From the Bayes-News list, Alexander Geisler writes,

I have a few questions about Model Diagnostic. I write my diploma thesis about an use case of a Bayesian logistic regression model for a rating function of a bank.

First of all I calculated a Bayesian logistic regression model with the function MCMClogit out of the MCMCpack (http://mcmcpack.wustl.edu).

Now there is the question for testing the results. Included in the MCMCpack for R is the CODA package. A few functions of the CODA package I use for testing my model.

Now my questions:
* Can model assumptions be tested? As an example the assumption of a normal prior? Are there any possibilities?
* I know the ANOVA from the classical statistics. Is there anything similar in Bayesian statistics?
* Are there any goodness-of-fit tests for Bayesian models?
* I know residual plots in classical statistics. Is there anything similar in Bayesian statistics?

For the theoretical part of my diploma thesis my reference is the book Bayesian Data Analysis from Gelman. Is there only the CODA-package for testing in R? Or are there other packages for model diagnostic of Bayesian models in R?

My response:

Yes, model assumptions can be tested! I’d start by simulating replicated data from the model and comparing graphically to the actual data (as in chapter 6 of our book). Also, this paper on exploratory data analysis and this earlier paper on logistic regression talk about Bayesian versions of residual plots. Finally, the question about software is a good one. That was part of Jouni’s thesis–to set up a Bayesian software environment in which model checking could be done more naturally–but it’s not part of any operational system. Yet.