Title: Bayesian Lasso-related variable selection methods for prediction and signal detection Abstract: High-throughput laboratory techniques are producing vast amount of genetic marker data. Linear regression model is often considered to link study phenotypes and these marker measurements to each other for: (1) gene mapping of trait-associated loci, (2) predicting genomic breeding values in plants/animals, and (3) predicting (disease associated) phenotypes in humans. We consider recent extension of Bayesian LASSO which we call extended Bayesian LASSO (Mutshinda and Sillanpää 2010). This method is related to so called adaptive LASSO where the tuning parameter has been made locus specific. We further divide this locus specific parameter into two parts, common and locus specific parts. This extension provides many nice properties like the method becomes robust to tuning and it can separate model sparseness and shrinkage in individual parameters. We also derive formal Bayes factor for decision making based on these locus-specific hyper parameters (Mutshinda and Sillanpää 2012). The nice performance of the method in general is illustrated with few examples (e.g. Kärkkäinen and Sillanpää 2012) including detection of gene-gene interactions (Li and Sillanpää 2012).
About the presenter: Mikko Sillanpää is professor of Statistics at University of Oulu. He is affiliated at Department of Mathematical Sciences and Biocenter Oulu. Sillanpää's research considers statistical methods and their applications in biology. His present research interest address statistical variable selection methods and their use in identifying genetic determinants and prediction of individual's genetic value (merit / risk) to the quantitative, qualitative and function-valued human, plant and animal traits based on genome-wide sets of molecular markers. Newer interests are variance and precision matrix inference methods.
Last updated on 20 Apr 2015 by Mats Sjöberg - Page created on 20 Apr 2015 by Mats Sjöberg