Abstract
Identifying causal interactions in neural ensembles is, generally, a challenging task when multiple variables, causal loops, self-interactions, and strong nonlinearities are present. The Granger causality (GC) along with its multivariate and nonlinear descendants recently gained popularity. GC paradigm is based on removing one variable at a time from the ensemble and estimating how this affects predictability of the remaining variables. This approach cannot be used to evaluate self-interactions and, besides, has a strong tendency to weaken causality estimates when applied to non-separable systems (both linear and nonlinear), where a variable’s information is not completely removed by simply removing the variable. For example, consider a variable A, which causally influences variables B and C, A->C having a longer time lag. A->C causality will be underestimated by GC-based measures compared to A->B causality because A’s information is encoded in B even with A eliminated. Here a non-GC approach, which avoids such artifacts, is proposed. The variable tested for causality is never removed from the ensemble, its causal input is estimated from the full ensemble predictions instead. Based on this approach two new causality measures are proposed, one (linear) based on multivariate autoregressive fit (CARMA), the other (nonlinear) based on higher-order mutual information estimates (CIM). The measures were tested on various simulated and biological/neural datasets and compared favorably with GC-based measures. When applied to visually evoked potentials CARMA and CIM produced similar results, hence validating them given the very different methods. Strong causal interactions were observed between occipital (OP) and lateral-occipital (LO) scalp regions, when the stimulation (contrast reversal) was contralateral to the ROIs. For both measures the LO->OP causality peaked around 100 - 150 msec from the stimulus onset. CIM indicated ~5 msec time lag of the causal interaction, possibly, a feedback from the LOC/MT complex to early visual areas V1/V2/V3.
Meeting abstract presented at VSS 2013