Retiring for Better Health? Evidence from Health Investment Behaviors in Japan
Meng Zhao* Assistant Professor, Department of Economics, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo 113-0033, Japan,
Yoshifumi Konishi Associate Professor, Faculty of Liberal Arts, Sophia University, 7-1 Kioi-cho, Chiyoda-ku, Tokyo 102-8554, Japan
Haruko Noguchi Professor, Faculty of Political Science and Economics, Waseda University, 1-6-1 Nishiwaseda, Shinjuku-ku, Tokyo 169-8050, Japan,
*Corresponding
author. Tel.: +81 3 5841 5530; fax: +81 3 5841 5521. E-mail address:
[email protected];
[email protected] (M. Zhao). We gratefully acknowledge the helpful comments from the seminar participants at the National Institute of Population and Social Security (IPSS), the National Graduate Institute for Policy Studies (GRIPS), the Japanese Economic Association Meeting, the Hong Kong Economic Association Meeting, and the Western Economic Association International Meeting. The study was supported by the Grants-in-Aid for Young Scholars from the Japan Society for the Promotion of Science (JSPS).
Abstract The economic literature argues that increasing pension eligibility ages may not restore the financial viability of a social security system if it negatively affects the health outcomes of the affected population. This paper analyzes recent longitudinal data from four waves of Health and Retirement Survey in Japan and examines the causal effects of retirement on three types of health investment behaviors: smoking, drinking and exercising. We employ two econometric methods to distinguish the impacts of two types of retirement: (i) a fuzzy regression-discontinuity (RD) design for the effects of retirement from permanent employment (RPE); and (ii) an instrumental variable (IV) approach that exploits exogenous variation in Japan’s national pensionable ages for the effects of complete retirement (CR). Our results indicate that individuals increase participation into regular exercising and drinking, yet reduce smoking intensity significantly after retirement. The effects are stronger for CR than for RPE. We also find the differential effects of retirement by gender.
JEL classification: C14; C26; I12; J26 Keywords: retirement, health behaviors, Japan, pensionable age
1
1. Introduction One inevitable consequence of aging population is an increase in the dependency rate. The ratio of people aged 65 or above in all OECD countries is expected to rise substantially from 22% in 2000 to 47% in 2050 (OECD, 2007). With increasing financial burdens on government budgets, a number of developed countries are starting or have started raising the eligibility ages for publicly funded old-age pensions. Japan is one of them: It has repeatedly changed pensionable ages from 55 to 60 for women in 1985, and further to 65 for both men and women in 1994 and 2000. Economists have long debated, however, whether delaying pension eligibility ages would actually reduce government expenditures on social security programs. One counteracting effect is the potentially negative impact of late retirement on population health. One’s health may deteriorate both physically and mentally due to long working hours. If so, increasing pensionable ages may increase health care costs, and the extra financial burden may more than offset the savings in pension expenditures. If instead individuals stay healthier when working than after retirement, delayed retirement may have additional benefits besides saving pension expenditures. Over the last few decades, a number of studies have investigated the relationship between health and retirement. However, their findings vary wildly. Some studies find that retirement has a positive effect on some measures of physical or mental health (e.g. Bound and Waidmann, 2007; Charles, 2004; Coe and Zamarro, 2011; Neuman, 2008), yet others find null or negative effects (e.g. Kerkhofs and Lindeboom, 1997; Dave, et al., 2008; Lindeboom, et al., 2002; Johnston and Lee, 2009). (See Section 2 for more detailed literature review). These earlier studies have revealed two primary difficulties associated with the effort to identify the causal relationship between retirement and health. First, retirement and health are obviously endogenous. Not only there are unobserved third factors at individual levels that can affect both outcomes but also the causality between them can run in both directions. Second, even if one was able to identify the causal link between retirement and health outcomes, the economic mechanisms that generate the link would still be unidentified. Indeed, the existence of several competing pathways that can go in the opposite directions may be the very reason for the earlier studies’ mixed findings. For this reason, this study investigates the causal effects of retirement on health investment behaviors instead of health outcomes, accounting for the endogeneity of retirement decisions. In the health capital model of Grossman (1972), health behaviors are considered as inputs for production of health. By examining how retirement could 2
affect the inputs rather than the outcomes of health production, this study attempts to provide important insights about the channels through which retirement affects health outcomes. Specifically, we investigate the effects of retirement on three types of health investment behaviors, namely, smoking, drinking, and exercising, as they are the major risk factors of chronic health conditions as identified by the World Health Organization, accounting for 80% of total chronic diseases. To correct for the bias caused by the endogeneity of retirement, we employ two econometric approaches. Each of the approaches exploits the quasi-experimental setups created by the unique institutional background in the Japanese labor market. Our first approach makes use of the fact that as of 2008, more than 94% of the Japanese companies who employ 30 or more employees adopt a mandatory retirement system, which forces their employees to retire from their permanent employment (MHLW, 2009). Though a majority of the forced retirees still find alternative employment elsewhere, they typically work in the less intense work environment that require fewer hours of work per week. We use a fuzzy regression-discontinuity (RD) approach to identify the effect of the retirement from permanent employment (RPE) by exploiting the fact that the probability of RPE in Japan significantly increases at the age 60 which used to be the Basic Pension eligibility age for men by 2000 and for women by 2005. Our second approach is the instrumental variable (IV) method, exploiting the exogenous variation in pensionable ages over time and across gender (due to the aforementioned pension reforms in Japan) to instrument endogenous retirement choices. These two approaches also allow us to distinguish the effects of RPE, in which individuals may still continue to stay in the labor force as fixed-term, part-time, or self-employed employees after the mandatory retirement, and the effects of complete retirement (CR), in which case individuals exit the labor force and work zero hours.1 We analyze recent longitudinal data from the Health and Retirement Survey (HRS) collected during 2008-2011 in Japan and our results indicate several important findings. First, upon CR, individuals significantly reduce the intensity of smoking by 2.3 cigarettes a day, and the effect is greater for men, 4.8 cigarettes per day. The effect of RPE for male is even greater, a reduction of approximately 7.5-9.3 cigarettes per day, but less robust for the whole sample. Second, individuals are more likely to drink after CR: The probability of drinking and over-drinking increases by more than 20%. However, the effect is weaker for RPE. Lastly, the probability of engaging in regular 1
Permanent employees are defined as those who work for wage without a pre-determined time limit. They are usually assured certain rights, e.g. subsidized health care and paid holidays, but also subject to mandatory retirement. 3
exercising is approximately 55 percentage point higher for individuals who have retired completely. Again, the effect is less robust for RPE and is statistically significant for only women. These results are consistent with our hypothesis that retirement generally has health-inducing as well as leisure-inducing effects, which may go in the opposite direction for some health investment behaviors. These heterogeneous effects of retirement on different health behaviors across different subpopulations can help explain the mixed findings on the effects of retirement on health outcomes. The rest of the paper is organized as follows. Section 2 discusses the related literature. Section 3 presents an organizing framework for empirical analysis. Section 4 discusses the identification and estimation strategies. Section 5 provides institutional background and describes the data. The main results are presented in Section 6. The last section concludes. 2. Literature Review Economists have long been interested in the inter-relationship between health and labor supply. A large body of literature exists that established the causal impact of health on labor supply and retirement. More recently, however, studies have investigated the causal relationship running in the opposite direction: the effect of retirement on health. Existing studies differ in terms of methodologies they use to control for the endogeniety of retirement, and their findings differ accordingly.2 The first strand of literature uses an instrumental variable (IV) approach and relies on exogenous variations in retirement regulations or retirement-driving policies as instruments (e.g. Charles, 2004; Neuman, 2008; Kuhn, et al., 2010; Coe and Zamarro, 2011). Charles (2004) and Neuman (2008) used age-specific retirement incentives provided by the U.S. Social Security regulations as instrumental variables, and independently found a positive effect of retirement on self-reported measures of health. A more recent study by Coe and Zamarro (2011) used early and full-time retirement ages in eleven European countries as instrumental variables to assess the effects of retirement on self-reported measures of health, mental health, and cognitive ability. They found a long-lasting health-preserving effect of retirement but no significant effect on mental health or cognitive ability. The second line of literature exploits cross-sectional variation in the panel data. 2
There are other non-economic studies in epidemiology or public health that examined the effects of retirement on health behaviors (Perreira and Sloan, 2001; Lang et al., 2007; Henkens et al., 2008). These studies, however, often fail to correct for the endogeneity of retirement or time-varying unobservables. 4
Using the panel data from Netherland collected in 1993 and 1995, Kerkhofs and Lindeboom (1997) and Lindeboom et al. (2002) found evidence that health tends to deteriorate with increased working efforts, but found no significant effect of retirement on mental health. Dave et al. (2008) explored the longitudinal data from the Health and Retirement Study in United States and showed that complete retirement can have negative effects on mobility, illness conditions, and mental health. However, these estimates may be biased if some unobservable time-varying factors can influence both retirement decision and health (e.g. increasing health concern). More recently, several studies have used a regression discontinuity (RD) approach (e.g. Bound and Waidmann, 2007; Johnston and Lee, 2009). For example, Johnston and Lee (2009) employed the RD approach to analyze the short-term effects of retirement on health, exploiting the fact that people are significantly more likely to retire once they reach the pensionable age. Based on the cross-sectional data from the Health Survey for England, they found that retirement increases the sense of well-being and mental health. However, they did not find any significant effect on physical health. There are few relevant empirical studies in Japan (mainly published in Japanese). A majority of the studies focus only on the correlation between health and retirement (e.g. Sugisawa et al., 1997; Osada and Ando, 1998). Kajitani (2011) used an IV approach to take care of the endogeneity of retirement, and showed that working in old age has a positive effect on health for males aged over 60, suggesting a negative health effect of retirement. However, the study used self-employment status and marital status as IVs, both of which are individual-level variables that are likely to be correlated with unobservable factors of health. 3. Empirical Framework 3.1 Empirical Regularities on Retirement in the Japanese Health and Retirement Survey We first start with some empirical regularities found in the Health and Retirement Survey (HRS) in Japan that are useful in guiding our empirical strategies. The data are for 1,334 individuals aged 45-80 collected through four waves from 2008 to 2010 (See Section 5 for more details). Figure 1 show the percentages of individuals who fall into four types of employment status, plotted against age: (a) permanent employee, (b) fixed-term or part-time employee, (c) self-employed, and (d) non-working. Figure 2 shows the average working hours per week, by age and by gender, for permanent 5
employees. Figure 3 shows the national averages of wage rates, by age and by gender, reported by the Japanese government. These figures suggest important empirical regularities for our empirical analysis. First, there is a discontinuous drop in the percentage of individuals working as permanent employees at age 60 (which used to be the Basic Pension eligibility age for men by 2000 and for women by 2005 and is the pensionable age for the Employee’s Pension or Mutual Aid Pension until 2013) whereas a discontinuous rise in the percentage of individuals working for fixed-term, part-time, or self-employment. Second, there is no apparent discontinuity in the percentage of non-working individuals at age 60, though it generally increases after around age 55. Third, there is a substantial drop in the average working hours and the average wage rate at age 60 among permanent employees, and the decrease in working time is larger for female employees. These empirical regularities suggest that a large proportion of permanent employees retire from their permanent employment at age 60 when they reach their pensionable age, yet many of them, particularly male employees, still find other employment, either as fixed-term, part-time, or self-employed employees, at reduced wage rates for reduced working hours. These empirical regularities are consistent with the institutional background for the Japanese labor market: In Japan, most employers in formal sectors set the mandatory retirement age at 60 in accordance with the pensionable ages. Thus in the Japanese labor market, individuals face an exogenous change in the employment opportunities (equivalent to an exogenous change in the expected wage rate) at age 60, which then induces them to choose labor participation and labor hours endogenously. See also the detailed description of the institutional background in Section 5. 3.2 Linking Empirical Regularities to the Empirical Framework We now present an organizing framework for our empirical analysis. One important puzzle here is why rational individuals may make such a discontinuous change in health investments upon retirement. Pensionable ages and mandatory retirement ages at companies are known to employees well ahead of their actual retirement. Thus individuals must be able to foresee and plan accordingly their retirement ages. To the extent this is true, shouldn't rational individuals also optimally smooth health investments over time as well? Indeed, retirement and health investment decisions can be thought of as the flip sides of the same coin. That is, the same economic mechanism that would induce a 6
change in labor participation may also induce a change in health investment behaviors. The key here is that there is one important feature of health investments that is substantially different from other types of investments. That is, time required for health investments cannot be lent and borrowed over time, at least not in the same way as money can. Whenever individuals engage in health-increasing activities (e.g. exercising, walking, healthy cooking etc), they need to use their own time. Of course, one can argue that rich enough individuals could afford to buy labor to free up their own time for these activities, but that option does not seem available for a vast majority of individuals. The mandatory retirement system in the Japanese labor market adds a set of important features into this context. First, mandatory retirement forces employees to retire from their permanent employment and, for those who still wish to work, to search for alternative jobs that pay substantially lower salaries. In this sense, mandatory retirement can be thought of as an exogenous change in the employees' (expected) wage rates. This change in the wage rates implies that retirees face lower opportunity costs of time for their health investments. Second, because mandatory retirement often occurs at the government-mandated pensionable ages, retirees start receiving pension payments at the same time. This means that retirees have higher non-labor income flows than non-retirees. The impacts of these two features of mandatory retirement can be analyzed in the spirit of Grossman's health investment model (1972). A consumer's preferences in each period are defined over a numeraire consumption good c, the stock of health capital h, and a composite of health investments y. We allow for the possibility that consumer's preferences, particularly on health, may change over time. One-period utility is thus given by ( , ℎ , ; ). Note here that we allow y to influence utility directly. The consumer is endowed with a fixed amount of time Ω and allocates her time between labor l and a composite of health investments y. The consumer incurs two kinds of health investment costs: monetary and time costs. Let be the monetary cost and be the time required per unit of the health investment activity. Following Grossman, assume that the number of sick days s is inversely related to health capital h. Hence, the consumer's time constraint in each period is given by: +
(ℎ ) +
∙
= Ω.
(1)
The stock of health capital follows the standard transition equation: ℎ
= ( ; ) + (1 − )ℎ ,
(2) 7
where f is the time-varying health production function and δ is the depreciation rate of health capital. For simplicity, we assume that there is no intertemporal borrowing and saving. Thus the consumer faces the temporal budget constraint: +
∙
≤
+
,
(3)
where nl is her non-labor income (including her pension) and w is the wage rate, both of which are assumed exogenous as in Grossman (1972). In the Appendix, we solve the consumer’s dynamic optimization problem, subject to (1)-(3). The solution is in general characterized by the following first-order condition: (
+ ) ( )=
( )+
( + 1) ( ),
(4)
where arguments in the derivatives of per-period utility u, health production f, and value function v are suppressed. Equation (4) is the standard condition that an optimal health investment must equate the marginal cost of investment, in terms of the monetary and time costs, with the marginal benefit, in terms of the marginal utility of health investment and the discounted value from the increased health capital in the future. As a result, the optimal level of current-period health investment is a function of the health stock h, the wage rate w, and the non-labor income nl at time t: ∗
= "#ℎ ,
,
; ( ), ( )$.
(5)
Note that given y*, the optimal level of labor supply l* is implicitly defined by time constraint (1). Representing equation (4) in the (c, y) space, we see that there are two alternative paths the consumer can take in adjusting her labor-health investment behavior in response to the mandatory retirement from permanent employment. In Figure 4, path (a) depicts the case where the consumer immediately ceases labor force participation in response to the mandatory retirement (which is represented by the increase in non-labor income from and the decrease in the wage rate, or the change from &'( to &' ). Path (b), on the other hand, depicts the case where the consumer does not retire immediately and instead gradually adjust labor hours. The continuously increasing trend in the share of non-working status in Figure 1-(d) indicates that the latter is more prevalent. In the previous empirical literature, retirement is often modeled and defined as 8
non-participation into the labor force (e.g. French, 2005). That is, an individual is deemed to have retired when labor hours l = 0. We call “complete retirement” (CR). In such a case, we are interested in the effect of retirement on health investment behaviors in the following form: )*+ = +, - ./-01 = 12 − +, - ./-01 = 02 = +4 - | - = 06 − +4 - | - > 06
(6)
where yi is the variable representing i-th individual’s health investment behavior and
/-01 is the indicator variable representing i-th individual’s complete retirement status. Our empirical framework makes it clear why the exogenous variation in pensionable ages, which in turn induces changes in non-labor incomes and (expected) wage rates,
would be a good instrument for the endogenous variable /-01 . See Section 4 for more detailed discussion. The empirical regularities also suggest that some individuals remain working as fixed-term or part-time employees after retiring from their permanent jobs. In such case, we identify the effect of retirement as: )*+ = +4 - ./-189 = 16 − +4 - ./-189 = 06 = +4 - . - (/-189 = 1)6 − +4 - | - (/-189 = 0)6.
(7)
In the empirical analysis below, we consider three types of health behaviors: smoking, drinking, and regular exercise. Though not explicitly considered in the model, implications of the model for different health behaviors can be gleaned from Figure 4 or Eq. (4). First, a reduction in labor hours due to retirement would free up time for health investment behaviors. This implies that the effect of retirement is likely to be more prominent in health investment behaviors that are time consuming. We thus expect retirement to have a larger impact on regular exercise than the other two. Second, less time for labor implies not only more time for health investment but also for other types of leisure. Drinking is often considered a leisure activity in the Japanese culture while reducing it would be a health investment. Thus there are competing effects of retirement on drinking, and individuals may even increase drinking if the leisure-inducing effect is stronger than the health-inducing effect. The same argument may apply to smoking, though possibly to a lesser extent. Lastly, given the overall increase in y, different individuals would engage in different types of health investments differently depending on the perceived marginal utility uy and the marginal health effect fy. A few caveats are in order. First, Figure 4 also makes it clear that we are much less likely to observe a change in labor participation and health investment behaviors if the 9
pensionable age only means the increase in the non-labor income from pension payments and does not represent the change in employment opportunities (and the associated change in the expected wage rate). Second, though we focused on retirement induced by the mandatory retirement system, individuals may choose to retire for other reasons such as childbearing, marriage, early retirement bonuses, or health problems. Hence, the average treatment effect we attempt to identify is not limited to the effect of mandatory retirement. Lastly, there is a human-capital-investment aspect of health investment. If the wage rate w is a function of health and the marginal effect of health on w becomes smaller after mandatory retirement, it is possible that individuals will decrease health investments. Hence, the sign and magnitude of the ATE are in general ambiguous. 4. Identification Strategies We estimate the ATE (6) or (7) in a random sample of permanent employees with the data on an outcome variable :; , which describes health investment behaviors and a retirement indicator /; , which equals one if i-th individual (i = 1,…, N) retires and zero otherwise. The standard parametric econometric specification would be: :; = < + /; + γ>; +
; (8)
where β measures the average retirement effect, >; is a set of other factors that affect the outcome, and ; represents the unexplained variation in :; . If the assignment of retirement is purely random, then β could be consistently estimated by Ordinary Least Squares (OLS). However, retirement is often an endogenous choice for individuals made for a variety of reasons. Therefore, retirement status /; may be endogenous in a non-experimental setup, and the OLS estimate will be most likely biased. We thus use two econometric approaches to estimate equation (8): the RD approach and the IV approach. Both have advantages and disadvantages, which we discuss below. 4.1. The Regression Discontinuity Approach Our first empirical strategy exploits the fact that individuals tend to retire from their permanent employment at the pensionable age exogenously set by the Japanese government. As shown in Figure 1-(a), the probability of retirement from permanent employment significantly increases at the pensionable age. This allows us to estimate 10
the causal impact of retirement in a regression discontinuity (RD) framework. Consider the individuals who are within a small interval in the neighborhood of the cutoff age (i.e. the pensionable age of 60). If other factors in >; do not change systematically before and after the cutoff age, the average characteristics of the two samples slightly below and slightly above the pensionable age are likely to be the same, as age cannot be controlled or influenced. Therefore, the average outcomes for the two samples should be the same in the absence of the government-mandated pensionable age. Thus, in the small neighborhood of the cutoff age, our RD design mimics a randomized experiment. There are two types of RD design, namely “sharp” vs. “fuzzy” RD designs. A “sharp” RD design would be applicable if retirement is a deterministic function of age relative to the pensionable age: i.e. everybody retires only after she reaches the pensionable age. However, in many cases, individuals may retire before the pensionable age or may continue working after it. Thus the “sharp” RD is not suitable for our setup. On the other hand, the “fuzzy” RD only requires that there is a discontinuity in the probability of retirement at the cutoff age and, therefore, is well suited for our setup. Furthermore, as shown in Figure 1-(d), mandatory retirement from permanent employment does not induce individuals to cease their participation in the labor force. Because many of them continue working as fixed-term, part-time, or self-employed workers, the labor participation rate does not change discontinuously at the cutoff age. As a result, our treatment variable Ri in the RD approach is not whether one is working or not, instead it is whether one retired from permanent employment or not. Thus, intuitively, we are comparing the average outcomes between those who retired from permanent employment versus those who did not, immediately after the pensionable age 60. In other words, the RD estimate gets at the effect of shifting working status from permanent employment, which often requires longer and more intense working hours, to non-permanent employment. In the recent literature, the RD estimation is often done by a non-parametric procedure (Lee and Lemieux, 2010). Following Hahn et al. (2001), we obtain the RD point estimate by the following estimator, using only the data on individuals whose ages are slightly above and below the pensionable age: = where :
Y − YB , R − RB
(9)
and : B are the one-sided limits of the outcome as age approaches the 11
cutoff age from right and from left. Similarly, /
and / B are the right and left limits
of the probability of retirement evaluated at the cutoff age. Following the literature, we rely on a local linear regression (LLR) to estimate the : E s and / E . Specifically, the
LLR estimators for Y , Y B , /
and / B are given by
JK ,
JL ,
MK
and
ML
from
the following optimization: (
JK , NJK )
≡ argminδ,θ ∑;:[\]^ _`4:; −
− N(UVW; − )X 6Y; ,
(10a)
(
JL , NJL )
≡ argminδ,θ ∑;:[\]^ a 4:; −
− N(UVW; − )X 6Y; ,
(10b)
(
MK , NMK )
(
ML , NML )
≡ argminδ,θ ∑;:[\]^ _`4/; −
≡ argminδ,θ ∑;:[\]^ a 4/; −
− N(UVW; − )X 6Y; ,
(10c)
− N(UVW; − )X 6 Y; ,
(10d)
where c denotes the cutoff age and Y; is a kernel function. We use a uniform kernel instead of a triangular kernel as in previous studies, because using a triangular kernel will result in the loss of all the observations at the ending points, which can be a serious problem when we focus on the sample aged two or three years different from the cutoff age.3 The RD approach compares the outcomes of the two subsamples sufficiently close to the pensionable age. There is a question of “how close” is sufficiently close to ensure the two subsamples to be comparable in the absence of retirement, without efficiency loss. Following Imbens and Kalyanaraman (2009), we estimate the optimal bandwidth to balance the bias due to the bandwidth that is too large and the efficiency loss due to the bandwidth that is too small. Because age is a discrete variable, there is also a concern that the estimator of the optimal bandwidth may not be appropriate in our setup. Hence, we also constrained our samples to be sufficiently close to the cutoff age: i.e. individuals aged 57 to 63, as in Johnston and Lee (2009).4 Following the literature, we apply the same bandwidth in the estimation of equations (10) and obtain the robust standard errors from a standard 2SLS where retirement status is regressed against age and an indicator of whether above pensionable age in the first stage. 4.2. Instrumental Variable Approach
3
Choice of kernel functions typically has little impact on estimates (Lee and Lemieux, 2010). Individuals aged 60 and 61 are grouped into one, as the discontinuity in the probability to retire is observed at both ages 60 and 61 in Japan. More details will be discussed in Section 5.
4
12
One disadvantage of the RD approach is that the retirement variable Ri is based on whether one retires from permanent employment or not, instead of whether one works or not (i.e. one’s labor hours are zero or not). To identify the ATE in Equation (6), we also adopt an instrumental variable (IV) approach, using the working status as the retirement variable. We refer to this as “complete retirement” in contrast to “retirement from permanent employment”. The question then is what instruments would work. Previous studies often use the retirement-inducing features of pension systems as IVs (e.g. Charles, 2004; Bound and Waidmann, 2007; Neuman, 2008; and Coe and Zamarro, 2011). We follow these studies and exploit the exogenous variations in the pensionable ages due to a series of pension reforms introduced in Japan. The Japanese pension reforms offer a quasi-experimental setup by creating the exogenous variations in eligibility ages for both the Basic Pension and the Employee’s Pension or Mutual Aid Pension (EPMAP) over time and across individuals of different genders enrolled in the two mandatory pension systems in Japan (Table 1). We exploit these variations as the instruments for retirement. In order to qualify as valid IVs, they must be correlated with retirement but uncorrelated with the unobservables in the health investment. The changes in the pensionable ages in Japan were made by the government. Therefore, they are unlikely to correlate with the error term in equation (8). On the other hand, the IVs are likely to affect retirement decisions. First, a higher pensionable age is likely to induce people to stay in the workforce longer to make up for the delayed pension benefits, even after they retire from their permanent employment. A delay in the Basic Pension payments may have a stronger effect than that in the EPMAP, not only because it is the major income source for most retirees but also because Japanese employers often set retirement ages in accordance with the eligibility age for Basic Pension. Second, delayed pensionable ages for the EPMAP may discourage working, because it implies a reduction in real wage rates as the enrollees pay their pension premiums in proportion to their total annual incomes. Formally, for each individual i and time period t: :;b = < + /;b + γ>;b + /;b = c
;b
1 if e- = 0 0 if e- > 0
e- = fg;b +
(8’) (11a)
;b (11b)
where e- is her labor supply, g- is a set of variables including the IVs and the factors 13
in >- , f is a vector of parameters, and ;b the disturbance term. g’s are assumed to be exogenous and uncorrelated with both and , but corr( ;b , ;b ) ≠ 0 because of the common unobservable factors. We employ a two-step estimation strategy to estimate the model (8’) and (11): In the first stage, we estimate a probit model of retirement against all the exogenous variables including the pensionable ages; and in the second stage, we use the predicted probability of retirement as one of the regressors and estimate a probit model for binary measures of health behaviors (i.e. currently smoking, drinking or exercising) and a linear regression model for the continuous measure (i.e. smoking intensity). Although the two-step estimation provides consistent estimates (Heckman, 1978), robust standard errors need to be corrected for their biases. Moreover, the two-step estimation strategy assumes that idiosyncratic terms - and - are time independent. To correct for potential serial correlation, we controlled for the lags of time-variant explanatory variables in all of the regressions. This seems to have reduced the serial correlation significantly (See Section 6 for more detailed discussions). 5. Institutional Background and Data Japan’s national pension system was established in 1961 as the “universal coverage” public pension system, mandating every individual to be enrolled in the system. Under the system, the enrollees must contribute to the pension fund for a minimum of 25 years in order to be eligible for the pension. The system consists of three-layer plans (MHLW, 2010a). The first layer “Basic Pension” provides basic benefits for all enrollees after the pensionable age. Everyone must participate in this first layer. The second layer “Employee’s Pension or Mutual Aid Pension (EPMAP)” supplements the National Pension. The participation in the EPMAP is mandatory for those working at companies with five or more employees, but not for self-employed individuals. Both employers and employees contribute to the EPMAP in proportion to their income level. Lastly, individuals and companies may voluntarily choose to participate in some optional retirement plans run by non-government organizations. This makes up the third layer of the pension system. To maintain the sustainability of the system, the Japanese government has launched a series of pension reforms, which raised pensionable ages for the Basic Pension and the EPMAP several times over the last three decades. The pensionable ages for females were increased from 55 to 60, starting from 1988, to match with those for men. Another amendment was made in 1994, which increased the pensionable age for the Basic 14
Pension every two years from 60 to 65, starting in 2001 for men and in 2006 for women. The most recent amendment to the pensionable ages was made in 2000, in which the pensionable age for the EPMAP was also increased from 60 to 65, starting from 2013 for male and from 2018 for female (Table 1). Retirement is often highly regulated in most Japanese enterprises in formal sectors. Most enterprises set their retirement ages the same as the eligibility age for receiving Basic Pension. In 2008, retirement is mandatory at about three-quarters of all Japanese enterprises. The percentage of the enterprises with mandatory retirement regulations increases with the size of firms, and the share reaches more than 94% for enterprises with more than 30 employees. About 67% of those requiring mandatory retirement had uniform regulations on retirement age, 82% of whom set retirement age at 60 in 2008, down 6 percentage points from 88% in 2004 (MHLW, 2009). Though the government made it legal obligation for enterprises either to delay (or abolish) mandatory retirement since 2006, the mandatory retirement custom is still persistent in the Japanese formal sectors. Because our identification strategies rely mainly on the exogenous variations that originate from these institutional setups in the permanent employment sector, we limit our empirical analysis to the sample of current and former permanent employees. The data used for this study come from the Health and Retirement Survey (HRS), a longitudinal study conducted by the Central Research Services (CRS) in Japan. The data were collected through four waves of the survey from 2008-2010. In 2008, the CRS drew a random sample of 2,747 individuals aged 45 to 80 from all the residents across Japan, following a three-step sampling method while carefully adjusting for the gender ratio and age structure of the total population. The CRS then contacted these individuals via mail or phone call to invite them to participate in the survey. About 40% of the selected sample (1,074 observations) responded with their willingness to participate. In return, the participants received a 500-yen (approximately $4.5) coupon for book purchase. As 60% of the target sample chose not to participate in the survey, the findings of this study may suffer from sample selection bias. We thus compared some major characteristics of the 1,074 observations to their national averages. Since they appear very similar, we think that the sample selection bias, if there is any, may not be serious.5 5
We have compared the major observable characteristics such as average age, gender composition, and family size with the national averages calculated using the 2010 Japanese census data (Statistical Bureau of Japan, 2010). We find very small gap in these indicators between our sample and the national averages. For example, males constitute 50% of our sample, compared to an average of 49% for the population aged 40-80. The average age of our sample is 61.2, while it is 59 for the corresponding age group surveyed in the 2010 Census. To 15
The second wave of the survey was conducted in 2009. About 20% of the original sample dropped, and the remaining sample of 862 were re-interviewed. To make up the decreased sample size, an additional sample of 257 individuals was newly selected by similar sampling procedures, resulting in a total sample of 1,119 in 2009. In 2010, 954 of the 2009 sample were interviewed again, with an attrition rate of about 15%. In the fourth wave in 2011, the sample size further dropped to 844, approximately 88% of the sample in the previous round. Although the attrition rates are generally not very large (12%-20%), we have further examined whether the dropped sample is systematically different from the remaining one.6 Our analysis confirms little difference between these two, except that the remaining sample tends to have higher level of asset holdings. Since asset holdings as well as their lags are controlled for in our analysis, we think the potential sample attrition bias is not serious. In sum, there are 1,334 individuals (3,991 observations) in our sample, among which 262 was interviewed for only once, 145 for twice, 257 for three times, and 667 for four times. The HRS has collected comprehensive data on individuals’ health behaviors and labor decisions (e.g. employment, weekly working hours, retirement, etc.). As shown in Table 2, approximately 55% of the whole sample was currently working, with 21% as a permanent employee in formal sectors, 18% as a contract or part-time employee, and 16% self-employed. Of those who were not working, more than 60% were former permanent employees. 7 They retired for various reasons, the major one being mandatory retirement (50%), followed by marriage and childbearing (10%), health problems (6%), taking care of sick family members (6%), early retirement (6%) and check for any systematic bias in unobservables, Lee and Lemieux (2010) suggest examining the outcome variables, i.e. health behaviors. Thus we compare the drinking and smoking rates of our sample to the national averages. We also find very small differences. For example, according to the National Survey of Tabaco Use conducted by Japan Tabaco, the smoking rates for men in their 40s, 50s, and 60s were 46%, 45% and 28%, respectively, in 2007, compared to the corresponding rates of 47%, 41% and 33% in our sample. The data on national smoking rates by age are available at http://www.health-net.or.jp/tobacco/product/pd090000.html, accessed on Feb 29, 2012. As for drinking, 82% of males in our sample reported to be currently drinking, consistent with the national average of 80-85% in 2003. Data on drinking rates are from http://www.mhlw.go.jp/topics/tobacco/houkoku/061122b.html, accessed on Feb 29, 2012. 6 We have run a regression of an indicator of dropping out of the sample against various personal and household characteristics, controlling for time trend and geographic heterogeneity. The probit estimates of most of the explanatory variables are insignificant, except for those of asset holdings. 7 In fact, the 2008 HRS did not ask retirees questions about their former employment. We use the information from later rounds to retrieve this information for the 2008 sample. However, since 212 observations dropped after the first round, we are not able to retrieve the details of their former employment. We group them into the “other” category in Table 2. Therefore, the share of former permanent employees is under-reported. 16
others. However, one needs to be cautious in taking these self-reported reasons which are often subjective and may not reflect the real reasons. Table 3 presents the descriptive statistics of the key variables used in the analysis based on the pooled sample. In addition to the sample averages, we have also made two comparisons: (a) working versus non-working individuals in the third and fourth columns, and (b) current versus former permanent employees in the last two columns. As shown in the table, the mean age of the HRS sample is 62.4, suggesting that the samples above and below 60 are relatively balanced. Although the gender distribution is quite even for the whole sample, the share of males is clearly much greater for the working individuals (58%), especially for the current permanent employees (75%), reflecting the well-known lower labor participation rate for women in Japan. On average, the working individuals have a higher level of education and a larger family than their non-working counterparts. The whole sample has a mean annual income of 3.05 million yen (approximately $27,450). Current permanent employees earn the most (6 million yen, or $54,000), followed by the overall working sample (4.01 million yen, or $36,090). We observe a much smaller gap between the income of retired permanent employees and that of non-working individuals. Household surveys are known for under-reporting of incomes in kind. Hence, following the literature, we use the total household assets instead as a more reliable indicator of households’ economic resources. In general, non-working individuals tend to hold a higher level of assets. The bottom panel of Table 3 presents the descriptive statistics of five measures of health behaviors. On average, about 18% of the total sample reported to be currently smoking and smoked an average of 19.1 cigarettes per day.8 Both the participation and intensity of smoking for the non-working samples are clearly lower than those of their working counterparts. Drinking is much more prevalent: About 56% of the total sample drinks some kinds of alcohol and 17% can be considered as over-drinking.9 Working individuals are more likely to drink and over-drink. Lastly, 48% of the sample reported 8
Although not shown in the table, the smoking rate has been decreasing steadily over years: 20.67% in 2008, 19.12% in 2009, 17.96% in 2010, and 15.22% in 2011. A similar pattern is observed for the intensity of smoking, dropped from 20 cigarettes per day in 2008 to 17 cigarettes per day in 2011 for those who smoke. Unlike smoking, we do not observe an obvious decreasing trend in drinking rate over years. 9 According to USDHHS (1995), over-drinking is defined as drinking more than two units of alcohol per day for men aged below 65 or drinking more than one unit of alcohol per day for women aged below 65 and for all the individuals aged above 65. The standards are similar to those set by Japan’s Ministry of Health, Labor and Welfare (MHLW), thus are applicable to Japanese population. 17
that they exercise regularly. It is clear that the rate is much higher for non-working individuals than for their working counterparts. All the comparisons in health behaviors above are statistically significant at the 1% level in a two-sample t-test. However, these differences may be driven by observable (e.g. age, sex, income) or unobservable confounders (e.g. health concerns and risk preference). Hence, we use the identification strategies discussed in Section 4 for causal inferences. 6. Results 6.1 Preliminaries The empirical regularities found in Figures 1 and 2 suggest that the RD design would be an appropriate approach to analyzing the causal effect of retirement from permanent employment, since we observe a noticeable discontinuity in the probability of retirement from permanent employment around age 60. The validity of the RD design also relies on the assumption that other factors do not change abruptly around the cutoff, i.e. no potential confounders that may bias the RD estimate. To verify this, we examined the distributions of various major covariates of health behaviors. We generally found them distributed continuously around the cutoff age. 10 Because there may be unobserved confounding factors, we also examined the distribution of the assignment variable, age, itself (Lee and Lemieux, 2010) and found it distributed continuously around the cutoff age. We now examine the distributions of five measures of the health behaviors. In Figures 5-7, the average of each measure is plotted against age for the full, female and male samples, respectively. The solid curves are the kernel-weighted local linear smoothers, calculated with the bandwidth of 2 years. Figure 5 shows that the proportion of those who were currently smoking decreases with age, with no obvious discontinuity at the cutoff (the top panel). As for smoking intensity, however, a substantial reduction in the number of cigarettes smoked per day is observed at the cutoff age for the male sample, but not for the female sample (the bottom panel). Figure 6 shows no noticeable drop in the share of drinking for the full sample. But interestingly, we observe a sharp decrease in drinking participation for the female sample, but an increase for the male sample (the top panel). On the other hand, the distributions of over-drinking appear relatively smooth (the bottom panel). Lastly, the share of regular exercising increases 10
We have examined sex, education, assets holdings, and family size. The results are available upon request. 18
slightly at the cut off age, this time for both men and women (Figure 7). 6.2 The RD Estimates: The Effect of Retirement from Permanent Employment Turn now to the RD point estimates in Table 4. We first estimate the optimal bandwidth to decide the focus sample. Since the estimation indicates an optimal bandwidth of 2.4-2.6 years for the participation of smoking, drinking and exercising, we focus on the sample aged 58 to 62 (i.e. bandwidth = 2 years) and that aged 57 to 63 (i.e. bandwidth = 3 years).11 The estimate of the optimal bandwidth is larger for smoking intensity (5.8 years), implying that the results based on the bandwidth of 2 or 3 years are consistent but may be less efficient. For each type of health behaviors, the RD point estimates based on the bandwidths of 2 or 3 years are presented in Table 4. We report the results based on the full sample in the first two columns, and those on the female and male subsamples in columns 3-6, respectively. Recall that, for the reason explained in Section 4, the denominator in Equation (9) is estimated as the discontinuity in the probability of retiring as a permanent employee. First, as shown in the top panel in Table 4, the point estimates of the retirement effect on participation in current smoking are positive for women but negative for men. They are statistically insignificant. Second, the estimated effects on smoking intensity are negative for the full sample, but are statistically insignificant. Although not shown, the estimate gets greater in absolute terms (-3.75 cigarettes per day) and also turns statistically significant (t = 1.96) if the bandwidth is increased to 6 years (closer to the optimal bandwidth of 5.8 years). When the sample is restricted to women, the estimate turns positive, yet still statically insignificant regardless of bandwidth choice. Interestingly, however, the estimate becomes much larger in absolute terms and statistically significant for men. Male retirees tend to reduce their smoking intensity right after retirement from permanent employment by 7.5-9.3 cigarettes per day. This estimated retirement effect is quite large, considering the sample mean of 19 cigarettes smoked per day. These results are consistent with the patterns observed in Figure 5. Third, retirement does not seem to affect the probability of current drinking for the full sample. However, female retirees reduce the probability of drinking after retirement from permanent employment (at the 10% significance level). In contrast, the estimated effect for men is positive and its statistical significance is sensitive to the bandwidth. 11
With the concern of the low precision due to a small sample size, we decide not to use the sample aged 59 to 61, yet the point estimates are similar to those obtained from a larger sample. 19
Similar patterns are observed for the probability of over-drinking, though the estimates are generally insignificant. Last, our results reveal that women are more likely to increase regular exercise after retirement from permanent employment: Female retirees increase the probability of regular exercising by 55-60% at the 5% significance level whereas the estimates for the male sample are very small in absolute terms and they are all statistically insignificant. In sum, our results suggest that the health-inducing effects of retirement from permanent employment are heterogeneous across gender: (a) women tend to reduce drinking and increase regular exercising, and (b) men tend to cut their smoking intensity. We believe these differential effects of retirement across gender are coming from the difference in the effects of RPE on work time allocation by gender. That is, female employees tend to retire completely (i.e. work zero or very few hours) upon mandatory retirement from their permanent employment whereas male employees tend to find alternative employment and work more hours, as evidenced in Figure 2). For this reason, we expect the IV estimates, which capture the effects of complete retirement, to be stronger in magnitude and significance particularly for the full and male samples. 6.3 The IV Estimates: The Effect of Complete Retirement12 6.3.1 Validity of IVs In the first stage estimation, we estimate a pooled probit model of retirement decision for permanent employees, with the pensionable ages for the Basic Pension and the EPMAP as the IVs.13 As noted before, the retirement variable in the IV approach is 12
As discussed elsewhere, we constrain our sample to permanent employees only. We also estimated the same regressions for the whole sample, including fixed-term, part-time, self-employed employees, and those who have never worked. The results are similar, except that the coefficients are generally smaller, which is well aligned with our expectation because non-permanent employees are less likely to be responsive to the increase in free time due to retirement. The results on the full sample are available upon request. 13 We choose the pooled specification rather than a fixed-effects or random-effects specification for the following reasons. We first conduct several statistical tests. Since our data are longitudinal, we are concerned if there are any individual-level unobservables that are correlated with other covariates. If that were the case, a fixed-effects model would be ideal for consistent estimates. The Hausman test indicates, however, that the random-effects estimates are not significantly different from the fixed effects estimates (p-values range from 0.591 to 0.776), suggesting that a fixed-effects model may not be necessary. We then compared a random-effects model against the pooled model in which time trends and geographical locations are controlled for. We find that the random-effects estimates tend to be more efficient in some regressions. However, the relatively small sample we have, especially for female, results in small variation 20
whether an individual is currently working or not (i.e. labor hour is zero or not), and thus it is an indicator of “complete retirement” for permanent employees. The results are presented in Table 5. The probit estimates for the permanent employees are in Columns 1-3. For comparison, we also present the results for non-permanent employees in Columns 4-6. The results are in line with our expectations. First, the Basic Pension eligibility age has a strong negative effect on retirement, statistically significant at the 1% level for permanent employees whereas the effect is much weaker for non-permanent employees and completely insignificant for male. This is what we expect, because non-permanent employees are not subject to mandatory retirement. Second, other covariates are generally insignificant for permanent employees whereas they tend to be significant for non-permanent employees, suggesting that nonpermanent employees tend to choose not to work for reasons other than pensionable ages. Third, the EPMAP pensionable age is insignificant for all samples probably because it was fixed at age 60 for most individuals until 2013. Following Kan (2007), we also implement a likelihood ratio test to examine the explanatory power of our IVs. Under the null hypothesis that the pensionable ages have no explanatory power, the test statistic follows an F distribution. According to the Staiger-Stock (1997) criterion, the F-statistic should be close to or greater than 10. As shown in the bottom of Table 5, the F-statistics on the sample of permanent employees are greater than 10, except for male employees (7.75). The F-statistic is far smaller for male non-permanent employees (0.22). We thus think that our IVs have sufficient explanatory power to predict complete retirement for permanent employees, and that the weak instrument bias, if there is any, would not be serious. Lastly, we also implement the over-identification test against the correlation between the IVs and the error term in equation (8’), assuming both the first stage and the second stage estimations as linear. The p-values of the Sargan’s test range from 0.12 to 0.90, implying that we cannot reject the hypothesis that our IVs are exogenous. In the bottom of Table 6, only the test statistics for the number of cigarettes smoked are reported, but the test statistics for binary variables are similar.
in outcome variables once all the covariates are controlled for. This made the estimation of a random-effects model non-convergent in some of the regressions. Hence, in order to consistently compare our estimates across different regressions, we choose to estimate the pooled probit model. Following Wooldrige (2002), we implement a Wald test for serial correlation in the idiosyncratic errors and confirm that the errors are independent over time. The inclusion of the lagged household asset holdings reduces the serial correlation substantially: The F-statistic of the Wald test increases from 0.773 (p-value = 0.379) to 10.164 (p-value = 0.00) when asset holdings are excluded. 21
6.3.2 The Effect of Complete Retirement on Health Behaviors Table 6 presents the estimates of the effect of complete retirement on smoking, drinking and exercising for permanent employees. Each regression is estimated for different genders as well as for the full sample. The first three columns report the results of the probit (for binary outcomes) or the OLS (for continuous outcomes) regressions without correcting for the endogeneity of retirement, whereas the last three columns present the IV results, the results from the two-step estimations. Robust standard errors are reported in the parentheses. First, the IV results for the full sample indicate that individuals tend to reduce smoking intensity by 2.3 cigarettes per day after complete retirement (significant at the 10% level). However, the estimated effects are quite different across genders. Women tend to increase the probability of smoking, yet their smoking intensity does not significantly increase. On the other hand, male workers do not significantly change smoking participation, but instead, reduce their smoking intensity significantly, by approximately 4.8 cigarettes per day. It may appear surprising to find that women increase the chance of smoking at retirement. However, it is indeed reasonable: Women who were prohibited from smoking in the male-dominant permanent employment may start smoking for leisure. On the other hand, although male smokers would not be able to quit smoking easily, they tend to cut the amount of smoking because of either less stress from work or higher demand for health investment. These effects are consistent with the RD results discussed in Subsection 6.2. Now turn to the regressions of drinking behaviors. The probit estimates on the full sample indicate that once controlling for endogeneity of retirement, individuals increase the probability of drinking and over-drinking (significant at the 5 % level). And this effect is driven mainly by male employees. Recall that in Subsection 6.2, the RD estimates indicate that women tend to reduce the probability of drinking after RPE, and men do not respond to RPE. This is in line with our expectation because we expect the effect of CR to be stronger than that of RPE. As discussed in Section 3, more free time associated with retirement can have health-inducing effects as well as leisure-inducing effects, and in the case of drinking, they tend to go in the opposite direction. The empirical results suggest that the leisure-inducing effect is stronger for male employees, and, therefore, they increase drinking after CR more than after RPE. Moreover, the RD estimates should be more accurately interpreted as the local treatment effect of retirement for those aged around 60 within a short period of time after retirement. This may also help to explain these differences. 22
Lastly, we observe significantly positive effects of CR on the probability of regular exercising (significant at the 1% level). The estimates turn even larger in absolute terms and more statistically significant after controlling for endogeneity, which suggest a downward bias caused by omitted variables (e.g. individuals who retire due to health problems tend to engage in less exercising). Recall that the RD estimates were significant positive for only female employees. Again, the results are consistent, not only with our hypothesis that the effects of CR are stronger than RPE, but also the fact that female employees reduce working time more dramatically after RPE than male employees who gradually withdraw from the labor force (see in Figure 2). Based on the IV estimate on the full sample (Column 4), the likelihood of regular exercising increases by 55 percentage points upon complete retirement at the sample mean, which is almost the same as the RD estimate for female. 6.4 Robustness For robustness, we also estimate a bivariate probit model of retirement and binary in equations (8’) and (11b) follow a health investments, assuming the errors u and bivariate normal distribution. The results on the health behaviors are similar to the IV estimates in terms of the signs and the statistical significance.14 Moreover, the Basic Pension eligibility age has a strong retirement-inducing effect yet the estimates of the EPMAP eligibility age are generally insignificant, in line with the first-stage estimates presented in Table 5. Lastly, the estimated correlations between the error terms are statistically significant, suggesting that retirement is indeed endogenous. Provided that an important health-inducing effect of retirement comes from the associated increase in the availability of time for health investments, one might wonder why we do not examine the effect of working hours directly. Indeed, we did. Following a similar two-step estimation strategy, we find that increased working time is associated with a lower probability of drinking and regular exercising and greater intensity of smoking for male employees. However, we choose not to follow through this strategy, because we believe the estimates could be seriously biased and no justifiable methods are available to us to take care of the bias. Working hours are not only censored at zero, but are observed for only non-retirees. Hence, we need good instruments that can predict not only the labor participation but also working hours. Unfortunately, we do not have good candidates for such instruments. Conditional on positive labor supply, it is 14
We do not report the results because the estimations for female drinking and smoking fail to converge due to the small sample size, but they are available upon request. 23
unlikely that individual significantly adjust their working time according to pensionable ages. 7. Conclusions With the longest life expectancy in the world and a constantly low fertility rate, Japan now has the largest share of the elderly population in the world. More than 23% of the Japanese population was aged 65 or above in 2010, by far the highest in the world, and it is projected to increase to 40% in 2050 (Statistics Bureau, 2011). Since most other developed countries are now moving towards an aging society, the analysis of recent Japanese data will provide important policy implications for other countries. This study investigated the causal effects of retirement on three types of health investment behaviors in Japan: smoking, drinking, and exercising. We took advantage of the two empirical strategies suggested in the earlier economic literature to take care of the endogeneity of the retirement decision. The first is the RD approach that exploits the empirical regularity that the probability of retirement from permanent employment substantially increases at age 60 due to the mandatory retirement custom in the Japanese labor market. Because a majority of the retirees from permanent employment still finds alternative employment elsewhere as fixed-term, part-time, or self-employed employees, the RD approach can only get at the effect of retirement from permanent employment. The second is the IV approach that exploits the fact that the Japanese government introduced a series of pension reforms that altered the pensionable ages for both the Basic Pension and the EPMAP. These reforms offered exogenous variations in pension eligibility ages over time and across gender, which we used as instruments for the retirement variable. Our first stage estimation shows that the exogenous changes in pensionable ages were significantly associated with the retirement decisions for permanent employees, but not for those working as fixed-term, part-time, or self-employed employees. Because the changes in pensionable ages can affect labor participation (i.e. zero labor hours or not), the IV approach allowed us to explore the effect of complete retirement for permanent employees (instead of retirement from permanent employment only). We found empirical evidence that, upon complete retirement, permanent employees significantly increased drinking and regular exercising (for both male and female), while decreased smoking intensity (for male only), and the effects of CR are generally stronger than those of RPE. On one hand, our evidence is in support of the idea that retirement is good for time-consuming health investment behaviors, and is thus related 24
to a strand of literature that finds individuals tend to increase exercising upon layoffs. On the other hand, our results suggest that this freed-up-time effect of retirement may have reverse incentives for health behaviors. More time for health investment can also mean more time for leisure, and some individuals may increase drinking, smoking, or possible other health-decreasing activities, because they may consider them as “leisure” for which more time is available. If this leisure-inducing effect is larger than the health-inducing effect, then more time after retirement could deteriorate the health of the affected pollution. Our result that individuals increased drinking or female former permanent employees increased smoking after retirement is indicative of such a concern. In an attempt to restore financial viability of pension systems, regulatory agencies would consider these important pathways in evaluating the impacts of pension reforms. Our study thus offers unique empirical evidence in this context by focusing on health behaviors rather than health outcomes and by exploiting the quasi-experimental setups in the Japanese labor market.
25
References: Banks, J., Blundell, R., Tanner, S., 1998. “Is there a retirement-savings puzzle?”, American Economic Review 88(4): 769-788. Battistin, E., Brugiavini, A., Rettore, E., Weber, G., 2009. “The retirement consumption puzzle: evidence from a regression discontinuity approach”, American Economic Review 99(5): 2209-2226. Bound, J., Waidmann, T., 2007. “Estimating the health effect of retirement”, University of Michigan Retirement Research Center Working Paper 2007-168. Charles, K.K., 2004. “Is retirement depressing? Labor force inactivity and psychological well-being in later life”, Research in Labor Economics 23: 269-299. Dave, D., Rashad, I., Spasojevic, J., 2008. “The effects of retirement on physical and mental health outcomes”, Southern Economic Journal 75: 497-523. Coe, N.B., Zamarro, G., 2011. “Retirement effects on health in Europe”, Journal of Health Economics 30(1): 77-86. Deaton, A., 1992. Understanding Consumption. Oxford: Oxford University Press. French, E., 2005. “The effects of health, wealth, and wages on labour supply and retirement behavior”, Review of Economic Studies 72: 395-427. Grossman, M., 1972. “On the concept of health capital and the demand for health”, Journal of Political Economy 80(2): 223-255. Hahn, J., Todd, P., Klaauw, W.V., 2001. “Regression discontinuity”, Econometrica 69: 201-209. Heckman, J., 1978. “Dummy endogenous variables in a simultaneous equations system”, Econometrica 46 (4): 931–959. Henkens, K., Van Solinge, H., Gallo, W., 2008. “Effects of retirement voluntariness on changes in smoking, drinking and physical activity among Dutch older workers”, European Journal of Public Health 18(6): 644. Imbens, G., Kalyanaraman, K., 2009. “Optimal bandwidth choice for the regression discontinuity estimator”, NBER Working Paper 14726. Imbens, G., Lemieux, T., 2008. “Regression discontinuity designs: A guide to practice”, Journal of Econometrics 142(2): 615-635. Johnston, D.W., Lee, W.S., 2009. “Retiring to the good life? The short-term effects of retirement on health”, Economics Letters 103: 8-11. Kajitani, S., 2011. “Working in old age and health outcomes in Japan”, Japan and the World Economy 23: 153-162. Kan, K., 2007. “Cigarette smoking and self-control”, Journal of Health Economics 26: 26
61-81. Kerkhofs, M., Lindeboom, M., 1997. “Age related health dynamics and changes in labor market status”, Health Economics 6: 407-424. Kuhn, A., Wuellrich, J., Zweimüller, 2010. “Fatal attraction? Access to early retirement and mortality”, IZA Discussion Paper Series 5160. Lang, I., Rice, N., Wallace, R., Guralnik, J., Melzer, D., 2007. “Smoking cessation and transition into retirement: analyses from the English Longitudinal Study of Ageing”, Age and Aging 36: 638-643. Lee, D.S., Lemieux, T., 2010. “Regression discontinuity designs in economics”, Journal of Economic Literature 48(2): 281-355. Lindeboom, M., Portrait, F., van den Berge, G.J., 2002. “An econometric analysis of the mental-health effects of major events in the life of older individuals”, Health Economics 11(6): 505-520. Midanik, L., Soghikian, K., Ransom, L., Tekawa, I., 1995. “The effect of retirement on mental health and health behaviors: the Kaiser Permanente Retirement Study”, Journals of Gerontology Series B: Psychological Sciences and Social Sciences 50(1): 59-61. Mein, G., Martikainen, P., Hemingway, H., Stansfeld, S., Marmot, M., 2003. “Is retirement good or bad for mental and physical health function? Whitehall II Longitudinal Study of Civil Servants”, Journal of Epidemiology and Community Health 57: 46-49. Ministry of Health, Labor and Welfare, 2009. Survey on Employment Conditions of Elderly Persons in 2008, http://www.mhlw.go.jp/toukei/itiran/roudou/koyou/keitai/ 08/index.html. Ministry of Health, Labor and Welfare, 2010. Annual Health, Labor and Welfare Report 2009-2010, http://www.mhlw.go.jp/english/wp/wp-hw4/dl/pension_security/2011072 501.pdf. Neuman, K., 2008. “Quit your job and live long? The effect of retirement on health”, Journal of Labor Research 29(2): 177-201. OECD, 2007. Society at a Glance: OECD Social Indicators 2006. Osada, H., Ando, T., 1998. “Effects of retirement on mental health and psychological well-being”, Job Stress Research 5: 106-111. (in Japanese) Perreira, K., Sloan, F., 2001. “Life events and alcohol consumption among mature adults: A longitudinal study”, Journal of Studies on Alcohol 62: 501-508. Staiger, D., Stock, J.H., 1997. “Instrumental variables regression with weak instruments”, Econometrica 65(5): 557-586. 27
Statistics Bureau of Japan, MIC, 2011. Statistical Handbook of Japan 2011. Tokyo: Ministry of Internal Affairs and Communication of Japan. Statistics Bureau of Japan, 2010.National Results of the 2010 Census, http://www.e-stat. go.jp/SG1/estat/List.do?bid=000001034991&cycode=0. Stokey, N., Lucas, R., Prescott, E., 1989. Recursive Methods in Economic Dynamics. Harvard University Press, Cambridge, Massachusetts. Sugisawa, A., Sugisawa, H., Nakatani, Y., Shibata, H., 1997. “Effect of retirement on mental health and social well-being among elderly Japanese”, Japanese Journal of Public Health 44(2):123-30 (in Japanese). U.S. Department of Health and Human Services (USDHHS), Public Health Service, National Institutes of Health, National Institute on Alcohol Abuse and Alcoholism, 1995. “The Physicians’ Guide to Helping Patients with Alcohol Problems”, NIH Publication No. 95-3769. Wooldrige, J., 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge: the MIT Press. Yamada, A., 2010. “Labor force participation rates of older workers in Japan: Impacts of retirement policy, steep age-wage profile, and unionization”, The Japanese Economy 37(1): 3-39.
28
Figure 1: Working Status by Age (b) Contract/Part-time Employee 1 .8 0
.2
.4
.6
.8 .6 0
.2
.4
Proportion
1
(a) Permanent Employee
40
50
60
70
80 Age
50
60
70
80 Age
.8
1
(d) Not Working
0
.2
.4
.6
.8 .6 0
.2
.4
Proportion
1
(c) Self-employed
40
40
50
60
70
80
40
50
60
70
80
Age
Age
Figure 2: Working Time by Age (Permanent Employees Only) Age
40
50
60
70
80
Age
Age
10
20
30
40
50
Male
0
10
20
30
40
50
Female
0
10
20
30
40
50
Age
0
Mean Working Time (hour)
Both Genders
40
50
60
70
80
Age
40
50
60
70
80
Age
29
Figure 3: Monthly Wage Rates by Age and Gender
Source: The Basic Statistics of National Wage Structure, Ministry of Health, Labor and Welfare, 2011.
Figure 4: The Effect of Mandatory Retirement on Health Investment
c
( )+
ℎ
=
1 +
(b) (a)
nl1
BC1
nl0
BC0 y
30
Figure 5: Smoking Participation and Intensity by Age
1
.8 .6 .2 0
.2 0 45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
Intensity of Smoking
#
10
10
0
5
5 0
5
10
15
15
#
15
#
0
# of Cigarettes Smoked Per Day
Male
.4
.6
.8
Female
.4
.6 .4 0
.2
Proportion
.8
1
Full
1
Participation of Smoking
45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
Age
Age
Age
Figure 6: Drinking and Over-drinking by Age Participation of Drinking
.8 .6 .4 0 45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
Participation of Over-Drinking
Quantity of Drinking
.5 .4 %
.2
0
0
.1
.1
.2
%
.3
.3
.4
.4
.5
.5
Quantity of Drinking
.3 .2 .1 0
Male
.2
.2 0 45 50 55 60 65 70 75 80
Proportion
Percentage of Drinking 1
1 .6 .4
.6 .4 0
.2
Proportion
Female
.8
Full
.8
1
Percentage of Drinking
45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
Age
Age
45 50 55 60 65 70 75 80
Age
Age
31
Figure 7: Regular Exercising by Age Female 1
Male
0
.2
.4
.6
.8
1 .8 .6 0
.2
.4
.6 .4 .2 0
Proportion
.8
1
Full
45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
45 50 55 60 65 70 75 80
Age
Age
Age
32
Table 1: Changes in Pensionable Ages in Japan Female
Male
Birth Year
Basic Pension
EPMAP Pension
Birth Year
Basic Pension
EPMAP Pension
before 1932 1932-1933 1934-1935 1936-1937 1938-1939 1940-1945 1946-1947 1948-1949 1950-1951 1952-1953 1954-1955 1958-1959 1960-1961 1962-1963 1964-1965 after 1965
55 56 57 58 59 60 61 62 63 64 65 65 65 65 65 65
55 56 57 58 59 60 60 60 60 60 60 61 62 63 64 65
before 1932 1932-1933 1934-1935 1936-1937 1938-1939 1940-1945 1941-1942 1943-1944 1945-1946 1947-1948 1949-1950 1953-1954 1955-1956 1957-1958 1959-1960 after 1960
60 60 60 60 60 60 61 62 63 64 65 65 65 65 65 65
60 60 60 60 60 60 60 60 60 60 60 61 62 63 64 65
Source: Ministry of Health, Labor and Welfare, 2010.
33
Table 2: Working Status by Sampling Year Working
2008 2009 2010 2011 Total
Not Working
Permanent Employees
Contract or Part-time Employees
Selfemployed
Retired as a Permanent Employee
Other
248 (24) 249 (22) 197 (21) 158 (19)
202 (19) 206 (19) 174 (19) 148 (18)
167 (16) 178 (16) 141 (15) 139 (17)
164 (16) 234 (21) 147 (16) 123 (15)
261 (25) 246 (22) 268 (29) 249 (30)
1,042 (100) 1,113 (100) 927 (100) 817 (100)
852 (21)
730 (18)
625 (16)
668 (17)
1,100 (28)
3,975 (100)
Total
Notes: 1. In the parentheses are the percentages within each row. 2. “Other” includes former fixed-term, part-time, self-employed employees as well as those who have never worked or have no information on former employment status.
34
Table 3: Descriptive Statistics
Obs.
Mean
Mean
Mean
Current Permanent Employees (852) Mean
3,975 3,975 3,890 3,853 2,671 3,462
62.4 0.50 2.35 3.06 305 4.73
58.5 0.58 2.50 3.30 401 4.57
67.7 0.39 2.15 2.73 173 4.96
54.0 0.75 2.88 3.47 600 4.62
68.3 0.65 2.40 2.68 228 5.21
3,952 711 3,924 3,924 3,933
0.18 19.1 0.56 0.17 0.48
0.23 19.9 0.65 0.19 0.38
0.12 17.0 0.45 0.15 0.61
0.29 19.9 0.76 0.23 0.36
0.15 17.1 0.56 0.20 0.67
Full Sample (3,975)
Working (2,207)
Nonworking (1,768)
Retired Permanent Employees (668) Mean
Personal Characteristics Age Sex (1 = male) 1
Education level Family size
2
Personal annual income (10,000yen) 3 Total household assets Health Behaviors Currently smoking (1 = yes) 4
# of cigarettes smoked per day Currently drinking (1 = yes) 5
Over-drinking (1 = yes) Regular exercising (1 = yes)
Notes: 1. Education levels are coded as: 1 = lower secondary; 2 = upper secondary; 3 = vocational; 4 = tertiary. 2. Averages from only three sampling years, 2009-2011, since the survey only asked individuals to report the range of their incomes in 2008. 3. Total household asset is the sum of the current values of the household’s savings, stocks, bonds, and real estate. Asset levels are coded as: 1 = less than 3 million yen; 2 = 3-8 million yen; 3 = 8-12 million yen; 4 = 12-20 million yen; 5 = 20-30 million yen; 6 = 30-50 million yen; 7 = 50-100 million yen; 8 = 100-150 million yen; 9 = more than 150 million yen. 4. Averaged over those who were currently smoking. 5. Over-drinking is defined as drinking more than two units of alcohol per day for males aged below 65. Women aged below 65 and all the individuals aged above 65 are considered to over-drink if drinking more than one unit of alcohol per day.
35
Table 4: The RD Estimates The Effect of Retirement from Permanent Employment on Health Behaviors Full Bandwidth Currently smoking (%) RD point estimate Standard error Obs. # of cig. smoked per day RD point estimate Standard error Obs. Currently drinking (%) RD point estimate Standard error Obs. Over-drinking (%) RD point estimate Standard error Obs. Regular exercising (%) RD point estimate Standard error Obs.
Note:
Female
Male
2
3
2
3
2
3
0.098 0.136 690
0.108 0.106 953
0.039 0.195 349
0.068 0.158 480
-0.040 0.151 480
-0.022 0.117 473
-1.574 3.560 690
-2.884 2.857 953
1.311 3.495 349
1.678 2.993 480
-7.470 * 4.110 480
-9.303 *** 3.249 473
0.146 0.165 685
0.132 0.127 945
-0.546 * 0.329 347
-0.461 * 0.271 475
0.153 0.125 475
0.161 * 0.097 470
0.199 0.124 694
0.165 * 0.097 958
-0.054 0.201 352
-0.015 0.164 483
0.183 0.135 483
0.145 0.106 475
0.178 0.168 682
0.198 0.129 942
0.043 0.159 475
0.024 0.120 467
0.553 * 0.338 346
0.598 ** 0.278 475
* significant at the 10% level; ** significant at the 5% level; *** significant at the 1% level.
36
Table 5: The First-Stage Estimations of Retirement Decision Permanent Employees
Basic Pension eligibility age EPMAP eligibility age Age Sex (1 = male) Senior high school (1 = yes) Vocational school (1 = yes) University or above (1 = yes) Family size Asset level dummies Lags of Asset level dummies 1 Year dummies 2 Location dummies Obs. 3 Weak IV tests
Non-Permanent Employees
(1) Both
(2) Female
(3) Male
(4) Both
-0.592 *** (0.069) 0.010 (0.090) 0.030 * (0.016) 0.160 (0.226) 0.076 (0.306) 0.410 (0.376) -0.042 (0.315) -0.104 ** (0.049)
-0.679 *** (0.114) 0.130 (0.229) -0.011 (0.023) 0.678 (0.499) 0.779 (0.554) 1.292 ** (0.623) -0.009 (0.100)
-0.619 *** (0.130) 0.098 (0.125) 0.074 ** (0.032) -0.199 (0.410) 0.272 (0.500) -0.379 (0.422) -0.082 (0.067)
Yes Yes Yes Yes
Yes Yes Yes Yes
Yes Yes Yes Yes
Yes Yes Yes Yes
Yes Yes Yes Yes
Yes Yes Yes Yes
851 42.09 [0.000]
241 26.65 [0.000]
610 7.75 [0.000]
1,208 10.47 [0.000]
661 13.64 [0.000]
547 0.22 [0.801]
-0.231 (0.058) 0.013 (0.071) 0.029 (0.013) -0.801 (0.115) -0.282 (0.121) -0.396 (0.173) -0.213 (0.165) -0.068 (0.038)
***
*** *** ** **
*
(5) Female -0.329 (0.080) 0.004 (0.098) -0.009 (0.019) -0.518 (0.188) -0.554 (0.242) -0.387 (0.285) -0.113 (0.052)
(6) Male
***
-0.068 (0.116) -0.058 (0.187) 0.089 *** (0.027) *** -0.031 (0.182) ** -0.644 * (0.376) -0.168 (0.229) ** -0.085 (0.057)
Notes: 1. Three year dummies for sampling years. 2. Twelve dummies for 13 regions across Japan. 3. F-statistics of the tests against the explanatory power of the excluded IVs. 4. * significant at the 10% level; ** significant at the 5% level; *** significant at the1% level. 5. Robust standard errors are included in the parenthesis and p-values in the brackets.
37
Table 6: The IV Estimates The Effect of Complete Retirement on Health Behaviors Without Accouting for Endogeneity (1) Both
(2) Female
Currently smoking (1 = yes) Coefficient 0.263 0.211 Standard error (0.168) (0.534) Obs. 849 188 # of cig. smoked per day Coefficient 0.291 0.171 Standard error (0.851) (0.975) Obs. 849 240 Currently drinking (1 = yes) Coefficient -0.157 -0.554 Standard error (0.141) (0.260) Obs. 842 233 Over-drinking (1 = yes) Coefficient 0.205 1.491 *** Standard error (0.154) (0.554) Obs. 837 140 Regular exercising (1 = yes) Coefficient 0.654 *** 0.857 *** Standard error (0.138) (0.230) Obs. 844 232 Over-identification test
(3) Male
Accouting for Endogeneity (4) Both
(5) Female
(6) Male
0.122 (0.181) 609
0.207 (0.335) 849
1.441 ** (0.702) 188
-0.313 (0.393) 609
0.476 (1.074) 609
-2.284 * (1.470) 849
1.057 (1.371) 240
-4.762 ** (2.436) 609
-0.164 (0.165) 603
0.701 ** (0.289) 842
0.046 (0.429) 233
0.848 ** (0.408) 603
0.029 (0.166) 603
0.788 ** (0.291) 837
0.701 (0.585) 140
0.729 (0.404) 603
0.355 ** (0.163) 607
1.400 *** 1.466 *** (0.280) (0.391) 844 232
1.535 *** (0.385) 607
2.394 [0.122]
0.507 [0.477]
0.781 [0.377]
Notes: 1. Each regression has controlled for all the explanatory variables in the first-stage estimation. 2. Over-identification tests are reported for the number of cigarettes smoked only; Results for other regressions are similar. 3. * significant at the 10% level; ** significant at the 5% level; *** significant at the 1% level. 4. Robust standard errors are included in the parentheses and p-values in the brackets.
38
Appendix: Labor Supply and Health Investment Decisions Given the primitives of the model laid out in Section 3, the consumer chooses an optimal consumption and investment path i , ℎ , j and the final period T, maximizing the discounted sum of utilities: ∑kl(
( ,ℎ ,
; ),
subject to the constraints (1)-(3) and given her initial health capital ℎ( , where β represents the consumer's time preference. We assume strict concavity on u and f and a strictly increasing depreciation rate to ensure that death takes place at some finite period. As discussed in Section 3, the signs of the marginal utility of health investments could be positive or negative. As in Grossman, death takes place at period T when ℎk ≤ ℎm-n . We assume ( , ℎ , ; ) = 0 for ℎ ≤ ℎm-n for all ( , ℎ , ). Let χ be an indicator variable
such that χ = 0 if ℎp ≤ ℎm-n for some ≤ and one otherwise. This last assumption captures the idea that once death takes place, the consumer can gain no utility thereafter. Under some regularity conditions, the maximization problem can be reformulated as a recursive dynamic programming problem (Stokey, Lucas, and Prescott, 1989) as follows: (ℎ; ) = iqUr s 4 ( , ℎ, ; ) +
(ℎE ; + 1): χ = 16j,
where primes indicate variables' values in the next period. Note that the last period’s value is (ℎm-n ; *) = ( , ℎm-n , ; *) = 0 by assumption. The optimal last period can be readily derived by the one-step-look-ahead rule. That is, conditional on ℎ > ℎm-n , choose health investments such that ℎE ≤ ℎm-n if and only if it is at least as good or better to die now (i.e. at t+1) than to continue living one more period (i.e. until t+2): max ( , ℎ, ; ) ≥ maxi ( , ℎ, ; ) +
( E , ℎE ,
E
; )j.
Substituting the constraints (1)-(3), we can rewrite this problem further in terms of the current decision variables. (ℎ; ) = max s i ( (v − − ∙ ) −
∙
+
, ℎ, ; ) 39
+ Hence, for (
( ( ; ) + (1 − )ℎ; + 1): χ = 1j.
< *, the first-order condition for health investment
+ ) ( )=
( )+
( + 1) ( ).
is:
40
