Off-Pathway White Matter as a Control for Hidden Confounding in Microstructural Mediation

Brain imaging can now detect damage to white matter, the brain's wiring, years before a radiologist sees anything abnormal. Diffusion MRI measures how water moves through tissue, and summaries such as fractional anisotropy (how strongly water movement follows the fiber direction) or free-water fraction pick up this early damage. Studies then use these measures to argue for a mechanism: visible lesions degrade specific fiber tracts, and that degradation causes cognitive decline. The problem is that the same property that makes these measures useful for early detection also makes them easy to confound. They respond to diffuse, whole-brain processes such as small-vessel disease, systemic inflammation, and biological aging. Those processes also lower cognitive scores directly, and they are almost never measured. A tract can look like it carries the lesion-to-cognition effect simply because it tracks a hidden global vulnerability.

We propose to turn that diffuseness into a tool. If the hidden confounder acts on the whole brain, then white matter tracts that are anatomically unrelated to the cognitive outcome carry information about it without lying on the causal path. Two such off-pathway measurements, with independent noise, can act as the paired proxies required by proximal causal inference, a method that removes bias from an unmeasured confounder using two imperfect stand-ins for it. This lets us estimate how much of a claimed "microstructural mediation" survives once confounding between the mediator and the outcome is accounted for. That link is the one existing sensitivity methods assume away.

Core Claim

Mediation analysis splits an exposure's total effect into an indirect part that runs through an intermediate variable (the mediator) and a direct part that does not. In neuroimaging the mediator is usually a tract-level diffusion measure. The mediator–outcome link cannot be randomized, so it is the weakest point. Our claim is that the spatial organization of the brain supplies usable proxies for the confounder on exactly that link. Take an exposure $X$ such as white matter hyperintensity volume, a mediator $M$ such as anterior thalamic radiation integrity, and an outcome $Y$ such as processing speed. Microstructure in tracts with no plausible route to $Y$ serves as an outcome-side proxy $W$. A second off-pathway measure, from a different tract or diffusion compartment, serves as a mediator-side proxy $Z$.

With these proxies, and with observed covariates $C$ (age, sex, scanner site, head motion), the confounding can be removed by solving for an outcome bridge function $h$:

$$E[Y \mid Z, M, X, C] = E[h(W, M, X, C) \mid Z, M, X, C].$$

Here $h$ is a function of the outcome-side proxy that reproduces the outcome's dependence on the hidden confounder. Once $h$ is found, it stands in for $Y$ inside the usual mediation formula. We do not claim to have invented proximal mediation as a statistical idea. The new parts are three. First, a concrete, anatomically motivated construction of the proxies. Second, testable diagnostics for whether those proxies are valid. Third, an empirical reassessment of published microstructural mediation findings.

Relation to Existing Work

The motivating evidence comes from imaging studies that report mediation without deconfounding. White matter integrity mediates the associations between white matter hyperintensities and cognitive function in patients with silent cerebrovascular diseases used tract-level fractional anisotropy across 20 tracts and linear mediation with the bootstrap (the PROCESS macro). It reported indirect effects of −0.099 through the right anterior thalamic radiation and −0.038 through the left superior longitudinal fasciculus on processing speed. It assumed no unmeasured confounding on any path. Cognitive implications of white matter microstructural changes in individuals with low heart rate variability: a NODDI study found widespread, not tract-specific, free-water elevation in people with low heart rate variability. Its evidence for mediation was correlational. That widespread pattern is exactly the signature of a diffuse confounder, and it motivates our proxy construction.

The methods that could address hidden confounding each stop short of this setting. Sharp Bounds for Generalized Causal Sensitivity Analysis derives tight bounds under a bounded amount of hidden confounding and includes an algorithm for mediation paths. It explicitly assumes no unmeasured mediator–outcome confounding, and it produces bounds rather than a corrected estimate. On Partial Identification of the Natural Indirect Effect bounds indirect effects for three-level mediators. It shows that relaxing its cross-world assumption triples the interval width, pushing it across zero. That result constrains us: assumption-light bounds are too wide to be informative, which argues for bringing in extra information through proxies. On semiparametric estimation of a path-specific effect in the presence of mediator-outcome confounding gives estimators that stay consistent if any one of several nuisance models is correct. It requires the mediator–outcome confounder to be observed. We borrow its efficient-influence-function construction for the estimation stage.

On the confounding side, Demystifying Proximal Causal Inference supplies the bridge-function method, but for total effects. Deep Multi-Modal Structural Equations For Causal Effect Estimation With Unstructured Proxies learns latent confounders from images, also for total effects only. A Negative Control Outcome Regression Accounting for Unobserved Confounding and Lagged Causal Effects is noted in our research file as illustrating how badly ignored confounding biases estimates. Counterfactual Graphical Models for Longitudinal Mediation Analysis With Unobserved Confounding gives the graphical rule (the "recanting district criterion") for when a path-specific effect is identifiable. We use it to check identification in our graph, which adds proxy nodes the paper does not consider. Front-door methods such as Finding and Listing Front-door Adjustment Sets require an unconfounded mediator–outcome link, so they fail precisely where we work.

Research Program

The first component is identification. We write the causal graph with a latent diffuse vulnerability $U$ that affects $X$, $M$, $Y$, $W$ and $Z$. We then state the conditions under which the natural indirect effect is identified from the bridge equation. The two key conditions are that $W$ has no effect on $Y$ except through $U$, and that $W$ is independent of $M$ and $Z$ given $U$, $X$ and $C$. We also need a completeness condition, meaning the proxies must vary enough with $U$. We check which path-specific effects remain identified using the recanting district criterion. We are explicit that the condition "off-pathway tracts do not affect $Y$" is an anatomical assumption, not a theorem. Motion and site induce correlated measurement noise across all tracts, so they enter $C$. Otherwise they would violate the independence of $W$ and $Z$.

The second component is estimation. We implement two versions of the bridge. The linear one is essentially two-stage regression with $Z$ as the instrument for $W$. The nonparametric one uses kernel regression with a min-max objective. Both feed into a multiply robust indirect-effect estimator adapted from the semiparametric path-specific estimator above, with cross-fitting and bootstrap confidence intervals.

The third component is proxy diagnostics. When several candidate off-pathway tracts exist, different $(Z, W)$ pairs should give the same indirect effect if all are valid. Their disagreement is a built-in test that some are invalid, much like an overidentification test for instruments. We also check a negative-control mediator: an off-pathway tract used in the mediator's role should show an indirect effect near zero after correction.

Baselines

The naive reference point is linear mediation with the bootstrap in the style of the PROCESS analysis above. We adjust it for the same observed covariates $C$ and run it on identical subjects and tract definitions. The only reported numbers from this literature are the indirect effects above, −0.099 through the anterior thalamic radiation and −0.045 through the right inferior fronto-occipital fasciculus for executive function. We will reproduce this style of analysis as the uncorrected estimate. Beating it means showing lower bias and nominal confidence-interval coverage on the benchmarks where the truth is known.

The strongest existing method for hidden confounding in mediation is the sharp-bounds approach. We run it with its released algorithm on the same data and report the smallest confounding strength $\Gamma$ at which the indirect effect's bounds include zero. It assumes an unconfounded mediator–outcome link, so we apply it in two ways: as published, and to the exposure–outcome part only. No reported result was found on Lacuna for this method on imaging mediators. The only figure in our file is its runtime of 0.4 seconds versus 137 seconds for grid search, which does not bear on accuracy. We count our method as better if it yields a corrected point estimate whose confidence interval is narrower than the bounds at the confounding strength implied by the held-out confounder (see below), while keeping coverage.

The key ablation is naive proxy adjustment. Here we add the off-pathway tract measures, or global mean fractional anisotropy, as ordinary regression covariates. This uses exactly the same information as our method but not the bridge function. It isolates whether the proximal machinery matters or whether the anatomical proxy idea alone does the work. A second ablation adjusts for T1 brain-age gap. We compute it from the pretrained network in Accurate brain age prediction with lightweight deep neural networks, which reports a test mean absolute error of 2.14 years on UK Biobank, using the UKBiobank_deep_pretrain code, as a learned summary of global vulnerability. All methods share the same cross-fitting folds, covariates, and a fixed hyperparameter budget chosen on held-out folds.

Why not just…?

Why not just adjust for global white matter integrity or brain age? This is the most serious competitor. If off-pathway microstructure measured the confounder without noise, direct adjustment would be enough. It does not: diffusion measures are noisy proxies. Adjusting for a noisy proxy of a confounder removes only part of the bias, and the leftover shrinks only as proxy noise vanishes. The bridge function corrects for that noise by using the second proxy, the same logic by which instruments correct for errors in regressors. The ablation tests this directly. If naive adjustment matches the bridge estimator in bias and coverage on the planted-confounder benchmark, we will report that the anatomical proxy idea is what matters and the proximal machinery is unnecessary.

Why not just run a sensitivity analysis and report robustness? Sensitivity bounds answer "how much confounding would overturn this?" but not "what is the corrected effect?" The sharp-bounds method also assumes away mediator–outcome confounding. The indirect-effect bounds of On Partial Identification of the Natural Indirect Effect tripled in width and crossed zero once assumptions were relaxed. Bounds alone are likely to be uninformative for effects as small as those reported for tracts. We still report bounds alongside our estimates, so readers can see both.

Why not just use longitudinal data with baseline mediator measurements? Temporal ordering fixes reverse causation. It does not fix a stable diffuse vulnerability that drives both baseline tract integrity and later decline. UK Biobank's repeat imaging lets us combine the two approaches, and we will report the longitudinal estimate as an additional comparison rather than a replacement.

Why not just use a deep multimodal classifier? Models such as HA-HI (0.825 accuracy on ADNI) and the ensemble in A multimodal MRI-based machine learning framework for classifying cognitive impairment in cerebral small vessel disease (validation AUC 0.878) answer a prediction question. They do not say which tract's damage, if reduced, would change cognition. Deep Multi-Modal Structural Equations shows that discriminative models fitted to rich image data give poor causal estimates: an effect error of 4.067 for inverse-probability weighting versus 0.096 for their generative approach.

Proposed Experiments

The first experiment is a semi-synthetic test on real UK Biobank imaging. We keep real tract diffusion measures and real white matter hyperintensity volumes. We plant a latent vulnerability by adding a smooth, whole-brain perturbation to all tracts, then generate cognition with a known indirect effect and a known confounding strength. We report bias, root-mean-square error, and 95% interval coverage of the indirect effect for all baselines. We sweep confounding strength and proxy noise, which should show naive adjustment degrading as noise grows while the bridge estimator holds.

The second experiment, and the decisive real-data check, is a held-out confounder test. We choose a measured variable with a strong diffuse effect on white matter and cognition, such as age or a vascular risk score, and hide it from every estimator. The reference answer is the mediation estimate obtained when that variable is included. A method succeeds if, without the variable, it recovers the reference estimate within its confidence interval. Naive mediation should fail this test visibly. Cross-proxy agreement across several off-pathway tract pairs is the second decisive check. If different valid-looking pairs give inconsistent corrected effects, the proxy assumption is failing, and we will say so rather than report a single number.

The third experiment reassesses published findings. We reproduce the lesion-volume, tract, and processing-speed analysis of the silent cerebrovascular disease study in UK Biobank, and a heart-rate-variability, free-water, and cognition analysis where the variables are available. In each case we report the naive indirect effect, the proximally corrected effect, and sharp bounds. We then replicate on ADNI's diffusion subset, which comes from a different population and scanners, to check portability across sites. The substantive result is the fraction of each reported microstructural mediation that survives correction. That tells the early-detection field which sensitive biomarkers are candidate mechanisms and which mainly track global vulnerability.