Introduction

Retirement is an important event in the life of an individual. Among many other factors (socio-economic and financial), individual health status may have a strong influence on retirement decisions. For the last few decades, population aging has foreshadowed serious policy problems throughout the world, including within the United States. With the rapid rise in aging population in U.S, an increasingly high proportion of individuals are fast approaching their full retirement age (65 years). According to Social Security Administration (Office of Policy), the proportion of people over 65 years of age in 2009 was 12.9% as compared to 8% in 1950. Over next few decades, as the baby boom generation (born during 1946–1964) enters their elderly years the proportion of individuals over 65 years of age is projected to rise to 20% (in 2040).

These demographic changes suggest financing challenges for transfer programs such as the Social Security program. As a result, policymakers may promote policies designed to retain productive older workers in the workforce in order to defer their Social Security payments. The success of such policies depends in part on the ability to identify the key determinants of a worker’s decision to retire and the magnitude of their impact. One obvious factor that plays into a worker’s decision to continue working or retire is their health status. Policies that improve the health of workers may encourage them to continue working and defer the start of their Social Security payments.Footnote 1 The purpose of this paper is to examine the role of physical and mental health conditions in determining the duration of full-time employment for older Americans. In other words, this paper measures the extent to which health influences the decision to retire among older U.S. workers.

There is a literature focused on identifying the causal effect of health on retirement (Anderson and Burkhauser 1985; Bazzoli 1985; Bound 1991; Bound et al. 1998; Disney et al. 2006; Dwyer and Mitchell 1999; Lindeboom and Kerkhofs 2002). These studies primarily used self-reported subjective measures of health. Such measures of health may be plagued with problems that lead to bias. First, self-reported measures of health are based on subjective judgments and there is no reason to believe that these judgments are comparable across individuals. Second, since poor health may represent a legitimate reason for a person of working age to be outside the labor force, respondents who are not working may cite health problems as a way to rationalize behavior (the “justification hypothesis”). A final issue is that many papers in the literature are forced to rely on relatively short panels due to limitations in data availability at the time of the studies.

This paper adds to the existing literature by using duration analysis on a panel dataset of older Americans (those at least 50 years old and working full-time in 1992) and utilizes a wider variety of health indices to estimate the impact of health on the duration of full-time employment. Unlike much of the previous literature, the long panel nature of the Health and Retirement Study dataset is exploited here, making it possible to observe more cases of actual retirement than retirement plans (expectations) and the potential to observe multiple spells of employment over a 20-year timeframe.Footnote 2 Excluding younger individuals and those initially working less than full-time reduces concern about the justification bias. The issue of subjectivity in health outcomes is addressed by constructing eight relatively objective health indices (factors) through principal component analysis (PCA) that are based on a broad range of subjective as well as objective health measures.Footnote 3 Unlike previous studies, it is possible to observe how different health measures load in these indices. Besides the health indices created by PCA, changes in physical and mental health between consecutive waves are considered.

The main results for the overall study sample suggest that among physical health factors, an increase in the functional limitations factor increases likelihood of exit from full-time work by 18.88% overall, which is largely driven by exits via complete retirement or disability routes. However, the probability of exit for complete retirement is much larger in magnitude. On the other hand, cancer leads to a 7.71% decrease in likelihood of exit via complete retirement. Among mental health factors, an increase in cognitive disorders has no significant effect on exit via the complete retirement route but leads to a 1.14% increase in likelihood of exit via disability. An increase in depression factor leads to 9.06%, 3.04 and 0.90% increase in likelihood of exit via complete retirement, part-time work and unemployment routes respectively. An increase in the risky lifestyle behavior factor (smoking, drinking) leads to an increase in likelihood of exit via complete and partial retirement routes.

The rest of this paper is organized as follows: section 2 reviews previous literature and section 3 presents the data. A discussion of the empirical methodology is provided in section 4 followed by a discussion of results in section 5 and ending with a conclusion in section 6.

Literature Review

An early paper on this topic, Boskin (1977), does not find any significant effect of health on retirement, but a large effect of Social Security income. Unlike Boskin (1977), Quinn (1981) finds that the presence of a health limitation reduced the probability of labor force participation by 20 percentage points. Bazzoli (1985) looks at the determinants of early retirement and the impact of various measures of health. She also addresses the issue of the relative importance of health and economic factors in influencing early retirement. She finds economic factors rather than health factors play the major role in retirement decisions.

In attempting to identify the causal effect of health on retirement decisions, the use of subjective measures of health has been a focus of much attention in the literature (Anderson and Burkhauser 1985; Bazzoli 1985; Bound 1991; Bound et al. 1998; Disney et al. 2006; Dwyer and Mitchell 1999; Lindeboom and Kerkhofs 2002). Such measures of health may be plagued with problems that lead to bias. First, self-reported measures of health are based on subjective judgments (leading to bias) and there is no reason to believe that these judgments are comparable across individuals. The size of the bias present in self-reported health measures is documented in Benítez-Silva et al. (2004). Second, self-reported health may not be independent of labor market status. Third, since poor health may represent a legitimate reason for a person of working age to be outside the labor force, respondents who are not working may cite health problems as a way to rationalize their behavior (“justification hypothesis”). Fourth, for individuals for whom the financial rewards of continuing in the labor force are low there exists a financial incentive to report poor health as means of obtaining disability benefits. This is often cited as the “disability route” into retirement (Marmot et al. 2002; Riphahn 1999). For example, in a study of social security benefit programs in the Netherlands, Kerkhofs and Lindeboom (1995) show that recipients of disability insurance systematically overstate their health problems.

A large literature attempts to address this concern about subjective reports of health status. To mitigate response bias, authors have attempted to use arguably more objective measures of health, such as the observed future death of respondents (Anderson and Burkhauser 1985; Parsons 1982). Parsons (1982) and Stern (1989) find those who withdraw from the labor market are likely to cite poor health as the cause even if they are not in poor health, simply because they may be rewarded for doing so through eligibility for transfers. Their findings suggest that the traditional measure of self-reported health is endogenous due to “justification hypothesis.” In other words, people who intend to keep working will downplay their health problems while ones who dislike work and wish to exit from the labor force will exaggerate their health problems. Dwyer and Mitchell (1999) implement an instrumental variable approach to deal with endogeneity using the first four waves of HRS. The authors used parental health and mortality, respondent’s age, number of children, and BMI as instruments for self-assessed health. They do not find evidence to support the justification hypothesis. This could be because in the first four waves the HRS sample is still relatively young and does not yet consist of a majority of retired individuals.

Bound et al. (1998) examined the relationship between health and alternative labor force transitions like retirement, job change, and application for disability insurance. Their analysis not only considers health status, but also declines in health and its effect on the work behavior of individuals. According to the authors, retirement is often a last resort. Prior to such an outcome, workers may resort to increased effort, putting in more time, requesting a reduction in performance standards, or changing jobs in order to accommodate their physical limitations.

Using the first few waves of the HRS, McGarry (2004) models the labor market behavior and retirement probabilities of older workers prior to their eligibility for early retirement benefits and Social Security. She finds that changes in retirement expectations are driven more by health than economic variables. The effect of subjective measures of health is strong even when objective measures are included. Miah and Wilcox-Gök (2007) also use HRS data to determine how chronic illness affects asset accumulation and retirement. They find that the vast majority of the chronically ill population does not report their general health to be poor nor do they report functional limitations in activities of daily living. Nevertheless, their results indicate that chronic illness leads these people to accumulate fewer assets during their working years and consequently retire later.

While most studies in the existing literature use a fixed effect approach on panel data from different countries to investigate the effect of health on retirement, Meghir and Whitehouse (1997), Christensen and Kallestrup-Lamb (2012) and Siddiqui (1997) use duration models to study effect of poor health on labor force transitions for UK, British and Danish panel surveys respectively. In sum, these findings strongly suggest that poor health is a determinant of retirement or exit from the work force.

Data

The analysis presented in this paper exploits a long panel of data for Americans (1992–2010) from the Health and Retirement Study (HRS) conducted by the Institute for Social Research at the University of Michigan.Footnote 4 The HRS is an ongoing longitudinal survey, which began in 1992, and is conducted in biennial waves. Prior to 1998, the main HRS cohort included individuals born between 1931 and 1941, and another distinct cohort, the Study of Assets and Health Dynamics among the Oldest Old (AHEAD), included individuals born before 1924. Since 1998, the data for these two cohorts is collected jointly, and the sample frame has been expanded to include cohorts born between 1924 and 1930 and those born between 1942 and 1947. The HRS is administered for the specific purpose of studying life-cycle changes in health and economic resources, and includes detailed information on various subjective and objective health outcomes. In this paper, I focus on a sample of older individuals who were at least 50 years of age who were working full-time in 1992. The sample consists of 4128 individuals having multiple records that generate 15,442 person-wave observations. This was obtained by strictly dropping all cases with inconsistent or missing information for health measures and socio-demographic or economic variables.

Employment Spells

I consider two types of employment spells in my analysis. I start by focusing on initial employment spells by following the 4128 individuals in my sample starting in wave 1 (1992) over the subsequent waves until their first observed exit from full-time employment occurs (initial exit model).Footnote 5At the end of each two year wave of the HRS, an individual who was working full-time in the previous wave could either continue working (and thus be treated as a right censored observation for that wave) or exit full-time work via one of the five different routes: complete retirement, partial retirement, part-time work, unemployment and disability. The exit routes are defined using the labor force participation variable.Footnote 6 Individuals leave full-time employment through one of the five exit routes mentioned above or leave the survey via attrition or for some other reason, such as death. Those that remain in full-time employment or that exit the survey are treated as right censored employment spells.Footnote 7 Overall, 3.4% of the sample remains continuously full-time employed across all of the waves we observe. Since I am focusing here on initial employment spells, any subsequent transitions back to full-time work are ignored. This implies I am considering retirement to be an absorbing state.

The other type of employment spell I consider in my analysis is a subsequent employment spell (multiple exit model). In other words I drop the assumption that retirement is an absorbing state and allow members of the sample that retire to contribute additional spells of full-time employment if they re-enter the labor market. This adds 740 additional employment spells to the 4128 initial employment spells described above, for a total of 4868 full-time employment spells generated by my sample.Footnote 8

Figure 1 depicts the baseline hazard rate for individuals to exit over time via any route in general and also via the five different exit routes. It is observed that the hazard rate for exiting full-time employment in general (via any route) cumulatively rises over time with a distinct peak occurring in Wave 7 (2004). This is largely driven by the rise in hazard rate of exiting via the complete retirement route which also peaks in Wave 7. This is probably because the individuals who are 50 years of age or older in 1992 become eligible for retirement benefits around the same time. The figure also indicates that the baseline hazard rate for the different exit routes is non-linear and not constant over time which calls for its parametric estimation using a suitable distribution.Footnote 9 The Kaplan-Meir survival estimates indicate that the probability of surviving in full-time employment declines over time. However, this decline is larger in the initial waves. This is true for all routes combined and the complete retirement and partial retirement routes. For the part-time work, unemployment and disability or not in labor force routes, the decrease in survival rate almost flattens out over time.

Fig. 1
figure 1

Cumulative hazard rates for different exits (1992–2010)

Full size image

Health Measures

I use a wide range of health measures in this study in an attempt to address the concern that many health measures, such as self-reported health, are based on individual perceptions and may be plagued by misreporting and measurement error. Some of these measures are relatively more objective than others are and have not been used in the previous studies. Below I discuss these health measures by grouping them into four broad categories.

Self-reported health: This is the most subjective measure of health and has been widely used in existing studies. In the HRS, individuals may report their health as excellent, very good, good, fair and poor.

Physical health: Some of my measures of physical health have been used in prior studies, including counts of difficulties with activities of daily living (ADL) and diagnosed chronic conditions. The ADL difficulties include difficulties with daily chores like bathing, eating, getting dressed, getting in or out of bed, and walking across a room. The chronic conditions include diseases like blood pressure, diabetes, lung disease, heart disease, stroke, cancer, arthritis and psychological disorders. The dichotomous variables created for these conditions are based on whether the individual has been medically diagnosed and whether he has been using prescription drugs or undergoing therapy to treat this specific condition in the previous two years, to get relatively more objective indicators of physical health.

Mental health: The existing literature has mostly overlooked measures of mental health. The studies that do include mental health only account for depression while paying no attention to cognitive disorders. I measure mental health by using information on depression and cognition as well as other diseases like Alzheimer’s and Dementia. These mental health conditions are also validated with prescription drug use information. In the HRS, depression is measured in a 0–8 scale, as defined by the Center for Epidemiologic Studies on Depression (CESD). This CESD score measures the sum of adverse mental health symptoms for the past week, based on if the respondent felt depressed, felt that everything was an effort, had restless sleep, was not happy, felt lonely, felt sad, could not get going, and did not enjoy life. Studies have confirmed this to be a valid and reliable indicator for incidence of major depression in older adults (Irwin et al. 1999). Information on measures of cognitive functioning is also included in my analysis. The cognitive functioning measures include immediate and delayed word recall, the serial 7’s test, counting backwards, naming tasks (e.g., date-naming), and vocabulary questions. In addition to the individual cognitive functioning measures, the HRS also derives three cognition summary indices. The total recall index which is available for all waves is a concise summary of the immediate and delayed word recall tasks. The mental status index adds the scores from counting, naming, and vocabulary. To maintain consistency across waves, I have used the total cognition score in this study, which sums the total recall and mental status indices and thus ranges from zero to thirty-five.

Other health measures: In addition to self-reported health and measures of physical and mental health, some other measures of health (ignored in existing studies) are also used in this study. These measures include body mass index (BMI), work related stress, physical effort at work, extent of physical exercise, number of nights at hospital, number of doctor visits, risky behaviors like smoking and drinking, and out-of-pocket-medical expenditure.

Descriptives

One way to analyze the impact of health status on the decision to exit full-time employment is to investigate whether or not there are baseline differences (in 1992) in health status between those that are observed working full-time throughout the sample and those that are observed exiting full-time work via one of the routes. Table 1 presents such a comparison for some important standard measures of health. In general, those that subsequently exit from full-time work seem to have worse baseline measures of health than those that remain working full-time. For example, those that exit from full-time employment via complete retirement are more likely to report poor health, ADL difficulties, multiple chronic conditions, depression, and psychological problems. A similar pattern is observed for individuals who subsequently exit via other routes. There is statistically significant difference in means of the health outcome measures for samples of working individuals and individuals who exit via one of the routes as reported in Table 1. The baseline differences in socio-demographic and economic variables for the individuals working full-time across all waves and those that subsequently exit via one of the routes are reported in Table 2. The individuals who subsequently exit via complete retirement route are older, more likely to be male, married and in blue-collar jobs with lower individual and household income as compared to those that remain in labor market full time. For most health measures baseline difference in means for individuals who continue in full-time work and ones who subsequently exit is significantly driven by exit categories-complete retirement and disability. For few health measures, there is statistically significant difference in mean for exit categories partial retirement and part-timework (compared to full-time work) but for unemployment exit route, the differences are not statistically significant. However, mixed results are observed for socio-demographic variables. Baseline difference in a majority of socio-demographic and economic variable means for individuals who continue in full-time work and ones who subsequently exit are driven by exit categories part-time work and disability.

Table 1 Descriptive statistics for some health outcome variables measured in 1992
Full size table
Table 2 Descriptive statistics for socio-demographic variables measured in 1992
Full size table

Health Indices

To mitigate the potential difficulties arising due to use of subjective self-reported measure of health, Bound (1991, Bound et al. 1998) suggested an approach that involves estimating a model of self-reported health as a function of more objective measures of health to define a latent ‘health stock’ variable. This health stock variable is then used as a measure of health in a model of retirement / exit from full-time employment. This idea of constructing a health stock is in many ways analogous to using objective measures of health to instrument for the endogenous and potentially error-ridden self-reported health variable. In this paper, the latent health stock variable is predicted by using an ordered probit model for self-reported health, where the ordered measure of self-reported health is regressed on a set of relatively more objective health measures reported in Table 1. More health problems are associated with a lower order of the latent health stock. Unfortunately, this method for creating a latent health stock is not very effective at suggesting how the different individual health measures are weighted. This can be a problem because clearly neither high blood pressure nor diabetes is the same as cancer. Physical health outcomes are also clearly different from mental health outcomes. On the other hand, including every health measure separately in a regression model will make it cumbersome. Hence principal component analysis is used as a comprehensive way to extract eightFootnote 10 meaningful factors (indices) from twenty-eight individual health outcomes. For each factor, it is possible to note how the different individual health measures weigh.Footnote 11 It is important to note that self-reported health does not load heavily in any factor, which implies that it is not the best indicator of health of an individual. Based on the health outcomes that load heavily in each factorFootnote 12 I label them: Factor 1: Has chronic conditions, Factor 2: Has functional limitation, Factor 3: Hospital stay, Factor 4: Has cognitive functioning problems, Factor 5: Has depression, Factor 6: Lack of physical exercise, Factor 7: Has cancer, and Factor 8: Has lifestyle behavioral problems.

Empirical Method

In this paper, a standard proportional hazard model is used to estimate the impact of health on the duration of full-time employment, where time is measured in two-year waves. In some specifications, only initial employment spells are included in the sample, while in other specifications I include subsequent employment spells as well. Another way in which I differentiate the model is to combine all five exit routes in some specifications (a combined risk or lumped risk model) while other specifications each exit route is treated separately (a competing risk model).

More formally, the competing risk proportional hazard model is given by:

$$ {\mathrm{H}}_{\mathrm{j}}\ \left(\mathrm{t}\right)={\mathrm{H}}_0\ {\left(\mathrm{t}\right)}^{\ast }\ \exp\ \left({{\mathrm{X}}^{'}}_{\mathrm{it}}\upbeta \right) $$
(1)

Here, j is an index for each of the five exit routes and Xit is the vector of covariates that vary with time while H0(t) is the baseline hazard that only depends on time but not individual covariates which means it is common for all units. The impact of the observable characteristics is parametrically estimated using the standard proportional hazard functional form exp. (X’itβ).

Given that the hazard is not constant over time (time-dependency of hazard rates), it is important to choose a suitable parametric distribution for estimating the baseline hazard. If the chosen distribution correctly characterizes the time-dependency, then the parameter estimates are likely to be more precise than the parameter estimates of semi parametric or non-parametric models where the time-dependency is left unspecified. Hence, there are advantages of using a suitable parametric model. But problems may arise if a wrong parametric form is chosen. The most common approach for choosing an appropriate parametric model is based on using the Akaike Information Criterion (AIC). It is based on penalizing the log likelihood to reflect the number of parameters being estimated by different models (distributions) and comparing them. Although the best fitting distribution is the one with the largest log likelihood, the one with smallest AIC is most preferred. Table 3 presents the log likelihood and AIC information for different parametric models. Given the smallest AIC, the Weibull distribution is chosen for parametrically estimating the baseline hazard.Footnote 13According to the proportional hazard specification stated earlier, the Weibull hazard rate is given as:

$$ \mathrm{H}\ \left(\mathrm{t},\mathrm{X}\right)=\uplambda\ \mathrm{p}\ \left(\lambda t\right)\ \mathrm{p}-1 $$
(2)
Table 3 Log likelihood and akaike information criterion for different parametric models
Full size table

Where, λi = eXiβ and p is the shape parameter.

In all specifications, in addition to the socio-demographic and economic variables reported in Table 2, each specification I estimate includes spousal health and work status, occupations, census regions, expected longevityFootnote 14 and controls for general economic conditions (through wave dummies). In order to estimate the model with standard software, an independence assumption across the exit routes is imposed. Then this independence assumption is tested by estimating Martingale residuals for each exit route and checking their correlations for statistical significance, as in Borgan and Langholz (2007) and Marton et al. (2010).

Hazard models may be plagued by duration dependence, which arises due to unobserved heterogeneity.Footnote 15 Ignoring unobserved heterogeneity may exaggerate the rate of failure for some individuals and underestimate the rate of failure for others. In this context, unobserved heterogeneity is addressed through the addition of an additional random parameter or “frailty term” to the model. In the proportional hazards model, the hazard rate increases or decreases with the covariates. The problem is that if there are unmeasured or unobserved ‘frailties’, then the hazard rate will be a function of the covariates and the frailties. Eq. (1) may be rewritten as:

$$ Hj\ \left(\mathrm{t}\right)={\mathrm{H}}_0\ {\left(\mathrm{t}\right)}^{\ast } \exp\ \left({{\mathrm{X}}^{'}}_{\mathrm{i}\mathrm{t}}\upbeta +{\upvarepsilon}_{\mathrm{i}}\right) $$
(3)

Or,

$$ Hj\ \left(\mathrm{t}\right)={\mathrm{H}}_0.{\mathrm{v}}_{\mathrm{i}}. \exp\ \left({{\mathrm{X}}^{'}}_{\mathrm{i}\mathrm{t}}\upbeta \right) $$
(4)

Where vi = exp. (εi).

So, the frailty term acts multiplicatively on the hazard rate. The hazard rate is conditional on both the covariates and the frailty. For identification purposes, it is assumed that mean of v = 1 and the variance is equal to some unknown finite parameter θ. Then the unobserved heterogeneity or frailty is modeled using a gamma distribution and effectively the frailty variance θ is estimated. The hypothesis that θ = 0 may be tested using a likelihood ratio test to determine whether unobserved heterogeneity is something to worry about in the model.

This is equivalent to the inclusion of a random effects term in a standard panel data model. In some specifications, I include a frailty term that is modeled parametrically using the gamma distribution.Footnote 16There is no hazard model equivalent to a fixed effect panel data model. In the results section below, I investigate this potential limitation by estimating standard linear probability models of retirement with both fixed and random effects to see if there is a big different in the coefficients.Footnote 17

Results

Controlling for Unobserved Heterogeneity in Hazard Models Vs. Panel Data Models

I start by estimating a simple linear probability model of complete retirement using the standard health measures from existing studies and only the first five waves of the HRS.Footnote 18 Both fixed effect (FE) and random effect (RE) models are presented in panel A of Table 4. These results suggest poor self-reported health, multiple ADL difficulties and chronic diseases, heart disease and stroke lead to an increase in probability of complete retirement. In this case, the Hausman-Wu test rejects the null hypothesis that the difference in the coefficients in the FE and RE models is not systematic. This implies that the FE model does a better job of controlling for unobserved heterogeneity. Panel B illustrates that adding an additional five waves to the sample does not affect the coefficient estimates in a major way, except that having psychological problem now leads to a statistically significant increase in probability of complete retirement. The Hausman-Wu test again rejects the hypothesis that the RE and FE coefficients are the same.

Table 4 Fixed and random effect estimates in linear model
Full size table

As mentioned in Section 3.4, there are advantages associated with using the health indices (factors) that come from a PCA to control for health status, rather than the limited individual health measures typically used in the literature. Panel C includes the eight health indices described in Section 3.4 rather than the typical health indicators from the literature.Footnote 19 When we use 10 waves of the HRS and include our PCA health indices to measure individual health status, the Hausman-Wu test cannot reject the hypothesis of equality of the RE and FE coefficients. Given that the RE and FE coefficients are so similar in the linear probability model framework, I have confidence that the lack of a fixed effects specification within a hazard model framework is not a serious limitation in my subsequent analysis presented below. In other words, this implies that the individual random effects (frailty) in my hazard model will control for the same unobserved factors as would individual fixed effects.

Hazard Model

The hazard ratios from parametric hazard model (Weibull) are reported for initial exit and multiple exit models with and without frailty for combined risk competing risks specification. Table 5 reports the hazard ratios for the health indices (factors) for both the combined risk and competing risks specifications estimated without a frailty term over all initial employment spells. The latent health stock variable has statistically significant hazard ratios 0.845 and 0.796 for the combined risk model and the complete retirement exit route, respectively. This implies an increase in one’s latent health stock (i.e. better self-reported health) makes an individual 15.5% less likely to exit from full-time employment in general and 20.4% less likely to exit via complete retirement route. The functional limitations factor has statistically significant hazard ratios of 1.101, 1.136 and 1.278 (significant at 1% level) for the combined risk model, the complete retirement exit route, and the disability/not in labor-force exit route, respectively. This implies that multiple functional limitations (ADL difficulties) that limit mobility and work, increase the probability of exit from full-time employment by 10.1% in general, 13.6% via the complete retirement exit route, and 27.8% via the disability exit route. More generally, a hazard ratio for an independent variable greater than 1 implies that the presence of (or an increase in) that variable leads to an increase in the likelihood of an exit (i.e. worse chances of survival in full-time employment). The opposite is true if the estimated hazard ratio is less than 1 (i.e. better chances of survival in full-time employment). The magnitude of change is calculated as (1-Hazard Ratio). A positive value would signify better chances of survival (lower likelihood of exit) while a negative value would signify worse chances of survival (higher likelihood of exit).The p-values are for the hypothesis test that the hazard ratio for the variable in question is equal to 1 (i.e. no effect). The magnitudes of these effects can be difficult to interpret, because they are relative probabilities. Therefore, the absolute effectFootnote 20associated with each independent variable area also reported. For each exit route, these absolute effects can be compared to the average probability of exiting via that route.Footnote 21 For example, the absolute effect associated with the functional limitations factor in the combined risk model suggests that an increase in functional limitations makes an individual 18.10% more likely to exit from full-time employment, which is greater than the average exit probability of 16.44% for every period. An increase in functional limitations also makes an individual 9.58% and 1.08% more likely to exit via complete retirement and disability routes respectively, which is higher than the average probabilities to exit via those routes (8.44% and 0.85%, respectively) every wave. Other statistically significant health factors include depression, risky lifestyle behavior like drinking and smoking and cancer. It is observed that an increase in the cancer factors leads to a lower probability of exit via any route in general and the complete retirement route in particular.

Table 5 Competing risk model (weibull distribution) for initial exit from full-time employment (without frailty)
Full size table

In Table 6, I present the same models but estimate them using both initial and subsequent employment spells (multiple exit model). While this increases the sample size, it does not generate large changes in the coefficient estimates. Higher latent health stock (better self-reported health) still makes an individual less likely to exit. For physical health conditions, an increase in functional limitation factor makes an individual more likely to exit via any route in general and via the complete retirement and disability routes in particular. Among the mental health factors, an increase in depression raises the probability of exit via any route in general and through complete retirement route in particular. It is interesting to note that cognitive functioning disorders factor have no statistically significant effect on exit through the complete retirement route, but an increase in problems related to cognitive functioning makes an individual 1.12% more likely to exit via the disability route. Increase in risky behaviors makes an individual 9.47% more likely while cancer makes one 7.78% less likely to exit via the complete retirement route.

Table 6 Competing risk model (weibull distribution) for exiting full-time employment allowing for multiple spells of employment (without frailty)
Full size table

In Table 7, I re-estimate the models from Table 6 on the same sample, but now include a frailty term in the model to control for unobserved heterogeneity (this represents the most complete model).Footnote 22 These results suggest that increase in latent health stock makes an individual 13.72% less likely to exit from full-time employment in general, 6.44% less likely to exit via complete retirement and 0.37% (significant at 1% level) less likely to exit via disability route. These are slightly lower than the average probabilities associated with these exit routes. For the indices of physical health, an increase in the functional limitations factor increases the likelihood of exit from full-time work by 18.88% overall, by 9.50% for the complete retirement exit route and by 1.07% for the disability exit route (significant at 1% level). These are higher than the average exit probabilities associated with these exit routes (bottom of Table 7). For mental health factors, an increase in depression factor increases the likelihood of exit from full-time employment for an individual by 18.68% in general, 9.06% via the complete retirement route (significant at 1% level), 0.90% via unemployment route (significant at 5% level) and 3.04% via part-time work route (significant at 10% level). Increases in cognitive problems factor have no statistically significant effect on the likelihood of exit via complete retirement, but increases the likelihood of exit via the disability exit route by 1.14% (significant at 1%). The risky behavior factor leads to 9.59% (significant at 1% level) and 4.82% (significant at 5% level) higher probability of exit via complete retirement and partial retirement respectively, while the cancer factor leads 7.71% lower likelihood of exit via complete retirement. The likelihood ratio test for the estimates reported in Table 7 rejects the null hypothesis of “no frailty”, which implies the existence of unobserved heterogeneity that needs to be accounted for.

Table 7 Competing risk model (weibull distribution) for exiting full-time employment allowing for multiple spells of employment (with frailty)
Full size table

The results from the parametric model are compared to the semi parametric Cox proportional hazard model estimates to check for the consistency of the estimates.Footnote 23 The hazard ratios from the Weibull parametric model are directly comparable to the hazard ratios reported in the Cox model. The hazard ratios obtained in the Cox proportional hazard model, are qualitatively similar, consistent in statistical significance but slightly smaller in magnitude for the different health indices for both combined risk and competing risks case. The inference drawn about the effect of the health factors on probability of exit from full-time employment for both Weibull and Cox models is similar. Among physical health factors, functional limitation leads to higher likelihood of exit via complete retirement and disability, with the magnitude of the effect being much smaller for the disability route. While for mental health factors, depression leads to higher likelihood of exit via complete retirement and unemployment while cognitive problems lead to higher probability of exit via the disability route. Lifestyle risky behavior also increases the likelihood of exit via complete retirement. In the Cox proportional hazard model, unlike the parametric Weibull model, cancer does not statistically significantly decrease the probability of exit from full-time employment via complete retirement route.

As a robustness check, I examine the impact of changes in health outcomes between waves, rather than simply looking at levelsFootnote 24 for the overall sample (i.e. without splitting by age). Changes in self-reported overall health, counts of chronic conditions, counts of functional limitations (ADL difficulties), and the onset of memory related diseases between waves are considered.Footnote 25 Extreme reductions in these measures between waves serve as a proxy for exogenous changes in health. The results suggest that a major reduction in overall self-reported health increases the likelihood of exit from full-time employment via complete retirement, and disability. Increases in counts of chronic conditions and onset of memory related diseases, between waves increase the likelihood of exit via the complete retirement and the disability exit routes. While increases in functional limitations between waves have no statistically significant effect on exiting via complete retirement, but increases the likelihood of exit via disability.

In summary, the overall sample results indicate that physical health problems (functional limitations) lead to increases in the likelihood of exit from full-time employment in general, which one can attribute to the increase in the likelihood of exit via the complete retirement route and the disability route. The magnitude of the effect is much smaller for the disability route. As for mental health problems, depression increases the likelihood of exit via complete retirement, while cognitive disorders increase the likelihood of exit via the disability exit route (with no statistically significant effect on the likelihood of exit via complete retirement).

Conclusion

This study contributes to the existing literature by empirically modeling the duration of full-time employment of older Americans using a long panel from the Health and Retirement Study. I distinguish between the different exit routes from full time employment and allow for multiple employment spells. Moreover, this study addresses the inherent problem of the subjectivity of health measures in surveys by constructing relatively objective comprehensive indices of physical and mental health that take into account a wide variety of health indicators based on both medical diagnosis and medication. The PCA method used for construction of the health factors (indices) is not only an effective method of data reduction but also helps to get uncorrelated explanatory variables (health factors). This is particularly important because physical and mental health outcomes are likely to be highly correlated which can lead to endogeneity problem and hence biased estimates. The PCA analysis helps to address this issue although the causal effect of the constructed health factors is not strongly established. Moreover, unlike existing studies, I am able to distinguish between different dimensions of physical and mental health (functional limitations versus chronic conditions and depression versus cognition) and their impact on continued employment.

Consistent with the findings of most existing studies, the main inferences drawn from the results of the most complete model (multiple spells with frailty) indicate that better self-reported health decreases the likelihood of exit from full- time employment. It is also found that physical and mental health problems are both impediments to continued work. Few studies that have explored similar research question based on HRS are divided in their findings. Dwyer and Mitchell (1999) find evidence that overall poor health leads to early retirement plan among men. But few other studies like McGarry (2004) and Miah and Wilcox-Gök (2007) find opposite results. McGarry (2004) finds that poor health has very strong correlation with the decision to remain employed. Similarly, Miah and Wilcox-Gök (2007) find that chronic illness is associated with higher probability of retiring later because individuals with chronic illnesses are able to accumulate less assets over time. Given the lack of consensus in existing literature it is useful to explore this association between health and retirement. The findings of this paper are more informative. I am able to disaggregate the wide variety of health outcomes and observe their differential impact on duration of full-time work. Among physical health factors while functional limitations lead to a higher likelihood of exit from full-time employment, incidence of cancer leads to a decrease in likelihood of exit from full-time work. Among mental health factors, depression leads to an increased likelihood of exit via complete retirement while cognitive problems have no statistically significant impact. Moreover, due to the competing risks specification in hazard analysis framework I am able to distinguish between different routes of exit i.e. some health conditions may decrease the likelihood of continuing in full-time employment but need not imply full-retirement i.e. due to certain health problems like depression an individual may exit from full-time work into part-time work or unemployment while cognitive disorders leads to an increase in likelihood of exit from full-time employments via disability route.

These results produce targets for policies that seek to improve the health of older working Americans. Improving the health of older workers means they can be retained in the labor force for an extended period of time, which would result in decreased training costs for replacement workers, the ability to maintain the experience and productivity of these older workers, and the ability to defer their Social Security benefits.

Limitations of the study stem from not being able to adequately mitigate the existence of “justification bias” although there are mixed empirical findings about the existence and magnitude of such bias. It would also be important to include future leads regarding health for the older workers, in addition to measuring past and current health. This study has opened interesting possibilities for future research. For example, it would be interesting to further investigate transitions in and out from full-time employment to the different exit routes. This could indicate whether improvements in health bring retirees (ones who have exited) back into the labor force full time.