Health and retirement age: Comparison of expectations and actual retirement

Good health is not equally distributed in society. Individuals and socioeconomic groups with frequent health problems are likely to face higher poverty risk. With aging this risk increases if those with poor health retire early with low pensions. They are also affected when early retirement channels are abolished and old-age pensionable ages are raised. Poor health may lead to withdrawal from the labor market if there are no early exit options, or to forced continuation at work despite a lower work capacity. However, the increase in the healthy lifespan has been at odds with the downward trend in the labor supply of older workers [1]. The relationship between health and retirement is likely to be affected by attitudes and preferences as well as the pension and labor market institutions. There is clearly a need for better understanding of the role of health in retirement.

Extensive research evidence indicates that health status plays an important role in retirement decisions. There are many mechanisms that shape the relationship between health and retirement [1,2]. First, poor health is likely to increase the disutility of work, regardless of age. These negative effects on labor supply are likely to gain strength as people get older since aging typically brings with it an increasing number of health risks. This effect is typically amplified by pension and social security systems that often include early exit options, which thus reinforce the effects of health deterioration on retirement. Second, health problems may also hamper work performance and labor productivity and contribute negatively to labor demand and earnings, as individuals who have poor health may have fewer job opportunities. In a more dynamic setting, health status can also influence individuals’ expectations of their life expectancy and their plans on how to use their presumed remaining time in and outside of the labor market. Accordingly, health status already at a prime working age can influence expectations related to the length of the working career and timing of retirement. In line with this, changes in health status make individuals reformulate their expected labor supply choices. The analysis of the connection between health and retirement is, in practice, made difficult by various statistical problems, including appropriate measurement of health and possible two-way causality between health and work status. A large literature exists on different ways of solving these problems [1,2].

Earlier studies on the health–retirement connection mostly use actual retirement time (e.g. [3–10]), sometimes expected retirement time [11,12], and sometimes both (e.g. [13–17]). Expected retirement has been used both because this information may be easily available from surveys and because the anticipations of individuals—the expected retirement age—can be seen as good predictors of their actual behavior. Comparison of expectations and realization has indeed been one issue in the studies.

This paper analyzes the role of health in retirement decisions by modeling the relationships between self-assessed health (SAH) status, expected full-time retirement age, and actual full-time retirement age. Compared to the earlier studies that use both expectations and realizations, we emphasize a systematic analysis of the measurement issues. Specifically, the purpose of this paper is as follows: (a) to compare SAH as a predictor of actual retirement age, expected retirement age, and expectation error; (b) to examine the use of other health measures as alternative variables or as instruments for SAH to solve the problem of SAH as an incomplete measure of health; (c) to examine the role of future health shocks for retirement to solve the problem of reverse causality; and (d) to investigate potential health-related sample-selection problems as retirement timing is more likely to be observed when health is poor. We use Finnish survey data combined with longitudinal register data. This allows us to have information on expectations and SAH from the survey, and actual retirement time from the follow-up data.

The studies on labor supply and retirement decisions of older workers typically find that health status variables, together with incentives, have the most predictive power. This result is not surprising since health typically declines with age, and at some stage this decline is likely to create a mismatch between the work requirements of an individual and their work capacity. This creates a link between poor health and earlier retirement. Some studies have particularly aimed to compare the relative importance of financial incentives and health in retirement decisions [8], finding that the former are the most important in relation to other early retirement schemes, and the latter is dominant in explaining transitions to disability and unemployment insurance.

Furthermore, the effect of health on retirement is clearly dependent on various institutions like pension systems and on various factors linked to the position of the individual in the labor market. The role of institutions is analyzed, for example, in studies where different retirement channels are analyzed separately [8]. There are also interactions between health, work ability, and working conditions. While chronic health problems are likely to affect productivity negatively, these effects can be buffered if resources and demands at work are adapted to the needs of employees suffering from health problems.

While the studies are quite consistent in finding a relationship between health and retirement, there is still a lot of uncertainty regarding it. The most disputed questions are related to measuring the health status and to the endogeneity of health in the labor supply analysis. The most common method of measurement is to use self-reported health status [18], and questions arise about the accuracy of this subjective indicator.

There are various possible sources of bias in the estimation of the relationship between health and (actual or anticipated) retirement [2]. First, SAH may be correlated with the unobservables in the error term, because health and work status can be simultaneously determined. That is, just being at work, especially the content of work and working conditions, will likely affect health status. Therefore unobservable, work-related effects correlate with the health measure. For example, the factors that affect health positively may also influence the willingness to continue at work beyond the official retirement age.

Second, SAH may be correlated with the error term because of justification bias. People who have psychic or financial incentives to retire early may overstate their health problems to justify their non-employment status.

Third, there is likely to exist an errors-in-variables problem, as the SAH measure is not perfect and does not necessarily measure well work-related health which is relevant for retirement [3,9,19]. It has been suggested that the biases caused by endogeneity and measurement error may be in different directions [3]. Further, the health measure based on survey data is typically measured in a Likert scale (e.g. 1 to 5). Using this ordinal measure as if it were cardinal creates a measurement error. One potential problem is related to comparability of individuals when assessing the scale for the health status. They may use different scales for defining poor and good health [7,9].

Fourth, when the relationship between health and retirement age is studied, there are sample selection issues. Since in an analysis of retirement behavior people are still working at the beginning of the data period, those with good health are likely to be overrepresented as those with the poorest health have already withdrawn from the labor market; for example, through disability retirement. This is the so-called healthy worker effect [20]. On the other hand, among those in the data, actual retirement age is more likely to be observed for those with poor health if healthier individuals are more inclined to continue working. Therefore, people with good health are underrepresented in the sample of completed employment spells. Ignoring this selection problem leads to a bias in the health–retirement age connection.

Earlier research has suggested various ways of dealing with some of these issues. Endogeneity of health has been tackled by modeling the health stock. In this case, health is modeled as a function of other health measures, personal characteristics, past health, etc. [5,8,11]. Use of panel data also helps in wiping out time-invariant factors but does not work if unobservables, like true health, change over time [3]. Health shocks, like self-reported big health changes, accidents, or newly diagnosed health events [4,6,10,12], are health measures which are less likely to be endogenous, although the results are restricted to fairly extreme cases and cannot necessarily be extended to small changes in health.

More objective measures of health—for example, indicators for diagnosed diseases or sickness absences—have been used instead of SAH [3,7,19,21]. These may have less justification bias, but they may also have measurement errors or may not measure work-related health. This has motivated studies where one compares results when using both self-reported and more objective measures for health. These studies do not give unambiguous results, but some studies [8] find that subjective health measures overstate the effect of health on retirement. On the other hand, the alternative health measures could be used as instrumental variables for SAH [22]. This is based on the idea that two erroneous measures can be used as instruments for one another if the measurement errors are not correlated [23]. In addition, modeling of the response behavior to health questions has been examined as a way of taking into account the variation in scales between respondents [9].

Finally, the use of selection models to deal with sample selection is possible, although it may be difficult to find variables that can be argued to affect selection into the sample (retirement observed) but not the timing of retirement. An alternative is to treat retirements after the observation period as censored observations.

There are several studies where the connection between retirement expectations and actual retirement has been examined. These studies use the terms retirement “expectations”, “anticipations”, “intentions”, “preferences”, and “plans”. Sometimes these are used as synonyms or the wording comes from the survey that is used. The term that is used may also reflect the process of retirement. In organizational psychology it has been suggested that retirement is a process with steps from retirement intentions to decision to retire and from the decision to actual retirement [24]. Further, it has been argued that retirement considerations, preferences, and decisions represent different levels of firmness of retirement intentions [16].

In economics the emphasis has been on the properties of retirement expectations in relation to actual retirement and on how shocks or new information affect expectations and/or decisions [25]. Some studies that have used both expectations and realizations of retirement include comparisons of actual and expected retirement dates or ages as continuous variables. Bernheim [26], who pioneered this kind of analysis, tested the rationality of retirement expectations by regressing realized years of retirement on expected years. Rationality was clearly rejected (as also in Haider and Stephens [27]), and his tabulation of actual retirement times (years) by fixed expected times showed that the distributions were skewed and that the modes corresponded best to the expected times. He therefore concluded that retirement expectations are “most probable retirement ages” rather than the means of a distribution. This has also been confirmed in other studies [13,14]. Bernheim [26] also estimated models where expectation error (actual minus expected retirement year) was explained by the characteristics of the individuals, including health. Individuals with better than average health had lower expectation errors.

There are also studies where the purpose has not been to examine the discrepancy of retirement expectations and realizations quantitatively but by coding actual retirement as being before, at, or after expected retirement age. This trichotomous variable has then been explained by the characteristics of the individual, including health, as explanatory variables [13,14,16]. Sometimes multinomial models are used for different combinations of expectations and realizations of retiring or continuing to work [15].

The studies have concluded that poor self-rated health (at the time of the survey) was positively associated both with retiring at an earlier age than preferred and with retiring earlier than decided [16], deterioration of health between the survey of expectations and actual retirement had a positive effect on the probability of retiring earlier than expected [13,14], and those who planned to retire and did so were in worse health, and negative health shocks made realized retirement timing differ from the anticipated one [15].

One variant of this kind of analysis is the estimation of a duration model for actual retirement time with categorical indicators of expected retirement as covariates [17]. Even less formal comparisons of actual and expected retirement are studies where the outcome is a binary indicator for (actual) early retirement and where one of the covariates is an indicator for having early retirement intentions [28,29] or where there are separate estimations for the probabilities of expecting and actually retiring after a certain age [30]. The general result from this kind of study is that indicators for having retirement intentions are positively associated with subsequent retirement.

Our data period covers the years 2003–2013. There was a major pension reform in 2005, so we describe both the pre-reform system and the post-reform system. In the old system, the official retirement age in the private sector was 65 years. Most public sector (state or municipality) employees and some special groups had a lower retirement age, often 60 years or even lower. The pension accrual rate for the private sector employees was 1.5% of annual income earned before the start of the year when the person became 60 years old, and the rate was 2.5% after that. In the public sector, the accrual rate for new employees was the same as in the private sector, but somewhat higher for employees with longer tenures.

However, even in the private sector, actual retirement ages were much lower than 65 due to different early exit routes. The routes of early exit from the labor market were related to the following: (a) early old-age pension, (b) disability, (c) unemployment, and (d) reduction of working hours. The early old-age pension meant the possibility to retire at ages 60–64, with a somewhat lowered pension. On the other hand, it was possible to delay retirement from 65 with a somewhat increased pension. The disability pension was connected to reduced work ability and required medical verification. With no particular age limit, it concerned large population groups, but its incidence decreased during the past decades. Also, part-time disability retirement was possible. The early exit route related to unemployment was a scheme with extended unemployment-linked benefits for older workers who lose their job to bridge the time until they receive the old-age pension. The lower age limit for this so-called unemployment pension tunnel was 55 years. The part-time pension was the third form of early retirement, which had the aim of bringing flexibility to the later years of working life. It could be granted to an employed person, 58 years or older, who changes from full-time to part-time work.

In 2005 a pension reform was carried out, with the purpose of increasing the retirement age. The main changes in the system concerned the mandatory retirement age and the pension accrual rate. After the pension reform, old-age retirement became flexible between ages 63 and 68 years. The private and public retirement ages were harmonized for new employees, but incumbent public sector employees were given a personal retirement age; that is, an age when they receive full public pension. These personal retirement ages increased gradually, based on age and the length of tenure, from the previous retirement ages to the new ones.

After the reform, the new pension accrual rate was 1.5% until age 52, 1.9% for ages 53–62, and 4.5% for ages 63–68. The purpose of the “super accrual” from age 63 was to improve incentives to continue working longer. As to the early exit routes, after the reform, early old-age retirement with lower pension was available only for 62-year-olds. The full- and part-time disability pension and part-time pension remained the same as before and the lower age limit of the unemployment pension tunnel was gradually increased. In practice, the pension reform led to an increase in the average retirement age, but also to the concentration of retirements to certain ages. After our data period, a new reform took place in 2017, which concerned the mandatory retirement age, pension accrual, and part-time pension, but full- and part-time disability pension remained the same as before. Therefore, the current system partly differs from that described above.

Figure 1 shows the distributions of expected and actual retirement ages and the expectation error by SAH status for the sample where retirement ages are observed. The error is defined as Expectation error = Actual retirement age – Expected retirement age. Because of the small number of observations in the lowest health category (SAH = 1), categories 1 and 2 have been combined in the figure. The distribution of expected retirement ages is concentrated on values 60, 63, and 65, which are sort of “focal points.” Those with the poorest health expect to retire earlier, and among them 60 years is the mode rather than 63 years as among those with a better health status. Explanations for the concentration of retirement at certain ages include social norms, certain ages as default options, and reference dependence (planned retirement income as a reference point) [33]. In the Finnish case, age 63 is understandable as the lowest retirement age in the post-2005 pension system and 65 as the “traditional” official retirement age in the old system. Age 60 may be related to some public sector pensions or just to the use of a round number as an approximate retirement age.

Actual retirement ages are concentrated in Figure 1 on the range 63–65, with a clear mode at 63 years, except for those with lowest health status. In our data the argument of expectations as “most probable retirement ages” [26] was not fully supported, as the modes were not systematically closer to the expected retirement ages than the means. This may be related to the relatively small number of observations for some expected retirement times and the high concentration of retirements at age 63. The distribution of expectation errors is skewed: people are more likely to work longer than they expect, but the mode is still at zero error. Those with poor health expect to retire early and often do, which means that both expected and actual retirement ages are low. However, this does not mean that there is no room for making errors. In fact, the variances of both retirement ages and expectation errors are negatively related to SAH (Panel A of Table II). One interpretation of this is that for people with low SAH the timing of retirement is more unpredictable.

Table II presents the unconditional connection of health and retirement in slightly aggregated form (but in our statistical analysis we treat the variables as continuous). The expectation error is classified into three groups: retirement before expected age (negative error), retirement at expected age (zero error), and retirement after expected age (positive error). Again, SAH categories 1 and 2 are combined. Panel A shows the connection of the error with health in the sample where censored retirement times are excluded. Among those with the poorest health, 46% retire later than they had expected, and 23% earlier than expected. Of those with the best SAH status, 45% work longer than they had anticipated, and 18% retire earlier than they had expected. The differences in working longer than expected between the health statuses are, therefore, relatively small. The main difference between the health levels comes from those with poor health retiring earlier than expected and, therefore, more often making negative expectation errors.

In Panel B of Table II, we include into the group “After expected age” also those who have not yet retired but who have passed their expected retirement age. Panel B also shows the share of those who did not reach their expected retirement age by the end of the follow-up period and had not yet retired. For them we cannot tell when retirement happens compared to the expected retirement age. This group is clearly more common among those with better health. Comparison of the two panels shows that there may be a selection problem in the health–retirement relationship because retirement age is more likely to be observed for those with poor health.

The table shows that it is most difficult for persons with poor health (SAH 1 to 3) to estimate their timing of retirement. This group includes both persons who retire earlier than they expected and have negative errors and also persons who underestimate their future ability or willingness to continue to work and, hence, make positive forecast errors. The discrepancy between the expectations and realizations may also be related to being in a position where one has to continue working for financial reasons or because of a rejected disability pension application. Positive forecast errors are common also in the other health groups, but perhaps for different reasons. If good health is positively correlated with socioeconomic position, those with good health may have more opportunities to choose the timing of retirement, and they may also have more enjoyable working conditions.

We have estimated regression models where actual retirement age, expected retirement age, and expectation error are regressed on SAH and the controls. Since the same variables are used in all of the models, the estimates of the expectation error model are simply the differences between the coefficients of the actual and expected retirement age models. The baseline results are shown in Table III.

The estimates concerning the health variables are collected in Table IV. The estimates are presented there so that they tell the connection of one standard deviation change in a health variable with months of retirement age, expected retirement age, and expectation error. The baseline results in the first row of Panel A of Table IV (based on the first line of Table III) show that health is clearly correlated with both actual and expected retirement ages, when these are regressed on health and controls. Health has a stronger connection with expected retirement age, which produces a negative connection between health and expectation error. In terms of months, the results imply that one standard deviation improvement in health (0.87) is associated with two months later retirement, 3.6 months later expected retirement, and 1.6 months smaller expectation errors. However, as the descriptive evidence in the previous section showed, these results may have been influenced by sample selection, which leaves out many of those with good health who are working longer than they expected.

Among the other variables (reported in Table III), age is positively related to expected and actual retirement age. Gender, education, and earnings have no effect on the results. The pension accrual rate is positively related to expected retirement age, which supports the incentive role of the age-related pension accrual. Those who are already in partial retirement or whose spouse is retired tend to have lower actual retirement ages, but these variables have a weak relationship to expected retirement age. Public sector pensions are positively related to both actual and expected retirement age. There is evidence that in the smallest plants (reference group under 10 employees) employees expect to retire later, but the relationship of plant size to actual retirement age is not as clear. Finally, the indicator for the 2008 survey has a negative and significant coefficient for actual retirement age. This may be related to the pension reform, which has made the lower limit 63 years a clear focal point in retirement behavior. This is so even though average retirement age has increased partly through restrictions on some of the early exit routes (average observed retirement age in our data is 62.8 for those in the 2003 survey and 63.2 for those in the 2008 survey). The focal point has been observed also in other studies [34].

As discussed above, there are potential sources of bias. We give attention to biases arising from measurement error, endogeneity, and sample selection. The results of the health variables are presented in different panels of Table IV.

First, we address the issue of measurement error in health. It can be argued that the indicators for chronic illness and pains and the number of days of sickness absence are more objective measures of health even when they are self-reported. Therefore, they themselves may contain some measurement error. We first use them as health variables instead of SAH. We use both self-reported absences in the past 12 months and register information on absences for which sickness allowance has been paid during the calendar year preceding the survey. Since the variables are measured differently (SAH from 1 to 5, pains and chronic illness as binary indicators, and absences as days), we report the results in terms of one standard deviation change in each health variable. Rows 2–4 of Panel A in Table IV summarize the coefficients of the health variables scaled in this way and are expressed in months. Naturally, SAH has a sign which is different from the signs of the other health variables, as a higher value of SAH means better health, whereas the reverse is true for the other variables. Generally, SAH and sickness absences as explanatory variables give fairly similar results, whereas the other health variables give lower values for the relationships. The main difference between self-reported and register-based sickness absences is that the former has a stronger relationship with the expected retirement age. The result that SAH overstates the relationship compared to objective health measures is in line with the results of some earlier studies [8]. On the other hand, the number of sickness absence days may be a better summary measure of health than the binary indicators, and it gives results similar to those with SAH.

An alternative way to proceed is to use these other health measures as instruments for SAH. This is the so-called multiple indicator solution [23]. The true health status is unobserved, and when SAH is used as a proxy for health the measurement error in SAH is part of the error term of the model. Assume that there are two health indicators that have uncorrelated measurement errors, which are ignorable in the sense that if true health could be controlled, the indicators would have no explanatory power for retirement. Under these assumptions, one health indicator is a valid instrument for another. Panel B of Table IV shows the associations of one standard deviation change in SAH with the retirement variables when indicators for pains and chronic illness are used as alternative instruments. The point estimates of the coefficient of health increase (in absolute value) in instrumental variable estimation, which is consistent with attenuation caused by measurement error. Using sickness absences (either self-reported or register-based) as an instrument produced very highest estimates in absolute value (not reported in the table).

The under-identification and weak identification tests showed that the instruments are relevant; that is, sufficiently correlated with SAH. When at least two instruments are used, overidentification can be tested; that is, that the set of instruments is not correlated with the error term. We tried using different pairs of two instruments at the same time. Hansen’s J test led to the conclusion that the instruments are valid in all of the three equations only if pains and chronic illness are used together as instruments. In this case, the estimate for the relationship of one standard deviation improvement in health with actual retirement age is 3.5 months, and with expected retirement age 3.8 months. When the number of absences was used together with either of the other instruments, the J-statistic was high in the equation for expected retirement age, and the null hypothesis of no correlation with the errors was rejected. Sickness absence seems to be a problematic variable as an instrument and it may be correlated with the measurement error in SAH, as both are summary measures of health.

Second, we have examined the role of health shocks. They are health-related events in the follow-up period that are more exogenous because they happen after the survey. Health measured in the survey may not play a big role if there are later severe health events. We measure the health shock by the average number of annual sickness absence days between the survey year and retirement, using register information on absences for which the Social Insurance Institution has paid sickness allowance. Use of absences may be problematic, since there may be latent health symptoms that affect both SAH and later sickness absences. To reduce this problem, we concentrate on initially healthy individuals. These are defined as those who have not had absences with sickness allowance during a two-year period (survey year or the year before), have answered in the survey that they have no chronic illnesses, and have a good health status (SAH equal to 4 or 5). These restrictions leave 466 observations, among which the average number of annual sickness absence days in the follow-up period has mean 4.96 and standard deviation 14.54. The correlation of the shock measure with SAH is negative and low in absolute value (–0.05), so the problem of latent health effects should not be serious. Even this measure is subject to justification bias, if people who would like to retire early, for example because of dissatisfaction with their work, take sickness absences to justify the need for early retirement even when their health is not poor. However, obtaining sickness allowance requires a medical certificate, and the allowance is paid only after 10 days of sickness. Therefore, just low job satisfaction without any observed medical symptoms is not sufficient.

Figure 2 shows the unconditional relationship between the shock and actual retirement age. Panel C of Table IV shows that in the retirement age model the health shock is significant with a negative coefficient. An adverse health shock, measured as a one standard deviation increase in absences, is associated with 2.7 months earlier retirement. This is probably closest to a causal effect that we can get with our data. The estimate is close to the ordinary least squares (OLS) estimates, with absences as the health variable. One interpretation of this result is that the endogeneity of health is not a serious problem, at least when the number of absence days is the health measure. The shock is naturally also negatively related to the expectation error; that is, adverse health shocks lead to earlier retirement than expected. However, the future shocks should not be related to the (earlier) expectations. This is confirmed by the estimations, which supports the assumption that the health shock measure is not related to latent health that is likely to influence retirement expectations.

Third, we have investigated the sample selection issue. We expect that there is a selection bias toward zero in the coefficient of health in the retirement age equation when only observed retirement ages are used, since those with better health are overrepresented among those who have not yet retired during the observation period. For the expected retirement age there is no selection effect, as it is observed for all. The individuals with unobserved retirement times can be treated as censored observations. In practice, the censoring limit—that is, the highest age at which a non-retired individual is observed—is age at the end of the follow-up period in 2013. The limit, therefore, varies across individuals. We estimate the censored regression model using interval regression [23]. When the retirement age is observed, we have point data (actual retirement age), whereas for the censored observations we know that retirement happens in the interval from age at the end of 2013 to infinity.

The results in Panel D of Table IV show that censored regression indeed produces higher estimates. One standard deviation change in SAH is now associated with 3.2 months later retirement. As the expected retirement age is observed for all, the censored regression naturally produces the same estimates for the expected retirement age equation as would OLS. The difference in this case between Panels A and D is caused by the larger sample in Panel D. The estimate for health in the expectation errors equation is no longer the simple difference between the estimates in the retirement age and expected retirement age equations, since the models are nonlinear. When censoring is taken into account, health is no longer significantly related to the errors. The negative relationship between SAH and expectation errors in the baseline OLS estimates seems to be caused by more common censoring of retirement ages among healthy individuals.

Finally, as another estimation, where censoring is taken into account, we have used interval regression using information on the age intervals when retirements happen, rather than retirement age as a continuous variable. Using ordered models (ordered probit or logit) for retiring before, at, or after expected retirement age would not take into account the fact that individuals are observed at different ages and at different distances from their expected or actual retirement. In interval regression we can take this information into account. Specifically, retirement can happen in the following cases: in the interval between age at time of survey and expected retirement age; at expected retirement age (point data); in the interval between expected retirement age and age at the end of the follow-up in 2013; or retirement age can be right-censored. This estimation, therefore, combines information on both expected and actual retirement. Since the values for the limits of the intervals are used, the estimation produces association of health with actual retirement age. According to the results (Panel E of Table IV), one standard deviation increase in SAH is associated with 3.6 months later retirement. Since the intervals are based on expected retirement ages, the estimation does not produce results on expected retirement ages or expectation errors.

We have carried out various robustness analyses, which we briefly describe without presenting the results in a table. So far, we have used SAH as a cardinal measure. An alternative is to use indicators for different levels of health. We used indicators for SAH levels 3, 4, and 5, with levels 1 and 2 combined as the reference. The coefficients of the indicators had an increasing pattern in the models for actual and expected retirement age and a decreasing (negative) one in the model for expectation error. Cardinality implies certain restrictions on the coefficients of the indicators: coefficients of adjacent indicators increase by the same amount and there are multiplicity constraints; for example, the coefficient of the fourth indicator is twice the coefficient of the second one. These constraints were accepted in the estimations with the indicators for health levels. Therefore, the use of a cardinal health measure is justified.

To investigate the heterogeneity of the health–retirement relationships, we tried including interactions of the health variable with education (indicators for secondary and tertiary education), gender, and income. The only statistically significant result from this exercise was that the interaction of the indicator for tertiary education and health was positive and significant in the model for actual retirement age. In the expected retirement age equation, even this interaction was insignificant. However, in the censored regression model none of the interactions were significant.

We have examined the sensitivity of the results to the choice of the sample. Since there may be disability retirements among those below age 60, we tried including in the estimations also those in the 2008 survey who were 50 to 54 years old at the end of 2008 and who, therefore, did not reach age 60 by the end of the follow-up. The baseline results were close to those reported above, as one standard deviation change in SAH was associated with 2.4 months later retirement. The similarity of the results is natural, since only 19 additional observations with observed retirement age were included. However, at the same time the extension of the sample added 547 new censored observations. Censored regression with this larger sample resulted in a somewhat higher estimate for SAH than in the baseline model. However, it is difficult to tell whether this was due to the larger sample size or the higher share of censored observations. As another robustness check we estimated the model using only the 2003 survey. Censoring should be a smaller problem in this group. On the other hand, the number of observations in the OLS estimations dropped to 792. The estimate for the health effect dropped and became insignificant, whereas the relationship between SAH and expected retirement age remained almost the same as with the larger sample. In the censored regression model, the health effect on retirement age was again significant. A one standard deviation change in SAH was associated with three months later retirement. This is close to what we obtained in the baseline estimations. Again, it is hard to tell whether the differences in some of the results are due to less censoring or smaller sample size.

A limitation of our analysis is that we deal with full-time retirement only. This restriction is dictated by the survey, which does not ask about expected part-time retirement age. We can still speculate how part-time retirement might affect the results. Had the survey respondents been asked about both expected part-time and full-time retirement age, it is possible that we would find a stronger relationship between health and expected retirement, as those with poor health would be more likely to plan to take part-time retirement. For actual retirements we can examine how part-time retirement affects the relationship between health and retirement. To do this we dropped from the data those who were already in part-time retirement during the surveys, added those who were observed to retire part-time but not full-time, and defined the actual retirement age according to part-time retirement for those for whom both part-time and later full-time retirement were observed. The analysis of our baseline model with this sample showed that one standard deviation increase in SAH was associated with 2.4 months later retirement, which indicates 0.4 months stronger association than in panel A of Table 2. It is difficult to say how part-time retirement would affect the connection of expectation errors and health if both expected and actual retirement age were affected. Including more flexible retirement channels would possibly lead to more negative expectation errors (retirement earlier than expected) for those with poor health.