Comparing Propensity Score Weighting Schemes with CBPS and XGBoost in Causal Effect Estimation
Propensity score is the conditional probability of an individual being in the treatment group based on its pre-treatment covariate values (Rosenbaum & Rubin, 1983a). It can be used to equate the treatment and control groups on distribution of covariates to accurately estimate treatment effect in a non-randomized design. There are several methods to estimate the propensity scores. Traditionally, it is estimated with logistic regression, which requires a correctly specified propensity score model. With the advance of machine learning techniques, non-model-based methods such as XGBoost (Chen & Guestrin, 2016) has been the state-of-art algorithm in propensity score estimation. Another propensity score estimation method that is robust to model misspecification is the covariate balancing propensity score (CBPS) method (Imai & Ratkovic, 2014), which models the propensity score while optimizing the covariate balance using the generalized method of moments (GMM) framework. This study compared logistic regression, XGBoost and CBPS as propensity score estimation methods and evaluated their performance with five different propensity score weighting schemes under various conditions of covariates and outcome model complexities. A large comprehensive simulation study was conducted to achieve this goal.
Teng, Yue, "Comparing Propensity Score Weighting Schemes with CBPS and XGBoost in Causal Effect Estimation" (2023). ETD Collection for Fordham University. AAI30486229.