Multidimensional Fully Latent Principal Stratification

• In Event: Thursday Roundtable Session 9:00 am
  In Roundtable Session: AI, Psychometrics, and Data Analysis (Table 24)

Abstract

This research presents a multidimensional principal stratification (MFLPS) framework for inference on latent measurement variables in randomized control trials (RCTs), extending a unidimensional framework and exploring its performance through Monte Carlo simulation. The aim is to enhance the analysis of rich, multivariate data derived from RCTs, especially in the context of computer-based educational interventions. Preliminary simulation results underscore the crucial role of factor reliability and sample size for accurate estimation of principal effects. With a sample size of at least 500, MFLPS correctly identified a medium effect size in over 80% of cases. This study's findings provide valuable guidelines for researchers interested in RCTs, offering insights into MFLPS application and determining optimal conditions for its use.

Authors

• Sooyong Lee, WIDA

• Adam C. Sales, Worcester Polytechnic Institute

• Hyeon-Ah Kang, University of Texas at Austin

• Tiffany A. Whittaker, University of Texas at Austin