📊 statistics

Counterfactual Optimization of Policy Interventions: Lexical Ordering and Leapfrogging

This paper proposes a harm-aware policy optimization framework that prioritizes the ethical principle of "first do no harm" by designing policy transitions with a lexical leapfrogging structure to improve overall welfare while strictly limiting the risk of individual harm, as demonstrated through a reanalysis of the I-SPY2 breast cancer trial.

Original authors: Martina Scauda, Tobias Freidling, Qingyuan Zhao

Published 2026-08-24

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Martina Scauda, Tobias Freidling, Qingyuan Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of medicine and public policy, a fundamental tension exists between doing good for the many and avoiding harm to the few. For decades, data-driven systems designed to assign treatments or interventions have operated on a simple principle: maximize the average benefit. If a new drug saves more lives on average than the old one, the algorithm recommends it for everyone who fits the profile, regardless of how it might affect specific individuals. This approach, while efficient, overlooks a critical ethical reality: a policy that looks good on a spreadsheet can still be disastrous for a substantial fraction of the people it touches. The ancient medical principle of "first do no harm" suggests that we should not simply chase the highest average score if doing so inevitably crushes a vulnerable minority. The challenge for modern scientists is to build decision-making tools that respect this limit, finding a way to improve the overall outcome without crossing a line where too many people are made worse off.

This is the precise problem tackled by researchers at the University of Cambridge and École Polytechnique Fédérale de Lausanne. They developed a new method for designing policy changes that explicitly accounts for the risk of individual harm. Instead of asking only "what works best on average?", their approach asks, "how can we move from our current treatment to a better one while ensuring that the worst-case chance of hurting someone stays below a specific, acceptable limit?" The researchers found that when you impose this strict safety constraint, the optimal way to change policies is not a gradual, step-by-step improvement. Instead, the best strategy often involves a "leapfrogging" structure. This means that for any specific group of patients, the system should either keep them on their current treatment or immediately switch them directly to the single best available option for their specific profile. It skips over any intermediate treatments that might seem like a safe middle ground but ultimately fail to provide the best protection against harm.

To understand how this works, imagine a hospital with several different treatment options for breast cancer, each with varying success rates depending on a patient's specific biological markers. Traditionally, a doctor might look at the average success rate of a new drug and decide to offer it to everyone who shows a slight statistical benefit. However, the new method reveals that this average can hide a dangerous truth: for some patients, the new drug might be a miracle cure, while for others with a slightly different biology, it could be ineffective or even harmful. The researchers' model calculates a priority score for every possible patient-treatment combination. This score balances the potential gain of switching treatments against the worst-case risk of causing harm. If the risk of harm is too high for a particular group, the model dictates that they should not be switched at all, even if the average benefit looks promising.

The study was tested using real data from the I-SPY2 breast cancer trial, a large-scale study involving nearly a thousand patients and ten different treatment arms. The researchers applied their new "harm-aware" logic to this data and compared the results against standard methods that only look at average benefits. The difference was stark. The standard approach, which prioritizes based on average gains, suggested a certain order for which treatments to keep or drop. In contrast, the new harm-aware approach produced a completely different ranking. In some cases, a treatment that looked like a clear winner under the old system was deprioritized because the model identified a significant risk that it would harm a specific subgroup of patients. The researchers demonstrated that by accepting a small, controlled limit on how many people might be negatively affected, they could create a policy path that is ethically safer and, in many scenarios, more robust than the traditional average-maximizing approach.

A key insight from their work is the nature of the "leapfrogging" solution. When the researchers allowed for multiple treatment options, they discovered that the optimal policy rarely involves moving patients through a series of intermediate steps. For instance, if a patient is currently on Treatment A, and Treatment C is the best option, the model often dictates that the patient should jump straight to Treatment C, rather than moving to Treatment B first. This happens because intermediate steps can introduce unnecessary risks of harm without providing the full benefit of the best option. The mathematical structure of their solution ensures that every transition is either a direct move to the best possible care for that specific patient or no move at all. This creates a clear, hierarchical order of priority, where the most critical switches happen first, and those with higher risks of harm are left alone until the safety budget allows for them.

The researchers also explored how different assumptions about the relationship between potential outcomes affect these decisions. In the real world, we rarely know exactly how a patient would have reacted to a treatment they did not receive. The model accounts for this uncertainty by considering a range of possible scenarios, from the most optimistic to the most pessimistic regarding how treatments interact. They found that even when accounting for this deep uncertainty, the leapfrogging structure held true. Whether they assumed the worst-case scenario for how treatments might fail or a more moderate correlation between outcomes, the result was the same: a policy that prioritizes direct, safe transitions over gradual, risky ones. This suggests that the finding is not a fluke of a specific mathematical assumption but a fundamental property of trying to do good while strictly limiting harm.

In their analysis of the breast cancer data, the researchers visualized these findings by grouping patients into "tiers" of priority. They showed that depending on which method you use to rank the treatments, the top priority for discontinuing a failing therapy could be entirely different. Under the standard average-benefit method, a specific combination of drug and patient type might be the first to be dropped. Under the new harm-aware method, a different combination takes that spot, often because the standard method failed to see the hidden danger to a small but significant group. This reordering has profound implications for how clinical trials are run and how treatments are approved. It suggests that regulators and doctors might need to reconsider which treatments are safe to continue offering, not just based on who benefits the most, but on who is least likely to be hurt by the change.

The work does not claim to solve every ethical dilemma in medicine, nor does it provide a magic formula that eliminates all risk. The researchers are clear that their method controls the expected harm at a population level; it does not guarantee that every single individual will be protected. There are scenarios where the math dictates that a small number of people must be exposed to risk to achieve a larger benefit for the rest. However, by making this trade-off explicit and quantifiable, the method moves the conversation from vague ethical concerns to concrete, manageable limits. It allows policymakers to say, "We will improve the average outcome, but only if we can keep the worst-case harm below this specific threshold."

Ultimately, this research offers a new lens through which to view the complex machinery of medical decision-making. It challenges the long-held belief that the best policy is simply the one that yields the highest average score. Instead, it proposes that the most ethical and effective path forward is one that respects the limits of human safety, prioritizing direct, high-confidence improvements while avoiding the slippery slope of gradual, uncertain changes. By focusing on the "leapfrogging" nature of optimal transitions, the study provides a practical framework for navigating the difficult balance between doing good and doing no harm, ensuring that the pursuit of progress does not come at the cost of the vulnerable.

Technical Summary: Counterfactual Optimization of Policy Interventions: Lexical Ordering and Leapfrogging

Problem Statement
Existing data-driven policy learning methods typically aim to maximize average population outcomes (welfare), often overlooking the possibility that a policy beneficial on average may still inflict significant harm on a substantial fraction of individuals. This creates a tension between the utilitarian goal of beneficence and the ethical imperative of "first do no harm" (Primum non nocere). While some recent works incorporate counterfactual harm, they often focus on binary treatments, deterministic baselines, or specific resource constraints.

This paper addresses the problem of designing a policy transition (a change from a baseline policy $d$ to a new policy) that maximizes the expected population utility while ensuring the worst-case expected counterfactual harm remains below a specified bound $B$ . The framework operates under partial identification, where the marginal distributions of potential outcomes are known (e.g., from randomized trials), but the within-individual dependence (coupling) between potential outcomes is unknown. The goal is to find an optimal transition kernel $\kappa(z'|z, x)$ that maps current treatments to new treatments for individuals with covariates $x$ , subject to a constraint on the supremum of expected harm over a class of possible joint distributions $\Gamma$ .

Methodology
The authors formulate the problem as a constrained optimization:
$\max_{\kappa} E_{\kappa}\{Y(Z')\} \quad \text{subject to} \quad \sup_{\gamma \in \Gamma} E_{H_{\kappa,\gamma}} \le B$
where $h(y)$ is a harm function (typically the negative part function $h(y) = \max\{-y, 0\}$ ), and $\Gamma$ represents the class of admissible joint distributions of potential outcomes (e.g., the Fréchet class or classes with restricted dependence like non-negative correlation).

Key methodological steps include:

Lexical Ordering in Binary Settings: For binary treatments, the problem is reduced to a continuous knapsack problem. The authors show that the optimal transition follows a lexical ordering based on a priority score $\Lambda(x)$ , defined as the ratio of the Conditional Average Treatment Effect (CATE) gain to the worst-case expected harm of the transition.
Leapfrogging Lemma for Multi-valued Treatments: A central technical contribution is Lemma 2, which establishes that under conditions of subadditive harm and monotone recoupling (where the class $\Gamma$ allows for specific rearrangements of potential outcomes), an optimal policy transition need not proceed through intermediate treatment levels. Instead, it exhibits a "leapfrogging" structure: individuals are moved directly from their current treatment to the conditionally optimal treatment, skipping any suboptimal intermediate arms.
General Lexical Leapfrogging Theorem: Theorem 3 generalizes the result to multi-valued treatments and general outcomes. It proves that under the Fréchet class and stochastic dominance conditions, the optimal transition is determined by a priority score $\Lambda(z, x)$ :
$\Lambda(z, x) = \frac{E\{Y(z^*(x)) - Y(z) \mid X=x\}}{\sup_{\gamma \in \Gamma} E_\gamma\{h(Y(z^*(x)) - Y(z)) \mid X=x\}}$
where $z^*(x)$ is the conditionally optimal treatment. The transition assigns individuals to $z^*(x)$ with probability determined by a thresholding rule on this score until the harm budget is exhausted.
Closed-Form Priority Scores: The paper derives explicit formulas for the priority score under specific dependence models:
- Binary Outcomes: Under the Fréchet class (no dependence constraints) and under non-negative correlation assumptions.
- General Outcomes: Under Gaussian copula assumptions, providing a continuous expression for the worst-case harm that converges to the Fréchet bound as correlation approaches -1.

Key Results

Structure of Optimal Policy: The optimal policy transition is not a gradual improvement but a "leapfrog" to the best available arm for each subgroup. This structure arises because intermediate transitions can be replaced by direct transitions to the optimal arm without increasing worst-case harm or decreasing average utility, provided the dependence class satisfies specific recoupling properties.
Priority Score Definition: The optimal ranking of subgroups is determined by a ratio of expected benefit to worst-case expected harm. This differs from standard CATE-based ranking, which ignores the magnitude of potential harm.
Numerical Illustration (I-SPY2 Trial): The authors reanalyze data from the I-SPY2 breast cancer platform trial. They demonstrate that rankings of treatment-subgroup pairs vary significantly depending on the dependence assumption (CATE vs. non-negative correlation vs. Fréchet class).
- Under the Fréchet class (most conservative), certain treatment arms that appear optimal under CATE are deprioritized due to high potential counterfactual harm.
- The study shows that finite-sample uncertainty can lead to substantial variability in rankings, suggesting the use of "priority tiers" rather than strict linear rankings for decision-making.
Pareto Efficiency: Under total stochastic ordering of potential outcomes, the lexical leapfrogging solution is shown to be Pareto efficient for maximizing worst-case expected benefit while minimizing worst-case expected harm.

Significance and Claims
The paper claims to advance the literature on harm-aware policy learning by providing a general framework that accommodates:

Transitions from stochastic baseline policies (not just control).
Multi-valued treatments.
General harm measures and various dependence restrictions on potential outcomes.

The authors position their work as occupying a "middle ground" between unconstrained aggregate-welfare maximization (utilitarianism) and strict lexical priority (Rawlsian maximin). By imposing a bound on worst-case counterfactual harm, the framework generates a lexical solution path toward the optimal policy. This is distinct from standard resource-constrained optimization, where constraints are exogenous costs; here, the constraint is an intrinsic, causal, moral loss.

The paper emphasizes that counterfactual harm provides a perspective distinct from standard precision medicine. It shifts the focus from "who is worst off in absolute terms" to "who would be made worse off by the contemplated transition." The "leapfrogging" property is highlighted as a non-trivial finding that emerges specifically from the interaction between the harm constraint and the structure of the dependence class, distinguishing this approach from standard knapsack problems with linear costs.

The authors conclude that while their framework offers a principled way to incorporate ethical constraints into policy learning, it relies on scalar outcomes and single linear constraints. They note that extending this to multidimensional safety outcomes or multiple constraints would likely result in more complex optimization problems without the elegant single-score threshold solution.

Technical Summary: Counterfactual Optimization of Policy Interventions: Lexical Ordering and Leapfrogging

More like this