Post-Stimulus Choice Decoding Is Stronger From Regional Firing Rates Than From Inter-Area Correlations

Neural decoders are often used to compare candidate representations of perceptual choice, but their interpretation is difficult when stimulus category, behavioral choice, and motor response are coupled by the task. I reanalyzed 39 simultaneous multi-region Neuropixels sessions from 10 mice performing a visual contrast-discrimination task. I compared two matched-dimensional feature families: regional mean firing rates and trial-wise inter-area correlations. Preprocessing was restricted to the training folds, performance was quantified by balanced accuracy, and inference was conducted on mouse-level summaries. Neither feature family decoded choice before stimulus onset (regional rates: 49.5%; inter-area correlations: 50.1%). After stimulus onset, regional firing-rate features were substantially more predictive than correlation features (0-250 ms: 64.2% versus 53.4%; 250-500 ms: 67.3% versus 50.1%; Holm-adjusted paired p = 0.0078 in both windows). The feature-family-by-time interaction was also reliable. Additional analyses showed that post-stimulus regional-rate decoding increased with stimulus discriminability and was much stronger on correct than error trials, indicating a substantial stimulus-linked component. Nevertheless, regional rates decoded choice above chance on equal-contrast trials, so stimulus side alone cannot explain the result. The correct-to-error stimulus-generalization analysis was inconclusive after multiplicity correction and was limited by the smaller number of error trials. Thus, the data support a time-dependent difference in decodability from these two feature representations, not a causal switch from distributed to local computation. The results also illustrate why raw accuracy, fixed stimulus-response mappings, and model-dependent decoder performance must be handled explicitly when inferring neural mechanisms.