Social Capital as a Moderator for Caregivers’ Psychological Distress: A Dynamic Panel Data Model Analysis in Japan

Background: The adverse impact of caregiving on caregivers’ mental health and the positive impact of social capital (SC) on health are both well understood. This study examined the moderating effect of SC on the association between family caregiving and caregivers’ psychological distress (PD). Methods: We used longitudinal data from 27,869 individuals born between 1946 and 1955. The data were collected from a 14-wave nationwide longitudinal survey conducted from 2005 to 2018. We estimated dynamic panel data (DPD) models, which could control for an individual’s time-invariant attributes in a dynamic framework. We did this to examine how SC moderated the association between informal caregiving and a caregiver’s PD (dened by a Kessler score of 13 or higher). We also examined how the results varied over time, as the caregiver’s age advanced. Results: Of the respondents aged 50–73 years, 12.5% of women and 8.4% of men provided care to their older parents or spouses. The DPD model results showed that the onset of caregiving increased the probability of PD (M 3.4%) by 2.1% (95% condence interval [CI]: 1.6%–2.7%) and 1.1% (95% CI: 0.5%–1.6%) for women and men, respectively. SC moderated the association between caregiving and a caregiver’s PD by 53.4% (95%: 30.4%–76.4%) and 84.9% (95% CI: 62.0%–107.8%) for women and men, respectively. We also observed that the moderating effect of SC on a caregiver’s PD increased as the caregiver’s age advanced in both women and men, preventing a deterioration in the psychological impact of caregiving. Conclusions: The results underscore the moderating effect of SC on the association between caregiving and PD. This suggests the need to keep family caregivers from being socially isolated, especially as they get older.


Study sample
We used data obtained from a nationwide 14-

Materials Caregiving
The survey asked individuals if and to whom they were providing informal care. We constructed a binary variable of caregiving by allocating one to the individuals who answered that they were caring for at least one of their parents, parents-in-law, or spouse -regardless of coresidence with them -and zero to others.

Psychological distress (PD)
We measured PD using Kessler 6 (K6) scores [22,23]. Earlier studies have con rmed the reliability and validity of this score in the psychological analyses of Japanese people [24,25]. The survey asked participants to answer a six-item psychological distress questionnaire. The questionnaire included: "During the past 30 days, how often did you feel a) nervous, b) hopeless, c) restless or dgety, d) so depressed that nothing could cheer you up, e) that everything was an effort, and f) worthless?" These were rated on a 5-point scale (0 = none of the time to 4 = all of the time). Furthermore, the sum of the reported scores (range: 0-24) was calculated and de ned as the K6 score. The Cronbach's alpha for the entire study sample was 0.897. K6 scores ≥13 indicates PD in a Japanese sample, as validated by previous studies [23,25]. We constructed a binary variable of PD by allocating one to K6 scores ≥13 and zero to those < 13.

Social capital (SC)
To construct the SC variable, we utilized answers to the question about participation in social activities, as used by previous studies in Japan (e.g., [16]). The survey asked respondents whether they participated in six types of social activities: (1) hobby or entertainment; (2) sports or physical exercises; (3) community activities; (4) childcare support, or educational or cultural activities; (5) support for the elderly; and (6) others (multiple answers permitted). If respondents answered yes, they were asked to indicate with whom they participated in each activity by choosing: (1) alone, (2) family members or friends, (3) workplace colleagues, (4) members in a neighborhood association, or (5) members in a non-pro t organization or public-service corporation. Multiple answers were permitted). We considered that respondents who chose at least one item from (2)-(5) in at least one of the six social activities (1)-(6) had SC.

Covariates
For individual-level covariates, we considered marital status, employment status, household spending, current tobacco smoking, educational attainment, and age at baseline (in 2005). We constructed binary variables for having a spouse, having a paid job, and current smoking by allocating one to those who answered that they were married, have a paying job, and currently smoke, respectively, and zero to those who answered "no" to those questions. Regarding household spending as a proxy for household income, we divided the reported monthly household spending by the square root of the number of household members to adjust it by household size [26]. We categorized it into quartiles and constructed binary variables for each quartile. To the respondents who did not answer for household spending, we allocated a binary variable for unanswered. Regarding educational attainment, we constructed ve binary variables for graduating from junior high school, high school, junior college, college or above, and other. We also constructed binary variables for each age at baseline. We further included binary variables for each survey year to control for survey-year-speci c factors. In the DPD regression, ages at baseline and educational attainment were automatically removed because they were time-invariant.

Statistical analysis
We began by using descriptive analysis to describe unadjusted associations across caregiving, SC, and PD by comparing the prevalence of PD across four types of respondents: non-caregivers with SC, non-caregivers with no SC, caregivers with SC, and caregivers with no SC. We did so using pooled cross-sectional data.
For regression analysis, we limited the study sample to those who provided no care in the previous year and estimated a DPD model [21] in the form of a linear probability model (LPM) [27,28], which linearly predicts the probability of individual i's PD in wave t as: PD it = β 0 + β 1 Caregiving it + β 2 SC it + β 3 Caregiving it × SC it + β 4 PD it−1 + (covariates) + u i + ε it , where u i indicates individual i's time-invariant attributes and is controlled for in the DPD model. We predicted β 1 > 0, β 2 < 0, β 3 < 0, and β 4 < 0. β 1 and β 1 + β 3 represent the association of caregiving with no SC and caregiving with SC, respectively. The ratio -β 3 /β 1 indicates the proportion of the association between caregiving and PD moderated by SC. If the interaction term (Caregiving× SC) is not included in the regression model, the coe cient of caregiving (β 1 ) represents the association between caregiving and SC on average for caregiving both with and without SC.
We rst estimated the DPD model separately for men and women using the data over the entire wave (with the age in the range of 51-72 years) and compared the results with those of pooled cross-sectional models. We then estimated the same DPD model for the 10-year age range group by sliding the age range by one year from 55-64 years to 63-72 years, and compared the results across nine age groups.
We employed LPM rather than logistic or probit models to predict a binary variable for PD. We did this in consideration of the general validity of LPM [27,28], the straightforward interpretation of the coe cient of the interaction term [29], and a substantial reduction in our sample size due to the required exclusion of the respondents who experienced no change in a binary variable for PD over the study period from logistic xedeffects model regression [30].
For both descriptive and regression analyses, we compared the results between women and men. We did this because observations in previous studies provide evidence of gender differences in caregiver stressors [31][32][33]. For all statistical analysis, we used the Stata software package (Release 15; STATA Corp, College Station, TX, USA). Table 1 summarizes the proportion of caregivers out of the pooled observations (individuals × waves) and examines how the prevalence of PD differed according to the combination of caregiving and SC. Of the entire sample, 12.5% of women and 8.4% of men were caring for older adults, while 72.2% of women and 68.9% of men had SC. Figure 1 compares the prevalence of PD across four combinations of caregiving and SC, based on the results reported at the bottom of Table 1. For both women and men, the prevalence of PD was higher among caregivers (6.4% for women and 4.7% for men) than among non-caregivers (3.4% for women and 2.8% for men). The prevalence of PD was also higher among those with no SC (6.7% for women and 5.1% for men) than those with SC (2.6% for women and 2.0% for men), as we expected. Equally important, Fig. 1 shows that, for women, the difference between caregivers with SC and non-caregivers with SC (4.6% − 2.4%) was somewhat smaller than the difference between caregivers who experienced PD while having no SC and non-caregivers with no SC (11.5% − 6.1%). The same was true of men (3.3% − 1.8% vs. 8.4% − 4.9%). These comparisons suggest that SC has a moderating effect on the association between informal caregiving and PD. This effect can be formally con rmed by the two-way analysis of variance, which showed that caregiving, SC, and their interaction all signi cantly explained the variance of the probability of PD (p < 0.001, not reported in Table 1 or Fig. 1). It should be noted that the results in Table 1 and Fig. 1 were not adjusted for other factors.

Descriptive analysis
Regression analysis Table 2 summarizes the regression results using the entire study sample. This compares the estimated associations with the probability of PD between the pooled cross-sectional and DPD models. The upper part of the table presents the results obtained when we did not include the interaction term (Caregiving × SC), while the lower part presents the results with the interaction term. The full regression results are presented in Table S1 in the Supplementary le. As seen in the upper part of Table 2, the estimated coe cients of caregiving (β 1 ) indicated that caregiving -regardless of with or without SCcorresponded to 2.1% (95% con dence interval [CI]: 1.6-2.7%) and 21.1% (95% CI: 0.5-1.6%) higher probability of PD for women and men, respectively. The magnitudes of associations were somewhat lower than those observed in pooled cross-sectional models, 2.6% and 1.6% among women and men, respectively. Still, they were substantial compared to the actual prevalence of PD, 3.8% and 3.0% among women and men, respectively, as seen in Table 1. The negative coe cients of SC (β 2 ) indicated a negative association between SC and PD, although its magnitude was much more limited in the DPD models than in the pooled cross-sectional models.
The lower part of Table 2 shows that the estimated coe cients (β 3 ) of the interaction term (Caregiving × SC) were negative in all models. This con rms the moderating effect of SC on the association between caregiving and PD. The proportions of the association moderated by SC were 53.4% (95% CI: 30.4-76.4%) and 84.9% (95% CI: 62.4-107.8%) for women and men, respectively. These results were slightly lower for women and higher for men when compared to the pooled cross-sectional model results. Table 3 summarizes the estimated associations of the probability of PD with (i) caregiving without SC (β 1 ), (ii) the interaction between caregiving and SC (β 3 ), and (iii) caregiving with SC (β 1 + β 3 ), as well as the estimated proportion (%) of the moderating effect of SC (-β 3 /β 1 ). All of these were obtained from DPD models for nine 10-year age ranges. Figure 2 graphically illustrates how the estimated associations of PD (i) caregiving with no SC and (ii) caregiving with SC) changed according to a caregiver's age. In Table 3 and Fig. 2, we observe that for women, the association of PD with caregiving with no SC increased as a caregiver got older. Meanwhile, the association of PD with caregiving with SC declined slightly. The moderating effect of SC increased as age advanced; the proportion of the effect increased from 39.3% for women aged 55-64 years to 82.8% for those aged 62-71 years. In the case of men, the increase in the association of PD with caregiving without SC as age advanced was much more limited than for women. The association of PD with caregiving with no SC remained largely unchanged and nonsigni cant over the entire age range for men. The proportion of the moderating effect of SC was somewhat larger and showed a less clear increase as age advanced compared to the results found in women. a Based on dynamic panel data models adjusted for covariates (see Table 2). b Social capital. c Con dence interval.

Discussion
This study examined the extent to which SC moderated the association between caregiving to older parents or a spouse and caregiver's PD.
Unlike most preceding studies, this study applied a DMD model analysis, which allowed us to control for an individual's time-invariant attributes in a dynamic framework. This approach, combined with the limitation of the sample to those who provided no care in the previous year, mitigated the biases that cannot be controlled for when using cross-sectional regression. In addition, we compared the results across different age groups, considering the possibility of a change in the relevance of SC for caregivers' mental health as caregivers aged. The key ndings and their implications are summarized as follows: First, the results con rmed the moderating effect of SC on the association between caregiving and PD for both women and men, a result consistent with previous studies that showed the favorable impact of SC on health [16][17][18][19]. Even after controlling for individual-speci c xed effects and other factors, the DPD model results suggest that SC mitigated a substantial portion of the adverse impact of caregiving on caregivers' mental health. This implies that socially isolated caregivers may face higher risks of deterioration in their psychological well-being, pointing to the need for policy measures to help middle-aged and older adults create and enhance their SC.
Second, the moderating effect of SC increased as caregivers aged. The adverse impact of caregiving tended to be more serious for older caregivers, especially for women. This result highlighted health problems related to the issue of "elderly-for-elderly care." It should be noted, however, that the moderating effect of SC tended to offset the enhanced adverse impact of caregiving on PD. As a result, the probability of PD Third, we observed a substantial difference in the moderating effect of SC. We found that the effect differed substantially between women and men. As already observed in previous studies, there are consistently higher levels of PD in women [31][32][33]. Notably, the proportion of the association moderated by SC (-β 3 /β 1 ) was somewhat more limited for women than for men. Considering the magnitude of the interaction effect between caregiving and SC (|β 3 |) was almost in the same range or even larger for women (especially for older age groups), as seen in Table 3, an SC's lower moderating effect for women can be accounted for by a closer association between caregiving and PD, that is, higher value of β 1 , for women. This appears to have partly re ected that women undertake more intensive caregiving; indeed, the average hours for caregiving per week was 19.0 hours (SD: 23.7 hours) among female caregivers, as compared to 13.2 hours (SD: 23.7 hours) among male caregivers. Hence, the lower moderating effect of SC for women did not mean a limited importance of SC for women. In fact, that effect increased substantially as female caregivers got older, as clearly seen in Fig. 2. Meanwhile, our results showed a more limited increase in the moderating effect of SC for men as they got older. This may re ect that a man spends longer in the workplace, which was likely to reduce the chances of enhancing SC, as compared to women.
We recognize that this study has several limitations and that many issues remain to be addressed. Importantly, we did not fully address the potential endogeneity of SC. DPD model analysis controlled for the confounding effects of an individual's attributes on the associations between key variables, and we found that the estimation results differed substantially from pooled cross-sectional models, as already suggested by previous xed-effects model studies [34,35]. However, we could not exclude the possibility that caregiving affects SC during the caregiving process [36], especially as we focused on the individual-level SC. Second, we did not examine the evolution of caregivers' PD over time. Instead, we focused on the relatively short-term association between the onset of caregiving and a caregiver's PD in order to mitigate potential simultaneity biases. Going forward, we should extend the analysis to the evolution of a caregiver's PD over time [37][38][39]. Examining that must involve a two-way interaction with SC, as mentioned above. In addition to these methodological limitations, caution should be exercised when generalizing the results. The relevance of caregiving and caregivers' mental health may depend heavily on formal long-term care provisions, family care norms, and other socio-cultural backgrounds [40].

Conclusions
This study con rmed the moderating effect of SC on the association between family caregiving and caregivers' PD for both women and men. The results also con rmed that this effect prevented PD from continued deterioration as age advanced, which suggests the need to keep older caregivers from becoming socially isolated.