Fixed and refitted OLS correlations differed by more than 0.01 after one-scan removal in 11 of 12 AOMIC task runs

We tested removal of each scan in 12 fixed AOMIC working-memory runs. We used 100 cortical regions. The primary design used unweighted ordinary least squares (OLS), six motion columns, an intercept, and 128-second cosine drift. We compared a new fit on retained scans with removal of a residual row from the first fit. Both results used centered residuals. The largest absolute difference was 0.009454 to 0.061473 across runs. Eleven runs had a difference above 0.01. A fixed-fit screen omitted 104,962 of 746,484 scan-pair cases whose refitted correlation change was at least 0.01. These dependent counts are descriptive. An established regression identity gives all removal results from one fit. Separate retained-row fits checked every one of 9,263 removals across five fixed designs. The largest correlation mismatch was 4.44e-15. All original surface files and selected motion values passed source checks. The package preserves code, input arrays, source recipes, full results, and repeat instructions. This fixed sample gives no population estimate or proof of neural accuracy.

functional MRI · ordinary least squares · scan removal · nuisance regression · correlation audit