Matching Strategies for Observational Studies with Multilevel Data in Educational Research
This project seeks to develop, test, and disseminate matching strategies for causal inference with observational multilevel data in educational research. Matching strategies for non-equivalent control group designs with multilevel data are required since randomized experiment cannot always be conducted—due to ethical reasons, for instance. In order to remove selection bias and draw causal conclusions from observational data, matching techniques like propensity score (PS) matching or Mahalanobis distance matching have gained increasing popularity during the last two decades.
The project investigates different within- and across-cluster matching strategies for estimating causal treatment effects across and within clusters. In using theoretical investigations, simulation studies, within-study comparisons, and analyses of ECLS–-K data, we pursue two main goals. First, determine the conditions under which within- and across-cluster matching strategies produce consistent estimates of average treatment effects. Second, find which matching strategies and analytic approaches will work best in educational research practice. In evaluating the matching strategies, we will also probe different PS estimation methods (fixed vs. random effects models) and various matching techniques including PS matching, inverse-propensity matching, PS stratification, or Mahalanobis distance matching.