Causal Inference for Measures of Health Disparities

Abstract

There is increased interest in the evaluation of health disparities between different socioeconomic groups using data from observational studies. However, in the absence of randomization, the results and conclusions may be limited to associations rather than causal effects. The causal inference framework allows us to estimate causal measures for such situations. Using generalized propensity scores, we introduce inverse probability weighting (IPW), doubly-robust (DR) and covariate-adjustment (CA) estimators for the vector of the marginal means of the distributions of the potential outcomes corresponding to multiple socioeconomic groups. We estimate the variance of the IPW, DR and CA estimators using an M-estimation approach. The variances of the estimators for the causal measures of health disparities are subsequently estimated using the multivariate delta method, and 95% confidence intervals (CIs) are constructed accordingly. A bootstrap method to construct the 95% CIs is also considered for comparison purposes. In simulation studies, the 95% CIs based on the analytical method had empirical coverage probabilities close to the nominal level and had advantages over the bootstrap-based 95% CIs under certain scenarios. We illustrate the proposed methods using a real data set from the National Health and Nutrition Examination Survey I Epidemiologic Follow-up Study.