Data mining for mutagenic potency estimation in hierarchical linear models

Potency estimation is a general concept in toxicology, viewed as a quantitative measure of an agent's relative toxic capability. Properly calibrated, a positive potency indicates a positive toxic potential, while a zero or negative potency is not considered detrimental. The potency estimates are taken from a ubiquitous bioassay known as the Ames/Salmonella assay used for purposes of environmental hazard identification and risk assessment. There is a broad spectrum of potency estimators in this arena from which to choose, however, it is not clear which estimators are best to use for statistical inferences and models. This dissertation investigates the distributional properties of ten different potency estimators through a large simulation study. From this study, $bD1$, b_lnDUI, and $bKL3$appear to be consistent with a fairly symmetrical distribution.; These three potency estimators are used in simple hierarchical linear models to combine information across different experiments. Both the full hierarchical Bayes and empirical Bayes approach are investigated using three criteria: reduction in standard errors, shrinkage, and changes in inference. Evaluation of the hierarchical models under the three criteria did indeed indicate that information is gained through use of the models and these techniques may be useful in obtaining more precise interval estimates for individual potency estimators.