Estimation of causal effect using propensity score and weighted-average method

Abstract

Abstract Although many studies focus on causal relationships in causal and consequential variables, relationships between variables do not necessarily hold for only two variables regarding cause and effect. Thus, attempting to determine the effect of the cause on a result empirically requires the removal of as many confoundings as possible that affect both the cause and the result. The effect of a cause on an outcome without the effects of the confoundings can be empirically analyzed via the weighted average method where propensity scores are estimated using confoundings. A typical weighted average method is the inverse probability weights (IPW) estimator. However, the IPW estimator has two problems: the estimation of causal effects is biased when the propensity score estimation model is not correct (Problem 1), and the estimation is susceptible to extreme propensity scores (Problem 2). The doubly robust (DR) estimator and the overlap weighting (OW) estimator have been proposed to address problems 1 and 2, respectively. However, the DR estimator cannot address problem 1, and the OW estimator cannot address problem 2. Hence, using the propensity score and weighted-average method, this study compares the DR and OW estimators and proposes a new estimator to solve problems 1 and 2. This study confirms the usefulness of the proposed method by means of a simulation and applies it to real data to ground the relevance of the findings.