Reliability of Neural Population Geometry
Recording yield confounds and session noise limit cross-region comparisons of neural population geometry
Session-to-session reliability of 20 geometric metrics computed from Neuropixels population recordings in mouse cortex (Steinmetz et al. 2019), measured by ICC(1,1) and tested against optogenetic silencing ground truth (Zatka-Haas et al. 2021) after controlling for neuron count and recording yield. Metrics below a Spearman–Brown-derived reliability threshold of ICC = 0.34 are excluded before correlation with silencing effects, attenuation ceilings bound the largest correlation observable given the measurement noise, and norm-based metrics, which inherit √n scaling with neuron count, are corrected by partialling out log neuron count.