Abstract:
|
Understanding associations among features is one of the more important problems in modern statistical and machine learning literature. Graphical models give us the language and the toolkit to model and interpret these associations. In many practical scenarios however, a graphical structure is already supplied by experts - the statistical validity of which would then be needed to be ascertained. In this work, we employ the newly-developed directional likelihood method to construct chi-squared type tests for the null edges of the pre-specified graph structure under Gaussian data. We also establish that this method can be easily adapted for Gaussian copula graphical models via a pseudo-likelihood approach by employing non-parametric measures of association such as Kendall's tau. We delineate the performance of this procedures through extensive simulation studies.
|