The Causal Effect of Retirement on Health-Care Expenditures

Transcription

1 The Causal Effect of Retirement on Health-Care Expenditures Wolfgang Frimmel Johannes Kepler University of Linz Gerald J. Pruckner Johannes Kepler University of Linz (Draft paper: January 30, 2016) Abstract Pension systems and their reforms are often discussed in the context of financial viability. These debates grow in intensity with the aging of the population in industrialized countries. However, changes in retirement age may create unintended side effects for retirees health or health care costs. This paper empirically analyzes the effect of (early) retirement on individual inpatient and outpatient health-care expenditures in Austria. We use comprehensive labor market and retirement data from the Austrian Social Security Database combined with detailed individual inpatient and outpatient health expenditures for the province of Upper Austria. To account for the endogeneity in retirement decisions, we exploit exogenous variation in the early retirement age induced by two Austrian pension reforms in 2000 and We find significant effects of retirement on health expenditures. For both sexes, retirement decreases subsequent expenditures for outpatient medical attendance and hospitalization. The quantitative effects are stronger for females. A further analysis on a disaggregated level as well as a separate analysis for groups with different socio-economic status indicate a positive causal effect of retirement on individual health. However, the analysis also reveals behavioral changes in the utilization of health-care services that do not necessarily reflect health status improvement. The latter argument seems in particular relevant for white-collar workers. JEL Classification: I11, I12, J26, H51 Keywords: retirement, health-care expenditures, instrumental variable The usual disclaimer applies. This research was funded by the Christian Doppler Laboratory Aging, Health and the Labor Market.

2 1 Introduction With the aim of making pension systems more financially sustainable, most OECD countries have introduced reforms that encourage or enforce people to work for longer periods of their lifetimes. Population aging, a decline in fertility, and the recent economic crisis have certainly increased the pressure for decisive action. However, pension reforms are typically being accompanied by adverse social and economic effects that may influence individual well-being. For example, a change (increase) in retirement age may create unintended side effects for retirees health or health-care costs. In fact, a series of studies suggest a positive effect of early retirement on individual health or a decrease in mortality due to retirement. Most of the literature on health effects of retirement focuses on subjective self-reported health status. Such studies can be criticized, not least due to the fact, that answers to questions about mental and physical health can be expected to vindicate the retirement decision. In this paper, we study retirement effects on inpatient and outpatient healthcare expenditures using administrative health-care data. One of the advantages of using such data is their objective nature. First, the effects of retirement on out-of-pocket healthcare expenditures are important from the perspective of financing the health-care system. Causal empirical evidence would allow the health insurance funds an informed assessment of future health-care costs triggered by (early) retirement reforms. Second, health-care expenditures serve as a proxy for the individual health status. However, different expenditure categories reflect individual health differently. For example, the utilization of certain health-care services such as routine dental visits or other medical check-ups have a clear preventive character. Health-conscious people can be expected to utilize such services more often than are less health-conscious patients. As a consequence, higher expenditures for the utilization of health-care services may not necessarily indicate a worse health status, but also reflect more risk-averse behavior by patients and/or physicians. Even it is difficult to unequivocally distinguish health effects from behavioral (risk-aversion) effects, the level of detail in our register data provides evidence whether different components of expenditures reflect the utilization of curative (to improve poor health) or preventive (to maintain good health) services. For instance, we interpret the number and length of hospital stays and the consumption of medical drugs as better indicators of health status than expenditures on medical attendance in the outpatient sector. In our empirical analysis, we try to identify retirement effects for these different expenditure categories. To empirically analyze the effect of (early) retirement on individual inpatient and outpatient health-care expenditures in Austria, we use labor market and retirement data from a comprehensive employee-employer matched dataset and combine this information with detailed individual health register data for the province of Upper Austria. To account for the endogeneity in retirement decisions, we identify a causal effect of retirement by 2

3 exploiting exogenous variation in the early retirement age induced by two Austrian pension reforms in 2000 and In these reforms, the eligibility age of early retirement was gradually increased from 60 to 65 for men and 55 to 60 for women, respectively. We find significant effects of retirement on health expenditures. For both sexes, retirement decreases subsequent expenditures for outpatient medical attendance and hospitalization. The quantitative effects are stronger for females. Retirement of women reduces their expenditures for outpatient doctor visits by approximately 18.5% of the standard deviation and for hospital treatment by 11.9% of the standard deviation (6.8% and 7.9% for males, respectively). However, we do not find significant effects of retirement on expenditures for medication. On a disaggregated level, we find reductions for males in outpatient doctor expenditures for GPs, ENT specialists, orthopedists, psychiatrists, and diagnostic services. Female outpatient expenditures decrease for internists, dentists, orthopedists, psychiatrists, and diagnostic services. The results indicate a positive causal effect of retirement on individual health. However, the analysis also reveals behavioral changes in the utilization of health-care services that do not necessarily reflect health status improvement. The latter argument seems in particular relevant for white-collar workers. Literature The available empirical literature on retirement effects on health provides conflicting evidence. Those who find a positive impact of retirement stress the workplace burden in physical and mental terms that will be eliminated by retirement. In contrast, negative retirement effects refer to the loss of professional responsibilities, the lack of physical and mental activities after retirement, and unhealthy lifestyles including alcohol abuse. A judgement of the existing evidence is difficult since the studies rely on different outcomes and identification strategies. This questions external validity of single results together with the fact that pension systems in different countries vary widely. Cross sectional studies find in general that people who retire early have worse health after retirement. Obviously, these results are highly questionable since (many) persons who retire early can be expected to do so due to health reasons. The selection into retirement is not adequately addressed in such studies, and the results cannot be interpreted causally. There is a growing literature that addresses the selection into retirement by using longitudinal data or quasi-experiments. Most of these studies often based on subjective health measures report positive effects of retirement on health. In their longitudinal study of civil servants in the UK, Mein et al. (2003) find that retiring at the mandatory age of 60 (as compared to continuing working) had no effects on self-reported physical health, but was associated with improved mental health, in particular among high socio-economic status (SES) groups. Coe and Lindeboom (2008) use an early retirement window offer to instrument for retirement and do not find detrimental health effects for men due to early retirement. The authors report temporary increases in self-reported health and health improvement for well-educated workers. The GAZEL cohort study for 3

4 older workers from the French national gas and electricity company (Westerlund et al., 2003) suggests a relief in the burden of self-perceived ill-health by retirement in all groups of workers apart of the ones with perfect working conditions. 1 Studies that utilize SHARE (Survey on Health, Aging, and Retirement in Europe) data support the positive effects of retirement on self-reported health. For example, Coe and Zamarro (2011) exploit variation in retirement ages in several European countries and find a reduced probability for reporting deterioration in health after retirement. Shai (2015) identifies the retirement decision by an exogenous increase in the mandatory retirement age for men in Israel. The author agues that compulsory employment at older ages impairs self-reported health and that the effects are stronger among lower SES groups. These findings are supported by data from the Israeli Household Expenditure and Health Surveys. In utilizing SHARE data and two other datasets for the UK (English Longitudinal Study of Ageing, ELSA) and the U.S. (Health and Retirement Study, HRS), Horner (2014) finds a large positive effect of retirement on subjective well-being that fades over subsequent years. Two studies that exploit changes in pension regulation in France and the Netherlands also find positive health effects (Blake and Garrouste, 2012) and a reduction in the probability of death (Bloemen et al., 2013). The most recent quasi-experimental evidence is provided by Hallberg et al. (2015). The authors analyze a retirement offer to Swedish military officers 55 years of age or older and find a significant reduction in days of inpatient care and a lower risk of mortality. The reduction in days of hospitalization is greater for lower SES groups. The given interpretation is that the effect may be linked to less stress and less exposure to workplace hazards. Among the papers that identify negative health consequences of retirement, Dave et al. (2008) find an increase in difficulties in mobility, daily activities and mental illness. Rohwedder and Willis (2010) who exploit changes in retirement policies in the U.S. and in European countries, find that retirement is associated with a deterioration in cognitive abilities. A very convincing identification strategy is offered by Kuhn et al. (2010). An exogenous change in Austrian unemployment rules that allowed workers in eligible regions to withdraw from the workforce up to 3.5 years earlier as compared to their counterparts in non-eligible areas increases the probability of dying before the age of 67 by 13 percent for males and has no effect on females. The analysis of death causes for men indicates a higher incidence of cardiovascular disorders. In a similar vein, based on ELSA database for the UK, Behncke (2012) finds that retirement increases the risk of cardiovascular disease and cancer which is also reflected in increased risk factors such as body mass index (BMI and blood pressure). Finally, Hernaes et al. (2013) exploit data from a series of changes in retirement policy in Norway. Based on administrative data that cover the population of Norway, IV estimates show no effect of retirement age on mortality. The remainder of the paper is as follows. Section 2 covers briefly the institutional 1 Using the same database, Vahtera et al. (2009) find in addition a decrease in sleep disturbances. 4

5 background of the Austrian pension and health-care system. In Section 3, we describe our data (3.1), discuss our estimation strategy (3.2) and show descriptive statistics (3.3). Section 4 presents the estimation results. Section 5 provides preliminary conclusions and the future research agenda. 2 The Austrian pension and health-care system The Austrian health-care system Austria represents a Bismarckian-type health-care system that provides for the whole population universal access to medical services. mandatory health insurance covers all expenses for medical care in the inpatient and outpatient sector including those for medication. The Health insurance is offered by nine provincial health insurance funds (in German, Gebietskrankenkassen ) depending on occupation and place of residence. The insurance funds are responsible for all private employees and their dependents and represent approximately 75 % of the population. 2 Patients pay a prescription charge for medical drugs, and several insurance funds charge a small deductible or copayment. The expenses for doctor visits and medication are funded by wage-related social security contributions of employers and employees. The expenditures for hospitalization are co-financed by social security contributions and general tax revenues from different federal levels. After retirement, insured persons still have unlimited access to health-care services. contributions any longer. The Austrian pension system However, retirees do not pay social security The Austrian public pension system covers all private sector workers and provides early retirement pensions 3, old-age pensions and disability pensions. Public pensions are by far the most important income source of retirees in Austria. The amount of the pension is based on the number of insurance months collected during working life and the assessment base, which consists of the 15 best annual earnings years for most individuals in our sample. 4 The average pension in Austria for men is 76.6 percent of an average worker s earnings (as compared to the total OECD average of 5.5 percent, in 2012). There are capped penalties for retiring before the statutory retirement age of 65 years. Austrian men retire on average at the age of 60.6 (in 2014). This rather low retirement age implies extraordinarily high take-up rates for early retirement options and disability pensions. The latter is disproportionately high among men, as this is the only feasible way to retire before 60. Austrian women retire on average at the age of 58.5, hence only 1.5 years below their statutory retirement age of 60. Only in very recent 2 The rest of the population is covered by special social insurance institutions providing health insurance for farmers, civil servants, and self-employed people. 3 The most common form of early retirement is due to long periods of insurance. There was an early retirement option for the long-term unemployed until 2004 and for disabled workers until May With the pension reform 2003, the system has been changed to a so-called pension account where all contributing insurance years are part of the assessment base. 5

6 years, average retirement age has increased due to legal restrictions for early retirement, however, during our main sample period, average retirement age of men was still clearly below 60. The low labor force participation among the elderly can be attributed partly to disincentives provided by the Austrian pensions system (Hofer and Koman, 2006). Hanappi (2012) computed the social security wealth and accrual rates for Austria and finds that the social security wealth peaks at the age of 63 for men, hence creating strong disincentives for working beyond 63. In order to smooth the transition into retirement, the Austrian government introduced old-age part-time schemes for older employees in the early 2000s, where working time reductions of elderly workers are subsidized. This scheme often ends up in early retirement (Graf et al., 2011). Firms play also an important role in the retirement decision. Special severance payments (golden handshakes) paid to the worker in case of leaving the job bring along tax advantages for the employer and the employee. Frimmel et al. (2015) show that steeper seniority wage profiles in firms lead to earlier job market exit. During our sample period 1998 to 2012, several major reforms of the Austrian pension system took place. While the later reforms (2004 or later) do not strongly affect retirees in our sample because of typically long transition periods, earlier reforms changed retirement eligibility, early retirement age and financial incentives for workers in our sample. In particular, we use the gradual increase of early retirement age for different quarters of birth cohorts as an exogenous variation in the probability of retirement (see section 3.2). The suspension of certain early retirement schemes, i.e. early retirement due to reduced working capacity or due to long-term unemployment, also affect our sample such that earlier retirement became increasingly difficult. Over time, early retirement was no longer possible except in case of disability 5 or long insurance periods (45 years for men, and 40 years for women). 6 Due to harsh public protest against the new eligibility restrictions, a corridor pension at age 62 was introduced in Research design In this section, we present the data to be used in the empirical analysis and discuss the estimation strategy to identify a causal effect. The section also provides descriptive statistics for our variables of interest. 5 Workers need an external doctor to verify their reduced working capacity 6 Both options were further restricted in

7 3.1 Data The empirical analysis is based on several administrative data sources for the Austrian province of Upper Austria. All labor market and retirement-related information is gathered from the Austrian Social Security Database (ASSD). It is a matched employeeemployer dataset collected to verify pension claims for all Austrian workers in the private sector (Zweimüller et al., 2009) and contains detailed information on workers employment and earnings histories and basic socio-economic characteristics such as age, broad occupation, experience or tenure. The ASSD also records information on the start of the pension and pathways into retirement, i.e. disability pension, early retirement or regular retirement. We then combine the labor market and retirement information with inpatient and outpatient health-care expenditure data from the Upper Austrian Health Insurance Fund. These register data include detailed information on expenditures for doctor visits and for medication at the ATC ( Anatomical, Therapeutic, Chemical ) classification code level 7 in the outpatient sector. Inpatient information covers hospital expenditures, the number of days in hospitalization, and admission diagnoses for each individual. We focus on male and female private sector workers born between 1933 and 1957 and observe their health-care expenditures in the period between 1998 and We observe individuals on a quarterly base, hence health-care expenditures cover all expenses within a quarter. Individuals are required to retire after 1998, and we exclude individuals who have special retirement regulations, i.e. heavy labor workers, workers with more than 45 insurance years and public sector workers. We do not exclude individuals retiring due to disability pension, even if this may indicate a health problem. However, as a robustness check, we also present results if we exclude those who retire through disability pensions. This leaves us with 1, 319, 969 individual-quarter observations for men and 2, 073, 845 individual-quarter observations for women. This corresponds to 46, 999 men and 81, 916 women, respectively. The panel is unbalanced if individuals died before the end of the observation period. 3.2 Estimation strategy We examine a series of important health-care expenditure variables for the inpatient and outpatient sector to analyze the effect of retirement for male and female workers separately. We estimate the following empirical model: expenditures iq = β 0 + β 1 retired iq + β 2 age iq + β 3 age 2 iq + time q + µ i + ɛ iq (1) 7 We do not have data for medication for prescription-free medication, such as headache pills. 7

8 where we explain health expenditures of individual i in quarter q with a binary indicator whether the individual is retired in quarter q. We further control for a second-order polynomial of age in months (age, age 2 ) and add year-quarter fixed effects (time) to account for trends or time-related effects in health-care expenditures or retirement behavior. We do not include further socio-economic characteristics, since the longitudinal structure of our dataset allows us to estimate individual fixed-effects captured by the parameter µ i. The individual fixed-effect controls for all time-invariant individual characteristics, such as occupation, industrial sector, educational attainment, ability, general health status or genetic endowment. The fixed-effect, however, cannot account for time-varying heterogeneity, which may influence individual s health-care expenditures and the retirement decision. So even with an individual fixed-effect, we cannot perfectly rule out a remaining correlation between retirement and further time-varying confounding factors of ɛ iq. In order to account for this potential omitted variable bias, we suggest an IV approach where we exploit exogenous variation in the decision to retire from two Austrian pension reforms in 2000 and Pension reforms To guarantee fiscal sustainability of the public pension system, the Austrian government implemented two large pension reforms in 2000 and One important feature of both reforms was the gradual increase of the eligibility age for early retirement. The first reform in 2000 increased the early retirement age by 1.5 years. This increase was conducted step-wise by different quarters of birth cohorts. More precisely, men born before October 1940 were still eligible for early retirement at age 60, whereas for men born in the fourth quarter of 1940 the eligibility age was increased by 2 months. For every subsequent birth quarter, the eligibility age was further raised until the total increase of 1.5 years was reached. The same rule applied for women, where women born after September 1945 had a 2 months higher eligibility age than women born before. Overall, the 2000 pension reform aimed at increasing eligibility age for early retirement from 60 to 61.5 for men, and from 55 to 56.5 for women. In the second reform in 2003, the eligibility age for early retirement was further raised from 61.5 (56.5 for women) to 65 (60) by a similar step-wise increase for each quarter of a birth cohort. Figure 1 shows the development of early retirement age over birth quarters for men, Figure 2 shows the equivalent picture for women. It must be noticed that the introduction of the corridor pension at age 62 circumvents the gradual increase of early retirement age. Hence the eligibility age is practically capped at age 62 for men. Men (women) with more than 45 (40) insurance years were exempted from this reform. Further relevant changes of the reform was the step-wise extension of the assessment base from the best 15 earning years to lifetime earnings, higher penalties for early retirement from 2 to 4 percent per year (capped at 10%) and also a temporary extension of unemployment benefits for certain birth cohorts from 52 weeks to 78 weeks. Staubli and Zweimüller (2013) analyzed the employment effects of both pension reforms and find 8

9 that raising the early retirement age increased employment by 9.75 percentage points for men and 11 percentage points for women, and that the reform had substantial spillover effects on the unemployment insurance program. Spillover effects to disability insurance were reported to be small (+1.3 percentage points). Effects were largest for high-wage and healthy workers. Instrumental variable To identify the causal effect of retirement, we suggest to exploit exogenous variation induced by the two pension reforms described above. Individuals of different birth quarter cohorts are endowed with a different exogenously determined eligibility age of early retirement. For the first-stage regression, we partly follow Staubli and Zweimüller (2013). We define our instrumental variable as a binary indicator being 1 if the individual is below the early retirement age in quarter q, conditional on quarter-year fixed effects and a secondorder polynomial of age in months. Hence, the first-stage estimation can be written as retired iq = γ 0 + γ 1 1[age iq < era i ] + γ 2 age iq + γ 3 age 2 iq + time q + µ i + η iq (2) with era i as an individual s eligibility age for early retirement. Our parameter of interest in the first-stage is γ 1 which measures the impact of the individual-specific early retirement age (with respect to the birth quarter of an individual) on the probability of being retired. Based on the findings of Staubli and Zweimüller (2013), we expect γ 1 to be negative; a higher required early retirement age is expected to decrease the probability of being retired in quarter q. Identifying assumptions First, the validity of the instrument requires is a significant effect of the early retirement age on the probability of being retired, so γ 1 0. Second, we need to assume that the change in early retirement age affects health-care expenditures through the changed probability of retirement only, and there is no direct channel of the reform on health-care expenditures. Third, the exclusion restriction of the instrument requires the individual-specific early retirement age to be uncorrelated with any confounding factors included in ɛ iq, conditional on covariates and individual fixed-effects. There may be some potential concerns with respect to the credibility of the instrument. One major objection against the instrument may be that the impact of an exogenous increase in early retirement age on health-care expenditures does not only capture the effect of retirement but also includes age effects or time trends. We control in our estimation for a second-order polynomial of age measured in months and for year-quarter fixed effects. These variables control for potential age and time effects of the increased early retirement age. We further plan to conduct a placebo reform test to convincingly show that we really capture a reform effect. Second, retirement is associated with a reduction in earnings, because pension claims are typically below earnings. Moreover, the pension reforms raised penalties for retirement 9

10 before the statutory retirement age. One may be worried that the reform induces negative income effects which may spill-over to the health-care sector. The change in financial incentives induced by the reform should be negligible since these reforms include relatively long transition periods, and are not relevant for the majority of our sample. However, there may still be an income effect through longer employment periods induced by the reform. In a robustness check, we additionally include earnings (labor income or pension payment if already retired) into our model to check for those potential income effects Descriptive statistics Our sample comprises of 46, 999 men born between 1940 and 1955 and 81, 916 women born between 1945 and Table 1 summarizes the health-care expenditure outcomes and the covariates for men (Column (I)) and women (Column (II)). 41 percent of men and 52 percent of women retire until the end of our observation period in The higher retirement incidence among women can be explained by the lower statutory retirement age of 60. Men are more likely blue-collar workers, and their monthly average income is almost twice as high than those of women. We observe significant differences in utilization of health-care services between men and women. Overall, women have on average higher health-care expenditures than men in basically all dimensions. Women spend on average of e per quarter for doctor visits as compared to e 85.7 for men. Similarly, female drug expenditures of e 72.9 per quarter outweigh men s expenditures of e Only for hospital expenditures, we observe a slightly higher amount for men. Disaggregated expenditures for medical attendance reveal further significant differences in utilization. While average quarterly expenditures for GPs, internists, psychiatrists and orthopedists are relatively similar, women have significantly higher expenditures for diagnostic services and dentists. For disaggregated drug expenditures, we find significantly higher expenditures for females for psychotropic drugs, anticancer drugs, and drugs for muscular diseases. In other drugs categories, differences stay quite modest. Figures 3 (for men) and 4 (for women), depict aggregate doctor expenditures for three different birth cohorts. Less surprisingly, expenditures substantially increase in age. Although women have higher doctor expenditures at the age of 55, the subsequent increase by age is steeper for men. We also see that the younger male birth cohorts have on average higher expenditures, particularly at younger age. After the age of 62, the birth cohorts converge in their health-care utilization behavior. The pattern for women is similar, however, convergence of cohorts occurs earlier. The picture for cohort differences in medication expenditures is less clear, as shown in Figures 5 and 6. Cohorts resemble more across age. For men, the youngest cohort seems to have on average higher medication expenditures after the age of We refrain from including income into our baseline model, as we consider income as a potential bad control. 10

11 4 Results Table 2 provides first-stage fixed-effects estimation results, Table 3 summarizes fixedeffects and fixed-effects IV results for aggregated health-care expenditures. The results for disaggregated expenditures are shown in Table 4 for medical attendance and in Table 5 for drug use. All estimation results are presented separately for men and women. The estimations include individual fixed-effects, a second-order polynomial of age in months, and year-quarter dummies. We report clustered standard errors on individuals throughout the paper. 4.1 First stage estimation results Table 2 depicts our first-stage results of equation (2) for the sample of men and women separately. As was pointed out before, the instrument is equal to one if the individual s age in quarter q is still below the early retirement age. We find - conditional on the covariate and individual fixed effects - a statistically significant negative effect of early retirement eligibility in quarter q on the probability of being retired in the same quarter. The estimated effects differ between men and women; retirement probabilities decrease by 16.3 percentage points for men, and by 8.3 percentage points for women. The very high F-statistic of the instrumental variable indicates that it is sufficiently strong. Our first-stage results are in line with the findings by Staubli and Zweimüller (2013). 4.2 Aggregate health expenditures Table 3 presents our estimates for aggregated quarterly health-care expenditures for doctors, medication, hospitalization, and for hospital days. Fixed effects estimation results The first columns in each of the four health-care measures present the impact of retirement based on the individual fixed-effect estimation. For men, we find significant effects for all outcomes. Doctor expenditures decrease by approximately e 2.6, medication expenditures increase by almost e 6, hospital expenditures decrease by e 19, and hospital days also decrease slightly. We find similar fixed-effects regression results for females including a slight decrease in medication expenditures. Overall, the results indicate positive health effects of retirement, at least in terms of expenditures. The decrease in expenditures for doctors may either reflect a positive health effect or a change in behavior, in that people simply go less often to doctors once they are retired. In particular, the reduction in hospital expenditures together with fewer hospital days can be interpreted as a positive health effect of retirement. Nevertheless, these results could still be biased by time-varying confounding factors, i.e. negative health shocks at work. Fixed effects instrumental variable estimation results The second columns for each 11

12 outcome show the results of an instrumental variable estimation. As compared to simple fixed-effects estimators, the IV-results yield qualitatively similar results. There are, however, quantitative differences, indicating that unobserved time-varying factors indeed have an impact on health-care expenditures and the probability of retirement. For men, doctor expenditures decrease by e 11.4, which is roughly 7 percent of a standard deviation of doctor expenditures. We do not find a statistical significant effect on medication expenditures; the IV estimator is close to zero indicating that men do not take more or less drugs in retirement. The causal effect of retirement on hospital expenditures has remarkably increased to minus e 87.8, which is approximately 8 percent of a standard deviation of hospital expenditures. This effect is highly significant and also large in size. There is - in contrast to the results without accounting for endogeneity - no reduction in hospital days. These somewhat contradicting results for the inpatient sector raises questions about the interpretation of these effects. On the one hand, the reduction in hospital expenditures may clearly indicate a positive health effect of retirement. On the other hand, we do not find an equivalent effect on the days of hospitalization, which might indicate behavioral changes either by retirees, who are in hospitals, or by (the doctors in) the hospitals. In order to separate health from behavioral effects, we require a more thorough analysis on a disaggregated level, i.e. whether we can observe substantial changes in diagnoses and treatments. For women, the results get even more pronounced. Doctor expenditures decrease by e 35.4, whereas the impact on medication expenditures remains insignificant. Hospital expenditures decrease by e (11.8 percent of a standard deviation), and hospital days are reduced by 0.16 days. The effect on hospital days is significant at the 10 percent level only. The larger effects for females imply that women benefit more from retirement than men, or their behavioral changes are stronger. 4.3 Disaggregated doctor expenditures Table 4 summarizes IV estimates for different doctor expenditure categories. We consider general practitioners (GPs), internists, diagnostic services, psychiatrists and psychologists, orthopedists, dentists, ENT (ear, nose, throat) specialists, eye doctors, and a catch-up category for all remaining specialists. We find that the lower doctor expenditures at the aggregate level result from decreased expenditures in several categories. For men we see expenditure reductions for GPs ( e 4, diagnostic services (( e 1.6), psychiatrists (( e 0.9), orthopedists (( e 1.2) and ENT specialists (( e 0.9). Given the mean values of the outcome variables, the size of those effects is significant. The cost reductions for psychiatrists or orthopedists may indicate improved mental health of retirees due to lower stress levels and a decrease in heavy work. It is less evident to interpret the reduction for ENT specialists and GPs as a health effect, since the findings might rather reflect 12

13 behavioral changes in the utilization of health-care services (e.g. no need for GP visits for sick leave after retirement). Equivalently for women, we find expenditure reductions for diagnostic services ( e 2.5), psychiatrists ( e 3), orthopedists ( e 3.3) and most surprisingly dentists ( e 10). While reduction for diagnostic services, psychiatrists and orthopedists may again reflect a positive health effect, we consider the substantial reduction for dentist expenditures much more as a change in female health behavior after retirement. 4.4 Disaggregated drug expenditures On an aggregate level, we do not detect significant changes in medication expenditures induced by retirement. This either means there is no causal effect on drug expenditures, or the effects for different types of drugs cancel out. In order to test these hypotheses, we estimate the causal effect of retirement on different drug categories according to the ATC classification code, i.e. antiinfectives, drugs for cardiovascular diseases, muscular diseases, respiratory diseases, sensory organ diseases, anticancer drugs, psychotropic drugs, and a catch-up variable for the remaining drug expenditures. Table 5 summarizes the IV estimates for the different drug categories. For men, we find a negative effect for psychotropic drugs and muscular diseases and a positive effect for respiratory diseases. For women, we only find a reduction in expenditures for muscular diseases. Overall, we cannot provide strong evidence for an effect of retirement on medication expenditures for men and women. 4.5 Heterogeneous effects for selected health expenditures In this section, we analyze heterogeneous treatment effects of retirement to identify socioeconomic groups who may be affected by retirement differently. First, we distinguish between blue-collar and white-collar workers since these occupation groups obviously differ in physical and psychological work load (Table 6). Second, we differentiate between income groups in dividing men and women into groups with labor income above and below the sexspecific median income 9 (Table 7). In our analysis, we focus on expenditures for resident doctors and hospitals as well as expenditures for GPs and dentists. We find substantial differences between male blue-collar and white-collar workers. Doctor expenditures for blue-collar workers are reduced by e 28 per quarter, while the effect for white-collar workers remains insignificant. Similarly, the reduction in hospital expenditures for blue-collar workers ( e 144) is more than twice as large as the reduction for white-collar workers ( e 66). The equivalent pattern can be observed for GP expenditures. Overall, male blue-collar workers seem to benefit from retirement substantially 9 Income is measured at the beginning of our sample, hence at the age of 55 for men and the age of 50 for women. 13

14 more than their white-collar counterparts. For women, we find a reduction in doctor expenditures for both blue-collar and whitecollar workers of approximately e 36. Moreover, there is a significant reduction in hospital expenditures and GP expenditures for female blue-collar workers, whereas the effect for white-collar workers is insignificant. Interestingly, the observed negative coefficient for dentist expenditures is only apparent for female white-collar workers, which might indicate, that behavioral changes in the utilization of health-care services are more prevalent among this group, whereas female blue-collar workers tend to benefit from a positive health effect. A very similar pattern can be observed if we distinguish between high and low income individuals. As can be seen from Table 7, we find reductions in expenditures for resident doctors, hospitals and GPs for low-income men, while there is only a small reduction of GP expenditures for high-income males. Also for women, we observe a clear income gradient of retirement effects. Low-income women seem to benefit much more from retirement than high-income women. Again, the dentist effect only occurs for high-income women. 4.6 Further robustness checks Our baseline sample consists of individuals who either retire through an old-age pension, early retirement or due to disability. The latter way of retiring depends on the healthstatus of the individual and has become a common way of retirement particularly of male blue-collar workers. 10 To see whether and to what extent our results are driven by the less healthy group of workers, we conduct a robustness check where we exclude all individuals who retire through disability pensions. 11 Table 8 summarizes the results of our estimates based on the reduced sample. We find that reductions in health-care expenditures induced by retirement are only partly driven by the less healthy group of disability pensioners. For both men and women, our previous estimates remain qualitatively the same, but are somewhat smaller in size. The estimates for men react stronger than for women, which simply reflects the fact that men are more likely to use disability pensions as a pathway into retirement than women. In order to check whether our estimates are influenced by potential income effects generated by the retirement decision or the pension reform, we include income 12 as another covariate and re-estimate our model for all outcomes. inpatient and outpatient outcome variables are shown in Table 9. The results for the aggregated One interesting finding is that income appears to be rather important in the first- 10 Approximately one third of male blue-collar workers retire through disability pensions (Frimmel et al., 2015). 11 Moreover, as was shown by Staubli and Zweimüller (2013), there are small spillover effects of an increase in early retirement age on the take-up of disability pensions. 12 Income may consist of either labor income, unemployment benefits or pension payments for retired individuals. 14

15 stage regression, which is also reflected by a lower F-statistics of the instrument. The coefficients for income in the second-stage are insignificant for males and significant but very small in size for women. Income obviously plays an important role for the retirement decision, but is less relevant for the utilization of health-care services. The latter insight can be explained by the full coverage of the Austrian health-care system which is totally independent from individual income. Comparing the estimates from the model with income as a control variable with the estimates of our baseline model in Table 3, we see qualitatively and quantitatively very comparable results. We conclude that our estimates should not be biased from a potential income effect. 5 Conclusions and outlook In this paper, we look at the causal effect of retirement on health-care expenditures. We identify the causal effect by exogenous variation induced by two pension reforms in Austria, which gradually increased the early retirement age for men and women. We find significant effects of retirement on health-care expenditures. For both sexes, retirement decreases subsequent expenditures for outpatient medical attendance and hospitalization. The quantitative effects are stronger for females. Retirement of women reduces their expenditures for outpatient doctor visits by approximately 18.5% of the standard deviation and for hospital treatment by 11.9% of the standard deviation (6.8% and 7.9% for males, respectively). We do not find significant effects of retirement on expenditures for medication. On a disaggregated level, we find reductions for males in outpatient doctor expenditures for GPs, ENT specialists, orthopedists, psychiatrists, and diagnostic services. Female outpatient expenditures decrease for internists, dentists, orthopedists, psychiatrists, and diagnostic services. The results indicate a positive causal effect of retirement on individual health. However, the analysis also reveals behavioral changes in the utilization of health-care services that do not necessarily reflect health status improvement. The latter argument seems in particular relevant for white-collar workers. At least two (preliminary) conclusions can be drawn from our analysis. First, the attempt of the Austrian government to increase the (early) retirement age of workers is likely be accompanied by negative health effects. Second, it is important to be noticed that retirement has not only an impact on people s health status, but may also change individual health behavior. The behavioral changes in the utilization of health-care services may have something to do with the availability of time, shirking at the workplace (doctor consultations during working hours), and eventually lower incentives for preventative measures after retirement. In the next steps, we would like to differentiate in more detail between health effects and behavioral changes in health-care utilization. For this purpose, we (i) plan to take a closer look at the inpatient sector and see whether treatments and diagnoses in hospitals 15

16 may have changed, and (ii) analyze the effect of retirement on screening participation, which clearly reflects a behavioral effect. Due to the skewed distribution of several disaggregated outcome variables, we also consider alternative methodological approaches. For example, we will check the robustness of our results by Tobit model estimations. Moreover, we plan further robustness checks with respect to time. It is important for the identification of the causal effect that we properly control for time trends. To make sure that we unambiguously capture the effect of the reform and not a time effect, we apply a placebo test by assuming a hypothetical reform for different - in the real world unaffected - birth cohorts. 16

17 References Behncke, S. (2012). Does Retirement Trigger Ill Health? Health Economics, 21, Blake, H. and Garrouste, C. (2012). Collateral Effects of a Pension Reform in France. Health econometrics Data Group, working paper 12/16. Bloemen, H., Hochguertel, S. and Zweerink, J. (2013). The Causal Effect of Retirement on Mortality: Evidence from Targeted Incentives to Retire Early. IZA discussion paper, no Coe, B. and Lindeboom, M. (2008). Does Retirement Kill You? Evidence from Early Retirement Windows. IZA discussion paper, no and Zamarro, G. (2011). Retirement Effects on Health in Europe. Journal of Health Economics, 30, Dave, D., Rashad, I. and Spasojevic, J. (2008). The Effects of Retirement on Physical and Mental Health Outcomes. Southern Economic Journal, 75 (2), Frimmel, W., Horvath, T., Schnalzenberger, M. and Winter-Ebmer, R. (2015). Seniority Wages and the Role of Firms In Retirement Decisions. Tech. rep., WP1505, Department of Economics. Graf, N., Hofer, H. and Winter-Ebmer, R. (2011). Labour Supply Effects of a Subsidised Old-Age Part-Time Scheme in Austria,. Zeitschrift fuer Arbeitsmarktforschung. Hallberg, D., Johansson, P. and Josephson, M. (2015). Is an early retirement offer good for your health? Quasi-experimental evidence from the army. Journal of Health Economics, ( Hanappi, T. P. (2012). Retirement Behaviour in Austria: Incentive Effects on Old-Age Labor Supply. Tech. rep., The Austrian Center for Labor Economics and the Analysis of the Welfare State, Johannes Kepler University Linz, Austria. Hernaes, E., Markussen, S., Piggott, J. and Vestad, O. (2013). Does Retirement Age Impact Mortality? Journal of Health Economics, 32, Hofer, H. and Koman, R. (2006). Social security and retirement incentives in Austria. Empirica, 33 (5), Horner, E. M. (2014). Subjective Well-Being and Retirement: Analysis of Policy Recommendations. Journal of Happiness Studies, 15, Kuhn, A., Wuellrich, J. and Zweimueller, J. (2010). Fatal Attraction? Access to Early Retirement and Mortality. IZA working paper, no Mein, G., Martikainen, P., Hemingway, H., Stansfeld, S. and Marmot, M. (2003). Is Retirement Good or Bad for Mental and Physical Health Functioning? Whitehall II Longitudinal Study of Civil Servants. Journal of Epidemiology Community Health, 57,

18 Rohwedder, S. and Willis, R. (2010). Mental Retirement. Journal of Economic Perspectives, 24 (1), Shai, O. (2015). Is Retirement Good for Men s Health? Evidence Using a Change in the Retirement Age in Israel. Mimeo, Department of Economics, Hebrew University of Jerusalem. Staubli, S. and Zweimüller, J. (2013). Does raising the early retirement age increase employment of older workers? Journal of Public Economics, 108 (1), Vahtera, J., Westerlund, H., Hall, M., Sjsten, N., Kivimki, M., Salo, P., Ferrie, J., Jokela, M., Pentti, J., Singh-Manoux, A., Goldberg, M. and Zins, M. (2009). Effect of Retirement on Sleep Disturbance: The GAZEL prospective cohort study. Sleep, 32, Westerlund, H., Kivimki, M., Singh-Manoux, A., Melchior, M., Ferrie, J. E., Pentti, J., Jokela, M., Leineweber, C., Goldberg, M., Zins, M. and Vahtera, J. (2003). Self-rated Health Before and After Retirement in France (GAZEL): a cohort study. The Lancet, 374 (Issue 9705), Zweimüller, J., Winter-Ebmer, R., Lalive, R., Kuhn, A., Ruf, O., Büchi, S. and Wuellrich, J.-P. (2009). The Austrian Social Security Database (ASSD). NRN Working Paper, University of Linz. 18

19 6 Figures and tables Early retirement age Men 1940q1 1945q1 1950q1 1955q1 Birth cohort Early retirement age Women 1945q3 1948q3 1951q3 1954q3 1957q3 Birth cohort Figure 1: Early retirement age of men Figure 2: Early retirement age of women Avg. doctor expenditures age Cohort 1940q3 Cohort 1945q3 Cohort 1942q3 Figure 3: Doctor expenditures for men Avg. doctor expenditures age Cohort 1945q3 Cohort 1950q3 Cohort 1947q3 Figure 4: Doctor expenditures for women Avg. medication expenditures Avg. medication expenditures age Cohort 1940q3 Cohort 1945q3 Cohort 1942q3 Figure 5: Drug expenditures for men age Cohort 1945q3 Cohort 1950q3 Cohort 1947q3 Figure 6: Drug expenditures for women 19

20 Table 1: Descriptive statistics (I) Men (II) Women Aggregated outcome variables Doctor expenditures (167.65) (191.83) Medication expenditures (242.37) (272.52) Hospital expenditures (1, ) (1, ) Hospital days 0.26 (1.95) 0.29 (2.08) Disaggregated doctor variables GP (general practitioner) (38.79) (43.94) Internist 4.84 (25.59) 4.93 (25.45) Diagnostic services 5.95 (20.22) (28.44) Psychiatry 1.20 (11.32) 2.02 (16.99) Orthopedics 2.45 (19.93) 3.34 (23.28) ENT specialist 4.33 (15.60) 5.14 (17.09) Dentist (113.61) (131.36) Other doctor 7.69 (28.64) (34.32) Disaggregated medication variables Cardiovascular diseases (39.10) (93.55) Antiinfectives 2.14 (78.71) 2.42 (49.63) Psychotropics 3.60 (33.81) 7.71 (16.99) Muscular diseases 1.72 (10.38) 4.28 (20.78) Cancer 4.39 (140.63) (201.35) Respiratory diseases 3.89 (28.27) 3.75 (27.45) Sensory organ diseases 0.47 (6.44) 0.54 (6.40) Other drugs (72.40) (62.36) Covariates Age in years (3.05) (3.06) Retired until (0.492) (0.499) Income per month 2, (1, ) 1, (830.81) Blue-collar worker (0.492) (0.348) No. of observations 1,319,969 2,073,845 Descriptive statistics for outcome variables per quarter and selected individual characteristics. Expenditures and income in e. Standard errors in parentheses. 20