Estimating Distributions and Ranks of Cluster-Specific Effect Parameters: New Bayesian Tools for Value-Added Modeling (Poster 16)

• In Event: AERA Poster Session Saturday 11:45 am
  In Invited Poster Session: Excellence in Education Research: Early Career Scholars and Their Work Poster Session

Abstract

Standard random-effects models and empirical Bayes shrinkage often fail to address key inferential goals—such as estimating effect distributions, identifying outliers, or establishing reliable rankings—especially when effect sizes correlate with their standard errors or distributions deviate from normality. This project develops new Bayesian tools for value-added modeling, motivated by evaluating 1,524 Alabama pre-K classrooms. First, it proposes flexible deconvolution methods capturing non-normal, precision-dependent effect distributions. Second, it tailors posterior predictions to ranking goals, creating "report card" measures aligned with policymakers' loss functions. Third, it extends these techniques to multivariate settings, jointly modeling multiple outcomes to improve precision and reveal cross-domain relationships. These methods are broadly applicable to any setting requiring identification of high- or low-performing units.

Author

• Joon-Ho Lee, University of Alabama