Investigating an extreme precipitation network with a threshold on the interest factor

Climate network theory aims to discover and understand the physical mechanisms that influence atmospheric circulation. The first step to design a climate network is the definition of the links that connect different nodes and this can be done considering the correlation/association between the nodes. Different association measures in a network design may lead to different topologies. Here, in the case of a climate network of binary variables (obtained making a threshold on the precipitation), the use of the interest factor, instead of the ϕ coefficient, or the event synchronization, is discussed, and a threshold on the interest factor, based on a significance level, is proposed. The application to an extreme precipitation network (obtained using the 80th percentile of wet days as binary threshold) and the adaptation of the threshold on binary Markov chains support the hypothesis that using a threshold on the interest factor may be a viable alternative measure of association to the ϕ coefficient, if the aim of the network analysis is to assess the co-occurrence of specific events, and not the dependence of the whole time series. The comparison of the results with the event synchronization method, which is the most used measure of association in the case of binary variables with extreme behavior (i.e., with low probability of occurrence), does not show significant differences. The proposed method identifies a valid measure of association, in case of binary variables, that takes into account the autocorrelation, greatly reducing the computational time of constructing large networks respect to the event synchronization. •Different association measures leads to different network topological structures.•Interest factor is a sound association measure to study single type dependence.•Interest factor thresholds are a very fast way to detect large networks.•Significance level based thresholds overcome the problems of the interest factor.•Significance level based thresholds solve Markov chain overdispersion issues.