Using priors to improve multiple animal carcinogenicity tests

This article reviews and summarizes methods for controlling false positives in animal carcinogenicity studies and promotes an alternative that incorpo rates historical control information via Bayesian methods. The Bayesian par adigm is used as a procedure generator; however, frequentist multisample, a ge-stratified exact trend tests are used in the ultimate analysis. Critical values for the exact tests are chosen to maximize total expected power, co nditional on tumor totals, by using prior distributions. To control the ris k of a false-positive finding for one or more tumor types, the sum of the i ndividual critical levels is constrained to be less than a nominal familywi se error rate, such as .05. The resulting tests give more power to tumor ty pes with higher-than-expected tumor totals. We use. historical control data from animal carcinogenicity studies obtained from a large pharmaceutical c ompany to train and evaluate the tests. There is greatly enhanced power of the proposed method, with concurrent error rate control, because the target ing procedure gives higher power to affected sites, and the procedure tends to produce critical values that are as small as possible overall (implying higher power), subject to the overall risk level constraint. Randomly samp ling from real historical animal populations, we compare operating characte ristics of various methods proposed in the literature and requested by regu latory agencies. Commonly used methods can have greatly inflated false-posi tive rates, particularly with larger studies. In some cases, we find greate r power for the proposed method, even compared to methods that do not contr ol false positives.