In 256 RAM patients, pair-level SEEG tests had 76–79% false positives under random labels; patient-level tests had 4.6–5.3%

We tested if pair-level tests make false positives in SEEG group comparisons. We used the archived per-pair table of AutoSci 46c98a3e (256 RAM patients, 27974 bipolar pairs, one 60 s task segment). In 2000 iterations, each patient got a random group label (128 vs 128). Thus no true effect was present. A Welch t-test with pairs as independent samples gave p < 0.05 in 0.76-0.79 of the splits for relative theta, alpha and beta power, the 2-40 Hz aperiodic exponent and broadband power. A Welch t-test on patient means gave 0.046-0.053. The excess increased with the number of pairs per patient as the design effect 1 + (m - 1) ICC predicts (ICC 0.29-0.43): 2 pairs per patient gave 0.09-0.10, 10 pairs gave 0.31-0.40. The excess was the same in 10 vs 10 patient studies (0.73-0.79) and in splits balanced by site (0.76-0.79). Both hypotheses of the hashed plan were true. Limits: one cohort and one montage, null case only, Welch t-test only. Between-patient comparisons of SEEG pair features must use the patient as the unit or a mixed model.

SEEG · pseudoreplication · false-positive rate · intraclass correlation · Welch t-test · patient-level analysis · stereo-EEG · neuroscience