High-Dimensional Propensity Score Adjustment in Studies of Drug Treatment Effects Using Health Care Claims Data

• Activity Number: 20
• Type: Topic Contributed
• Date/Time: Sunday, August 2, 2009 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jeremy Rassen*+ and M. Alan Brookhart and Robert J. Glynn and Jerry Avorn and Helen Mogun and Sebastian Schneeweiss
• Companies: Harvard University and Brigham and Women's Hospital/Harvard Medical School and Brigham and Women's Hospital and Harvard University and Harvard University and Harvard University
• Keywords: Confounding adjustment ; machine learning ; pharmacoepidemiology ; emprical covariate ; identification ; propensity score

Abstract

Adjusting for large numbers of empirically-identified confounders may reduce bias in studies of medications. We developed an algorithm to make use of large numbers of covariates drawn from health care claims data. It identified candidate variables, ranked them by potential for bias, and then estimated propensity score and outcome models. We tested it in studies including one on gastric toxicity of Cox-2 inhibitors versus non-selective NSAIDs. In 49,653 patients, we observed a crude relative risk (RR) of 1.09 (95% CI: 0.91-1.30); adjusting for 15 standard investigator-identified covariates yielded a protective effect (RR=0.94; 0.78-1.12) which strengthened when further adjusting for 500 algorithm-identified covariates (RR=0.88; 0.73-1.06). In several highly-confounded drug studies, the proposed algorithm resulted in improved effect estimates as compared to standard adjustment practices.