Satisfaction with job
Satisfaction with job was assessed using a six-item scale measuring one’s satisfaction with current income, job security, working environment, working time, promotion opportunity, and general perception on a five-point Likert scale from 1 (very unsatisfied) to 5 (very satisfied). The item measuring satisfaction with job opportunity was excluded because of a large amount of missingness at all waves vs other items (e.g., 65.4% vs. 21.8% for general satisfaction with job in 2018). An example item is “In general, how satisfied are you with this job?”. Item scores were averaged with higher scores indicating greater satisfaction with job. The internal consistency of the scale is good (Cronbach αs > 0.82).
Satisfaction with marriage
Satisfaction with marriage was assessed using three items (“In general, are you satisfied with your current marriage/cohabitation?”, “Are you satisfied with the economic contribution that your spouse/partner makes to the family?”, and “Are you satisfied with the contribution on housework that your spouse/partner makes to the family?”) on a five-point Likert scale from 1 (very unsatisfied) to 5 (very satisfied). Item scores were averaged with higher scores indicating greater satisfaction with marriage. The internal consistency of the scale is acceptable (Cronbach αs > = 0.76).
Satisfaction with medical services
Satisfaction with medical services was assessed using one item (“Are you satisfied with the overall medical service?”) on a five-point Likert scale from 1 (very unsatisfied) to 5 (very satisfied).
Covariates
In all models, we adjusted for a wide range of covariates. Categorical covariates in the present study included baseline sex, education level, residential possession or hukou [20], and Chinese Communist Party membership, with female (vs. male), high school or below (vs. some college or above), rural residents (vs. urban residents) and non-party member (vs. party member) as the reference group respectively. Continuous covariates included age and log-transformed family income [21]. We included age, sex, education, residential possession, and family income because they had been linked to health and were commonly controlled for in the literature [17, 22]. We included Chinese Communist Party membership as a covariate because it being a nationally integrated political party has the founding and dominant status in China, and its membership may confer better access to resources. The membership has been conceptualized as an indicator of social capital linked to higher health care utilization [23]. Furthermore, perceived physical health at baseline was additionally controlled for in the multivariable linear regression models predicting perceived physical health at follow-up.
Outcomes
Perceived physical health
Perceived physical health was assessed with the question, “How would you rate your health status?” on a five-point Likert scale from 1 (excellent) to 5 (poor). Item scores were reverse scored so that higher scores indicate better health.
Chronic health condition onset
Chronic health condition onset was measured with the question, “During the past six months, have you had any doctor-diagnosed chronic disease?”. Possible answers were “Yes” and “No”.
Hospitalization
Hospitalization was measured with the question, “In the past year, were you ever been hospitalized due to illness?”. Possible answers were “Yes” and “No”.
Statistical analysis
All analyses were conducted in R version 4.1.3 (R Project for Statistical Computing) via base R functions and car, psych, forestplot, bruceR, mice, and miceadds packages [24,25,26,27,28,29,30].
To examine if increases in domains of life satisfaction predict better physical health and lower odds of chronic condition onset and hospitalization above and beyond their baseline levels, we calculated the difference of participants’ scores on each domain’s life satisfaction at baseline and at follow-up such that a positive score indicates that a participant’ satisfaction has increased over time. These values, along with their baseline levels, were then used as predictors of physical health in (logistic) regression models. Levels of multicollinearity were low across all models (variance inflation factors < 2.5).
In total, we built one regression model with perceived physical health (continuous) as the outcome and two binary logistic regression models with chronic condition onset (binary) and hospitalization (binary) as the outcomes, respectively. In each of the models, we entered both changes and levels of three domains of life satisfaction (job, marriage, and medical services) simultaneously while controlling for covariates including age, sex, education level, residence possession, log-transformed family income, and membership of Chinese Community Party. To predict chronic health condition onset and hospitalization, in each logistic regression model, we removed participants who reported chronic health conditions and hospitalization in the baseline wave, respectively. Doing so resulted in 8542 and 8975 cases in these two models, respectively. Lastly, because some individuals were nested within families, clustered robust standard error was calculated in all regression models at the family level using family identifiers via the (g)lm.cluster () functions of the miceadds package [30]. All continuous predictors were standardized (mean [M] = 0, standard deviation [SD] = 1) to facilitate effect size comparison and interpretation of study findings.
Sensitivity analyses
To evaluate the robustness of the findings, we conducted the following sensitivity analyses: (1) reanalysis of all models while additionally adjusting for Big-Five personality traits assessed by a validated brief Chinese version in 2018 [31]; (2) reanalysis of all models using only the items measuring general satisfaction with job and marriage in addition to the single-item question measuring satisfaction with medical services (i.e., three single-item questions corresponding to each domain);(3) reanalysis of all models using 10 imputed datasets by chained equations with binary variables imputed using logistic regression and continuous variables imputed using predictive mean matching via the mice() function of the mice package; and (4) reanalysis of all models using the alternative baseline CFPS wave (2014) and two different observation periods (i.e., from 2014 to 2018 and from 2014 to 2020).