Performance evaluation of dereferencing methods for estimating information flow in laminar connectivity models

Multi-layered brain structures contain canonical microcircuits that specialize in region-specific function. Information flow across the layers is typically inferred using multivariate techniques that operate on local field potentials (LFPs). These methods (e.g., Granger Causality (GC)) are sensitive to the presence of a common reference that corrupts LFPs recorded with a multi-contact electrode and introduces spurious covariations. Using models of reference-noise corrupted signals with laminar interactions, we evaluated the efficacy of three dereferencing methods - bipolar subtraction, current source density (CSD), and common average referencing. We examined which method best recovered the underlying functional interactions between layers. Each dereferencing method introduced different types of error, often to alarming levels of false prediction. While CSD and bipolar subtraction methods performed best, they often predicted spurious connections, exhibited GC power in incorrect frequency bands, and/or missed salient relationships between layers. Though the confounds in this model may not be present in evaluations of functional connectivity between brain areas, these findings call for a reassessment of dereferencing methods used in the context of evaluating information flow within laminar neural tissue.