acceptodds
Under review as a conference paper at ICLR 2027

Fixed-Confidence Identification of Capacity-Constrained Treatment Allocations with Prognostic Covariates

Abstract

Randomized experiments are often used to guide treatment allocation under limited capacity. Existing work has studied fixed-confidence identification under structured or constrained decision sets and covariate adjustment for improving treatment effect estimation. However, it remains unclear whether individual-level covariates that predict outcomes, which we call prognostic covariates, can reduce the sample complexity of identifying an optimal capacity-constrained group-level treatment allocation. To address this gap, we study fixed-confidence identification of capacity-constrained group-level treatment allocations with prognostic covariates. We characterize the first-order expected stopping-time benchmark when individual covariates are retained and show that it can be strictly smaller than the benchmark based on group-level information alone. We then develop an adaptive procedure that learns the unknown prognostic functions from accumulating data while controlling the allocation-selection error in finite samples, and prove that its expected stopping time asymptotically attains the same first-order benchmark as when the prognostic functions are known. Experimental results confirm the predicted information gap and show the learned procedure approaching the full-covariate benchmark.

Primary area:
causal reasoning
Keywords:
Causal Inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GYvdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies fixed-confidence identification of optimal capacity-constrained treatment allocations using individual-level prognostic covariates. It derives an information-theoretic sample complexity lower bound, shows that observing prognostic covariates can reduce the required sample size, and proposes an adaptive procedure that learns unknown outcome regression functions while maintaining finite-sample error control. The authors claim they may attain the optimal first-order expected stopping time.

Critical strengths

The problem is practically relevant, and the paper provides a coherent connection between constrained experimental design, covariate adjustment, and fixed-confidence best-arm identification. Also, the use of predictable likelihoods and composite-null profiling may offer a principled way to maintain sequential error control while learning nuisance functions.

Critical weaknesses
  1. The proposed framework largely combines existing ideas from covariate adjustment, best-arm identification, and sequential inference, without a clearly distinctive methodological contribution.

  2. The theoretical guarantees rely on pretty restricted conditions such as Gaussian outcomes and known residual variances and covariate distributions. The deployment objective also requires clearer justification.

  3. Experiments seem to be very limited to a single simple synthetic setting with only two feasible allocations, providing little evidence of general applicability. Further, the proposed method outperforms the benchmark only at extremely stringent confidence levels, leaving its advantages in realistic settings unclear.

  4. The adaptive design is initialized at its oracle optimum and remains fixed, while comparisons rely on an oracle benchmark rather than practical competing methods.

Questions for authors

See the weakness section above.

Reviewer_8TphOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies capacity-constrained group-level treatment allocations using prognostic covariates with fixed-confidence identification. The paper investigate whether such covariates can reduce sample complexity, even when they cannot be used for treatment randomization or final deployment decisions. The paper establishes a theoretical improvement in sample complexity from retaining prognostic covariates and proposes an adaptive algorithm with finite-sample error control and asymptotic optimality. Synthetic experiments were also conducted.

Critical strengths
  1. The paper studies an interesting and specific problem, as whether individual-level covariates can improve the sample efficiency of identifying optimal group-level allocations under capacity constraints. The distinction between using covariates for statistical efficiency and using them for treatment assignment is particularly interesting.
  2. The synthetic experiments provide supporting evidence for the theoretical information gains and prognostic learning behavior.
Critical weaknesses
  1. The empirical evaluation is limited to a simplified synthetic setting with two groups, two treatments, and a single binary prognostic covariate. Moreover, as stated in the setup "We set q(0) = q∗, which gives qt = q* at every round. ", the experiments fix the treatment allocation probabilities, so they do not evaluate the adaptive design-tracking component. It remains unclear how the complete adaptive procedure performs with more groups, higher-dimensional covariates, or more complex prognostic functions.
  2. The experiments primarily compare the proposed method with oracle variants and a numerical coarsened-data benchmark. There are no direct comparisons with existing or suitably adapted methods from related best-arm identification or sequential experimental design literature.
Questions for authors

Following weakness.

  1. Could the authors provide additional experiments with more groups and higher-dimensional prognostic covariates? And evaluation on the adaptive design-tracking component since it is also part of the proposed algorithm.
  2. Could the authors include additional baselines, potentially by adapting existing sequential identification methods to the same allocation and sampling constraints if possible?
Reviewer_5QDMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies fixed-confidence identification of optimal capacity-constrained treatment allocations using individual-level prognostic covariates. The authors investigate whether retaining covariates that predict outcomes, but do not affect treatment randomization or deployment decisions, can reduce the sample complexity of identifying the optimal allocation. They establish an information-theoretic comparison between full-covariate and coarsened experiments, showing that TF∗≤TC∗T_F^* \leq T_C^*, and with strict improvement possible under capacity constraints. Under a centered Gaussian RKHS model, they derive the full-covariate characteristic time and show that learning unknown prognostic functions does not affect the first-order benchmark. They further propose an adaptive sequential procedure with finite-sample δ\delta-correctness and asymptotically optimal expected stopping time. Synthetic experiments illustrate the theoretical information gain and show that the learned procedure approaches the optimal benchmark as δ→0\delta \to 0.

Critical strengths
  1. I think the paper has a good identification of a well motivated problem. It addresses the practically relevant problem of identifying optimal treatment allocations under capacity constraints and investigates how prognostic covariates can improve sample efficiency.

  2. The paper provides an information-theoretic characterization of sample complexity and establishes finite-sample δ\delta-correctness and asymptotically optimal expected stopping time for the proposed procedure.

  3. Coherent theoretical and empirical framework: The paper connects information-theoretic lower bounds, adaptive experimental design, and sequential inference in a well-structured framework. The synthetic experiments illustrate the predicted information gains and convergence toward the theoretical benchmark.

Critical weaknesses
  1. I find the theoretical novelty somewhat limited. The inequality TF∗≤TC∗T_F^* \leq T_C^* follows directly from the data-processing inequality, and the general observation that prognostic covariates improve statistical efficiency is already well established.

  2. The result that unknown prognostic functions do not affect the first-order characteristic time is interesting, but it appears to rely heavily on the centered additive Gaussian model. In particular, the least-favorable alternatives preserve the prognostic functions because of the orthogonal KL decomposition. I would like to better understand how much of this result extends beyond the specific model considered.

  3. Several assumptions seem quite restrictive: such as the analysis assumes that the covariate distribution P_X and residual variances are known. These quantities would generally need to be estimated in practice, and it is unclear whether the main guarantees continue to hold in that setting.

  4. The asymptotic optimality theorem is a useful result, but the paper provides limited insight into finite-sample efficiency. An explicit nonasymptotic stopping-time bound would help clarify the additional sample cost of learning the prognostic functions and when the asymptotic benefits become practically meaningful.

  5. The empirical evaluation is rather limited. All experiments are based on essentially one synthetic setting with two groups and two treatments. It would be helpful to consider more complex allocation problems, different prognostic functions, and varying levels of covariate informativeness.

  6. the authors considered the experiments that initialize the sampling design at the true information-optimal solution, q(0)=q∗q^{(0)}=q^*, if I understand this correctly this means the design remains fixed throughout the experiment? If so, as a result, the experiments do not seem to be good at evaluating the adaptive design-tracking component, which is an important part of the proposed procedure and its theoretical analysis.

Questions for authors

Please refer to questions in Weakness 2, 3, 5 and 6 above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.