A0581
Title: Signpost testing to navigate the parameter space of the Gaussian graphical model with high-dimensional data
Authors: Wessel van Wieringen - Amsterdam University Medical Centers (Netherlands) [presenting]
Abstract: A hypothesis test is presented to guide the search for the precision parameter of the Gaussian graphical model (GGM) from high-dimensional data. This parameter is unknown, but often information on it is available from a different setting. The unknown and different settings are assumed to be represented quantitatively by a parameter value. The direction between the values is a signpost in the parameter space. The hypothesis test evaluates the signpost for the true parameter. The test's null and alternative hypotheses state that no or some, respectively, relevant information for the parameter can be found in the direction of the signpost. The line segment is parameterized between the two settings by a one-dimensional parameter. The test statistic optimizes, within the class of regularized precision matrix estimators, the loss over this line segment. It measures the relevance of the signpost's direction for the problem at hand. The test statistic's null distribution is constructed or approximated asymptotically. The signpost test's power is evaluated, and it is compared to the likelihood ratio test. The GGM's precision matrix of a pathway is learnt in a low-prevalent breast cancer subtype. In addition to in-house gene expression data, external data on a more prevalent subtype is available. The latter comes to mimic federated learning under the EU's GDPR, as a parameter estimate. The signpost test assesses this estimate's relevance for the in-house estimation problem.
