VENI-onderzoek Geert Mesters: Netwerktesten

VENI-onderzoek Geert Mesters: Netwerk Testen

Door: Geert Mesters

Afgelopen zomer heeft Geert Mesters een VENI beurs ontvangen ter waarde van €250.000 van de Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO). Met deze beurs krijgt hij de kans om drie jaar lang onderzoek te doen naar netwerktesten. Op 1 september is hij begonnen met het onderzoek aan onze faculteit. 

Geert MestersGeert Mesters is in 2015 gepromoveerd aan onze faculteit en aan de Rechtenfaculteit (tegelijkertijd) en was daarna werkzaam als Universitair docent aan de Universitat Pompeu.

Samenvatting van zijn onderzoek (in het Engels):

Over the last years, network analysis has become an active topic of research in economics and finance. From an economic perspective, the interest in networks has been boosted by findings that show that individual entities can have a non-negligible effect on the aggregate economy when the system has a high degree of interconnectedness. In finance it has been shown that dense interconnections serve as a mechanism for the propagation of shocks; leading to a more fragile financial system. 

In a nutshell, network analysis is concerned with representing the interconnections of a large number of variables as a graph: the nodes of the graph represent the variables of interest, and the presence of an edge between two nodes denotes the presence of some appropriate measure of dependence between the two variables. 

A key difficulty in economics and finance is that the networks of interest are typically not observed and need to be estimated using outcome data. To this extent, recently developed estimation methods exploit the link between the (long run) partial correlation network and the inverse (long run) covariance matrix, or concentration matrix. This relationship transforms the network estimation problem into a large inverse covariance matrix estimation problem. This insight is usefull as many estimators for high dimensional covariance matrices have been recently developed.  

The estimation of unobserved networks is obviously the first step for learning about them. At the same time the sole focus on estimation has severe limitations for policy analysis. Importantly, it implies that the uncertainty around the estimated edges of the network is not quantified and therefore drawing conclusions from such estimates is difficult if not impossible. What is needed is an inferential theory to guide policy analysis based on network estimates. 

This is exactly where the current veni research comes in. It develops hypothesis tests for testing functions of edges that are estimated from high dimensional time series panels. We concentrate on testing functions of groups of edges that are relevant for economics and finance. The two leading examples are tests for degrees and tests for communities. 

Community structures, or block structures, imply that certain groups of variables are connected while being distinct from other groups. These structures arise naturally in many fields of social science and typically being able to test for the existence of such structures is of crucial importance. The degree of a node corresponds to the number of edges attached to the node. Degrees can be interpreted as measuring the overall influence of variables in the system. As such tests for degrees are tests for the stability of the network.