Dose-response assessment for developmental toxicity. III. Statistical models.

Although quantitative modeling has been central to cancer risk assessment for years, the concept of dose-response modeling for developmental effects is relatively new. The benchmark dose (BMD) approach has been proposed for use with developmental (as well as other noncancer) endpoints for determining reference doses and reference concentrations. Statistical models appropriate for representing the unique features of developmental toxicity testing have been developed and applied (K. Rai and J. Van Ryzin, 1985, Biometrics 41, 1-9; L. Kupper, C. Portier, M. Hogan, and E. Yamamoto, 1986, Biometrics 42, 85-98; R. Kodell, R. Howe, J. Chen, and D. Gaylor, 1991, Risk Anal. 11, 583-590). Generalizations of those models (designated the RVR, LOG, and NCTR models, respectively) account for the correlations among observations in individual fetuses or implant within litters; the potential for variables other than dose, such as litter size, to affect the probability of adverse outcome; and the possibility of a threshold dose below which background response rates are unaltered. The generalized models were applied to a database of 607 endpoints with significant dose-related increases in response rate. It was determined that the models were generally capable of fitting the observed dose-response patterns, with the LOG model appearing to be superior with respect to fit. A significant contributor to the ability of the LOG model to fit the data was its flexibility with respect to the representation of the dependence of response probability on litter size, a trait not shared by the other two models. Litter size appeared to be a significant covariable for predicting response rates, even when intralitter correlation was accounted for by assuming a beta-binomial distribution for the observations among individual fetuses. In contrast, a threshold dose parameter did not appear to be necessary to adequately describe the observed dose-response patterns. BMD estimates (corresponding to 5% additional risk) from all three models were similar to one another and to BMDs estimated from other, generic dose-response models (not specifically designed for developmental toxicity testing) that modeled average proportion of fetuses affected. The BMDs at the 5% level of risk were similar to no observed adverse effect levels determined by statistical tests of trend. Greater emphasis on and further examination of dose-response modeling for developmental toxicity testing are needed; biologically based approaches that consider the continuum of developmental effects induced in such tests should be encouraged.