CMStatistics 2016: Start Registration
View Submission - CMStatistics
B1147
Title: Graph-based multiple testing with correlated and high-dimensional data Authors:  Florian Klinglmueller - Medical University of Vienna (Austria) [presenting]
Abstract: Graph-based multiple testing procedures permit to tailor the testing procedure such that it reflects the importances and contextual relations of different study objectives. The procedure is specified using a directed weighted graph, where nodes correspond to hypotheses and edges determine the algorithm for reallocating the significance level. Being based on the Bonferroni test, these procedures suffer from considerable conservativeness, especially in situations where test statistics are correlated. We present extensions of graphical approaches that account for the correlation between observations even if the joint distribution is unknown. This is achieved by using multivariate permutation tests and by adapting the weighting strategy using the blinded observations. We illustrate the approach with applications from neuroscience and genetics.