Retirement and its Consequences for Health in Australia Kostas Mavromaras, Sue Richardson, and Rong Zhu 31 January 2014. Draft: please do not cite or quote. Abstract This paper estimates the causal effect of retirement incidence and duration on health outcomes among mature age Australians. We utilize pension eligibility age in the context of Fixed Effects Instrumental Variables estimation. We find that becoming retired has positive and significant effects on health, both physical and mental, and more so for women. We find that longer time spent in retirement confers clear additional health benefits for women, but not clearly so for men. We also find robust evidence that health improvements due to retirement can only be traced for people with below median health status. JEL classif ication: I12; J1; J24; J26 Keywords: Retirement, health, Age Pension Funding from the Australian Research Council Discovery Project Grant DP0987972 is gratefully acknowledged. This paper uses unit record data from the Household, Income and Labour Dynamics in Australia (HILDA) Survey. The HILDA Project was initiated and is funded by the Australian Government Department of Social Services (DSS) and is managed by the Melbourne Institute of Applied Economic and Social Research (Melbourne Institute). The findings and views reported in this paper, however, are those of the authors and should not be attributed to either DSS or the Melbourne Institute. E-mail: k.mavromaras@flinders.edu.au. E-mail: sue.richardson@flinders.edu.au. Corresponding author. E-mail: rong.zhu@flinders.edu.au. 1
1 Introduction Countries around the world are experiencing rises in the average age of their populations, caused in part by rising life expectancy. The fiscal challenge that these present to governments has led many to consider strategies to extend the age at which people retire from the workforce. In the design of such strategies, it is important to understand why people choose to retire and what influences the timing of their decision. Where retirement is not compulsory, people retire for good reason. If governments press people to continue to work beyond the point at which they would have otherwise chosen to retire, this is likely to have consequences, some of which adverse. These consequences should be understood and, if possible, quantified. They will affect both the efficacy of alternative strategies and their impact on the wellbeing of the affected groups. Health is an important example. It both affects the decision to retire and is affected by retirement. While it is clear that poor health induces earlier retirement, the effect of retirement on health is more ambiguous, both theoretically and empirically. Retirement can exert a positive influence on people s health, for example through reduced stress and increased enjoyment of life. But it can also have a negative effect, for example through a reduced sense of purpose and the loss of social interactions. Understanding the net and causal effects of retirement on health is important, especially as many developed countries have already started raising the statutory state pension eligibility age, with the intention to induce postponed retirement (Hering and Klassen, 2010). For example, the Australian Age Pension eligibility age for women has been increased from 60 in 1995 at the rate of six months every two years reaching the (hitherto unchanged) eligibility age of 65 for males by 2014. From 2017, the pension qualifying age for the two genders will start rising again from 65 years old, by six months every two years, until it reaches the age of 67 in 2023. The success of such policies in reducing government expenditure on social security programs partially depends on the potential health effects of delayed retirement. It is possible that postponed retirement may lead to deteriorations in population health, in which case the alleviation of government financial stress from reduced pension expenditure may not be enjoyed in full, because of an offsetting increase in health care spending. 2
Establishing the causal effect of retirement on health is confronted with several difficulties. First, the selection into retirement is not a random process. Second, unobserved individual heterogeneity and time-varying confounding factors can affect both retirement decisions and health outcomes. Finally, a reverse causality problem can exist in the sense that retirement is triggered by poor health (Bound et al., 1999; Disney et al., 2006). The endogenous relation between retirement and health means that the causal impact of retirement on health cannot be easily established empirically (Dave et al., 2008; Bonsang et al., 2012). A recent development in the literature shows that the age of eligibility for social security benefits including pension benefits, can be used as an exogenous instrument that may influence the decision to retire without having a correlation with health. For example, in the United Kingdom, a significant portion of retirement income becomes available at age 65 for men and 60 for women (Bound and Waidmann, 2007). In the United States, 62 is the earliest age at which people can receive social security retirement benefits and 65 is the age at which retired people can claim full social security benefits (Bonsang et al., 2012). The variation in retirement induced by social security benefits has been used to identify the causal health effects of retirement. Using the second wave of the English Longitudinal Study of Ageing (ELSA), Bound and Waidmann (2007) find that retirement status has a small and positive causal effect on physical health of men, and the effect is not significant for women. With data from the US Health and Retirement Survey (HRS), Charles (2004) and Neuman (2008) both find that being retired preserves subjective health, while Bonsang et al. (2012) highlight a significant negative effect of retirement on cognitive functioning. A few other studies have utilised cross-country differences in the eligibility age for retirement benefits as the exogenous source of variation in retirement rates within Europe. With data from the Survey of Health, Ageing and Retirement in Europe (SHARE), Rohwedder and Willis (2010) and Mazzonna and Peracchi (2012) find that the causal effect of retirement on cognitive functioning is negative and statistically significant, while Coe and Zamarro (2011) find a large and positive effect of retirement on self-assessed health as well as a health index constructed using a rich set of health variables. The econometric challenges 3
of dealing with endogeneity in its different manifestations (i.e. non-random selection, two way causality and unobserved factors) most probably help to explain the conflicting results of the causal effect of retirement on health that are found in the literature. This paper takes several important steps to identify causality in a robust way. These steps include the use of an Instrumental Variable approach combined with Fixed Effects estimation (FE IV) to control for unobserved heterogeneity and for two-way causality. We use the Age Pension eligibility to instrument for retirement, following the logic established in the literature, namely, that eligibility age influences the retirement decision but does not influence health (Rohwedder and Willis, 2010). In contrast with most existing studies, we differentiate further between the causal effect on health of (i) retirement status (being retired or not, as a discrete event) and (ii) retirement duration (length of time spent in retirement, as a cumulative process). 1 We find that retirement status has positive and highly significant effects on the simpler measure of self-reported health, as well as the more complex measure of SF-36 and its physical and mental health constituents, and that these effects are in most instances stronger for women than for men and of high statistical significance for both men and women. The effect of retirement duration is also found to be positive and highly significant for Australian women and smaller with only a weak statistical significance for Australian men. When we examine the effect of retirement on the different components of physical and mental health of the SF36, we find considerable diversity in both strength and statistical significance. We pay particular attention to the different effects of retirement status and retirement duration for people with different health status. Specifically, we examine whether the health effects of retirement are different for individuals whose health is above or below the median value of each of the SF 36 health indices we use. Our analysis reveals substantial dispersion of the health effects of retirement, and we find that it is mainly the people with lower levels of health that experience the strongest health improvement from the transition into retirement status and from longer exposure to retirement. 1 The only study that has analyzed the health effects of both retirement status and retirement duration is Bonsang et al. (2012). Other existing studies on the causal health effects either focus solely on retirement status (Charles, 2004; Neuman, 2008; Rohwedder and Willis, 2010; Coe and Zamarro, 2011) or retirement duration (Coe et al., 2012; Mazzonna and Peracchi, 2012). 4
The remaining paper is organized as follows. Section two describes the Age Pension system in Australia. Section three describes the data and presents summary statistics. Section four discusses the empirical approach. Section five presents the estimation results and section six concludes. 2 Age Pension System in Australia The Australian system of retirement income support has three components (Barrett and Tseng, 2008; Agnew, 2013; Atalay and Barrett, 2013; Ryan and Whelan, 2013). First, the publicly funded Age Pension, second, the income from the mandatory employer-contributed superannuation, and third, the voluntary private retirement savings. There is no compulsory retirement age threshold in Australia, so that the eligibility age for accessing the different components of potential retirement income will depend on its source. Here we focus on the eligibility age for accessing the publicly funded age pension, as this is the age threshold we will use as our instrument for estimation. The eligibility for the publicly funded Australian Age Pension is subject to three qualifying conditions. These are that one has to have been an Australian resident for ten years, one has to pass an income and assets means test, and one has to have reached a certain age threshold. About 70 percent of the elderly population receive some Age Pension, and of those recipients, approximately two thirds receive the full amount of Age Pension, which in 2013 was AU$751.70 per fortnight for individuals and AU$1333.20 for couples. The qualifying Age Pension age threshold is presently 65 for men and women and is planned to start rising in July 2017 till it reaches 67 by 2023. Based on the thinking that eligibility thresholds cannot in themselves influence health outcomes, but can influence retirement decisions, this paper utilizes the relationship between the Age Pension eligibility thresholds and actual retirement timing and duration decisions to construct an Instrumental Variable in order to identify the causal effects of retirement on the health outcomes of Australians. 5
3 Data 3.1 The Household, Income and Labour Dynamics in Australia (HILDA) Survey This study uses data from the Household, Income and Labour Dynamics in Australia (HILDA) survey, which is the first and only large-scale, nationally representative household panel survey in Australia. Starting from 2001, HILDA collects annually rich information on people s demographics, education, labour market dynamics and health status (Richardson, 2013). This paper uses the first eleven waves (2001-2011) of the unconfidentialised HILDA, which contains information on the exact date of birth and the date of survey for each individual, which, together with the Age Pension eligibility ages enable us to identify accurately whether an individual has reached the pension qualifying age at the interview date. A few sample restrictions are applied to facilitate the analysis. First, we follow Bonsang et al. (2012) and focus on mature age people aged between 51 and 75. Second, as described in the previous section, the eligibility for the Age Pension is subject to a residency condition of living in Australia for at least ten years, so we exclude a few observations that do not meet the condition. Finally, observations with missing information on core variables used in this study are dropped. Our final sample is an unbalanced panel consisting of 36,713 observations for 7,269 persons in Australia between the years 2001 and 2011. 3.2 Variables and Descriptive Statistics Our definition of retirement follows those in French (2005) and Bonsang et al. (2012). An individual is defined to be retired if he/she is not in the labour force. HILDA contains a rich set of variables measuring individual health status. Respondents were asked to rate their health on a five-point scale: excellent, very good, good, fair and poor, based on which, we generate a dummy variable for subjective good health that is equal to one if an individual s self-assessed health is good, very good or excellent, and zero if fair or poor. 6
Table 1: SF 36: Physical and Mental Health Components Summary Measures Scales Measuring Physical Functioning (PF) Limitations to daily activities Physical Health Role Physical (RP) Limitations in work or activities caused by physical health Mental Health Bodily Pain (BP) General Health (GH) Social Functioning (SF) Role Emotional (RE) Vitality (VT) Mental Health (MH) Pain and limitations therefrom Health perception Social limitations Limitation in work or activities due to emotional health Fatigue and energy scales Feelings of anxiety and depression The self-assessed general health measure may provide a comprehensive picture of one s overall well-being; however, it is subjective and may suffer from reporting bias. For example, if people from different socioeconomic backgrounds perceive their health status differently, then the estimates of the effects of retirement on (self-reported) health can be biased. To alleviate this potential concern, we also use the health measures derived from the 36-item Short Form Health Survey (SF-36), an internationally tested and widely used tool for measuring health (Hemingway et al., 1997). In each wave of HILDA, respondents were asked all SF-36 questions about their physical and mental health. Out of the 36 questions, 22 fall in the category of physical health and 14 in the category of mental health. The physical health measures are grouped into four sub-categories (Physical Functioning (PF), Role-Physical (RP), Bodily Pain (BP) and General Health (GH)). Similarly the mental health measures are also divided into four sub-categories (Social Functioning (SF), Role-Emotional (RE), Vitality (VT) and Mental Health (MH)). These eight sub-categories are provided in a standardized form in the HILDA survey with a range from 0 100, the higher scores indicating better health. For the current analysis, we generate a physical health measure for each observation by calculating an average of the four physical health sub-categories and another one for the four mental health ones. We also create an overall health variable that summarizes both physical and mental health, by taking the simple average of all eight sub-categories. Summary statistics by gender are presented in Table 2. The overall retirement rate is 49 percent in the sub-sample (aged 51 75), with the gender-specific rate being 42 percent for 7
males and 56 percent for females. The average ages for men and women are similar, being slightly over 61 years old. As the pension eligibility ages were lower for women during the survey periods, the proportion of women (38 percent) meeting the age requirement of the Age Pension is found to be seven percent higher than that for men (31 percent). Furthermore, men are slightly better educated, more likely to be married, and have a larger family size than women. In terms of health measures, 76 percent of the age group reported having a good, very good or excellent health status, and we find no difference between the two genders. The summary statistics of SF-36 health measures show that Australian men are in slightly better physical, mental and overall health than women. Table 2: Summary Statistics All Male F emale Mean SD Mean SD Mean SD Demographic characteristics Retired 0.49 0.50 0.42 0.49 0.56 0.50 Age 61.27 6.73 61.31 6.72 61.23 6.74 Age eligible for Age Pension 0.35 0.48 0.31 0.46 0.38 0.49 Schooling 11.67 2.40 11.97 2.38 11.41 2.39 Married 0.74 0.44 0.81 0.40 0.69 0.46 Family size 2.24 1.05 2.37 1.12 2.13 0.97 Health measures Self reported good health 0.76 0.43 0.76 0.43 0.76 0.42 Physical health (SF 36) 69.78 24.00 70.92 23.61 68.73 24.29 Physical Functioning (PF) 76.01 23.88 78.11 23.44 74.09 24.12 Role Physical (RP) 70.90 40.24 72.37 39.59 69.55 40.78 Bodily Pain (BP) 67.33 25.39 68.99 24.93 65.80 25.71 General Health (GH) 64.87 22.41 64.22 22.21 65.48 22.57 Mental health (SF 36) 75.19 20.26 76.38 19.89 74.10 20.54 Social Functioning (SF) 81.36 24.59 82.49 23.94 80.34 25.12 Role Emotional (RE) 82.32 33.92 83.21 33.29 81.50 34.47 Vitality (VT) 61.05 20.33 62.58 19.91 59.65 20.62 Mental health (MH) 76.03 17.13 77.25 16.72 74.91 17.43 Overall health (SF 36) 72.48 20.65 73.65 20.43 71.41 20.79 N 36,713 17,527 19,186 N ote: Data Source: HILDA 2001 2011. 8
4 Empirical Approach To investigate whether retirement causally affects the health outcomes of mature age people in Australia, we consider the following health production function H it = R it β + X itγ + u i + ɛ it (1) where H it denotes a health measure, and R it is a binary variable indicating the retirement status of individual i at time t. X it is a vector of control variables including age, age squared, years of schooling, a binary variable indicating marital status and family size. A full set of state of residence dummies and wave dummies are also included. Time-invariant unobserved heterogeneity is controlled for by the individual fixed effect, u i, and ɛ it is the idiosyncratic error term. An ordinary least squares (OLS) estimation of equation (1) will yield a consistent estimate of β, only under the assumption that R it, our main variable of interest, is uncorrelated with both µ i and ɛ it after conditioning for covariate vector X it. This assumption is not likely to hold, for a number of reasons. For example, people with higher unobserved ability, which is relatively stable over a short time span, may be more likely to stay in the labor force, and in the meanwhile, continue to take better care of their health. Ignoring this individual heterogeneity (u i ), as OLS estimation does, can lead to inconsistent estimates of the potential health effects of retirement, principally due to the correlation between the unobserved individual heterogeneity u i and retirement status R it. Fixed effects panel estimation can consistently estimate β, provided we assume that there is no correlation between ɛ it and the retirement decision R it. However, this assumption can be violated too, for several reasons. For example, a negative family shock can induce both a retirement decision and a decline in health status. Or it could be that a health shock can induce retirement in a direct way. To address the endogeneity problem resulting from the correlation between ɛ it and R it, we employ the method of Fixed Effects Instrumental Variable estimation (FE IV). This method requires the availability of an instrumental variable that satisfies two conditions. 9
First it has to be sufficiently correlated with the retirement variable, what is often referred to as the relevance condition, and second, it has to be orthogonal to ɛ it, what is often called the orthogonality or exclusion condition. The two conditions imply that retirement is the only channel through which individual health outcomes can be affected by the instrumental variable. The FE IV estimation consists of two stages. In the first stage, we estimate the following retirement equation using Fixed Effects panel regression R it = E it θ + X itλ + u i + ε it (2) where E it is the instrument for the retirement variable R it. In the second stage of the estimation, we estimate the following health equation H it = R it β + X itγ + u i + ɛ it (3) The specification of equation (3) is the same as in equation (1) except that we have replaced R it with the predicted retirement variable R it, derived from the first stage estimation. Then β can be consistently obtained with the fixed effects panel estimation of equation (3). It should be noted that while the instrument in equation (3) needs to be uncorrelated with ɛ it, it is not necessary that R it be also uncorrelated with u i, as such correlation can be adequately handled by the Fixed Effects panel estimation in each stage of the FE IV estimation. The instrument E it we use is defined as E it =I(Age it Age p t ), where I is an indicator function which takes the value one when the condition in the bracket is true, Age it is the age of individual i at time t, and Age p t is the pension eligibility age at time t for a person of the relevant sex. The instrument is a binary variable indicating when the age of an individual reaches the corresponding Age Pension eligibility age. Many previous studies have shown that social security benefits are an important determinant of the time of retirement (Hurd, 1990; Anderson et al., 1999; Rohwedder and Willis, 2010; Coe et al., 2012). Controlling for individual age, the binary instrumental variable indicating whether 10
people reach certain specific ages is unlikely to have an effect on people s health except through the channel of people s retirement decisions. While this exclusion condition is not testable, the literature does not report any mechanism other than the retirement decision, through which the pension age eligibility could affects individual health. Many papers argue in a convincing manner that the variations in retirement that are induced by social security incentives are exogenous (Bound and Waidmann, 2007; Rohwedder and Willis, 2010; Mazzonna and Peracchi, 2012) and we follow that logic. 5 Results 5.1 The Causal Effects of Retirement Status on Health Table 3 reports the results of equation (1) estimated using OLS regression. Retirement status is found to be associated with significantly lower health levels for each of the four health measures. For example, on average, a retired individual is 22.5 percentage points less likely to report a good, very good or excellent health status than a labor force participant. The SF-36 physical health and mental health measures are also lower for retirees. The relationship is stronger for men than for women. The Fixed Effect (FE) panel estimation results in Table 3 show that the associations between retirement status and health measures are negative and significant but much smaller in magnitude than those estimated using OLS regression. There are two possible explanations for this finding. The first is that FE estimation controls for unobserved timeinvariant individuals heterogeneity. For example, people with higher unobserved ability may be more likely to stay in the labor force and also to take better care of their health. It follows that OLS estimation, can lead to an overstatement of the negative relationship between retirement and health. An alternative explanation could be that FE estimates are more susceptible to measurement error problems than OLS estimates, which can result in estimates of smaller magnitude and larger attenuation bias (Griliches and Hausman, 1986). A more fundamental problem remains in both OLS and FE estimates, in that neither of them accounts for and deals with the possibility of reverse causality between health 11
Table 3: OLS and FE Estimation Results OLS Estimation FE Estimation Self reported Physical Mental Overall Self reported Physical Mental Overall good health health health health good health health health health All 0.225*** 15.521*** 10.698*** 13.110*** 0.051*** 3.843*** 2.607*** 3.225*** (0.010) (0.625) (0.521) (0.545) (0.008) (0.398) (0.396) (0.350) Male 0.280*** 18.484*** 12.824*** 15.654*** 0.046*** 3.982*** 2.510*** 3.246*** (0.016) (0.932) (0.811) (0.834) (0.011) (0.570) (0.541) (0.505) F emale 0.184*** 13.562*** 9.148*** 11.355*** 0.057*** 3.727*** 2.695*** 3.211*** (0.014) (0.830) (0.675) (0.711) (0.011) (0.549) (0.523) (0.480) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables include: a dummy variable indicating that self assessed health is good, very good or excellent, physical, mental and overall health measures constructed from SF 36 (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Huber White robust standard errors clustered at the individual level are reported in parentheses. 12
outcomes and retirement decisions or with the problem of unobservables that change over time and which may affect both health and retirement, thus contaminating our estimates. These problems prevent us from identifying the causal effects of retirement on health with either OLS or FE estimation, and are the prime reason why we use FE-IV estimation. Table 4: First Stage Results of FE IV Estimation All Male F emale Instrument Age eligible for 0.101*** 0.105*** 0.084*** the Age Pension (0.007) (0.010) (0.009) F statistic on the 231.39 117.64 82.45 excluded instrument N 36,713 17,527 19,186 Note: * p<0.1; ** p<0.05; *** p<0.01. The dependent variable is a binary variable indicating being retired or not. Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Standard errors are reported in parentheses. Fixed Effect Instrumental Variable (FE-IV) estimation results are obtained in order to establish the potential causal effect of the retirement decision on health. We begin with the investigation of the first-stage regression of equation (2) with retirement as the dependent variable and Age Eligibility (our instrument) in the right hand side with other covariates. The results from the first stage of the two stage FE IV estimation are reported in Table 4. Table 4 helps us assess whether the age eligibility variable which we use as our instrument satisfies the relevance condition for a valid instrument. The coefficient estimates of the age eligibility variable are highly significant in the full sample and in each of the gender sub-samples, indicating that the age eligibility for the Age Pension is a strong predictor of retirement behavior in Australia. All first stage F statistics on the excluded instrument far exceed the Staiger and Stock (1997) rule-of-thumb threshold of 10, indicating that age eligibility as an instrument is not weak and has sufficient explanatory power. Being over the Age Pension eligibility age increases the probability of being retired by 10.5 percentage points for males and 8.4 percentage points for females. 13
Table 5: The Causal Effects of Retirement Status on Health Self reported Physical Mental Overall good health health health health All 0.310*** 14.523*** 14.364*** 14.444*** (0.074) (3.382) (3.258) (2.950) Male 0.293*** 13.482*** 10.863*** 12.173*** (0.100) (4.556) (4.181) (3.890) F emale 0.390*** 18.464*** 22.201*** 20.332*** (0.129) (6.005) (6.101) (5.410) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables include: a dummy variable indicating that self assessed health is good, very good or excellent, physical, mental and overall health measures constructed from SF 36 (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Standard errors are reported in parentheses. Table 5 reports the second stage of the FE IV regression results, which estimates the causal contemporaneous effect of being retired on the health measures we use. Unlike the findings from OLS and FE estimations, we find positive and highly significant causal health effects of retirement on health, suggesting that OLS and FE estimates are inconsistent and misleading. Being retired significantly leads to a 31 percentage point increase in the likelihood of reporting good, very good or excellent health status. The positive effect of retirement on health is also confirmed when using the SF 36 health measures. Both physical and mental health are improved when transiting from labor force participation into retirement. For example, the SF 36 physical health measure of a typical individual in our sample would be 14.5 higher in retirement than in labor force, which corresponds to about 20.8 percentage point increase in physical health when compared with the sample average. We also find that the health effects of retirement status are uniformly larger for women than for men. The message from the FE IV estimations is very clear, in that it suggests the presence of a strong and positive causal effect of retirement on the health of the retirees. We now move to the potential impact of the next dimension of retirement, its duration, on the health of the retirees. 14
5.2 Does Retirement Duration Affect Health? In the previous section, we considered retirement as a discrete change in lifestyle. Results in Table 5 were based on equation (3) which estimates the potential impact on health of a one-off change from employment to retirement. This section views retirement as a cumulative process of exposure to being out of the labor force. We investigate how health can be affected by the length of time spent in retirement. As the retirees reported the age at which they retired from the labour market in the HILDA data, the retirement duration variable is constructed as RET duration it = Age it RET age it (4) which measures the elapsed time between the self-reported retirement age (RET age it ) and the age at the time of survey (Age it ). For those who are not retired, the value of retirement duration is set to zero. To address the endogeneity issue of the retirement duration variable, we generate a measure of the duration of being age eligible for the Age Pension at the time of survey ELIduration it = Age it ELIage it (5) where the duration is calculated as the difference between the age at which the individual first became age qualified for Age Pension (ELIage it ) and the current age of each individual at interview time (Age it ). If an individual is too young to be age eligible for the pension, the eligibility duration is equal to zero. We use the log form of the duration variables in the following equations H it = Log(RET duration + 1) it β + X itγ + µ i + ɛ it (6) Log(RET duration + 1) it = Log(ELIduration + 1) it θ + X itλ + u i + ε it (7) and estimate equations (6) and (7) using FE IV regression. 2 The first-stage estimation 2 We also estimated equations (6) and (7) using a linear form of duration variables (RET duration it, 15
Table 6: The Effects of Retirement Duration on Health (OLS and FE Estimates) OLS Estimation FE Estimation Self reported Physical Mental Overall Self reported Physical Mental Overall good health health health health good health health health health All 0.088*** 5.695*** 3.827*** 4.761*** 0.006 0.325 0.135 0.230 (0.005) (0.319) (0.259) (0.277) (0.004) (0.208) (0.201) (0.183) Male 0.131*** 8.201*** 5.694*** 6.948*** 0.002 0.103 0.141 0.019 (0.008) (0.491) (0.425) (0.438) (0.007) (0.321) (0.318) (0.284) F emale 0.069*** 4.645*** 3.024*** 3.835*** 0.008 0.466* 0.295 0.381 (0.007) (0.408) (0.325) (0.350) (0.006) (0.272) (0.260) (0.238) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables include: a dummy variable indicating that self assessed health is good, very good or excellent, physical, mental and overall health measures constructed from SF 36 (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Huber White robust standard errors clustered at the individual level are reported in parentheses. 16
results are displayed in Table 7, and the second-stage results are in Table 8. The results from the OLS and FE estimation of equation (7) are displayed in Table 7 for comparison purposes. OLS estimation results in Table 7 show that retirement duration has a negative and statistically significant association with health measures. A doubling of the time in retirement is associated with 8.8 percent decrease in the probability of reporting good health. The magnitudes of the coefficients on the retirement duration variable are considerably smaller when using the FE estimation method, and almost all estimates are not significant. 3 Table 7: First Stage Results of Equation (6) All Male F emale Instrument Log(ELIduration+1) 0.196*** 0.162*** 0.199*** (0.011) (0.017) (0.016) F statistic on the excluded instrument 299.70 96.54 154.66 N 35477 17,040 18,437 Note: * p<0.1; ** p<0.05; *** p<0.01. The dependent variable is the log of years spent in retirement. Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Standard errors are reported in parentheses. The first stage results of the FE IV estimation reported in Table 7 show that the duration of pension age eligibility is a strong predictor for the retirement duration. A doubling of time being age eligible for the Age pension is found to lead to a 16 percentage points increase in the time spent in retirement by males, and the effect is slightly larger for females (20 percentage points). The causal results displayed in Table 8 show that the time spent in retirement has a positive and significant effect on each of the four health measures for the pooled sample. ELIduration it ), in addition to the log form (Log(RET duration+1) it, Log(ELIduration+1) it ). Estimations showed that the instrument variable is much weaker when using the linear form, and the related first stage F statistics on the excluded instrument are respectively 28.31, 25.26 and 5.36 for the full sample, males and females respectively, which indicates that we have a weak instrument problem for the sub-sample for women. The first-stage estimation results using the log form of duration variables reported in Table 7 show that the retirement duration variable is more likely to have a non-linear relationship with the duration of being Age eligible for the pension. 3 The only exception is that the association between the physical health of female and retirement duration is negative and significant at the 10% level when using fixed effects estimation. 17
Table 8: The Casual Effects of Retirement Duration on Health Self reported Physical Mental Overall good health health health health All 0.136*** 5.167*** 6.843*** 6.005*** (0.037) (1.663) (1.628) (1.444) M ale 0.124* 4.752 5.694* 5.223* (0.070) (3.134) (2.940) (2.678) F emale 0.167*** 6.346*** 9.197*** 7.771*** (0.049) (2.225) (2.275) (1.976) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables include: a dummy variable indicating that self assessed health is good, very good or excellent, physical, mental and overall health measures constructed from SF 36 (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Standard errors are reported in parentheses. A doubling of the time in retirement will lead to 13.6 percent increase in the probability of reporting good, very good or excellent health status. While we find highly statistically significant health effects of retirement status for men, we do not find evidence that the physical health of men is affected by the length of exposure to retirement. The effects of retirement duration on self-reported health, mental health and overall health are positive but only significant at the 10% level for men. For women, the effects of retirement duration on health outcomes are positive and highly significant. Similar to the gender differences in the health effects of retirement status reported in Table 5, we find that the effects of retirement duration are also larger for women than for men. 4 5.3 The Effects of Retirement Status and Retirement Duration on Health Constituents Table 5 and 8 report significant causal effects of retirement status and retirement duration on health, however, it is not necessarily the case that the effects are uniform for the 4 Unreported estimation results show that for the full sample both retirement status and retirement duration lead to more physical activities and less smoking, indicating that health-related behaviors can be the channels through which retirement affects people s health in Australia. This finding is also primarily driven by the large and significant retirement effects on the health investment behaviors of women. Neither participation in nor the intensity of exercising or smoking of men are affected by their retirement status or retirement duration. The gender difference in the effects of retirement on health-related behaviors may explain why the health effects of retirement are larger for women than for men. 18
different components of health, which summarise different aspects of health. For example, among the four constituents of the SF 36 physical health component, Role Physical (RP) measures the limitations in work or activities caused by physical health problems, while quite differently, Bodily Pain (BP) measures directly physical pain. We estimated separately the causal effect of retirement status and retirement duration on each constituent of physical and mental health components, and present the results by gender in Table 9. We find that being retired significantly benefits physical health in three of its four components (Physical Functioning (PF), Role Physical (RP) and Bodily Pain (BP)). The effects of retirement status on the three constituents are different for each gender. For example, retirement status has the largest benefit for the Physical Functioning (PF) of males, which measures the limitations to daily activities. For females, there is a very large effect of retirement status on Role Physical (RP). Indeed, it is this that drives the result that the estimated effect of retirement status on physical health as a whole is larger for women than for men. In terms of the mental health effect of retirement status, the four constituent estimates are all larger for females than for males, and for each gender group, the effects of being retired differ between mental health constituents. The Social Functioning (SF) benefit of being retired is larger than the gains in the other three mental health components for men, while for females, the benefit of being retired is the largest on Role Emotional (RE), which measures the limitations in work or activities caused by emotional health. The effects of retirement duration on health constituents are also found to be not uniform. While Table 8 documents a significant effect of retirement duration on the physical health measure for the full sample, Table 9 shows that this result is mainly driven by the significant retirement duration effect on Physical Functioning (PF) and Role Physical (RP). For males, the effects of retirement duration on different constituents of physical and mental health are mostly insignificant, and we only find a significant effect on Physical Functioning (PF) that measures daily activity limitations and Mental Health (MH) that measures feelings of anxiety and depression. For females, most coefficient estimates of retirement duration are positive and statistically significant, showing the influences of 19
Table 9: The Causal Effects of Retirement Status and Retirement Duration on Health Constituents Physical component Mental component Physical Role Bodily General Social Role Mental Vitality Functioning Physical Pain Health Functioning Emotional health The Effects of Retirement Status All 17.040*** 25.981*** 12.064*** 3.006 13.882*** 18.070*** 13.523*** 11.982*** (3.532) (7.462) (4.080) (2.786) (4.307) (6.614) (3.103) (2.772) M ale 21.165*** 16.890* 13.511** 2.362 12.309** 11.260 7.993** 11.890*** (5.084) (9.857) (5.658) (3.815) (5.689) (8.642) (3.986) (3.682) F emale 15.102** 41.416*** 11.560* 5.866 19.187** 31.504*** 22.624*** 15.489*** (5.874) (13.519) (6.942) (4.819) (7.733) (12.504) (5.822) (4.996) The Effects of Retirement Duration All 7.152*** 8.784** 2.662 2.071 5.365** 10.494*** 6.915*** 4.602*** (1.747) (3.723) (2.054) (1.435) (2.168) (3.372) (1.576) (1.380) M ale 7.510** 7.236 2.929 1.334 5.074 6.323 4.116 7.262*** (3.416) (7.007) (3.967) (2.755) (4.016) (6.180) (2.850) (2.602) F emale 8.433*** 10.957** 2.456 3.539* 7.023** 15.545*** 9.569*** 4.648** (2.283) (4.968) (2.672) (1.890) (2.939) (4.641) (2.195) (1.867) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables are the four scales of SF 36 physical health component and the four scales of the SF 36 mental health component (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, a full set of state of residence dummies and wave dummies. Huber White robust standard errors clustered at the individual level are reported in parentheses. 20
retirement duration are exerted on most aspects of female physical and mental health. The results of Table 9 also highlight that retirement status and retirement duration affect individual health in a variety of diverse ways. 5.4 Retirement Benefits Mainly those with Lower Health Levels This section examines whether the health effects of retirement differ for people with different health status. To this purpose, for each of the SF 36 health indices we use, we group individual observations into two groups by whether their individual health index in the previous wave was above or below the median value, and we then repeat the (FE IV) estimation separately for the two health sub-groups. Results are displayed by gender in Table 10. Table 10: Heterogeneity in the Health Effects of Retirement Physical health Mental health Overall health Below Above Below Above Below Above median median median median median median The Effects of Retirement Status All 44.136*** 9.747*** 33.972*** 2.158 35.191*** 5.774* (14.880) (3.743) (9.012) (3.375) (10.499) (3.129) M ale 47.264** 8.179 30.175** 0.810 29.931** 2.885 (20.276) (5.082) (13.086) (3.603) (12.842) (4.136) F emale 43.634* 13.934** 38.923*** 7.028 42.474** 11.823* (26.256) (6.315) (12.964) (8.302) (18.354) (5.625) The Effects of Retirement Duration All 8.513*** 6.028** 11.305*** 3.753** 8.478*** 4.793** (3.105) (2.531) (3.442) (1.826) (1.992) (1.855) M ale 10.574*** 9.145 13.985** 2.753 7.568 5.578 (5.996) (5.817) (6.938) (3.048) (5.256) (3.845) F emale 8.118** 6.369** 12.532*** 5.515** 10.428** 5.795** (3.939) (3.080) (4.528) (2.637) (4.047) (2.376) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables are the physical, mental and overall health measures constructed from SF 36 (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Standard errors are reported in parentheses. We find that the positive health effects of retirement status and retirement duration displayed in Table 5 and Table 8 are mainly driven by the large and significant positive effects among people with below median health, and this finding holds for all SF 36 health 21
measures, and also for each gender. For people with above average health in the relevant domain, the retirement effects are significant for only women with small magnitudes. 5 To check whether the finding of heterogeneous effects of retirement on health is robust or not, we repeat what we have done in Table 10, but with the dependent variable as each of the four SF 36 physical health scales (Physical Functioning (PF), Role Physical (RP), Bodily Pain (BP), General Health (GH)) and each of the four mental health scales (Social Functioning (SF), Role Emotional (RE), Vitality (VT) and Mental Health (MH)), which measure different aspects of physical and mental health. The results obtained tell a consistent story with Table 10: when there is a positive and significant effect of retirement on health outcomes, it is mainly people in poor health who benefit most from the transition into retirement and longer time spent in retirement. 6 5.5 A Robustness Check: Including Household Income and the Value of Housing Assets as Additional Controls We conduct a sensitivity analysis by including the household income and the value of household housing assets into our modeling. Two reasons justify our approach here. First, a change in income or wealth may affect health services utilizations and other health related behaviors that may have an impact on the health outcomes of Australian people. Second and more importantly, our instrument, which describes whether the age of an individual meets the age requirement of the Australian Age Pension, now has a different interpretation. As the observations in our sample all satisfy the pension residency condition, when we include the income levels and the values of housing assets in the first stage estimation, whether an individual is age eligible for the Age Pension is equivalent to whether the individual is eligible for the Age Pension. The results are reported in Table 11. The health effects of retirement status and retirement duration reported herein, which are further conditional on household income 5 It should be noted that the insignificant results displayed in Table 10 are not driven by the weak instrument problem as the related first stage F statistics on the excluded instrument are all greater than the rule-of-thumb cutoff of 10 for weak identification not to be considered a problem (Staiger and Stock, 1997). 6 Detailed results are available upon request. 22
Table 11: Robustness check using Household Income and Housing Wealth as Additional Controls Physical health Mental health Overall health Self reported Below Above Below Above Below Above good health All All All median median median median median median The Effect of Retirement Status All 0.298*** 14.014*** 41.781*** 9.362*** 13.945*** 32.976*** 2.103 13.980*** 33.832*** 5.527* (0.071) (3.270) (13.771) (3.597) (3.152) (8.641) (3.224) (2.851) (9.903) (3.008) M ale 0.289*** 13.252*** 45.089** 8.153 10.665*** 29.170** 0.878 12.172*** 28.928** 2.884 (0.099) (4.490) (19.098) (5.075) (4.123) (12.497) (3.581) (3.890) (12.292) (4.131) F emale 0.356*** 17.211*** 39.152* 12.816** 20.970*** 37.486*** 5.985 19.091*** 40.154** 10.535** (0.120) (5.615) (22.990) (5.762) (5.701) (12.302) (7.282) (5.043) (16.823) (5.108) The Effect of Retirement Duration All 0.130*** 4.994*** 8.355*** 5.983** 6.709*** 11.165*** 3.669** 5.851*** 8.353*** 4.618** (0.036) (1.630) (3.067) (2.443) (1.595) (3.410) (1.766) (1.415) (3.649) (1.792) M ale 0.121* 4.618 10.286* 8.979 5.577* 13.985** 2.865 5.097* 7.511 5.581 (0.069) (3.085) (5.904) (5.764) (2.894) (6.837) (2.995) (2.663) (5.174) (3.796) F emale 0.158*** 6.076*** 7.657** 6.106** 8.981*** 12.265*** 5.290** 7.528*** 10.099** 5.420** (0.048) (2.168) (3.860) (2.930) (2.217) (4.449) (2.530) (1.924) (3.972) (2.265) N ote: * p<0.1; ** p<0.05; *** p<0.01. The dependent variables include: a dummy variable indicating that self assessed health is good, very good or excellent, physical, mental and overall health measures constructed from SF 36 (standardized to be between 0 and 100). Control variables include age, age squared, years of schooling, a married dummy, family size, and a full set of state of residence dummies and wave dummies. Standard errors are reported in parentheses. 23
and housing wealth, are very similar to those reported in Table 6, 9 and 11. We find very robust evidence that retirement as a discrete event and as a cumulative process can both lead to significant improvement in self reported health and physical and mental health outcomes. We also find that it is overwhelmingly the individuals with poorer health in the previous wave (namely, whose corresponding physical, mental and overall health measures were below the relevant medians in the last year) that experience the most health gains from retirement incidence and duration. 6 Conclusion The paper has been motivated by the general trend for prolonging working life and postponing retirement in many developed economies. While there is a wide literature to show how worse health can increase the probability of retirement, there is little evidence on the reverse possibility, namely that retirement may influence health in a causal way. Most of the studies that show an association between higher probability to retire and worse health do not provide evidence on whether this relationship is causal and how. Being motivated by the prolongation of working life and the associated increased chances of health deteriorating with age, we focus on the investigation of whether retirement influences health in a causal fashion. We embark on this exercise appreciating that there are valid priori reasons why retirement could improve health and other similarly valid reasons why retirement could harm health. This paper contributes to the literature by estimating the causal effect of retirement on health outcomes in several important ways. First, we show that with the use of appropriate econometric methodology, causal effects can indeed be clearly identified. To achieve this we applied Fixed Effects Instrumental Variable estimation on individual longitudinal data from the Household, Income and labor Dynamics in Australia (HILDA) survey. Using information on the age eligibility of individuals for the publicly funded Australian Age Pension as our instrument, we identified the causal effect of becoming retired and or remaining retired on health outcomes. Second, we show that both retirement status and retirement duration have positive and 24