Discussion of “ Multivariate Bayesian Logistic Regression for Analysis of Clinical Trial Safety Issues ” by W . DuMouchel

We would like to comment on this article by William DuMouchel, as it gives an interesting application of logistic regression to clinical safety data. Not to underscore the scope of the multivariate Bayesian logistic regression (MBLR) model, but the use of numerical integration is arguably its most important feature. Avoiding Markov chain Monte Carlo (MCMC) sampling techniques for other data-mining tools, such as the Multiple-item Gamma Poisson Shrinker (DuMouchel, 1999), has proven successful for Dr. DuMouchel in their acceptance among nonstatisticians. With MBLR this should not be an exception. As most statisticians lack the clinical insight required to specify the appropriate MBLR model inputs, it makes MBLR an ideal tool for use by the clinicians. However, targeted users may not appreciate some subtleties of MBLR, which we present below. We also present findings from our empirical evaluation of the MBLR algorithm. This commentary provides some perspective that we have gained through multiple interactions with Dr. DuMouchel