Sample
Data were derived from a survey of respondents drawn from a nationally representative web-based sample of 6909 who participated in the Hankook Research Master Sample Panel. The respondents, adult men and women 20 years of age and older currently residing in South Korea, were proportionally selected by gender and age based on census data. The sample households were distributed according to the ratio of houses and apartments for each cluster (metropolis, city, and town). The final respondents were recruited by extracting 229 clusters from 15 provinces with a stratified cluster sampling method. The final response rate was 80.0%. Missing values of survey questions for key analytical variables were excluded using a pairwise method.
Analytical framework
The contextual multilevel models employed in this study attempt to identify the effects of social integration on individual health status. We adopted a multilevel proposition, what is referred to here as a contextual multilevel approach, which shows the effect of the macro-level variable z on the micro-level variable y while controlling for the micro-level variable x [22]. The contextual multilevel approach permits a simultaneous examination of how individual- and group-level variables are related to individual-level health outcomes using community-level predictors. It accounts for the possibility of a residual correlation between individuals within groups and enables an examination of between-group variability and the factors associated with it [22, 23].
Measures
The contextual effects model of this study is based on Blakely and Subramanian [24]. We used two types of indicators which show the integrative characteristics of community-level contextual effects. An aggregate indicator refers to the effect of a derived group-level variable on an individual-level outcome (e.g., mean educational attainment). An integral indicator refers to the effect of group-level variables that can apply to any situation involving lower-level units nested within higher-level units (e.g., gross regional domestic product; GRDP) [23, 25]. We employed the ratio of one-person households (ROPH) and the total fertility rate (TFR) as aggregate indicators and the number of four (murder, robbery, theft, violence) major crimes (NFMC), the number of beds in medical institutions per 1,000 people (NBMI), and GRDP as integral indicators.
Dependent variable
The health outcome of this study is the self-rated health (SRH) status of residents. Respondents were asked to rate their own general health status on a ten-point Likert-type scale ranging from very good (10) to very bad (1) when responding to the question “How is your health in general?” The answers were eventually grouped into two categories. Respondents reporting 1 to 5 points for self-rated health status were coded as 1 and were considered as the low-SRH group, while those reporting 6 to 10 points self-rated health status were coded as 0 and were considered as the high-SRH group. In prospective studies, this general health question has been validated as a good predictor of morbidity and mortality, with a differential relationship between consecutive categorical ratings of SRH and the probability of mortality [26, 27].
Independent variables
Individual-level variables: Four dimensions of structured tools developed to measure social quality were used [28]: social communication, political participation, social participation as a form of neighborhoods and organizations, and social inclusion as a type of in- and out-of-network. social participation (SP) refers to the degree of informal ties by which an individual is connected to his community [29-31], and social inclusion (SI) has been developed as a proxy for social efficacy [32]. All other variables were used in the Korean General Social Survey (KGSS; http://kgss.skku.edu/).
Community-level variables: Indicators were collected based on the European social quality (SQ) frame by Maesen and Walker [33]. This frame points out four conditions of a good society: socio-economic security, social cohesion, social inclusion and social empowerment, and proposes ninety-five sub-indicators. Based on this frame, we systematically collected usable indicators from 230 local governments in South Korea. The indicators were primarily collected from the central government statistics portal, which provides social indicators to local governments. The secondary sources were major public organizations followed by the Korea Public Information Disclosure System (https://www.open.go.kr). The portals used to collect indicators were the Korean Statistical Information Services (http://kosis.kr), the Local Administration Integrated Information System (http://www.laiis.go.kr), and the National Geographic Information Institute (http://nationalatlas.ngii.go.kr). We also used the expanded national database from the Ministry of Public Administration and Security, Ministry of Health and Welfare, Ministry of Culture, Sports and Tourism, the National Police Agency, the National Election Commission, the Health Insurance Review & Assessment Service, the Korea Education Development Institute, and the Korean Film Commission. Under the principles of conceptual suitability, clarity, reliability, consistency, changeability, and comparability, 81 indicators collected in total. Five indicators were selected using the Delphi method [7]. The variables representing social integration are the total fertility rate and GRDP, and the variables representing social disorganization are the number of four major crimes and the ratio of one-person households.
Standardization of Indicator
The indicators selected were transformed by the imputation of missing values, ESS (the European Social Survey) standardization, and GIS (Geographical Information System) transformation. First, the following were considered for missing values: 1) when there were no data from the year under examination, they were substituted with data from the previous year; 2) when there were no data from the previous year, they were substituted with data from the year before that; and 3) in other cases, the data were substituted with the average values for the province in which the region under examination was located. However, when the average values were simple sums, they were substituted with 1/n of the number of regions; additionally, when they were rate values, they were used in their entirety. Indicators were inputted after two-stage computations were calculated. First, we standardized the z-score of each indicator with the transformation formula of the European Social Survey in order to unify the units of measurement and form a normal distribution. The ESS methods enable us to minimize the bias of outlier observations as well as model complexity by transforming an indicator’s minimum and maximum values into 0 and 10, respectively, and its average into approximately 5 (http://essedunet.nsd.uib.no).
See equation 1 in the supplementary files.
The medical resource factor consisted of two measures. The number of physicians is the aggregate number of certified physicians per 1,000 people in a certain region. The number of general hospitals was computed by utilizing the GIS method so that we could take into consideration the availability of hospitals in neighboring regions to measure medical accessibility.
The GIS-based number of general hospitals = the number of general hospitals in a certain region + (the neighboring region's number of general hospitals / the number of neighboring regions) + (the secondary neighboring region's number of general hospitals / the number of secondary neighboring regions)
For example, suppose that there are three general hospitals in region X; at the same time, six general hospitals are in four neighboring regions and eleven general hospitals are in seven secondary neighboring regions. By substituting figures into the above formula, 3+(6/4)+(11/7), we get a value of 6.07. This is the number of GIS-based hospitals in region X.
Potential confounders
The individual-level confounders were gender, age, annual income, educational attainment, and district type (metropolis, city, and town). The community-level confounders were the number of beds in medical institutions per 1,000 people.
Statistical analyses
The unconditional slope and conditional model of the multilevel model were devised for binary variables using the Bernoulli response and logit link respectively. Based on this model, the relationships between social integration/disorganization and low-SRH status at the individual and community levels were reviewed for statistical significance even after controlling for the relevant covariates.
See Models in the supplementary files.
In the logistic models, we calculated the intra-class correlation (ICC) using the formula σ2/(σ2+3.29), where σ2 is the area-level variance. The estimated size of the ICC based on the above model was 4.79%. The effect size was 0.22 (Nagelkerke value). The analysis procedure proceeded in the following order: descriptive statistics, factor and reliability analyses, and hierarchical generalized logistic regression. In the multilevel analysis, model 1 was the unconditional model, model 2 the unconditional slope model, and model 3 accounted for community-level and individual-level effects of social integration and disorganization. All analyses were conducted using HLM for Windows v.7.0.
Ethics Statement
This project was approved by the institutional review board of Dongduk Women’s University, Seoul, Korea (Sep 24, 2019; DDWU1909-01). Subjects provided written informed consent to participate in this study. No information that can publicly identify individual participants was collected during the data collection process.