Identifying Baseline Covariates for Use in Propensity Scores: A Novel Approach Illustrated for a Non-randomized Study of Recovery High Schools

摘要

There are many situations where random assignment of participants to treatment and comparison conditions may be unethical or impractical. This article provides an overview of propensity score techniques that can be used for estimating treatment effects in non-randomized quasi-experimental studies. After reviewing the logic of propensity score methods, we call attention to the importance of the strong ignorability assumption and its implications. We then discuss the importance of identifying and measuring a sufficient set of baseline covariates upon which to base the propensity scores and illustrate approaches to that task in the design of a study of recovery high schools for adolescents treated for substance abuse. One novel approach for identifying important covariates that we suggest and demonstrate is to draw on the predictor-outcome correlations compiled in meta-analyses of prospective longitudinal correlations.