|
Abstract:
|
Increasing genome-wide association studies for complex human diseases have drawn interests of statisticians to the issue of dimension reduction. Various variable selection methods, including information criteria, penalized likelihood functions, and Bayesian models, have been proposed for finding a modest-sized subset of markers in the search of associations. However, there has been a lack of guidance in practice on how to choose from these methods. This paper compares these methods through simulation as an attempt to answer this question. Data sets simulated by re-sampling from actual genomic data are used. Practical issues are considered such as estimation biases, computation complexity, and the circumstance when observations are missing. The results are discussed and suggestions for selecting methods are made upon difference conditions.
|
