions on health care financing. For cross-sectional based financing, it is important to know whether, and how, the intergenerational distribution will evolve over time. We studied these changes over time, and the changes could be linked to advances in medical technology, in the case of hospital care. For a longitudinal based financing approach (e.g. health savings accounts), the interpersonal distribution of health care expenditures as seen over a lifetime is of major interest. A method was proposed to estimate this distribution of total health care expenditures, as well as sector-specific health care expenditures. All above mentioned problems were tackled using a wide range of statistical methods. The methods included panel data techniques (generalized estimating equations) and multiple equations models (the two and three-part model), as well as methods for missing data (multiple imputation by chained equations) and resampling techniques (nearest neighbor resampling). These methods can be used in multiple contexts. Generally, we could make a distinction between the following three approaches:  Descriptive modeling (chapters 2, 7, 8)  Explanatory modeling (chapters 5, 6, 9)  Predictive modeling (chapters 3, 4, 10, 11) We have shown that each of these approaches has its merits, and all contribute to the understanding of health care expenditure in various ways. One of the main goals of descriptive modeling in this thesis was to determine which (combinations of) diseases were associated with high health care expenditures. The main advantage of this approach was that it yielded many 236

GENERAL DISCUSSION

combinations of diseases that are possibly new and provide opportunities for further research. This approach can be useful when not much a priori knowledge on possible associations is available, as it can generate hypotheses in many cases. Our analysis of the relationship between comorbidity and health care expenditures could provide additional information that allows health economists, health care researchers and clinicians to set their priority in future research, with the ulterior goal of reducing health care expenditures, optimizing health services for comorbid patients, and understanding disease etiology. The disadvantages are the timeconsuming character of this approach, and possibly steep sample size requirements in case of multiple potential associations. The availability of large-scale datasets may also be an issue. Explanatory modeling was mainly used to answer very specific research questions. In chapter 6 we investigated whether the disabling effects of morbidity changed throughout time, while in chapter 9, we examined whether the age-specific growth in hospital care utilization could be explained by advances in medical technology. In both cases, an underlying theoretical framework or a decent amount of a priori knowledge was available, and empirical support was given for (parts of) this theoretical framework. This approach is highly common in the social sciences. Finally, predictive modeling was used to forecast future macro-level expenditures (chapter 4), and the lifetime distribution of health care expenditures (chapters 10 and 11). These models are characterized by their data-driven nature, rather than relying on a causal framework. For instance, the relationship between health care expenditures and the last year of life is clearly not a causal one. Nor are past health care expenditures necessarily causal factors in future health care expenditures, as these should probably be seen as byproducts of a persistent poor health status. In predictive models these types of variables are perfectly acceptable, as long as they make an important contribution to the prediction. Particularly, in chapters 10 and 11 we demonstrate how past health care expenditures can be effectively used to construct full life cycles of health care expenditures. Clearly, the main disadvantage of this approach is that it does not necessarily add to the understanding of underlying causalities (e.g. health in the case of health care expenditures). Many of the models in this thesis have one theme in common. They often require large-scale datasets. These include the Dutch hospital register, which has a nationwide coverage, and the Vektis insurance dataset, which covers a large fraction of the Dutch population. Particularly, chapters 3, 8, 9 and 10 show how the combination of advanced statistical techniques, computing power and largescale data can contribute to new findings that might have never been realized

237

CHAPTER 12

otherwise. Developments in these three fields go a long way into advancing health economics, and applied sciences as a whole.

I. The ‘Red Herring’ phenomenon Discussion of findings In terms of causality, using proximity to death as a predictor is an odd concept, as conditioning health care expenditures on a future event makes interpretation difficult. However, proximity to death should not really be interpreted in a literal sense, but it should rather be viewed as an approximation of the processes that lie underneath. Particularly, in terms of the cure sector, one can think of morbidity and comorbidity underlying proximity to death (Chapter 2), while disability plays a similar role in long-term care (de Meijer et al., 2011). One should realize though, that this approximation is very rough at best. Proximity to death fails, as a predictor, to capture those individuals who have high health care expenditures, yet are able to survive for a relatively long period. Examples would be individuals with a chronic disease, such as stroke, chronic obstructive pulmonary disorder and dementia. Not surprisingly, many of these individuals reside in long-term care institutions, which also explain why the proximity to death effect is much smaller in long-term care. Since these individuals do not die in the short-term, their incurred health care expenditures are probably attributed to age in proximity to death models. Furthermore, critics of the Red Herring have pointed towards potential estimation problems surrounding the presence of reverse causality (or in more economic terms, the presence of endogeneity). Proximity to death is used as a variable to predict individual health care expenditures, but at the same time, as more health care is consumed, the probability of survival is likely to increase. The reverse causality becomes a particular problem when the association between proximity to death and health care expenditures is time-varying. For instance, if the health care expenditures (in fixed prices) needed to prolong life increase over time, or in other words, if the cost-effectiveness of health care decreases over time, then the differences in average health care expenditures between deceased and survivors is likely to decrease, and as a result, the net relative effect decreases as well. On the other hand, if macro-level health care expenditures spending is mainly targeted at postponing death with just a few years for patients with incurable disease (e.g. cancer in advanced stages), then the proximity to death effect might become stronger. While Seshamani & Gray (2004) have found the proximity-to-death relationship to be stable over time, and the effect of proximity to death is unlikely to disappear altogether, the magnitude of the effect will probably change over time. In this thesis we have given another explanation, next to a potential changing cost238

GENERAL DISCUSSION

effectiveness of health care, in the form of changing morbidity patterns. Proximity to death is a mixture of disease-specific proximity to death relationships, where each disease is weighted by their prevalence. This implies that if disease prevalences change over time, then the proximity to death effect will change as a result. Note that reverse causality might also arise when actual morbidity indicators are being used instead of proximity to death, as the presence of diseases not only influences the level of health care expenditures, but also vice versa. Implications for research and policy Proximity to death models are mainly being used for two purposes: to estimate the effect of aging on macro-level health care expenditures, and to predict macro-level health care expenditures in the future. However, there may be some issues with using proximity to death. First, the effect of aging is based on the current relationship between age and health care expenditures. As was found in this thesis, the health care expenditures growth is likely to be higher for the elderly, in absolute as well as in relative sense. This suggests that the relationship with age and health care expenditures might become stronger over time and thus, the effect of aging increases as well. Second, current forecasts based on proximity to death models often use growth rates based on annual growth in aggregate health care expenditures (e.g. Wickstrøm et al., 2002). Our results indicate that the actual growth rate is higher when accounting for proximity to death, as declining mortality rates act as a dampening factor on aggregate health care expenditures. When using this higher growth in health care expenditures projections, forecasts with proximity to death were very similar to those without proximity to death (Chapter 4). The implication is here to that proximity to death models are much more suited to estimate the effect of aging, than they are to forecast health care expenditures, with the caveat that the relationship between age and health care expenditures will probably change over time. Results from this thesis also help in debunking misconceptions surrounding the health care expenditures in the last year of life. Many studies have found that the health care expenditures are higher in the last years of life. Polder et al. (2006) found that, in the Netherlands, the health care expenditures are over 13 times as high in the last year of life. An important thing to note here is that this relationship is true based on averages. In other words, this relationship can vary immensely between individuals. In our lifetime simulations (Chapters 10 and 11), we found that for men [women], on average, 32% [25%] of all acute health care expenditures during lifetime fall in the last years of life, but had a considerable standard deviation of 21% [17%]. Thus, the last year of life is not expensive for all individuals. The share for acute health care also implies that the majority of lifetime acute health care expenditures fall in the other years of life. In other words, an individual does not have to be near-death in order to incur heavy expenditures. In 239

CHAPTER 12

fact, only 11% of all annual aggregate health care expenditures is related to the last year of life (Polder et al., 2006). Another misconception is that the health care expenditures in last years of life are a ‘waste of resources’, and subsequently, some have argued to spend less on those who die. If this were indeed a case of a ‘waste’, then this would amount to the aforementioned 11% at most. But this line of thinking is far too rigid, as the finding that average health care expenditures are highest in the last year of life is, in a sense, misleading. Cutting on death-related health care expenditures is only possible if one could predict a priori who will die and who will survive. Furthermore, if there indeed was a case of ‘waste’, it is only after we look at the resources spent on those who manage to be cured, or postpone death – and a lot of individuals do survive as a result of more health care – that we can perceive the health care expenditures in the last year of life from a more reasonable perspective. Policy makers should be aware of these misconceptions, and realize that the costs in the last year of life have merely been used as a tool for estimating aggregate health care expenditures, and estimate the effects of aging on health care expenditures.

II. The relationship between the presence of one or more diseases, and health care expenditures Discussion of findings Using Dutch self-reported data for elderly, we examined trends for a selection of chronic diseases. The assumption was made that the screening and diagnosing of these diseases did not improved considerably during this time. While the results were not the same for all investigated diseases, we found that for diabetes, stroke and cancer the prevalence increased during 1990-2007. We also found that the prevalence of comorbidity, defined as the presence of at least two out of our selected diseases, increased over time, suggesting that the health of a senior in 2007 may be worse than the health of a senior in 1990 (Chapter 6). Not only does comorbidity have consequences for patients’ health, it is also widely associated with higher health care expenditures (Gijsen et al., 2001). Thus, it should be taken into account that comorbidity could increase the demand for health care even more in the future. The extent to which this will lead to health care expenditures, will probably depend on the combination of diseases involved. Within hospital care, we find that in about 7.5% of all investigated disease combinations one disease will increase the expenditures for another. For any combination of disease A and disease B, A might increase the expenditures for B, but not necessarily vice versa. The implication is here that comorbidity will lead to additional health care expenditures only for the diseases as investigated in this study (Chapter 8). 240

GENERAL DISCUSSION

Furthermore, the Dutch self-reported data indicated that the prevalence activity limitations in the Dutch population remained fairly stable. Since the prevalence of morbidity has increased over time, one might expect the relationship between morbidity and disability to have weakened, i.e., diseases to have become less disabling. Although this was found in some surveys, by and large, the relationship was stable as well. As disability is a socially defined construct, which depends on the person’s expectations of daily life, changing expectations might have clouded these trends. Older persons might be less inclined to accept a deterioration in their functioning, therefore potentially masking a true decline in activity limitations. Development of more objective measures for disability will be needed to truly examine these trends. This knowledge helps in anticipating future health care needs, as disability acts as an important driver for long-term care utilization. Implications for research and policy It was found that conditions like dementia, stroke and hip fractures strongly predict the need for long-term care (Chapter 5). Not surprisingly, these conditions are also known to cause significant disability. Currently, long-term care makes up for around 33% of all health care expenditures in the Netherlands. Given how age still plays a large role into explaining long-term care, this share might increase in the coming decades. Investing in prevention of dementia, stroke and hip fractures is therefore an attractive option on paper. However, policy makers should be aware that prevention might not lead to cost-savings in the long run, as living longer means there is a longer time to incur health care expenditures. However, living longer in good health should be a goal, rather than to save health care expenditures (Wouterse et al., 2011). Another good reason would be to relieve pressure on health care personnel in long-term care, which faces enormous pressure with the population aging. The comorbidity cost estimates might give researchers an idea on the scale of cost reduction that might be possible within comorbid patients. There have been reported cases of unnecessary hospitalizations (e.g. drugs that are effective for stroke, but not when diabetes is also present). Thus, treatment can be further optimized to prevent such cases. In other words, the cost estimates provide an idea as to which comorbidities have large consequences for individual health care expenditures, and thus deserve further attention in health care research. However, this will not necessarily lead to a reduction of aggregate health care expenditures. In the Netherlands, hospital expenditures are largely controlled by the global budgets as set by the government. Thus, targeting a cost reduction within comorbid patients should mainly have the purpose of increasing cost-effectiveness, as the resources that are saved here can be spent on other patients. A more direct application of the results may be found in the context of risk equalization in health insurance. In the Netherlands, a health insurer receives ex 241

CHAPTER 12

ante compensation if its clients are relatively unhealthy, when compared to other health insurers. This system was implemented to allow for competition while in the meantime reducing the incentives for risk selection amongst health insurers. Currently, the risk profiles of clients are based on age, sex, socioeconomic status, medication, as well as clinical conditions (Schut and de Wildt, 2011). The risk profile has an additive nature. For every clinical condition a client might have, a fixed amount is added to the total subsidy. As our results indicate, in 7.5% of all possible comorbidities (based on our selection of diseases) there might be additional hospital expenditures (next to the expenditures for each disease separately). In other words, the risk equalization scheme could be improved by including these specific comorbidities in the formula.

III. The interpersonal and -generational distribution of health care expenditures Discussion of findings Large-scale insurance data from the Netherlands indicated that the growth in health care expenditures during 1997-2005 was strongest within elderly, in absolute as well as relative sense. Dutch self-reported data and a Dutch hospital register suggested that the probability of health care consumption also increased over time. This was the case for many sectors (hospital care, specialist care, and dental care). Particularly, we found that advances in technology were statistically associated with a higher probability of hospital admission, and that this association was stronger for higher ages. In other words, elderly benefit more from technology (Chapter 9). A plausible reason can be given for this by looking at the supply side. Let us assume the perspective from a company in the business of medical technology, whose main goal is to increase profits. There is a clear demand for technology, to combat or deal with severe diseases, and this demand is largely unmet. The company focuses on the part of the market where the demand is likely to be highest: the elderly. Note that this might be an unconscious decision through targeting diseases that are most common within elderly. Examples are artificial kidneys, which prolong the life by several years, and beta blockers, which reduce the risk of congestive heart failure. Health care institutions and health insurance companies play their part by acknowledging the great demand under elderly and the importance of meeting this demand, and ultimately, by covering for these costly technologies. Thus, in estimating the health care expenditures associated with age, the interaction between medical innovation, age and health care expenditures should be taken into account. Individual life cycles in health care expenditures were simulated to gain insight into the lifetime dynamics of health care expenditures (Chapter 10 and 11). These 242

GENERAL DISCUSSION

life cycles were simulated for a cohort of newborns in 2005 whilst assuming that health care expenditures and mortality rates remain fixed at the levels of 2005. Though this is not realistic, the resulting numbers give a good indication of the spread in lifetime expenditures. The median lifetime acute health care expenditures was around €84,000 and €105,000 for men and women, respectively. About 73% of these health care expenditures is spent on hospital care. The lifetime expenditures are higher for women than men. This effect is likely to be larger for long-term care (Wong et al., 2008). These differences can be attributed to differences in life expectancy between the sexes. Not only do women incur more expenditures because they are longer at risk, they are also less likely to live with a partner at old ages. The presence of a spouse means that informal care can be given, which acts as a substitution of formal long-term care (Chapter 5). These differences between men and women might become smaller in the future, as the life expectancy of men is likely to increase faster than for women (Polder et al., 2008). However, all of these means are paired with a large standard deviation. Thus, the variation coefficient, which gives the ratio between the standard deviation and the mean, has a value near one for all sectors, which suggests considerable variation between individuals around the mean. Thus, mean expenditures are unlikely to reflect expenditures for any given individual. Implications for research and policy One contribution of this thesis is the methodology for estimating lifetime health care expenditures. Researchers might consider nearest neighbor resampling as a way to extrapolate short observed panels of individual health care expenditures into full life cycles. While this method is not new (it has been applied in other scientific disciplines, such as physics and hydrology), its application to this context is. It has many favorable properties, of which three are of particular interest to analysts of health care expenditures. First, it is almost completely non-parametric, and requires no a priori assumptions on the model structure (e.g. linear dependence of the response variable on covariates, distribution of the response variable). Given the complex nature of health care expenditures over time, this is quite desirable. Second, with enough data available, higher order serial correlation can easily be implemented with this method. Many researchers (e.g. French and Jones, 2004) have mainly restricted their models to first-order correlation, but we found that first-order led to a worse performance in terms of recreating health care expenditures over the long run. Finally, the method easily allows for multivariate extrapolation. It is known that mortality and health care expenditures are strongly correlated (Liu, 2009). This correlation can be taken into account by sampling mortality and health care expenditures simultaneously, conditional on past health care expenditures. 243

CHAPTER 12

In this thesis, we found that the gap between elderly and non-elderly in the probability of using hospital care becomes larger as technology advances more quickly. This might also affect intergenerational solidarity in health care. It is widely known that aging of the population could endanger the solidarity in health care. Because of the growth of the elderly population, large pressure will be exerted on the working non-elderly, who have to contribute more, in terms of net payments (i.e., payments minus the value of care received), to finance the rise in health care utilization that follows. Acceleration of medical innovation might induce higher trends in health care expenditures for elderly, and ultimately, to greater net transfer payments between generations, affecting the inequality in net payments. Policy makers might need to consider that the intergenerational solidarity, as found in a cross-sectional health care financing system (such as health insurance), might be greater in the future. Perceptions of solidarity in health care are often based on annual distributions of health care expenditures. These distributions suggest that there are a lot of individuals with low expenditures, and a select few who have a large share in total health care expenditures. For instance, in 2005 the top 10% men [women] were associated with 72% [64%] of all acute health care expenditures. But policy makers should be aware that this inequality is not necessarily the same as seen over a lifetime. As seen over the lifetime, the top 10% men [women] were associated with 26% [23%] of all acute health care expenditures. While the inequality is still large, it clearly is less skewed than the inequality based on a cross-section. Three explanations can be given for this. First, individuals of all ages are included in a cross-section. When considering a complete life span, the age component is diminished, and the variation in health care expenditures between individuals becomes relatively smaller. Secondly, the last years of life are expensive for many individuals, which will decrease the inequality. Thirdly, health care expenditures have a dynamic character. The risk for suffering a health shock and incurring high health care expenditures becomes larger as one lives longer. We argue that solidarity is best reflected by lifetime expenditures, as a large share of crosssectional expenditures involve the non-elderly lending out money to finance the health care of the elderly, only to be remunerated in the form of health care at some point in the future. In other words, cross-sectional expenditures reflect temporary solidarity to a large degree. Not only do the synthetic life cycles in health care expenditures help in assessing lifetime solidarity in health care, but they also can be used as an aid in determining the feasibility of health savings accounts. In the Netherlands, health care savings accounts have been proposed by the Dutch Council for Public Health and Health Care to potentially reduce moral hazard in health care, as well as increase individual responsibility whilst relieving some of the pressure on the public finances (Jeurissen, 2005). These health care savings accounts implicitly 244

GENERAL DISCUSSION

assume that the variation in lifetime health care expenditures is manageable (in the sense that all individuals have a considerable amount of baseline health care expenditures during lifetime), and that an optimal amount can be saved. In reality, this optimal amount is not known. The life cycles provide some empirical basis for estimates. Furthermore, the results show that the lifetime variation in health care expenditures differs by health care sector. Interpersonal variation in the acute health care sector was found to be moderate. Almost everybody uses hospital care and general practitioner care at some point during their lifetime. The top 25% of all men [women] are associated with 38% [50%] of all general practitioner care expenditures. By contrast, the paramedic care expenditures are more skewed. The top 25% here are associated with 65% [61%] with all expenditures for medical devices. This suggests that the financing of health care could be further separated for each health care sector (in case of the Netherlands, remain separate), as there are differences between sectors in the number of individuals who would save for health care they rarely need, within a health care savings account context. Finally, the results show that health care expenditures not only vary between individuals, but also within individuals. Within a life span health care expenditures exhibit a highly volatile behavior, characterized by large health shocks for many individuals. This variation during a lifetime can only be partially explained by age, as the onset of diseases play a much larger role. Given the large health shocks, it seems entirely possible that individuals are unable to set aside enough money to finance the health care needed to cope with the health shocks. Therefore, it seems an implementation of health savings accounts can only succeed if it is paired with a health insurance system.

The value of using large datasets What binds many of the analyses in this thesis is the use of large-scale datasets. Particularly, the large datasets were absolutely necessary for chapter 7, 8, 10 and 11. Chapters 7 and 8 use a mass-exploratory approach, that requires a lot of data to find new (at least, unknown) comorbidities. Similarly, Chapters 10 and 11 use a non-parametric prediction model that is only efficient when using large datasets. More generally, data mining is an increasingly popular technique amongst all discipline. The advent of statistical techniques, computing power and data collection is what drives this trend. There is no question that more data mining could prove to be useful, and advance health economics as a discipline. Policy makers should be aware of this trend, and invest in data collection (in the area of morbidity, disability and health care expenditures) that will ultimately help gain the knowledge needed for evidence based policy: although knowledge has its price, making ill-informed decisions may be even costlier in the long run. 245

CHAPTER 12

Future Research With increasing availability of data and advances in statistics and computing power, possibilities for new research will open up. Some topics that are worth researching:  Costs of comorbidity, based on a strong causal theory. In this thesis we have looked for the pairs of diseases associated with high health care expenditures, without paying attention to any (lack of) causal links. In the future, these links should be investigated. These types of models should concentrate on pairs for which our exploratory analysis showed that they contributed most to health care expenditures.  Causal models for technology. Some technologies might increase expenditures, while others are cost cutting. At the same time, new technology leads to many health benefits. A framework should be developed for this, to explain the role of technology in expenditures growth and in health.  Generic health care profiles can be deduced from the life cycles developed in Chapters 10 and 11. These health care profiles can be used to further understand typical health care utilization patterns.  The substitution between health care sectors should be investigated more. For instance, to what extent can drugs help prevent a hospital admission, and help to contain costs?  Using the synthetic life cycles, the effect of health care savings accounts on the welfare of the population, as well as the health care cost reductions and loss in national tax income should be modeled. The feasibility of health care savings accounts could be based on such a model.

246

Bibliography Ahmed A, Allman RM, DeLong JF. 2003. Predictors of nursing home admission for older adults hospitalized with heart failure. Archives of Gerontology and Geriatrics 36(2): 117-126. Ahmed A, Lefante CM, Alam N. 2007. Depression and nursing home admission among hospitalized older adults with coronary artery disease: a propensity score analysis. American Journal of Geriatric Cardiology 16(2): 76-83. Alemayehu B, Warner KE. 2004. The lifetime distribution of health care costs. Health Services Research 39(3): 627-642. Ameriks J, Caplin A, Laufer S, Van Nieuwerburgh S. 2011. The joy of giving or assisted living? using strategic surveys to separate public care aversion from bequest motives. Journal of Finance 66(2): 519-561. Anderson TW. 1962. On the distribution of the two-sample Cramér-von Mises criterion. The Annals of Mathematical Statistics 33(3): 1148-1159. Baan CA, Van Baal PH, Jacobs-Van der Bruggen MA, Verkley H, Poos MJ, Hoogenveen RT, Schoemaker CG. 2009. Diabetes mellitus in the Netherlands: estimate of the current disease burden and prognosis for 2025. Nederlands Tijdschrift voor Geneeskunde 153: A580. Bech M, Christiansen T, Khoman E, Lauridsen J, Weale M. 2011. Ageing and health care expenditure in EU-15. European Journal of Health Economics 12(5): 469-78. Beddhu S, Bruns FJ, Saul M, Seddon P, Zeidel ML. 2000. A simple comorbidity scale predicts clinical outcomes and costs in dialysis patients. The American Journal of Medicine 108(8): 609-613. Benjamini Y, Hochberg Y. 1995. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society Series B 57(1): 289-300. Besseling P, Shestalova V. 2010. Forecasting public health expenditures in the Netherlands. CPB Report. Netherlands Bureau for Economic Policy Analysis: The Hague. Beyer K, Goldstein J, Ramakrishnan R, Shaft U. 1999. When is nearest neighbor meaningful? International Conference on Database Theory: 217-235. Bharucha AJ, Pandav R, Shen C, Dodge HH, Ganguli M. 2004. Predictors of nursing facility admission: a 12-year epidemiological study in the United States. Journal of the American Geriatrics Society 52(3): 434-439. Blough DK, Ramsey SD. 2000. Using generalized linear models to assess medical care costs. Health Services and Outcomes Research Methodology 1: 185-202.

BIBLIOGRAPHY

Breeze E, Sloggett A, Fletcher A. 1999. Socioeconomic and demographic predictors of mortality and institutional residence among middle aged and older people: results from the Longitudinal Study. Journal of Epidemiology and Community Health 53(12): 765-774. Breslow NE, Day NE. 1987. Statistical methods in cancer research. Volume II-the design and analysis of cohort studies. IARC Scientific Publications 82: 1-406. Breyer F, Felder S. 2006. Life expectancy and health care expenditures: A new calculation for Germany using the costs of dying. Health Policy 75(2): 178-186. Brouwer WB, Van Exel NJ, Koopmanschap MA, Rutten FF. 1999. The valuation of informal care in economic appraisal. A consideration of individual choice and societal costs of time. International Journal of Technology Assessment in Health Care 15(1): 147-160. Brouwer WB, Van Exel NJ, Van den Berg B, Van den Bos GA, Koopmanschap MA. 2005. Process utility from providing informal care: the benefit of caring. Health Policy 74(1): 85-99. Brown JR, Finkelstein A. 2008. The interaction of public and private insurance: Medicaid and the long-term care insurance market. American Economic Review 98(3): 10831102. Buchner F, Wasem J. 2006. “Steeping” of health expenditure profiles. Geneva Papers on Risk and Insurance: Issues and Practice 31(4): 581-599. Buhlmann P. 2002. Bootstraps for time series. Statistical Science 17(1): 52-72. Buishand TA, Brandsma T. 2001. Multisite simulation of daily precipitation and temperature in the Rhine basin by nearest-neighbor resampling. Water Resources Research 37(11): 2761-2776. Buishand TA. 2007. Estimation of a large quantile of the distribution of multi-day seasonal maximum rainfall: the value of stochastic simulation of long-duration sequences. Climate Research 34(3): 185-194. Buntin MJB, Escarce JJ, Goldman D, Kau H, Laugesen MJ, Shekelle P. 2004. Increased Medicare Expenditures for Physician Services: What are the causes? Inquiry 41(1): 83-94. Cai Q, Salmon JW, Rodgers ME. 2009. Factors associated with long-stay nursing home admissions among the U.S. elderly population: comparison of logistic regression and the Cox proportional hazards model with policy implications for social work. Social Work in Health Care 48(2): 154-168. Cameron AC, Trivedi PK. 1998. Regression analysis of count data. Cambridge University Press: Cambridge. Campbell SE, Seymour DG, Primrose WR. 2004. A systematic literature review of factors affecting outcome in older medical patients admitted to hospital. Age and Ageing 33(2): 110-115. Carstensen B, Kristensen JK, Ottosen P, Borch-Johnsen K. 2008. The danish national diabetes register: Trends in incidence, prevalence and mortality. Diabetologia 51(12): 2187-2196. Casdagli M. 1989. Nonlinear prediction of chaotic time series. Physica D 35(3): 335-356. Casdagli M, Eubank S, Farmer JD, Gibson J. 1991. State space reconstruction in the presence of noise. Physica D 51: 52-98. Casdagli M. 1992. Chaos and deterministic versus stochastic non-linear modelling. Journal Royal Statistical Society Series B 54(2): 303-328.

248

BIBLIOGRAPHY

Charlson ME, Pompei P, Ales KL, MacKenzie CR. 1987. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. Journal of Chronic Diseases 40(5): 373-383. Charlson ME, Charlson RE, Peterson JC, Marinopoulos SS, Briggs WM, Hollenberg JP. 2008. The Charlson comorbidity index is adapted to predict costs of chronic disease in primary care patients. Journal of Clinical Epidemiology 61(12): 1234-1240. Christensen K, Doblhammer G, Rau R, Vaupel JW. 2009. Ageing populations: the challenges ahead. Lancet 374(9696): 1196-1208. Christiansen T, Bech M, Lauridsen J, Nielsen P. 2006. Demographic changes and aggregate health care expenditure in Europe. ENEPRI Research Report 32. ENEPRI: Brussels. College voor Zorgverzekeringen. 2006. Rapport richtlijnen voor farmaco-economisch onderzoek; evaluatie en actualisatie. College voor Zorgverzekeringen: Diemen. Cornell JE, Pugh JA, Williams JW Jr, Kazis L, Lee AFS, Parchman ML, Zeber J, Pederson T, Montgomery KA, Noël PH. 2008. Multimorbidity clusters: clustering binary data from multimorbidity clusters: clustering binary data from a large administrative medical database. Applied Multivariate Research 12(3): 163-182. Crimmins EM. Trends in the health of the elderly. 2004. Annual Review of Public Health 25: 79-98. Crimmins EM, Beltrán-Sánchez H. 2011. Mortality and morbidity trends: is there compression of morbidity? The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 66(1): 75-86. Crom B. 2005. De invloed van externe budgetparameters op de interne budgettering van academische ziekenhuizen: Verklaringen voor verschillen in budgetteringssystemen en hun effecten. PhD-thesis. University of Groningen: Groningen. Cui J. 2007. QIC program and model selection in GEE analyses. Stata Journal 7(2): 209220. Cutler D, Deaton A, Lleras-Muney A. 2006a. The determinants of mortality. The Journal of Economic Perspectives 20(3): 97-120. Cutler DM, Rosen AB, Vijan S. 2006b. The value of medical spending in the united states, 1960-2000. New England Journal of Medicine 355(9):920-927. Davison AC, Hinkley DV, Young GA. 2003. Recent developments in bootstrap methodology. Statistical Science 18(2): 141-157. De Bruin A, De Kardaun JWPF, Gast F, Bruin EI, De Sijl M, Van Verweij GCG. 2004. Record linkage of hospital discharge register with population register: experiences at Statistics Netherlands. Statistical Journal of the United Nations ECE 21: 23-32. De Kok IM, Polder JJ, Habbema JD, Berkers LM, Meerding WJ, Rebolj M, Van Ballegooijen M. 2009. The impact of healthcare costs in the last year of life and in all life years gained on the cost-effectiveness of cancer screening. British Journal of Cancer 100(8): 1240-1244. De Meijer C, Koopmanschap M, Bago d’ Uva T, Doorslaer E. 2011. Determinants of long-term care spending: Age, time to death or disability? Journal of Health Economics 30(2): 425-438. De Nardi M, French E, Jones JB. 2010. Why do the elderly save medical expenses? Journal of Political Economy 118(1): 39-75.

249

BIBLIOGRAPHY

De Pablo P, Losina E, Phillips CB, Fossel AH, Mahomed N, Lingard EA, N Katz J. 2004. Determinants of discharge destination following elective total hip replacement. Arthritis & Rheumatism 51(6): 1009-1017. De Waegenaere A, Melenberg B, Stevens R. 2010. Longevity risk. De Economist 158: 151-192. Deb P, Trivedi PK. 1997. Demand for medical care by the elderly: a finite mixture approach. Journal of Applied Econometrics 12(3): 313-326. Deb P, Holmes AM. 2000. Estimates of use and costs of behavioral health care: a comparison of standard and finite mixture models. Health Economics 9(6): 475489. Deeg DJ, Van Tilburg T, Smit JH, De Leeuw ED. 2002. Attrition in the Longitudinal Aging Study Amsterdam. The effect of differential inclusion in side studies. Journal of Clinical Epidemiology 55(4): 319-328. DerSimonian R, Laird N. 1986. Meta-analysis in clinical trials. Controlled Clinical Trials 7(3): 177-188. Deyo RA, Cherkin DC, Ciol MA. 1992. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. Journal of Clinical Epidemiology 45: 613619. DG-ECFIN. 2009. The 2009 ageing report: economic and budgetary projections for the EU-27 member states (2008-2060). Office for Official Publications of the European Communities: Luxembourg. Dormont B, Grignon M, Huber H. 2006. Health expenditure growth: reassessing the threat of ageing. Health Economics 15: 947-963. Dow WH, Norton EC. 2003. Choosing between and interpreting the Heckit and twopart models for corner solutions. Health Services and Outcomes Research Methodology 4: 5-18. Dunlop DD. 1994. Regression for longitudinal data: a bridge from least squares regression. The American Statistician 48: 299-303. Dutch Hospital Data. 2009. Available at: http://www.dutchhospitaldata.nl/category/lmr_en_lbz/. Engelfriet PM, Hoogenveen RT, Boshuizen HC, van Baal PHM. 2011. To die with or from heart failure: A difference that counts. European Journal of Heart Failure 13(4): 377-383. Farmer JD, Sidorowich JJ. 1987. Predicting chaotic time series. Physical Review Letters 59(8): 845-848. Feenstra TL, Van Baal PH, Gandjour A, Brouwer WB. 2008. Future costs in economic evaluation: a comment on Lee. Journal of Health Economics 27(6): 1645-1649. Feigin VL, Lawes CMM, Bennett DA, Anderson CS. 2003. Stroke epidemiology: A review of population-based studies of incidence, prevalence, and case-fatality in the late 20th century. The Lancet Neurology 2(1): 43-53. Felder S, Werblow, A. 2008. Does the age profile of health care expenditure really steepen over time? New evidence from Swiss cantons. Geneva Papers on Risk and Insurance: Issues and Practice 33: 710-727. Felder S, Werblow A, Zweifel P. 2010. Do red herrings swim in circles? Controlling for the endogeneity of time to death. Journal of Health Economics 29(2): 205-212. Field MJ, Jette AM, eds. 2007. The future of disability in America. The National Academies Press: Washington, DC.

250

BIBLIOGRAPHY

Fitzmaurice GM, Laird NM, Rotnitzky AG. 1993. Regression models for discrete longitudinal responses. Statistical Science 8: 284-309. Fleishman JA, Cohen JW. 2010. Using information on clinical conditions to predict highcost patients. Health Services Research 45(2): 532-552. Fogel R. 2004. The escape from hunger and premature death, 1700-2100: Europe, America, and the Third World. Cambridge University Press: New York. Folmer K, Douven R, Van Gameren E, Mannaerts H, Mot E, Ooms I, Westerhout E., Woittiez I. 2006. Zorg in model. Netherlands Bureau for Economic Policy Analysis memorandum 146. Netherlands Bureau for Economic Policy Analysis: The Hague. Folmer C, Westerhout E. 2008. Financing medical specialist services in The Netherlands: Welfare implications of imperfect agency. Economic Modelling 25(5): 946-958. Forget EL, Roos LL, Deber RB, Walld R. 2008. Variations in lifetime healthcare costs across a Population. Healthcare Policy 4(1): 148-167. Freedman VA. 1996. Family structure and the risk of nursing home admission. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 51(2): 61-69. Freedman VA, Martin LG. 2000. Contribution of chronic conditions to aggregate changes in old-age functioning. American Journal of Public Health 90(11): 17551760. Freedman VA, Schoeni RF, Martin LG, Cornman JC. 2007. Chronic conditions and the decline in late-life disability. Demography 44(3): 459-477. French E, Jones JB. 2004. On the distribution and dynamics of health care costs. Journal of Applied Econometrics 19: 705-721. Fried LP, Bandeen-Roche K, Kasper JD, Guralnik JM. 1999. Association of comorbidity with disability in older women: the Women’s Health and Aging Study. Journal of Clinical Epidemiology 52: 27-37. Fried LP, Ferrucci L, Darer J, Williamson JD, Anderson G. 2004. Untangling the concepts of disability, frailty, and comorbidity: implications for improved targeting and care. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences 59: 255-263. Fries JF. 1980. Aging, natural death and the compression of morbidity. New England Journal of Medicine 303(3): 130-135. Fuchs Z, Blumstein T, Novikov I, Walter-Ginzburg A, Lyanders M, Gindin J, Habot B, Modan B. 1998. Morbidity, comorbidity, and their association with disability among community-dwelling oldest-old in Israel. Journals of Gerontology Series A: Biological Sciences and Medical Sciences 53(6): 447-455. Gandjour A, Lauterbach KW. 2005. Does prevention save costs? Considering deferral of the expensive last year of life. Journal of Health Economics 24(4): 715-724. Garber AM, Phelps CE. 1997. Economic foundations of costeffectiveness analysis. Journal of Health Economics 16(1): 1-31. Gaugler JE, Duval S, Anderson KA, Kane RL. 2007. Predicting nursing home admission in the U.S: a meta-analysis. BMC Geriatrics 7: 13. Gerdtham UG, Sögaard J, MacFarlan M, Oxley H. 1998. The determinants of health expenditure in the OECD countries. Health, the Medical Profession, and Regulation. Kluwer: Dordrecht.

251

BIBLIOGRAPHY

Gerdtham UG, Jönsson B. 2000. International comparisons of health expenditure: Theory, data and econometric analysis. Handbook of Health Economics 1: 11-53. Elsevier: North-Holland. Getzen TE. 2000. Health care is an individual necessity and a national luxury: applying multilevel decision models to the analysis of health care expenditures. Journal of Health Economics 19(2): 259-270. Getzen TE. 2001. Aging and health care expenditures: A comment on Zweifel, Felder and Meiers. Health Economics 10(2): 175-177. Getzen TE. 2006. Aggregation and the measurement of health care costs. Health Services Research 41(5): 1938-1954. Gijsen R, Hoeymans N, Schellevis FG, Ruwaard D, Satariano WA, Van den Bos GA. 2001. Causes and consequences of comorbidity: a review. Journal of Clinical Epidemiology 54(7):661-674. Glader EL, Stegmayr B, Norrving B, Terént A, Hulter-Asberg K, Wester PO, Asplund K; Riks-Stroke Collaboration. 2003. Sex differences in management and outcome after stroke: a Swedish national perspective. Stroke 34(8): 1970-1975. Goldacre M, Kurina L, Yeates D, Seagroatt V, Gill L. 2000. Use of large medical databases to study associations between diseases. QJM 93:669-675. Gray A. 2005. Population ageing and healthcare expenditure. Ageing Horizons 2: 15-20. Greene VL, Ondrich JI. 1990. Risk factors for nursing home admissions and exits: a discrete-time hazard function approach. Journal of Gerontology 45(6): 250-258. Greene WH. 2008. Econometric Analysis. Prentice Hall: New York. Grundy E, Glaser K. 1997. Trends in, and transitions to, institutional residence among older people in England and Wales, 1971-91. Journal of Epidemiology and Community Health 51(5): 531-540. Häkkinen U, Martikainen P, Noro A, Nihtilä E, Peltola M. 2008. Aging, health expenditure, proximity to death, and income in Finland. Health Economics, Policy and Law 3: 165-195. Hall RE, Jones CI. 2007. The value of life and the rise in health spending. The Quarterly Journal of Economics 122(1): 39-72. Hancock R, Arthur A, Jagger C, Matthews R. 2002. The effect of older people's economic resources on care home entry under the United Kingdom's long-term care financing system. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 57(5): 285-293. Härdle W, Horowitz J, Kreiss JP. 2003. Bootstrap methods for time series. International Statistical Review 71(2): 435-459. Harrel FE. 2001. Regression modeling strategies: with applications to linear models, logistic regression and survival analysis. Springer-Verlag: New York. Harris Y, Cooper JK. 2006. Depressive symptoms in older people predict nursing home admission. Journal of the American Geriatrics Society 54(4):593-597. Harris Y. 2007. Depression as a risk factor for nursing home admission among older individuals. Journal of the American Medical Directors Association 8(1): 14-20. Hastie T, Tibshirani R, Friedman J. 2009. The elements of statistical learning: data mining, inference, and prediction. Second Edition. Spring-Verlag: New York. Heijink R, Noethen M, Renaud T, Koopmanschap M, Polder J. 2008. Cost of illness: an international comparison. Australia, Canada, France, Germany and the Netherlands. Health Policy 88(1): 49-61.

252

BIBLIOGRAPHY

Himes CL, Wagner GG, Wolf DA, Aykan H, Dougherty DD. 2000. Nursing home entry in Germany and the United States. Journal of Cross-Cultural Gerontology 15(2): 99118. Hoeymans N, Wong A, Van Gool CH, Deeg DJH, Nusselder WJ, De Klerk MMY, Van Boxtel MPJ, Picavet HSJ. 2012. Are diseases becoming less disabling? A study of five Dutch surveys on trends in diseases, activity limitations and their interrelationships. American Journal of Public Health 102(1): 163-170. Hogan C, Lunney J, Gabel J, Lynn J. 2001. Medicare beneficiaries’ costs of care in the last year of life. Health Affairs 20: 188-195. Hoogendijk E, Broese Van Groenou M, Van Tilburg T, Deeg D. 2008. Educational differences in functional limitations: comparisons of 55-65-year-olds in the Netherlands in 1992 and 2002. International Journal of Public Health 53(6): 281289. Hosmer DW, Lemeshow S. 2000. Applied logistic regression. Second edition. Wiley: New York. Hseih D. 1991. Chaos and nonlinear dynamics: application to financial markets. Journal of Finance 46(5): 1839-1877. Hu FB, Goldberg J, Hedeker G, Flay BR, Pentz MA. 1998. Comparison of populationaveraged and subject specific approaches for analyzing repeated binary outcomes. American Journal of Epidemiology 147: 694-703. Jagger C, Matthews RJ, Matthews FE, Spiers NA, Nickson J, Paykel ES, Huppert FA, Brayne C; Medical Research Council Cognitive Function and Ageing Study (MRCCFAS). 2007. Cohort differences in disease and disability in the young-old: findings from the MRC Cognitive Function and Ageing Study (MRC-CFAS). BMC Public Health 7: 156. Jeune B, Brønnum-Hansen H. 2008. Trends in health expectancy at age 65 for various health indicators, 1987-2005, Denmark. European Journal of Ageing 5(4): 279-285. Jeurissen P. 2005. Houdbare solidariteit in de gezondheidszorg. Government Report. Council for Public Health and Health Care. John R, Kerby DS, Hennessy CH. 2003. Patterns and impact of comorbidity and multimorbidity among community-resident American Indian elders. Gerontologist 43: 649-660. Jolles J, Houx PJ, Van Boxtel MP, et al, eds. 1995. Maastricht Aging Study: determinants of cognitive aging. Maastricht, the Netherlands: Neuropsychology Publishers. Jones CI. 2002. Why have health expenditure as a share of GDP risen so much. NBER working paper 9325. NBER: Cambridge. Knecht JH, Oomen PGM. 2006. Lifesciences & Gezondheid: Trends in octrooiaanvragen in de Medische Biowetenschappen. Dutch Patent Office Research Report. Dutch Patent Office: Rijswijk. Koissi MC, Shapiro AF, Hognas G. 2006. Evaluating and extending the lee-carter model for mortality forecasting: Bootstrap confidence interval. Insurance: Mathematics and Economics 38(1): 1-20. Kommer GJ, Wong A, Slobbe LCJ. 2010. Determinanten van de volumegroei in de zorg. RIVM-briefrapport 270751021. RIVM: Bilthoven. Koopmanschap M, De Meijer C, Wouterse B, Polder JJ. 2010. Determinants of health care expenditures in an aging society. Panel Paper 22. Netspar: Tilburg.

253

BIBLIOGRAPHY

Kunst A, Meerding WJ, Varenik N, Polder JJ, Mackenbach JP. 2007. Sociale verschillen in zorggebruik en zorgkosten in Nederland 2003. Zorg voor Euro’s 5. RIVM: Bilthoven. Kuo TC, Zhao Y, Weir S, Kramer MS, Ash AS. 2008. Implications of comorbidity on costs for patients with Alzheimer disease. Medical Care 46(8): 839-846. Kuo RN, Lai MS. 2010. Comparison of Rx-defined morbidity groups and diagnosisbased risk adjusters for predicting healthcare costs in Taiwan. BMC Health Services Research 10: 126. Lall U, Sharma A. 1996. A nearest neighbor bootstrap for resampling hydrologic time series. Water Resource Research 32: 679-693. Lee RD, Carter LR. 1992. Modeling and forecasting US mortality. Journal of the American Statistical Association 87(419): 659-671. Lee R. 2000. The lee-carter method for forecasting mortality, with various extensions and applications. North American Actuarial Journal 4(1): 80-93. Lee RH. 2008. Future costs in cost effectiveness analysis. Journal of Health Economics 27(4): 809-818. Li Q, Racine JS. 2003. Nonparametric estimation of distributions with categorical and continuous data. Journal of Multivariate Analysis: 86: 266-292. Liang KY, Zeger SL. 1986. Longitudinal data analysis using generalized linear models. Biometrika 73: 13-22. Librero J, Peiro S, Ordinana R. 1999. Chronic comorbidity and outcomes of hospital care: length of stay, mortality, and readmission at 30 and 365 days. Journal of Clinical Epidemiology 52: 171-179. Liu L. 2009. Joint modeling longitudinal semi-continuous data and survival, with application to longitudinal medical cost data. Statistics in Medicine 28: 972-986. Long MJ, Marshall BS. 2000. The relationship of impeding death and age category to treatment intensity in the elderly. Journal of Evaluation in Clinical Practice 6: 63-70. Lubitz J, Beebe J, Baker C. 1995. Longevity and medical expenditures. New England Journal of Medicine 332: 999-1003. Lubitz J, Cai L, Kramarow E, Lentzner H. 2003. Health, life expectancy, and health care spending among the elderly - three decades of health care use by the elderly, 19651998. New England Journal of Medicine 349(11): 1048-1055. Madsen J, Serup-Hansen N, Kristiansen IS. 2002. Future health care costs--do health care costs during the last year of life matter? Health Policy 62(2): 161-172. Majer IM, Nusselder WJ, Mackenbach JP, Klijs B, van Baal PHM. 2011. Mortality risk associated with disability: A population-based record linkage study. American Journal of Public Health 101(12): 9-15. Manning WG, Mullahy J. 2001. Estimating log models: to transform or not to transform? Journal of Health Economics 20(4): 461-494. Manning WG, Basu A, Mullahy J. 2005. Generalized modeling approaches to risk adjustment of skewed outcomes data. Journal of Health Economics 24: 465-488. Manns B, Meltzer D, Taub K, Donaldson C. 2003. Illustrating the impact of including future costs in economic evaluations: an application to end-stage renal disease care. Health Economics 12(11): 949-958. Marquardt DW. 1970. Generalized inverses, ridge regression, biased linear estimation, and nonlinear estimation. Technometrics 12: 591-612.

254

BIBLIOGRAPHY

Martin LG, Freedman VA, Schoeni RF, Andreski PM. 2009. Health and functioning among baby boomers approaching 60. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 64(3): 369-377. Martin LG, Freedman VA, Schoeni RF, Andreski PM. 2010. Trends in disability and related chronic conditions among people ages fifty to sixty-four. Health Affairs 29(4): 725-731. McCullagh P, Nelder JA. 1989. Generalized linear models. Second Edition. Chapman & Hall: London. McGrail K, Green B, Barer M, Evans R, Hertzman C, Normand C. 2000. Age, costs of acute and long-term care and proximity to death: evidence for 1987-88 and 1994-95 in British Columbia. Age and Ageing 29: 249-253. McGuire TG, Pauly MV. 1991. Physician Responses to Fee Changes with Multiple Payers. Journal of Health Economics 10(4): 385-410. McWhinnie JR. 1981. Disability assessment in population surveys: results of the OECD common development effort. Revue d Épidémiologie et de Santé Publique 29(4): 413-419. Meara E, White C, Cutler DM. 2004. Trends in medical spending by age, 1963-2000. Health Affairs 23(4): 176-83. Meerding WJ, Bonneux L, Polder JJ, Koopmanschap MA, Van der Maas PJ. 1998. Demographic and epidemiological determinants of healthcare costs in Netherlands: cost of illness study. British Medical Journal 317(7151): 111-115. Meltzer D, Egleston B, Stoffel D, Dasbach E. 2000. Effect of future costs on costeffectiveness of medical interventions among young adults: the example of intensive therapy for type 1 diabetes mellitus. Medical Care 38(6): 679-685. Meltzer D. 2008. Response to ‘Future costs and the future of cost-effectiveness analysis’. Journal of Health Economics 27(4): 822-825. Miller EA, Weissert WG. 2000. Predicting elderly people's risk for nursing home placement, hospitalization, functional impairment, and mortality: a synthesis. Medical Care Research and Review 57(3): 259-297. Ministry of Health Welfare and Sports. 2009. Long-term care in the Netherlands - The Exceptional Medical Expenses Act. Available from: http://www.minvws.nl/en/folders/lz/2009/host-country-report.asp Ministry of Welfare, Health and Sport. 2010. Health insurance system. Accessible at http://english.minvws.nl/en/themes/health-insurance-system/default.asp. Molenberghs G, Verbeke G. 2000. Models for discrete longitudinal data. Springer-Verlag: New York. Mullahy J. 1998. Much ado about two: reconsidering retransformation and the two-part model in health econometrics. Journal of Health Economics 17: 247-281. Murabito JM, Pencina MJ, Zhu L, Kelly-Hayes M, Shrader P, D’Agostino RB Sr. 2008. Temporal trends in self-reported functional limitations and physical disability among the community-dwelling elderly population: the Framingham Heart Study. American Journal of Public Health 98(7): 1256-1262. Mustard C, Finlayson M, Derksen S, Berthelot JM. 1999. What determines the need for nursing home admission in a universally insured population? Journal of Health Services Research & Policy 4(4): 197-203. National Institute for Health and Clinical Excellence. 2008. Guide to the methods of technology appraisal. NICE: London.

255

BIBLIOGRAPHY

National Public Health Compass. 2009. Available from: http://www.rivm.nl/vtv/object_document/o2308n18838.html Nelder JA, Mead R. 1965. A simplex algorithm for function minimization. Computer Journal 7(4): 308-313. Newhouse JP. 1992. Medical care costs: how much welfare loss? Journal of Economic Perspectives 6(3): 3-21. Newson R. 2000. sg151: B-splines and splines parameterized by their values at reference points on the X-axis. Stata Technical Bulletin 57: 20-27. Nieboer A, Stolk E, Koolman X. 2006. Toekomstvisie langdurige zorg en ondersteuning vanuit burgerperspectief. Report No.: 06.79. Erasmus MC: Rotterdam. Nuyen J, Schellevis FG, Satariano WA, Spreeuwenberg PM, Birkner MD, Van den Bos GA, Groenewegen PP. 2006. Comorbidity was associated with neurologic and psychiatric diseases: a general practice-based controlled study. Journal of Clinical Epidemiology 59: 1274-1284. Nuyen J, Spreeuwenberg PM, Groenewegen PP, Van den Bos GA, Schellevis FG. 2008. Impact of preexisting depression on length of stay and discharge destination among patients hospitalized for acute stroke: linked register-based study. Stroke 39(1): 132138. Nyman JA. 2004. Should the consumption of survivors be included as a cost in costutility analysis? Health Economics 13(5): 417-427. OECD. 2006. OECD economic outlook 79: Analyses and projections. OECD: Paris. OECD. 2010. OECD Health data. Available at http://www.oecd.org/health/healthdata. Oeppen J, Vaupel JW. 2002. Broken limits to life expectancy. Science 296(5570): 10291031. Oh EH, Imanaka Y, Evans E. 2005. Determinants of the diffusion of computed tomography and magnetic resonance imaging. International Journal of Technology Assessment in Health Care 21(1): 73-80. Ohwaki K, Hashimoto H, Sato M, Tokuda H, Yano E. 2005. Gender and family composition related to discharge destination and length of hospital stay after acute stroke. Tohoku Journal of Experimental Medicine 207(4): 325-332. Okunade AA, Murthy VNR. 2002. Technology as a ‘major driver’ of health care costs: a cointegration analysis of the Newhouse conjecture, Journal of Health Economics 21(1): 147-159. Onder G, Liperoti R, Soldato M, Cipriani MC, Bernabei R, Landi F. 2007. Depression and risk of nursing home admission among older adults in home care in Europe: results from the Aged in Home Care (AdHOC) study. Journal of Clinical Psychiatry 68(9): 1392-1398. O'Neill C, Groom L, Avery AJ, Boot D, Thornhill K. 2000. Age and proximity to death as predictors of GP care costs: results from a study of nursing home patients. Health Economics 9(8): 733-738. Orosz E, Morgan D. 2004. SHA-based national health accounts in thirteen OECD countries: a comparative analysis. OECD health working papers no. 16. OECD: Paris. Ottenbacher KJ, Campbell J, Kuo YF, Deutsch A, Ostir GV, Granger CV. 2008. Racial and ethnic differences in postacute rehabilitation outcomes after stroke in the United States. Stroke 39(5):1514-1519.

256

BIBLIOGRAPHY

Pan W. 2001. Akaike’s information criterion in generalized estimating equations. Biometrics 57(1): 120-125. Parker MG, Ahacic K, Thorslund M. 2005. Health changes among Swedish oldest old: prevalence rates from 1992 and 2002 show increasing health problems. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences 60(10): 1351-1355. Parker MG, Thorslund M. 2007. Health trends in the elderly population: getting better and getting worse. Gerontologist 47(2): 150-158. Payne G, Laporte A, Deber R, Coyte PC. 2007. Counting backward to health care's future: Using time-to-death modeling to identify changes in end-of-life morbidity and the impact of aging on health care expenditures. Milbank Quarterly 85(2): 213257. Peden EA, Freeland M. 1998. Insurance effects on US Medical Spending 1960-1993. Health Economics 7(8): 671-687. Peijnenburg K. 2011. Consumption, savings, and investments over the life cycle. Tilburg University: Tilburg. Pettitt AN. 1976. A two-sample Anderson-Darling rank statistic. Biometrika 63(1): 161168. Picavet HS, Hoeymans N. 2002. Physical disability in The Netherlands: prevalence, risk groups and time trends. Public Health 116(4): 231-237. Pischke JS, Velling J. 1997. Employment effects of immigration to Germany: an analysis based on local labor markets. Review of Economics and Statistics 79(4): 594-604. Pitacco E, Denuit M, Haberman S, Olivieri A. 2008. Modeling longevity dynamics for pensions and annuity business. Oxford University Press: USA.. Polder JJ, Meerding WJ, Koopmanschap MA, Bonneux L, Van der Maas PJ. 1998. The cost of sickness in the Netherlands in 1994: the main determinants were advanced age and disabling conditions. Nederlands Tijdschrift voor Geneeskunde 142(28): 1607-1611. Polder JJ, Bonneux L, Meerding WJ, Van der Maas PJ. 2002. Age-specific increases in health care costs. European Journal of Public Health 12(1): 57-62. Polder JJ, Barendregt JJ, Van Oers H. 2006. Health care costs in the last year of life - the Dutch experience. Social Science & Medicine 63(7): 1720-1731. Polder JJ, Wong A, Schols JMGA. 2008. Toename levensverwachting remt uitgaven ouderenzorg. Economisch Statistische Berichten 93(4539): 422-425. Politis DN. 2003. The impact of bootstrap methods on time series analysis. Statistical Science 18(2): 219-230 Poos MJJC, Smit JM, Groen J, Kommer GJ, Slobbe LCJ. 2008. Kosten van Ziekten in Nederland 2005. Accessible at: http://www.kostenvanziekten.nl. Portrait F, Lindeboom F, Deeg D. 2000. The use of long-term care services by the Dutch elderly. Health Economics 9: 513-531. Prismant. 2008. Available from: http://www.prismant.nl/. Puts MT, Deeg DJ, Hoeymans N, Nusselder WJ, Schellevis FG. 2008. Changes in the prevalence of chronic disease and the association with disability in the older Dutch population between 1987 and 2001. Age and Ageing 37(2): 187-193. Rajagopalan B, Lall U. 1999. A k-nearest-neighbor simulator for daily precipitation and other variables. Water Resources Research 35: 3089-3101.

257

BIBLIOGRAPHY

Rappange DR, Van Baal PH, Van Exel NJ, Feenstra TL, Rutten FF, Brouwer WB. 2008. Unrelated medical costs in life-years gained: should they be included in economic evaluations of healthcare interventions? Pharmacoeconomics 26(10): 815-830. Rasmusen E. Moral Hazard in Risk-Averse Teams. 1987. RAND Journal of Economics 18(3): 428-435. Reijneveld SA, Spijker J, Dijkshoorn H. 2007. Katz’ ADL index assessed functional performance of Turkish, Moroccan, and Dutch elderly. Journal of Clinical Epidemiology 60(4): 382-388. Reinhardt U. 2003. Does The Aging Of The Population Really Drive The Demand For Health Care? Health Affairs 22(6): 27-39. Reuben DB, Valle LA, Hays RD, Siu AL. 1995. Measuring physical function in community-dwelling older persons: a comparison of self-administered, intervieweradministered, and performance-based measures. Journal of the American Geriatrics Society 43(1): 17-23. Rijken M, Van Kerkhof M, Dekker J, Schellevis FG. 2005. Comorbidity of chronic diseases: effects of disease pairs on physical and mental functioning. Quality of Life Research 14: 45-55. Rizzo JA, Blumenthal D. 1994. Physician labor supply: do income effects matter? Journal of Health Economics 13(4): 433-453. Roe CJ, Kulinskaya E, Dodich N, Adam WR. 1998. Comorbidities and prediction of length of hospital stay. Australian & New Zealand Journal of Medicine 28(6): 811815. Rogers WH. 1993. Regression standard errors in clustered samples. Stata Technical Bulletin 13: 19-23. Romano PS, Roos LL, Jollis JG. 1993. Adapting a clinical comorbidity index for use with ICD-9-CM administrative data: differing perspectives. Journal of Clinical Epidemiology 46: 1075-1079. Rubin DB. 1987. Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons: New York. Rundek T, Mast H, Hartmann A, Boden-Albala B, Lennihan L, Lin IF, Paik MC, Sacco RL. 2000. Predictors of resource use after acute hospitalization: the Northern Manhattan Stroke Study. Neurology 55(8): 1180-1187. Salas C, Raftery JP. 2001. Econometric issues in testing the age neutrality of health care expenditure. Health Economics 10: 669-671. Salive ME, Collins KS, Foley DJ, George LK. 1993. Predictors of nursing home admission in a biracial population. American Journal of Public Health 83(12): 17651767. Schoeni RF, Freedman VA, Martin LG. 2008. Why is late-life disability declining? Milbank Quarterly 86(1): 47-89. Scholz J, Seshadri A, Khitatrakun S. 2006. Are Americans saving "optimally" for retirement? Journal of Political Economy 114(4): 607-643. Schram MT, Frijters D, Van de Lisdonk EH, Ploemacher J, De Craen AJ, De Waal MW, Van Rooij FJ, Heeringa J, Hofman A, Deeg DJ, Schellevis FG. Setting and registry characteristics affect the prevalence and nature of multimorbidity in the elderly. Journal of Clinical Epidemiology 61: 1104-1112. Schulz R, Beach SR. 1999. Caregiving as a risk factor for mortality: the Caregiver Health Effects Study. Journal of American Medical Association 282(23): 2215-2219.

258

BIBLIOGRAPHY

Schut E, De Wildt JE. 2011. Risicoverevening: hoe zit het precies? De eerstelijns 4. Schwartz M, Iezzoni LI, Moskowitz MA, Ash AS, Sawitz E. 1996. The importance of comorbidities in explaining differences in patient costs. Medical Care 34: 767-782. Seshamani M, Gray AM. 2004a. Ageing and health care expenditure: the red herring argument revisited. Health Economics 13: 303-314. Seshamani M, Gray AM. 2004b. A longitudinal study of the effects of age and time to death on hospital costs. Journal of Health Economics 23: 217-235. Seshamani M, Gray AM. 2004c. Time to death and health expenditure: an improved model for the impact of demographic change on health care costs. Age and Ageing 33(6): 556-561. Shang B, Goldman D. 2008. Does age or life expectancy better predict health care expenditures? Health Economics 17(4): 487-502. Shmueli G. 2010. To explain or to predict? Statistical Science 25(3): 289-310. Shmueli G, Koppius O. 2011. Predictive analytics in information systems research. MIS Quarterly 35(3): 553-572. Shwartz M, Iezzoni LI, Moskowitz MA, Ash AS, Sawitz E. 1996. The importance of comorbidities in explaining differences in patient costs. Medical Care 34(8): 767782. Slobbe LCJ, Kommer GJ, Smit JM, Groen J, Meerding WJ, Polder JJ. 2006. Costs of Illnesses in the Netherlands 2003: Zorg voor euro’s - 1. RIVM: Bilthoven. Slobbe LCJ, Smit JM, Groen J, Kommer GJ, Poos MJJC. 2011. Kosten van Ziekten in Nederland 2007. RIVM: Bilthoven. Southern DA, Quan H, Ghali WA. 2004. Comparison of the Elixhauser and Charlson/Deyo methods of comorbidity measurement in administrative data. Medical Care 42: 355-360. Spillman BC, Lubitz J. 2000. The effect of longevity on spending for acute and long-term care. New England Journal of Medicine 342: 1409-1415. Stansbury JP, Jia H, Williams LS, Vogel WB, Duncan PW. 2005. Ethnic disparities in stroke: epidemiology, acute care, and postacute outcomes. Stroke 36(2): 374-386. Stasinopoulos DM, Rigby RA. 2007. Generalized additive models for location scale and shape (GAMLSS) in R. Journal of Statistical Software 23: 7. Statacorp. 2007. Stata Statistical Software Release 10 Reference I-P: 315-334. StataCorp LP: Texas. Statistics Netherlands. 1996. Netherlands Health Interview Survey 1981-1995. SDU Publishers/CBS-Publications: The Hague. Statistics Netherlands. 2010. Statline. Available from: http://www.statline.nl. Stearns SC, Norton EC. 2004. Time to include time to death? The future of health care expenditure predictions. Health Economics 13: 315-327. Stoop I. 2000. Covering the population: sample surveys at the social and cultural planning office. Kwantitatieve Methoden 21: 69-84. Struijs JN, Baan CA, Schellevis FG, Westert GP, Van den Bos GA. 2006. Comorbidity in patients with diabetes mellitus: impact on medical health care utilization. BMC Health Services Research 6: 84. Stuck AE, Walthert JM, Nikolaus T, Büla CJ, Hohmann C, Beck JC. 1999. Risk factors for functional status decline in community-living elderly people: a systematic literature review. Social Science and Medicine 48: 445-469.

259

BIBLIOGRAPHY

Sundararajan V, Henderson T, Perry C, Muggivan A, Quan H, Ghali WA. 2004. New ICD-10 version of the Charlson comorbidity index predicted in-hospital mortality. Journal of Clinical Epidemiology 57: 1288-1294. Tabeau E. 2001. A review of demographic forecasting models for mortality. Forecasting Mortality in Developed Countries. Kluwer: Dordrecht. Taş U, Verhagen AP, Bierma-Zeinstra SM, Odding E, Koes BW. 2007. Prognostic factors of disability in older people: a systematic review. British Journal of General Practice 57(537): 319-323. Taubes G. 1995. Epidemiology faces its limits. Science 269: 164-169. Terza JV, Basu A, Rathouz PJ. 2008. Two-stage residual inclusion estimation: addressing endogeneity in health econometric modeling. Journal of Health Economics 27: 531543. Tomiak M, Berthelot JM, Guimond E, Mustard CA. 2000. Factors associated with nursing-home entry for elders in Manitoba, Canada. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences 55(5): 279-287. Uijen AA, Van de Lisdonk EH. 2008. Multimorbidity in primary care: prevalence and trend over the last 20 years. European Journal of General Practice 14(1): 28-32. Van Baal PH, Feenstra TL, Hoogenveen RT, De Wit GA, Brouwer WB. 2007. Unrelated medical care in life years gained and the cost utility of primary prevention: in search of a ‘perfect’ cost-utility ratio. Health Economics 16: 421-433. Van Baal PH, Van den Berg M, Hoogenveen RT, Vijgen SM, Engelfriet PM. 2008. Costeffectiveness of a low-calorie diet and orlistat for obese persons: modeling long-term health gains through prevention of obesity-related chronic diseases. Value in Health 11(7): 1033-1040. Van Baal PH, Feenstra TL, Polder JJ, Hoogenveen RT, Brouwer WB. 2011a. Economic evaluation and the postponement of health care costs. Health Economics 20(4): 432-445. Van Baal PHM, Wong A, Slobbe LCJ, Polder JJ, Brouwer WBF, De Wit GA. 2011b. Standardizing the inclusion of indirect medical costs in economic evaluations. Pharmacoeconomics 29(3): 175-187. Van Buuren S, Groothuis-Oudshoorn K. 2011. MICE: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software 45: 3. Van de Vijsel A. 2009. De vraag naar zorg in ziekenhuizen. TNO Report 034-UT-200901236 rpt-czb. TNO: Delft. Van de Vijsel AR, Engelfriet PM, Westert GP. 2011. Rendering hospital budgets volume based and open ended to reduce waiting lists: Does it work? Health Policy 100(1): 60-70. Van den Akker M, Buntinx F, Knottnerus JA. 1996. Comorbidity or multimorbidity: what’s in a name? A review of literature. European Journal of General Practice 2: 65-70. Van den Akker M, Buntinx F, Metsemakers JF, Roos S, Knottnerus JA. 1998. Multimorbidity in general practice: prevalence, incidence, and determinants of cooccurring chronic and recurrent diseases. Journal of Clinical Epidemiology 51(5): 367-375. Van den Akker M, Buntinx F, Roos S, Knottnerus JA. 2001. Problems in determining occurrence rates of multimorbidity. Journal of Clinical Epidemiology 54: 675-679.

260

BIBLIOGRAPHY

Van den Berg M, Van Baal PH, Tariq L, Schuit AJ, De Wit GA, Hoogenveen RT. 2008. The costeffectiveness of increasing alcohol taxes: a modeling study. BMC Medicine 6: 36. Van Exel NJ, Koopmanschap MA, Van den Berg B, Brouwer WB, Van den Bos GA. 2005. Burden of informal caregiving for stroke patients. Identification of caregivers at risk of adverse health effects. Cerebrovascular Diseases 19(1): 11-17. Van Exel J, De Graaf G, Brouwer W. 2007. Care for a break? An investigation of informal caregivers' attitudes toward respite care using Q-methodology. Health Policy 83(2-3): 332-342. Van Exel NJA, De Ruiter M, Brouwer WBF. 2008. When time is not on your side: Patient experiences with waiting for home care and admission to a nursing or residential home. Patient: Patient-Centered Outcomes Research 1(1): 55-71. Van Gool CH, Picavet HSJ, Deeg DJH, De Klerk M, Nusselder W, Van Boxtel M, Wong A, Hoeymans N. 2011. Trends in activity limitations: the Dutch older population between 1990 and 2007. International Journal of Epidemiology 40: 1056-1067. Van Oostrom SH, Picavet SJ, Van Gelder BM, Lemmens LC, Hoeymans N, Verheij RA, Schellevis FG, Baan CA. 2011. Multimorbiditeit en comorbiditeit in de Nederlandse bevolking - gegevens van huisartsenpraktijken. Nederlandse Tijdschrift voor Geneeskunde 155: 3193. Verbrugge LM, Jette AM. 1994. The disablement process. Social Science and Medicine 38(1): 1-14. Verschuren WMM, Blokstra A, Picavet HSJ, Smit HA. 2008. Cohort profile: the Doetinchem cohort study. International Journal of Epidemiology 37(6): 1236-1241. Vita AJ, Terry RB, Hubert HB, Fries JF. 1998. Aging, health risks, and cumulative disability. New England Journal of Medicine 338(15): 1035-1041. Ware JE Jr, Sherbourne CD. 1992. The MOS 36-item short-form health survey (SF-36): I. Conceptual framework and item selection. Medical Care 30(6): 473-483. Ware JE, Kosinski M, Bayliss MS, McHorney CA, Rogers WH, Raczek A. 1995. Comparison of methods for the scoring and statistical analysis of SF-36 health profile and summary measures: summary of results from the Medical Outcomes Study. Medical Care 33(4): 264-279. Weaver F, Stearns SC, Norton EC, Spector W. 2008. Proximity to death and participation in the long-term care market. Health Economics 18(8): 867-883. Weisbrod BA. 1991. The health care quadrilemma: an essay on technological change, insurance, quality of care and cost containment. Journal of Economic Literature 29(2): 523-552. Werblow A, Felder S, Zweifel P. 2007. Population ageing and health care expenditure: a school of ‘red herrings’? Health Economics 16(10): 1109-1126. Westert GP, Satariano WA, Schellevis FG, Van den Bos GA. 2001. Patterns of comorbidity and the use of health services in the Dutch population. European Journal of Public Health 11: 365-372. Wickstrøm J, Serup-Hansen N, Kristiansen IS. 2002. Future health care costs - Do health care costs during the last year of life matter? Health Policy 62: 161-172. Williams RL. 2000. A note on robust variance estimation for cluster-correlated data. Biometrics 56: 645-646. Winnink JJ. 2006. EP-octrooien Parameters van het verleningsproces bij het EOB. Dutch Patent Office Research Report. Dutch Patent Office: Rijswijk.

261

BIBLIOGRAPHY

Wolinsky FD, Callahan CM, Fitzgerald JF, Johnson RJ. 1992. The risk of nursing home placement and subsequent death among older adults. Journal of Gerontology 47(4): 173-182. Wong A, Kommer GJ, Polder JJ. 2008. Levensloop en zorgkosten: solidariteit en de zorgkosten van vergrijzing. RIVM: Bilthoven. Wong A, Baal PHM van, Boshuizen HC, Polder JJ. 2011a. Exploring the influence of proximity to death on disease-specific hospital expenditures: a carpaccio of red herrings. Health Economics 20(4): 379-400. Wong A, Boshuizen HC, Schellevis FG, Kommer G, Polder JJ. 2011b. Longitudinal administrative data can be used to examine multimorbidity, provided false discoveries are controlled for. Journal of Clinical Epidemiology 64(10): 1109-17. Wong A, Boshuizen HC, Polder JJ. 2011c. Modeling the distribution in health care expenditures: a nearest neighbor resampling approach (submitted). Wong A, Wouterse B, Slobbe LCJ, Boshuizen HC, Polder JJ. 2012. Medical innovation and age-specific trends in health care utilization: findings and implications. Social Science and Medicine 74(2): 263-272. Wooldridge JM. 2002. Econometric analysis of cross section and panel data. MIT Press: Cambridge. World Health Organization. 2002a. Lessons for long-term care policy: World Health Organization. World Health Organization. 2002b. Medical Savings Accounts: lessons learned from limited international experience. World Health Organization discussion paper number 3. Available at: http://www.who.int/entity/health_financing/documents/dp_e_02_3med_savings_accounts.pdf. World Health Organization. 2010. WHO International Shortlist of Hospital Morbidity Tabulation. Available at http://www.who.int/classifications/icd/implementation/ morbidity/ishmt/en/index.html. World Health Organization. 2011. International Classification of Functioning, Disability and Health. Available at: http://apps.who.int/classifications/icfbrowser. Wouterse B, Meijboom BR, Polder JJ. 2011. The relationship between baseline health and longitudinal costs of hospital use. Health Economics 20(7): 355-362. Xu K, Evans DB, Kawabata K, Zeramdini R, Klavus J, Murray CJ. 2003. Household catastrophic health expenditure: a multicountry analysis. Lancet 362(9378): 111-117. Xuan J, Kirchdoerfer LJ, Boyer JG, Norwood GJ. 1999. Effects of comorbidity on health-related quality-of-life scores: an analysis of clinical trial data. Clinical Therapeutics 21: 383-403. Yakowitz S. 1987. Nearest neighbour methods for time series analysis. Journal of Time Series Analysis 8: 235-247. Yakowitz S. 1993. Nearest neighbor regression estimation for null-recurrent Markov time series. Stochastic Processes and their Applications 48: 311-318. Yang Z, Norton ED, Stearns SC. 2003. Longevity and health care expenditure: the real reasons older people spend more. Journal of Gerontology: Social Sciences 58B: S2S10. Zekry D, Loures Valle BH, Lardi C, Graf C, Michel JP, Gold G, Krause KH, Herrmann FR. 2010. Geriatrics index of comorbidity was the most accurate predictor of death

262

BIBLIOGRAPHY

in geriatric hospital among six comorbidity scores. Journal of Clinical Epidemiology 63(9): 1036-1044. Zweifel P, Breyer F. 1997. Health Economics. Oxford University Press: New York. Zweifel P, Felder S, Meier M. 1999. Aging of population and health care expenditure: a red herring? Health Economics 8(6): 485-496. Zweifel P, Felder S, Werblow A. 2004. Population ageing and health care expenditure: new evidence on the red herring. Geneva Papers on Risk and Insurance: Issues and Practice 29(4): 653-667. Zweifel P, Steinmann L, Eugster P. 2005. The Sisyphus Syndrome in health revisited. International Journal of Health Care Finance and Economics 5(2): 127-145.

263

BIBLIOGRAPHY

264

Summary The annual growth of health care expenditures is a phenomenon common to all Western developed countries. That growth is both substantial and highly persistent. In the Netherlands, the share of health care expenditures in the Gross Domestic Product (GDP) is around 14% in the year 2011. This share is expected to rise even further. Not surprisingly, there is much debate in the research community and among decision makers on the trends and future of health expenditure. The most important questions that remain unanswered include the following: Are current health care systems sustainable? If this is not the case, what kinds of policy options are available? And what changes can be best implemented to change the tides? Without a doubt, the key to making well-informed decisions in this area is having detailed knowledge of the dynamics and drivers of health care expenditures. Current research has already contributed many important insights. On the supply side, GDP has been found to have a long-term relationship with the macroeconomic level of health care expenditures. With increasing national income the ability and willingness to devote ever more resources to health care also increases. Advances in medical technology, which may include anything from medical devices to surgical techniques, are likely to spur health care demand and supply even further. Although the overall impact of medical innovation is more or less known, the causal mechanisms are various and are still to a large extent unexplained. Some instances of technology are believed to be cost-increasing, while others are believed to be cost-cutting. On the demand side, aging is seen as a key determinant of health care expenditures. Population aging – i.e., the increase in the proportion of the elderly – is caused by increasing longevity, in combination with decreasing fertility rates. Furthermore, in the next three decades population aging will accelerate by the Post World War II baby-boom cohort entering the retirement phase. The economic consequences will be considerable, and diverse in nature. These include decreasing benefits from income taxes, and, even more important, an increasing gap between pension obligations and the resources required to fund them, since the working age

SUMMARY

population decreases. But perhaps above all, the financing of health care comes under further pressure. In insurance-based financing schemes, population aging leads to greater intergenerational payments: the young working population is required to finance the growing health care needs of a large share of the nonworking elderly. One should note however, that it is not necessarily increasing longevity in itself that is responsible for growing individual lifetime health care expenditures, and at the same time for growing aggregate health care expenditures. In numerous studies it has been shown that health care expenditures in the last year of life are, on average, much higher than in the preceding years. If the expenditures would not increase with age, then increasing longevity would imply that the extra expenses of the last year of life are merely postponed. In that case, lifetime expenditures only increase with the more or less constant and relatively low average yearly costs preceding the last year of life. The attention that longevity is given in much health economic research has therefore recently been referred to as focusing on a ‘Red Herring’, as it diverts the attention from other factors that are likely to be more important, such as advances in medical technology. However, generalizing this finding to all individuals is complicated by the fact that health care expenditures over the life course tend to vary greatly between individuals, even in the last year of life. Clearly, the explanation has to be sought within differences in health status. Impairment of health is commonly measured by the presence of morbidity, comorbidity (the presence of two or more diseases), and disability. In public health research all three conditions are generally believed to be associated with higher health care expenditures, but the exact quantitative nature of each relationship remains unclear. This thesis identifies several aspects of health care expenditures that deserve further exploration. These mainly fall under three themes:  The ‘Red Herring’ phenomenon. This was investigated in order to understand the role of aging, and to make projections of health care expenditures based on demographic changes;  The relationship between the presence of one or more diseases and the amount of health care expenditures. This relation was analyzed to understand the large variation in health care expenditures between individuals;  The study of interpersonal and –generational distribution of health care expenditures, in the context of health care financing schemes. The goal of this thesis was to adapt – and modify where necessary – existing statistical techniques to gain a fuller understanding of these themes. The techniques as used in this thesis are diverse, but were mainly used in the following three ways: 266

SUMMARY

 Descriptive modeling, which is aimed at summarizing or representing the data in a compact manner. The focus is on discovering associations between variables, without necessarily presupposing a causal and theoretical framework. The knowledge gained in this manner can be used to form hypotheses, or to lay a foundation for future research;  Explanatory modeling, which is used to test causal hypotheses. This approach is driven by an underlying theoretical framework, which more or less requires the structure of the model to be defined a priori. Testing a hypothesis will not necessarily prove the causality (as most statistical models are not necessarily causal models), but rather provide empirical support for the causality (which follows from the theoretical framework);  Predictive modeling, where new observations are being predicted from explanatory variables. A special subclass of predictive modeling is forecasting, where only future observations are predicted. Predictive modeling is similar to explanatory modeling, with the exception that the model structure is largely data driven. Predictions could facilitate decisionmaking, or provide input for further research. In each chapter of this thesis, we will tackle problems belonging to one or more themes, using one or a combination of the aforementioned approaches. Each approach is chosen based on the nature of the problem involved. The problems and findings in each chapter are outlined in the following paragraphs. In Chapter 2, the relationship between proximity to death and individual average disease-specific health care expenditures was explored for 94 diseases. Using a large Dutch hospital register, individual hospital care expenditures were decomposed into disease-specific hospital care expenditures. For each disease a two-part model was fitted, which comprised the probability of non-zero health care expenditures as well as the conditional amount of health care expenditures. Results revealed that the disease-specific hospital care expenditures were associated with proximity to death for the majority of the investigated diseases. Though the strength of this association was linked to the degree to which the disease is lethal, even the expenditures for diseases that are generally not considered to be lethal in the public perception were found to be related to proximity to death. Expenditures were also positively associated with age for most diseases. However, this relationship was found to be less strong than the relationship between time to death and hospital care expenditures. The great variability according to disease in the relationship between hospital care expenditures, proximity to death and age suggests that shifts in disease prevalences in the future will probably change the relation between proximity to death and hospital care expenditures.

267

SUMMARY

Much effort in the field of health economics is devoted to cost-effectiveness analyses, in which health care interventions are compared against each other in terms of benefits (e.g. life years gained) and costs. The latter, in principle, include both the costs of the intervention itself, and the medical costs that are incurred in life years gained. A distinction can be made between related and unrelated medical costs: costs are considered unrelated, or indirect, when they are not directly related to the intervention itself. A current shortcoming of many economic evaluations is that they mostly do not include the unrelated medical costs in life years gained, or even disregard the intervention related costs in life years gained. Methodological issues, related to the ‘red herring’ phenomenon of high health care expenditures in the final year of life, play an important role in not taking into account such cost. However, it is not theoretically sound to omit these costs. In Chapter 3, a conceptual model is proposed to estimate costs of unrelated diseases in life-years gained for cost-effectiveness analyses. The approach presented here differs from other standardized methods to calculate unrelated medical costs, because it includes the fact that health care expenditures tend to rise quickly with proximity to death. Furthermore, this model goes beyond the standard proximity-to-death model, as it takes into account that the relationship between health care expenditures and proximity to death depends on the disease involved. Using the results from Chapter 2, as well as various Dutch cost-of-illness data, a standardized method was developed and embedded in a software package called 'Practical Application to Include future Disease costs' (PAID). PAID can be used as a supporting tool by researchers in their cost-effectiveness analyses. In Chapter 4, an introduction is given to the use of the ‘time to death’ parameter in projecting aggregate health care expenditures on the one hand, and estimating the share of aging in the growth of health care expenditures on the other. Where time to death is normally used in a micro-data (individual level) context, its relationship to health care expenditures has not been used on a macro-level for prediction purposes. Given that health care expenditures are higher in the last years of life, it is theoretically derived that, ceteris paribus, the aggregate health care expenditures are positively associated with mortality rates. Using data on aggregatelevel expenditures from the period 1981-2007, the growth in hospital care expenditures was indeed shown to be linked to changes in aggregate mortality rates. Forecasts were made for the years 2008-2020, and compared to forecasts that exclude any link with changes in mortality. In contrast to the findings of other studies, the inclusion of mortality seemed to have little impact on the performance of the forecast. An explanation for this was found in the reliance of the annual growth in health care expenditures on the inclusion of mortality: this growth was found to be higher with the inclusion of mortality. The main conclusion is that time to death models are of limited value in forecasting HCE and that their main 268

SUMMARY

strength lays in estimating the effect of aging in the growth of health care expenditures. As the population is aging, the demand for long-term care will probably increase within the near future. Given the limited resources (in terms of financing as well as the availability of personnel), finding ways of lowering long-term care utilization is an important policy issue. In Chapter 5 we investigated the factors that predict long-term care utilization after a hospital discharge for elderly aged 65 or above who lived at home prior to hospitalization. Large-scale Dutch hospital and longterm care registers were used. The statistical model contained as explanatory variables a wide range of demographic variables as well as hospital diagnoses. Cerebrovascular diseases were the strongest disease predictor of nursing home admission, and fractures of the ankle or lower leg were strong determinants of admission to a home for the elderly. Lung cancer was the strongest determinant of discharge to home with home care. The results of this study show that trends in hospital morbidity will influence the need for elderly care. Policy-makers and others involved in health care planning can use these insights in their scenarios for the future of chronic care, and also determine which diseases deserve to be prioritized in terms of disease prevention. The research reported in Chapter 6 studied the trends in the prevalence of chronic diseases and of disability, also analyzing the association between chronic diseases and the disability. Eight chronic diseases were considered: diabetes, heart disease, peripheral arterial disease, stroke, lung disease, joint disease, back problems, and cancer. Self-reported data from five Dutch surveys amongst non-institutionalized older persons aged 55-84 years were pooled over the period 1990-2008 using a meta-analysis approach to study these trends. It was found that prevalence rates of chronic diseases increased over time, whereas activity limitations were stable or slightly decreasing, depending on the definition of disability used. The associations between chronic diseases and activity limitations were also fairly stable, with great variation between studies. Thus, no strong evidence was found that diseases may have become less disabling over time, a hypothesis which has its roots in public health research. In Chapter 6 we found that the prevalence of chronic diseases increased over time, and we demonstrated that this conclusion not only holds for individual diseases, but also for combinations of chronic diseases. In other words: the prevalence of co-morbidity increases over time as well. Although it is known from literature that co-morbidity is often associated with higher health care expenditures, which specific co-occurrences of diseases are particularly common remains largely unexplored, as well as what potential additional health care expenditures they lead 269

SUMMARY

to. Chapter 7 describes how administrative data can be used to study comorbidity. The example of a Dutch hospital register is given to illustrate this. For various pairs of diseases the frequencies of the simultaneous occurrence of these diseases (“co-occurrences”) were counted and it was calculated whether they occurred more often than expected on the basis of chance alone (i.e., based on the prevalence of each respective disease). The results from the study could be used to confirm previously reported disease pairings while also revealing hitherto unnoticed associations, to find out which pairings cluster most strongly and to gain insight into which diseases cluster frequently with other diseases (providing incentives for prevention of combinations of diseases). Some caveats with this method are also described. One major pitfall is the issue of multiple testing. With a p-value of 0.05, on average one out of twenty associations will be identified as significantly by chance alone. Multiple testing corrections have been proposed in the literature, but these are often too conservative in terms of rejecting the null hypothesis. In this chapter we offer a simple alternative in the form of checking the significance of the association in subsequent years, to deal with chance findings.

Chapter 8 further investigated the relationship between co-morbidity and hospital care expenditures. For a broad range of diseases, the hypothesis was tested that hospital care expenditures for a principal disease are increased by the presence of a given co-morbid condition. Using a Dutch hospital register, the two-part model of Chapter 3 was expanded into a three-part model, comprising the probability of hospital admission, the conditional frequency of hospital admissions and the conditional cost per admission. It was found that hospital expenditures for a specific disease A were not increased by all co-occurring diseases, but only in roughly 7.5% of all disease pairings. It was found that for a specific disease pair A and B, B could lead to higher expenditures of A, but not vice versa. The implication here was that general trends in co-morbidity might lead to additional per capita hospital care utilization, if and only if the diseases that were identified in this study are underlying the trend. Thus, the relationship between comorbidity and health care expenditures is difficult to characterize in a general way.

Chapter 9 investigated the hypothesis that advances in medical technology will mainly benefit the elderly. In other words, the elderly will experience a large growth in health care utilization over time as result of advances in medical technology. In this chapter, age-specific trends in the probability of health care utilization were studied for different health care sectors in the Netherlands, using aggregated data from a Dutch health survey for the period 1981-2009. Moreover, for the hospital sector a model was formulated that included the following elements that follow from established theory: demographics, health status, supply 270

SUMMARY

and institutional factors. Using this model, the link between the trend in the probability of health care utilization and the state of medical technology was explored. It was found that for most health care sectors, the trend in the probability of health care utilization is highest for ages 65 and above. For hospital care, greater advances in medical technology were found to be significantly associated with a higher growth of hospitalization probability, particularly for the higher ages, suggesting that advances in medical technology indeed mainly benefit the elderly. These age-specific trends might exert additional pressure on the sustainability of intergenerational solidarity in health care. For hospital care utilization, this pressure might be strengthened even further by advances in medical technology. With many questions being raised on the sustainability of current health care financing systems, policy makers are interested in alternative ways of financing. In the Netherlands, it has been suggested that health savings accounts should be considered. These are individual savings accounts with balances that are used solely for purposes of financing health care. This system relies on a good estimation of future health care expenses, as well as on the notion that each individual incurs roughly the same amount of expenditures during lifetime. In order to study the feasibility and potential of health savings accounts, we developed a statistical method in Chapter 10 to generate synthetic life cycles in health care expenditures, and to simulate a lifetime distribution. The statistical method was adapted from existing applications in other scientific fields and modified to allow generating synthetic life cycles in health care expenditures. The model was then extensively validated on an insurance dataset of acute health care expenditures. We found it to be a suitable method to extrapolate short observed individual panels to full life cycles. Finally, Chapter 11 built on the method proposed in Chapter 10 and further modified it to examine the health care expenditures in each specific health care sector. It was concluded that while the distributions of lifetime expenditures in each sector were strongly skewed to the right, differences were observed in the degree of inequality between each sector. Of all acute health care expenditures, only hospital care, GP care and dental care featured moderate inequality. The implication is that when individual health savings accounts are considered as a form of health care financing, only expenditures of specific health care sectors with a moderate degree of inequality may be suitable for this type of financing. Furthermore, the frequent occurrence of ‘catastrophic health care expenditures’ (i.e., health care expenditures exceeding a large fraction of the disposable household income) suggests that combining individual savings with additional health care insurance will be necessary. 271

SUMMARY

As this thesis shows, insights into health care expenditures were obtained using a wide variety of approaches. A descriptive approach was found to lead to knowledge that may generate new hypotheses (e.g. in chapter 2 the description of disease-specific health care expenditures in the last year of life led to the hypothesis that changing prevalences of diseases will lead to a different relationship between time to death and health care expenditures), or that can be used in priority setting for either policy making or further research (e.g. the prevalent en expensive disease combinations as identified in chapter 7 and 8 deserve extra attention for prevention policies and research). The explanatory approach allows examining whether there is empirical support for existing hypotheses (e.g. in chapter 9 the hypothesis that mainly elderly benefit from medical innovation is tested, and not rejected). Finally, the predictive approach yields numbers that are useful as input for further research, or that facilitate decision making for policy-makers (e.g. the simulated life cycles in chapter 10 and 11 will help policy-makers in assessing the feasibility of health savings accounts). The main common theme of the approaches in this thesis is how to turn the increasingly available large datasets into productive use. Many of the results would not have been obtained otherwise. The advances in statistical techniques as well as computing power will result in a stronger empirical basis for decision-making in many areas of society, including health economics.

272

Samenvatting De jaarlijkse groei van de zorguitgaven komt in alle westerse landen voor. Die groei is niet alleen aanzienlijk, maar ook zeer persistent. In Nederland is het aandeel van de zorguitgaven in het bruto binnenlands product (BBP) in de definitie van de Zorgrekeningen van het CBS ongeveer 14% in het jaar 2011. Dit aandeel zal naar verwachting verder stijgen. Er is dan ook veel discussie onder wetenschappers en beleidsmakers over de toekomst van de zorguitgaven. Onder de belangrijkste en onbeantwoorde vragen vallen de volgende: Zijn de huidige zorgstelsels houdbaar? Indien dit niet het geval is, welke beleidswijzigingen behoren dan tot de mogelijkheden? En welke veranderingen leiden dan tot de beste resultaten, zowel met betrekking tot de volksgezondheid als de zorguitgaven? Voor de juiste beslissingen is gedetailleerde kennis van de dynamiek en de determinanten van zorguitgaven benodigd. Bestaand onderzoek heeft reeds veel belangrijke inzichten opgeleverd. Aan de aanbodzijde is vastgesteld dat het BBP een lange-termijn relatie heeft met de totale zorguitgaven. Bij toename van het nationale inkomen nemen ook het vermogen en de bereidheid toe om steeds meer middelen te besteden aan de zorg. Medische innovaties, die kunnen uiteenlopen van medische hulpmiddelen tot operatietechnieken, zullen waarschijnlijk niet alleen het aanbod maar ook de vraag naar gezondheidszorg verder bevorderen. Hoewel het effect van medische innovatie min of meer bekend is, zijn de causale mechanismen divers en nog voor een groot deel onverklaard. Sommige vormen van medische technologie worden verondersteld te leiden tot stijgingen in de zorguitgaven, terwijl andere vormen misschien weer tot kostenbesparingen kunnen leiden. Aan de vraagzijde wordt de vergrijzing vaak gezien als een belangrijke determinant van zorguitgaven. Vergrijzing van de bevolking – d.w.z., de toename van het aandeel van de ouderen – wordt veroorzaakt door de toenemende levensverwachting. Bovendien zal in de komende drie decennia de bevolking sneller vergrijzen, doordat de naoorlogse geboortegolf (de “baby-boomers”) op het punt staan de pensioengerechtigde leeftijd bereiken. De gevolgen van deze dubbele vergrijzing voor de economie zullen aanzienlijk zijn, en divers van aard. Deze

SAMENVATTING

omvatten onder meer dalende belastinginkomsten voor de overheid, en, nog belangrijker, een groeiende kloof tussen de pensioenverplichtingen en de middelen die benodigd zijn om die te financieren, aangezien de beroepsbevolking afneemt. Wellicht is de grootste consequentie van de vergrijzing dat de financiering van de gezondheidszorg verder onder druk komt te staan. In het geval dat de zorg gefinancieerd wordt door een verzekeringsstelsel, leidt de vergrijzing tot grotere betalingen tussen generaties: de jonge beroepsbevolking wordt geacht om de groeiende zorgvraag van een groot deel van de niet-werkende ouderen te financieren. Het is echter niet de vergrijzing die het grootste aandeel heeft in de stijging van de totale jaarlijkse zorguitgaven. In talrijke studies is gebleken dat de uitgaven voor de gezondheidszorg in het laatste levensjaar gemiddeld veel hoger is dan in voorgaande jaren. Als de uitgaven niet zou toenemen met de leeftijd, zou de toenemende levensverwachting impliceren dat de extra kosten van het laatste levensjaar alleen maar wordt uitgesteld. In dat geval zouden de uitgaven gedurende de levensloop alleen maar toenemen met de relatief lage, en min of meer constante, kosten voorafgaande aan het laatste levensjaar. De aandacht voor de toenemende levensverwachting wordt in de gezondheideconomie daarom gezien als een dwaalspoor (“Red Herring”), omdat het de aandacht afleidt van andere factoren die waarschijnlijk belangrijker zijn, zoals de ontwikkelingen in de medische technologie. Het is echter de vraag of dit voor alle individuen geldt, want de zorguitgaven variëren sterk tussen individuen, zelfs in het laatste levensjaar. Deze variatie kan worden toegeschreven aan verschillen in gezondheidstoestand. De mate van gezondheid wordt gewoonlijk gemeten door indicatoren zoals de aanwezigheid van morbiditeit, co-morbiditeit (de aanwezigheid van twee of meer ziekten), en lichamelijke beperkingen. In de literatuur wordt elk van deze indicatoren geassocieerd met hogere zorguitgaven, maar de exacte kwantitatieve aard van elk afzonderlijke relatie blijft onduidelijk. Dit proefschrift identificeert verschillende aspecten van de zorguitgaven die meer aandacht verdienen. Deze aspecten vallen onder drie thema's:  Het 'Red Herring' fenomeen. Dit wordt onderzocht om de rol van het ouder worden beter te begrijpen, en om projecties van de zorguitgaven op basis van demografische veranderingen te maken;  De relatie tussen de aanwezigheid van één of meer ziekten en de hoeveelheid zorguitgaven. Deze relatie wordt geanalyseerd om de grote variatie in de zorguitgaven tussen individuen beter te begrijpen;  Een verkenning van de interpersoonlijke en –generationele verdeling van zorguitgaven, in het kader van zorgfinancieringssystemen.

274

SAMENVATTING

Het doel van dit proefschrift is om moderne statistische technieken in te zetten – en waar nodig aan te passen – voor het inzichtelijk maken van deze thema’s. De technieken zoals gebruikt in dit proefschrift zijn divers, maar zijn voornamelijk gebruikt op de volgende drie manieren:  Beschrijvende modellering, gericht op het samenvatten van gegevens op een compacte manier. De nadruk ligt op het ontdekken van verbanden tussen variabelen, zonder noodzakelijkerwijs uit te gaan van een causaal en theoretisch kader. De opgedane kennis kan vervolgens worden gebruikt om hypothesen te formuleren of kan een basis vormen voor toekomstig onderzoek;  Verklarende modellering, die wordt gebruikt om causale hypothesen te toetsen. Deze aanpak wordt gestuurd door een onderliggend theoretisch kader, dat min of meer vereist dat de structuur van het model a priori is gedefinieerd. Het testen van een hypothese zal niet per se de causaliteit bewijzen, maar biedt empirische ondersteuning voor de causaliteit (die volgt uit het theoretisch kader);  Voorspellende modellering, waarbij nieuwe waarnemingen worden voorspeld op grond van verklarende variabelen. Een speciaal geval hiervan is “forecasting”, waarbij alleen toekomstige waarnemingen worden voorspeld. Voorspellende modellen zijn vergelijkbaar met verklarende modellen, met als belangrijke uitzondering dat de modelstructuur grotendeels bepaald wordt door de data. Voorspellingen kunnen de besluitvorming ondersteunen, of input geven voor verder onderzoek. In elk hoofdstuk van dit proefschrift is een afzonderlijk onderwerp geanalyseerd met behulp van één of meerdere modelmatige benaderingen. Elke benadering is gekozen op basis van de aard van het probleem. In Hoofdstuk 2 is de relatie tussen het laatste levensjaar en individuele ziektespecifieke zorguitgaven verkend voor 94 ziekten. Met behulp van een grote registratie van ziekenhuisopnamen werden individuele ziekenhuisuitgaven opgesplitst naar ziektespecifieke ziekenhuisuitgaven. Voor elke ziekte werd een “two-part” model geschat. In het eerste deel wordt de kans op zorguitgaven geschat, en in het tweede deel worden de zorguitgaven, gegeven zorguitgaven groter dan nul, geschat. Voor de meerderheid van de onderzochte ziekten werd een verband gevonden tussen de ziektespecifieke ziekenhuiszorguitgaven en de tijd tot overlijden. Hoewel dit verband sterker werd naar mate de ziekte dodelijker was, waren zelfs de uitgaven voor ziekten die meestal niet als dodelijk worden beschouwd hoger in de laatste levensjaren. Er werd voor de meeste ziekten ook een positief verband gevonden tussen de uitgaven en leeftijd gevonden; deze relatie bleek echter minder sterk te zijn dan de relatie tussen zorguitgaven en tijd tot 275

SAMENVATTING

overlijden. De grote variabiliteit tussen ziekten in termen van de relatie tussen de ziekenhuiszorg uitgaven, tijd tot overlijden en leeftijd doet vermoeden dat epidemiologische verschuivingen in het vóórkomen van ziekten waarschijnlijk zullen leiden tot een verandering in de relatie tussen tijd tot overlijden en de (niet ziekte-specifieke) ziekenhuiszorguitgaven. In de gezondheidseconomie wordt veel aandacht besteed aan kosteneffectiviteitsanalyses. Hierin worden interventies in de gezondheidszorg met elkaar vergeleken in termen van effecten (bijvoorbeeld gewonnen levensjaren) en de kosten. De kosten omvatten in principe zowel de kosten van de interventie zelf, als de medische kosten die worden gemaakt in gewonnen levensjaren. Er kan hierbij onderscheid worden gemaakt tussen verwante en niet-verwante medische kosten: de kosten worden beschouwd als niet-verwant wanneer ze niet direct gerelateerd zijn aan de interventie zelf. Een tekortkoming van veel economische evaluaties is dat ze de niet-verwante medische kosten in gewonnen levensjaren buiten beschouwing laten. Daarnaast komt het ook nogal eens voor dat zelfs de aan de interventie verwante kosten in gewonnen levensjaren genegeerd worden. Het is vanuit de theorie gezien echter niet verdedigbaar om deze kosten weg te laten. Methodologische kwesties, die verband houden met het 'Red Herring' fenomeen, spelen een belangrijke rol in het niet rekening houden met nietverwante kosten. In Hoofdstuk 3 is een conceptueel model voorgesteld om de kosten van niet-verwante ziekten te schatten in gewonnen levensjaren voor kosteneffectiviteitsanalyses. De hier gepresenteerde benadering wijkt af van andere gestandaardiseerde methoden voor de berekening van niet-verwante medische kosten, omdat het ook rekening houdt met de hoge zorguitgaven in het laatste levensjaar. Bovendien gaat dit model verder dan het standaard tijd tot overlijden model, omdat dit model tevens er van uit gaat dat de relatie tussen de zorguitgaven en tijd tot overlijden afhangt van de ziekte in kwestie. Met behulp van de resultaten van Hoofdstuk 2 en gegevens uit de Kosten van Ziekten studie werd een gestandaardiseerde methode ontwikkeld en omgezet naar een software pakket, genaamd 'Practical Application to Include future Disease costs' (PAID). PAID kan als een ondersteunende tool gebruikt worden bij kosteneffectiviteit analyses.

Hoofdstuk 4 gaat over het gebruik van tijd tot overlijden bij de projectie van totale zorguitgaven en de schatting van het aandeel van de vergrijzing daarin. Waar tijd tot overlijden normaliter wordt gebruikt in een microdata-setting, is deze relatie niet gebruikt op macroniveau om zorguitgaven te voorspellen. Gezien het feit dat de zorguitgaven gemiddeld hoger zijn in de laatste levensjaren, wordt hier afgeleid dat, ceteris paribus, de totale zorguitgaven met afnemende sterfte zullen dalen. Met behulp van gegevens betreffende totale ziekenhuisuitgaven uit de periode 19812007 werd een verband gevonden tussen de groei in de ziekenhuisuitgaven en de 276

SAMENVATTING

veranderingen in de totale sterfte. Ziekenhuisuitgaven werden voorspeld voor de jaren 2008-2020, en vergeleken met voorspellingen op basis van een model dat geen gebruik maakte van de tijd tot overlijden. In tegenstelling tot andere studies, leek het rekening houden met veranderende sterftepatronen weinig invloed te hebben op de voorspelkracht. Een verklaring hiervoor is dat de jaarlijkse groei van de zorguitgaven afhangt van de sterfte: deze groei bleek hoger te zijn naar mate de sterfte hoger is. De belangrijkste conclusie is dat de modellen die gebaseerd zijn op tijd tot overlijden van beperkte waarde zijn bij het voorspellen van zorguitgaven, en dat hun grote kracht voornamelijk ligt in het schatten van het aandeel van de vergrijzing in de groei van de zorguitgaven. Naarmate de bevolking meer vergrijst, neemt de vraag naar langdurige zorg toe. Gezien de beperkte middelen (in termen van financiering, alsmede de beschikbaarheid van voldoende personeel), is het vinden van manieren om de vraag naar langdurige zorg te verlagen een belangrijke politieke kwestie. In Hoofdstuk 5 hebben we factoren geïdentificeerd die de vraag naar langdurige zorg goed voorspellen voor ouderen van 65 jaar of ouder die ontslagen worden uit het ziekenhuis. Voor dit onderdeel werden grootschalige ziekenhuis- en langdurige zorgregistraties gebruikt. Het statistische model bevatte als verklarende variabelen een breed scala aan demografische variabelen en ziekenhuisdiagnoses. Cerebrovasculaire aandoeningen waren de sterkste voorspeller van verpleeghuisopnamen, en fracturen aan het onderbeen waren sterke determinanten van een opname in een verzorgingshuis. Longkanker was de sterkste determinant van thuiszorg. De resultaten tonen aan dat trends in ziekten de vraag naar ouderenzorg kunnen beïnvloeden. Beleidsmakers in de gezondheidszorg kunnen gebruik maken van deze inzichten voor het maken van scenario's voor de toekomst van de chronische zorg, en voor het bepalen van de ziekten die de grootste aandacht verdienen bij preventie. In Hoofdstuk 6 zijn de trends in de prevalentie van chronische ziekten en lichamelijke beperkingen bestudeerd, evenals de relatie tussen chronische ziekten en lichamelijke beperkingen. Acht chronische ziekten/ziektegroepen werden onderzocht: diabetes, hart- en vaatziekten, perifeer arterieel vaatlijden, beroerte, longaandoeningen, gewrichtsaandoeningen, rugklachten, en kanker. Zelfgerapporteerde gegevens van vijf Nederlandse enquêtes onder nietgeïnstitutionaliseerde ouderen van 55-84 jaar werden door middel van een metaanalyse aanpak gebruikt om deze trends over de periode 1990-2008 te bestuderen. De prevalentie van chronische ziekten bleek in de loop der tijd toe te nemen, terwijl de prevalentie van lichamelijke beperkingen stabiel bleef of licht daalde, afhankelijk van de gebruikte definitie van lichamelijke beperkingen. De associaties tussen chronische ziekten en beperkingen van de aktiviteit bleken redelijk stabiel te 277

SAMENVATTING

zijn. Hierbij werd er een grote mate van variatie tussen de verschillende studies gevonden. Zo werd er geen sterk bewijs gevonden voor de gangbare hypothese dat de onderzochte ziekten in recentere jaren minder beperkingen met zich mee zijn gaan brengen. In hoofdstuk 6 hebben we vastgesteld dat de prevalentie van een aantal chronische ziekten toeneemt in de tijd. Daarnaast bleek ook de prevalentie van comorbiditeit (combinaties van ziekten) toe te nemen. Hoewel het bekend is dat comorbiditeit vaak geassocieerd wordt met hogere zorguitgaven, is het gelijktijdig voorkomen van ziekten, en de mogelijke consequenties daarvan voor zorguitgaven, voor een groot deel nog niet kwantitatief onderzocht. Hoofdstuk 7 beschrijft hoe administratieve gegevens kunnen worden gebruikt om het vóórkomen van comorbiditeit te bestuderen. Ter illustratie wordt hier de Landelijke Medische Registratie van ziekenhuisopnamen gebruikt. Voor verschillende paren van ziekten werden de frequenties van het gelijktijdige optreden van deze aandoeningen geteld, en werd berekend of deze paren vaker komen dan verwacht op basis van toeval (d.w.z. op basis van de prevalentie van elke respectievelijke ziekte). De resultaten van deze studie kunnen niet alleen worden gebruikt om eerder gerapporteerde ziekteverbanden te bevestigen, maar ook om (tot nu toe) onbekende ziektecombinaties te onthullen, om te achterhalen welke ziektecombinaties het sterkst clusteren en om inzicht te krijgen in welke ziekten het meest vaak met andere ziekten clusteren (hetgeen informatie oplevert over welke combinaties van ziekten potentieel meer aandacht verdienen voor preventie). Bij deze methode past echter wel een aantal kanttekeningen. Een belangrijke valkuil is het probleem van “multiple testing”. Met een p-waarde van 0.05, zal gemiddeld één op de twintig verbanden alleen al op basis van toeval als statistisch significant worden beschouwd. Er bestaan reeds vele multiple testing correcties in de literatuur, maar deze zijn vaak conservatief in termen het verwerpen van de nulhypothese. In dit hoofdstuk doen wij een voorstel voor een eenvoudig alternatief voor. Gevonden verbanden worden alleen als relevant beschouwd als deze in meerdere jaren voorkomen. In Hoofdstuk 8 is de relatie tussen comorbiditeit en ziekenhuiszorguitgaven onderzocht. Voor een breed scala van aandoeningen werd de hypothese getest dat ziekenhuiszorguitgaven voor een hoofdziekte worden verhoogd door de aanwezigheid van een bepaalde nevenaandoening. Met behulp van de Landelijke Medische Registratie voor ziekenhuisopnamen werd het “two-part” model van hoofdstuk 3 uitgebreid tot een “three-part” model, bestaande uit de kans op ziekenhuisopname, het aantal ziekenhuisopnames, gegeven dat er minimaal 1 ziekenhuisopname heeft plaats gevonden, en de kosten per opname. Uit de resultaten bleek dat ziekenhuisuitgaven voor een bepaalde ziekte lang niet altijd 278

SAMENVATTING

werd verhoogd door de aanwezigheid van een andere ziekte. Dit was slechts het geval in ongeveer 7,5% van alle ziekteparen. Voor een specifiek ziektepaar A en B bleek het bovendien mogelijk dat B kan leiden tot hogere uitgaven voor A, maar niet vice versa. De implicatie hier is dat de algemene trends in co-morbiditeit kan leiden tot additionele per capita ziekenhuisuitgaven, als de ziekten die in deze studie zijn geïdentificeerd ten grondslag liggen aan de trend. Maar dit betekent ook dat er niet een enkele algemeen geldende relatie bestaat tussen comorbiditeit en zorguitgaven. In Hoofdstuk 9 is de hypothese onderzocht dat ouderen de meeste baat hebben bij de ontwikkelingen in de medische technologie. Met andere woorden, onder de ouderen vindt er een grotere groei in zorggebruik plaats als gevolg van de ontwikkelingen in het zorgaanbod. In dit hoofdstuk werden leeftijdsspecifieke trends in de kans op zorggebruik onderzocht voor verschillende zorgsectoren, met behulp van geaggregeerde gegevens van een Nederlandse gezondheidsenquête voor de periode 1981-2009. Bovendien werd er voor de ziekenhuissector een model geformuleerd dat onder meer de volgende aspecten bevatte: demografie, gezondheid, zorgaanbod en institutionele factoren. Met behulp van dit model, werd de link tussen de veranderingen in de opnamekans en de stand van de medische technologie onderzocht. Voor de meeste zorgsectoren was de toename in de kans op zorggebruik het hoogst voor ouderen van 65 jaar en ouder. Voor de ziekenhuiszorg werd er een verband gevonden tussen de ontwikkelingen in de medische technologie en de toename van de opnamekans. Dit verband was sterker voor hogere leeftijden, hetgeen suggereert dat de ontwikkelingen in de medische technologie voornamelijk de ouderen ten goede komen. De leeftijdsspecifieke trends zouden extra druk kunnen uitoefenen op de houdbaarheid van de intergenerationele solidariteit in de gezondheidszorg. Voor de ziekenhuiszorg geldt dat dit versterkt wordt vanwege de ontwikkelingen in de medische technologie. Omdat de houdbaarheid van het huidige zorgstelsel onder druk staat, zijn beleidsmakers geïnteresseerd in alternatieve manieren van financiering. In Nederland is gesuggereerd dat het zorgsparen moet worden overwogen. Dit zijn individuele spaarrekeningen met saldi die uitsluitend worden gebruikt voor zorgkosten. Dit financieringssysteem gaat ervan uit dat toekomstige zorguitgaven goed kunnen worden ingeschat, en dat iedereen ongeveer even veel zorguitgaven maakt gedurende de levensloop. Om de haalbaarheid van zo’n spaarsysteem te onderzoeken, is in Hoofdstuk 10 een statistische methode ontwikkeld om synthetische levenslopen in termen van zorguitgaven te genereren, en daarmee een verdeling van zorguitgaven te simuleren. Deze statistische methode is gebaseerd op bestaande toepassingen in andere wetenschappelijke disciplines, en aangepast om de levenslopen te genereren. Het model werd uitgebreid gevalideerd aan de hand 279

SAMENVATTING

van verzekeringgegevens. De resultaten suggereren dat het een geschikte methode is om korte waargenomen tijdsintervallen te extrapoleren naar een volledige levenscyclus. Tot slot is in Hoofdstuk 11 de in hoofdstuk 10 voorgestelde methode verder aangepast om sectorspecifieke zorguitgaven te onderzoeken. Hoewel de verdelingen van de levensloop zorguitgaven in elke zorgsector scheef over de bevolking verdeeld bleken te zijn, werden er wel verschillen waargenomen in de mate van scheefheid tussen de verschillende zorgsectoren. Van alle curatieve zorguitgaven bleek de scheefheid in de ziekenhuiszorg, huisartsenzorg en tandheelkundige zorg redelijk gematigd te zijn. De resultaten suggereren dat alleen specifieke zorgsectoren geschikt om via een spaarsysteem gefinancierd te worden; namelijk die sectoren waarbij de scheefheid beperkt is. Bovendien suggereert het regelmatig voorkomen van “catastrofale zorguitgaven” (d.w.z. zorguitgaven die een aanzienlijk deel van het beschikbaar huishoudinkomen in beslag nemen) dat zorgsparen niet op zichzelf kan staan, maar op zijn minst met een zorgverzekering (voor catastrofale uitgaven) zal moeten worden gecombineerd. In dit proefschrift is inzicht in de zorguitgaven verworven met behulp van een breed scala aan benaderingen. Een beschrijvende aanpak levert kennis op die tot nieuwe hypothesen kan leiden (zoals het beschrijven van de ziektespecifieke kosten in het laatste levensjaar leidde tot de hypothese dat de ‘tijd tot overlijden’ relatie met de zorgkosten zal veranderen als gevolg van verschuivingen in de ziekteprevalenties [hoofdstuk 2]), of gebruikt kan worden voor het stellen van prioriteiten bij het maken van beleid of het uitvoeren van verder onderzoek (de gevonden veelvoorkomende en dure ziekteparen in hoofdstuk 7 en 8 kunnen de prioriteit krijgen bij preventiebeleid en verder onderzoek). De verklarende aanpak maakt het mogelijk om na te gaan of er empirische steun bestaat voor de gestelde hypothesen (in hoofdstuk 9 is de hypothese dat vooral de ouderen profijt hebben van medische innovaties getoetst en niet verworpen). Ten slotte levert de voorspellende benadering cijfers op die gebruikt kunnen worden in verder onderzoek, of de besluitvorming kunnen ondersteunen voor beleidsmakers (in hoofdstuk 10 en 11 kunnen de voorspelde levenslopen helpen met besluitvorming over zorgsparen). De gebruikte benaderingen in dit proefschrift hebben één thema gemeen: er wordt gebruik gemaakt van grote datasets met een schat aan waardevolle informatie. De beschikbaarheid van dergelijke datasets is in de afgelopen jaren sterk toegenomen. Zonder die data zouden de meeste behaalde resultaten in dit proefschrift niet zijn verkregen. De ontwikkelingen in beschikbare datasets, statistische methoden en rekenkracht zullen tot een sterkere empirische basis leiden voor de besluitvorming op veel terreinen van de samenleving, met inbegrip van de gezondheidseconomie. 280

Acknowledgements Health economists are often interested in what happens over the course of a life. How does health evolve over life? What are the lifetime health expenditures? And how are these related? Amongst other things, this thesis presents a statistical approach to simulating lifecycles in terms of health care expenditures. To me, this has been the most intriguing part, in many ways. As I was examining these lifecycles, I found that health care expenditures not only vary strongly between individuals, but also within individuals. At any given moment, an individual might go from low costs to high costs, and back again. If you break a leg in an accident, you will end up in a hospital. If you discover a bump on your body somewhere, you will have it removed by surgery. What probably should be obvious isn’t actually obvious to many: almost all individuals incur high health care expenditures at some point during life 1. A lot of it is unpredictable. And an argument could be made that this applies to many other facets of life. If anything, my career path has exhibited a little bit of ‘randomness’ as well. Had there not been an opening for an internship at that exact moment, I would have not ended up working at the National Institute for Public and the Environment (RIVM). And without the RIVM, I would have not met my supervisors, probably not pursued a PhD degree, and surely not made this corny introduction 2. The truth is: the one thing that has remained stable is that I have been fortunate enough to be surrounded by trustworthy and reliable people. In the following section I would like to express my gratitude towards those. Although I cannot thank all of them personally here, it goes without saying that I am grateful. First off, I would like to thank my supervisors and the other members of the PhD committee. My supervisors have made my journey relatively smooth, despite the little bumps here and there. When starting at the RIVM, pursuing a doctorate Admittedly, this seems like a rather morose way of thinking. This is completely unintended, however. I would rather turn this argument on its head. Being in good health is, for many, a prerequisite for overall happiness. With the knowledge that anything can happen, one should treasure every moment of being in good health. This should be obvious, but somehow it is often taken for granted. 2 I apologize if this, and what will follow, may come off as cliché (because, well, it probably is!). 1

ACKNOWLEDGEMENTS

had never really entered my mind. Johan, without your persuasion, this thesis would have never come into fruition. Your moral support, and your great faith in me, has certainly spurred me on. Your ability to convey complex matters to those not familiar with the source material never ceases to amaze me, and has been an eye-opener for me. Thank you for providing some of the means to my personal development 3. Hendriek, you have made it possible for me to combine both the PhD research and work as a statistician. I thank you for this unique opportunity, especially because the diversity has allowed my mind to stay fresh for this research. Many times, you have shown to be very supportive. Despite you being busy with other things, I could always drop by for any question, and I was always amazed by how many times you had an answer ready for me. The extent of your knowledge is vast, I am convinced. And in the rare case you did not have an answer ready, you nudged me in the right direction. I would also like thank prof. dr. S. Felder, prof. dr. R.T.J.M. Janssen, dr. M.A. Koopmanschap and prof. dr. Kooreman for their willingness to serve as members of the PhD committee, and their critical evaluation of this thesis. I have learned a great deal from working with many other people. I would like to thank the following co-authors for their valuable input in various areas: Martin van Boxtel, Werner Brouwer, Dorly Deeg, Coen van Gool, Rianne Elderkamp-de Groot, Job van Exel, Nancy Hoeymans, Mirjam de Klerk, Wilma Nusselder, Susan Picavet, François Schellevis, en Ardine de Wit. The following four people I would like thank in particular. Geert Jan Kommer, you set things into motion when you hired me as intern, and you introduced me to Johan. Thank you for that, and thanks for the interest you’ve shown. Lany Slobbe, your great knowledge and attention to detail has helped me to get started – which is always the toughest part. Peter Engelfriet, thank you for your willingness to revise several parts of this thesis, which has undoubtedly improved the quality of this thesis. Despite my best efforts, I could not omit ‘PAIDvader’ Pieter van Baal here. In all seriousness, your intuition in matters that seem to be vague at first turned out to be very valuable. And your great sense of humor never made working with you dull – I will certainly miss working with you. This work could not be done without funding from various sources: the RIVM (a major thanks to Bennie Bloemberg and Hans van Oers), the Dutch Ministry of Health, Welfare and Sport (VWS) and the Netherlands Organization for Health Research and Development (ZonMW). This thesis has relied exclusively on the availability of good and reliable data. I would like thank CAK, Dutch Hospital Data, Kiwa-Prismant, Octrooiencentrum, Vektis and Zorgverzekeraars Nederland for their great data resources. Statistics 3

Although not necessarily the only and/or the best way to obtain skills, doing a PhD is certainly an excellent vehicle for personal development.

282

ACKNOWLEDGEMENTS

Netherlands have integrated most of these sources, and made them accessible to researchers. Their efforts have been instrumental in advancing research in the Netherlands. A special thank you goes to Onno van Hilten, Agnes de Bruin, Janneke Ploemacher, Marije Berger-van Sijl, Gerard Verweij, Mirthe Bronsveld-de Groot and Jos de Ree, who have made it possible to work – and pleasantly at that, I should stress – at their department when Remote Access was not available yet. At the RIVM I have had a great time. I would like to thank everybody for their interest and the conversations. Particularly, I am indebted towards my (ex-)office mates. Monique and Katia, you have put great effort into making me feel at home at the RIVM; I am fortunate to have shared a room with such sweet persons. Thank you for the many conversations. Jan en Maarten, I highly enjoyed our frequent (awful and yet hilarious) banter. Although it sometimes has been detrimental to the amount of work done, I would argue it has helped me in the long-run. Also thanks for answering many statistics-related questions, including how to make neat graphs in R. Arnold, thanks for some of the more lengthier discussions (particularly in Bordeaux), and always pointing out with your sometimes-not-so-subtle jokes when I have lost sight of what truly matters! José, you have had – and still have – a great influence on me. I should thank you for rekindling some of my interest for mathematics. I am truly envious of your ability to come up with mathematical proof on the fly. You have shown to me, in your words, ‘that there is more to statistics than just regression’, which has resulted in the substantial role of the nearest-neighbor method in this thesis. The countless days of re-running simulations almost got to me in the end, but you have helped me through all of this, by providing your expertise, as well as changing the subject of our discussions towards other (often lighter) matters, such as tennis, science and politics. Thank you. I have also had the pleasure of working at the department of Scientific Centre for Care and Welfare (Tranzo) within Tilburg University. What struck me, and pretty much everybody who had been there for any reasonable amount of time, is the great atmosphere there. There were plenty of nice drinks, and dinners. I would like to thank Henk Garretsen for the opportunity to work here. Jacqueline, thanks for the help on the ZonMW grant. The proposal for that grant has improved tremendously with the useful tips of Rosalie and Marloes, whom were also brief but great office mates. The fellow economists Richard and Arthur should be thanked for organizing the ‘economenuurtjes’, and for always providing some really good laughs. Marie Jeanne, I have always enjoyed our dinners together. Your laugh is truly contagious. Kees, thanks for reminding me things are not as bad as they seemed, and for the good times at the pub. Dung, thanks for the numerous long conversations we had. I really appreciate your sense of humor. Also thanks to Emely, Theo, Marja, Marjolein, Maartje et al. (2008-2011) for the drinks, chatter and so on. 283

ACKNOWLEDGEMENTS

Thanks to my friends who have helped me keep my mind off the PhD trajectory and relax a little bit. Thanks to Shenyu, Yenting, Ronald K., Eeka, Ino and Asheng for some of those nights. Also thanks to Jong, Weijun, Lucas, Linda, Sander, Linma, Johan en Zhiying for the nice dinners. Particularly Lijsaan has been very kind. Jeroen, Maarten K., Willem V. and Ronald L. are thanked for their support, particularly in the ‘old days’. Finally, my two ‘paranymphs’ deserve a special mention. Willem, the last year or so has been quite eventful. Thanks for all the conversations, and your support. Now that I have a little bit more time to hang out, it is somewhat ironic that you will probably have a little less time on your hands in the near future! Bram, with all the diversity within Tranzo, I feel fortunate to have encountered somebody who is – more or less – doing the same type of research. Although we may not have worked together as much as I would have liked, the opportunity to exchange ideas to each other has surely paid off. Furthermore, we share a good (or poor, depending on your viewpoint) sense of humor, which allows us to get along really well. So well, that we have travelled nearly 40,000 kilometers together to international congresses and beyond. I look back fondly on this period. When it comes to being surrounded with people you can rely on, it starts with family. Fortunately, my family was extremely supportive, even though they did not always understand what I was doing (and why on earth I was taking so long!). Hiuyee, Eduard, Hiuli and Floris, Thijmen and Merrijn, Merel and Lisa, thank you for giving me a boost whenever I needed it. Even if the distance prohibits us from seeing each other on a more frequent basis, I always feel at home when visiting you guys. And speaking of distances, Hiuli, thank you for the incredible cover (representing the Nearest Neighbor method), and the pieces of artwork as found in this thesis. They have far exceeded what I ever had in mind. Finally, I would like to thank my parents for everything. Since certain nuances tend to get lost in translation, I will convey my message in traditional Chinese (thanks to Donald and Justin for their help):

爸爸、 媽媽,我很感謝你們養育之恩。你們灌輸給我的 誠信、努力和决心, 才能使我完成學業。更重要的是你們 多年來不斷的循循善誘,無薇不至,偉大的慈愛。所以這 編畢業論文不但是我的,是屬於我們大家的。 And so my journey ends…here. A little bit tired, but ultimately, it was a journey well worth taking. –Albert Wong, Bilthoven, November 2011 284

Curriculum Vitae Albert Wong (honzi: 黃家俊; jyutping: wong4 gaa1 zeon3) was born on September 21, 1980, in Enschede, the Netherlands. Born to Chinese parents, he was raised bilingually, learning Cantonese at home, while receiving education in Dutch. He was awarded a Certificate of Proficiency in English by Cambridge ESOL in 1997, and graduated from secondary school at the Stedelijk Lyceum Zuid in 1998. Afterwards, he studied Applied Mathematics at the University of Twente, with a specialization in Statistics and Probability. As part of his degree requirements, he had an internship at the National Institute for Public Health and the Environment (RIVM), where he wrote a thesis on the estimation of ambulance travel times. In 2006 he obtained an ingenieur degree, which is equivalent to the Master of Science degree of present-day. After completing his thesis, he worked at the Centre for Public Health Forecasting within the RIVM in February 2006, where he did a project on the estimation of health care expenditures. In August 2006 he relocated to the Statistics and Mathematical Modeling department within the RIVM. Here, he worked as a statistician, making statistical models and advising other researchers on statistical problems. His modeling subject areas were mainly health economics, epidemiology, and public health. From September 2008 on, he became a part-time PhD candidate at the department of Scientific Centre for Care and Welfare (Tranzo) within Tilburg University. His research involved the description, explanation and prediction of health care expenditures, with the goal of facilitating decision-making for policy makers within this area. His results have been used by the RIVM in the quadrennial publication ‘The Dutch Public Health Status and Forecasts Report’, as well as by the Netherlands Bureau for Economic Policy Analysis (CPB) to test the feasibility of health care savings accounts as a health care financing system in the Netherlands. This work was partly funded by a grant from the Netherlands Organization for Health Research and Development (ZonMW), which he was awarded with in 2010. Since November 2011, he has made a gradual transition into full-time work at the RIVM as a statistical consultant. He will remain active in health economics, as he had cowritten a proposal on the quantitative relationship between advances in medical technology, health and health care expenditures, which led to a three-year grant by the RIVM.

CURRICULUM VITAE

286

Loading...

Tilburg University Describing, explaining and - Research portal

Tilburg University Describing, explaining and predicting health care expenditures with statistical methods Wong, A. Document version: Publisher's PD...

4MB Sizes 4 Downloads 15 Views

Recommend Documents

Tilburg University Macroeconomic policy - Research portal
Macroeconomic policy coordination during the various phases of economic and monetary integration in Europe van der Ploeg

Tilburg University Contesting religious authority - Research portal
Nov 10, 2015 - Users may download and print one copy of any publication from the public portal for the purpose of privat

Tilburg University Quantifying individual player - Research portal
Today, we can play games producing screenshots which are visually indistinguishable ..... 7.5.2 Comparison of theory-dri

Contesting religious authority - Research portal - Tilburg University
1 Tilburg University Contesting religious authority Sunarwoto, Sunarwoto Document version: Publisher's PDF, also known a

Tilburg University Assessing the risk and prevalence - Research portal
Feb 23, 2016 - The ICVS 2005 data on hate crimes are based on samples of the general population of 15 European nations (

Tilburg University War, law and technology Dijkhoff - Research portal
20 Dec 2010 - Writing a thesis as this one seems a quite solitary activity. ... §3.2 The Answers. 18. §3.3 Method. 19.

Tilburg University The Endogeneity Bias in the - Research portal
is strongly positively correlated with firm heterogeneity. Synthesizing our findings, we ..... endogeneity as the case w

Levine, R.; Loayza, N. - Research portal - Tilburg University
While Levine, Loayza, and Beck (1999) use a very similar data set and identical econometric ..... from the Penn-World Ta

Tilburg University Es la teoria de la internalizacion - Research portal
Es la teoria de la internalizacion una teoria general de la empresa multinacional? El caso de la 'empresa de exportacion

Tilburg University Patient preference for counselling - Research portal
Tilburg University. Patient preference for counselling predicts postpartum depression. Verkerk, G.J.M.; Denollet, Johan;