The role of the design, analysis, and computation in addressing etiology in three types of studies in public health.

摘要

This dissertation comprises three distinct research papers. These papers, while quite diverse superficially, have a broad, common theme that each one of them addresses an aspect of delineating the etiology in studies of public health. The first paper focuses on two important methodological issues pertaining to the design and analysis in a relatively new epidemiological study design, called the case-crossover design. This design, which only uses the cases, can be useful in developing hypotheses regarding the etiology of an acute event by examining the association between a recurrent exposure and the acute event.; The second paper addresses an important analytic issue---model uncertainty in model based low dose extrapolation for microbial risk assessment. Here different dose-response models provide competing or alternate accounts of the etiology of microbial infection/illness. Inference on low-dose risk estimates is highly sensitive to model choice, and, hence, can be overly optimistic, if model uncertainty is ignored. Bayesian model averaging (BMA), which weights the inference from the competing models by the posterior model probabilities, provides a natural and rigorous framework for incorporating model uncertainty. However, in practice, its utility and objectivity are severely challenged by the need for specifying joint prior distributions on parameters in all the models. We propose a new approach called profiled Bayesian model averaging (PBMA), that uses the profile likelihood in Bayesian model averaging, and that only requires prior distribution on the target of inference.; We have developed a new class of computationally efficient, yet simple, numerical schemes, called the SQUAREM, which significantly improve the rate of convergence of the EM. These new schemes are based on extrapolation techniques from the numerical analysis literature. They can be applied very broadly to any nonlinear fixed point iterative scheme, and to the EM, in particular. When applied to the EM, they do not require any auxiliary quantities such as the complete- or incomplete-data log-likelihood, and/or their derivatives. We demonstrate the improved performance of the new schemes using a number of examples. The new computational schemes have potential utility in important scientific problems such as causal inference in longitudinal studies, latent variable regression models, mixed effects models, image reconstruction, and population genetics models. (Abstract shortened by UMI.)