Prefrontal MRI patterns fail to deliver robust cognition links

An exploratory analysis of structural MRI data from 743 healthy young adults found that patterns of variation across the prefrontal cortex did not yield a robust, prefrontal-specific link with cognition. The strongest candidate was an unexpected association between greater orbitofrontal surface area and faster performance on a working-memory task, but the relationship weakened and lost statistical significance when global brain surface area was removed. Two other associations—one involving lateral prefrontal thickness and inhibitory control, another involving orbitofrontal thickness and response speed across gambling contexts—were nominally significant before correction for multiple comparisons and disappeared under key robustness checks. The authors therefore described all three findings as preliminary effects requiring replication rather than evidence that distinct prefrontal networks support specific mental abilities.

The study used open-access Human Connectome Project data from participants aged 22 to 36 and older. After researchers excluded 21 people with missing analysis variables, the final sample included 743 adults, representing 97.3% retention of the eligible group. Magnetic resonance images had been acquired on a 3 Tesla Siemens Skyra scanner, and the cortex was reconstructed with FreeSurfer 5.3.0-HCP. Using the Desikan–Killiany atlas, the researchers extracted thickness and surface-area measurements from 22 bilateral prefrontal regions, including lateral frontal, orbitofrontal and anterior cingulate areas. The analysis was cross-sectional, so it could identify associations but not determine whether brain structure influences behavior, behavior influences brain structure, or both reflect shared genetic or environmental influences.

Rather than examining each brain region separately, the researchers used principal component analysis to identify recurring patterns in which prefrontal measurements covaried across people. This structural covariance approach describes how morphology varies together in a population; a person’s component score represents how strongly that individual expresses a group-wide pattern, not the strength of that person’s direct anatomical connections. Before the analysis, age, sex and intracranial volume were removed from the regional measurements. An oblique rotation was then used because neighboring prefrontal regions may share anatomical features and can legitimately covary. The investigators also examined whether the component solutions remained stable when the sample was repeatedly resampled.

The retained cortical-thickness solution contained three components and accounted for 53.7% of the variance. The largest, RC1, grouped the superior and middle frontal regions with the pars opercularis and was labeled global or lateral prefrontal thickness. RC2 combined the lateral and medial orbitofrontal cortex with the pars orbitalis, while RC3 was centered on the anterior and posterior portions of the anterior cingulate cortex. This three-part anatomy broadly resembled established lateral, orbital and medial divisions of the prefrontal cortex, supporting the study’s first hypothesis only partly. Parallel analysis actually favored a two-component solution, so the three-component arrangement was retained because it had been specified in advance. Bootstrapping also revealed uneven stability: congruence fell below 0.85 in 19.4% of resamples for RC1 and 28.0% for RC2, compared with 0.7% for RC3.

The thickness analysis initially produced two associations with cognition. People expressing more of the lateral-prefrontal thickness pattern RC1 obtained slightly lower scores on the NIH Toolbox Flanker test of inhibitory control and attention: the standardized coefficient was β = −0.103, with p = 0.027 and partial eta-squared of 0.007. The direction was opposite to the prediction that greater lateral-prefrontal expression would accompany better performance, and the result did not survive false discovery rate correction across the four directional hypothesis tests, where q = 0.054. It also had a straightforward alternative: RC1 correlated strongly with whole-cortex mean thickness, at r = 0.85. Mean thickness alone predicted lower Flanker scores, and after global thickness was regressed out of the regional data before repeating the principal component analysis, the RC1 association was no longer significant. The evidence therefore favored a global rather than specifically prefrontal relationship, although these sensitivity analyses were added after review and were not prespecified.

A second thickness result concerned the Gambling RT Difference, calculated as the median time to make a “larger” guess in reward blocks minus the corresponding time in punishment blocks. Because block context was not cued before each guess, the measure was interpreted as the carryover of recent outcomes into response speed, not as a direct measure of whether participants preferred larger rewards to larger punishments. Greater expression of the orbitofrontal thickness component RC2 was associated with a smaller difference, with β = −0.103, p = 0.025 and partial eta-squared of 0.007. Like the Flanker result, it failed false discovery rate correction at q = 0.054 across the prespecified family. It also disappeared under robust regression, after influential cases were removed, and when global mean thickness was regressed out before the component analysis. The corresponding relationship for “smaller” guesses pointed in the same direction but was not significant, so the data did not establish a reliable orbitofrontal effect.

The most notable behavioral result came from a separate analysis of surface area, which produced a three-component solution accounting for 39.9% of the variance. Its topology only partly matched the thickness solution: the largest Tucker’s congruence between corresponding components was 0.81, indicating that the two measures captured related but distinguishable patterns. The orbitofrontal surface-area component AC2 was associated with faster 2-Back reaction time, the speed of correctly processing working-memory information presented two trials earlier. The standardized coefficient was β = −0.125, p = 0.003 and partial eta-squared of 0.012. It survived false discovery rate correction across 12 surface-area component tests, where q = 0.032, but not across all 24 thickness and surface-area component tests, where q = 0.064. The association was not predicted in the original hypotheses and was therefore exploratory.

Most robustness checks supported the AC2 association. It remained significant with heteroskedasticity-consistent standard errors, robust regression, removal of influential cases, and adjustment for intracranial volume. It also was not a simple speed–accuracy tradeoff: greater AC2 expression was associated with higher 2-Back accuracy, and including accuracy as a covariate reduced but did not eliminate the reaction-time relationship. Yet the regional specificity of the finding remained uncertain. AC2 correlated with total cortical surface area at r = 0.38, and total surface area alone was associated with faster 2-Back reaction time before adjustment for intracranial volume. When total surface area was regressed out of the regional measurements before repeating the principal component analysis, the reconstructed AC2 closely resembled the original component, with Tucker’s congruence of 0.91, but its reaction-time coefficient was cut roughly in half to β = −0.065 and became non-significant at p = 0.078. The result could therefore not be cleanly separated from global brain size.

Several limitations constrain the interpretation. The HCP sample includes monozygotic and dizygotic twins and non-twin siblings, but the open-access release did not contain family identifiers, so related participants could not be modeled with mixed-effects analyses. Treating them as independent likely makes the reported p and q values anti-conservative. The sample also excluded education data available only in the restricted release, limiting adjustment for a factor known to influence cognitive performance. Some behavioral measures had limited reliability or could not test their intended hypotheses: for example, the block-level gambling choice score lacked the trial-level information needed to assess win-stay and lose-switch behavior. In addition, FreeSurfer thickness is an indirect morphometric measure influenced by myelination, neuronal structure, partial-volume effects and surface reconstruction, while the atlas groups some cytoarchitecturally distinct areas into coarse parcels. All behavioral models explained no more than 2.7% of variance, and the magnitude of the reported effects was small by conventional standards. Replication with independent samples, explicit family-structure modeling, education adjustment, finer anatomical parcellation and complementary diffusion or myelin imaging would be needed before the findings could be treated as evidence of distinct structure–function relationships in the prefrontal cortex.

Subject of Research: Exploratory analysis of prefrontal cortical morphometric covariance and cognitive performance in healthy young adults

Article Title: Prefrontal cortical morphometric covariance and cognitive task performance in healthy young adults: an exploratory analysis of Human Connectome Project data

Article References: Chivers, R. (2026). Prefrontal cortical morphometric covariance and cognitive task performance in healthy young adults: an exploratory analysis of Human Connectome Project data. Frontiers in Human Neuroscience, 20, Article 1976371. https://doi.org/10.3389/fnhum.2026.1976371

Image Credits: AI Generated

DOI: 10.3389/fnhum.2026.1976371

Keywords: Human Connectome Project, prefrontal cortex, structural MRI, cortical thickness, brain surface area, working memory, inhibitory control, orbitofrontal cortex, morphometric covariance, cognitive neuroscience, Prefrontal, cortical