Bulletin of the World Health Organization

Employment and health after retirement in Japanese men

Shohei Okamoto a, Tomonori Okamura b & Kohei Komamura c

a. Graduate School of Economics, Keio University, 2-15-45 Mita, Minato-ku, Tokyo, Japan.
b. Department of Preventive Medicine and Public Health, Keio University, Tokyo, Japan.
c. Faculty of Economics, Keio University, Tokyo, Japan.

Correspondence to Shohei Okamoto (email: shohei@z2.keio.jp).

(Submitted: 01 May 2018 – Revised version received: 06 September 2018 – Accepted: 10 September 2018 – Published online: 27 September 2018.)

Bulletin of the World Health Organization 2018;96:826-833. doi: http://dx.doi.org/10.2471/BLT.18.215764


In response to an ageing population and the associated financial challenges, many countries have introduced extensions to working life; however, research on the resultant effect on health is currently insufficient. In Japan, the age of income-related pension eligibility is gradually being increased from 60 years in 2013 by 1 year, every 3 years, to reach 65 years in 2030 for both men and women.1 Policies which set minimum mandatory retirement age (usually 60 years, but can vary from one company to another) have recently been updated, and workers are now free to continue in paid employment past this age if they desire.2

Although several studies have already investigated the association between health and retirement,311 they failed to identify causal relationships. As the age of retirement and condition of health can be interdependent, health selection exists;12,13 causal inference approaches are therefore required to measure the impact of working past current retirement age on health. A limited number of empirical studies have been conducted to determine causal impacts, but results are inconsistent. Some studies report negative effects on self-rated health, cognitive functioning and depressive syndrome because of retirement;1417 however, other studies have shown that retirement can lead to improvements in health as indicated by the numbers diagnosed with chronic conditions, activity limitations and self-rated health.18,19 Only a few studies have accounted for the healthy worker effect when assessing the health consequences of working past retirement age. These studies did not observe any health benefits.17,20

We contribute to this area of study in two ways. We first expand the health outcomes examined compared with those in previous publications,36,9,1719,21 which have generally only considered self-rated physical or mental health, activity limitations and cognitive decline, to consider the three additional health outcomes of death, stroke and diabetes. Analyses of stroke and diabetes are particularly beneficial in investigating the rising costs of health and long-term care in the Japanese population.22 Second, we quantify the effect of working past retirement age in terms of delay in years to the onset of these health outcomes. We believe this approach provides greater insight into healthy life expectancy compared with studies that investigate the likelihood of certain diseases by estimating hazard or odds ratios.



We retrieved data from the National Survey of the Japanese Elderly,23 a six-wave longitudinal prospective survey of a sample of Japanese men and women aged 60 years or older. The survey began with a population of 2200 in 1987, and was subsequently supplemented in later waves in 1990, 1993, 1996, 1999 and 2002 (Table 1). The sample was extracted from the Basic Resident Registration System, administered by the Japanese government, by stratified two-stage random sampling. Municipalities were first selected based on regions and population size, and then survey participants were randomly selected in a manner consistent with the age distribution of Japan; face-to-face interviews were performed at each wave. We have focused our analysis on men only, since labour force participation rates were lower for women (50.1% in 1998 versus 77.3% for men).24 Response rates at each wave are listed in Table 1.

Health outcomes

We used publicly available data, in which each survey participant was followed-up for the four health outcomes at each wave for at most 15 years. Information regarding the death of a participant, including date, was obtained from the Basic Resident Registration System. Assessment of the three other outcomes (cognitive function, stroke and diabetes) was made at each wave.

Cognitive functioning was assessed based on the short portable mental status questionnaire.25 This memory test comprises nine questions: the respondent’s address; date of interview; day of interview; mother’s maiden name; name of the current prime minister; name of the previous prime minister; a simple calculation; the subject’s birthday; and the subject’s age. This is a similar approach to standard memory tests used in the field of cognitive science.26,27 If the respondent made more than five mistakes, their cognitive functioning was recorded as moderately declined.25 If a participant was not actually able to complete an interview due to cognitive decline, they were assessed as having a cognitive impairment.

The onset of stroke or diabetes was determined according to self-reported symptoms in the survey questionnaire. The questionnaire also recorded information on whether disease onset was based on a diagnosis by a medical professional or by the participant.

Independent variables

We based the values of the independent variables used to adjust for confounders on the baseline (i.e. first) survey responses for each participant. The most important reason for this was that some participants had undergone a decline in cognitive functioning, reducing the reliability of their answers. Individuals older than 75 years at baseline were excluded from the analysis, since only 48 men were in employment.

We defined being in employment as a binary variable of value 1 if individuals were receiving a wage from their job at baseline, including all contract types (i.e. as a full- or part-time employee or being self-employed), and 0 if not employed.20 We included demographic information in the form of age and marital status, defining marital status as a binary variable of value 1 if married and 0 if not married (including those bereaved or divorced). We considered socioeconomic data in terms of home ownership (1 if people owned their own homes, 0 otherwise), equalized household income (the household income divided by the square root of the number of household members), educational attainment (proportion within one of four categories of years in education)28 and proportion within one of five longest-held employment types. Following Kan,23 a respondent’s longest-held occupation type was classified as either: professional or administrative; service or clerical; agriculture, forestry or fisheries; manual labour; or self-employed. We included condition of health in the model in terms of whether a smoker (1 if yes, 0 if not), self-rated health (1 for poor health, 0 otherwise) and body mass index (kg/m). To attempt to counteract the healthy worker effect, individuals of abnormal weight (i.e. a mean body mass index ± 3 standard deviations from the mean) were excluded from the analysis.

Propensity score method

We adopted the propensity score method to estimate the effects on health of being in employment when older than 60 years. The propensity score method was originally developed to investigate the presence of a causal inference in non-experimental (i.e. observational and therefore non-randomized) surveys, while attempting to reduce the bias due to confounding variables;29 previous investigations have demonstrated its use in dealing with potential observed endogeneity (i.e. outcomes which are not independent of baseline characteristics).30 We have applied the propensity score method here to account for the fact that individuals in poorer health may have retired earlier.12,13

We calculated the propensity score using Stata software, version 14.2 (StataCorp LLC, College Station, United States of America), as the conditional probability of being in employment or not at baseline. This conditional probability was obtained by logistic regression using the covariates describing the economic, sociodemographic and health characteristics of survey respondents. We then used the inverse probability of being in paid work (i.e. propensity score) to balance the covariates between those in employment and those not employed.

Incorporating the propensity score, we used survival analysis to estimate time to a particular health outcome from the first survey while assuming a Weibull distribution. By calculating the differences in times to an outcome for both those in employment and those not employed, we estimated the weighted average treatment effects on health of being in paid work past retirement age.


We retrieved data for 1288 men that were eligible for analysis on an available-case basis. The health outcome data for the participating men are provided in Table 2. Men not employed had a shorter time to death than men with employment (7.66 years versus 9.31 years; P < 0.001). We observed similar outcome for cognitive decline: men not employed had 7.58 years to onset, while men with employment had 11.20 years (P = 0.003). Men not employed had 5.84 years to the onset of stroke, which was less than the 8.03 years for men with employment (P = 0.001). Time to onset for diabetes was not significant between the groups (4.06 years for men not employed versus 3.96 years for men with employment; P = 0.319; Table 3).

The differences in values of independent variables between those in employment and those not employed are shown in Table 3. Participants in employment had higher incomes than those who were not employed (whose income was mostly in the form of pension benefits). Subjective feelings of being in poor health were significantly higher for individuals not in employment (average: 17%) compared with those who were employed (average: 7%). This example highlights how individuals not in employment may have been in that position due to poor health, which would reinforce the healthy worker effect unless adjusted for. We observed another major difference in the longest-held occupation type; the proportion of self-employed among individuals still working was much higher than those no longer in employment (averages of 68% versus 7%).

Table 4 provides the estimated average treatment effects of being in employment past retirement age compared with not being employed on the four different health outcomes. We cannot reject the null hypotheses that the covariates were balanced between the employed and not employed groups; this implies that confounders have been adequately adjusted for.

There were significant differences in all four outcomes between the two groups: participants in employment were likely to experience longer periods of good health regarding all four outcomes compared with those not employed. On average, individuals not employed died 9.40 years after their baseline surveys; those in employment died 1.91 years later (95% confidence interval, CI: 0.70 to 3.11). Participants not employed underwent cognitive decline 10.58 years after their baseline; those in paid work did not experience cognitive decline for another 2.22 years (95% CI: 0.27 to 4.17). Being in employment prolonged the disease-free period for stroke by 3.35 years (95% CI: 1.42 to 5.28) beyond the 11.08 years of onset among participants not employed. The onset of diabetes among individuals not employed began 8.52 years after baseline; for those in employment, it occurred 6.05 years later (95% CI: 4.44 to 7.65).

A substantial proportion of participants were self-employed. We therefore performed another analysis, separating those in employment into employees and the self-employed, and compared with those not employed (Table 5). Life expectancies were significantly longer among both employees (average treatment effect: 1.88 years; 95% CI: 0.35 to 3.73) and the self-employed (average treatment effect: 2.74 years; 95% CI: 1.30 to 4.17) compared with those not employed (onset of death at 9.17 years). Significant improvements in morbidity in terms of diabetes and stroke were only observed among employees, however, and not the self-employed. In comparisons with those not employed, extensions to healthy life were observed for employees for diabetes (average treatment effect: 6.30 years; 95% CI: 5.00 to 7.61) and stroke (average treatment effect: 3.90 years; 95% CI: 1.67 to 6.13). Health benefits for the self-employed for diabetes (average treatment effect: 0.19; 95% CI: –1.72 to 2.09) and stroke (average treatment effect: 1.64; 95% CI: –0.87 to 4.15) were not observed, however. We did not observe any significant difference between employees and the self-employed in terms of years to onset of cognitive decline.


In line with studies which evaluated the health impacts of retirement in Japan and the United States,1417 our estimates of average treatment effects imply that being in paid work past the current age of retirement has positive effects on health.

There are three main interpretations of the results. First, according to the human capital model,31 people may be incentivized to invest in health for the sake of increasing or maintaining their productivity and wages. If individuals have a problem with their health, they may experience presenteeism or absenteeism; they may even have to retire earlier than planned as a result of poor health. Second, studies have reported3234 that the social participation and networks that accompany most jobs can enhance health. Having increased social relationships is directly linked to improved physical and mental health outcomes, for example avoiding stress and depression.35,36 Third, the latent functions of employment, such as social status and self-respect, also contribute positively to health outcomes.3739 We did not investigate which of these three hypotheses has the greatest effect; these hypotheses need to be examined in detail in further studies before they can be usefully incorporated by policy-makers.

In an additional analysis, we examined the difference between effects on health for employees and the self-employed. Employees had better health than the self-employed regarding estimated onset to stroke and diabetes, whereas the estimated life expectancy of the self-employed was longer. One of the most significant explanations is that, in accordance with industrial safety and health law, employers in Japan are required to provide annual health check-ups; these are mainly focused on cardiovascular risk factors (e.g. hypertension and diabetes). Health check-up attendance among the self-employed is very low compared with employees, as these must be organized by the individuals.40 In addition, since the wages of employees can be determined by their employers according to their performance and human capital,31 which are in turn affected by their health, employees may be more motivated to manage their health than the self-employed. Policies which encourage the self-employed to seek and maintain better health are therefore necessary, such as financial incentives.41

Our study has some limitations. First, the case ascertainment of diabetes and stroke was based on self-reported symptoms, whereas death and cognitive decline were measured from demographic statistics and an established test. The questionnaire did record whether participants had been diagnosed with diabetes or as having had a stroke by a medical professional, but the source of diagnosis was not considered in our analysis. Caution should therefore be exercised when interpreting the results regarding these two health outcomes.

Second, our sample set comprised respondents of the National Survey of the Japanese Elderly, all of whom had survived to age 60 years or older, perhaps contributing to the healthy worker effect. We attempted to overcome this problem by using the propensity score method with survival analysis; by incorporating differences in health, measured from variables including body mass index and self-rated health, we have considered the consequences of behavioural health. In addition, the survival analysis method is focused on the onset of new health outcomes; those who were already unhealthy, that is, with a history of any of the health outcomes observed here, were excluded from the analysis.

Third, we did not consider changes in employment status. In an extreme case, a respondent could have changed their status from employed to not employed immediately after the baseline survey. There may also have been changes to health related to the total time in employment during the follow-up period. Further, beyond distinguishing between being employed or self-employed, we did not consider the frequency of work, job security or availability of any fringe benefits; such factors can also have an effect on health,42 and future investigations should consider these.

In conclusion, this empirical study implies that being in paid work in later life is beneficial for both mortality and some outcomes of morbidity, prolonging healthy periods. Although more research is needed to investigate the effects of transition of employment status and other precise health outcomes, our estimates of average treatment effects indicate that extending working lives has benefits to health. Policies designed for longer working lives should incorporate appropriate management of lifelong physical and mental health practices and conditions.


We thank faculty members of King’s College London, London School of Economics and Political Science, Keio University and Kyoto Sangyo University. The data for this secondary analysis, “the National Survey of the Japanese Elderly , the University of Michigan, and the Tokyo Metropolitan Institute of Gerontology,” was provided by the Social Science Japan Data Archive, Centre for Social Research and Data Archives, Institute of Social Science, The University of Tokyo.


This research is supported in part by a research assistantship of a Grant-in-Aid to the Program for Leading Graduate School for “Science for Development of Super Mature Society” from the Ministry of Education, Culture, Sport, Science, and Technology in Japan.

Competing interests:

None declared.