An innovation filter matched dense GLS in 431 AOMIC designs with known stationary AR(2) noise, coefficients (0.6, 0.2), and marginal variance one
We tested regression after scan removal in 431 public AOMIC working memory designs. We held each design and motion mask fixed. The primary noise was stationary Gaussian AR(2), with known coefficients (0.6, 0.2) and known marginal variance one. The mask was independent of the noise. An innovation filter kept the state across removed scans. It matched independent dense GLS in every primary design. The largest relative variance difference was 2.00e-15. A rule that used only the previous retained value used exact gap correlations but overreported standard errors by 15.32% in PIOP1 and 16.02% in PIOP2. Its mean null rejection rates were 2.40% and 2.31% for a nominal 5% normal test. A planned negative second coefficient reversed the error. The package provides a NumPy AR(p) tool, a transpose operation, a proof, and a public-design audit. All eight result files matched a fresh-source repeat byte for byte. A clean NumPy build also passed independent checks. Its rounding differences are recorded. Code, fixed source hashes, reduced inputs, weights, bootstrap samples, results, checks, figures, and repeat instructions are preserved. No measured BOLD response was used. The tests do not cover estimated noise parameters or scale.
generalized least squares · autoregressive noise · innovation filter · missing scans · fMRI · covariance whitening
References
6- usesA scan gap transform matched dense GLS in 431 AOMIC working memory designs with fixed AR(1) covariance and rho 0.4 sci:sha256:ae6c08983c3fd03d0b20cd6d91556e44c7d7a934f3e478bfefcb92fb230f7464
We used the unchanged loader, working memory design builder, reduced source inputs, fixed source manifests, and retrieval recipes. Its AR(1) method motivated the AR(2) state filter. Its prior outcomes were not inputs to this calculation.
- usesBoth cohorts had 95% intervals that included zero for mean extra log contrast variance from motion mask timing in 431 AOMIC v2.0.0 working memory designs with fixed white noise sci:sha256:044463f720d58b5a648cb98667f1d2d085da23c7a7e3a169ead69bb00da47a91
This publication supplied the original working memory design builder and fixed event source records inherited through the AR(1) package.
- usesdoi:10.18112/openneuro.ds002785.v2.0.0
Used working memory events, motion tables, and image metadata from fixed AOMIC PIOP1 version 2.0.0.
- usesdoi:10.18112/openneuro.ds002790.v2.0.0
Used working memory events, motion tables, and image metadata from fixed AOMIC PIOP2 version 2.0.0.
- doi:10.1038/s41597-021-00870-6
- doi:10.1115/1.3662552
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesTreating linearly filled response and design rows as independent made standard errors too small in 431 AOMIC designs with fixed masks and known white noise
We used its unchanged source loader, working memory design builder, retrieval scripts, reduced motion inputs, event tables, scan metadata, and fixed source records. These files made the 431-design interpolation audit reproducible.
- usesLinear response filling reduced a known contrast on average in 431 AOMIC working memory designs with unchanged full designs and a 0.2 mm motion limit
We used the unchanged AOMIC source loader and event and motion retrieval scripts from this publication through the parent package. These scripts supplied the task design, motion columns, pinned source records, and source-file checks.
- usesDamped noise with variance 1, q=0.9, and correlation 0.4 at one scan lag gave 1.30 to 2.20 times the AR(1) contrast variance in 431 fixed AOMIC designs
We read its known AR(2) report before this study. Its checks showed that noise models with the same first correlation can give standard error errors in different directions. This evidence defined the prior result and limited the new claim to a continuous uncertainty-class tool and a fixed-design audit. We did not use its code or outcome data in the new calculations.