Abstract
Interpreting correlations between neuronal activity and hemodynamic signals is complicated by their inherently strong autocorrelation. Standard parametric tests underestimate false positives, creating the appearance of widespread neurovascular coupling during rest. Here we present a surrogate-based statistical framework designed to calibrate inference in autocorrelated physiological signals. Using simultaneous recordings of cortical oxygen tension (PO₂), single-unit firing, and local field potentials (LFP) in awake rabbits, we applied amplitude-adjusted Fourier surrogates to generate null distributions that preserve temporal structure but remove cross-dependence. This workflow embeds lag optimization, controls for multiple comparisons across windows and units, and scales to population-level inference. Applying the method to 43 experiments, we found that PO₂ correlations with delta-band LFP and a minority of single neurons exceeded chance levels, while correlations with other LFP bands were not significant under surrogate testing. Aggregated activity such as multi-unit signals failed to predict PO₂, but small synchronized subpopulations produced robust associations, highlighting the role of limited synchronization rather than global activity. These findings refine resting-state neurovascular coupling: broad apparent correlations reduce to selective and reproducible effects once calibrated testing is applied. More broadly, the framework demonstrates how surrogate-based inference prevents misinterpretation of autocorrelated data and offers a generalizable approach for electrophysiology, neuroimaging, and other time-series domains where genuine interactions must be distinguished from random associations.