acceptodds
Withdrawn from ICLR 2027

Baseline-Calibrated Adaptive Gaussian Processes for Conditional Dose-Response Estimation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Estimating conditional dose–response functions from observational data requires capturing heterogeneous treatment responses while accounting for covariate-driven outcome variation. We propose Baseline-Calibrated Adaptive Gaussian Processes (BCA-GP), a structured Gaussian process estimator that combines an unpenalized covariate baseline with an adaptive dose kernel. The method learns dose directions from baseline-residualized outcomes and retains a common baseline space throughout direction learning, hyperparameter selection, and prediction. This construction prevents additive signals within the baseline space from altering dose adaptation. We establish exact baseline reproduction and equivariance of the complete fitting procedure: adding a baseline-space function to the outcomes produces the corresponding shift in predictions while preserving response-surface estimation error. We further characterize prediction as constrained kernel approximation and derive a deterministic error bound that separates function approximation, representation error, and adaptive noise propagation. Under standard causal identification assumptions, BCA-GP directly estimates conditional intervention responses. Experiments on semi-synthetic benchmarks show improved estimation of both conditional and average dose–response functions relative to the evaluated GP and neural baselines.

Primary area:
causal reasoning
Keywords:
Conditional dose–response estimation, Gaussian processes, Causal inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rxu4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BCA-GP for conditional dose-response estimation. The main idea is to separate the covariate baseline from the dose-response part, and then learn an adaptive dose kernel using residualized outcomes. The authors prove a baseline-shift invariance property and test the method on several semi-synthetic benchmarks.

Critical strengths
  1. The motivation is clear, and I think separating baseline effects from dose-response learning is a reasonable idea.

2。 The paper is generally well written and easy to follow. The baseline invariance result is also interesting.

  1. The experimental results look quite strong. BCA-GP outperforms all the reported baselines across the five benchmark tasks.
Critical weaknesses
  1. The method relies on a fixed spline-based dose representation and a limited baseline function space. I am not sure how well it works when the true data generation process does not fit these assumptions.

  2. I have concerns about the experiments. Even the basic GP performs better than VCNet and TransTEE on several benchmarks, which is quite surprising to me. Also, the authors did not use equal-budget hyperparameter tuning, and the semi-synthetic DGPs seem quite suitable for their method. I am not fully convinced by these comparisons.

  3. The main theorem proves invariance to shifts within the chosen baseline space, but this does not really guarantee better CDRF estimation. The error bound also does not show consistency or explain why the adaptive kernel should improve estimation accuracy.

Questions for authors
  1. What happens if the true baseline function is not well represented by the chosen baseline space? Can the authors test this?

  2. Can the authors try more complex DGPs, especially with sharp or highly nonlinear dose-response functions? I think the current benchmarks are quite friendly to the proposed spline-based kernel.

  3. I am quite surprised by the performance of VCNet and TransTEE compared with the basic GP. Could the authors explain this and provide results with better hyperparameter tuning for these baselines since I have tried to read the code but cannot see it via the anonymous link?

  4. Why not compare with a standard semiparametric GP or universal kriging using the same unpenalized baseline?

  5. The dose directions and kernel selection both use the same outcomes. Could this lead to overfitting? Is there any theoretical or experimental evidence that this adaptive selection is reliable?

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

The paper considers dose response estimation using RKHS methods. Compared to the prior literature, the innovation is to partial out some basis functions of the covariates (and not treatments) without penalization. The idea is to ensure that the final dose-response function estimate should be invariant to shifts in the outcome that add a function of covariates alone. To be precise, model the conditional mean outcome as E[Yi∣Xi=x,Ai=a]=f(x)′β+g(x,a)E[Y_i|X_i=x,A_i=a]=f(x)'\beta+g(x,a), where YiY_i is the outcome, XiX_i covariates, AiA_i treatments, and f(⋅)f(\cdot) a vector of known basis functions. Under standard identifying assumptions E[Yi∣Xi=x,Ai=a]E[Y_i|X_i=x,A_i=a] is the conditional mean potential outcome. Then the average potential outcome from a counterfactual treatment, relative to a baseline treatment level, is invariant to β\beta. The idea then is to estimate the dose-response so that its estimated value relative to baseline is likewise invariant to β\beta. For example, in series estimation, one would simply use OLS to partial out f(Xi)f(X_i) from YiY_i and whatever basis functions are used in the approximation of g(Xi,Ai)g(X_i,A_i). The key difficulty is to extend this idea to RKHS methods, note however that a finite vector of features f(⋅)f(\cdot) is still used in this case.

Critical strengths

The method makes sense, results appear to be correct. The performance roundly beats the competition across the experiments, which are fairly extensive. The idea is fairly natural and easy to understand.

Critical weaknesses

The writing is very dense with lots of brief technical statements and not much intuition. Frankly, the writing suffers many of the pitfalls that AI-generated technical text typically suffers from (this is not to say that the text was necessarily written using AI).

The main idea is easy to understand but seems like a fairly minor innovation. If there were a part of the model that should not be penalized, I would have thought it should be the additive effect of the treatments rather than the covariates: I worry penalization will shrink my dose response estimates so that I understate the impact of treatment, and so when I use penalization methods for treatment effect estimation I typically include features of AiA_i and do not penalize the coefficients on those features. Ultimately, the intuition for why leaving coefficients on f(Xi)f(X_i) unpenalized is helpful was not clear to me. That said, the experiments suggest the method outcompetes competitors, but I would still like to have a better sense of why.

Questions for authors

How does the method perform compared to simply using say, OLS including as regressors AiA_i, XiX_i, and their interactions (perhaps penalizing the latter)? It is typical to see only comparison to other ML methods but I often wonder how much (if at all) these actually improve on old-school parametric approaches.

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

This work is aimed at developing a method for treatment effect estimation. It proposes BCA-GP, a Gaussian Processes (GP) based estimator for conditional dose-response estimation. This method combines "adaptive dose kernel", "unpenalized" covariate baseline to obtain conditional dose-response estimation function. Authors also perform several experiments to show efficacy of their method and provide some theoretical analysis of their method.

Critical strengths
  1. Experimental results look promising.
Critical weaknesses
  1. My major concern is writing and presentation and that is a not "minor presentation issue". I did not understand the paper much, despite I have paper(s) in this area and I probably know atleast something in this area. I think many of the jargons are not at all explained clearly. I will list a few:
  • I really did not understand the abstract. What is “baseline calibrated?” What is baseline in the context?
  • Introduction mentioned “Our target is the conditional intervention mean, identified through outcome regression under consistency, conditional exchangeability, and adequate support.” I could not understand significance, context anything of these terms in this context.
  • In introduction, I suddenly encountered profile score: “These directions define candidate kernels, which are compared using a baseline-adjusted profile score.” I could not find mention of profile score anywhere except page 18 in Appendix.
  • Is “covariate landmark” a standard term?
  • What is “surface?” T_{\ell} or \mu? Or this is just a substitute for the word “term”? There are many more. Writing needs significant attention.
  1. The experiments section is very limited. One should provide statistical significance tests for their results? One should also vary seeds.
Questions for authors

Kindly check the weaknesses.

Since the writing is poor, I suspect my assessment is unlikely to change, as rewriting requires complete revamping of the paper.

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

Related papers

Discussion (0)

Sign in to comment.