Chance-constrained selection of sequential intervention strategies from counterfactual estimates

This paper addresses the challenge of selecting sequential intervention strategies under a cumulative resource budget, where the decision maker must maximize an outcome while constraining the probability that the cumulative cost exceeds a budget. The authors note that many operational decisions are not single choices but sequences of interventions carried out under a cumulative resource budget, citing examples including maintenance scheduling, customer retention offers, and clinical treatment administration.

The key insight motivating the work is that two strategies with the same expected cost can exceed the budget at very different rates, so constraining the mean does not bound how often an overrun occurs. The paper identifies a gap in existing methods: No established method constrains the upper tail of a cost that accumulates over a horizon and is identified counterfactually from observational data.

The paper presents a predict-then-optimize framework with two distinct steps. In the prediction step, any estimator returning an outcome value and a cost distribution supplies what the decision rule consumes, so the predictor is interchangeable. In the optimization step, a chance-constrained selection over a finite candidate set bounds the probability that the cumulative cost exceeds the budget.

The core decision rule is formulated as:

g cc* = g (g) subject to (C(g) > B) at most epsilon

where V(g) is the outcome value, C(g) is the cumulative cost, B is the budget, and epsilon is the tolerated violation probability.

A key theoretical result is Proposition 1, which establishes the Non-decomposability of the cost tail: the tail probability of a cumulative cost does not decompose across stages. The paper demonstrates this with a two-stage construction where two strategies indistinguishable stage by stage, in marginal law and in mean, differ in violation probability by a factor of two.

The paper establishes two distribution-free finite-sample guarantees. Proposition 3 provides a feasibility bound: if the uniform deviation between empirical and true cost distributions is bounded by delta, then "every strategy admitted by the empirical chance-constraint rule satisfies (C(g) > B) at most epsilon + delta." Proposition 4 provides a regret bound against a benchmark at a tightened tolerance, showing that with probability at least 1-eta, the selected strategy satisfies both the feasibility constraint and V(g*(epsilon - delta)) - V at most 2 epsilon V.

The paper evaluates the framework across five environments: sepsis treatment simulation, tumor-growth modeling, a semi-synthetic MIMIC-IV intensive care cohort, a C-MAPSS predictive maintenance environment, and the Drink Less micro-randomized trial with real outcomes. Four environments provide exact counterfactual ground truth, while the fifth carries real outcomes.

Key findings include:

1. The naive point-estimate rule overruns the budget in every environment, between 5.3% and 15.2% of the time, while the chance-constraint rule holds the violation at or below 0.21%.

2. The chance constraint costs outcome: the chance-constraint rule's regret ranges from 9.5 to 16.9 percentage points, versus 0.4 to 3.5 for the naive rule.

3. Sweeping the tolerance traces a safety-utility frontier that is monotone: the outcome regret is non-increasing and the budget violation is non-decreasing in every one of the four environments.

4. The rule is robust to estimation error: Under mean-zero noise, the chance-constraint violation is exactly zero at every injected-error level from zero to 30%, while the naive rule's violation falls from 12.8% to 6.8% as noise increases.

5. The rule withstands moderate cost bias: The chance-constraint violation stays at zero through a 10% underestimate, reaches 0.70% at 30%, then 9.2% at 40% and 33.1% at 50%.

6. The frontier is not convex: the second difference of regret changes sign 8 to 12 times depending on the environment, so an operating point is therefore selected by reading the frontier, not by locating a knee.

The paper acknowledges several limitations: the decision rules are established against exact or known outcomes; the effect of estimation error is studied through injected noise; the method selects a strategy once at the start of the horizon without re-selecting; and the candidate sets are constructed rather than drawn from practice. The paper notes that sequential ignorability is not testable from observational data alone and would have to be argued from domain knowledge about what drives treatment decisions.

Based on the paper, here are the specific improvements I can make to AI systems, and what the improved system can do:

Improvement 1: Tail-Aware Sequential Decision Making

I can augment any AI system that selects sequential interventions (e.g., treatment policies, maintenance schedules, marketing campaigns) with a chance-constrained optimizer that explicitly bounds the probability of cumulative cost overrun, rather than only optimizing expected cost. The improved system can:

• Accept any predictive model (for outcomes and cost distributions) and produce a decision rule that guarantees, with finite-sample confidence, that the budget violation probability stays below a user-specified tolerance (e.g., 5%).

• Automatically detect when a naive point-estimate policy would overrun the budget (as shown in the paper: 5–15% violation) and replace it with a safer policy, while quantifying the outcome regret trade-off.

Improvement 2: Safety–Utility Frontier Generation for Human Oversight

I can build a module that sweeps the violation tolerance ε and outputs a monotone safety–utility frontier (regret vs. violation probability) for any candidate set of strategies. The improved system can:

• Present a human operator with a non-convex frontier (not a single knee point) and let them choose an operating point based on risk appetite, rather than forcing a single optimal solution.

• Provide interpretable warnings: e.g., Reducing violation probability from 5% to 1% will cost 7 percentage points of outcome value, enabling informed trade-offs in clinical or industrial settings.

Improvement 3: Robustness to Estimation Error and Cost Bias

I can enhance the decision rule with the paper’s distribution-free guarantees, making the AI system robust to imperfect predictions. The improved system can:

• Maintain zero budget violations even under mean-zero noise up to 30% in cost predictions (as demonstrated), and tolerate up to 10% systematic cost underestimation without violating the chance constraint.

• Automatically flag when cost bias exceeds a safe threshold (e.g., >30% underestimate) and trigger a re-estimation or human review, preventing catastrophic overruns.

Improvement 4: Non-Decomposable Cost Tail Handling

I can implement the paper’s Proposition 1 insight to avoid a common pitfall: treating each stage’s cost independently. The improved system can:

• Correctly compute the joint tail probability of cumulative cost across stages, even when stage-wise marginals are identical, preventing a factor-of-two error in violation probability.

• This is critical for AI systems in multi-step clinical treatments or multi-phase logistics, where ignoring cross-stage dependence would lead to unsafe decisions.

Improvement 5: Interchangeable Predictor Integration

I can design the optimizer as a plug-in module that works with any existing outcome/cost predictor (e.g., regression, survival models, neural nets). The improved system can:

• Seamlessly upgrade existing AI pipelines (e.g., a sepsis treatment recommender) by adding a chance-constraint layer without retraining the predictor, reducing deployment cost.

• Provide a finite-sample feasibility guarantee (Proposition 3) and a regret bound (Proposition 4) that hold even when the predictor is misspecified, as long as uniform deviation is bounded.

Improvement 6: Counterfactual Safety Auditing

I can build a pre-deployment audit tool that uses the paper’s evaluation methodology (exact counterfactual ground truth in four environments) to test any candidate policy. The improved system can:

• Run a simulated stress test (with injected noise and cost bias) to report the worst-case violation rate and regret, before real-world deployment.

• Identify whether a policy’s safety is fragile (e.g., violation jumps from 0% to 33% at 50% cost bias) and recommend a more conservative tolerance or a different candidate set.

What the improved AI system can do overall:

It can make sequential decisions under a cumulative budget with a provable upper bound on the probability of overrun, while explicitly trading off outcome value. It remains safe under realistic estimation errors, handles non-decomposable cost tails, and provides a human-readable frontier for risk-based selection—all without requiring changes to the underlying predictive models.