Study participants
We used data from the AsiaBarometer Survey (2003–2006), which included information on individuals from 29 Asian countries on a vast range of subjects [21]. The countries included in our analysis were Afghanistan, Bangladesh, Bhutan, Brunei, Cambodia, China, Hong Kong, India, Indonesia, Japan, Kazakhstan, Kyrgyzstan, Laos, Malaysia, the Maldives, Mongolia, Myanmar, Nepal, Pakistan, Singapore, South Korea, Sri Lanka, Taiwan, Tajikistan, Thailand, the Philippines, Turkmenistan, Uzbekistan, and Vietnam. For the purpose of the study, Hong Kong and Taiwan were considered independent countries, in view of their socioeconomic characteristics. Prior ethics committee approval was obtained from the Chuo University. We received written informed consent from the survey participants.
Data collection
We used face-to-face interviews to administer structured questionnaires. The detailed content of the questionnaires has been published previously [21]. Data collection included demographics, marital status, socioeconomic factors (income, education, and occupation), self-rated health, interpersonal trust, and trust in the healthcare system and mass media, as well as information on political, environmental, and daily-life issues that were related to the AsiaBarometer Survey.
The individual-level independent variables included gender, age (range between 20 and 69 years), marital status, religious belief, income, education, employment, and individual-level social trust. Age was categorized into five groups of 20–29, 30–39, 40–49, 50–59, and 60–69 years old. Categories of marital status included single, married, divorced/separated, or widowed.
Annual household income was used as an income variable in this study. Categories of the income groups included low, middle, and high, based on the income distribution of each country (see Appendix A, in Additional file 1). For educational achievement, we also used three categories (low, middle, and high) based on the distribution of educational achievement in each country (see Appendix B, in additional file 1). For occupational status, six categorical classes were used: self-employed, employed, unemployed, retired, homemaker, and student. The self-employed group included: self-employed in agriculture, forestry or fisheries; business owner in mining or manufacturing industry of an organization with up to 30 employees; vendor or street trader; business owner or manager of an organization; and self-employed professional. The employed group included senior manager, employed professional or specialist, clerical worker, sales, manual worker, driver, and "other" worker.
In this study, self-rated health was defined as the individual's personal satisfaction with their overall health. In the survey, we asked "Please tell me how satisfied or dissatisfied you are with your health? Would you say you are very satisfied, somewhat satisfied, neither satisfied nor dissatisfied, somewhat dissatisfied, or very dissatisfied with your health?". These categories were collapsed to form a dichotomous outcome of self-rated health: poor health (1) for very dissatisfied, somewhat dissatisfied, or neither satisfied nor dissatisfied; and good health (0) for very satisfied, or somewhat satisfied.
Horizontal trust, a dimension of cognitive social capital, was measured by a composite index constructed from a factor (principal component) score of three questionnaire items related to general trust, interpersonal trust, and mutual help. The general trust question was, "Would you say that most people can be trusted or that you can't be too careful in dealing with people?". The question for interpersonal trust in merit-based utility was, "Would you say that most of the time people try to be helpful or that they are mostly looking out for themselves?". The question for mutual help was, "If you saw somebody on the street looking lost, would you stop to help?". For the last question, the responses were: "I would always stop to help", "I would help if nobody else did", and "It is highly likely I wouldn't stop to help". These questions have been widely used in previous studies to measure cognitive social trust [2, 5, 22, 23]. Factor analysis of these items provided a one-factor solution with an eigenvalue of 1.4. All items were loaded above 0.4 and no other factors exceeded unity. The individual scores were calculated using the regression equation with the factor loadings, and a higher score indicated lower trust. The scores were then standardized (mean 0; standard deviation 1). Before being included into the multivariable multilevel model, the scores were further collapsed to form a dichotomized variable: low social trust (0) for the values less than 0 and high social trust (1) for the values of 0 or more.
Trust in institutions (vertical trust) is an item that reflects the participant's trust in the healthcare system and in mass media (specified as newspapers and television). The item "Please indicate to what extent you trust the following institutions to operate in the best interests of society" offered the alternatives (a) the healthcare system and (b) mass media, with the six alternative responses: (1) "Trust a lot"; (2) "Trust to a degree"; (3) "Don't really trust"; (4) "Don't trust at all"; (5) "Haven't thought about it"; and (6) "I don't know".
Statistical analysis
Descriptive statistics were calculated and presented as the mean with standard deviation or the count number in proportion to the overall sample population where appropriate. Bivariate correlation analyses were conducted among the trust variables using Pearson's correlation coefficients.
We used the multilevel (mixed-effects) logistic regression model to analyze the relationship of individual characteristics to self-rated health by considering individuals nested in each country, as data structures in the Asia Barometer Survey were hierarchical multilevels (level 1, individual; level 2, country). The data provide information on individuals, while the individuals are also grouped in their countries. Analyzing hierarchical data at the individual level by conventional regression models does not meet the assumption of independence of observations. When ignoring the nesting of individuals in countries, the estimated standard errors would be smaller, thus inflating the risk of Type I errors [24]. The mixed-effects model can be used to analyze hierarchical data [24], and is used widely in social and epidemiological research. The random-effects covariance matrix was set to an unstructured form and we utilized three trust measures (horizontal trust, trust in the healthcare, and trust in mass media) as the random-effects parameters in the model. Variances and their standard errors were estimated for these random-effects parameters.
The model was constructed to evaluate the relations of trust in the healthcare system and mass media to self-rated health, adjusted for age, gender, marital status, income, education, occupation, and horizontal trust. We constructed a total of six models, including only baseline sociodemographic variables (base), such as age, gender, marital status, income, education, and occupation (Model 1), base plus horizontal trust (Model 2), base plus trust in the healthcare system (Model 3), base plus trust in mass media (Model 4), base minus income and education plus horizontal trust, trust in the healthcare system and trust in mass media (Model 5), and base plus horizontal trust, trust in the healthcare system and trust in mass media (Model 6; full model). Model 5 was constructed by eliminating income and education from the full model for examining the possible endogeneity to health of income and education.
No interaction terms were included in the model. To check the robustness of the model, we also conducted the logistic regression analysis including country fixed effects as well as the ordered probit model analysis using original dependent variable (self-rated health). The odds ratios (ORs) along with 95% confidence interval (CIs) were estimated in each variable for poor health. An OR value greater than one indicates greater effects that were positively related to poor health. All statistical analyses were performed using STATA 10.0 (College Station, TX, USA). Two-tailed P-values less than 0.05 were considered statistically significant.