Longitudinal Propensity Score Matching: A Demonstration of Counterfactual Conditions Adjusted for Longitudinal Clustering

Objectives Given the challenges of conducting experimental studies in criminology and criminal justice, propensity score matching (PSM) represents one of the most commonly used techniques for evaluating the efficacy of treatment conditions on future behavior. Nevertheless, current iterations of PSM fail to adjust for the effects of longitudinal clustering on participant exposure to treatment conditions. The current study presents and evaluates longitudinal PSM (LPSM) as an alternative method for assessing the effects of a treatment condition on future behavior. LPSM adjusts for the effects of longitudinal clustering (i.e., clustered error) by assuming that the association between a cross-sectional predictor and a treatment condition varies depending upon the time at which the treatment was administered. Methods Two general steps were taken to evaluate the validity of LPSM. First, we conducted a series of simulation analyses to illustrate the LPSM method. Second, we further demonstrate the method using data from 63,899 inmates incarcerated in Ohio prisons by assessing the effects of prison programming on recidivism over a three-year post-release period. Disparities in treatment effects were compared between cross-sectional PSM and LPSM. Results The simulation and demonstration analyses produced evidence of disparities in results between LPSM and cross-sectional PSM. LPSM appeared to provide the superior adjustment for longitudinal clustering relative to cross-sectional PSM. Conclusions LPSM provides a useful alternative to cross-sectional PSM when the probability of exposure to a treatment condition varies by the time at which the treatment was administered.