Exploration of structural and statistical biases in the application of propensity score matching to pharmacoepidemiologic data

Abstract

Certain pitfalls associated with propensity score matching have come to light, recently. The extent to which these pitfalls might threaten validity and precision in pharmacoepidemiologic research, for which propensity score matching often is used, is uncertain. We evaluated the “propensity score matching paradox” – the tendency for covariate imbalance to increase in a propensity score-matched dataset upon continuous pruning of matched sets – as well as the utility of coarsened exact matching, a technique that has been posed as a preferable alternative to propensity score matching, especially in light of the “propensity score matching paradox”. We show that the “propensity score matching paradox” may not threaten causal inference that is based on propensity score matching in typical pharmacoepidemiologic settings to the extent predicted by previous research. Moreover, even though coarsened exact matching substantially improves covariate balance, it may not be optimal in typical pharmacoepidemiologic settings due to the extreme loss of study size (and resulting increase in bias and variance) that may be required to build the matched dataset. Finally, we explain variability in 1:1 propensity score matching without replacement as well as methods that were developed to account for this variability, with application of these methods to an example claims-based study.