An Empirical Comparison of Three Methods to Establish Directional Relationships Using Multivariate Time Series: Effective Connectivity Analysis with Functional Magnetic Resonance Imaging Studies

Researchers in a variety of disciplines are interested in methods to establish causal links among several constructs. Statistical methods for modeling causal relations among brain regions in neuroscience are referred to as effective connectivity (EC) methods. Three popular effective connectivity methods are multivariate autoregressive modeling (MAR) with Granger Causality testing, structural equation modeling (SEM), and dynamic causal modeling (DCM). I conducted a head-to-head comparison of these three methods using both empirical and simulated data. The factors manipulated in the simulation study included (1) causality definitions (2) neural delays (3) variations in hemodynamic delay and (4) sampling frequencies. Results showed that none of the three EC methods showed satisfactory performance in terms of type I error rates and power across all conditions. With respect to interregional connectivity, the MAR method outperformed the other two EC methods. For bilinear effects which are the induced change by external stimuli, the SEM and the DCM method had higher average power than the MAR method. Regarding the direct effects of experimental inputs, the DCM method outperformed the other two methods. In the empirical example of integration of olfactory-visual threats, the three EC methods provided different connectivity patterns.