Benchmarking nonparametric Granger causality: Robustness against downsampling and influence of spectral decomposition parameters

Abstract

Brain function arises from networks of distributed brain areas whose directed interactions vary at subsecond time scales. To investigate such interactions, functional directed connectivity methods based on nonparametric spectral factorization are promising tools, because they can be straightforwardly extended to the nonstationary case using wavelet transforms or multitapers on sliding time window, and allow estimating time-varying spectral measures of Granger–Geweke causality (GGC) from multivariate data. Here we systematically assess the performance of various nonparametric GGC methods in real EEG data recorded over rat cortex during unilateral whisker stimulations, where somatosensory evoked potentials (SEPs) propagate over known areas at known latencies and therefore allow defining fixed criteria to measure the performance of time-varying directed connectivity measures. In doing so, we provide a comprehensive benchmark evaluation of the spectral decomposition parameters that might influence the performance of wavelet and multitaper approaches. Our results show that, under the majority of parameter settings, nonparametric methods can correctly identify the contralateral primary sensory cortex (cS1) as the principal driver of the cortical network. Furthermore, we observe that, when properly optimized, the approach based on Morlet wavelet provided the best detection of the preferential functional targets of cS1; while, the best temporal characterization of whisker-evoked interactions was obtained with a sliding-window multitaper. In addition, we find that nonparametric methods provide GGC estimates that are robust against signal downsampling. Taken together our results provide a range of plausible application values for the spectral decomposition parameters of nonparametric methods, and show that they are well suited to characterize time-varying directed causal influences between neural systems with good temporal resolution.