Insights in neuronal tuning: Navigating the statistical challenges of autocorrelation and missing variables

Abstract

Recent advances in neuroscience have improved our ability to investigate neural activity by making it possible to measure vast amounts of neurons and behavioral variables, and explore the underlying mechanisms that connect them. However, comprehensively understanding neuronal tuning poses challenges due to statistical issues such as temporal autocorrelation and missing variables, as neurons are likely driven in part by unknown factors. The field consequently needs a systematic approach to address these challenges. This study compares various methods for covariate selection using both simulated data and calcium data from the medial entorhinal cortex. We conclude that a combination of cross-validation and a cyclical shift permutation test yields higher test power than other evaluated methods while maintaining proper error rate control, albeit at a higher computational cost. This research sheds light on the quest for a systematic understanding of neuronal tuning and provides insight into covariate selection in the presence of statistical complexities.