Statistics Glossary → Bayes factor (BF₁₀)
Bayes factor (BF₁₀)concept
A ratio showing how many times more probable the observed data are under one hypothesis (H₁) than under the other (H₀); it measures the relative strength of evidence between two hypotheses.
The Bayes factor is the ratio of the marginal likelihoods of the data under two hypotheses: BF₁₀ = P(data | H₁) / P(data | H₀). A BF₁₀ of 6 says the data are six times more probable under H₁ than under H₀; a BF₁₀ of 1/6 (that is, BF₀₁ = 6) supports H₀ equally strongly. Unlike the p-value, the Bayes factor can quantify evidence in favour of the null hypothesis and can distinguish the case where the evidence is simply insufficient (BF around 1). On a common scale, 1-3 counts as weak, 3-10 as moderate and above 10 as strong evidence.
The Bayes factor depends on the prior distribution chosen for the effect size under H₁: if the prior is very wide, small effects can appear to support H₀. Standard priors such as the JZS (Jeffreys-Zellner-Siow) prior and their scale parameter (for example r = √2/2) should therefore be reported explicitly, with a sensitivity analysis over different scales where needed. A Bayes factor is not a posterior probability: the posterior odds of the hypotheses are the Bayes factor multiplied by the prior odds.
An everyday example
A friend claims their die is loaded so that it shows a six more often. In 30 rolls it shows a six 9 times. Comparing the fair-die hypothesis with the 'sixes are more likely' hypothesis gives a Bayes factor of, say, 2, so the data support a loaded die only weakly and more rolls are needed. As the number of rolls grows, evidence accumulates in favour of one hypothesis or the other.
In MerQur
The Bayesian t-Test and Bayesian ANOVA give the Bayes factor (BF10) with the JZS prior; Bayesian ANOVA also reports model and inclusion Bayes factors. Bayesian Correlation gives BF10 for a Pearson, Spearman or Kendall correlation. In Bayesian Linear Regression, the 'Bayes factor (JZS) + R² and JZS prior scale (r)' option gives the Bayes factor of the full model against the intercept-only model, while Hierarchical Bayesian Regression computes a BIC-approximated model Bayes factor.
Analyses where this concept appears
Bayesian t-Test (JZS Bayes Factor)Bayesian CorrelationBayesian ANOVABayesian Linear RegressionHierarchical Bayesian Regression
Related concepts
Prior distributionJZS (Cauchy) prior and the scale parameter r