The effect of filtering on Granger causality based multivariate causality measures.

In the past, causality measures based on Granger causality have been suggested for assessing directionality in neural signals. In frequency domain analyses (power or coherence) of neural data, it is common to preprocess the time series by filtering or decimating. However, in other fields, it has been shown theoretically that filtering in combination with Granger causality may lead to spurious or missed causalities. We investigated whether this result translates to multivariate causality methods derived from Granger causality with (a) a simulation study and (b) an application to magnetoencephalographic data. To this end, we performed extensive simulations of the effect of applying different filtering techniques and evaluated the performance of five different multivariate causality measures in combination with two numerical significance measures (random permutation and leave one out method). The analysis included three of the most widely used filters (high-pass, low-pass, notch filter), four different filter types (Butterworth, Chebyshev I and II, elliptic filter), variation of filter order, decimating and interpolation. The simulation results suggest that preprocessing without a strong prior about the artifact to be removed disturbs the information content and time ordering of the data and leads to spurious and missed causalities. Only if apparent artifacts like a current or movement artifact are present, filtering out the respective disturbance seems advisable. While oversampling poses no problem, decimation by a factor greater than the minimum time shift between the time series may lead to wrong inferences. In general, the multivariate causality measures are very sensitive to data preprocessing.