Dose-response data from experimental and epidemiologic carcinogenicity studies were analyzed in attempts to resolve basic questions in extrapolating from high to low doses and assessing human risk. Four models (Weibull, Mantel-Bryan, Marshall-Groer, and Mancuso-Stewart) were fit to 46 sets of experimental and 4 sets of epidemiologic data by maximizing the likelihood function with a Rosenbrock hill-climbing algorithm. The models were compared as to their adequacy in describing the data and analyzed to determine the effect of carcinogen and breeding category, species, and spontaneous tumor incidence. The shapes of the dose-response curves were analyzed, and errors in risk estimation from linear extrapolation through the origin were calculated. All models were shown to fit the data and to be comparable in accuracy. The dose-response curves were generally "stretched out," particularly for outbred strains, with one or two orders of magnitude of dose increase required to increase the proportion of tumor responders from 10% to 70%. Linear extrapolation through the origin generally underestimated the response at low doses, frequently by several orders of magnitude. The power dependence of tumor incidence on dose was generally found to be of the order of unity, substantially less than assumed in most mathematical models.

No abstract