For descriptive and comparative analysis between public and private providers, univariate, bivariate and multivariate analyses were performed using SPSS for Windows, version 11.0 and Stata 9.
Comparing providers of different ownership forms is inherently complicated by the differences in their characteristics. For example, smaller providers may have systematically lower performance if there are economies of scale and scope in performance, and private providers tend to be systematically smaller than government-owned providers. If mean performance scores are lower for private CHS than public CHS, to what extent is this difference because private CHS are smaller, rather than because they are private? To answer such questions, multivariate regressions that control for potentially confounding factors are useful.
We run descriptive regressions of the following form:
is the dependent variable of interest, such as the performance score for provider i in year t; λ is a constant; X represents a vector of observed characteristics of the provider; β
is the set of estimated coefficients for the characteristics in vector X; Private
denotes private ownership; Weifang
is a dummy variable equal to 1 if provider i is located in Weifang, rather than City Y; and u
denotes idiosyncratic errors. The primary estimated coefficient of interest is β
, representing the difference in the dependent variable between private and public providers in Weifang, compared to β
for City Y, controlling for other observable characteristics. If the sum of the estimated coefficients for Private
and Private * Weifang is statistically indistinguishable from zero, then there is no statistically significant difference in performance between public and private CHS in Weifang, after controlling for the fact that smaller CHS tend to have lower performance scores in both Weifang and City Y. Data limitations preclude any multi-year analyses; all our regressions are cross-sectional.
For example, consider our analysis of factors associated with CHS performance scores in 2009. In this analysis, t = 2009 and y
is the 2009 performance score for CHS i; the X vector includes such characteristics of the CHS as number of beds, number of staff, fixed assets, a dummy variable for being in the social health insurance network as an appointed provider, and a dummy variable for implementing the policy of separating prescribing and dispensing.
One empirical challenge is that several of the CHS characteristics are correlated (e.g., providers with more beds also have more staff and are more likely to be included in the insurance network as appointed providers). To enable qualitative statements about CHS of different ownership form with otherwise similar characteristics, we present results for multiple specifications of each regression, alternatively including different sets of explanatory variables. We confine our discussion of the primary estimate of interest—β
—to the case in which the estimate is relatively unchanged across these difference specifications. Stability of the estimate across different empirical specifications reflects robustness to different ways of controlling for the observable differences among providers.
These multivariate regressions are not intended to constitute an impact evaluation of the contracting reforms in Weifang. Several limitations of the available data preclude a research design that could evaluate impact or disentangle causality. For example, we lack baseline data for pre-reform trends in Weifang and the comparison group; and two years is probably too short a time frame for an evaluation of most dimensions of provider performance. Rather than an impact evaluation, the analyses document differences between ownership forms, controlling for observable factors (such as size). Such analyses enable statements of the form “public and private CHS differ in size and staffing, but once we control for these differences, public and private CHS no longer statistically differ in their performance scores in Weifang in 2009.”
We also use the survey data of CHS personnel to explore the factors associated with staff satisfaction and performance scores of community health stations in Weifang and City Y. In those regressions, in addition to the performance score y
for t=2009, we define a second dependent variable to be equal to 1 if the staff member reports being satisfied with his or her job. For the latter limited dependent variable, we use a logit analysis. The X vector of observed characteristics of staff member i includes a dummy variable if male; age and aged squared; years of work experience at the CHS; and dummy variables for educational attainment and job position (i.e., general practitioner, nurse, manager, pharmacist, technician).
A third and final set of multivariate analyses describes the association between resident characteristics and their utilization of CHS services. Suppose that we find, as we do, that on average more residents served by private CHS than public CHS say they are willing to do routine examinations at the CHS. Is this because private CHS provide better quality services than their public counterparts do, or because the residents in private-CHS communities tend to use community services more often than hospital outpatient departments? To address this question, we run logistic regressions of the following form:
The dependent variable Y
is a dichotomous variable about the health service utilization of individual i. For example, to estimate the probability of being served by a privately owned provider, Y
is equal to 1 if the resident lives in a community served by a private CHS; to estimate the probability that the resident is willing to use the community provider (rather than go to a hospital outpatient department), Y
is equal to 1 if the resident says he or she is willing to do a routine health exam at the neighborhood CHS, or is willing to visit the CHS for first-contact care when feeling ill. The dichotomous explanatory variables X
include male; uninsured; self-reported poor health status; low education; and low income (i.e., monthly income below 800 RMB). These regressions allow us to see whether utilization preferences are systematically associated with the individual socioeconomic characteristics of community residents. Standard errors are clustered at the CHS level.