Three of 26 primary ds000102 two-run model pairs have negative noise correlations at a specified variance ratio of four.

Using ds000102 flanker event tables from 26 participants, we reproduced the increase in two-run Euclidean contrast-weight cosine in all participants across 17 tests. The median increased from 0.0551 for run 2 to 0.5258 for two runs; with identity covariance, no two-run cosine was negative. In a conditional extension with specified AR(1) covariance and a run-1/run-2 variance ratio of 4, three participants had negative two-run noise correlations at each of five tested autocorrelation values, while six had negative Euclidean cosines. Across 7,280 repeated model pairs, 652 noise correlations were negative and 304 pairs had opposite signs for the two metrics. These are deterministic model calculations from event tables, not measurements of brain signals. The covariance values were specified rather than estimated, condition labels differed in 304 run-2 trials, and the correct displayed labels remain unknown. The work does not establish a biological effect, general proof, or novelty.

functional MRI · contrast weights · generalized least squares · noise covariance · AR(1) · reproducibility · ds000102