When patients do not take the drug they were assigned, a trial compares assignment rather than treatment, and small drug effects disappear into noise. A recent line of causal work shows that recording what patients actually took can sharpen the estimate. The method conditions on actual intake as a stepping stone between assignment and outcome, an approach called front-door adjustment, and it gives lower-variance estimates of per-patient effects than adjusting for baseline confounders alone. That work assumes intake is observed perfectly. In practice the most objective record of intake is a blood test for the drug, and every blood test has a detection floor, a quantification floor, run-to-run scatter and incomplete recovery. Analytical chemists measure exactly these properties when they validate a method, and then the numbers go unused in the downstream statistics.
This note proposes to join the two. We would turn an assay's validation report (its limit of quantification, precision and recovery) into an explicit error model for intake, derive and test when front-door estimation still beats ignoring intake, and report the answer as a design rule: the assay quality needed before measuring adherence is worth doing. The application regime is Alzheimer's disease, where symptomatic drug effects on cognitive decline are small, which is exactly where the front-door advantage should be largest. It is also where adherence is fragile, because the patients are cognitively impaired. Evaluation uses semi-synthetic treatment experiments built on real longitudinal trajectories from the Alzheimer's Disease Neuroimaging Initiative (ADNI).
Core Claim
The variance advantage of conditional front-door adjustment is not guaranteed once intake is measured by a real assay. It holds only when the assay's error is small compared with the spread in drug exposure between adherent and non-adherent patients. The specific failure mode is censoring below the quantification limit, which merges partial adherers with full non-adherers. Our claim is that this crossover can be predicted from four numbers every validation report already contains: the limit of detection, the limit of quantification, precision as a relative standard deviation, and recovery percentage. We further claim that an estimator which treats intake as a hidden continuous exposure, observed through that validated error model, recovers most of the lost advantage.
The closest work is Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence. That paper proves front-door adjustment has lower asymptotic variance than standard backdoor adjustment when the assignment effect is small. It introduces LobsterNet, a multi-task network with heads for assignment, intake and outcome. It treats intake as a known and usually binary variable. We keep its estimator and its theory as the method we build on. What we change is the observation model for intake, and the question we ask, which is when their guarantee survives measurement.
Why these three directions belong together
The motivating evidence comes from an unlikely place. Development and validation of a novel HILIC method for the quantification of low-levels of cuprizone in cuprizone-containing chow validated a chromatography method for a neurotoxin in rodent feed (limit of quantification 2.5 µg/mL). The authors found that a demyelination model had "failed" because one batch held 0.08% compound against a nominal 0.3%. The drug had silently gone undelivered. This is non-adherence at the level of the formulation, and it was invisible until a validated assay measured it. Human adherence is the same problem with a patient in place of a pelleting press.
The assays we would plug in are of the kind validated in Rapid quantification and validation of etomidate and dexmedetomidine blood concentrations in rats using a novel portable mass spectrometer. That work reports quantification limits as low as 5–6 ng/mL, accuracy of 86.3%–104.4%, and agreement with laboratory HPLC-MS of $r^2$ 0.972–0.986. It also names its own limits: acute dosing only, rat plasma, and two analytes. We use its numbers, together with the precision figures in Validation of Dissolution Method of Levofloxacin 750 mg Tablet by High Performance Liquid Chromatography (method precision RSD 1.68%), to define realistic tiers of assay quality. We make no claim that these are validated for Alzheimer's drugs.
The disease side supplies the outcome model and the heterogeneity. Feature selective temporal prediction of Alzheimer's disease progression using hippocampus surface morphometry gives a strong predictor of cognitive score decline in ADNI-1. Multimodal subtypes identified in Alzheimer's Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference gives five progression subtypes with different rates of decline, which we use as effect modifiers. Neither paper models treatment or adherence. Modeling treatment events in disease progression constrains the design. It shows that smoothing trajectories across treatment events badly misattributes effects, so the outcome model must represent treatment explicitly rather than absorb it into a smooth curve.
Research Program
The estimand is the conditional effect of assignment $T$ on the 24-month change in the Alzheimer's Disease Assessment Scale cognitive score $Y$, given baseline covariates $x$ (imaging, cerebrospinal fluid, demographics). Front-door adjustment writes the outcome under assignment $t$ as
$$E[Y \mid do(T=t), x] = \sum_a P(a \mid t, x) \sum_{t'} E[Y \mid t', a, x]\, P(t' \mid x),$$
where $a$ is actual intake. We replace the observed $a$ with a hidden cumulative exposure $C$, which is the patient's average plasma drug level over the interval. Intake is then seen only through the assay reading $\tilde{C}$:
$$\tilde{C} = R \cdot C \cdot e^{\varepsilon}, \quad \varepsilon \sim \mathcal{N}(0, s^2), \quad \text{reported as "below LOQ" when } R\,C\,e^{\varepsilon} < L.$$
Here $R$ is fractional recovery, $s$ is precision (relative standard deviation), and $L$ is the limit of quantification. All three are taken from the validation report rather than estimated from the trial.
The work has three parts. The first is analytical. We extend the variance comparison of Chen et al. to the case where the mediator is observed through this censored multiplicative channel. We then derive an approximate crossover condition, which states that front-door adjustment beats backdoor adjustment only if the information about $C$ lost to censoring and scatter is below a threshold. That threshold depends on the size of the assignment effect and on the distribution of exposure under the adherence patterns being modeled.
The second part is an estimator. It is LobsterNet's shared-backbone architecture, with its intake head replaced by a censored-likelihood model of $\tilde{C}$ given $(T, x)$ that has validation parameters fixed. The front-door sum then becomes an integral over the posterior of $C$.
The third part is a design tool. Given an assay's validation parameters and a planned trial's expected effect size, it predicts the variance ratio between the two estimators. That answers whether collecting blood samples is worth it and how much precision to demand.
Baselines
The naive reference point is intention-to-treat estimation with standard backdoor adjustment, meaning a T-learner that ignores intake entirely. It costs nothing and is immune to assay error, so any proposal must beat it where it matters. The strongest existing method is LobsterNet from Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence. That paper reports a 20% reduction in PEHE, the error in estimating per-patient effects, relative to backdoor adjustment on the AMR-UTI dataset, with intake observed exactly. No reported result was found on Lacuna for front-door or backdoor estimation on ADNI outcomes under non-adherence. We would run LobsterNet two ways: as an oracle given true intake, which is an upper bound, and in its realistic form given intake dichotomized at the quantification limit. The second is the fair head-to-head.
A generic measurement-error correction, regression calibration applied to $\tilde{C}$ with error variance set from the reported precision, isolates whether our censored likelihood adds anything beyond "use the right variance." The key ablation removes censoring from our error model while keeping scatter and recovery, which tests whether the quantification floor is the dominant harm.
All methods share the same backbone architecture and the same covariates. The outcome component uses the hippocampal surface features of the cFSGL study, whose 48-month rMSE of 3.164 against 4.367 for volumes confirms they are informative. All methods share the same splits and the same hyperparameter budget (equal numbers of random-search trials). Beating LobsterNet means lower PEHE at matched assay tiers, with confidence intervals over 50 simulation seeds excluding zero. It also means that the predicted crossover point falls within the empirical one's interval.
Why not just…?
Why not just use intention-to-treat and skip intake measurement? This is a real threat, and part of what we test. If assay error is large, backdoor adjustment may genuinely be better. The Chen et al. theory says front-door adjustment wins only for small effects, and that theory assumes a perfect mediator. Our crossover experiments are designed to locate exactly where intention-to-treat should be preferred. That is a useful answer on its own rather than a failure of the proposal.
Why not just plug the assay reading into LobsterNet, thresholding at the quantification limit? This is what a practitioner would do, and it is baseline two. We expect it to fail when drug holidays and partial adherence push true exposure near the limit. In that range, dichotomizing merges patients with different mediator values, and the front-door sum is then computed over the wrong intake groups. The no-censoring ablation will show whether this intuition holds.
Why not just apply regression calibration or another standard measurement-error correction? This could plausibly match our estimator when most readings sit well above the quantification limit, where error is roughly classical. For that reason we do not stake the contribution on the estimator alone. The design rule, which maps validation parameters to a predicted variance ratio, is useful regardless of which corrector is used. The experiments report whether the censored likelihood beats regression calibration specifically in the high-censoring regime.
Why not just use self-report or pill counts? These are cheaper, but their error is not characterized by any validation report and is plausibly biased toward reporting adherence. We simulate this as a one-sided misclassification channel at rates spanning the plausible range. We then compare it against the assay tiers on the same variance-ratio axis. No reported misreporting rates were found on Lacuna, so these rates are stated as assumptions.
Proposed Experiments
The main benchmark is semi-synthetic. We take the ADNI-1 MCI and Alzheimer's participants used in the cFSGL study, with their real baseline features and real untreated cognitive trajectories. On top of these we inject a randomized assignment with a small effect of 0.5–2 points of cognitive score. The effect is modified by SuStaIn subtype, so the "Subcortical" and "Cognitive" subtypes differ.
Adherence is simulated with a mix of full adherers, partial adherers and drug-holiday patterns, with adherence probability decreasing with baseline cognition. Plasma exposure comes from a one-compartment pharmacokinetic model. The exposure is then passed through three assay tiers: a laboratory tier (RSD about 2%, recovery about 100%), a portable tier (accuracy 86%–104% and a higher quantification floor, following the portable-MS paper), and a poor tier.
The decisive plot sweeps the quantification floor relative to median trough exposure, and sweeps precision, and shows the PEHE ratio of each front-door variant to backdoor adjustment. The proposal succeeds if three things hold. The analytic crossover must predict the empirical one. The validated-error estimator must keep a meaningful fraction of oracle LobsterNet's advantage in the portable tier, where thresholded LobsterNet should have lost it. And the ordering must hold across the five subtypes.
A second check reuses the synthetic IHDP benchmark from Chen et al., corrupting their intake variable with the same assay channel. This ensures the effect is not an artifact of our ADNI construction. It also reproduces their clean-intake result as a calibration point before any corruption is added.
References
- Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence
- Development and validation of a novel HILIC method for the quantification of low-levels of cuprizone in cuprizone-containing chow
- Rapid quantification and validation of etomidate and dexmedetomidine blood concentrations in rats using a novel portable mass spectrometer
- Validation of Dissolution Method of Levofloxacin 750 mg Tablet by High Performance Liquid Chromatography
- Feature selective temporal prediction of Alzheimer's disease progression using hippocampus surface morphometry
- Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference
- Modeling treatment events in disease progression