Online Program

Return to main conference page
Thursday, February 15
PS1 Poster Session 1 and Opening Mixer Thu, Feb 15, 5:30 PM - 7:00 PM
Salons F-I

Limitations of Propensity Score Methods: Demonstration Using a Real-World Example (303687)

View Presentation View Presentation

David T. Bearden, Oregon State University/Oregon Health & Science University College of Pharmacy 
Miriam R. Elman, OHSU/OSU College of Pharamcy 
Yoojin Kim, Oregon Health & Science University 
Kevin Langstaff, Oregon Health & Science University 
James S. Lewis II, Oregon Health & Science University 
Jessina C. McGregor, OHSU/OSU College of Pharamcy 
*Gregory B. Tallman, Oregon State University/Oregon Health & Science University 

Keywords: confounding, propensity scores, observational research, inverse probability of treatment weighting

Propensity score methods are used in comparative effectiveness research as a tool to control for confounding in observational studies, but details about propensity score diagnostics are often absent. We describe the failed application of propensity score methods in a cohort study evaluating the effectiveness of various antibiotics for prevention of surgical site infection. Multivariable logistic regression was used to construct a propensity score and diagnostics such as common support and standardized differences were used to evaluate it. Outcome frequencies were compared between treatment groups using inverse probability of treatment weighting (IPTW) to adjust for potential confounders. For comparison, a multivariable model was constructed using forward selection to adjust for confounders; a 20% change in odds ratio was the threshold for entry. Results showed discrepancies between IPTW and multivariable models. Propensity score diagnostics revealed limited common support and large standardized differences for most covariates. Propensity score methods may not be appropriate for all studies and careful evaluation of diagnostics is crucial to ensure validity of result interpretation.