Acknowledging the limitations of previous research, we used panel data from the Survey of Health, Ageing and Retirement in Europe (SHARE) to investigate the mental health effects surrounding retirement among the European population. The FE models were employed to control for unobserved time-invariant confounding factors, and instrumental variables (IVs) were included to reduce potential biases from the endogeneity problem. Moreover, the reasons for retirement were included as explanatory variables. Notably, instrumental variables representing “predictive retirement” and “retirement in the past” allowed us to examine mental health effects before and after actual retirement, respectively, which distinguishes this study from others on this topic.

An individual’s mental health outcome was represented by the variable EURO-D, developed by the EURODEP Concerted Action Programme [52]. EURO-D has been used in many studies investigating the mental health of the European population [6, 30, 53,54,55]. EURO-D uses a 12-item scale to measure where an individual is positioned on a range of being depressed or not depressed. The interviews used to obtain the EURO-D values were conducted in the local language and included questions regarding depression, pessimism regarding the future, suicidal feelings, guilt, sleeping difficulties, levels of interest, fatigue, irritability, concentration, appetite changes, and sadness and enjoyment. Every “Yes” answer to the questions was coded “1” and every “No” was “0.” The scores were summed for each respondent, and the resulting EURO-D score ranged from 0 (the respondent is not depressed at all) to 12 (the respondent is very depressed) [56, 57]. The imputation technique was applied for the EURO-D variable where observations were missing (see details in section A).

The central explanatory variable in this analysis was the retirement status, which took the value of “1” for retirees and “0” for employed and unemployed people. According to [12], retirement status has three different definitions. First, nonretirees are employed, unemployed but looking for work, homemakers, or permanently ill or disabled. The second definition includes homemakers and permanently ill or disabled persons with retirees if they report no paid work during the previous month, and the third includes only those who report being either retired or employed. However, from the literature, Heller-Sahlgren’s third definition is the most common approach [1], which has been used in various studies (e.g., [1, 6, 13, 18]).

Using this definition is necessary for a meaningful inference. The current job situation according to SHARE data includes six groups: (1) retired, (2) employed or self-employed, (3) unemployed, (4) permanently sick or disabled, (5) homemaker, and (6) other. We excluded the unemployed, sick/disabled, homemakers, and others because comparing retirees with the employed would be coherent in terms of the discussion on the mechanism impacts of retirement, such as more available leisure time, reduction in work-related stress, reduced contact with former colleagues, and losing sense of self-worth (Table 1).

The lead retirement variable (retirement in the next wave(s)) took the retirement variable value in the next wave(s). For example, a person who would retired in the next wave had the “lead retirement 1” variable’s value of 1; a person retiring in the next 2 waves had the “lead retirement 2” variable’s value of 1. Conversely, the lag variable (retirement in the previous wave(s)) took the retirement value in the previous wave(s). Assuming that people did not return to work after retiring, missing variables were recorded as “1” if they retired in the last wave, and “0” if they did not retire in the next wave. A typical observation is illustrated in Table 2.

Figure 1 illustrates the depression levels of retirees and non-retirees by age. Figure 1 shows that the average levels of mental ill-health over time develop a U- shaped pattern, with the middle-aged and very old individuals having higher levels of depression. The dip in retirees’ depression was located at 63–64 years, whereas that of the employed was at 65–67 years. The mean eligible pension age for an individual in our sample was 64.4 years; Fig. 1 suggests that people, regardless of their retirement status, might feel less depressed when they approached retirement age but eventually became unhappy when they got older. In addition, retired people seemed to have more mental health problems than the employed. However, a relation between retirement and mental health cannot be posited because old age and retirement status are strongly related to each other.

Observed depression level by age quantile

Ten reasons for retirement are listed in the SHARE data. The interviewees were asked whether they retired because they (1) became eligible for a public pension, (2) became eligible for a private occupational pension, (3) became eligible for a private pension, (4) received an offer of early retirement with special incentives, (5) lost their job due to redundancy (layoff), (6) were in poor health, (7) needed to care for an ill relative(s)/friend(s), (8) wanted to retire at the same time as their partner or spouse, (9) wanted to spend more time with their family, or (10) wanted to enjoy their life. However, we did not include each reason individually because there seem to be similar motivations across some reasons. Instead, we followed Robinson, who used 3 main reasons for retirement and divided the 10 SHARE reasons into 3 groups [58]. Reasons 1–4 formed the “positive circumstances” group; reasons 5–7 were included in the “negative circumstances” group; and reasons 8–10 were categorized as the “aspirational motivations” group. We classified respondents whose answers fell into more than one category as retiring due to “ambiguous reasons.”

Instrumental variables should meet two conditions: they are related to the explanatory variable and orthogonal to the exogeneity condition [3, 12]. Namely, retirement is the only “channel” through which the instrumental variables chosen in this study could alter the outcomes. Specifically, we followed studies that employed comparable models and chose the dummy of being past the eligible pension age and the dummy of being over the early pension age as the instruments [7, 30, 37].

Control variables consisted of demographic background, activities, and survey dummies. First, the sociodemographic variables used were age and marital status [6, 7, 30, 42], number of children [30], and the country’s GDP and unemployment rate, which represent the countries’ economic background [59]. Adopting Zon and Butterworth’s method, marital status was divided into “married” and “not married” [1, 19]. The former category included people who were married and living with or without a spouse as well as those who had registered partnerships; the latter included people who were never married and those who were divorced or widowed. Additionally, because the act of taking care of other individuals may affect mental health, we considered the number of children as a variable in the models, regardless of whether they were living alone or with a spouse/partner. Second, activities were variables that represented the frequency of engaging in vigorous activities, such as sports or heavy housework, as well as activities that required moderate energy levels, such as gardening, car washing, or walking [7, 30]. Third, we used the dummies of SHARE’s waves to control for the time effects.

The summary statistics for the main variables are shown in Tables 3 and 4. The age of individuals in our data set ranged from 39 to 88 years, with an average of 65.07 years. The share of both genders was almost equal, 50.6% female and 49.4% male. Most of them were married; each person had an average of more than two children regardless of their marital status. For the frequency of engaging in physical activities, 73.2% of the people reported doing activities requiring moderate energy more than once a week. However, for heavy activities, they either infrequently did (more than once a week, 38.8%) or hardly ever or never did (36.5%) them. Non-retired individuals reported engaging in both kinds of activities more frequently.

This research used data from the SHARE (see [57]), which is a longitudinal, multidisciplinary, and cross-national survey that collects data on the health and socioeconomic status of noninstitutionalized people aged over 50 years in 21 European countries and Israel, along with their social and family networks.

We extracted data obtained from Waves 1–7 of SHARE interviews and created panel data covering the 2004–2017 period. The Wave 3 and Wave 7 questionnaires contain SHARELIFE modules that focus on people’s life histories, including all the important aspects of respondents’ lives, but only Wave 7 has a regular panel questionnaire for all interviewees who previously answered SHARELIFE interview questions. Accordingly, we excluded Wave 3 because retrospective information was not considered in this study. The final unbalanced panel contains 182,142 observations from 6 survey rounds.

Being a large household survey, SHARE suffers from nonresponse issues [60], especially missing values [61]. SHARE release 7.0.0 provides five multiple imputations for the missing values of each variable [62]. Although there were few missing values of the variables used in our models (less than 5%), we accounted for them by using all five imputations from SHARE. Thus, all the variables in the analytical dataset came with five imputed values for each missing value. If a value is non-missing, the remaining associated imputed variables would have the same value as the base value. These variables in our data were imputed by SHARE’s hot-deck method. However, the estimated results were not much different from those of the dataset using single imputation or list-wise deletion (Details in Appendix 1).

As we exploited the longitudinal dimension of the SHARE database, the role of nonrandom attrition was concerning. Following Verbeek and Jones (see [63, 64]), we conducted two tests for attrition bias. First, we ran a regression of the depression score on an indicator counting the number of waves that an individual appeared in the panel. Second, we regressed the depression score on another indicator of whether an individual appeared in the next wave. Both regressions employed the pooled and random effect models with the unbalanced full sample, and the statistical significance of the two new indicators provided a test for nonresponse bias. In our study, all indicators’ coefficients were insignificant, thereby suggesting that there was no attrition bias (see Table 10 in Appendix 2). No matter how many times a respondent appeared in the panel, the EURO-D score was not systematically different. Other studies using the SHARE database have shown that nonrandom attrition was not an issue for the cognitive ability [65] or mental health represented by EURO-D [7].

To investigate the possible correlation between retiring and one’s mental health, FE models were employed as shown in Eq. (1):

$$M_= \alpha +\beta Retir_+\delta }_+_+ _$$

(1)

where MHit and Retireit denote a measure of mental health and retirement st0a1tus, respectively, of individual i at time t. \(}_\) is a combination of control variables that represent the individual’s demographic background (age, marital status, and number of children) [1, 6, 7, 19, 30, 42] and factors that could affect the well-being of individuals, such as limitations regarding daily activities and the frequency of playing sports [7, 30]. Finally, ui is the unobserved time-invariant heterogeneity with individual fixed effects, and ϵit represents distinctive error terms.

Coefficient β in the FE models is estimated under the assumption that ϵit is uncorrelated with Retireit (the retirement decision). However, many researchers believe that this condition is easily violated by the presence of omitted variables and potential reverse causality, thereby causing endogeneity biases [3, 44]. Therefore, following [3, 7], and [12], we applied the FE-IV estimator to control for time-variant unobservable factors and reverse causal impacts.

$$Retir_=_IVearl_+_IVnorma_+\delta }_+_+ _$$

(2)

Equation (2) is the first stage of the FE-IV models, where IV earlyit and IV normalit are instruments for Retireit. Each instrument is defined as Instrumentit = \(I(Ag_\ge Ag_^)\) where I is the indicator function, Ageit is the age of individual i at time t and Agep is the country- and sex-specific pension age. We used both early and standard pension ages for each country. I takes the value “1” if the condition is true, and “0” otherwise. The second stage in the FE-IV estimation is shown in Eq. (3), which is similar to Eq. (1), wherein \(\widehat_}\) is the predicted retirement status from the first stage function.

$$M_= \alpha +\beta \widehat_}+\delta }_+_+ _$$

(3)

The coefficient β in Eq. (3) represents the average effect of retirement on mental health in the year of the survey. The results of this model are presented in the last two columns of Table 5. This impact may include effects from the current and past retirement. Therefore, we separated the impact of retirement in the past from the impact of current retirement by adding three lags of \(Retir_\) to Eq. (3), indicating whether the individual retired in previous waves.

$$M_=\,\alpha +\beta \widehat_}+_}_+ _}_+ _}_+\delta }_+_+ _$$

(4)

Similar to Eq. (3), the lags of \(\widehat_}\) are instrumented by corresponding lags of early and normal IVs. For example, \(}_\) is intrumented by \(IVnorma_\) and \(IVearl_.\) Thereby, the \(\gamma\) coefficients show the impacts of past retirement on current mental health. In other words, they capture mental health impacts after retirement. Because the retirement event is predictable, the impact of the decision on mental health may not coincide with the exact retirement [18, 42]. Mental health may improve or decline in anticipation of retirement similar to Ashenfelter’s dip [41, 45,46,47,48,49,50,51]. Our study attempted to capture this potential effect by determining whether the level of mental health changed among those who knew they would retire in 2 years, 4 years, or 6 years compared with others in the workforce. The model is as follows:

$$M_=\,\alpha +\beta \widehat_}+_}_+ _}_+_}_+\delta }_+_+ _$$

(5)

where three leads of \(\widehat_}\) denote whether an individual will retire in the next two years (or the next four or six years for the longer lead time). Estimations of the lead \(\widehat_}\) are similar to those of the lag \(\widehat_}\), using corresponding leads of early and normal IVs. The λ coefficients indicate the impacts of predictive retirement on current mental health. This model is expected to reveal the effect of retirement before it happens. We call this the “Ashenfelter’s dip” of mental health. Results of pre- and post-impacts are shown in Tables 6 and 7. All statistical and econometrical investigations were carried out with Stata software, version 17.