The detection of directed interactions within networks derived from spatially highly resolved data, such as functional magnetic resonance imaging (fMRI) has been a challenging task for the last years. Commonly this is solved by restricting the analysis to a small number of representative network nodes (e.g. fMRI voxels), to regions of interest (e.g. brain areas) or by using dimension reduction methods like principal or independent component analysis. Recently, these problems have successfully been encountered by combining multivariate autoregressive models and parallel factor analysis. This approach involves a cascade of analysis steps, entailing a number of parameters that have to be chosen carefully. Yet, the question of an appropriate choice of analysis parameters has not been clarified so far—in particular for temporally varying models. In this work we fill this gap. Synthetic data with known underlying ground truth structure are generated to evaluate the correctness of results in dependence on the parameter choice. Resting state fMRI data are used to assess the influence of the involved parameters in the clinical application. We found that model residuals offer a good means for determining appropriate filter algorithm parameters; the model order should be chosen according to two aspects: the well-established information criteria and the fit between Fourier and estimated spectra.