531 – SPEED: Statistics in Epidemiology and Genomics and Genetics
Method Selection and Assumption Evaluation: Applications to Gene Expression Data
Demba Fofanna
University of Texas Rio Grande Valley
This paper proposes hybrid-testing procedures as a general class of methods that simultaneously addresses the problems of procedure selection and multiple testing. Hybrid-testing procedures apply a set of primary testing procedures to perform the tests of primary interest (t-test or rank-sum test to evaluate equality of univariate means across groups for a large number of variables) and a set of assumption testing procedures to statistically evaluate the assumptions (e.g. the normality of data as an assumption of the t-test for those same variables) of the primary test procedures. The results from each testing procedure are summarized as a set of p-values and empirical Bayesian probabilities (EBPs) of the corresponding null hypotheses. Prior knowledge of the statistical properties of the primary testing procedures according to the validity of the statistically evaluated assumptions is used to define an algorithm. A final EBP adjusts for multiple-testing and incorporates a formal evaluation of assumptions to combine the results of several hypothesis-testing procedures in a manner guided by prior statistical knowledge. The proposed procedures are applied to gene expression data.