Adolescent psychiatry has built a large literature on "chain" or serial mediation, the claim that a stressor such as bullying or childhood maltreatment harms mental health through an ordered cascade. The stressor first shifts emotion, the emotional shift changes behavior such as late-night phone use, and the behavior degrades sleep before symptoms appear. A separate machine-learning literature estimates a brain-age gap: a model trained on typical brains predicts age from an MRI scan, and the difference between predicted and actual age is read as a marker of advanced or delayed brain development.
The natural question is whether the brain-age gap is one of the dominoes in these cascades. The one large test, in 7,658 adolescents, found that it did not mediate adversity and later mental health. That study used a single mediator step, a single accuracy-optimized brain-age model, and pooled boys with girls even though it found effects of opposite sign by sex. This note proposes a longitudinal, pre-registered test on the same cohort. It asks whether the null survives three corrections applied together: adding the intermediate emotional and sleep steps, treating the choice of brain-age model as a measurement decision, and letting the effect differ by sex. It also reports an explicit upper bound if the null holds.
Core Claim
The null in Dimensional adversity, brain-age, & mental health can be explained in three ways that the field has not separated.
The first is that the design omits intermediate steps. The emotional mediation studies consistently route adversity through emotional and behavioral intermediaries, which a direct adversity-to-brain-age path would miss.
The second is measurement. Brain-age models with lower age prediction accuracy have higher sensitivity for disease detection shows that heavily regularized linear models give larger clinical group differences than deep networks. They do so by weighting diffuse global atrophy rather than localized age-specific features. The accuracy-optimized network (SFCN) that Shaul et al. used may therefore filter out exactly the variation a stress cascade would produce.
The third is cancellation. Shaul et al. found that deprivation and low socioeconomic status predicted younger-looking brains in exposed boys and older-looking brains in exposed girls. Pooled estimates can therefore average to zero.
Our claim is that a longitudinal chain ending in the change in brain-age gap exists only in a specific cell of this design, if it exists at all. That chain runs from baseline adversity, to later negative affect, to later sleep disruption, to the change in brain-age gap between scans. The cell is sex-specific and uses a strongly regularized model. The same study also has to show that, in that cell, the gap carries information beyond total gray matter volume. Otherwise "brain age" is only a relabeling of one volume measure.
What Prior Work Establishes and Leaves Open
The emotional mediation studies motivate the chain structure. Hu et al. found that emotion-regulation difficulty and depression fully mediated maltreatment and self-injury in 224 inpatients. Xin Zhang et al. found that phone addiction and sleep quality carried 64.01% of the effect of academic stress on depression, with sleep accounting for over 82% of the indirect effect. Li et al. split the path from bullying to self-injury into affect, sleep, and a serial affect-to-sleep chain carrying 17.56% of the total effect.
The same studies also constrain the design. All of them are cross-sectional surveys analyzed with the SPSS PROCESS serial-mediation model, so the ordering of mediators is assumed rather than observed. All measure sleep by self-report. Some serial effects are tiny: in Shen et al. the chained path carried 2.56% of the indirect effect. We borrow their hypothesized ordering (affect before sleep) but not their inference. A product of four small path coefficients, chained through an MRI-derived measure, may be too small to detect, and the design must plan for that.
The brain-age papers supply the method. Peng et al. provides SFCN, a lightweight 3D convolutional network. It reports a test mean absolute error of 2.14 years on UK Biobank adults, has public weights, and shows that a linear correction for regression toward the mean reduces the gap–age correlation from −0.39 to 0.01. Baecker et al. shows that classical regressors on voxel features reduced by principal components reach roughly 3.8–3.9 years of error. The regularization paper supplies the key manipulation: changing how strongly a ridge model is regularized.
Two further papers bear on the design. Bidirectional Mediation Effects of Sleep and Brain Volume on Cognition warns that sleep and brain structure can mediate each other in both directions in older adults, so direction must be tested rather than assumed. Pala et al. shows that structural-equation mediation into a biological-age measure is feasible, though for blood-derived age with no emotional variables.
Research Program
Data and waves. All analyses use the Adolescent Brain Cognitive Development (ABCD) cohort. Shaul et al. measured adversity at baseline (ages 9–10), MRI at the 2-year follow-up, and caregiver-reported symptoms at year 3. We use their four adversity dimensions: threat, deprivation, household instability, and socioeconomic status. We take negative affect and sleep and screen-use measures from the intervening survey waves; exact variable availability is to be confirmed. We use baseline and 2-year T1 scans so that the brain endpoint is the change in gap, adjusted for baseline gap. This separates "adversity is associated with brain maturity" from "the cascade shifted brain maturity after exposure."
Brain-age instruments. Three instruments are built. The first is SFCN initialized from the public weights at ha-ha-ha-han/UKBiobank_deep_pretrain and adapted to ages 9–14. The second is a ridge regression on imaging-derived regional measures, fitted at a grid of regularization strengths spanning the path studied in the regularization paper. The third is a non-learned reference: total gray matter volume residualized on age, sex, and head size.
The normative reference sample is a held-out set of low-adversity ABCD participants who are excluded from all mediation analyses. It is chosen on exposure only, never on symptoms. Every instrument receives the same linear age-bias correction. No instrument is tuned against affect, sleep, or symptoms. This matters because choosing the brain-age model by its mediation result would be a garden of forking paths.
Estimand. The primary estimand is the upstream chain, the effect of adversity on the brain-age gap change carried through affect and then sleep. In the linear case this is the product of path coefficients
$$\text{IE}_{\text{serial}} = a_1 \, d_{21} \, b_{\Delta}.$$
Here $a_1$ is the effect of adversity on negative affect, $d_{21}$ is the effect of affect on sleep disruption controlling for adversity, and $b_{\Delta}$ is the effect of sleep disruption on the gap change controlling for affect, adversity, and baseline gap.
The downstream link, from gap change to year-3 symptoms, is estimated separately. Multiplying it in would make the product too small to test with any realistic sample. We estimate the paths with interventional mediation effects from a sequential-mediator g-computation, fitted with cross-fitting and bootstrap intervals. Every estimate is reported separately for boys and girls and for their difference. A sensitivity analysis reports how strong an unmeasured confounder of sleep and brain change would have to be to erase the estimate.
Power and splits. Before touching outcomes, we run a simulation to find the smallest serial effect detectable at the achievable sample size. The simulation plugs in coefficients of the size reported in the mediation literature: roughly 0.1–0.2 for adversity to affect and affect to sleep, and a sleep-to-brain coefficient similar to the 0.12–0.16 range in the SG70 older-adult study.
The cohort is split by family into a discovery half and a confirmation half, stratified by site. All model choices are fixed on the discovery half. If the confirmation estimate is null, we report an equivalence bound: the largest serial effect the data can rule out.
Baselines
Naive reference. The do-nothing reference is the total effect of each adversity dimension on the brain-age gap change, adjusting for baseline gap and site, with no mediators. Any mediation claim must explain a nonzero share of an effect that actually exists. If there is no total effect for a dimension, the chain analysis for that dimension is reported only as descriptive.
Strongest existing method. This is a replication of Shaul et al.'s design: a single brain-age mediator from the SFCN-style model, pooled across sex, run on our discovery half with our preprocessing. The file reports no numeric effect size for this analysis; the paper describes only a non-significant mediation and a "relatively small" association between the gap and externalizing symptoms. No reported adolescent mediation number was found on Lacuna, so our replication defines the number to beat.
For brain-age accuracy, the reference points are SFCN's 2.14-year error on UK Biobank adults and the roughly 3.8–3.9 year error of voxel-plus-principal-components regressors in Baecker et al. Neither is from adolescents. We will report each instrument's own error and treat those figures as context, not targets.
What beats it. Beating it means a serial indirect effect whose interval excludes zero in discovery and has the same sign and overlapping interval in confirmation, in a prespecified sex-by-instrument cell. The effect must also remain after adding total gray matter volume change as a competing mediator.
Structural baselines. The first is a parallel-mediation model, in which affect and sleep each lead to the gap independently. It tests whether the ordering adds anything. The second is the gray-matter reference instrument run through the identical chain. It tests whether the learned gap is needed at all.
Ablations. Four ablations isolate the proposal's ingredients. Removing the affect step tests the chain. Pooling across sex tests the cancellation explanation. Sweeping ridge regularization tests the measurement explanation. Reversing the ordering, from baseline gap to later sleep to later affect, tests direction, as the bidirectional SG70 result demands.
All instruments and models share the same participants, covariates, bias correction, bootstrap count, and a matched tuning budget. Each learned instrument gets the same number of hyperparameter settings, chosen only by age error on the reference sample.
Why not just…?
Why not just run parallel mediation, which assumes no ordering? On cross-sectional data this objection is correct, and it is why we do not trust the published serial estimates as proof. The evidence offered for ordering in Xin Zhang et al. and Yang Liu et al. comes from same-time questionnaires. Our design measures affect and sleep at different waves, so ordering becomes testable. The parallel model is run as a baseline, and if it fits the confirmation half as well as the chain, we will report that the ordering is unsupported.
Why not just use total gray matter volume instead of a learned brain age? This is the most serious threat. The regularization paper explains the ridge model's clinical sensitivity by a shift of weight onto global gray matter. If our best cell is a heavily regularized ridge model, it may be a noisier proxy for one volume measure. The counterexample from ECG, where Lima et al. found that the age gap predicted mortality (hazard ratio 1.79) even in clinically normal tracings, comes from another organ and does not settle the question for adolescent brains. We settle it directly by running both as competing mediators. If gray matter carries the effect and the gap adds nothing, the conclusion is that "brain age" is the wrong construct here.
Why not just swap in a ridge brain-age model and rerun Shaul et al.'s single-mediator test? This could rescue the null without any chain, which would make the emotional-cascade half of the proposal unnecessary. The factorial design answers it. The single-mediator ridge model is one cell, and if it matches the chained estimate, we report that measurement choice, not missing intermediaries, explained the null.
Why not just run PROCESS serial mediation on a large one-time survey with an MRI added? That repeats the main weakness named across the mediation literature: imposed direction and shared method variance. It would also inherit the risk, raised by the SG70 work, that sleep and brain structure influence each other. The reversed-ordering test and the baseline-gap adjustment exist to address both.
Decisive Checks
The primary result is the sign and size of the upstream serial effect in each sex-by-instrument cell, estimated in confirmation after discovery-fixed choices.
The measurement explanation predicts a dose–response pattern: the serial effect should rise with ridge regularization strength and exceed the SFCN cell. If instead it tracks age accuracy, the paradox does not transfer to this setting.
The cancellation explanation predicts opposite-signed effects in boys and girls for deprivation and socioeconomic status, reproducing the sign pattern in Shaul et al., with a pooled estimate near zero.
The direction test requires that the forward chain clearly outperform the reversed chain. If the two are comparable, we report bidirectional coupling rather than a cascade.
Finally, for any dimension where every cell is null, we report the equivalence bound from the power simulation. That tells the field how large a brain-age-mediated stress cascade could be at most. It is useful whether or not the dominoes connect.
References
- Dimensional adversity, brain-age, & mental health: Differences in male and female adolescents
- Brain-age models with lower age prediction accuracy have higher sensitivity for disease detection
- Child maltreatment exposure and adolescent nonsuicidal self-injury: the mediating roles of difficulty in emotion regulation and depressive symptoms
- Perceived Academic Stress and Depression: The Mediation Role of Mobile Phone Addiction and Sleep Quality
- Adolescent campus bullying and non-suicide self-injury: chain mediating effect of negative affect and sleep quality
- The impact of childhood emotional maltreatment on adolescent insomnia: a chained mediation model
- Accurate brain age prediction with lightweight deep neural networks
- Brain age prediction: A comparison between machine learning models using region‐ and voxel‐based morphometric data
- Bidirectional Mediation Effects of Sleep and Brain Volume on Cognition in Older Adults
- Modeling the impact of socioeconomic disparity, biological markers and environmental exposures on phenotypic age using mediation analysis and structural equation models
- The chain mediating effect of anxiety and inhibitory control between bullying victimization and internet addiction in adolescents
- Deep neural network-estimated electrocardiographic age as a mortality predictor