Brown University Division of Biology and Medicine
Center for Statistical Sciences Seminar
Harvard School of Public Health
Candidate For Assistant Professor (Tenure Track) | |
Abstract: There is increasing interest in modeling disease on the basis of a moderate number of binary genetic factors, such as loss of heterozygosity (LOH) or methylation silencing. One natural approach is to use latent variable models, examples of which include latent class analysis (LCA) and item response theory (IRT). However, such techniques can fail when the number of subjects is of small to moderate size. We present a regularization approach that provides numerical stabilization and can improve estimation. In LCA, conditional on unobserved membership in one of k classes, item responses are assumed independent. We impose a penalty across the response probabilities conditional on class; the form of the penalty may differ depending upon whether responses have a natural ordering (e.g. LOH) or when it is of interest to identify items that distinguish class membership (e.g. methylation silencing). Variations of this approach include a linear-by-linear ordinal LCA. A similar reqularization approach may be used to improve estimation in IRT models, where an underlying latent normal variable is assumed to drive responses. We illustrate the techniques using LOH data obtained from brain tumors and methylation data obtained from bladder tumors.
<--- 2005 Index