| Summary: | This work was partially supported by the Canada Foundation for Innovation, by the Ministry of Research and Innovation of Ontario, and by the Faculty of Medicine of the University of Ottawa. I would like to thank Prof David Bickel, at the Department of Biochemistry, Microbiology and Immunology, Faculty of Medicine, University of Ottawa for his guidance throughout the pilot study and for the permission to use the R functions developed by him. I would also like to thank Corey Yanofsky, for the useful discussions related to the simulation study, while I was getting trained at Dr.Bickel’s Lab. Furthermore, the computational resources were provided by ACENET, the regional advanced research computing consortium for post-secondary institutions in Atlantic Canada. ACENET is funded by the Canada Foundation for Innovation (CFI), the Atlantic Canada Opportunities Agency (ACOA), and the provinces of Newfoundland and Labrador, Nova Scotia, and New Brunswick. Large scale data, such as the one collected in microarray, proteomics, MRI imaging, and massive social science surveys etc., often requires simultaneous consideration of hundreds or thousands of hypothesis tests, which leads to inflated type I error rate. A popular way to account for it is to use local false discovery rates (LFDR), which is the probability that a gene is truly not differentially expressed given the observed test statistic. The purpose of this report is to evaluate the Bootstrap Bias Corrected Percentile (BBCP) method proposed by Shao and Tu (1995) for estimating the lower bound for the LFDR. The method didn’t perform as expected. The overall coverage probability for null genes as well as non null genes was far from nominal coverage level of 50%. SUNY Brockport Mathematics Faculty Publications
|
|---|