Abstract:
|
Graphical modeling helps elucidate the complex interrelationships among a set of features. A key step in graphical modeling is to assess the conditional dependence between features. While common approaches for evaluating conditional dependence treat individual features as univariate variables, a multivariate approach could be advantageous in certain situations, e.g., when features are each composed of multiple sub-features. In particular, for a pair of features, if there are heterogeneous relationships present among their sub-features, an aggregated univariate approach might result in a loss of statistical power. Here we propose a flexible and nonparametric multivariate testing framework, Conditional RV, to assess the conditional dependence between two multivariate features in a graphical model. We demonstrate the performance of Conditional RV in the context of microbial association network construction, using both simulation studies and real data application. In the presence of heterogeneous relationships among sub-features, we show that Conditional RV has an improved power in detecting conditional dependence compared to univariate competing methods.
|