Computational NeuroscienceDoes partial Granger causality really eliminate the influence of exogenous inputs and latent variables?

Partial Granger causality was introduced by Guo et al. (2008) who showed that it could better eliminate the influence of latent variables and exogenous inputs than conditional G-causality. In the recent literature we can find some reviews and applications of this type of Granger causality (e.g. Smith et al., 2011; Bressler and Seth, 2010; Barrett et al., 2010). These articles apparently do not take into account a serious flaw in the original work on partial G-causality, being the negative F values that were reported and even proven to be plausible. In our opinion, this undermines the credibility of the obtained results and thus the validity of the approach.Our study is aimed to further validate partial G-causality and to find an answer why negative partial Granger causality estimates were reported. Time series were simulated from the same toy model as used in the original paper and partial and conditional causal measures were compared in the presence of confounding variables. Inference was done parametrically and using non-parametric block bootstrapping.We counter the proof that partial Granger F values can be negative, but the main conclusion of the original article remains. In the presence of unknown latent and exogenous influences, it appears that partial G-causality will better eliminate their influence than conditional G-causality, at least when non-parametric inference is used.

► Partial Granger causality is in theory better able to eliminate the influence of exogenous inputs and latent variables. ► We showed that in practice, much depends on the statistical approach used to construct confidence intervals. ► Non-parametric techniques are often the only possibility when the true causal network is not known. ► These techniques, such as the non-parametric bootstrap, have parameters such as the window size that in their turn influence results. ► Researchers should be aware that these practical issues are as important as theoretical considerations in their choice between conditional and partial granger causality.