acceptodds
Under review as a conference paper at ICLR 2027

Learning Subgroup-level Causal Actions with Maximum Interventional Probabilities

Abstract

Data-driven decision support asks who should receive which intervention to maximize a predefined outcome, a question with wide applications in critical areas such as healthcare, public policy, and economics. While existing methods mostly focus on a single actionable variable, real-world applications often involve many, possibly dependent, actionable variables, of which decision makers want to intervene on only a few while the others respond. We study the problem of subgroup-level causal action design, which aims to decide which variables to intervene on, what values to set, and which subgroup of the population to apply the intervention to, so as to maximize the interventional success probability. Such partial interventions are hard to evaluate, since the response of the untouched variables depends on the causal graph among them, which observational data do not identify. We show that the graph is nevertheless not needed to find the best subgroup-level action: under a partition-based model of the outcome, the problem reduces to a standard supervised learning problem, and we give a finite-sample guarantee that controls the regret of the learned action by the prediction error of the learned model. We instantiate the approach with a decision tree learner, the Subgroup-level Causal Action Tree (SCAT), and, to validate its empirical performance, construct a semi-synthetic benchmark with multiple treatments based on ACIC 2016. Compared to policy learning methods adapted to our setting, SCAT finds the subgroup-level action with the highest interventional success probability on most datasets, estimates that probability most accurately, and is substantially faster.

Primary area:
causal reasoning
Keywords:
Causal Action Design, Causal Inference, Policy Optimization, Policy Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HYWSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the selection of a population subgroup and a subset of treatment variables to intervene on, with the aim of maximizing the subgroup’s post-intervention success probability. The proposed SCAT method learns and prunes a classification tree using the minimum description length (MDL) criterion. Covariate conditions along a path define the subgroup, while treatment conditions define the intervention. A validation set is then used to select the candidate with the highest estimated success probability. Under an approximate partition assumption and the stated causal assumptions, the authors show that a near-optimal intervention can be obtained without recovering the causal graph among treatments. They also establish a finite-sample bound that relates action regret to prediction error.

Critical strengths

The paper is well motivated and brings subgroup selection, intervention-variable selection, and the choice of treatment values into a common framework. Its explicit treatment of partial in- terventions under causal dependence among multiple treatment variables has practical relevance. The method is straightforward, using the tree’s partition structure to extract subgroup conditions and intervention rules directly from its paths. These rules are intuitive and interpretable. The theoretical results also have value. Under the stated causal assumptions and partition approxi- mation conditions, they show that near-optimal subgroup interventions can be identified from an observational conditional response model without recovering the causal graph among treatments.

Critical weaknesses

W1. The separate contributions of the tree structure and MDL lack experimental validation The authors explain their reasons for using a tree structure and MDL, but direct experimental comparisons are missing. An ordinary predictive tree also allows subgroups and interventions to be extracted from its paths. A top-1 mixture-of-experts model with X as the gate input and Lasso penalties on parameters associated with T within each expert could also support adaptive grouping and treatment-variable selection within groups. The experiments do not include these structurally related alternatives, making it difficult to assess the respective benefits of the tree structure and MDL. Comparisons under the same action-selection and evaluation procedures would help establish their empirical contributions.

W2. The link between the optimization objective and practical intervention goals is insufficiently justified The introduction motivates the work through intervention decisions in healthcare, public policy, and related applications. Such applications often focus on the improvement an in- tervention provides relative to no intervention or an existing policy. The paper instead maximizes the absolute post-intervention success probability in the selected subgroup. This measures the outcome level after intervention, but does not directly quantify the benefit attributable to the intervention itself. This distinction is particularly relevant because the method jointly selects the subgroup and intervention. For a fixed population and reference policy, maximizing post-intervention success probability can be equivalent to maximizing improvement. This equivalence generally no longer holds when the population is also selected, because baseline success probabilities differ across subgroups. A subgroup with a high baseline success probability and only a small intervention benefit may receive a higher value than one with a lower baseline but a substantial intervention benefit. The reported high success probabilities therefore do not establish that the method identifies populations that benefit substantially from intervention. The authors should explain why this objective is appropriate for the motivating applications and report success probabilities under both the reference condition and the recommended intervention for the same selected subgroup. This would help distinguish a favorable baseline from an improvement caused by the intervention.

W3. The evaluation does not clearly separate subgroup selection from response esti- mation Each method is evaluated on its own selected subgroup and intervention using success prob- ability and its estimation error. These metrics characterize the final recommendation, but do not establish the accuracy of subgroup identification. Because the evaluation targets dif- fer across methods, a lower estimation error may also reflect the selection of a population whose response is easier to estimate, rather than stronger estimation performance. I suggest adding simulations with clearly defined ground-truth treatment-effect subgroups. The eval- uation should separately compare subgroup identification accuracy and response estimation error within the same ground-truth subgroup under identical intervention conditions. This would distinguish selection performance from estimation performance. It does not require a proof of consistent recovery of the true partition. Rather, simulations with known ground truth would provide direct evidence for the validity and interpretation of the learned groups.

W4. Practical relevance and interpretability lack concrete empirical support The introduction repeatedly emphasizes practical applications, but the experiments are lim- ited to semi-synthetic and fully synthetic data. There is no application with real treatments and real outcomes, which limits the evidence for the method’s practical value. The paper also does not present specific learned subgroup rules and intervention recommendations, leav- ing their simplicity and practical meaning unclear. The authors should show representative subgroup rules and interventions, examine whether they agree with the known mechanisms in simulations, and provide a domain-grounded example illustrating their interpretability and feasibility. Without such evidence, the claims of practical relevance and interpretability remain insufficiently supported.

Questions for authors

Why is a decision tree chosen to estimate the response function? Beyond the ease of extract- ing subgroup and intervention rules, is there theoretical or empirical evidence for its relative advantages? When the tree is deep and has many leaves, can MDL pruning and validation-set selection effectively control the number of candidates and maintain stable recommendations?

5 comments

We thank the reviewer for the encouraging and very sharp reviews which help us improve the paper.

W1. Separate contributions of the tree structure and MDL

We added both suggested alternatives, under the same action-selection and evaluation procedure as SCAT. For the predictive tree, CART (with cross-validation for pruning) replaces the MDL tree and nothing else changes. The mixture of experts follows the reviewer's description: a gate on XX with 2, 4, or 8 experts, and L1-penalized logistic experts on TT, both tuned by cross-validation.

True success probability and absolute estimation error of the selected action on the two benchmarks of the paper, average rank over the three methods in parentheses:

methodMT (13): successMT: errorACIC (30): successACIC: error
SCAT (MDL tree)0.698 (1.62)0.036 (2.00)0.975 (1.37)0.011 (1.57)
SCAT with CART0.697 (1.69)0.035 (1.77)0.969 (1.63)0.013 (1.97)
Mixture of experts0.590 (2.69)0.048 (2.23)0.775 (3.00)0.024 (2.47)

First, our main contribution is not a specific algorithm, but the reduction of the causal problem to learning a partition-based predictive model, and this reduction does not restrict which partition-based learner is used: the population result (Theorem 3.3) holds for any partition of the covariates and the treatments. SCAT is one instantiation, with which we validate the framework empirically. The comparison with CART confirms this: CART performs on par with the MDL tree on MT (SCAT is better on 7 datasets and worse on 6, p = 0.79), and the MDL tree is slightly ahead on ACIC (better on 17, tied on 4, worse on 9, p = 0.024). We chose the MDL tree for a practical reason: it has no hyperparameter and is deterministic, while the action selected with CART changes with the seed of the cross-validation folds on 5 of the 8 MT and ACIC datasets we checked (10 seeds on the same data split). Which partition learner suits which data is left as future work.

Second, the partition structure matters. The mixture of experts is clearly worse: SCAT has the higher success probability on 11 of the 13 MT datasets (p = 0.003) and on all 30 ACIC datasets (p = 2e-9). We think the reason is that each expert is additive in the treatments, so it cannot represent that a combination of treatment values is good while each value alone is not, which a tree on (X,T)(X,T) can. Further, the subgroups of a mixture of experts are regions of a linear gate and not rules, so they are also harder to read.

So the benefit comes from the partition structure and much less from the specific criterion, which fits our claim that the contribution is the reduction to a partition-based predictive model.

W2. The objective: absolute success probability versus improvement

We thank the reviewer for this point. We agree that the improvement over a reference is an important objective in many applications, and that a high success probability alone does not show how much of it is due to the action. Following the suggestion, we made two additions to the paper.

  1. We report, for the selected subgroup, the success probability under the reference and under the recommended action (Appendix G.7).
  2. If the goal is to optimize the improvement, our framework still applies, but it needs a different setup: the partition-based model must describe the improvement and not the success probability. We give preliminary theoretical results for this in the new Appendix A.2.

Both probabilities for the selected subgroup. The reference is the success probability of the same subgroup under the treatments its individuals actually received:

familydatasetsreferencerecommended actiondifference
MT (paper)130.6320.698+0.066
ACIC 2016 (paper)300.8180.975+0.156
MTAC200.6110.667+0.057
MTH200.4540.510+0.056
MTHAC200.4280.491+0.063
MTD (n = 20000)200.4740.773+0.299
MTDS200.4340.583+0.149
MTDAC200.4270.548+0.121

The difference is positive in every family and ranges from 0.06 to 0.30. So the reviewer is right that both cases occur: sometimes the selected subgroup has a favorable baseline, and sometimes the action carries most of the gain. This is expected, since SCAT maximizes the success probability and not the improvement.

Optimizing the improvement. For this goal, the partition-based model must approximate $P(Y\mid X,T) - P(Y\mid X)$ instead of P(Y∣X,T)P(Y\mid X,T), i.e., the learner models the difference directly. We show in Appendix A.2 (Proposition A.3) that the population result carries over under this assumption: the best cell of such a model gives the action with the largest improvement, up to the approximation error, and again without the causal graph among the treatments. The finite-sample result and the experiments are left as future work.

Why the paper studies the absolute success probability. We see the two objectives as complementary, as they answer different questions.

  1. The absolute success probability answers a question of its own. A clinician may ask whether there is a group of patients who, under a certain treatment plan, reach a very high survival probability (or, with the outcome reversed, remain at very high risk). The answer to this question is a level and not a difference, as it tells where a consistent plan is enough and where it is not.
  2. An improvement is relative to a reference, and it should be read together with it. In observational data the individuals already receive some existing treatment practice, so "no intervention" means keeping this practice and not the absence of treatment. A small difference from that practice then does not necessarily mean that the recommended action is unimportant, since it can also mean that the existing practice in that subgroup is already close to the recommended action.

W3. Separating subgroup selection from response estimation

Estimation error within the same subgroup. In our evaluation this error is identical for all methods by construction. A method only proposes candidate subgroup-level actions, and the estimate of any candidate is computed by one procedure that is shared by all methods: the Laplace estimate on the held-out instances that are in the subgroup and whose observed treatments match the action. For a fixed ground-truth subgroup and intervention, every method would therefore report the same estimate and the same error.

Hence the differences in estimation error between the methods come only from which subgroup-level action is selected, and never from a better or worse estimator. An action that is supported by more held-out instances has a more reliable estimate, and we regard selecting such an action as a merit of the method. We state this explicitly in the experimental setup.

Identification accuracy. We added a family with a known ground truth as suggested (MTD, 20 datasets, n = 20000): a subgroup defined by three covariate conditions is planted, in which two treatments set jointly are the best action. For the action selected by each method, we check whether its subgroup uses the planted covariates and whether its intervention contains the planted treatment values.

Share of runs (5 splits per dataset), average rank over the datasets in parentheses:

methodsubgroup and action recoveredaction recoveredtrue success probability
SCAT0.47 (2.20)0.55 (2.88)0.773
PT-Exact0.00 (4.08)0.32 (3.65)0.586
PT-Hybrid0.04 (3.65)0.33 (3.52)0.591
MB-IAMB0.04 (3.65)0.34 (3.45)0.594
MB-Lasso0.03 (3.78)0.26 (4.05)0.562
MB-DRScreen0.04 (3.65)0.34 (3.45)0.597

SCAT recovers both the subgroup and the action on more datasets than every competitor (better on 11 to 12 of the 20 datasets and worse on at most 1, p <= 0.002).

Reviewer_YMkJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces the Subgroup Causal Action Tree, a causality-inspired model for estimating the effect of interventions. The method partitions the population into subgroups based on observed covariates and recommends a partial intervention for each subgroup. The theoretical analysis shows that, given a sufficiently accurate predictor of the effect strength, the best action selected by this predictor approximates the true best action; a finite-sample bound is also provided. The proposed method, a minimum description length (MDL) decision tree variant, matches or exceeds current state-of-the-art performance while being 40–300x faster than competing approaches.

Critical strengths
  • The paper is well written, clearly presented, and easy to follow.
  • The theoretical results show that estimating the full structural causal model (SCM) is not necessary to identify the best action.
  • The method achieves strong performance while requiring substantially fewer computational resources than competing methods. This makes it practical for small labs and clinical settings with limited compute.
Critical weaknesses
  • The empirical evaluation is limited to a single dataset, and only semi-synthetic data is used. Evaluation on real-world data would considerably strengthen the paper’s claims of practical applicability.
  • The comparison includes only state-of-the-art subgroup-level treatment rule models. It remains unclear how the method compares to other state-of-the-art approaches for treatment effect estimation more broadly.
  • There is no ablation of the model such as the MDL, the growing and pruning, or the Laplace estimates.
Questions for authors
  • Can the authors provide results on additional datasets, ideally including real-world data?
  • How does the model compare to state-of-the-art methods beyond subgroup-level treatment rule models?
  • Which design choice of your model had the greatest impact on performance?
3 comments

We thank the reviewer for the careful reading and the encouraing and constructive review. Our replies are as follows:

W1 / Q1. Additional datasets and real-world data

Datasets. The submitted paper evaluates on 43 datasets: 13 with multiple dependent treatments (MT) and 30 from ACIC 2016. The revision adds six further semi-synthetic families with 20 datasets each (120 datasets), and three real datasets, which we discuss next.

Real-world data. We agree that results on real data would strengthen the paper. However, to our knowledge no real dataset can evaluate our setting, for the following reason. Our target is $P(Y=1\mid X\in S, do(T_A=t_A))$, in which only a subset TAT_A of the treatments is set and the untouched treatments are left free to respond. Scoring a method requires the true value of this quantity. The observational distribution alone does not determine it, so one needs either the structural causal model that generated the data, which is only known in a simulation, or an experiment that intervenes on some treatments while the others respond naturally. Available real data offer neither:

  1. trials with a single treatment have no untouched treatments;
  2. factorial trials randomize the treatments independently, so there is no dependence among the treatments, which is exactly the difficulty our paper addresses;
  3. observational data with several dependent treatments, e.g., health records, come without ground truth.

This is why our benchmark is semi-synthetic (real covariates and a simulated response), as in ACIC 2016.

We nevertheless added three widely used real datasets, although none of them fits our setting:

datasetntreatmentswhy it does not fit our settingground truth
IST19285aspirin, heparinrandomized independently, hence no dependence among the treatmentstrial estimate
ACTG 1752139ZDV, ddI, ddCa single treatment with four arms; only 4 of the 8 combinations existtrial estimate
Twins11984one (heavier twin)a single treatment with simulated assignmentexact (both outcomes are observed)

Here "trial estimate" is the mean outcome of the test units in the selected subgroup whose allocated arm agrees with the recommended action.

True success probability of the selected subgroup-level action, mean over five splits, with the average rank over the splits in parentheses. CART denotes our pipeline with the MDL tree replaced by CART (cross-validation for pruning). The runs of ListDTR on IST and Twins have not finished yet; if we get the results before the deadline, we will add them.

methodISTTwinsACTG 175
SCAT0.828 (1.5)0.971 (1.2)0.646 (4.0)
CART0.817 (1.5)0.957 (2.2)0.643 (4.2)
PT-Exact0.482 (3.9)0.835 (4.4)0.736 (1.6)
PT-Hybrid0.570 (3.1)0.869 (3.2)0.706 (2.1)
ListDTRnot finishednot finished0.694 (3.1)
EconML-PTn/a0.850 (4.0)n/a

Absolute estimation error, same layout (lower is better):

methodISTTwinsACTG 175
SCAT0.038 (2.3)0.004 (1.8)0.079 (2.8)
CART0.035 (2.1)0.005 (2.0)0.094 (3.6)
PT-Exact0.121 (2.9)0.018 (3.6)0.038 (2.6)
PT-Hybrid0.083 (2.7)0.032 (3.8)0.075 (2.2)
ListDTRnot finishednot finished0.094 (3.8)
EconML-PTn/a0.018 (3.8)n/a

So, outside the setting it is designed for, SCAT is the best method on two of the three datasets:

  1. On IST, SCAT is ahead of PT-Exact and PT-Hybrid on all five splits (by 0.35 and 0.26).
  2. On Twins, SCAT is ahead of every competitor (by 0.10 to 0.14), and it selects the same one-condition rule on most splits, e.g., IF gestat10 >= 4 THEN do(T=1) with coverage 0.44.
  3. On ACTG 175, PT-Exact is ahead of SCAT on all five splits (0.736 versus 0.646). We think this is expected, since a randomized trial with a single multi-arm treatment is the setting the doubly robust competitors are designed for: the propensities are known and there is no confounding, so their effect scores are unbiased. SCAT does not use this information, as it targets observational data with dependent treatments, where such scores are not available.

With five splits, the smallest possible two-sided p-value is 0.0625, which is reached on IST (SCAT versus PT-Exact and PT-Hybrid) and on ACTG 175 (PT-Exact and PT-Hybrid versus SCAT).

W2 / Q2. Comparison with treatment-effect estimation more broadly

We thank the reviewer for this question; we agree that the relation to treatment-effect estimation should be clearer in the paper. We think, however, that a direct comparison is not possible here, for three reasons, which we now make explicit in the related work.

  1. The task is different. Treatment-effect estimators return, for each individual, the effect of a fully specified treatment. Our task is to return a subgroup together with a partial intervention that maximizes the success probability of the subgroup, and an effect estimate is neither of the two.
  2. They do not apply to partial interventions on dependent treatments. To score do(TA=tA)do(T_A=t_A) with such an estimator, one has to supply values for the untouched treatments. Plugging in their observed values ignores that they respond to the intervention, and hence the result is biased unless the causal graph among the treatments is known, which is the gap our paper addresses.
  3. These estimators are already inside our competitors. Policy learning methods proceed in two steps: they first estimate the outcome under each treatment value (by regression, inverse propensity weighting, or doubly robust scores), and then choose a rule (Section 2). Our competitors are the second step applied to state-of-the-art first steps: PolicyTree and EconML use doubly robust scores, and the treatments given to every competitor are screened with a causal forest (Section 5). In other words, they are the versions of these estimators that produce the same kind of output as SCAT, which is what makes the comparison possible.

If the reviewer has a specific estimator in mind that can be turned into a subgroup-level action under partial interventions, we would be glad to include it in the final version.

W3 / Q3. Ablation, and the design choice with the greatest impact

We added the ablation, in which one design choice is changed at a time and everything else is as in SCAT.

13 MT datasets (average rank over the variants in parentheses, p against SCAT):

varianttrue success probabilitypestimation errorpfit time (s)
SCAT0.698 (4.42)0.036 (4.35)37
MDL criterion replaced by CART0.697 (4.77)0.790.035 (4.15)1.0014
growing and pruning replaced by early stopping0.694 (5.19)0.250.035 (4.38)1.001
exact pruning replaced by greedy pruning0.698 (4.42)same trees0.036 (4.15)0.7532
Laplace constant 1 to 0 (no smoothing)0.698 (4.42)same actions0.039 (5.62)0.0437
Laplace constant 1 to 0.50.698 (4.42)same actions0.037 (5.23)0.0637
Laplace constant 1 to 20.698 (4.65)1.000.034 (3.69)0.3437

30 ACIC datasets:

varianttrue success probabilitypestimation errorp
SCAT0.975 (2.68)0.011 (2.63)
MDL criterion replaced by CART0.969 (3.35)0.0240.013 (3.47)0.04
Laplace constant 1 to 0 (no smoothing)0.973 (3.17)0.110.010 (2.80)0.56
Laplace constant 1 to 0.50.974 (2.65)0.780.010 (2.50)0.18
Laplace constant 1 to 20.973 (3.15)0.0140.014 (3.60)0.013

What each design choice contributes:

  1. MDL criterion. It is on par with CART on MT and slightly ahead on ACIC. Its practical benefit is that it is deterministic and has no hyperparameter to tune, so the selected action does not depend on the seed of a cross-validation, which it does for CART on 5 of the 8 datasets we checked.
  2. Growing and pruning. Pruning is never worse than early stopping and is ahead on some datasets (Fork by 0.022, MT06 by 0.028), and greedy pruning returns the same trees as exact pruning.
  3. Laplace estimates. This is a minor choice. It does not change which action is selected on MT, and it slightly lowers the estimation error (0.036 versus 0.039 without smoothing, p = 0.04).

Answer to Q3. The choice with the greatest impact is the partition structure itself. To test it, we replaced the tree by a top-1 mixture of experts, an alternative suggested by Reviewer HYWS (a gate on XX with 2, 4, or 8 experts, and L1-penalized logistic experts on TT, both tuned by cross-validation), under the same action-selection and evaluation procedure (Appendix G.3). This lowers the true success probability from 0.698 to 0.590 on MT (p = 0.003) and from 0.975 to 0.775 on ACIC (p = 2e-9). This is in line with what the paper claims: the contribution is the reduction to a partition-based model, and a standard instantiation of it is sufficient.

I thank the reviewers for their detailed response. My concerns have be addressed. I maintain my already positive score.

Reviewer_ErDHOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper discusses how to choose a subgroup, which treatment variables to change for intervention, and their values to maximize the probability of a successful outcome. It proposes SCAT, which learns a decision tree from observational data and obtains the subgroup and intervention from a selected leaf. Under certain conditions and assumptions, this approach avoids learning the causal relationships among treatments.

Critical strengths

a) The paper addresses a practically relevant problem.

b) The theoretical analysis provides a useful contribution under the stated conditions.

c) The experiments seem well directed and are promising.

Critical weaknesses

a) The presentation makes the central idea unnecessarily difficult to understand. The introduction explains why evaluating partial interventions can require the causal graph, but then states that a predictive tree can identify the best action without clearly explaining what makes this possible. Technical terms and guarantees appear before the underlying intuition. Section 4 mixes detailed coding costs before explaining the complete action-selection procedure. Overall, I found the paper written in a very convoluted way making it a very frustrating read.

b) Some claims in the introduction are not sufficiently supported. The paper motivates partial interventions through changing only a few treatments and addressing sparsely observed treatment configurations. However, the experiments neither evaluate treatment budgets nor report how many treatments are changed. Statistical guarantee also still requires overlap across all complete treatment configurations (Theorem 3.4). The authors should carefully distinguish what partial interventions are allowed from what their method actually support.

c) The novelty needs clearer positioning. Learning Subgroups with Maximum Treatment Effects Without Causal Heuristics (Yang et al., 2026) already reduces causal subgroup discovery to supervised partition learning. While the method extends this to multiple dependent treatments, the paper should clearly explain what is added in the formulation and theoretical analysis compared with this closest paper, rather than simply emphasizing its single-treatment setting.

d) The comparisons need further clarification. I believe a standard predictive-tree baseline should be considered. Additionally, Appendix C.1 states that competitor candidates with fewer than 20 matching validation instances are excluded, while SCAT has no such restriction. Could this difference contribute to SCAT’s reported gains?

Questions for authors

Please refer the weaknesses.

4 comments

We thank the reviewer for the careful reading and the constructive feedback!

W(a). Presentation

We agree with all three points, and all of them are fixed in the revision.

  1. What makes the result possible. We now give the intuition in the introduction, right after the key finding: the success probability of any partial intervention is a weighted average of the observational response P(Y∣X,T)P(Y\mid X,T), and the unknown graph only sets the weights; an average can never exceed its largest term, and the action read off the best leaf attains it whatever the weights are.
  2. Intuition before the technical terms. Section 3 now opens with this intuition, and each assumption and theorem is preceded by one sentence that says in words what it means, e.g., "within a cell the success probability is nearly constant" for the approximate partition, "the best cell gives the best action" for the population result, and "a tree that predicts well gives a near-best action" for the finite-sample result.
  3. Section 4. It now starts with the complete procedure in four steps: fit a tree on the covariates and the treatments jointly, take each leaf as a candidate subgroup and action, rank the candidates on held-out data, and return the best one. The coding costs come afterwards, with the details moved to the appendix.

W(b). Number of treatments changed, budgets, and overlap

We first clarify what "partial" is meant to say, and then answer the three specific points.

Allowed versus supported. Our formulation allows an intervention on any subset of the treatments. SCAT returns the interventions that appear on a path of the tree, and hence how many treatments are set is learned from data and not fixed in advance. Further, by "partial" we mean flexibility and feasibility and not a cost per treatment: partial interventions include the full ones, and with many treatments a full intervention cannot be evaluated from data anyway. The word "cheaper" in the introduction suggests the cost reading, which is a bit misleading, and we rephrase it.

How many treatments are changed. The actions of SCAT are short. On the 13 MT datasets (20 treatments each, of which 3 or 4 are direct causes of YY), SCAT sets 1.15 treatments on average, and a single treatment in 89% of the runs, against 2.1 to 3.0 for the competitors:

methodmean number of treatments setshare of runs setting 1 / 2 / 3 / 4
SCAT1.150.89 / 0.06 / 0.05 / 0.00
PT-Exact2.380.18 / 0.35 / 0.35 / 0.11
PT-Hybrid2.310.18 / 0.38 / 0.37 / 0.06
MB-IAMB2.280.18 / 0.42 / 0.34 / 0.06
MB-Lasso2.110.23 / 0.43 / 0.34 / 0.00
MB-DRScreen2.260.17 / 0.43 / 0.37 / 0.03
ListDTR2.780.06 / 0.29 / 0.45 / 0.20
FlowOPT-DR3.000.00 / 0.15 / 0.69 / 0.15

Budgets. Finding the best action under a given budget is not the goal of the paper, but the fitted models allow this evaluation directly, so we report it as the reviewer suggests. We restrict the candidates of every method to actions on at most BB treatments; SCAT stays ahead at every budget (true success probability, 13 MT datasets):

methodB = 1B = 2B = 3B = 4
SCAT0.6700.6880.6980.698
PT-Exact0.5400.5800.6050.611
PT-Hybrid0.5640.5950.6000.606
MB-IAMB0.5570.5900.5950.603
MB-Lasso0.5560.5820.5950.595
MB-DRScreen0.5540.5890.5980.600
ListDTR0.4880.5810.6360.643
FlowOPT-DR0.5000.5610.6400.650

Overlap in Theorem 3.4. We think there is no conflict between the introduction and the theorem, but one sentence in the introduction made it look like one.

  1. Our main result, that the optimal subgroup-level action is found without the graph among the treatments (Theorem 3.3), only needs positivity, i.e., that every treatment configuration is possible. This is the standard requirement for learning from observational data. The overlap constant is used only for the finite-sample rate, and it is a sufficient condition and not a necessary one.
  2. "Possible" is different from "observed". The assumption does not ask the complete configurations to appear in the sample, and SCAT never uses them. On MT, 95% of the units have a complete configuration that no other unit shares, so no method can estimate the effect of a complete configuration from matching units, while the action selected by SCAT is estimated from a median of 63 held-out units that match the one or two treatments it sets. This is what the introduction means by sparsely observed configurations.
  3. The introduction says that for complete configurations "positivity fails", which is the wrong word for this. We rephrase it as "most complete configurations are not observed in a finite sample".

W(c). Novelty relative to Yang et al. (2026)

We agree that the difference should be stated beyond "a single treatment", and we add the following comparison to the paper.

Yang et al. (2026)this paper
tasksubgroup discovery for one given binary treatmentaction design: which subgroup, which treatments, and which values
objectivea contrast, P(Y=1∣do(T=1),X∈Q)−P(Y=1∣do(T=0),X∈Q)P(Y=1\mid do(T=1), X\in Q) - P(Y=1\mid do(T=0), X\in Q)a level, P(Y=1∣X∈S,do(TA=tA))P(Y=1\mid X\in S, do(T_A=t_A))
outputa subgroupa subgroup, a subset of the treatments, and their values
treatmentsone, fixed in advancemany and causally dependent; the subset to set is searched, and the others respond
identificationstandard: without hidden confounding the dodo can be droppedthe value of a partial intervention is not identified without the graph among the treatments, yet the optimum is
partitionexact, on the covariates onlyγ\gamma-approximate, on the covariates and the treatments jointly
population resultunder the exact partition, the best cell has the largest effectno action beats the best cell by more than γ\gamma, and the action read off the best cell is within 2γ2\gamma
finite-sample resultnoneregret bounded by the prediction error of the learned tree

What is added, in the formulation and in the theory:

  1. Formulation. The action itself is optimized, i.e., which treatments to set and to which values, and not only the subgroup for a given treatment.
  2. Identification. With one treatment, dropping the dodo is the standard step. With several dependent treatments, the value of a partial intervention is not identified without the graph, yet the optimal one is. This has no counterpart in Yang et al.
  3. Partition on (X,T)(X,T). The partition is on the covariates and the treatments jointly, and we prove that the subgroup and the action can be read off a cell.
  4. Approximate partitions. We allow an approximation error γ\gamma, with an explicit loss of γ\gamma and 2γ2\gamma, while Yang et al. require exactness and name this as a limitation.
  5. Finite-sample guarantee. We give a regret bound with explicit positivity and overlap assumptions; Yang et al. have no finite-sample result.

As for the relation between the two papers, both reduce a causal problem to supervised partition learning. Their partition model is the single-treatment, exact instance of our assumption. The objectives differ.

W(d). Predictive-tree baseline and the 20-instance floor

Predictive-tree baseline. We added it: CART (with cross-validation for pruning) replaces the MDL tree, and the action-selection and evaluation procedure is unchanged.

True success probability of the selected action on the two benchmarks of the paper, average rank over the two methods in parentheses:

benchmarkdatasetsSCAT (MDL tree)SCAT with CARTSCAT better / tie / worsep
MT130.698 (1.46)0.697 (1.54)7 / 0 / 60.79
ACIC 2016300.975 (1.37)0.969 (1.63)17 / 4 / 90.024

CART performs on par with the MDL tree on MT, and the MDL tree is slightly ahead on ACIC.

We want to emphasize that our main contribution is the reduction of the causal problem to learning a partition-based predictive model, and that this reduction does not restrict which partition-based learner is used: the population result (Theorem 3.3) holds for any partition of the covariates and the treatments. SCAT is one instantiation, with which we validate the framework empirically, and the result with CART confirms that a standard tree learner is sufficient as well.

We chose the MDL tree for a practical reason: it has no hyperparameter and is deterministic, while the action selected with CART changes with the seed of the cross-validation folds on 5 of the 8 MT and ACIC datasets we checked (10 seeds on the same data split).

The 20-instance floor. The floor does not explain the gains of SCAT, for two reasons: the same floor is applied to every method, SCAT included, and the results of SCAT are nearly the same without it (table below).

On the first point, the reviewer read Appendix C.1 correctly, and the sentence there was a mistake in our description: it said that SCAT needs no such floor, while in all our experiments a candidate action of any method, SCAT included, is only considered if it matches at least 20 validation instances. We have corrected Appendix C.1.

On the second point, we reran the selection of SCAT without the floor:

SCATMT, floor 20MT, no floorACIC, floor 20ACIC, no floor
true success probability0.6980.7050.9750.975
absolute estimation error0.0360.0370.0110.011
median number of matching validation instances6358136134
runs with the same selected action95% (62 of 65)99% (149 of 150)

The floor is there so that every reported estimate rests on enough data. The estimate of a candidate uses the validation instances in the subgroup whose observed treatments match the action, and for actions on several treatments this can be a handful of instances. Without the floor, the action selected by the competitors has fewer than 20 such instances in 54% to 79% of the runs on MT (SCAT: 5%), with a median of 6 to 19, and an estimate from so few instances is not reliable.

Dear Reviewer,

As the discussion period is coming to a close, we wanted to kindly follow up and ask whether our responses have addressed your concerns. We would be happy to provide any further clarification if needed.

Thank you again for your time and your thoughtful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.