Simulation Study Comparing Inverse Probability of Treatment Weighted (IPTW) Estimator and Multiple Regression Model

• Activity Number: 540
• Type: Contributed
• Date/Time: Thursday, August 2, 2007 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Jianmin Wang*+ and William Irish
• Companies: RTI Health Solutions and RTI Health Solutions
• Keywords: Marginal Structure Model ; Simulation ; inverse probability of treatment weighted (IPTW) estimator ; nonrandomized study ; Multiple regression model

Abstract

Bryan and colleagues (Biostatistics, 2004) compared naïve and IPTW estimators of treatment effect using simulation when treatment assignment is related to confounders. The authors concluded that IPTW produces an unbiased estimate of treatment effect whereas a naïve method is biased. Upon further review of the methods used, we found two serious limitations: 1) The naïve estimator is based on a mis-specified model; and 2) Response variable is determined as a function of observed covariates, and therefore, has no random component. Using more appropriate methods, we conclude that the estimator based on a correctly specified model is asymptotically unbiased whereas the IPTW estimator is biased. This conclusion is in direct contrast to that reported by Bryan et al, and has important implications for the use of IPTW in the setting of nonrandomized studies.