Decision on SIG-2026-0136

[MSOM Conference]

08-May-2026

Re: SIG-2026-0136, "Leveraging Observational Data for Adaptive Clinical Trials: A
Deconfounded Warm-Start Bandit Approach"

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 authors work on an important area --- leveraging rich non-experimental datasets to reduce regret in adaptive experiments. Their theory results do a good job of establishing the magnitude of potential regret improvements from their proposed approach versus LinTS.

Referee: 2
Strengths SIG Only: The paper addresses an important practical challenge in adaptive experimentation, which is how to reduce the amount of costly online learning needed before making good decisions.

Another major strength is the way it uses observational data in a creative way that works even in a high-dimensional setting with bias. More generally, the paper brings nice statistical ideas into sequential decision-making problems. I can see it being an attractive approach in many applications beyond healthcare and promoting interesting future research questions.

The theory also helps give insight on the benefit of their approach. Specifically, they show the value of effective dimension reduction and more informative initialization as the offline sample size increases.

Referee: 3
Strengths SIG Only: This manuscript presents a novel technique to be able to leverage observational data to more quickly determine personalized adaptive treatment, which is mathematically rigorous.

Referee: 1
Limitations: My limitations listed below are almost exclusively about the fit of this work with the Healthcare SIG.

The model setup has some hidden confounders in the large-scale observational data, but assumes that all confounders are fully observed when adaptively experimenting. I think there are certainly some real-world examples where this assumption holds, but they're probably pretty rare in healthcare settings. Doctors make treatment assignment decisions based on numerous factors that are not going to be fully observed in the EHR or collected in surveys as part of an adaptive experiment. Take the example of CVD management from the ACCORD study in Section 7. Providers consider many things when choosing intensive versus standard treatment -- patient history of missed appointments / medicine non-adherence, health literacy, practical constraints from job or living situation, family support, patient preferences of various types, and other tacit knowledge from past interactions with the patient. Few, if any, of these are going to end up being captured in the trial, yet they might have big effects on provider decision making in the observational data.

The numerical case study with the ACCORD trial (Section 7) was the main link to the healthcare SIG topic, but it had a synthetic feel to me. The authors chose hidden covariates that would actually generally be observed in the offline dataset (age, sex, race, comorbidities) and don't have data on the types of unobservables we're most concerned with the the healthcare setting (see above). Further, response-adaptive randomization is quite rare in clinical trials --- a recent systematic review [1] found only 65 examples over a nearly 4-decade time range. Ultimately the model didn't feel like a great fit for the healthcare context; it actually seems like it might be a better fit with some of its other motivations such as personalized recommendation systems, where the confounding assumptions may better hold and where bandits are widely used.

[1] Isabelle Wilson, Steven Julious, Christina Yap, Susan Todd, Munyaradzi Dimairo (2025). Response adaptive randomisation in clinical trials: Current practice, gaps and future directions. Statistical Methods in Medical Research 34(9): 1851-74.

Referee: 2
Limitations: Their approach utilizes some assumptions which may not be that practical. For instance, hidden confounders are assumed to be measurable in the online setting.

The numerical experiments lack insights beyond simply validating the overall performance of the approach at a limited scale. The setup of the numerical study also could be refined given it utilizes two different population datasets in a semi-synthetic way.

The algorithm and analysis depend on a threshold parameter for selecting variables to exclude for dimensionality reduction. There is not enough discussion of this dependence, given that it could cause the algorithm to start off in the wrong direction and lead to worst performance.

Referee: 3
Limitations: The manuscript is very detailed on the mathematical side of things, but the application itself could be better motivated. I would like to have heard more discussion of the actual applicability and benefit of using the developed technique in this drug approval setting, where I forsee there could be challenges in promoting the adoption of such a technique. The case study relied on some strong assumptions, and I think there should be clearer discussion of the choice of the multiplicative reduction factor. As a side note, I also found a few typos and broken references in the text; the authors should consider slightly more in-depth copy editing.

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.)