When the Mediator Goes Unmeasured: Path-Specific Effects Under Selective Blood-Gas Testing

Pulse oximeters, the clip-on sensors that estimate blood oxygen from light passing through the finger, read falsely high in many patients with darker skin. A recent analysis of two large intensive-care databases asked how much of the racial gap in ventilator care flows through this device error alone. The error is the difference between the oximeter reading and the true arterial oxygen level, and the true level is known only when a clinician orders an arterial blood gas test. That test is not ordered at random. Clinicians order it when they suspect the oximeter is wrong or the patient is getting worse. A patient whose oximeter reads falsely reassuring is, almost by definition, less likely to get the test that would expose the error. The quantity being studied therefore shapes whether it is ever observed.

This note proposes to put the measurement decision inside the causal model instead of filtering on it. We would identify and estimate the device-error pathway effect with a model of which patients get blood gases. The model sits in a doubly robust estimator, meaning one that stays consistent if either of two sub-models is right. We would add a sensitivity analysis for the case where ordering depends on things never recorded. We would also build a masking benchmark inside a sub-cohort that is measured almost continuously, which tells us whether any of this changes the answer. The core claim is testable. If ordering is driven by the displayed oximeter value, restricting to tested patients can bias the estimated device-error effect, which we call the VDE (the V-specific direct effect). Correcting for it could change the published conclusions on ventilation initiation and weaning.

What exists and what is missing

Some terminology first. A path-specific effect is the part of a cause's influence that travels along one chosen route through a causal diagram, while every other route is held at a reference value.

Path-specific effects for pulse-oximetry guided decisions in critical care defines the VDE as the effect of race on ICU decisions that passes through oximeter discrepancy $V = SpO_2 - SaO_2$. Here $SpO_2$ is the oximeter reading and $SaO_2$ the arterial truth. The paper estimates the VDE with a self-normalized doubly robust (SN-DR) estimator, which rescales its weights to tame variance. It reports a VDE on ventilation initiation of −0.21 percentage points in eICU and −0.46 in MIMIC-IV. It also reports ventilation lasting 2.3 hours longer in eICU and 7.4 hours longer in MIMIC-IV for Black versus White patients through this path.

This paper supplies our estimator and our target quantity. The file itself flags its gap: patients without blood gases could not have $V$ computed, which restricts who enters the sample. We keep the paper's static, whole-stay summaries on purpose. That isolates the missingness question from the separate open problem of time-varying graphs, and it means any change in estimates can be attributed to the measurement model alone.

Two lines of work shape the method. On semiparametric estimation of a path-specific effect in the presence of mediator-outcome confounding. derives the efficient influence function for a path-specific effect under exposure-induced confounding. An influence function is the first-order correction term that turns a plug-in estimate into one robust to errors in the sub-models. That paper builds a multiply robust estimator, and it shows parametric maximum likelihood failing badly under misspecification. We reuse its construction pattern: add one more nuisance model, here the measurement propensity, and the correction term that goes with it. Its negative result about maximum likelihood rules out a purely parametric imputation-based design.

The identification theory in Counterfactual Graphical Models for Longitudinal Mediation Analysis With Unobserved Confounding and Identifiability of Path-Specific Effects tells us which pathway effects are expressible from observed data. We extend those graphs with a measurement-indicator node.

Two papers motivate the problem without solving it. Risk factors for steroid-refractory in immune checkpoint inhibitor-induced colitis: a retrospective cohort study documents that elective diagnostics are ordered for sicker patients in retrospective cohorts. Step-by-step causal analysis of EHRs to ground decision-making warns that design choices around eligibility and time-zero can manufacture bias in MIMIC-IV. Restricting to blood-gas-tested patients is exactly such a design choice, but one that paper never analyzes, because it only estimates total effects.

Core Claim

Let $A$ be race, $Z$ pre-admission covariates, $W$ other post-admission physiology, $V$ the discrepancy, $Y$ the ventilation outcome, and $R \in \{0,1\}$ the indicator that a blood gas was drawn in the relevant window. The clinically plausible graph has an edge from the displayed reading $SpO_2$ to $R$. Since $SpO_2 = SaO_2 + V$, $R$ is a descendant of $V$ and of $A$. A complete-case analysis conditions on $R = 1$. It therefore estimates the pathway effect in a subpopulation whose composition itself depends on the pathway.

We assume measurement is at random given the observed history $H = (A, Z, W, SpO_2)$, that is, $R \perp SaO_2 \mid H$. Under that assumption, and under Zhang et al.'s positivity and no-unmeasured-confounding assumptions, the VDE becomes identified by re-weighting each tested patient by $1/\pi(H)$, where $\pi(H) = P(R = 1 \mid H)$. We also give an augmented form with an outcome-side model of $SaO_2$ given $H$. In the augmented form the measurement layer is consistent if either $\pi$ or the $SaO_2$ model is right, and the SN-DR layer keeps its own double robustness. Inside the measurement weights we self-normalize, as Zhang et al. do. We expect this to matter, because $\pi$ is likely small exactly for patients with falsely high readings.

Research Program

The first component is identification. We would add $R$ and its parents to the counterfactual graph and state the conditions under which the VDE is a functional of observed data. We would also say when it is not: for example, if race affects ordering through unrecorded clinician perception rather than through $SpO_2$ and $W$. For the non-identified case we would add a sensitivity parameter, an odds ratio linking $R$ to unobserved $SaO_2$ given $H$, and report the VDE as a curve over that parameter. The curve lets a reader see how strong selective ordering would have to be to flip the conclusions.

The second component is estimation. We would build the augmented influence-function estimator, fitting the nuisance models with cross-fitting so that flexible learners do not overfit their own correction term. Cross-fitting means each nuisance model is trained on one data split and evaluated on another.

The third component is the masking benchmark, which is the empirical heart of the proposal. In MIMIC-IV and eICU, patients with arterial lines, mostly those already ventilated, have frequent blood gases, so $V$ is nearly always observed for them. Within that sub-cohort we compute a reference VDE on the fully observed data. We then learn realistic ordering policies $\pi(H)$ from non-arterial-line patients and apply them to hide $SaO_2$ in the fully measured cohort. This produces known-truth missingness that mimics real clinical ordering, including dependence on displayed $SpO_2$. We would also impose deliberately non-ignorable policies, where masking depends on hidden $SaO_2$, to test the sensitivity analysis. The arterial-line cohort is sicker than the general ICU population, and that is a real limit on external validity. We state it rather than hide it.

Baselines

The naive reference point is the unadjusted difference in outcomes between Black and White patients, alongside a total-effect estimate from AIPW. AIPW is augmented inverse-probability weighting, a standard doubly robust estimator, run as in Step-by-step causal analysis of EHRs to ground decision-making. It shows how much of the gap exists before any pathway reasoning.

The strongest existing method is complete-case SN-DR from Path-specific effects for pulse-oximetry guided decisions in critical care, re-implemented on the same cohort definitions, covariate sets, and nuisance learners as our estimator. It will also get the same cross-fitting folds and hyperparameter search budget. On semi-synthetic data that paper reports an error of 0.003 for SN-DR against 0.024 for the X-learner, a standard two-stage method for estimating how treatment effects vary across patients. We therefore also run the X-learner, as an off-the-shelf reference that should trail both.

The published real-data targets we compare against are −0.21 and −0.46 percentage points for initiation and 2.3 and 7.4 hours for duration. In the masking benchmark, beating complete-case SN-DR means lower absolute bias against the full-data reference VDE together with 95% confidence-interval coverage near nominal. We require both, since a variance reduction alone would not count. Both are measured under the realistic, SpO₂-driven ordering policies.

Two ablations isolate the key ingredient. The first uses measurement weighting alone, with no $SaO_2$ imputation model. The second uses imputation alone, meaning multiple imputation of $SaO_2$ from $H$ fed into SN-DR. The full estimator should match the better of the two when that component is correctly specified. It should clearly beat each when that component is deliberately misspecified. If it does neither, the augmented construction is not earning its complexity.

Why not just…?

Why not just use complete cases, as the original analysis did? Complete cases are unbiased only if blood-gas ordering is unrelated to the pathway being measured. That is implausible when the displayed oximeter value drives ordering and itself contains the error. We do not assume the bias is large. The masking benchmark measures it directly. If complete-case SN-DR stays within its own confidence interval of the full-data reference under realistic ordering policies, we will report that the published estimates are robust, and that is a useful result in itself.

Why not just multiply impute $SaO_2$ and run the existing estimator? Under missing-at-random, correctly specified imputation is consistent, so this is a genuine threat. It is exactly our imputation-only ablation. Its weakness is dependence on getting the $SpO_2$-to-$SaO_2$ relationship right. That relationship is nonlinear and race-dependent, which is the whole phenomenon under study, and the Miles et al. results show how single-model approaches break under misspecification. If imputation-only matches the full estimator under both correct and misspecified conditions, we will recommend it as the simpler tool.

Why not just analyze the arterial-line patients, where measurement is near-complete? That population is not the question. The weaning result concerns ventilated patients, but the initiation decision concerns patients who are not yet ventilated and often lack arterial lines. Those are precisely the patients whose falsely reassuring readings would discourage testing. We use the arterial-line cohort as a measuring instrument for estimator bias, not as the answer.

Proposed Experiments

The decisive check is the masking benchmark in MIMIC-IV and eICU. For each realistic and adversarial ordering policy, we report bias, root-mean-squared error, and interval coverage for complete-case SN-DR, the X-learner, both ablations, and the full estimator. A second check re-estimates the four published quantities on the full eICU and MIMIC-IV cohorts. That covers initiation and duration in each database, now including untested patients through the measurement model, alongside sensitivity curves over the non-ignorability parameter. The proposal succeeds if the corrected estimates either move beyond the published confidence intervals, or provably do not move under any plausible sensitivity setting. In both cases the conclusion about device error rests on the whole ICU population rather than on the patients whose error someone chose to check.