The Use of Propensity Score Methods to Address Confounding by Provider Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 19, 2019
  • Hammill, Bradley
    • Affiliation: Gillings School of Global Public Health, Department of Biostatistics
  • For research questions regarding the real-world effectiveness and safety of medical therapies and devices, researchers must often rely on observational data. Unlike controlled clinical trials, the assignment of treatment to patients in routine medical practice is not randomized. One class of methods used extensively by researchers to address this selection problem is propensity score methods. The role of the healthcare provider has not typically been accounted for when propensity score methods are employed, despite the fact that provider, by imparting an effect on both patient-level treatment assignment and patient-level outcomes, is a potential confounding factor. When a healthcare provider has measurable impacts on both a patient’s treatment assignment and their downstream outcomes, simulation results demonstrated that not accounting for these provider effects could lead to biased estimates of treatment effect when using propensity score methods. This was true specifically when a provider’s direct effect on treatment was correlated with their effect on outcome; a situation that occurs when providers having better patient outcomes use therapies at higher (or lower) rates than other providers. Propensity score methods that incorporated provider were able to control this error. Even when provider effects on treatment and outcome were uncorrelated, it was still important to account for provider in the propensity score treatment model. Failure to do so resulted in confidence intervals around the estimated treatment effect that were either substantially too wide or too narrow, depending on the estimation methods used. A criticism of typical 1:1 propensity score matching, whether stratified by provider or not, is that the data from many patients are not utilized in the outcomes analysis. Full matching addresses this issue by optimally assigning all treated patients and all comparison patients into variably-sized matched sets. The result is closer matches between study groups than those obtained by other matching methods. Full matching is not currently utilized frequently because it is difficult to implement. A macro to perform full matching by leveraging SAS optimization procedures is presented.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Maciejewski, Matthew
  • O'Brien, Sean
  • Herring, Amy
  • Koch, Gary
  • Preisser, John
  • Doctor of Public Health
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2015
Place of publication
  • Chapel Hill, NC
  • There are no restrictions to this item.

This work has no parents.