Partial Directed Coherence (PDC) is a powerful estimator of effective connectivity. In neuroscience it is used in different applications with the aim to investigate the communication between brain regions during the execution of different motor or cognitive tasks. When multiple trials are available, PDC can be computed over multiple realizations, provided that the assumption of stationarity across trials is verified. This allows to improve the amount of data, which is an important constraint for the estimation accuracy. However, the stationarity of the data across trials is not always guaranteed, especially when dealing with patients. In this study we investigated how the inter-trials variability of an EEG dataset affects the PDC accuracy. Effects of density variations and of changes of connectivity values across trials were first investigated with a simulation study and then tested on real EEG data collected from two post-stroke patients during a motor imagery task and characterized by different inter-trials variability. Results showed the effect of different factors on the PDC accuracy and the robustness of such estimator in a range of conditions met in practical applications.