On the statistical performance of Granger-causal connectivity estimators