Studies of how childhood adversity shapes the developing brain increasingly rely on a single number: the brain-age gap. A model predicts a child's age from an MRI scan, and the difference between predicted and actual age is read as faster or slower maturation. Using this gap, Dai et al. reported a striking result in roughly 7,600 adolescents. Under deprivation and low socioeconomic status, boys' brains look younger than expected and girls' brains look older. But the gap is a model's error term. Adult audits such as Piçarra and Glocker show that brain-age models make systematically different errors for men than for women, and for Black than for White subjects. No one has checked whether the adolescent sex divergence is a fact about brains or a fact about the instrument.
This note proposes to treat several brain-age models as several imperfect rulers for the same hidden quantity, maturation. It borrows the logic of measurement invariance: the psychometric test of whether a questionnaire measures the same construct with the same scale and zero point in every group. The test asks whether the brain-age gap means the same thing across sex and race in childhood and adolescence. Only the part of an adversity effect that survives on an invariant latent maturation score would count as biology. The deliverable has three parts: a reusable invariance test for age clocks, a verdict on the sex-moderated adversity findings in the Adolescent Brain Cognitive Development (ABCD) cohort, and a demonstration that the usual fairness fix, equalizing error across groups, is the wrong target for this kind of science.
Core Claim
The fairness question that matters for developmental age-gap research is not whether each group gets the same mean absolute error (MAE). It is whether the model's error is independent of the exposure being studied within each group. A clock can have identical MAE for boys and girls and still produce a spurious sex-by-adversity interaction. This happens if its error depends on anatomical features, such as head size, tissue contrast, or motion, that differ by sex and also shift with deprivation.
Conversely, a clock trained to equalize subgroup error can erase genuine sex differences in maturation. One example is a model with demographic-specific normalization in the style of FairAdaBN. Girls enter puberty earlier, so a sex-conditioned clock absorbs exactly the signal that developmental vulnerability research is looking for. We therefore propose an alternative to both an unaudited clock and a "debiased" one. We build several clocks from inputs with largely separate error sources and fit a latent-variable model in which each clock's gap is a noisy indicator of one shared maturation factor. We then test whether the clocks' loadings and intercepts are equal across sex and race. Adversity effects are estimated on the latent factor, with non-invariant clocks explicitly modeled rather than silently averaged.
How This Differs From Existing Work
Two lines of work motivate the gap, and neither closes it. Piçarra and Glocker audited a 3D ResNet-34 brain-age model on 42,786 UK Biobank adults aged 44–82. They found significant MAE differences between Black and White subjects (p=0.022) and between males and females (p=0.008), and they found that a principal component of the penultimate layer separated subjects by race. They studied adults only and stopped at error rates, without asking what disparate error does to a downstream scientific estimate.
The ABCD studies supply the developmental question but treat the clock as fixed. Dai et al. used a frozen Simple Fully Convolutional Network regression model (SFCN-reg) trained on 53,542 lifespan scans and never reported its error by subgroup in ABCD. Chen et al. fine-tuned a model pretrained on the Philadelphia Neurodevelopmental Cohort and found that accelerated maturation tracks higher cognition. Dai et al. found that it tracks more externalizing problems. The two findings are in tension, and neither paper asks whether its gap measures the same construct as the other's.
A third line supplies tools we reuse or argue against. The tools we reuse are multi-model clock construction from Gong et al. and voxelwise local brain age from Popescu et al.. The tools we argue against are error-equalizing fairness methods: FairAdaBN, Fair Error-Bound Scaling, and latent orthogonalization. These were all built for 2D classification or segmentation, where the label is not itself the biology under study.
One negative result constrains the design. Du et al. showed that rebalancing brain MRI training data by sex and age did not remove sex disparities in reconstruction quality: female SSIM was 0.876 versus 0.868 for males, p<0.001 under every strategy. They attributed the gap to anatomy rather than sample counts. We take this as evidence that disparity is a property of what the scans look like, so we do not propose rebalancing as the fix.
Research Program
The work has four parts.
Build clocks with separate error sources. On ABCD, we build four age clocks with deliberately different inputs, trained only on children with low adversity scores.
- An SFCN ensemble on T1-weighted voxels, following Gong et al.
- A penalized linear model on FreeSurfer regional volume and thickness.
- A model on static functional network connectivity, the modality Chen et al. found to covary with structure (r=0.65–0.77).
- Regional gaps from a local brain-age U-Net.
The original frozen SFCN-reg clock is kept as a fifth indicator, so the Dai et al. analysis can be reproduced inside our model.
Fit the latent maturation model. Clock $k$'s gap for child $i$ in group $g$ is modeled as
$$\Delta_{ik} = \tau_{kg} + \lambda_{kg}\, M_i + \varepsilon_{ik},$$
The symbols mean the following:
- $M_i$ is latent maturation.
- $\tau_{kg}$ is a group-specific intercept: an offset the clock adds for group $g$ regardless of maturation.
- $\lambda_{kg}$ is a group-specific loading: how strongly the clock tracks maturation in that group.
- $\varepsilon_{ik}$ is clock-specific error.
Groups are sex crossed with parent-reported race/ethnicity. Invariance means $\tau_{kg}$ and $\lambda_{kg}$ do not depend on $g$. We test this in nested steps, first equal loadings and then equal intercepts. When a clock fails, we free its parameters for the offending group, which is the standard partial-invariance procedure. The model also yields, for each clock and group, how much of its gap is offset versus signal.
Re-estimate the adversity effects. We regress $M_i$ on the four adversity dimensions of Dai et al. (threat, deprivation, household instability, socioeconomic status), with sex interactions, site, and head motion as covariates. We then ask whether the sex divergence survives. The same latent score is used to revisit the Chen–Dai paradox: we check whether cognition and externalizing problems load on the same latent maturation, or on clock-specific residuals that one study's model happened to emphasize.
Account for puberty. Ages 9–14 straddle puberty, so we fit a second version in which chronological age is replaced by, or supplemented with, pubertal stage if available in the release we obtain. The aim is to separate "older-looking because further into puberty" from "older-looking because of adversity." The file establishes only that puberty timing varies across sexes, so this step depends on data access.
The central assumption is that the clocks' errors are conditionally independent given maturation and covariates. Head motion can contaminate both T1 and functional connectivity, which would break this assumption. We test it directly by checking residual correlations between clocks after including motion. If the residual correlation remains, we report the method as inapplicable for that pair of clocks rather than proceeding.
Baselines
Naive reference (do nothing). Reproduce the Dai et al. pipeline: frozen SFCN-reg, a single linear correction for regression toward the mean, and sex as a moderator. This is the analysis the field currently runs. Every other method is compared against the sex-by-adversity coefficients it produces.
Strongest existing practice. Fine-tune an SFCN ensemble on the low-adversity ABCD reference set and apply sex- and race-specific post-hoc age-bias correction. The ensemble mirrors the PAC 2019 winner, which reported MAE 2.90 years without bias removal and 2.95 years with bias removal at |r|<0.1 between gap and age (Gong et al.). These figures come from adults aged 17–90. For SFCN-reg error on ABCD, no reported result was found on Lacuna. We will report subgroup MAE ourselves before any downstream analysis. This baseline is strongest because group-specific correction is the cheapest fix that removes constant group offsets, which is exactly what the intercept terms in our model also capture.
Fairness-trained clock. A 3D SFCN with adaptive batch normalization conditioned on sex and race, plus a statistical-disparity loss, adapted from FairAdaBN. Fair comparison means the following:
- every clock sees the same reference subjects, preprocessing, and site splits;
- every clock gets the same hyperparameter search budget, tuned on reference-set MAE only, never on the downstream adversity coefficients.
Ablation. The latent model with all parameters forced equal across groups. This assumes invariance instead of testing it, and isolates whether the group-specific terms, rather than mere averaging over clocks, change the conclusions.
What counts as beating the baselines. In simulation with planted ground truth, the proposed estimator must recover the true sex-by-adversity interaction with lower bias and nominal confidence-interval coverage where the baselines fail. On real data there is no ground truth. "Beating" there means producing stable conclusions across reference-set resamples and passing the invariance diagnostics that the baselines cannot even run.
Why not just…?
Why not just add sex, race, and site as covariates and correct the gap separately per group? This is the most serious competitor, and it is our strongest baseline. It works if the clock's subgroup bias is a constant offset. It fails if the error slopes with exposure within a group, for example if deprivation shifts head size or motion and the clock misreads those shifts differently for boys and girls. Covariates on sex cannot remove an error that interacts with adversity. The simulation settles this. We plant three kinds of error: constant group offsets, group-specific scaling, and adversity-dependent error. If group-specific correction matches the latent model in all three, we will say the simpler method suffices and the contribution narrows to the diagnostic.
Why not just skip brain age and analyze FreeSurfer regions directly? This is also a real threat, because regional analyses avoid the residual-of-a-model problem entirely. Brain-age summaries persist because they collapse hundreds of regions into one interpretable, well-powered number. The test is whether regional sex-by-deprivation effects, after multiple-comparison correction, reproduce the latent-factor conclusion. If they do and the latent model adds nothing, we will report that. Popescu et al. found that global gaps wash out regional effects, with subcortical local gaps showing Cohen's d >1.5 for stable versus progressive cognitive impairment. That finding is why the regional clock is one of our indicators rather than a replacement for the framework.
Why not just train a fair clock that equalizes error across groups? Equal MAE is the wrong criterion here. Equalizing error by conditioning on sex can absorb real sex differences in maturation timing, the very effect under study. Fair Error-Bound Scaling and FairAdaBN target classification and segmentation, where the label is external to the group difference. We run the fairness-trained clock and check whether it shrinks the Dai et al. interaction toward zero even in simulations where we planted a true interaction. If it does, that shrinkage is evidence of erasure, not correction.
Why not just rebalance the training set? In brain MRI, Du et al. found that rebalancing left the sex gap significant and penalized the majority group. We include a rebalanced SFCN as a secondary check but do not expect it to differ from the strongest baseline.
Proposed Experiments and Decisive Checks
Simulation. We generate gaps from a known latent maturation with planted sex-by-adversity effects of zero and nonzero size, crossed with the three error types above. Error magnitudes are calibrated to the subgroup MAE differences we measure on ABCD's low-adversity reference set. Bias, interval coverage, and power are compared across the naive pipeline, the strongest baseline, the fairness-trained clock, the invariance-assuming ablation, and the full model.
First-stage audit on ABCD. We first audit every clock by subgroup on held-out reference children. We use balanced repeated subsampling and Kruskal–Wallis tests as in Piçarra and Glocker, which extends their adult audit into ages 9–14. We also run the invariance tests across sex and race.
Decisive check. We re-estimate the deprivation-by-sex and SES-by-sex effects on latent maturation. Three outcomes are possible:
- If both the latent model and the strongest baseline preserve the opposite-signed sex effects, the biological reading gains real support.
- If the effects vanish once the SFCN-reg clock's non-invariant intercepts are freed, the original finding was largely instrument error.
- If the effects survive but appear on only one clock, they are clock-specific.
Secondary check. We repeat the analysis with each clock left out in turn, to see whether any single input type drives the conclusion. We also report whether Chen et al.'s cognition association and Dai et al.'s externalizing association attach to the same latent factor.
All code and the invariance test will be released so that any study using an age clock in childhood can run the same audit before interpreting a gap.
References
- Dimensional adversity, brain-age, & mental health: Differences in male and female adolescents
- Analysing Race and Sex Bias in Brain Age Prediction
- FairAdaBN: Mitigating Unfairness with Adaptive Batch Normalization and Its Application to Dermatological Disease Classification
- Adolescent brain maturation associated with environmental factors: a multivariate analysis
- Optimising a Simple Fully Convolutional Network (SFCN) for accurate brain age prediction in the PAC 2019 challenge
- Local Brain-Age: A U-Net Model
- FairSeg: A Large-Scale Medical Image Segmentation Dataset for Fairness Learning Using Segment Anything Model with Fair Error-Bound Scaling
- Generalizing Orthogonalization for Models with Non-Linearities
- Unveiling Fairness Biases in Deep Learning-Based Brain MRI Reconstruction