Selecting, Translating, and Reporting Variance Priors
A practical workflow for scalar variance components in Bayesian multilevel models
Published Behavior Research Methods Liu, L.
Variance-prior decisions are often treated as a software detail. This tutorial reframes them as a sequence of transparent modeling decisions linking the target variance component, parameter scale, prior family, benchmark information, software implementation, checking, sensitivity analysis, and reporting.
Abstract
This tutorial provides a practical workflow for specifying scalar variance-component priors in Bayesian multilevel models. It reviews gamma, inverse-gamma, half-t, and half-Cauchy priors and clarifies how prior meaning changes across variance, precision, and standard-deviation parameterizations. A worked TIMSS 2019 example uses benchmark information from a non-overlapping reference sample to construct alternative priors for a smaller target analysis. The emphasis is on making the rationale, translation, checking, and robustness of prior choices visible rather than treating any default as automatically neutral.
Paper
Liu, L. (2026). A tutorial on selecting, translating, and reporting variance priors in Bayesian multilevel models. Behavior Research Methods, 58, Article 328.
Supplement
Supplementary tables and figures accompany the published article. They can be accessed through the published article. Reproducibility materials are also available in the public OSF repository.
Code
Reproducibility code and supplementary materials for the TIMSS 2019 tutorial example are publicly available on OSF →.
Practical Workflow
The published tutorial organizes variance-prior construction into five steps:
Determine which variance component needs a prior and whether the estimation software specifies it on the variance, precision, or standard-deviation scale.
Anchor plausible magnitudes in prior evidence or a suitable reference sample, and make the intended prior concentration explicit.
Use gamma and inverse-gamma relationships to connect benchmark information to interpretable variance- or precision-scale specifications.
Construct software-compatible specifications, including half-t and half-Cauchy alternatives, using an explicit calibration target such as a prior median.
Evaluate prior densities, prior predictive implications, posterior comparisons, convergence, and sensitivity to reasonable alternative choices.
Worked Example: TIMSS 2019
A two-level model of science achievement illustrates the workflow using a non-overlapping reference sample (7,408 students in 233 schools) and a smaller worked-example sample (400 students in 40 schools). The comparison examines priors for the between-school random-intercept and within-school residual standard deviations. Gamma, inverse-gamma, half-t, and half-Cauchy priors are calibrated to common benchmark medians on the standard-deviation scale and then evaluated using prior predictive checks, posterior comparisons, diagnostics, and sensitivity analysis.
Practical Safeguards
Check where mass is concentrated before fitting the target model.
Ask whether the prior implies plausible data-generating behavior.
Compare conclusions under alternative defensible specifications.
Make the prior reconstructable by another analyst.
Key Takeaways
- A numerical prior specification is not portable unless the parameter scale and software parameterization are also understood.
- Benchmark-informed calibration is most defensible when the source of information and its separation from the target analysis are explicit.
- Sensitivity analysis is part of prior specification, not a final cosmetic check.
- The goal is not to identify one universal prior family, but to make prior assumptions interpretable, inspectable, and reproducible.
Contribution to my research program
This tutorial complements my simulation-based research on variance-prior performance by providing an applied workflow for calibrating, checking, and reporting priors across model parameterizations and software environments.