The reliability of conditional Granger causality analysis in the time domain
Background. Brain function requires a coordinated flow of information among functionally specialized areas. Quantitative methods provide a multitude of metrics to quantify the oscillatory interactions measured by invasive or non-invasive recording techniques. Granger causality (G-causality) has emerged as a useful tool to investigate the directions of information flows, but challenges remain on the ability of G-causality when applying on biological data. In addition it is not clear if G-causality can distinguish between direct and indirect influences and if G-causality reliability was related to the strength of the neural interactions. Methods. In this study time domain G-causality connectivity analysis was performed on simulated electrophysiological signals. A network of 19 nodes was constructed with a designed structure of direct and indirect information flows among nodes, which we referred to as a ground truth structure. G-causality reliability was evaluated on two sets of simulated data while varying one of the following variables: the number of time points in the temporal window, the lags between causally interacting nodes, the connection strength between the links, and the noise. Results. Results showed that the number of time points in the temporal window affects G-causality reliability substantially. A large number of time points could decrease the reliability of the G-causality results, increasing the number of false positive (type I errors). In the presence of stationary signals, G-causality results are reliable showing all true positive links (absence of type II errors), when the underlying structure has the delays between the interacting nodes lower than 100 ms, the connection strength higher to 0.1 time the amplitude of the driver signal and good signal to noise ratio. Finally, indirect links were revealed by G-causality analysis for connection strength higher than the direct link and lags lower than the direct link. Discussion. Conditional multivariate vector autoregressive model was applied to 19 virtual time series to estimate the reliability of the G-causality analysis on the identification of the true positive link, on the presence of spurious links and on the effects of indirect links. Simulated data revealed that weak direct but not weak indirect causal effects could be identified by G-causality analysis. These results demonstrate a good sensitivity and specificity of the conditional G-causality analysis in the time domain when applied on covariance stationary, non-correlated electrophysiological signals.