Distance-based variable selection for partial correlations using the KOO method
Conference
64th ISI World Statistics Congress - Ottawa, Canada
Format: CPS Poster
Keywords: (n,p)-asymptotic, gaussian-graphical-model, variable-selection
Session: CPS Posters-03
Monday 17 July 4 p.m. - 5:20 p.m. (Canada/Eastern)
Abstract
In this study, we consider a selection problem which is to estimate the set of nonzero partial correlations.
Two new model selection criteria based on distance are derived.
We propose a knock-one-out (KOO) method for these criteria.
It is shown that these KOO methods have a consistency under the asymptotic framework that both the dimensionality and the sample size go to infinity and the ratio converges to a positive constant less than 1.
That is, our model selection methods choose true model for non-zero correlation with high-accuracy for sufficiently large sample and dimensionality case.
We do small-scale simulation for our proposed methods to confirm the consistency.