Decision on SIG-2026-0295

[MSOM Conference]

08-May-2026

Re: SIG-2026-0295, "Robust Logit-Based Analytics"

SIG Day Decision: Reject

Dear Author (this is to ensure anonymity):

We received many excellent submissions for the Healthcare Operations Management SIG-Day Conference. Unfortunately, we could not accept all of them to be included in the program, and we are sorry to say that your paper was not accepted to the SIG-Day conference.

If you also submitted an extended abstract of your paper to the main MSOM Conference, a decision on that submission will be made separately.


Sincerely,

Healthcare Operations;SIG Co-Chairs

MSOM Healthcare Operations Management SIG-Day Co-Chair

---------------------
Referee: 1
Strengths SIG Only: The paper proposes a framework for robust logit-based analytics, where they capture uncertainty in the estimated multinomial logit (MNL) parameters using robust optimization. The authors further establish statistical finite-sample guarantees and show that the formulated problem can be reformulated as a mixed-integer exponential conic optimization problem. Numerical experiments validated in product line design problem show improvements on the out-of-sample performance.

1. The manuscript proposes a well-motivated and practically important question. The existing predict-then-optimize approaches for logit optimization could indeed expand to include the important nuance of parameter estimation uncertainty. This is an area not previously systematically studied and explored, which poses as a literature gap.

2. The worst-case payoff structure is novel and interesting. Namely, this result shows an insightful observation that the worst case is a rebalancing process that depends on interactions between the payoffs and choice probabilities, which is qualitatively nontrivial.

3. Previous approach including Akchen and Mišić rely mostly on structured uncertainty sets that do not provide statistical justifications. The inclusion of a finite-sample guarantee result thus is novel in this robust logic optimization domain and provides important value to the audience.

Referee: 2
Strengths SIG Only: Not relevant

Referee: 3
Strengths SIG Only: The manuscript’s most important strengths are that it identifies a real and consequential weakness in the standard logit-based estimate-then-optimize pipeline and then offers a fairly complete response to that weakness. The paper clearly motivates the problem by arguing that treating estimated MNL parameters as ground truth leaves downstream decisions exposed to estimation error, model misspecification, and the optimizer’s curse. It then proposes a “unified framework for robust logit-based analytics” that explicitly incorporates parameter uncertainty into optimization. More importantly, the paper builds a substantial technical chain, including an exponential-conic reformulation of the MLE problem, a finite-sample performance guarantee, and an analysis that derives dual structure, a worst-case payoff characterization. The authors also did a good job of positioning the novelty, especially by distinguishing this paper from prior robust MNL work that mainly treats exogenous utility uncertainty and by emphasizing the attempt to handle multiple alternatives with a unified reformulation and statistical guarantees. Finally, the numerical section reports that PLD-RO “consistently improves” the PLD-ETO baseline for suitable radii. It shows that the framework is useful when data are limited, and that under model misspecification it achieves higher average profit with lower dispersion.

Referee: 1
Limitations: 1. Overall positioning: My biggest concern of this paper is its overall positioning. In particular, Kannan et al. (2024) and Bertsimas and Van Parys (2021) both provide a much more general framework of a similar class of problem, without restricting themselves to only the MNL model class. Thus, it remains to be shown by the authors, if they want to focus on the logit optimization problem, what special properties of this sub-class provides that are fundamentally more interesting. In its current form, the paper reformulates the problem via perspectification then McCormick linearization to derive the MIECO formulation, but these techniques are rather standard and do not reveal properties that were not previously understood. I would suggest the authors either 1) extend the framework to more general problem classes, or 2) delve a bit deeper into what this specific framework reveal that’s not generally observed. For example, Theorem 3’s reweighting payoff structure is genuinely interesting, but more elaborate discussions (currently only one short paragraph) are needed to fully bring forth this insight.
2. Computational Tractability: the key difficulty observed in the Akchen and Mišić paper’s two approaches is the fact that under their robust framework when dealing with multiple customer types, the separability property breaks down and thus the problem becomes computationally difficult. What the manuscript claims to resolve around this problem is via the above perspectification etc. approach. However, the empirical experiments adopt a scenario of N=1 (although, as the authors have pointed out, follows previous work), and thus seems to have altogether evaded the primary novelty point of the paper’s ability to handle cases where N>=1. In particular, with the nominal N=1 case, this seems to exhibit the same structure as what Akchen and Mišić has already shown in their Approach 1 (EC6.1). It thus remains to be shown if the claimed computational tractability still holds in more realistic scenarios with larger N that can be solved quickly in MOSEK.
3. Writing Structure: the paper does not follow traditional MSOM writing styles and have extensive proofs in the main body. It further lacks elaborate intuitions for why certain lemma, theorem piece together or what insights they demonstrate. It also does not have the important sections of managerial implications/operational insights where the authors can discuss how the current setup impacts the decision-making process for which stakeholders under what contexts.
4. Contribution: The paper follows a sequence of well-established literature investigating either robust logit optimization or robust contextual optimization but fails to concretely pinpoint exactly how the current manuscript compares against this stream of literature. Its current discussion of the contribution seems a bit too general. I would suggest either to highly why this is a technically challenging problem, or have more explicit experiments or discussions to better understand its potential for impactful decision making.

Referee: 2
Limitations: This paper is not a good fit for the healthcare SIG. It is not really a healthcare paper and only includes one small healthcare example.

Referee: 3
Limitations: The manuscript’s main limitations are that its claims are broader than the evidence it actually delivers. The paper presents itself as a “unified framework for robust logit-based analytics” with extensions to finite-mixture and even other discrete-choice models (3 examples was given), but the empirical validation is confined to a single application, product line design, and that study is entirely synthetic rather than based on field or real operational data; in fact, the experiments are generated from simulated MNL/MMNL environments and, in the product-line setup, the paper even fixes N=1, which makes the numerical setting much narrower than the broad framing would suggest. A second limitation is that the theoretical guarantee is built on fairly strong assumptions. Theorem 1 requires a well-specified MNL data-generating process, independence, positive definiteness of the empirical-loss Hessian, and a nontrivial sample-size condition. So the guarantee may be less informative in the misspecified, weak-data, or ill-conditioned settings that most motivate the paper in the first place. Relatedly, the paper itself concedes that in practice the uncertainty radius \gamma would be chosen by bootstrap or cross-validation rather than by the theorem’s bound, which weakens the practical force of the formal guarantee. A third limitation is that the tractability result is not as general as the high-level narrative suggests: the exact MIECO reformulation relies on a linear feature map and, “in many applications,” on a binary feasible set so that the bilinearities can be linearized exactly, meaning the computational story is strongest only for a narrower class of problems than the paper’s general framework language implies. More broadly, the manuscript is mathematically ambitious but the exposition is dense, so readers may struggle to see clearly which parts are genuinely general, which parts depend on special structure, and how much incremental value this adds beyond combining robust optimization, MNL estimation, and conic reformulation in one package.

Referee: 1

Comments to the Author
(There are no comments.)

Referee: 2

Comments to the Author
(There are no comments.)

Referee: 3

Comments to the Author
(There are no comments.)