Summary
the 2014 International Symposium on Nonlinear Theory and its Applications
2014
Session Number:B1L-B
Session:
Number:B1L-B4
Testing the randomness of causality networks from multivariate time series
Dimitris Chorozoglou, Dimitris Kugiumtzis,
pp.229-232
Publication Date:2014/9/14
Online ISSN:2188-5079
Summary:
In network analysis it is important to contrast the given or formed network to a random network, typically by means of significance testing of some network measures. This requires the generation of random networks that preserve certain properties of the original network, i.e. the total number of nodes and connections or even the number of connections of each node. In this paper we show that these schemes are not appropriate for correlation or causality networks formed from multivariate time series, which have nodes the observed variables and connections given by some measure of correlation (undirected connections) or Granger causality (directed connections). Further, we propose a scheme that performs the randomization on the time series rather than the network connections. Particularly for the networks formed by cross correlation, we generate surrogates for each time series preserving the marginal distribution and linear autocorrelation, and form the network from the surrogate multivariate time series. Simulations on multivariate time series with no inter-dependencies shows that the classical network randomization erroneously tends to reject the null hypothesis of random network, whereas the proposed scheme does not.