The use of propensity scores in pharmacoepidemiologic research

PURPOSE: To describe the application of propensity score analysis in pharmacoepidemiologic research using a study comparing the renal effects of two commonly prescribed non-steroidal anti-inflammatory drugs (NSAIDs). METHOD: Observational data were collected on the change in renal function, as measured by serum creatinine concentration, before and after use of two NSAIDs, Ibuprofen and Sulindac. To estimate the treatment effect of the different NSAIDs, we used the propensity score methodology to reduce the potential confounding effects caused by unbalanced covariates. After estimating the propensity scores (the probabilities of each patient being prescribed Sulindac) from a logistic regression model, we stratified the data based on sample quintiles of the propensity score distribution. The final estimate of the treatment effect was then obtained by averaging the treatment estimates from the stratified samples. RESULTS: Initially, 23 covariates differed significantly between the two treatment groups. Using the propensity score methodology, we were able to balance the distributions of 16 covariates. The imbalances in the remaining seven covariates were also greatly reduced. Although the use of either drug resulted in a decrease in renal function, overall differences between them were not statistically significant with respect to their effect on creatinine concentrations based on the propensity score analysis. CONCLUSION: Observational studies often produce treatment groups that are not directly comparable due to imbalances in covariate distributions between the treatment groups. Propensity score analysis provides a simple and effective way of controlling the effects of these covariates and obtaining a less biased estimate of the treatment effect. Copyright (c) 2000 John Wiley & Sons, Ltd.