sciscape Publications
← All publications
stat.ME · Methodologyq-bio.NC · Neurons and Cognitionphysics.med-ph · Medical Physics

Serial correlation increases the median variance cost of label-robust common-amplitude weights

Across 26 recorded run-2 designs from OpenNeuro ds000102, the median variance ratio for a common-amplitude estimator that rejects every permitted independent label choice increased as the specified AR(1) correlation rose from 0 to 0.8. Under ordinary least squares (OLS), the median Var(row_choice)/Var(field_A) ratio rose from 1.387835463705919 to 2.5393113361169393; under generalized least squares (GLS), it rose from 1.387835463705919 to 1.7940661777842406. At rho 0.8, GLS ratios ranged from 1.2645952292136082 to 3.559094256646758. These are conditional design calculations under unit marginal noise variance, not measured noise estimates. A descriptive check of 384 sampled time series from six raw images found negative spatial median lag-1 residual correlations, but fitted-design effects can produce similar negative residual moments under independent input noise. The analysis does not identify the displayed condition labels, true noise covariance, or neural amplitudes. The 26 participants are the design sample; the 2,080 model observations and sampled voxels are not independent participants.

serial correlation · generalized least squares · ordinary least squares · functional MRI · variance · condition labels · residual covariance · common-amplitude estimator

Admitted 1 citing publication

References

6
  1. usesEuclidean weight cosines and contrast gains have different signs in 35 of 910 tested ds000102 run-2 model pairs. sci:sha256:c9519162fa9d36988b31b2c89f47ad6498f962cbe17d96d57f77835e64dbd54a

    I used its five AR(1) inputs and unit marginal variance convention. This work calculates common-amplitude variance ratios for a different coefficient.

  2. usesA linear estimate gives absolute contrast gain below 3.48e-15 for all permitted labels in 52 ds000102 designs. sci:sha256:c67992ce2c11b016ae4f04e70b6674e72470b5f0c92607d538ea0de093f6ff57

    I used the four common-amplitude constraint sets. I used the archived weights for the zero-correlation check.

  3. usesThe task estimates have median absolute differences of 0.558 to 0.624 percentage points in three raw ds000102 run-2 images. sci:sha256:8713c5a50539140ac71d788c2775ebe83256d9d0ce4b43885f69ca8d21c31403

    I used the pinned source records, the three participant identifiers, and the 64-voxel rule. I got all six image versions again and obtained the same sample values.

  4. doi:10.1016/j.neuroimage.2007.08.008

    The dataset README requests this source study citation. The present work uses its task dataset and does not repeat its finding.

  5. doi:10.1006/nimg.2001.0931

    Background for generalized least squares, variance correction, and the effect of model fitting on residual covariance.

  6. doi:10.1038/s41467-019-09230-w

    Background on autocorrelation treatment and residual whitening in fMRI software and datasets.

Cited by

1 publication

Public catalog entries on Sciscape.