|
Abstract:
|
In this paper we propose a Bayesian framework for building prediction models using high dimensional data by combining a Bayesian variable selection method with generalized linear models(GLMs) for various phenotypes. Our strategy is 'divide and conquer', that is, we perform variable selection first and then, given selected variables we estimate parameters of interest such as regression coefficients in the prediction models. The rationale of this strategy is that when we have a few true predictors as well as many noisy ones, model uncertainty is a much bigger problem than parameter estimation. We consider binary, ordinal, count, and censored survival outcomes. To illustrate our approach we use data on a proteomic profiling study of melanoma, where serum proteins for each subject were analyzed using SELDI-TOF MS data. We also deal with highly correlated predictors in the variable selection.
|
