Dual controls, p-value plots, and the multiple testing issue in carcinogenicity studies.

Abstract

The interpretation of statistically significant findings in a carcinogenicity study is difficult, in part because of the large number of statistical tests conducted. Some scientists who believe that the false positive rates in these experiments are unreasonably large often suggest that the use of multiple control groups will provide important insight into the operational false positive rates. The purpose of this paper is 2-fold: to present results from two carcinogenicity studies with dual control groups, and to present and illustrate a new graphical technique potentially useful in the analysis and interpretation of tumor data from carcinogenicity studies. The experimental data analyzed show that statistically significant differences between identically treated groups will occur with regular frequency. Such data, however, do not provide strong evidence of extrabinomial variation in tumor rates. The p-value plot is advocated as a graphical method that can be used to assess visually the ensemble of p values for neoplasm data from an entire study. This technique is then illustrated using several examples. Through computer simulation, we present p-value plots generated with and without treatment effects present. On average, the plots look substantially different depending on the presence or absence of an effect. We also evaluate decision rules motivated by the p-value plots. Such rules appear to have good power to detect treatment effects (i.e., have low false negative rates) while still controlling false positive rates.