Bayesian ANCOVA
Bayesian Analysis of Covariance · Also known as: Bayesian ANCOVA, Bayesian analysis of covariance, B-ANCOVA, Bayesian covariate-adjusted group comparison
Bayesian Analysis of Covariance (Bayesian ANCOVA) extends classical ANCOVA by placing prior distributions on group effects and covariate slopes, then updating them with observed data to obtain posterior distributions and Bayes factors. It quantifies evidence for group differences on a continuous outcome after statistically adjusting for one or more continuous covariates, without relying on p-value thresholds.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian ANCOVA when you want to compare two or more groups on a continuous outcome while controlling for continuous covariates, and you need to quantify evidence rather than just test significance. It is particularly valuable when sample sizes are small or moderate (where p-values are unreliable), when you wish to accumulate evidence across studies, or when you want to distinguish between absence of evidence and evidence of absence. Do not use it as a drop-in replacement without understanding priors — results depend on prior choice and should be reported with sensitivity analyses. If covariates are categorical rather than continuous, use Bayesian factorial ANOVA or Bayesian MANCOVA instead. The homogeneity-of-regression-slopes assumption (covariate slope the same across groups) still applies and should be tested.
Strengths & limitations
- Provides a Bayes factor that quantifies evidence for both the alternative and the null hypothesis, unlike the p-value.
- Returns full posterior distributions, enabling richer inference including credible intervals and probability statements about effect sizes.
- Can incorporate prior knowledge from previous studies, improving estimation in small samples.
- Sequential updating: data can be added and the analysis updated without inflating false-positive rates.
- Transparently separates the contribution of the group factor from the covariate adjustment.
- Results depend on prior specification; default priors may be inappropriate for some research contexts.
- More computationally intensive than classical ANCOVA, especially with MCMC-based estimation.
- Bayes factors can be sensitive to prior scale choices, requiring sensitivity analyses.
- Requires familiarity with Bayesian concepts (posterior, prior, Bayes factor) for correct interpretation.
- Software implementations (JASP, BayesFactor R package) vary in prior defaults, complicating direct comparison across studies.
Frequently asked
How do I choose the prior for Bayesian ANCOVA?
The most common default is the JZS (Jeffreys-Zellner-Siow) prior with a Cauchy distribution on standardized effect sizes, scaled to r = 0.5. This default is implemented in JASP and the BayesFactor R package. If domain knowledge suggests a specific effect magnitude, you can shift the scale parameter. Always report which prior you used and run a sensitivity analysis varying the scale.
What counts as strong evidence in the Bayes factor?
Jeffreys' original scale: BF10 between 1 and 3 is anecdotal, 3–10 is moderate, 10–30 is strong, above 30 is very strong evidence for M1. The same thresholds apply inverted for evidence supporting M0 (BF10 < 1/3 is moderate evidence for no group difference). These are conventions, not hard thresholds.
Can I use Bayesian ANCOVA with more than two groups?
Yes. The model extends directly to k groups. The Bayes factor tests the overall group factor (omnibus test). For pairwise comparisons, you can compute separate Bayes factors for each pair, but apply a multiple-comparisons prior or correction strategy to control error accumulation.
Does Bayesian ANCOVA still require the homogeneity-of-slopes assumption?
Yes. The standard Bayesian ANCOVA model assumes the covariate slope is the same across groups. You should test this by comparing a model with group-by-covariate interaction to one without it, using a Bayes factor. A large BF in favour of the interaction model means the assumption is violated and the interpretation of adjusted group means is problematic.
How is this different from Bayesian linear regression?
Bayesian ANCOVA is a special case of Bayesian linear regression where at least one predictor is a categorical group factor. The distinction is conceptual: ANCOVA frames the question as group comparison adjusted for covariates, while general regression frames everything as continuous predictors. In software they use the same underlying model; ANCOVA simply emphasizes the group-factor Bayes factor and adjusted means.
Sources
- Rouder, J. N., & Morey, R. D. (2012). Default Bayes factors for model selection in regression. Multivariate Behavioral Research, 47(6), 877–903. DOI: 10.1080/00273171.2012.734737 ↗
- Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2014). Bayesian Data Analysis (3rd ed.). CRC Press. ISBN: 978-1439840955
How to cite this page
ScholarGate. (2026, June 3). Bayesian Analysis of Covariance. ScholarGate. https://scholargate.app/en/statistics/bayesian-ancova
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- ANCOVAStatistics↔ compare
- Bayesian Linear RegressionBayesian↔ compare
- Bayesian MANOVAStatistics↔ compare
- Bayesian one-way ANOVAStatistics↔ compare
- Robust ANCOVAStatistics↔ compare