Comparisons of Various Techniques in Propensity Score Estimation Using Simulation

• Activity Number: 134
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Jiaxiao M Shi*+ and Wansu Chen
• Companies: Kaiser Permanente and Kaiser Permanente
• Keywords: propensity score ; logistic regression ; GAM ; GBM ; CART ; NNET

Abstract

Propensity score (PS) method has been increasingly used to address confounding issues when evaluating causal effects. The score is often estimated using the logistic regression model (LR), an approach that is not convenient for handling non-linear relationships between the exposure and covariates/confounders or interactions. Several non-parametric methods have been suggested as alternative techniques to estimate PS. They include generalized additive models (GAM), classification and regression trees (CART), generalized boosted models (GBM) and neural networks (NNET). Simulation was conducted to compare the performance of LR, GAM, CART, GBM, and NN under varying degrees of non-linearity between the exposure and covariates, outcome rates and conditional odds ratios (OR). PS was applied into the outcome models by using regression adjustment and inverse probability of treatment weighting (IPTW). Our results showed that CART had the largest bias in most scenarios among the five techniques we evaluated except for models with interaction terms. The biases for all other techniques were comparable. When interactions existed, GBM seemed to perform the best (smallest bias and MSE).