Statistics Glossary → Prior distribution
Prior distributionconcept
In Bayesian analysis, the probability distribution expressing the information or uncertainty about a parameter before the data are seen.
By Bayes' theorem, the posterior distribution is proportional to the likelihood times the prior: posterior ∝ likelihood × prior. The prior can reflect information from earlier studies (an informative prior) or only sensible bounds on plausible values (a weakly informative or non-informative prior). As the amount of data grows, the likelihood dominates and posterior estimates become less sensitive to the prior.
Bayes factors, however, are clearly sensitive to the prior, in particular to the prior width of the effect: a very wide prior favours the null hypothesis. So when reporting a Bayes factor, the prior should be justified and a sensitivity analysis with several reasonable priors is recommended. In the Bayesian t-test, for instance, a Cauchy prior with scale parameter r on the standardized effect is common. The 'prior class probabilities' of classification methods rest on the same idea: the expected share of each class before the observation's data are seen, which affects the assignment rule.
An everyday example
If nothing is known about a newly opened restaurant, calling it excellent after two good reviews would be hasty. If the knowledge that most restaurants nearby score between 3.5 and 4.5 out of 5 is used as a prior, the first reviews are combined with that expectation; as reviews pile up, the restaurant's own data take over.
In MerQur
'JZS prior scale (r)', 'Prior type' (Cauchy r, medium, wide, ultrawide, informed t or normal) and 'Informed prior' in the Bayesian t-Test; 'Prior width κ (stretched beta)' in Bayesian Correlation; 'Prior type' (Conjugate NIG, Reference Jeffreys, Zellner g-prior), 'Prior precision' and 'Prior hyperparameters' in Bayesian Linear Regression; 'JZS prior scale r (fixed effects)' in Bayesian ANOVA. 'Prior probabilities' (Equal, Class proportions, User-defined) in Discriminant Analysis and 'Imbalanced classes: balanced weights' (equal prior probabilities) in the Decision Tree are prior class probabilities; the 'Firth correction' of Logistic Regression applies the penalized likelihood with the Jeffreys prior.
Analyses where this concept appears
Logistic RegressionDecision Tree (CART)Discriminant Analysis (LDA / QDA)Bayesian t-Test (JZS Bayes Factor)Bayesian CorrelationBayesian ANOVABayesian Linear Regression
Related concepts
Posterior distribution and posterior predictive checkJZS (Cauchy) prior and the scale parameter rNon-informative (flat, Jeffreys) prior