During the past several years a variety of methods have been developed to estimate the effective connectivity of neural networks from measurements of brain activity in an attempt to study causal interactions among distinct brain areas. Understanding the relative strengths and weaknesses of these methods, the assumptions they rely on, the accuracy they provide, and the computation time they require is of paramount importance in selecting the optimal method for a particular experimental task and for interpreting the results obtained. In this paper, the accuracy of the six most commonly used techniques for calculating effective connectivity are compared, namely directed transfer function, partial directed coherence, squared partial directed coherence, full frequency directed transfer function, direct directed transfer function, and Granger causality. These measures are derived from the coefficients and error terms of autoregressive models calculated using the dynamic autoregressive neuromagnetic causal imaging (DANCI) algorithm. These techniques were evaluated using magnetoencephalography recordings as well as several synthetic datasets that simulate neurophysiological signals, which varied on several parameters, including network size, signal-to-noise ratio, and network complexity, etc. The results show that Granger causality is the most accurate method across all experimental conditions explored and suggest that large multisensor data sets can be accurately analyzed using Granger causality with the DANCI algorithm.

past sever year varieti method develop estim effect connect neural network measur brain activ attempt studi causal interact among distinct brain area understand relat strength weak method assumpt reli accuraci provid comput time requir paramount import select optim method particular experiment task interpret result obtain paper accuraci six common use techniqu calcul effect connect compar name direct transfer function partial direct coher squar partial direct coher full frequenc direct transfer function direct direct transfer function granger causal measur deriv coeffici error term autoregress model calcul use dynam autoregress neuromagnet causal imag danci algorithm techniqu evalu use magnetoencephalographi record well sever synthet dataset simul neurophysiolog signal vari sever paramet includ network size signaltonois ratio network complex etc result show granger causal accur method across experiment condit explor suggest larg multisensor data set can accur analyz use granger causal danci algorithm