We used data from the Aichi Gerontological Evaluation study (AGES), which is a part of Japan Gerontological Evaluation Study (JAGES). The JAGES is an on-going longitudinal panel study which seeks to elucidate the social determinants of functional decline, cognitive impairment, and mortality among older adults aged 65 years and older. Figure 1 illustrates the flow chart of the study participants for this study. One wave of data were collected for this study in 2003. Of 29,374 people invited to participate, 14,804 people completed the questionnaire, corresponding to a response rate of 50.4 %. Those who did not respond with the baseline questionnaire were younger than 80 years old and were more likely to have higher household income, while there was no difference between men and women. Of the 14,804 respondents, we included 10,271 individuals in the present analysis, because 91 individuals were excluded due to deaths or functional disability before beginning of the follow-up, 45 individuals were excluded due to missing of the linkage variable before beginning of the follow-up, a further 382 individuals did not respond to questions on age or sex, and 4015 individuals did not join any civic organizations at the beginning of the follow-up. Detailed information on the design and data collection of the AGES baseline survey is available elsewhere [15, 16].
The JAGES protocol was reviewed and approved by the ethics committee in Research of Human Subjects at Nihon Fukushi University and has been carried out in compliance with Helsinki Declaration.
All-cause mortality data until May 2008 were obtained from the six municipalities participating in the AGES cohort and were linked by the researchers using the identification number which was assigned to each study participant. One of the strength of the AGES cohort is that there is no administrative loss during approximately 5-year follow-up.
Social participation was assessed by the statement “Do you belong to the following organization or group?(yes/no)” We inquired about the following types of community organizations: neighborhood association/senior citizen club/fire-fighting team, religious group, political organization or group, industrial or trade association, volunteer group, citizen or consumer group, hobby group, and sports group or club. This questionnaire on social participation in the AGES study was derived from JGSS (Japanese General Social Survey) to make the results of the study comparable to the general population.
For those who participated in one of the above organizations or groups, the respondent’s position in the participating organization was assessed by a dichotomous (yes/no) variable with the statement: “Do you serve as the head, manager or treasurer in the organization or group?” Those who answered “yes” were categorized as “leadership position” in terms of their social positions in the participating organization, whereas those who answered “no” were categorized as “regular” members.
Demographic variables included sex, age (65–69, 70–74, 75–79, 80–84, 85 years and over), marital status (married and their spouse was alive, other), and residence year (less than 5, 5–9, 10–19, 20–29, 30-39, 40-49, 50 years and more). Socioeconomic status included annual equivalized household income (less than 19,999 JPY, 20,000 JPY to 39,999 JPY”, 40,000 JPY or more”) and educational attainment (less than 6, 6–9, 10–12, 13 years and more, others). In Japan, six-three-three system of education is employed: 6 years of elementary school; 3 years of junior high school; and 3 years of high school.
In order to account for the health endowment effect, self-reported medical condition (No illness, Having illness but need no treatment, Having illness but discontinued treatment, Receiving some treatment), self-reported health status (Very good, Good, Poor, Very poor), self-reported physical condition (Very good, Good, Poor, Very poor), and Depression (Geriatric Depression Scale)  were considered as covariates.
First, we evaluated the differences in mortality between the groups according to the participants’ baseline characteristics. Next, in order to address potential treatment selection bias due to the differential chances of occupying leadership position within the organizations, we used a propensity score approach in an attempt to create pseudo-populations based on the probability of assignment to treatment for estimating the causal effect. We calculated the propensity score using a logistic regression model, where the dependent variable was the position in the community organization: regular members vs. leadership position. For explanatory variables, we selected important variables that potentially affect assignment to treatment (in this instance, occupying a leadership position within an organization), namely, age, sex, annual equivalized household income, educational attainment, marital status, self-reported health, self-reported physical function, self-reported medical conditions, Geriatric Depression Scale, and residence year. After examining whether weighting balanced measured covariates between groups (Table 2), the inverse of the propensity score was incorporated to the weighted Cox proportional hazards models to calculate hazard ratios for mortality according to the social position occupied within the organization(s). This approach is an alternative to implementing propensity score matching to statistically balance for confounding variables in non-randomized studies [18, 19].
The overall proportion of missing data was 36.3 %, if we excluded participants who had missing data on at least one of the variables considered in our analysis. As such, missing data on independent variable and covariates were imputed by multiple imputation with method under the missing at random assumption. Multiple imputations based on multivariate normal model was calculated using all the covariates as explanatory variables: age, sex, annual equivalized household income, educational attainment, marital status, self-reported health, self-reported physical function, self-reported medical condition, Geriatric Depression Scale, residence year and position in the participating organization(s). We produced ten imputed data sets and the estimates were combined. SAS 9.2 (SAS Institute, Cary, NC) was used for all statistical analyses.