A Bayesian solution to non-convergence of crossed random effects models

Huang, Mingya

Permalink

https://hdl.handle.net/2142/106411 Copy

Description

Title

A Bayesian solution to non-convergence of crossed random effects models

Author(s)

Huang, Mingya

Issue Date

2019-07-22

Director of Research (if dissertation) or Advisor (if thesis)

Anderson, Carolyn Jane

Committee Member(s)

• Kern, Justin
• Zhang, Jinming

Department of Study

Educational Psychology

Discipline

Educational Psychology

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

Crossed-Random Effects Models, convergence, MLE, REML, Bayesian model estimation

Abstract

When observations (trials) are nested within combinations of subjects and stimuli, crossed random effects models simultaneously take into account both fixed effects and random effects of the subjects and stimuli; however, maximum likelihood estimation (MLE) and restricted maximum likelihood (REML) estimation often encounter convergence problems, which in turn lead to researchers fitting simpler models (e.g., only random intercepts). If the random effect structure is too simple, tests of fixed effects are not valid; if the random effect structure is too complex, tests of fixed effects are inefficient. This study examines issues of estimation, convergence and problems inherent with MLE and REML. We investigated whether Bayesian estimation can solve the convergence problem through a simulation study, which makes a case for adopting a Bayesian approach for estimation, especially when using crossed random effects models. In our simulation study were found that both MLE and REML encountered convergence problems even when trying to fit the correctly specified model and simpler versions of it. Bayesian estimation of models converged 100% of the time, and for the correctly specified model, the parameter estimates were accurate estimates for both fixed and random effects and were essentially unbiased. In sum, the Bayesian approach is a viable alternative to MLE/REML, because models fit by Bayesian estimation solves the non-convergence problem and yields valid and efficient estimates.

Graduation Semester

2019-12

Type of Resource

text

Permalink

http://hdl.handle.net/2142/106411

Copyright and License Information

Copyright 2019 Mingya Huang

Owning Collections