Beyond pairwise: investigating higher-order statistical behaviour in network psychometrics.

N. Van Santen${}^a$, Y. Rosseel${}^a$ and D. Marinazzo${}^a$

${}^a$Department of Data-analysis, Ghent University

Networks have become a popular data-analytic tool for investigating item interactions in psychometric data [1]. This approach exemplifies a shift towards understanding the parts vs. whole contribution of items towards the construct under investigation, together with identifying central items or interactions for possible interventions. We believe that it is beneficial to extend this analysis beyond only pairwise statistics, since different items can become important at different orders [2]. By not restricting ourselves to pairwise associations, we can illuminate a richer landscape of behavioural pictures that can together inform possible underlying mechanisms. The focus shifts from an intractable network inference problem to shrinking the state space of possible mechanisms. We use higher-order information theoretical measures to reanalyze a dataset used to investigate the construct of empathy [3] and one used in personality research [4]. We compare our results with those drawn for correlation-based network psychometrics in the original publications.

Keywords: Network science, information theory, higher-order