Clinical trials show that a treatment works on average in the patients they enrolled. Those patients are younger, healthier, and less diverse than the people who later receive the drug. A hospital that wants to decide whom to treat faces two problems. The first is the known difference in patient characteristics, which can be corrected by reweighting. The second is unmeasured differences that change how well the treatment works, which reweighting cannot fix. The natural defense against the second problem looks like adversarial training. Pick the treatment rule that performs best when an adversary is allowed to distort the target population's outcomes within some plausible limit. Adversarial training is expensive because every learning step must first solve an inner "find the worst case" problem, usually by repeated gradient ascent.
This note argues that for a well-motivated clinical sensitivity model, the one in Sharp Bounds for Treatment Effect Generalization under Outcome Distribution Shift, the inner problem does not need to be solved iteratively. It separates across patients and treatment arms, and each piece has an exact answer found by sorting. The robust treatment rule can therefore be learned at the cost of ordinary weighted classification. We propose to build that learner with a debiased estimator of its worst-case value, test it against the strongest non-robust transport method, and measure how much the closed form saves over running the adversarial game directly.
Core Claim
The sharp-bounds paper models unmeasured outcome shift with a likelihood ratio. Let $r_a(x,y)$ be the ratio between the target-population and trial-population density of outcome $Y(a)=y$ under treatment $a$, among patients with covariates $x$. The ratio is constrained to $\Lambda^{-1}\le r_a\le\Lambda$ for a sensitivity level $\Lambda\ge1$, and it must integrate to one within each $(x,a)$. That paper proves the worst-case ratio has a threshold form. It equals $\Lambda$ below a conditional outcome quantile and $\Lambda^{-1}$ above it. From this it derives $O(n\log n)$ sharp bounds and a tipping point $\Lambda^*$ for a scalar average effect. It does not learn any decision rule.
Our observation concerns a stochastic treatment rule $\pi(a\mid x)$, meaning a function that gives each patient's probability of receiving each arm. The robust target value of such a rule is
$$V_\Lambda(\pi)=\min_{r}\;\mathbb{E}_{X\sim\text{target}}\Big[\sum_a \pi(a\mid X)\,\mathbb{E}_{\text{trial}}\big[r_a(X,Y)\,Y(a)\mid X\big]\Big].$$
Because $\pi\ge0$ and the constraints on $r$ are separate for each $(x,a)$, the minimum over $r$ moves inside the sum. The result is $V_\Lambda(\pi)=\mathbb{E}_{\text{target}}[\sum_a\pi(a\mid X)\mu^-_{a,\Lambda}(X)]$. Here $\mu^-_{a,\Lambda}(x)$ is the sharp lower conditional mean given by the threshold rule.
The min-max problem therefore becomes a single maximization against fixed, closed-form per-patient scores. Any extra cost of adversarial training comes from estimating those scores, not from an inner loop. The nontrivial parts are statistical. We must estimate $\mu^-$ well enough that the learned rule's guaranteed value is trustworthy, and we must compute the whole path over $\Lambda$ cheaply so that clinicians can see how much unmeasured shift a rule tolerates before it stops beating the default.
How This Relates to Prior Work
The motivation comes from audits showing that the gap is real and concentrated in particular patients. TrialTranslator emulated 11 oncology trials on about 153,000 real-world patients. High-risk phenotypes had 62% lower median survival and a 5.7-month smaller survival benefit than trial patients. Low-risk patients roughly matched trial results. GIST scored trial-to-EHR overlap and found that 52% of real-world diabetes patients had baseline HbA1c below a range most trials target. The heart-failure meta-analysis in Clinical outcomes of pharmacological therapies for heart failure in Black vs. White populations shows treatment benefit varying across groups. Together these say that who-to-treat rules, not just average effects, are what fail to transfer.
The method builds on two lines of work. The first is the threshold result of the sharp-bounds paper, described above. The second is the doubly robust transport machinery of DRCS and the unified transportability framework. "Doubly robust" means combining a model of who enters the trial with a model of outcomes, so the estimate stays consistent if either model is right. DRCS learns policies under measured covariate shift but assumes no unmeasured outcome shift. We keep its density-ratio and outcome-model components and add the $\Lambda$-worst-case correction. The unified framework shows that different doubly robust constructions diverge for non-linear effect measures, with estimating-equation forms more stable. That result warns us that our debiased estimate of $\mu^-$ must be built carefully rather than by naive plug-in.
The work that bridges the two directions uses expensive games. The clinical adversarial debiasing framework trains a predictor against an adversary network by alternating gradient steps across NHS sites. It targets prediction fairness, not treatment effects, and it pays the full cost of alternating optimization. Distributionally Robust Causal Abstractions runs a Wasserstein min-max game, meaning a search over all distributions within a fixed transport distance of the data. It is limited to linear maps with known causal graphs.
The efficiency literature on adversarial training uses a different lever. ATTA reuses attacks across epochs, reporting a 14.1× speedup over PGD-40 on CIFAR-10. Robustness tokens shrink the set of trained parameters. Both make the inner loop cheaper. We show that for this sensitivity model the loop can be removed entirely, which is a different and stronger claim, though it is limited to a narrower class of adversaries.
Research Program
The learner has three stages, each cross-fitted: estimated on one data fold and applied to another, so the same data are never used both to fit a model and to evaluate it.
The first stage estimates trial-participation weights $q(x)/p(x)$ between trial and target covariates, as in DRCS. It also fits per-arm conditional quantile and conditional tail-mean models for $Y(a)$ given $x$ on trial data.
The second stage builds a debiased pseudo-outcome for $\mu^-_{a,\Lambda}(x)$ in each target patient. The plug-in threshold value is corrected by an influence-function term. This is the standard derivative-based adjustment that cancels first-order errors in the quantile, tail-mean, and participation models. We will derive it for the threshold functional, which is an expected-shortfall-like quantity whose correction is known to be well behaved when the quantile is estimated at a reasonable rate. Following the unified-framework result, we will compare one-step and estimating-equation constructions.
The third stage learns $\pi$ by maximizing the estimated $V_\Lambda$. This is cost-sensitive classification with per-patient, per-arm scores. It covers constrained rule classes such as shallow trees or budget limits on the fraction treated, and it can use a neural network when covariates are rich.
Two analyses accompany the learner. The first computes the full robustness path. Because the threshold rule depends on $\Lambda$ only through a quantile level, all scores for a grid of $\Lambda$ values come from one set of fitted quantile models. From these we compute, for each learned rule, the tipping point $\Lambda^*$: the sensitivity level at which its guaranteed value drops below that of treating everyone or no one.
The second analysis tests the limits of the separability argument. If the adversary faces a single population-wide budget on shift instead of per-patient bounds, the inner problem couples patients. We conjecture it then reduces to a sort plus a one-dimensional search over a Lagrange multiplier, and we will check this both analytically and numerically. If the conjecture fails, that marks the boundary where genuine adversarial-training overhead returns.
Baselines
The naive reference point is the trial-only rule: learn $\pi$ on trial data, ignore the target population, and apply the rule to the target. Beating it is necessary but not interesting.
The strongest existing competitor is DRCS policy learning from Off-Policy Evaluation and Learning for External Validity under a Covariate Shift. It is the only method in the file that learns a treatment rule for a shifted target population with semiparametric efficiency guarantees. We will run it with the same participation-weight and outcome models, the same folds, the same rule class, and the same hyperparameter budget as our learner. The only difference will be whether the $\Lambda$ correction is present. That paper reports its mean squared error only graphically, so no reported result was found on Lacuna for a numeric DRCS value on SatImage, Vehicle, or PenDigits. We will therefore report our own reproduction.
Our learner beats DRCS if two things hold. First, it must achieve higher true target value when hidden shift is present at or above a modest $\Lambda$. Second, it must lose no more than a small, prespecified margin of value when no hidden shift is present. We fix that margin before running the experiments at 1% of the oracle-minus-treat-none value.
A stronger version of the same competitor is DRCS followed by post-hoc sharp bounds. We train DRCS rules across its hyperparameter grid, compute each rule's $\Lambda$-lower bound with the greedy algorithm of the sharp-bounds paper, and deploy the rule with the best bound. This isolates whether training against the worst case matters, or whether selecting by the worst case suffices.
A generic robust-optimization competitor optimizes a conditional value-at-risk objective, meaning the average of the worst $\alpha$ fraction of outcomes, over transported pseudo-outcomes, with $\alpha$ tuned on the same budget.
Two ablations isolate our ingredients. The first optimizes the identical $V_\Lambda$ objective by gradient descent-ascent over an explicitly parameterized $r$, matching the adversarial-training recipe of the clinical debiasing framework. We give it the same wall-clock budget and also run it to convergence. This isolates the closed form. The second uses plug-in rather than debiased scores, which isolates the correction term.
Why not just…?
Why not just use a doubly robust transport estimator like DRCS, with weight clipping? Clipping handles extreme participation weights, which are an estimation problem. Unmeasured effect modifiers are an identification problem. The sharp-bounds paper argues that such estimators become "precisely wrong as sample sizes increase" under that kind of shift, and TrialTranslator's high-risk gap persisted after reweighting within the eligibility criteria. The semi-synthetic experiments below inject a known hidden modifier, so the size of this failure is measured rather than asserted.
Why not just train a standard rule and report its sharp bound afterwards? This is the most serious threat, because the separability result means our learner is itself a maximizer of closed-form scores. The difference is that DRCS-then-bound can only choose among rules that were optimized for the point estimate. Our learner reallocates treatment toward patients whose benefit is still positive under $\Lambda$. The oncology finding that the gap concentrates in high-risk phenotypes suggests these sets differ materially. If the post-hoc selection baseline matches our learner in true target value across the $\Lambda$ sweep, we will report that. The contribution then shrinks to the debiased estimator and the robustness path.
Why not just run adversarial training or a Wasserstein robust game directly? Both are possible. The first ablation measures what they cost and whether they reach the exact optimum, since projected ascent on normalized, bounded ratios may stall. The causal abstraction work also needs linear maps and a known causal graph, which clinical data rarely provide. We concede that on small tabular problems the wall-clock savings may be modest. The more practical gain is computing an exact path over many $\Lambda$ values from one set of fits.
Why not just use a CVaR loss? A tail-average objective ignores the trial-to-target structure and uses a single tail level that clinicians cannot interpret. $\Lambda$, by contrast, is a bound on how differently outcomes may be distributed between trial and target patients who look alike. If tuned CVaR matches our learner's target value, the interpretability argument remains, but the performance claim fails. We will report whichever outcome occurs.
Proposed Experiments
The first testbed extends the four simulation data-generating processes from the sharp-bounds paper (linear, nonlinear, binary, and heavy-tailed) by adding heterogeneous treatment effects. We generate target populations with hidden shift at known $\Lambda_{\text{true}}\in\{1,1.5,2,3\}$. That paper reports coverage and tightness only graphically, so there is no reported number on Lacuna to match.
The second testbed converts the SatImage, Vehicle, and PenDigits bandit datasets used by DRCS in the standard supervised-to-bandit way. We build source and target splits by selecting on a covariate, then hide a reward-modifying feature from the learner so that the injected outcome shift is unmeasured and calibrated.
The metrics are true target value and regret against the oracle rule; the robust value at the deployed $\Lambda$; coverage of the debiased lower bound over 200 replications; tipping point $\Lambda^*$; and wall-clock time together with the number of gradient evaluations.
The real-data case study transports CRASH-3 tranexamic-acid trial results to the Traumabase registry, as in the unified-framework paper. No reported effect estimates were found on Lacuna for this pair, so we will report our reproduced point estimates, the learned rules, and their tipping points. There is no ground truth, so the purpose is to show whether a rule's advantage over "treat all" survives plausible $\Lambda$.
Three checks decide the proposal.
The first check tests the closed form. The converged descent-ascent ablation should match the closed-form optimum up to optimization error, while using many times more gradient evaluations. A persistent gap would mean our separability derivation is wrong, or that descent-ascent fails to converge.
The second check tests whether robust training helps. Our learner should beat both DRCS and DRCS-then-bound in true target value once $\Lambda_{\text{true}}\ge1.5$, and it should stay within the prespecified margin at $\Lambda_{\text{true}}=1$. Losing to DRCS-then-bound would reduce the method to post-hoc selection.
The third check tests the debiased estimator. Its lower bound should reach nominal coverage, while the plug-in version should undercover in the heavy-tailed process, where quantile errors bite hardest.
A final exploratory check, on the budgeted-shift variant, tells us whether the no-inner-loop result extends beyond per-patient bounds or marks where adversarial-training overhead becomes unavoidable.
References
- Sharp Bounds for Treatment Effect Generalization under Outcome Distribution Shift
- Evaluating generalizability of oncology trial results to real-world patients using machine learning-based trial emulations
- A Distribution-based Method for Assessing The Differences between Clinical Trial Target Populations and Patient Populations in Electronic Health Records
- Clinical outcomes of pharmacological therapies for heart failure in Black vs. White populations: a meta-analysis of randomized controlled trials of heart failure treatment
- Off-Policy Evaluation and Learning for External Validity under a Covariate Shift
- A Unified Framework for the Transportability of Population-Level Causal Measures
- An adversarial training framework for mitigating algorithmic biases in clinical machine learning
- Distributionally Robust Causal Abstractions
- Efficient Adversarial Training With Transferable Adversarial Examples
- Robustness Tokens: Towards Adversarial Robustness of Transformers