Study design and data
The data used in this study was derived from the 2014 National Internal Migrant Population Dynamic Monitoring Survey [18], which covered 348 cities in 32 provincial units and collected by the National Health Commission of China. The purpose of the survey was to investigate the utilization of health services among internal migrants. The sampling frame for this study was taken using the stratified multistage random sampling method by probability proportional to size (PPS) approach. All respondents in this study were aged 15-59 years who had been living in local residence without the ‘Hukou’ for more than one month, including migrants from both rural and urban areas. For more details on sampling, design and approvals of the survey, please refer to an earlier study [19]. The detailed sampling process was shown in Fig. 1. Finally, a total of 7,592 migrants with inpatient service need were included in this nationally representative analysis.
Dependent variable
In this study, the estimation of the need of inpatient services is based on doctor’s recommendations. Thus, migrants’ inpatient service need was measured by questions about whether they were asked to be hospitalized by a doctor’s diagnosis. Based on the inpatient service need, the outcome was categorized into unmet inpatient services need and met need. The unmet need for inpatient service referred to the proportion migrants who were asked to be hospitalized by a doctor but did not utilize it. The key independent variable was socioeconomic status (SES). American sociologist Duncan pointed out that income and education can directly represent socioeconomic status, that is, income represents economic status, and education represents social status [20]. Li et al. [21] also suggested that the SES estimated based on education and income was largely consistent with people's subjective evaluation of occupation prestige. Consequently, we used educational attainment (primary school or below, middle school, high school, and university or above) and economic level (measured by household income per month. Quartile 1 was the poorest and Quartile 4 was the richest) to present SES. In this study, we assessed the SES in two ways. First, in order to compare the inequity of the high-low SES in inpatient service utilization from a macro perspective, we integrated the educational level and economic status (household income per month) into a single SES index using principal component analysis (PCA) [22] method (details see in Appendix Table A1). Then, we used the two specific indicators to show the associations between SES and inpatient service utilization.
Independent variables
The confounders adjusted in this study are based on previous literature on health service utilization and Andersen health service utilization model [23-25]. The selection principle of control variables is related to both core explanatory variables (SES) and dependent variables (inpatient services). According to Andersen model, individuals deciding whether or not to use health services are mainly influenced by three factors: predisposing characteristics, enabling resources and actual need for care. Based on previous studies, predisposing factors mainly include gender, age, marital status, number of children, and duration of migration etc. Enabling resources mainly include the individual's SES, ethnic group, health insurance status, health records, and hukou etc. Thus, control variables in this study include gender, age, marital status (married or single), number of children, ethnic group (Han or ethnic minority), establishment of health records, Hukou types (urban or rural), health insurance, movement area (across province, city or county), duration of migration, region (east, central or west), and willingness for long-term residence of more than 5 years (yes, no, and not decided yet). Types of health insurance were divided into four subgroups: no health insurance, having New Rural Cooperative Medical Scheme (NCMS), having Urban Employee Basic Medical Insurance (UEBMI), and having Urban Resident Basic Medical Insurance (URBMI). Movement area was categorized into three types: migration across provinces; migration across prefectural cities but within a province and migration across counties but within a prefectural city. All the control variables were available through the 2014 National Internal Migrant Population Dynamic Monitoring Survey. In order to avoid collinearity caused by the possible highly correlation between control variables and SES, regression diagnosis was carried out. The correlation matrix among SES and confounding variables have shown that the largest absolute value of the correlation coefficient was only 0.4, showing a weak correlation. Further, we calculated the variance inflation factors and found all the factors were no more than 2, a trivial amount, suggesting that the correlation between SES and covariates is not a serious problem.
Analytical methods
Data analyses were conducted by using the STATA 14.2 (Stata Corporation, College Station, TX, USA). Descriptive analyses were performed to compare the inpatient service utilization across different subgroups of the participants using t-test or chi-square test as appropriate and reported their p-values. Sample weights were applied in all the analysis to represent the China population. To avoid possible regional variation in this study, we controlled regional variable (east, central or west) in all regressions, and all standard errors were clustered at the strata level (by using STATA command svy). To be specific, based on the study design of this survey, all provincial urban belt and key cities were stratified, and 119 strata were finally determined. For example, Beijing was divided into seven strata, namely Chaoyang district, Haidian District, Fengtai District, Daxing District and etc. Thus, all standard errors in this study were clustered at the strata level in order to eliminate the variations within strata level.
First, we estimated the concentration index (CI) and constructed a concentration curve (CC) to illustrate inequity in unmet inpatient service need among migrants. The CC graphs the cumulative percentage of the sample on the x-axis, ranked by SES index, beginning with the lowest. CI was used to quantify the magnitude of inequity in unmet need and corresponds to twice the area between the CC and the 45° line [26]. CC runs from -1 (over-diagonal) to +1 (under-diagonal), indicating whether the unmet inpatient service need is concentrated among the low-SES (CI < 0), the high-SES (CI >0), or equally distributed (CI = 0) [27].
The concentration index CM was calculated by the following formula:
Then, we adopted logistic regression method to investigate the SES disparities in multivariate analyses adjusted for confounding variables. Those who received inpatient services were defined as the reference group. In model 1, we examined the association between SES and inpatient service utilization without control variables. In model 2, we controlled for potential confounding factors and estimated the adjusted odds ratio and the 95% confidence intervals. The model was specified as:
Finally, the decomposition of the gap in inpatient service use between the high and low SES migrants was assessed using the Blinder-Oaxaca (BO) decomposition method. The BO decomposition method was originally developed to explain wage gaps between whites and blacks and between men and women since the seminal work of Oaxaca and Blinder in the early 1970s [30, 31]. The BO decomposition [32]was a counterfactual method with an assumption that "what the probability of unmet inpatient service need would be if low SES migrants had the same characteristics as their high SES counterparts". In this part, SES was created using a median split with low SES categorized as below the median of SES index total score and high SES categorized as above the median. Based on it, the SES inequity was divided into two parts by using BO decomposition as followed:
