Maximum Monte Carlo tests controlled false rejection on a stationary Gaussian AR(1) grid and failed under one unequal variance model in 24 AOMIC designs

A maximum Monte Carlo test has a 5% false rejection bound under stationary Gaussian AR(1) noise. The true correlation must be in the stated grid of 21 values. The design, contrast, and scan mask are fixed. We applied this established method to 12 AOMIC PIOP1 designs and 12 PIOP2 designs selected in earlier work. At correlation 0.6, fitted Student tests gave mean null rejection rates of 5.77% in PIOP1 and 5.57% in PIOP2. The maximum test gave 4.70% and 4.61%. Power was 2.59 and 2.35 percentage points below the known covariance controls. The power test used a contrast equal to twice the known covariance standard error. Two new calibration banks gave similar rates. One fixed Gaussian model with unequal scan variances gave 12.44% and 13.04% rejection for the maximum test. Correct covariance controls gave 4.80% and 5.08%. This artificial model is outside the stationary noise proof. The study measures simulated responses in real public designs. It does not measure neural effects in participants. The package preserves source tables, pinned URLs, hashes, code, a proof, and plans. It also preserves fitted statistics, thresholds, results, and instructions to reproduce the study.

Monte Carlo test · false rejection · AR(1) covariance · REML · fMRI · AOMIC · simulation calibration