Seminar by Chris Salahub

Searching Pairwise for Interesting Variables

When presented with a large data set containing many variables, a routine part of early analysis is to identify the variables of greatest interest for further investigation. This talk outlines the problem and presents several results relating to its aspects. Along the way, a new measure of dependence based on recursive binning is introduced and a framework for using pooled p-values to adjust for multiple tests is developed based on two common patterns of evidence. The measure of dependence and adjustment framework are combined under a simple genetic model to distinguish patterns of trait inheritance.  Finally, concepts from the pooled p-value framework are used to motivate an exploration of novel ways to combine parameter estimates in meta-analysis.