Daniele Orzechowski of the Federal University of Santa Catarina and Ronaldo Martins da Costa of the Federal University of Goiás have built and released a containerised pipeline for Linear Mixed-Effects (LME) analysis of task fMRI data, and they have stress-tested it against one of the largest neuroimaging datasets in existence: the Human Connectome Project. Their work, published in September 2026, is notable not only for what the new model finds but for how transparently the tool reports its own behaviour. The pipeline requires the user to declare which fitting engine is being used, refuses to silently substitute one for another, versions every output file so repeated runs cannot overwrite each other, and writes a machine-readable manifest recording the software versions, hardware, and parameters behind every step. In a field still reckoning with reproducibility failures, those design choices are as consequential as the statistics themselves.

The validation is unusually thorough. The authors cross-validated their Python implementation, built on the statsmodels library, against the R reference standard lme4 by fitting identical data from 21 brain regions under both engines in separately archived executions. The two implementations agreed on the intraclass correlation coefficient, a key variance measure, to a mean absolute difference of 2.2 multiplied by ten to the power of minus six, with significance concordance in all 21 regions. Monte Carlo simulations across five variance regimes confirmed that the pipeline recovers known parameters with only tiny biases. The only practical divergence between the two backends was inferential rather than numerical: R’s lmerTest applies Satterthwaite degrees of freedom while statsmodels uses Wald z-tests, which placed a single region, the right hippocampus, on opposite sides of the 0.05 threshold even though the underlying estimates agreed to within one part in a million.

For the case study, the team turned to the emotion task from the Human Connectome Project S1200 release, in which participants view fearful faces and match neutral shapes in a blocked design. After quality control excluded three runs with excessive head motion, 142 healthy adults aged 22 to 35 contributed 283 functional runs acquired on a 3 Tesla scanner with 2-millimetre resolution. Each run produced a beta map for the fearful-face-versus-shape contrast, and all 283 maps entered the mixed-effects model, which modeled a fixed intercept, a fixed effect for run type, a random intercept per subject, and residual noise. The intraclass correlation coefficient then quantified how much of the variance was between subjects rather than within them.

The headline result is a surprise that reframes a long-running methodological debate. Compared head to head over an identical mask of 215,265 voxels with the same false discovery rate correction, the full LME model flagged 74,004 significant voxels while the conventional GLM flagged 108,627, a ratio of 1.47, with a Dice overlap of 0.77. That looks like the familiar story of mixed-effects models being more conservative. But the authors then fitted an intermediate LME model that kept the hierarchical structure yet dropped the run covariate, and it produced 109,273 significant voxels, almost indistinguishable from the GLM, with a correlation of 0.994 and a Dice of 0.993. The entire sensitivity difference, in other words, comes from the run covariate absorbing systematic run-related variance, not from hierarchical variance decomposition itself. Previous reports of reduced LME sensitivity may have conflated two distinct effects.

The run effects themselves proved scientifically interesting. They were significant in 13 of the 21 regions of interest, with the largest in the right amygdala, and they survived adjustment for framewise displacement, a head-motion measure, at the group level, on top of the twelve motion regressors already included in every first-level model. Because the Human Connectome Project acquires its two runs in a fixed order without counterbalancing, the authors caution that this effect may partly reflect practice, adaptation, or fatigue rather than the phase-encoding direction of acquisition, and they suggest reading it as a run-and-order effect. Still, the finding demonstrates the kind of signal that summary-statistics approaches simply average away.

Perhaps the most reusable product of the study is a set of whole-brain, voxelwise ICC maps covering all 215,265 voxels, released unthresholded. The maps reveal a pronounced subcortical-to-cortical gradient in within-subject reliability: the mean voxelwise ICC was 0.199, with values reaching 0.412 in the left amygdala and around 0.37 in the amygdala on average, while lateral frontal cortex sat near 0.10 to 0.14 and the frontal pole was effectively zero. High-ICC regions such as the amygdala are the ones most suitable for brain-behavior correlation studies and biomarker development, whereas low-ICC regions show high within-subject variability that conflates genuine neural fluctuation, measurement noise, and model misspecification. The authors also quantify how imprecise two-run ICC estimates are, with a simulation root-mean-square error of 0.081, improving substantially with three or more runs per subject.

Computationally, the pipeline makes whole-brain LME genuinely feasible. Benchmarked on synthetic data, the mixed model cost 19 times more per voxel than the GLM, 37.2 milliseconds against 1.99, but on real data the full analysis of 215,265 voxels completed in 5 hours and 41 minutes on twelve processor cores, with 100 percent convergence as read from the optimizer’s own flag. The authors note that convergence rates computed merely from the absence of exceptions, a common shortcut, overstate model fit. Stability checks reinforced confidence in both methods: split-half reliability exceeded 0.98 for the GLM and the LME alike, and leave-one-out cross-validation showed the group estimates were robust to dropping any individual subject.

So when should researchers abandon the GLM? The study’s answer is nuanced. For single-run designs or rapid exploratory work, the GLM remains a practical and nearly equivalent choice, since hierarchical modeling alone produces negligible divergence from summary statistics. LME earns its cost when multiple runs are available, when run-level covariates need explicit modeling, when conservative inference is a priority, or when within-subject reliability is itself the question. The decomposition strategy the authors propose, fitting full, intermediate, and standard models over an identical voxel set, generalises to any run-level covariate and can serve as a routine diagnostic for attributing sensitivity differences between statistical frameworks.

The pipeline is released under an MIT license with a Dockerfile and pinned environment manifests, and it is designed to travel: the task labels and atlases are configuration values rather than code, the statistical core is indifferent to whether data arrive as volumes or surface vertices, and the authors flag the two dataset-specific conventions, motion-regressor file layouts and rotation units, most likely to break outside the Human Connectome Project. Limitations remain, including a single contrast and dataset, a random-intercept-only specification forced by the two-run design, and volumetric rather than surface-based analysis. But the message to the field is clear: mixed-effects modeling of task fMRI is no longer a statistical luxury or a computational gamble. It is a validated, benchmarked, reproducible tool that reveals what the standard approach has been quietly discarding, and it now runs on hardware most labs can access overnight.

Subject of Research: Linear mixed-effects modeling of task fMRI data from the Human Connectome Project

Article Title: Beyond the General Linear Model: Linear Mixed-Effects Modeling of fMRI Data from the Human Connectome Project

Article References: Orzechowski, D., & Martins da Costa, R. (2026). Beyond the General Linear Model: Linear Mixed-Effects Modeling of fMRI Data from the Human Connectome Project. Neuroinformatics, 24(4), Article 64. https://doi.org/10.1007/s12021-026-09820-2

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09820-2

Keywords: fMRI, linear mixed-effects models, General Linear Model, Human Connectome Project, intraclass correlation coefficient, neuroimaging, reproducibility, statsmodels, lme4, emotion task, brain reliability, Neuroinformatics