A cross-sectional survey was conducted using a Qualtrics online survey platform (Qualtrics Inc, Provo, UT, USA). Qualtrics is a well-known online survey company that offers various services for the research, which includes access to the Qualtrics research panel facilitating participant recruitment for online data collection. The Qualtrics research panel consists of pre-arranged individuals recruited from various sources who have agreed to respond to Qualtrics online surveys in exchange for compensation. The Qualtrics research panel is increasingly recognized as an acceptable online data source for the research  and has been successfully used for health-related studies involving older adults . For the present study, the recruitment of participants as well as the administration of the survey were operated by Qualtrics, Inc. Adults in the pool of Qualtrics research panel who met the following eligibility criteria were invited to the study: 1) adults aged ≥40 years old; 2) residents of Virginia in the U.S.; and 3) can read English. Upon electronic informed consent, participants were asked to complete the online survey including a set of questionnaires about demographic characteristics, social capital, and PA levels at their own pace and at any time or place where they can have internet access to complete the survey in either web- or mobile-based platforms. Of the initial respondents of 803, 17.56% (n = 141) were excluded due to the failure to pass the pre-designed quality check questions and to provide valid data on key variables, leaving the final sample of 662 (Mean age: 58.11±10.59 years old; Range = 40 – 85 years old);. Participants who completed the survey received the compensation in various forms (e.g., gift cards, electronic coupons) from the Qualtrics Inc. The study protocol was approved by the Institutional Review Board at Texas Tech University (IRB#: 2019-307).
Personal Social Capital
Individual-level social capital was assessed using the Personal Social Capital Scale (PSCS). PSCS was developed by Chen et al.  to measure the two subtypes of individual-level social capital, bonding and bridging social capital, that are derived from a network perspective. Bonding social capital refers to links between community residents whose social identities are similar while bridging social capital involves connections between community residents with differing status and power . Thus, bonding social capital indicates the cohesion within a group, enhancing the homogeneity of the group’s characteristics and helping to mobilize reciprocity and solidarity among actors. By this means it strengthens the linkages between the members of the group. In contrast, bridging social capital is directed outside the group by linking actors of different networks. PSCS consists of 10 composite items calculated based on 42 subitems measured by a five-point Likert scale with 1 = ‘none or a few’ to 5 = ‘all or a lot’.
In the initial development, the scale’s internal consistency was proven to meet common standards with Cronbach’s alphas of 0.87 for the overall scale and 0.85 and 0.84 for the bonding and bridging social capital subscales, respectively, among 128 Chinese adults aged between 18 and 50 years old . Additionally, the authors demonstrated acceptable construct validity by examining known-group differences as well as predictive validity. PSCS has repeatedly been tested for its validity and reliability across different population groups [27, 28] including older adults . In the present study, the scale was checked for reliability with satisfying levels of internal consistency (Cronbach’s α = 0.82 & 0.81 for bonding and bridging scale, respectively). Additionally, confirmatory factor analysis was conducted and showed acceptable fit indices (Standardized root mean squared residual= 0.059; Root mean square error of approximation = 0.098; Comparative fit index = 0.911), indicating that the PSCS is an adequate tool to measure bonding and bridging social capital among the sample of this study.
The outcome variable of self-rated health was measured with a single question, asking how they would rate their health in general using a five-point Likert scale ranged from 1 (poor) through 5 (excellent). Self-rated health question is a simple and easy to administer measure of general health, allowing respondents to prioritize and evaluate different aspects of their health. The self-rated health question has been widely used as a measure of self-rated health in the elderly population . Although the reliability of the item has been debated in some contexts, its adequate predictive validity in terms of several objective health measures has been verified in numerous studies [31, 32]. For instance, Baćak and Ólafsdóttir  examined the concurrent validity of self-rated health item among 19 European countries and found the measure to be a significant predictor of both mental and physical health (0.386 < r < 0.768). For the present study, participants were categorized into ‘not-good’ (fair/poor) and ‘good’ (good/very good/excellent) health groups, which is in line with the majority of literature examined self-rated health with social capital .
Leisure-Time Physical Activity
The long form of the International Physical Activity Questionnaire (IPAQ) was used to assess the level of PA. Initially developed by the International Consensus Group for the development of an internationally standardized measure of PA , the IPAQ has been extensively used worldwide with well-established evidence of the reliability and validity across different settings [36, 37]. The participants were asked to disclose frequency and time they have spent in PA at moderate- and vigorous-intensity levels during the past 7 days across the four domains (i.e., domestic, leisure-time, work-, and transport-related activities). For the present study, participant’s responses to leisure-time PA were used to calculate total energy expenditures in metabolic-equivalent units (METs-minutes/week) by the IPAQ scoring guidelines . Participants were categorized into three groups based on the current recommendations : ‘no LTPA’, ‘<600 MET-minutes/week’, which is equivalent to 150 minutes of moderate PA, and ‘≥600 MET-minutes per week’.
Following common operational definitions , SES was assessed among the dimensions of education and household income. Education level was assessed through a seven-point scale with the response categories of ‘less than high school diploma’, ‘high school degree’, ‘some college credit, no degree’, associate degree’, ‘bachelor’s degree’, ‘master’s degree’ and ‘doctorate’. Annual household income was assessed through a six-point scale with the response categories of ‘<$20k’, ‘$20k to <$35k’, ‘$35k to <$50k’, ‘$50k to <$75k’, ‘$75k to <$100k’ and ‘≥$100k’. Both variables were recoded to ‘low’ (‘high school degree’ or below), ‘middle’ (‘some college credit, no degree’ and ‘associate degree’) and ‘high’ (‘bachelor’s degree’ or above) for education level, and ‘low’ (<$35k), ‘middle’ ($35k to <$75k) and ‘high’ (≥$75k) for household income, respectively.
Other Study Covariates
Demographic characteristics including age (<60 years old vs. ≥60 years old), gender (male vs. female), marital status (married or domestic partnership vs. single or no partnership), and ethnicity (white vs. others) were collected. Additionally, self-reported history of chronic diseases they have ever confirmed by health professionals such as stroke, asthma, cancer, arthritis, diabetes, kidney disease were used to categorize the participants into three groups (‘no’, ‘one’, and ‘one or more’ chronic diseases).
Descriptive statistics of study variables were calculated differentiated by self-rated health status, followed by chi-square test of independence examining between-group differences in the proportions for categorical variables. The differences in social capital scores by the key explanatory variables including LTPA levels and SES characteristics were examined using one-way analysis of variance with Tukey’s pairwise comparisons. Additionally, given the ordered nature of LTPA and SES variables, linear trends were tested using orthogonal polynomial contrasts.
Using the self-rated health (‘good’ vs. ‘not good’) as a dependent variable, logistic regression model was established predicting the likelihoods of reporting ‘good’ self-rated health based on a set of explanatory variables. The models were hierarchically developed to address the proposed research questions. The model 1 was established to examine the association of social capital with self-rated health without adjustment of other study covariates, followed by the model 2 adjusting for the demographic variables including age, gender, ethnicity and marital status, as well as for the history of chronic diseases. LTPA levels were then introduced in the model 3 to test the possible mediating effect on the relationship between social capital and self-rated health as suggested by Baron and Kenny [40, 41]. Specifically, two nested logistic regression models (model 2 and 3) were statistically compared using the Karlson-Holm-Breen (KHB) method  to test the significance of the indirect effect (i.e., mediating effect of LTPA). Lastly, SES variables including education and household income were additionally included in the model 4. After applying the overall model, the possible moderating effects of SES were tested by including interaction terms between social capital and SES variables (i.e., bonding-by-education, bonding-by-household income, bridging-by-education, and bridging-by-household income) in the model. If the interaction effect is significant regarding the specific SES variable, follow-up simple effect logistic regression analyses were conducted to examine the differences in the effect structures by the level of SES variable. Throughout the analyses, the results are presented as odds ratio (OR) along with 95% confidence interval (CI). As a supplement, average marginal effects (AME) expressing the average influence of independent variable in a form of probability of reporting “Good” self-rated health was reported. Additionally, linear trends analyses were performed using orthogonal polynomial contrasts for the ordered categorical variables including LTPA and SES variables. The goodness of fit of the logistic models were reported with the Nagelkerke’s pseudo-R2, representing the reduction of deviance due to the inclusion of predictors, and the Hosmer-Lemeshow goodness-of-fit test. Multicollinearity was tested by the variance inflation factors for each model. All data analyses were performed using the SAS v9.4 (SAS Institute, Cary, NC, USA) and STATA v13.1 (Stata Corp, College Station, TX, USA). Statistical significance was set at P ≤.05.