We used data from the 2010 and 2013 waves of Japan Gerontological Evaluation Study (JAGES) project. Detailed descriptions of JAGES are available elsewhere [21, 22]. In brief, the JAGES conducted postal surveys targeting physically and cognitively independent community-dwelling older adults aged ≥65 years in Japan. The baseline survey was conducted in 2010 and we sampled 112,123 older adults residing in 31 municipalities in Japan (Response rate = 66.3%). The second survey was conducted in 2013 in which we sampled 137,736 older adults residing in 30 municipalities in Japan (Response rate = 71.1%). Twenty-four municipalities participated in both years. Sixty-two thousand four hundred thirty-eight participants responded to both surveys.
The 2013 wave of JAGES dataset consisted of five modules. The question inquiring about participants’ sense of meaning in life was measured only in one of the five modules, which was only mailed to one-fifth of the total sample. We used respondents to this module (N = 12,487) to test the interaction between retirement and two types of social participation, as well as whether a sense of meaning in life mediated this relationship. All variables used in this study were self-reported. Ethical approval of this study was obtained from the ethics review board of the University of Tokyo Medical School.
Changes in depressive symptoms
Depressive symptoms were measured at both waves using the validated Japanese short version of the Geriatric Depression Scale (GDS-15)  Total scores could range from 0 to15, where higher scores indicate more depressive symptoms. GDS-15 has been commonly used as a screening tool for depression among older adults with a cutoff of 5 or above to indicate clinical depression [23, 24]. We used changes in total GDS-15 scores from 2010 to 2013 as the continuous outcome variable in this study.
Changes in working status
Information on working status (either “currently working” or “retired and not currently working”) was collected in both 2010 and 2013. Thus, no one among the subjects was considered to be partially retired. We categorized subjects as “Kept working (working/working)”, “Retired (working ➔ not working)”, “Started work (not working ➔ working)”, or “Continuously retired (not working ➔ not working)” based on patterns of changes in working status from 2010 to 2013.
Changes in social contacts and social support
We used two variables to tap changes in social contacts. First, we asked the number of friends and acquaintances the subjects met in the past month at both waves, and then, changes in the answers were modeled as a continuous variable. Second, we asked subjects “What is the relationship between you and the person you often meet?” At each wave, we focused on individuals who maintained social contacts with colleagues in the workplace. As for social support, we collected information at both waves on whether or not subjects received emotional (‘Do you have anyone who will listen to you when you have worries and complains?’) and instrumental (‘Do you have anyone who can take care of you when you are sick in bed for a few days?’) social support from their friends. Between-wave changes in perception of social supports from friends were made into a categorical variable.
The occupation that subjects had engaged in for the longest period of time was asked at both waves. Since it is unlikely that the longest job of older people aged 65 years or older change within 3 years (from 2010 to 2013), we treated this as a time-invariant variable. Professionals and managers were categorized as “higher occupational class”. Clerical support workers, service and sales workers, craft and related trades workers, skilled agricultural, forestry, and fishery workers, and others were categorized as “lower occupational class”.
Social participation/a sense of meaning in life
Participation in different clubs/groups was asked at both waves. We defined “social participation with roles” as participation in volunteering clubs or neighborhood councils. We defined “recreational social participation” as participation in sports organizations, hobby clubs, or older adults clubs. A sense of meaning in life was measured in 2013 by asking “Do you feel a sense of meaning in life” (yes/no).
Changes in equivalised household income, marital status, instrumental activities of daily living, incidence of serious illnesses and family caregiving were treated as time-varying covariates. These variables are potential confounding factors, which could have influence on both retirement decision and mental health. Household income at each wave was equivalised in order to adjust for the number of members within households. Marital status of subjects at each wave was defined either “married” or “non-married (including being widowed, divorced, or single)”. Subjects’ instrumental activities of daily living (IADL) were measured using the Tokyo Metropolitan Institute of Gerontology (TMIG) Index of Competence as a broad indicator of physical health. TMIG Index of Competence was validated using the data of Japanese older community residents  and its total scores could range from 0 to13, where higher scores indicates more independency in daily livings. Incidence of serious illnesses and need for caregiving was assessed in 2013 by administering a checklist for events occurring in the past year.
As for missing values in some variables used in the analysis, we executed multiple imputation by chained equation method to create 100 data sets without missing value. We used the FCS statement of the MI procedure in SAS version 9.3 to create imputed data sets. Then, we excluded those who indicated depression in 2010 (GDS score of 5 or more as described below), those who never worked, and those who answered inconsistently to the question about the occupational class at the two time points. We ran all models using each data set and, using MIANALYZE procedure, we obtained the final estimates combining 100 estimates derived from 100 analyses.
We used first-difference models stratified by gender to investigate the main effects of changes in working status on changes in depressive symptoms with full-sample data (Model 1). The first-difference models tool the changes in all variables in the models, which enable to can control for all observed and unobserved time-invariant confounders (e.g. age and education) by cancelling out the estimates of time-invariant confounders. In model 2, we added variables regarding changes in social contacts and social support to test whether these changes could explain the association between changes in working status and changes in depressive symptoms. In model 3, we added an interaction term between changes in working status and occupational class. In model 4, we added an interaction term between transition to retirement and changes in marital status.
In the analysis involving a sense of meaning in life, we analyzed only one of the five sub-samples of JAGES data because it is the only module that contains information on participants’ sense of purpose in life. In model 5, we used the first-difference model again to test the main effect of changes in working status on depressive symptoms among this sub-population. In the model 6, we added the interaction term between transition to retirement and two types of social participation. Lastly, in model 7, we added sense of meaning in life as a covariate to check whether it mediated the interaction between retirement and social participation.
We stratified data by gender instead of modeling gender interaction effects to make it easy to interpret our results and prevent the problems of multicollinearity on interaction terms.