Adaptive Frequency-domain Granger Causal Inference from Neuronal Ensemble Data

Abstract

Granger causality is an increasingly prevalent tool in extracting the functional networks that underlie neural processes. While its time domain formulation yields useful insights into these functional networks, the inferred Granger causal influences leave the spectral properties of said functional networks ambiguous: this is a question of particular interest when neural processes exhibit oscillatory behavior. The frequency-domain formulation of Granger causality proposed by Geweke has addressed the spectral properties of functional links between stationary processes. Based on Geweke’s method of conditional Granger causality, we introduce a framework to derive direct spectro-temporal causal interactions in a population of neurons from multivariate ensemble spiking observations, using point process modeling, state space estimation and multitaper spectral analysis. Further, we propose statistical tests that characterize the significance of these functional links. The utility of our methods is demonstrated through application to simulated and real data.