<hw>Propensity Scores

Abstract Propensity scores represent the probability that an individual is assigned to a treatment group given the individual’s scores on a set of covariates. Propensity scores are used in nonrandomized studies to equate treatment and control groups on a large number of covariates measured at baseline. Assumptions necessary for propensity score analysis to yield a valid estimate of the causal effect are presented. The five steps involved in a propensity score analysis are reviewed: (a) selecting the covariates, (b) estimating the propensity scores, (c) conditioning on propensity scores, (d) checking covariates for balance, and (e) estimating the causal effect. Complexities including designs with more than two groups, covariates with missing values, multilevel designs, and sensitivity analysis to check for the potential effects of unmeasured covariates are briefly considered. Finally, a short evaluation of the performance of the propensity score approach is presented.