The investigator develops methods that control the multiple comparison error rate called generalized Familywise Error Rate (gFWER). Whereas control of the traditional Familywise Error Rate (FWER) guarantees with a high probability that no incorrect inference in multiple comparison will be made, control of gFWER guarantees with a high probability that the number of incorrect inferences in multiple comparison will be kept to no more than a fixed number m. Thus control of gFWER is less stringent than control of FWER, if m is greater than zero. Controlling gFWER may be more appropriate in some bioinformatics situations than controlling the False Discovery Rate (FDR), which is the expected proportion of false discoveries. The investigator uses the generalized Partitioning Principle to develop modeling-based methods controlling gFWE. The generalized Partitioning Principle makes specific use of the joint distribution of test statistics to derived methods more powerful than methods that ignore such information.
The new methodologies the investigator develops can have a broad impact in the area of bioinformatics. For example, microarrays probing all human genes can discover which genes are differentially expressed in a disease situation. As a follow-up study, SNP (single nucleotide polymorphism) chips can be used to confirm which genes have mutations causing increased susceptibility to the disease. In such situations, the cost associated with falsely rejecting true null hypotheses (of non- differential expression of genes) is roughly proportional to the number of incorrectly rejected true null hypotheses. Controlling gFWER controls such costs. In collaboration with genomic researchers, the investigator is developing and testing his gFWER-controlling multiple comparison methods for microarray data analysis.
