To better explore the abundant information of brain regions involved in functional magnetic resonance imaging (fMRI) data during task status,especially the pattern of information transfer concerned in effective connectivity,many ways and methods have been used,such as psycho-physiological interactions (PPI),structural equation modeling (SEM),multivariate autoregressive models (MAR),dynamic causal modeling (DCM) and granger casual analysis (GCA).Because it roles in neuron-level directly and takes into account the nonlinear and dynamic properties of nervous system,DCM is one of the most widely used method in effective connectivity study.In DCM,Bayesian estimation is used to evaluate the intrinsic connectivity among selected regions of interest (ROI),in which Bayes factors are used to compute different neuro-physiological models with intrinsic connectivity structures and then to select the optimal model.However,DCM is a modeldriven method essentially.Therefore,the wrong assuming models will draw the wrong conclusions in DCM.GCA is a data-driven method,which based on multiple linear regressions for investigating whether one time series could correctly predict another.This paper presents the methodology that combining GCA to DCM to improve the accuracy of DCM and reduce the workload of users.Our established procedures were to: 1) Extract the time series of ROI,and adopt surrogate data methodology to generate a certain number of surrogate time series for each ROI; 2) Conduct pair-wise GCA of surrogate time series by F value which is based on the decrease of the variance of residual to investigate the causal effect; 3) Sort F value from low to high,and get the threshold T with confidence interval from 0.95 to 1,then determine the existence of casual relationship when the F value of two initial ROI time series greater than T; 4) Consider the GCA compute result as the prior model architecture of DCM; 5) Use Bayes factors to select the optimal model.