A Computationally Efficient and Statistically Robust Optimal Discovery Procedure Estimator

• Activity Number: 136
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Sangsoon Woo*+ and John Storey
• Companies: University of Washington and Princeton University
• Keywords: optimal discovery procedure ; Kullback-Leibler divergence ; Likelihood

Abstract

Real data shows that group of genes share similar regulation pattern in gene expression and the genes are also biologically related. We developed a clustering optimal discovery procedure (cODP) method using this common expression pattern. We cluster genes with Kullback-Leibler distance, estimate each cluster's likelihood and estimate the statistics for each gene. The cODP shows consistent results as the full ODP approach (introduced by John). Because we evaluate each gene over all clusters and the number of clusters is much less than the number of genes, the new approach substantially reduce the computing time. Because each cluster's likelihood is estimated using genes grouped, we get more robust estimator. Simulation study and Moroccan study show that cODP is not sensitive to the choice of number of clusters. And it shortens running time dramatically than the full ODP for the same data.