MerQuR+Statistics Glossary

Statistics Glossary → JZS (Cauchy) prior and the scale parameter r

JZS (Cauchy) prior and the scale parameter rconcept

The default family of prior distributions for effect size used when computing Bayes factors in Bayesian t-tests, ANOVA, correlation and regression, named after Jeffreys, Zellner and Siow.

A Bayes factor needs a prior stating which values the effect could take under H₁. In the JZS approach to the t-test, the standardised effect size δ is given a Cauchy(0, r) prior (Rouder et al., 2009). This prior is concentrated around zero but, thanks to its heavy tails, still allows large effects. r sets its width: it says the effect lies between −r and +r with probability of about one half. Common named values are 'medium' √2/2 ≈ 0.707, 'wide' 1 and 'ultrawide' √2. In ANOVA and regression the same idea is generalised to many parameters through an inverse-gamma prior on g (Rouder et al., 2012).

The Bayes factor is sensitive to r. A wider prior gives more probability to large effects under H₁, so when the observed effect is small it shifts the evidence toward H₀. r should be chosen from the expected effect size before looking at the data, and the sensitivity of the result to different r values (a Bayes factor robustness plot) should be reported. The 'default' prior is not objective; it is only a reasonable and comparable starting point in many situations. The posterior distribution is usually far less sensitive to the prior than the Bayes factor is.

An everyday example

A psychologist expects a new memory technique to have a small-to-medium effect and uses r = 0.707. The Bayes factor is BF₁₀ = 3.2: the data are about 3 times more likely under H₁ than under H₀. With r = 1.414 the same data would give BF₁₀ of about 2.4, a difference that can be shown in the report as a robustness analysis.

In MerQur

In the Bayesian t-test, 'JZS prior scale (r)' sets the width of the δ ~ Cauchy(0, r) prior, and the default box value of 0.707 is treated as exactly √2/2. In Bayesian ANOVA, 'JZS prior scale r (fixed effects)' is the scale of the g ~ IG(1/2, r²/2) prior (medium 0.5). In Bayesian linear regression, 'Bayes factor (JZS) + R² and JZS prior scale (r)' gives the JZS Bayes factor of the full model against the intercept-only model.

Analyses where this concept appears

Bayesian t-Test (JZS Bayes Factor)Bayesian ANOVABayesian Linear Regression

Related concepts

Bayes factor (BF₁₀)Prior distribution