Bayesian Penalized Regression Methods for Matched Case-Control Data

• Activity Number: 538
• Type: Invited
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: International Indian Statistical Association
• Author(s): Jaya Satagopan*+ and Ananda Sen and Qin Zhou
• Companies: Memorial Sloan Kettering Cancer Center and University of Michigan and Memorial Sloan Kettering Cancer Center
• Keywords: Case Control Study ; Reduced Rank Regression ; Shrinkage Estimation ; Bayesian Lasso ; non-Hodgkin's lymphoma

Abstract

Matched case-control designs are used in epidemiology studies to assess the effect of exposures on binary traits. Modern epidemiology studies increasingly enjoy the ability to examine a large number of exposures. However, the risk factors often tend to be related in a non-trivial manner, undermining the efforts to identify the disease risk factors due to inflated type I errors and possible masking of effects. Data reduction methods such as principal component analyses are useful, but the results are frequently not replicated in a new wave of samples. Therefore, epidemiologists group the prognostic factors into themes based on biological considerations. A data reduction technique based on these themes can provide critical insights into the disease risk factors. However, it is important to account for potential misspecification of the themes to avoid false positive findings. To this end, we propose novel Bayes and empirical Bayes methods to evaluate the risk factors by making use of the themes, while simultaneously accounting for their potential misspecification. We illustrate these methods using data from a matched case-control study of organochlorines and non-Hodgkin's lymphoma.