When a study of historical patient records reports that Black patients received a treatment less often than White patients "after adjusting for clinical factors," the size of that gap, and even whether one exists, depends on a choice the study rarely defends. Some variables are counted as legitimate medical reasons for treating patients differently and are adjusted away. Others are counted as unfair circumstances whose effect on treatment should count toward the disparity. Comorbidity burden, tumor stage at diagnosis, insurance status and length of stay all sit in a gray zone. Each is a real clinical signal, and each is also partly produced by unequal access to care. We propose to make this choice the object of a formal sensitivity analysis. Instead of reporting one adjusted disparity, a retrospective cohort audit would report how the disparity changes across every defensible way of sorting the contested variables into "legitimate indication" and "unfair circumstance." It would also report the smallest normative concession that would make the disparity disappear.
The analogy is to sensitivity analysis for unmeasured confounding. Epidemiologists already ask how strong a hidden confounder would have to be to erase an effect. We ask a parallel question: how many contested variables would have to be reclassified as legitimate to erase a disparity, and which ones. A finding that survives most reasonable classifications is robust. A finding that flips when one variable is reclassified is, in the language of the fairness-ethics literature, fragile. Either way, readers learn exactly where the disagreement lies.
Core Claim
Fairness theory has argued that equality claims are empty until someone specifies which differences are legitimate grounds for different treatment. Algorithmic Unfairness Through the Lens of EU Non-Discrimination Law makes this argument from law. A Moral Framework for Understanding Fair ML through Economic Models of Equality of Opportunity formalizes it by splitting variables into "circumstances," which a person cannot be held responsible for, and "effort," which they can, and shows that standard parity criteria quietly commit to particular splits. Where Is the Normative Proof? and Are There Exceptions to Goodhart's law? add that a fairness claim which collapses under small changes to these assumptions is not trustworthy. These papers motivate the proposal, but none supplies a procedure an analyst can run on a real cohort.
Retrospective cohort methods make the same choice implicitly whenever they decide what to condition on. In the propensity-score analysis of high-grade prostate cancer by Tewari et al., comorbidity, stage, and income all entered the adjustment set. The same study separately documented that Black men were less likely to receive radical prostatectomy (20.0% vs. 31.0%). The adjustment that makes a treatment-effect comparison valid is not automatically the right one for a fairness comparison.
Our claim is that the partition of covariates into legitimate and circumstantial sets should be treated as an uncertain input, varied over a constrained space, and summarized. The goal is an honest map of which normative commitments a disparity verdict rests on, not a single number.
Relation to the Closest Work
The nearest work is Causal Fairness Assessment of Treatment Allocation with Electronic Health Records. It audits coronary bypass surgery allocation in the Columbia/NewYork-Presbyterian record database by comparing patients within "principal strata," meaning groups predicted to benefit equally. It reports that women were 4.4%–7.7% and Black patients 5.4%–8.7% less likely to receive surgery within equal-benefit strata, both with and without social-determinant variables.
That paper performs one binary sensitivity check (social determinants in or out). It fixes one normative baseline: equal benefit justifies equal treatment, which is a Rawlsian equal-opportunity commitment. It does not justify that baseline. We differ in three ways:
- We vary the full space of admissible partitions rather than one switch.
- We report robustness summaries and minimal concessions rather than a point estimate.
- We cross the partition axis with an allocation-principle axis, comparing equal-benefit allocation against prioritarian allocation, which favors the worst-off (Distributive Justice as the Foundational Premise of Fair ML).
Other intersection work leaves this question open. Fairness without Demographics on Electronic Health Records clusters training gradients to find disadvantaged groups without labels, which hard-codes its own unexamined baseline. MIMIC-IF found shorter mechanical-ventilation durations for Black and Hispanic ICU patients in MIMIC-IV. It explicitly could not tell whether this reflects clinical indication or discrimination. Our method is designed to answer that kind of question conditionally: under which classifications of the contested variables does the gap remain?
Research Program
Setup. Let $G$ be group membership, $T$ the treatment decision, and $V$ the set of contested covariates. A partition assigns a subset $L \subseteq V$ to "legitimate indication," together with an always-legitimate core $A$ such as acute physiology. The partition-specific disparity is
$$ \Delta(L) = \mathbb{E}_{X_{A \cup L}}\big[\, P(T=1 \mid G=1, X_{A\cup L}) - P(T=1 \mid G=0, X_{A\cup L}) \,\big], $$
where the outer expectation is over the reference group's distribution of the legitimate covariates $X_{A \cup L}$. Covariates left out of $L$ are not conditioned on, so their group differences count toward the disparity. This is the standard standardization logic of cohort studies, applied once per partition. We estimate each $\Delta(L)$ with a doubly robust, cross-fitted estimator. This combines an outcome model and a group-propensity model, so the estimate remains valid if either model is correct.
Admissible partitions. Admissibility constraints come from two sources: clinical reasoning (for example, oxygenation is always legitimate for a ventilation decision) and causal ordering (a variable cannot be legitimate if it is solely downstream of a variable already declared circumstantial). With 6–10 contested variables grouped clinically, the lattice has a few hundred to a thousand cells. That is tractable with shared nuisance models.
Summaries. We report three quantities:
- Robustness fraction: the share of admissible partitions whose confidence interval excludes zero.
- Minimal concession: the smallest-cardinality set $L$ that renders $\Delta(L)$ indistinguishable from zero.
- Per-variable attribution: a Shapley-style average change in $\Delta$ when each variable moves into $L$, which tells readers which single disagreement matters most.
Allocation-principle axis. We repeat the analysis under a prioritarian target. Here the reference allocation ranks patients by baseline risk of the bad outcome rather than by predicted benefit, and the disparity is measured relative to that reference.
Unmeasured confounding. We cross the normative lattice with a standard bound on hidden bias. This produces a two-dimensional frontier: how much hidden confounding, combined with how many normative concessions, is needed to explain a gap away. This addresses the ignorability weakness that Zhang et al. acknowledge.
Baselines
Naive unadjusted disparity. The raw difference in treatment rates between groups. It is the do-nothing reference point and corresponds to the partition where every contested variable is a circumstance. Any method that cannot tell readers more than this is pointless.
Full adjustment. The standard retrospective practice, as in Tewari et al., adjusts for every measured covariate by propensity or standardization. It is the strongest conventional competitor because it is what clinical journals currently accept. It is also the other endpoint of our lattice. We run it with the same estimator, the same cross-fitting folds, and the same nuisance models. The only difference is that it uses a single partition.
Principal-fairness audit. The method of Causal Fairness Assessment of Treatment Allocation with Electronic Health Records, with and without social determinants. It is the strongest existing method for exactly this question. Its reported result is the 4.4%–7.7% gap for women and the 5.4%–8.7% gap for Black patients in bypass allocation. That database is not public, so we reimplement the method on MIMIC-IV with matched tuning budgets. For the benefit-prediction component we use an IMV-LSTM-style model (MIMIC-IF reports 0.955 AUROC for in-hospital mortality).
No reported result for any normative-baseline sensitivity method was found on Lacuna, so there is no prior number to beat on that axis.
Endpoints-only ablation. Reporting only the naive and fully adjusted estimates plus the social-determinant toggle. This isolates the key ingredient, the interior of the lattice.
Unconstrained-lattice ablation. The same lattice without admissibility constraints. This tests whether the clinical and causal constraints change conclusions or only reduce compute.
What counts as beating the baselines. In semi-synthetic data, the lattice must correctly identify the variable driving a verdict flip in cases where the endpoints give an ambiguous or misleading picture. In real data, the interior of the lattice must reveal flips or variable-specific dependencies that the endpoints and the social-determinant toggle do not.
Why not just…?
Why not just report minimally adjusted and fully adjusted models, as epidemiologists already do? This is the real threat. If disparities are typically monotone in the number of adjusted variables, and a single variable drives most of the change, the two endpoints already tell the story. The endpoints-only ablation tests this directly. We will measure on MIMIC-IV, and in semi-synthetic settings, how often the interior of the lattice contains verdict flips that the endpoints miss, or that are attributable to a specific variable the endpoints cannot name. If this almost never happens, the honest contribution shrinks to the minimal-concession and attribution summaries, and we would say so.
Why not just fit a causal graph and compute path-specific effects, in the spirit of Counterfactual Fairness? A path-specific analysis requires deciding which paths are unfair. That is exactly the normative choice in question, made once and hidden inside the graph. We reuse that machinery where a graph is available, since each lattice cell is a path-specific contrast. What we add is the enumeration and the robustness summary. This responds to the pluralism concern in Where Is the Normative Proof?: we do not aggregate judgments into one answer.
Why not just run the principal-fairness audit, which already conditions on benefit? Zhang et al.'s social-determinant toggle left disparities intact, which shows the result is not fragile along that one dimension. It says nothing about stage, comorbidity, or length of stay, which are classified as legitimate by default. It also cannot express a prioritarian alternative. Running their method as one cell of our lattice shows directly whether its verdict is typical or an edge case.
Proposed Experiments
Experiment 1: semi-synthetic ground truth. We build data from MIMIC-IV covariates, simulating treatment so that a known contested variable, such as an insurance-linked delay that worsens admission severity, carries structural disadvantage. The lattice should recover that variable as the minimal concession and as the top attribution, with calibrated coverage of $\Delta(L)$. We compare against endpoint reporting as the strength of the planted effect varies.
Experiment 2: MIMIC-IV mechanical ventilation. We revisit MIMIC-IF's finding on ventilation duration, and on receipt of invasive ventilation, by ethnicity. Contested variables include insurance, comorbidity indices, code status, and admission source. The decisive check is whether the gap survives under prioritarian targets and under a majority of admissible partitions. It is also whether a single concession, for example classifying code status as legitimate, accounts for most of it.
Experiment 3: generalization to a second clinical decision. We repeat the analysis on MIMIC-III with a second treatment decision, to see whether attribution patterns are decision-specific.
Throughout. Results are reported jointly with the hidden-bias frontier. The proposal is falsified if robustness summaries add no information beyond the two endpoints in both semi-synthetic and real settings.
References
- Algorithmic Unfairness Through the Lens of EU Non-Discrimination Law
- A Moral Framework for Understanding Fair ML through Economic Models of Equality of Opportunity
- Where Is the Normative Proof? Assumptions and Contradictions in ML Fairness Research
- Are There Exceptions to Goodhart's law? On the Moral Justification of Fairness-Aware Machine Learning
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- Causal fairness assessment of treatment allocation with electronic health records
- Distributive Justice as the Foundational Premise of Fair ML: Unification, Extension, and Interpretation of Group Fairness Metrics
- Fairness without Demographics on Electronic Health Records
- MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset
- Counterfactual Fairness