Statistical modeling for genetics: Pharmacogenetics, molecular evolution and complex traits

In this dissertation, three projects exploring statistical applications to genetic problems are conducted. All the methods investigated have application in the fields of pharmaceutical and medical research.; The first study is in collaboration with researchers at GlaxoSmithKline (RTP, NC) to evaluate recursive partitioning as a useful tool in modeling pharmacogenetic data. It is conducted in three stages using two data sets created to simulate pharmacogenetic clinical trials and one real set of data from a clinical trial sponsored by GlaxoSmithKline. Analysis plans, including recursive partitioning and regression methods, are constructed to compare modeling techniques. Strategies and procedures are recorded throughout the investigation. Experience gained at one stage contributes to the development of the analysis plan for the next stage. A final strategy is proposed to implement efficient modeling methods for prediction and explanation of pharmacogenetic data.; The goal of the second study is to find the best evolutionary model for HIV sequence data, based on maximum-likelihood estimation. This involves calculating maximum-likelihood phylogenies of HIV-1 molecular sequences using Hidden Markov Models and comparing the resulting models to classical likelihood models. Eighteen sets of DNA sequences, used in studies by Poss et al. (1997, 1998), are chosen for analysis. They are sampled from three Kenyan women infected with clade A HIV-1. Maximum likelihood phylogenies using Hidden Markov Models are constructed with the software program PHYLIP, and the classical models, e.g. Jukes and Cantor (1969), are constructed using PAUP. These common models are generally based on a continuous-time Markov chain. The Akaike Information Criterion is used to compare the models.; In the third study, an association-based linkage test is derived by extending the transmission probabilities found in Sham and Curtis (1995) to the two-locus model for candidate genes. For the two-locus model, there are five parental genotypes providing useful information, where at least one locus is heterozygous. In this setting, the transmission probabilities are defined as the parameters in a binomial or multinomial distribution depending on parental mating type. A likelihood function is developed based on these distributions. The power of this test is compared to the two-locus TDT (Fan et al., 2000).