Variance Priors for Bayesian Multilevel Models
Mode-matched inverse-gamma priors for variance components when higher-level sample sizes are limited
Published online · 2026 Multivariate Behavioral Research Liu, L.
When higher-level sample sizes are small, variance components can be weakly identified and prior choices can materially affect Bayesian multilevel estimates. This study introduces a transparent inverse-gamma construction that matches the prior mode to an interpretable variance target while keeping prior strength explicit.
Abstract
The paper proposes a mode-matched inverse-gamma strategy for variance components in Bayesian multilevel models. A substantively plausible or empirically grounded variance value is used as the target prior mode, while a separate shape parameter controls prior concentration. The method was evaluated in a Monte Carlo study and in a disjoint reference-sample illustration using TIMSS 2019 U.S. Grade 8 data. The broader goal is not to prescribe a universal prior, but to make prior location, prior strength, calibration source, and sensitivity explicit.
Paper
Liu, L. (2026). Mode-matched inverse-gamma priors for variance components in Bayesian multilevel models. Multivariate Behavioral Research. Advance online publication. Read the published article →
Published online October 21, 2026.
Supplement
The published article includes online supplementary material with mathematical derivations, Mplus model specifications, extended simulation results, and TIMSS sensitivity analyses. View supplementary material through the journal →
Code
Mplus prior syntax and model specifications are documented in Supplement S2, available through the published article. A separate public code repository is not currently linked here.
Simulation Design
Crossing these factors produced 162 two-level random-intercept conditions, with 200 replications per condition. Seven estimation approaches compared maximum likelihood, Bayesian baseline priors, IG(0.01, 0.01), mode-matched inverse-gamma priors, and corresponding fixed-effect prior benchmark combinations. The study included both an idealized true-mode benchmark and a disjoint reference-sample calibration illustration.
generated simulation datasets
non-overlapping TIMSS reference and target samples
bias, RMSE, interval coverage, and interval width
variance components versus fixed effects
Key Findings
- Mode-matched priors improved variance-component estimation in the benchmark conditions. Gains were most evident for between-level residual variance with small or moderate numbers of clusters, including lower RMSE and mean absolute relative bias than the Bayesian baseline and IG(0.01, 0.01) specifications.
- Reference-informed calibration helped, but was not automatically optimal. In the disjoint reference-sample illustration, the calibrated prior reduced RMSE and interval width relative to the baseline, while relative-bias performance was less favorable when the number of clusters was very small.
- Fixed-effect point estimates were comparatively stable across Bayesian prior choices. Prior sensitivity was concentrated more strongly in variance components than in fixed effects.
- Calibration and sensitivity analysis are part of the method. The results support documenting where prior information comes from, separating reference and target data when possible, and checking how conclusions change across plausible prior strengths.
Contribution to my research program
This paper anchors my broader work on Bayesian calibration under limited information. Related projects extend prior calibration to multilevel item-response models, software translation, random-effect covariance structures, and applied guidance for selecting and reporting variance priors.