New weights gave lower maximum contrast variance than AR(1) in 24 fixed AOMIC designs over damped cosine mixtures with known variance=1, lag-one covariance=0.4, and q=0.9

We tested new unbiased linear contrast weights in 24 fixed AOMIC working memory designs. The largest allowed variance was 19.8 to 37.6 percent lower than with AR(1) weights. The mean reduction was 32.8 percent. The assumed noise class contained all damped cosine covariance mixtures with q=0.9. Noise variance 1 and covariance 0.4 at one original scan lag were known. The AR(1) comparison used correlation 0.4 and kept scan gaps. Every primary fit had a lower and upper bound gap of at most 0.02 percent. If AR(1) was exactly correct, the new weights increased variance by 12.6 percent on average. Five q values gave 120 fits, including an exact AR(1) control. The study used public task events, motion tables, and scan metadata. It used no measured brain signal values. The package preserves sources, code, bounds, weights, checks, and reproduction instructions. Independent numerical checks passed. A resumed repeat reproduced all 124 result files byte for byte.

minimax estimation · covariance uncertainty · contrast variance · generalized least squares · AR(1) · fMRI · working memory · AOMIC