acceptodds
Under review as a conference paper at ICLR 2027

Covariate Balancing for Estimating Continuous Treatment Effects Using Supremum of Integral Probability Metrics

Abstract

Estimating causal effects from observational data requires controlling for confounding. Weighting approaches seek to remove dependence between treatment and confounders by reweighting the units. We propose Covariate Balancing using Supremum of Integral Probability Metrics (CBSupIPM), a weighting method for continuous treatments. CBSupIPM learns weights by minimizing the largest discrepancy, over treatment levels, between the weighted conditional covariate distribution and the population covariate distribution. We establish uniform convergence of the weighted Nadaraya-Watson estimator of the average dose-response function (ADRF) based on CBSupIPM weights. When an ideal weight function is sufficiently smooth and the true response surface belongs to a suitable class, the estimator matches the nonparametric convergence rate of the corresponding regression estimator in an unconfounded setting, up to logarithmic factors. Also, uniform consistency is guaranteed even when the true response surface belongs to a broader class, such as uniformly Lipschitz continuous functions in the covariates. For the numerical experiments, we construct a realistic semi-synthetic dataset using the National Medical Expenditure Survey (NMES). We also use repeated simulations with synthetic data to assess method performance more reliably and examine how it changes as different data-generating factors vary. In these numerical experiments, CBSupIPM shows the most favorable empirical performance in terms of ADRF estimation error among the comparators in most settings.

Primary area:
causal reasoning
Keywords:
Continuous Treatment, Average Dose-Response Function, Covariate Balancing, Integral Probability Metric

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HDZoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies estimation of the average dose-response function (ADRF) for continuous treatments under standard assumption (no unmeasured confounder, positivity and consistency). The authors propose Covariate Balancing using Supremum of Integral Probability Metrics (CBSupIPM), which minimizes the largest discrepancy across treatment levels between the weighted conditional covariate distribution and the marginal covariate distribution:

\operatorname{IPM}_{\mathcal{F}} \left(P^W_{X\mid T=t},P_X\right).$$ The authors use a neural network to learn positive balancing weights by minimizing the worst conditional covariate imbalance across treatment levels. Kernel smoothing defines the local weighted covariate distributions, and a class of single-ReLU discriminator functions measures their discrepancy from the marginal covariate distribution. The authors derive uniform convergence rates for a weighted Nadaraya-Watson ADRF estimator. Under sufficiently strong smoothness conditions, the claimed rate is $n^{-2/5}$ up to logarithmic factors. Uniform error control also yields a bound for estimating the maximum ADRF value. Numerical evaluation includes synthetic experiments and a semi-synthetic NMES experiment. Tables 15 and 16 additionally report augmented estimators using OLS and RF outcome models.
Critical strengths

Controlling the worst conditional imbalance across treatment levels is a natural objective when estimating an entire dose-response curve. Uniform error control is also relevant when the estimated curve is subsequently optimized. Hence, the proposed method provides a valid alternative to existing methods like EBT and DCOW.

Critical weaknesses

1. The contribution is not sufficiently distinguished from closely related work.

The approach closely connects IPM-based covariate balancing for binary treatments, particularly Kong et al. (2023, ICML) and Park et al. (2025), with discrepancy balancing for general treatments (Martinet, 2020) and continuous-treatment independence weighting (Huling et al., 2024). The ReLU discriminator itself follows Park et al. (2025).

Under the paper's setting, exact conditional balance PX∣T=tW=PXP^W_{X\mid T=t}=P_X is equivalent, almost everywhere, to PX,TW=PX⊗PTP^W_{X,T}=P_X\otimes P_T. Thus, the main distinction is the conditional supremum loss and its associated uniform analysis, rather than a new population balancing target. This distinction can be meaningful, but the paper needs to explain more convincingly what is gained beyond the existing frameworks. A pointwise theorem for a competing estimator does not, by itself, show that the estimator cannot satisfy a uniform guarantee under additional conditions.

2. The scope of the theorem needs clarification.

The scope of the theorem may need to be separated from the experimental settings. Assumption B.6 requires a regular treatment density, whereas the main synthetic settings and NMES include a point mass at zero. The CDF-trimming rule does not necessarily remove this atom. The dimension-dependent smoothness requirements for the advertised rate, including α≥(dX+3)/2\alpha\ge(d_X+3)/2 and β≥2d\beta\ge 2d, also warrant discussion.

3. The implemented estimator and optimization procedure are insufficiently specified.

Equation (8) and the theoretical estimator use target-dependent weights evaluated at (Xi,t)(X_i,t), while Equation (9) uses fixed unit-level weights. Section 2.2 states that Equation (9) is used in the experiments, leaving it unclear which estimator produces the CBSupIPM results and whether Theorem 4.1 applies to that implementation.

Section 3.3 also replaces the continuum supremum by a maximum over GnG_n and describes the solver only as iterative gradient updates. Some details of these procedures should be provided.

4. The augmented results need substantially more explanation.

The manuscript does include DR/augmented comparisons in Tables 15 and 16, but these results are not integrated into the main empirical discussion, and the augmented estimator and nuisance-fitting procedure are not adequately defined. Huling et al. (2024) already develop and evaluate augmented estimation with independence weights, and Kennedy et al. (2017) provide an established continuous-treatment DR framework. The relationship of Tables 15 and 16 to these estimators should be explicit.

These tables also show that CBSupIPM's advantage is not uniform after augmentation. In the linear setting, EBCT has lower RMSE than CBSupIPM with both OLS augmentation (0.1062 versus 0.1075) and RF augmentation (0.1036 versus 0.1040). In the nonlinear RF setting, the RMSEs are 0.1136 for CBSupIPM, 0.1141 for EBCT, and 0.1145 for Equal, with reported Monte Carlo standard errors around 0.004. These are small descriptive differences, not evidence of equivalence or statistically established superiority.

Weighting-only comparisons are useful for evaluating the balancing procedure, but broader claims of practical superiority should also address these augmented results. The augmented comparisons should be extended to NMES.

Questions for authors

1. Relationship to prior work and the experimental benchmark.

Could the authors explicitly discuss Kong et al. (2023), Park (2025) and identify the precise methodological and theoretical advances beyond the existing IPM and continuous-treatment balancing frameworks? Please also explain how the NMES construction relates to Huling et al. (2024). Can an ablation hold the network, discriminator, bandwidth, and downstream estimator fixed while comparing the supremum and treatment-averaged conditional IPM objectives?

2. Theorem 4.1 and mixed discrete-continuous exposures.

Is Appendix B.6 intended first to establish Lemma B.1 under B.9? Please provide the missing step transferring the ReLU-IPM bound to the Hölder response class, including uniform control of its constants over treatment levels. Separately, is t=0t=0 included in the balancing and evaluation grids, and is the atom treated separately from positive exposures? Which guarantees apply at the atom and which apply only away from zero?

3. Interpretation and implementation of Tables 15 and 16.

What is the exact augmented estimating equation? Do these results implement the augmented DCOW estimator of Huling et al. (2024), the pseudo-outcome approach of Kennedy et al. (2017), or another construction? Please specify the OLS features, RF settings, normalization, bandwidth selection, and any sample splitting or cross-fitting. Are nuisance fits shared across weighting methods within each replication? Why are corresponding augmented NMES results not reported, and how do the authors interpret the small differences from EBCT and Equal after RF augmentation?

Moreover, as discussed in the weaknees, under DR (table 15, 16), the proposed method did not perform better than the existing methods, any insights?

4. Design of the maximum-value experiments.

What data-generating mechanism produced Figure 1 and Table 2? How does the curved ADRF relate to Equation (25), and were parameters fixed rather than redrawn? Please also clarify whether the reference maximum is defined over the full treatment support or the trimmed evaluation region.

5. The NMES benchmark requires stronger justification and robustness checks.

Like Huling et al. (2024), the authors conduct a semi-synthetic experiment using NMES, but they adopt a different outcome-generating mechanism. This modification is not inherently problematic; however, its rationale is unclear, and it remains uncertain whether the reported advantage over DCOW is robust or depends on the particular outcome construction.

Could the authors include the NMES design from Huling et al. (2024) as an additional benchmark and examine alternative outcome generators with varying smoothness and treatment-effect heterogeneity? Please also report performance separately at zero exposure and over positive doses, clarify whether repeated zero-valued evaluation points are counted multiple times in the error metrics, and provide augmented/doubly robust comparisons for NMES. These analyses would help determine whether the reported improvement reflects a broader advantage in confounding adjustment or is specific to the chosen semi-synthetic design.

Reviewer_N4dmOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper propose a covariate balancing method named CBSupIPM. The core idea is to find a balancing weight function that ensure the weighted distribution of covariate conditional on treatment is equivalent to marginal distribution. The object function is the supremum of IPM under all treatment levels. Theoretically, previous methods give pointwise consistent estimation, while this method provide uniform consistency.

Critical strengths
  1. The motivation to control the error between conditional distribution and marginal distribution of covariate uniformly in all treatment level is quite reasonable.
  2. The method proposed is clear with detailed theoretical analysis.
  3. The experimental performance of the method is very encouraging.
Critical weaknesses
  1. The presentation need to be polished. Some definitions should be made more clear, and explicitly distinguished from small propositions. There are even some duplicated sentences in the manuscript.
  2. The authors admit that the method suffers from substantial computation time. But no specific runtime comparison is provided.
  3. The convergence rate depends on the dimension of covariate and smoothness of response surface. If the dimension is high or the response surface is not smooth enough, the convergence can be slow.
  4. The optimal depth and width configuration of the network depend on β\beta, which cannot be accruately measured or estimated.
  5. The ultimate problem is a three-layer min-max optimization, which is not easy to solve.
Questions for authors

See above.

Reviewer_22SwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a new weighting method for continuous treatments. It minimizes the worst covariate imbalance across treatment levels and uses the resulting weights to estimate the dose-response curve.

Critical strengths

The idea of controlling imbalance across the treatment range makes sense. The uniform convergence result is useful for studying the whole curve and its maximum.

Critical weaknesses
  1. It is unclear how much the supremum objective adds. Under the paper’s normalization, exact balance makes the weighted joint distribution equal to P_X\otimes P_T, which is also the target of existing independence-weighting methods. The difference is mainly the localized, sup-over-t criterion. It is helpful to see a comparison that replaces the supremum with an average over t and keeps everything else fixed.
  2. The simulation description and the supplied data do not agree. Appendix C.1 describes a linear outcome model and a randomly generated neural-network outcome model, which vary across replications. In the bundled data, both designs instead use the same fixed ADRF, identical across replications and designs. Which design produced the reported tables and figures? This needs to be resolved before the simulation results can be assessed.
  3. The rate comparison needs more qualification. The stated n^{-2/5} rate is obtained under beta>=2(d_X+1), so five covariates require beta>=12 for the exact balancing score. Assumption B.2 also requires smoothness after applying psi^{-1}, which does not follow from ordinary smoothness of the response surface. These are strong conditions, and the paper should discuss when they are plausible.
  4. There is a gap between the theory and the experiments. The code uses ELU networks, a Gaussian kernel, and a fixed 5% nearest-neighbor bandwidth. The theorem uses constrained ReLU networks, a compactly supported kernel, and a bandwidth tending to zero. The treatment distributions also have a point mass at zero, contrary to the smooth-density assumption. In NMES, the outcome model allows a jump at zero, and zero is included in the supremum evaluation grid. A setting satisfying the assumptions, along with separate results at zero and positive exposure, would be helpful.
  5. The doubly robust results deserve discussion in the main text. In Tables 15 and 16, adding an outcome model makes the differences between weighting methods much smaller. A practical question is whether the proposed weights offer a meaningful improvement over simpler augmented weighting methods.
Questions for authors
  1. Does taking the supremum help compared with averaging the imbalance over t?
  2. Which simulation design was used for the reported tables and figures?
  3. Do the gains remain with outcome adjustment?
  4. How do the results differ between zero and positive exposure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.