The Effect of Having a Regular Doctor as a Primary Care Provider on Emergency Room Utilization in South Korea

Lee, Su-Young; Lim, Hyeong-Seok

In observational studies such as the current study, differences in the distribution of observed baseline covariates between treatment and control groups may occur. (Table 1 ) This selection bias prevents justifiable causal inferences without controlling for the bias. To estimate the causal effect by adjusting for unbalanced confounding covariates between treatment and control groups, this study used a counterfactual framework (also known as potential outcome framework) and propensity score (PS) analysis. 11, 12) In the PS analysis, the probability of treatment conditional on baseline confounding covariates was first estimated and these conditional probability estimates were used for estimation of treatment effect. 13, 14) We estimated potential-outcome means (POMs) and average treatment effect (ATE) in the general population, patients with hypertension, and those with diabetes mellitus (DM) using an augmented inverse-probability-weighted (AIPW) estimator that combined elements of regression adjustment on PS and PS weighting estimation among several PS analysis methods. 15) In the counterfactual framework, each subject had two potential outcomes.