Application of a propensity score approach for risk adjustment in profiling multiple physician groups on asthma care

Health Services Research,  Feb, 2005  by I-Chan Huang,  Constantine Frangakis,  Francesca Dominici,  Gregory B. Diette,  Albert W. Wu

• E-mail
• Print
• Link

Analytic Framework for Comparing Different Risk-Adjustment Methods

We adjusted for exogenous factors, i.e., factors for which providers have no influence (mainly patient characteristics, such as age, sex, education, baseline severity, etc.). We did not include race in the risk-adjustment model. Evidence suggests that African-American patients may receive poorer quality of care than white patients (Kahn et al. 1994). If this makes patient satisfaction across physician groups differ because of different race distribution, these are differences we want to capture, and adjusting for race here would mask them. In this case, examining patient satisfaction separately within race groups would highlight such inequalities (Iezzoni 1997). Moreover, we did not adjust for endogenous factors, i.e., factors that providers can affect (mainly physician group characteristics, such as physician group specialty, number of supplementary staff, etc.) (Welch, Black, and Fisher 1995). Adjusting for endogenous factors may mask true performance of physician groups because these factors can influence the patient outcomes.

We evaluated patient satisfaction among physician groups using two analytic methods, thus also assessing sensitivity of the results to different risk-adjustment approaches (Table 1). We used the first physician group as the reference group for comparisons among different methods.

For method 1, we implemented a hierarchical outcome regression model without propensity scores. At the first stage (patient level) we used a logistic regression model for estimating the group-specific log odds ratio (OR) of patient satisfaction (outcome) as a function of patient characteristics, including age, sex, education level, type of insurance, prescription drug coverage, asthma severity, number of comorbidities, and health status. At the second stage (group level), we modeled the variation of the log OR across 20 physician groups. The hierarchical outcome regression approach takes into account clustering of patients within physician groups and the different number of patients within each physician group (reliability). Under the hierarchical outcome model the group-specific estimates of performance are shrunk toward an average performance common to all physician groups to address the regression-to-the mean that arises with comparison of multiple groups (Morris 1983; Christiansen and Morris 1997; Sullivan, Dukes, and Losina 1999). The hierarchical outcome regression model is detailed in Appendix 1.

With this method, the relative performance of physician groups was assessed by estimating the risk-adjusted OR of satisfaction with care (greater versus less satisfaction) attributable to the jth physician group relative to the first physician group (reference group) by exponentiating the difference between the estimated provider-specific random intercept of the jth (j = 2, ..., 20) and the first physician group (DeLong et al. 1997; Katon et al. 2000).