Online Program

How to use Propensity Scores to Strengthen Estimates of Treatment Effects: A Guided Tour of the Propensity Score Landscape for Real-world Analysts (302608)

 View Presentation

Keywords: Propensity scores, hierarchical linear modeling, observational studies

In the absence of available experimental data, analysts interested in treatment effects must resort to nonexperimental (observational) data. To reduce bias in treatment effect estimates, analysts may condition their observed data based upon calculated propensity scores (PS) in an effort to statistically approximate randomization. A PS estimates the probability, given a set of known covariates, an observed unit of analysis would be found within a specific treatment condition in the study. To effectively utilize PS, however, analysts must make some important methodological choices including the selection of covariates, estimation of the PS model, evaluation of balance achieved, and conditioning methods. PS conditioning methods include using PS to match units across conditions, to stratify units into comparable groups, to derive weights for a weighted analysis, or to create a single composite covariate for an ANCOVA. These methodological choices are particularly pressing when nested structures are present in the dataset. In this poster, the authors provide some practical suggestions for analysts based upon the results of their recent simulation work.