Bayesian Estimation of MIMIC Models Under Small-Sample Conditions and Latent Outlier Contamination

• In Event: AERA Roundtable Session Sunday 1:45 pm - Gold 3
  In Roundtable Session: RT1: Measurement and Related Issues in Quantitative Research Designs, Analysis, and Simulation

Abstract

Structural Equation Models (SEMs), particularly the Multiple Indicators Multi-
ple Causes (MIMIC) variant, are widely used in the social and behavioral sciences.
However, Maximum Likelihood (ML) estimation, default method for SEM, per-
forms poorly under small-sample conditions and non-normal data. This study evaluates
Bayesian estimation as a robust alternative, examining the impact of prior specifica-
tion on parameter recovery, coverage, power, and Type I error rates in MIMIC models
with latent outlier contamination (5%–20%) and small samples (30–200). Results
indicate that correctly specified informative priors yield the lowest bias, optimal
inference, while misspecified priors on factor loadings lead to severe estimation degra-
dation. Findings emphasize importance of careful prior selection, particularly for
measurement parameters, in small-sample SEM applications.

Author

• Nancy Alila, University of Georgia