Likelihood methods for clustered discrete and continuous outcomes in developmental toxicology

Appropriate statistical methods are needed for analyzing developmental toxicity data. Many proposed methods make assumptions with broad implications for dose-response modeling. The first chapter outlines a history of proposed methods for developmental toxicology data while discussing their advantages and disadvantages with an emphasis on variance-covariance modeling and quantitative risk assessment.; In the second chapter, we propose a likelihood-based approach for clustered trinomial outcomes in developmental toxicology. This method does not assume conditional independence between the live and non-live outcomes. Quantitative risk assessment including benchmark dose estimation and lower limit calculation is a feature of the method.; In the third chapter, we propose a likelihood-based approach for modeling clustered discrete and continuous outcomes based on a latent variable assumption. This method is unique in incorporating discrete and continuous outcomes without relying on a conditional independence assumption between the live and non-live endpoints within a litter, a limitation of some other methods that does not appear to hold in at least some data sets. Benchmark dose estimation and joint quantitative risk assessment is described and lower limit calculations based on the likelihood ratio test statistic are available.; In the fourth chapter, we compare the full likelihood method proposed in Chapter 3 to a simpler, two-stage estimation technique via a simulation study. Two-stage estimation may be an alternative estimation method for allowing conditional dependence in the model without a large degree of computational complexity. Comparisons between estimates from the two models and their biases are performed under 96 combinations of parameter settings reflecting a wide range of underlying event rates and correlations. In addition, quantitative risk assessment is performed to evaluate the regulatory implications which might result from these two methods.