A Dirichlet Process Elaboration Diagnostic for Binomial Goodness of Fit.pdfVIP

A Dirichlet Pro cess Elab oration Diagnostic for Binomial Go o dness of Fit Cinzia Carota Istituto di Statistica Facolta di Scienze Politiche Universita di Pavia Strada Nuova Pavia Italia Giovanni Parmigiani Institute of Statistics and Decision Sciences Box Duke University Durham NC USA Decemb er Abstract Useful mo del checking to ols can b e constructed by measuring the distance b etween a prior distribution that concentrates most of its mass around a mo del of interest and the resulting p osterior distribution In this pap er we use this approach to construct a diagnostic measure for detecting lack of t in discrete data with sp ecial fo cus on binomial data We b egin by constructing a suitable probability mo del around the mo del of interest via a Dirichlet Pro cess elab oration We derive the resulting diagnostics and show that approximately it is the sum of two terms the rst is the logarithm of the Bayes factor and the second is prop ortional to the Pearson chisquare statistics We give details of a simulation algorithm for computing the diagnostic and illustrate its use in an application to biomedical data Keywords Bayesian mo del criticism binomial data logarithmic divergence chisquare statistics Running Title Binomial Go o dness of Fit Intro duction Diagnostic measures are often used to assess the validity of a mo del Recently in Carota Parmigiani and Polson we prop osed an approach to developing diagnostic measures that compromises b etween a fully coherent utilitybased assessment of mo delling alternatives and