Should I Stop or Should I Go: Early Stopping with Heterogeneous Populations

Randomized experiments often need to be stopped prematurely due to the treatment having an unintended harmful effect. Existing methods that determine when to stop an experiment early are typically applied to the data in aggregate and do not account for treatment effect heterogeneity. In this paper, we study the early stopping of experiments for harm on heterogeneous populations. We first establish that current methods often fail to stop experiments when the treatment harms a minority group of participants. We then use causal machine learning to develop CLASH, the first broadly-applicable method for heterogeneous early stopping. We demonstrate CLASH's performance on simulated and real data and show that it yields effective early stopping for both clinical trials and A/B tests.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer tX2R6/10 · confidence 4/52023-07-01

Summary

This work proposes a method for adapting stopping tests of randomized experiments in heterogeneous populations. Specifically, the authors motivate the problem, namely why heterogeneous treatment effects lead to late stopping of randomized experiments, for instance when a minority group is harmed. They then propose a two stage method which first predicts a weighting of the original test statistic components used in stopping decisions, then and then uses these statistics to make the stopping decision. The methodological contribution is well-motivated and supported by theoretical results which analyse the convergence behavior of these weights, and the probability of stopping under the assumption of knowing the group membership.

Strengths

* There is a lack of machine learning methods that addresses the heterogeneous early stopping problem, and this paper provides a possible, first solution while making few assumptions. This renders the work an original, well-motivated contribution. * The paper is very clear and well written. In particular, the links between each of the sections are very clear. * The experimental results, including the simulated scenarios, are convincing and interesting. For instance Figure 2 makes clear why the proposed approach is advantageous over homogeneous stopping tests.

Weaknesses

* The task is somewhat niche. It is furthermore unclear to what degree the stopping task in randomized experiments could be reformulated as a similar task in another domain, i.e. to what degree this or similar problems have been solved in other contexts. * Overall, the work makes various idealised assumptions in both the theoretical results, and in the (synthetic) experiments considered. It is an interesting proof of concept, but there is a lot of work which would need to be done to make this method applicable in practical settings, for instance real-world clinical trials which this method is motivated by, but does not evaluate on. Two further points to note is the lack of performance on high-dimensional data that the authors themselves notes, which is present in some clnical trials. Second, the method crucially relies on treatment effect estimation methods in Stage 1, which themselves are far from being widely applicable in practice.

Questions

* The problem setup of CLASH stops the entire experiment if a stopping decision has been made. However, in clinical trials for example, it may not be ethical to stop the experiment for a majority group which benefits from the treatment. Could CLASH be adapted to an online setting where the experiment is only stopped for the harmed subgroup? It would be interesting if the authors could comment on this.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

* Prop. 3.1 assumes the group membership (and CATE) is known, which is never the case in practice. The authors state this, yet it is unclear how one would efficiently infer group membership in large-scale scenarios in practice, and whether the proposed solution in Stage 1 of the algorithm would work. It is consequently unclear how well the proposed tests would generalise if group membership is not known, or how the asymptotic behavior analysed in this Proposition would change in this case. It would be helpful if the authors could comment on this. * The experimental settings are all limited to two groups. In many real-world settings such as clinical trials, we would expect more than two groups with somewhat homogeneous treatment effect. How this method would perform in such cases is unclear, and neither discussed in the paper.

Reviewer R8bL7/10 · confidence 3/52023-07-05

Summary

The paper focuses on stopping tests for harm in clinical trials or A/B testing where heterogeneous treatment effect is involved. The proposed method contains two phases: First, the population harmed by the trials is identified via conditional treatment effect estimation. Second, weighted versions of widely-used test statistics are computed to determine if the trial needs to be stopped early. Theoretical analyses are carried out to show that the proposed method meets the desired requirements: producing high probability when a subgroup is harmed while limiting the unnecessary stopping. Experiments using simulation and real-world data show improvement over homogeneous stopping tests and the existing heterogeneous ones.

Strengths

+ The paper is well-written and easy to follow. The research problem is clearly defined and the motivation for the proposed method is nicely presented. + Theoretical analyses show that the proposed method could achieve the desired properties. The convergence property (Thm. 3.2.) is particularly important and well-explained.

Weaknesses

- In lines 191 to 193, $(x_i, y_i)$ is excluded from training set when estimating $\tau(x_i)$. Does it mean that the CATE estimation model needs to be trained for $n$ times during stage 1? Are there any comparisons of the running time between baseline models, say SUBTLE?

Questions

- In the lower part of Fig. 2, when the treatment has no effect on the majority, why CLASH has lower probabilities of stopping early than the homogeneous approach when the harmful treatment effect on the minority group is below 0.4?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The limitations of the proposed method are discussed in the last section. It would be great if the authors could add a few sentences about how they plan to address the limitations in future work.

Reviewer SD946/10 · confidence 3/52023-07-17

Summary

The authors propose an approach to early stop clinical trials in order to prevent subgroup level harms. Their approach involves first estimation of sequential estimation of a an individualized treatment effect using machine learning methods followed by reweighting the test statistic at each iteration with the estimated mean and standard deviation of the CATE. Furthermore their choices are such that they do not need to make apriori assumptions on what groups represent the minority subgroups. The authors present results around optimality of the metrics using the estimated CATE and perform extensive real world and synthetic experiments to demonstrate the effectiveness of their approach.

Strengths

* The problem of early stopping of a clinical trial to prevent aggregate and subgroup level harms is an important one. The contribution is timely and relevant. * The demonstration of the proposed method on the time-to-event (survival) setting is welcome since outcomes in most real world clinical trials are censored time-to-events. * The extensive theoretical insights and real world and synthetic experiments are thoughful and welcome.

Weaknesses

* The paper requires estimates of the uncertainity in CATE at an *individual* level. CATE is itself hard to estimate, and its uncertainity is even harder in practice to estimate. The authors propose to use "bootstrap" to compute this quantity. This seems practically impossible especially with high dimensional covariates. * While the paper is motivated strongly there is a fundamental question that remains to be addressed: The current method is such that it does not require apriori assumptions on which covariates specify the subgroups. It is unclear as to how in practice the group would be specified implicitly by the model. For instance, in the case of a large RCT there would always exist trivially small subgroups that are harmed. However these subgroups would not be generalizable and such results would not transport. Ideally this should be reflected in the uncertainity estimates upto generalization error. However this maybe violated in small sample size regimes.

Questions

* The cdf for chosen weights, $w$ corresponds to a normal distribution. I think this is because the estimated weights tend to a normal distribution by central limit theorem, but I was unable to find a result for this in the paper can the authors point to this. * Can the authors address the second weakness pointed out in more details. Specifically does the current setup allow a practioner to specify something along the lines of a "minimum" group size to prevent the model from raising false alarms with trivial subgroups? * Finally, the current setup it seems does not allow for anyway of specifying how the covariates across different subgroups are related. For a subgroup to be actionable there should be similarities in there covariates. Is there a way to enforce or obtain such similarities from the experiment. Overall I am willing to readjust my scores favorably based on answers to my questions as well as discussions and deliberations during the author rebuttal phase.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The paper does raise an interesting concern over whether current trials should be stopped for the entire population if a single subgroup is found to be harmed from the intervention. As the authors point out, this is a question of medical ethics and is largely beyond the scope of this manuscript, and hence I am inclined to not consider it as a "limitation". I think however that perhaps some examples of where such decisions can lead to different outcomes should be included in the manuscript.

Reviewer K2p17/10 · confidence 3/52023-07-18

Summary

The authors propose CLASH, a method for early stopping in RCTs and A/B tests on heterogeneous populations. They some theoretical results that show that CLASH works, and provide simulations and one real experiment.

Strengths

- Clear writing - Good motivation - Excellent exposition of theoretical results for readers who may not be able to fully understand proof details - Convincing (albeit limited) experimental evidence

Weaknesses

The biggest issue with this paper has to do with the experiments. EDIT: I have raised my score to a 7 after the authors answered my questions. - Unrealistic settings in the simulation - see the questions section. - Your real-world data experiment is nice, but it's a single experiment with a single outcome - you could be doing as well as Oracle purely by chance. You could run it multiple times by randomizing the order of the data and reporting a distribution of stopping times against oracle, SUBTLE, and homogeneous. - "Note that stopping the experiment just in one region would affect statistical inference at the end of the experiment, as the treatment would no longer be randomly assigned across regions. Practitioners can use covariate adjustment, inverse probability weighting, or adaptive sampling methods [13] to adjust for this selection." I would recommend actually performing this calculation and reporting the outcome.

Questions

- In the simulation you have "X maps deterministically to G" -> is this not unrealistically simple? What happens if you use a stochastic function with varying degrees of noise? - Similarly, "recruiting one treated and one untreated participant at each step" is unrealistic. Why not try to mimic real recruitment procedures and have batch recruitment? Does this affect your results at all? - In your simulation "Y is normally distributed" with a fixed standard deviation of 1.0. What happens if the standard deviation is larger or smaller? Is it important to know this standard deviation in practice? - You write "All performance increases are robust to an increase in the number of covariates" but also "CLASH works better with a relatively small number of covariates": these two statements conflict, and I did not see any experiments that demonstrate the latter. It's important to add experiments that quantify how your models perform (or don't) in higher dimensions. What happens when the dimensionality is 100 and 1000 in the simulation? - "We only sample from Region and Regions 5-8; this gives us one harmed group (Region 1) that comprises 28% of the total population" - why do this? Why not sample uniformly at random from the entire dataset?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes

Reviewer af9S7/10 · confidence 4/52023-07-20

Summary

This paper proposes a method to decide when to stop a clinical trial or an A/B study in cases where the treatment/intervention only causes harm to a minority group of participants, in which traditional methods can fail to detect the need for stopping an experiment. The paper includes a thorough theoretical analysis of the method, which suggests that it has desirable properties (stopping when a treatment is harmful and not stopping when it is not harmful) for large enough n. It also includes extensive simulations demonstrating that the method outperforms established methods in most cases and performs well in a real-world application.

Strengths

- The paper is well written and easy to follow. - The research question is interesting and important also from a fair AI perspective as this research protects minority groups. - The paper includes a very extensive analysis of the problem, including theory, simulations and an empirical application.

Weaknesses

- The simulations could include a wider range of sample sizes, i.e. sample sizes of 50-200 participants which are common in clinical trials.

Questions

First of all, I want to say that this paper was a pleasure to read. I was impressed by the thoroughness and extend of the analyses. I only have a few suggestions. I hope you will find them helpful and constructive. **Major points** - I would like to see how the method holds up for much smaller sample sizes in simulations and on the empirical data, since there are many clinical trials with sample sizes around 50, 100 or 200 participants. Even if the method does not work well for this, it would be an important information to have so that practitioners do not apply this method when it’s not appropriate. - On p. 3, you write “For example, consider a situation in which there are two equally sized groups with equal but opposite treatment effects, that is, p(G = 0) = p(G = 1) and τ (0) = −τ (1). The ATE is zero, and so any stopping test with H0 : ATE ≤ 0 is designed to continue to completion at least (1 − α)% of the time.” However, later you focus on Gaussian outcomes. Can your method handle the first scenario as well (i.e., bimodal distributions)? Or could it be extended to Gaussian mixtures for example? - In the simulations shown in the supplement (Figure S6), there are certain scenarios for which the homogenous baseline outperforms CLASH (e.g., panel A: Bayesian Estimation MaxSPRIT, no treatment effect on the majority and small harmful effects or panel B: MaxSprit for small effects). Can you speculate on why this is the case? I do not think it is a problem if your method is not the best in every scenario, but it is important to outline and understand the conditions under which it is outperformed by other methods. **Minor points** - How would this method be used to decide for whom a trial should be stopped early in practice? Would this be based on the probability of being harmed, does this require a threshold? Perhaps you can elaborate more on this. - Out of curiosity, could this method be extended to multiple groups that show different harm (e.g., a strongly and a weakly harmed group)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

The authors adequately discuss the limitations of their work. Depending on the performance for much smaller n in simulations it might be worthwhile expanded on the required sample size as well.

Reviewer wJ1C7/10 · confidence 4/52023-07-24

Summary

A two-stage approach CLASH was proposed to determine the early stopping time of a randomized experiment when the treatment is harmful to a subset of the population. The indicators of the harmed groups are estimated by causal machine learning methods in stage 1, and the early stopping time is determined using the weighted test statistic in stage 2. Theoretical properties of the proposed method were established. Simulation studies were conducted to show the existing homogenous method's failure and CLASH's success. Finally, an illustration of the proposed methods was presented by analyzing real data from a digital experiment. Overall, the proposed methodology is useful, and the paper is well-written.

Strengths

The method is solid, clearly described, and useful in different areas. With the accommodation of both Gaussian and time-to-event experimental outcomes and considering the limited prior work on this specific area, the proposed method can make a good impact. The results of the simulation studies and real data analysis are convincing and support their conclusions.

Weaknesses

The simulation settings seem relatively simple and difficult to interpret: only several binary covariates were included, and there was only 1 harmed group. It would be interesting to see some simulation results under settings closer to the real data analysis (multi-level categorical covariates and multiple harmed groups). Including several sentences to interpret the simulation settings would also be helpful.

Questions

See Weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The author provided a paragraph to discuss the limitations.

Reviewer tX2R2023-08-10

Response to rebuttal

My points of concern are largely resolved, my questions answered. I have also read through the other reviews and their responses. I think this paper particularly shines in tackling an interesting, understudied problem, for which no good solutions seem to exist. It is simple and straight-forward, yet solid. I vote for accepting the work.

Reviewer K2p12023-08-10

Good response

Thank you for the complete response. I will raise my score to a 7.

Reviewer wJ1C2023-08-10

Thank you for your response. I have read the other reviews and rebuttals. I will keep my original score and agree to accept this paper.

Reviewer af9S2023-08-11

All questions answered and thank you

Thank you very much for addressing my concerns and including these additional experiments. I stand by my first score highlighting that this is a strong paper. I am certain that this work will be a great contribution to the conference. I am looking forward to reading more about your work and wish you a great conference.

Reviewer R8bL2023-08-18

My concerns have been addressed

Thank the authors for the response. I have read the responses, as well as the discussion between the authors and other reviewers. I think my main concerns have been clarified/addressed. Overall, this is a solid paper, so I am happy to raise the rating to 7.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC