Performance of Propensity Scores in the Analysis of Rare Events

• Activity Number: 452
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jessica M. Franklin*+ and Sebastian Schneeweiss
• Companies: Brigham & Women's Hospital/Harvard Medical School and Brigham & Women's Hospital/Harvard Medical School
• Keywords: propensity score ; pharmacoepidemiology ; logistic regression ; software ; boosting

Abstract

The propensity score (PS) is an important tool for confounder adjustment in nonrandomized studies. However, it was developed for estimation of treatment effects on continuous outcomes, and little is known about its performance in the common epidemiologic scenario of many confounders and few binary outcome events. To evaluate analytic performance, we utilized an existing cohort study of incident exposure to anticonvulsant medications. Patients were followed for cardiovascular events for 90 days; no biologically plausible treatment effect exists within that timeframe, so treatment effect estimates were compared to an assumed gold standard hazard ratio of 1.0. We found that PS estimation and subsequent treatment effect estimation became highly unstable as variables were added sequentially to the PS model, as small changes in PS values had large impacts on treatment effect estimates. In particular, differing convergence criteria in the logistic regression implementations in SAS and R led to differing results with just 35 variables in the PS model, despite more than 12,000 exposed patients. Boosted regression for PS estimation led to more stable results.