Statistics Glossary → Posterior distribution and posterior predictive check
Posterior distribution and posterior predictive checkconcept
In Bayesian analysis, the probability distribution obtained by combining prior information with the data, which expresses updated belief about the possible values of a parameter; the posterior predictive check tests the model's fit with data generated from this distribution.
Bayes' theorem defines the posterior as proportional to the product of the prior and the likelihood: p(θ | data) ∝ p(data | θ) · p(θ). The posterior usually cannot be computed analytically, and samples are drawn from it with simulation methods such as MCMC. From these samples one computes the posterior mean or median (point estimate), the posterior standard deviation and credible intervals (e.g. the 95% highest density interval, HDI). Unlike a frequentist confidence interval, a 95% credible interval can be read directly as 'given the data and the prior, the probability that the parameter lies in this interval is 95%'.
In a posterior predictive check, new data sets are simulated using parameter values drawn from the posterior and compared with the observed data. If the simulated data cannot reproduce important features of the observed data (mean, spread, excess zeros, extreme values), the model is inadequate. With small samples the posterior can be sensitive to the choice of prior, so a sensitivity analysis with different priors should be carried out. Convergence of the MCMC samples should also be checked with diagnostics such as R-hat and the effective sample size.
An everyday example
Looking at a patient's symptoms, a doctor initially puts the probability of a disease at 20% (the prior). A test result then updates that estimate and the probability rises to 60% (the posterior). To see how good her model is, she also asks: 'if my estimates are right, what kind of test results should I have seen in similar patients?' and compares that with what she actually saw. That second step is a posterior predictive check.
In MerQur
It is the core output of the Bayesian t-Test (BEST), Bayesian Correlation, Bayesian ANOVA and Bayesian Linear Regression. In GLMM the VB (mean-field) estimate approximates the posterior, and in Multiple Imputation the BayesianRidge option performs posterior sampling; Discriminant Analysis, Latent Class and Latent Profile Analysis give posterior probabilities of class or profile membership.
Analyses where this concept appears
Generalized Linear Mixed Models (GLMM)Discriminant Analysis (LDA / QDA)Latent Class Analysis (LCA)Latent Profile Analysis (LPA)Bayesian t-Test (JZS Bayes Factor)Bayesian CorrelationBayesian ANOVABayesian Linear RegressionMultiple Imputation
Related concepts
Prior distributionCredible interval and highest density interval (HDI)