The Role of PS Estimation and Implementation in the Presence of Heterogeneous Effects of Measured and Unmeasured Covariates on Treatment

• Activity Number: 22
• Type: Topic Contributed
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Til Stürmer*+ and Richie Wyss and Mark Lunt
• Companies: The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill and University of Manchester
• Keywords: propensity scores ; instrumental variables ; interaction ; confounding ; sensitivity analysis ; trimming

Abstract

Propensity scores (PS) can be misspecified e.g., due to omittance of interactions between measured covariates and unmeasured confounders leading to treatment contrary to prediction (TCTP). We simulated a dichotomous treatment, Poisson outcome, two confounders, and an instrumental variable. Predictors of treatment were either dichotomous or normally distributed. PSs were estimated including confounder main effects only and implemented using inverse probability treatment weighting (IPTW) and matching. For TCTP we simulated a strong unmeasured confounder at the tails of the PS based on measured covariates. For scenarios with continuous predictors of treatment, the percentage bias ranged from 4%-61% for omitted confounder interactions, and from 1%-20% for IPTW and 1%-8% for matching with confounder-instrument interactions. With dichotomous covariates, effect estimates were biased only with omitted confounder interactions and IPTW. Trimming of those TCTP reduced bias in most scenarios assessed. The omittance of interaction terms in PS models generally, but not always, leads to bias. Trimming of those TCTP can be used as a sensitivity analysis to reduce bias from unmeasured confounders.