Mental Health Effects of Retirement

Abstract

We study the retirement effects on mental health using a fuzzy regression discontinuity design based on the eligibility age to the state pension in the Netherlands. We find that the mental effects are heterogeneous by gender and marital status. Retirement of partnered men positively affects mental health of both themselves and their partners. Partnered female retirement has hardly any effect on their own mental health or the mental health of their partners. Single persons retirement does not seem to have an effect on their mental health status.

Appendices

Appendix 1: Details on Our Data and Sample

1.1 Sample Selection

Between 5072 and 6698 individuals were interviewed each year for the core study on health from 2007 until 2018, resulting in a total of 63,603 records. We matched each record on the basis of the information on the year and month of interview to the corresponding information about the retirement status and age in months coming from the background variables. We were able to match 62,502 records, belonging to 12,958 different individuals.¹² Given the aim of this paper, we restricted the sample to individuals close to the moment of the state pension eligibility. After defining according to the rules outlined in Sect. 2 a variable that measures the distance in months from the month in which an individual becomes eligible to the state pension, we kept all the observations who were within 84 months away from the month of the state pension eligibility at the moment of the interview. The sample size shrank therefore to 16,551 observations. Since the aim is, not only to unveil the effect of retirement on his/her-own mental health, but also to identify the impact on partner’s outcome variables, we restricted the sample to couples in which both partners answered the questionnaire on health (3497 couples) and to singles (4339 observations).

Finally, we dropped from the samples 128 couples for which at least one partner is interviewed in the month in which the eligibility to the state pension was attained, or in the previous or in the subsequent month. Similarly, we eliminated 79 singles. This refinement is due to a kind of heaping problem (Barreca et al. 2016) or rounding error (Dong 2015). From 1 January 2012, the state pension eligibility is indeed received from the day in which one satisfies the age requirement. Since we do not have the day of birth, but only the month in which an individual becomes eligible to the state pension, we cannot be sure, for those interviewed in the month in which they become eligible to the state pension, whether they are already eligible at the moment of the interview or they will be soon eligible to the state pension. Although this kind of error is likely to be randomly distributed across those observations interviewed in the month of state pension eligibility, it is present only above the cutoff (Lee and Card 2008). Given the small number of such observations, omitting them from the sample is the easiest way of facing the problem and getting unbiased estimates of the treatment effect for all the others (Barreca et al. 2016). The remaining sample has 3369 records of couples and 4260 singles.

Along the paper, we use different subsamples, depending on the chosen bandwidth. For example, when the chosen bandwidth is 42 (for couples satisfied by both partners), the sample is made up of 1250 couples and 2120 singles (861 men and 1259 women).

1.2 Definition of Variables and Descriptive Statistics

Table 5 clarifies the discrete nature of the outcome variables and the meaning attached to the numeric values. Information on happiness, calmness, depression, anxiety and feeling down in the last month is collected by asking individuals whether in the last month they felt, respectively, “happy”, “calm and peaceful”, “depressed and gloomy”, “very anxious” and “so down that nothing could cheer me up”. They could choose one the following six: (1) never, (2) seldom, (3) sometimes, (4) often, (5) mostly and (6) continuously. Since the top or the bottom categories were sometimes indicated rarely by respondents, we had to group them as detailed in the last column of Table 5. These five variables measuring mental health, with their original scores from 1 to 6, are the ones used to build the MHI-5 scale.¹³ We summed up the scores of the five mental health variables after reversing depression, anxiety and feeling down, subtracted 5 and multiplied by 4. By doing so, the resulting MHI-5 scale spans from 100, for those reporting all 6s, to 0, for those reporting all 1 s. Larger numbers represent therefore better mental health conditions.

Table 5 Outcome indicators

Table 6 provides descriptive statistics of the outcome variables. Table 7 reports summary statistics of the retirement indicator, the number of months from the age of state pension eligibility and the other control variables used in the econometric analysis. Finally, Fig. 4 shows the distribution of the age difference between husbands and wives in the sample used for our baseline estimates. The age difference is predominantly positive: males are on average older than their female partners. In almost 77% of the couples the male is older than the female. For 30% of the couples the male is at least two years older.

Table 6 Descriptive statistics of the outcome variables

Table 7 Descriptive statistics of the covariates

Fig. 4

Distribution of the age difference between partners are within the bandwidth of 42 months (age man minus age woman). Note: The sample is limited to 1250 couples with both partners within the bandwidth of 42 months

Appendix 2: Validity and Falsification Tests

As suggested by McCrary (2008), a jump in the density of the running variable at the threshold would be direct evidence of the failure of the local randomization assumption. Figure 5 displays the local polynomial density estimate of the running variable described in Cattaneo et al. (2018). The graphs show that there is no evidence of discontinuity in the population density at the cutoff, for both genders and whether living in a couple or not.

Fig. 5

Graphical density test of the running variable, Notes: The solid line is the the local polynomial density estimate of the running variable described in Cattaneo et al. (2018). The local polynomial is of order 3. The robust bias-corrected test proposed in Cattaneo et al. (2018) cannot reject the null hypothesis of the absence of discontinuity: p value equal to 0.711 (0.293) for men (women) for couples; p value equal to 0.161 (0.872) for men (women) for singles

If the retirement probability is locally randomized near the cutoff, then the treatment should not have an effect on the pre-treatment covariates, i.e. the treated units should be similar to control units in terms of observed characteristics. We follow Lee and Lemieux (2010) and test if the discontinuity influences our predetermined variables, by estimating a seemingly unrelated regression (SUR) with one equation for each of the predetermined variables, with the same bandwidth (42 months) and local polynomial regression as in the baseline estimates. After the estimation of the SUR model, we performed joint and individual tests of the significance of the discontinuities. They are reported in Table 8. The single tests do not show systematic jumps at the cutoff: only 2 discontinuities out of 39 are significantly different from 0 at the usual 5% confidence level. All the joint tests do not reject the null hypothesis. Since we are testing the presence of discontinuities for many covariates, the joint tests suggest that the two significant discontinuities are so by random chance (Lee and Lemieux 2010).

Table 8 Falsification test: discontinuity of the predetermined variables at the cutoff for couples (SUR estimation, bandwidth equal to 42)