Statistical support of the proposed National Toxicology Program protocol.

Statistical issues related to the design, analysis and interpretation of carcinogenicity studies are considered. A proposed modification of the "standard" National Cancer Institute (NCI) design of carcinogenicity studies in rodents is discussed. The primary objective was to obtain a design that maintained the power of the NCI design for detecting carcinogenic effects, while increasing the precision with which the underlying dose-response model could be estimated. "Optimum" three-dose designs are presented and discussed. Data analysis issues are also considered. Comparisons of tumor incidence should include adjustments for survival differences, and, if possible, cause of death determinations should be made for each individual animal. In some instances historical control data can aid an investigator in the overall evaluation of experimental results. Several important problems are discussed that must be overcome before historical control data can be used in a meaningful way. One critical issue in the interpretation of tumor incidence data is the issue of false positives. Data from 25 recent long term feeding studies are used to derive a statistical decision rule that closely approximates the scientific judgment process used in these studies. The overall false positive rate associated with this decision rule was found to be no more than 7-8%.