Department of Statistical Science
Duke University
presents:
Ina Hoeschele
Virginia Polytechnic Institute and State University
"Statistical Gene Mapping Using Bayesian Analysis and Markov Chain Monte Carlo Algorithms and Current Problems in Variance Components Estimation in Generalized Linear Models"
Abstract: A Bayesian method for the mapping of linked Quantitative Trait Loci (QTL) using multiple linked genetic markers is presented. Parameter estimation and hypothesis testing was implemented via Markov chain Monte Carlo (MCMC) algorithms. Parameters included were allele frequencies and substitution effects for two biallelic QTLs, map positions of the QTLs and markers on a chromosome allele frequencies of the markers, and polygenic and residual variances. Missing data were polygenic effects (continuous) and multi-locus marker-QTL genotypes (discrete). Three different MCMC schemes for testing the presence of 0, 1, or 2 linked QTLs on the chromosome were compared. The first approach includes a model indicator variable representing two unlinked QTLs affecting the trait, one linked and one unlinked QTL, or both QTLs linked with the markers. The second approach incorporates an indicator variable for each QTL into the model for phenotype, allowing or not allowing for a substitution effect of a QTL on phenotype, and the third approach is based on model determination by reversible jump Markov chain Monte Carlo. The methods were evaluated empirically by analyzing simulated granddaughter designs. All methods identified correctly a second, linked QTL if present, and did not reject the one-QTL model when there was only a single QTL and no additional or an unlinked QTL, besides polygenic effects, in the simulated data.
Several problems arising in variance components estimation in generalized linear mixed models, e.g., for binary data, are discussed, including biases, improper posterior distributions, and inefficiency of MCMC algorithms.
Friday, September 27, 1996
4:00 - 5:00 pm
130 Soc/Psych Building Any questions concerning the seminar may be addressed to Cheryl McGhee @[919] 684-8029, e-mail cheryl@stat.duke.edu