Knowledge of neural interactions amongst cortical sites is important for understanding higher brain function. We studied such interactions using Granger causality (GC) to analyze auditory event-related potentials (ERPs) recorded directly and simultaneously from two physiologically identified and functionally interconnected auditory areas of cerebral cortex in human neurosurgical patients. Two methods of GC analysis were used and the results compared. Both approaches involved adaptive autoregressive modeling but differed from each other in other ways. Results obtained by using the two methods also differed. Fewer false-positive results were obtained using the method that suppressed the ERP non-stationarity and that expressed the GC as the sum of model coefficients, which suggests that this is the more appropriate approach for analyzing ERPs recorded directly from the human cortex.