Identification of directed interactions in networks

Abstract

Multichannel data collection in the neurosciences is routine and has necessitated the development of methods to identify the direction of interactions among processes. The most widely used approach for detecting these interactions in such data is based on autoregressive models of stochastic processes, although some work has raised the possibility of serious difficulties with this approach. This article demonstrates that these difficulties are present and that they are intrinsic features of the autoregressive method. Here, we introduce a new method taking into account unobserved processes and based on coherence. Two examples of three-process networks are used to demonstrate that although coherence measures are intrinsically non-directional, a particular network configuration will be associated with a particular set of coherences. These coherences may not specify the network uniquely, but in principle will specify all network configurations consistent with their values and will also specify the relationships among the unobserved processes. Moreover, when new information becomes available, the values of the measures of association already in place do not change, but the relationships among the unobserved processes may become further resolved.