Benchmarking surrogate nulls for neural–choice dependence under drift and selection

Abstract

Pre-stimulus decoding tests whether neural activity is associated with an upcoming choice, but drift and response-dependent trial inclusion complicate interpretation. We benchmark label-surrogate procedures using simulations and exploratory analyses of public Steinmetz and International Brain Laboratory recordings. We distinguish source-population independence from dependence induced by selection. Under independent drift, circular shifts rejected in 7.65% of 2,000 simulated sessions (95% Wilson interval 6.56–8.90%), above nominal 5%. At applied stratum lengths, no evaluated label-surrogate procedure met the stated tolerance for excess rejection across all arms. With source-population coupling absent, strong selection raised circular-shift and action-kernel rejection to 43.10% and 40.90%. Reapplying the known selection mechanism reduced these rates to 9.35% and 5.30%; the circular procedure remained inflated. A separate normalization audit quantified the size and power of implemented-rank and conservative truncated-shift procedures. Recording analyses were descriptive, and fitted-retention results were model-dependent sensitivity analyses. The benchmarks identify failure modes and conditional guidance for matching the inferential target, temporal null, inclusion mechanism and scored length. They establish neither a calibrated correction for these recordings nor the absence of weak pre-stimulus signals.