acceptodds
Under review as a conference paper at ICLR 2027

Orthogonal Learning for Conditional Geodesic Treatment Effects

Abstract

Estimating conditional average treatment effects (CATEs) is a fundamental task in causal inference, and orthogonal learners reduce sensitivity to nuisance estimation. In many scientific applications, however, outcomes are naturally represented on Riemannian manifolds, where Euclidean differences do not define geometry-respecting treatment effects. Existing work develops Fr\'echet regression for manifold-valued outcomes and geodesic causal effects, but leaves open how to perform Neyman-orthogonal pointwise estimation and inference for covariate-dependent geodesic effects from observational data. To address this, we define the conditional geodesic treatment effect through treatment-specific conditional Fr\'echet means and develop the Geodesic Orthogonal learner (GO-learner) for its estimation. We derive a joint error expansion in which errors in estimating the two means are propagated through the logarithm map, propensity and conditional regression errors enter through products, and the resulting bounds yield convergence guarantees and asymptotically valid pointwise confidence sets. Experimental results demonstrate that (i) GO-learner has lower average MSE among the three evaluated estimators with sample size increasing; (ii) the predicted bias reduces when either nuisance class is correctly specified; and (iii) inference-specific bandwidths attain near-nominal pointwise coverage at the largest sample sizes.

Primary area:
causal reasoning
Keywords:
Causal Inference, Geodesic Outcome, Treatment Effect Estimation, Geodesic Causal Inference

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_m5pFOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studeis causal effect estimation when the outcome lies on a Riemannian manifold. The target is the CGTE at a covariate value xx,

$ \tau(x) = Log_{\mu_0(x)}, \mu_1(x) \in T_{\mu_0(x)} \mathcal{M}$

where μa(x)=arg⁡min⁡mE[d2(Y(a),m)∣X=x]\mu_a(x)=\arg\min_{m}\mathbb{E}[d^2(Y(a),m)\mid X=x] is the treatment-specific conditional Fréchet mean and Log\mathrm{Log} is the Riemannian logarithm map. Reporting the effect in a tangent space makes it a Euclidean object that can be averaged, projected, and tested.

The proposed Go-learner has following good components:

  • 1, Neyman-orthogonal moment functions for the two treatment conditional Fréchet means.
  • 2, A joint linearization of the logarithm map that propagates the estimation errors of both arms into a common tangent space, with a joint error expansion showing that the first-order nuisance bias cancels.
  • 3, Four-way sample splitting to estimate the nuisance functions and the effect, plus pointwise asymptotic normality and studentized confidence sets.

The paper additional proves a negative result, separately smoothing the two arms with Fréchet regression introduces an extra bias term whenever the two arms have different local covariate distributions, the bias is removed by the joint construction.

Critical strengths

1, The proposed method is novel, Using Neyman-orthogonal pointwise esitmatoin and inference procedure for conditional geodesic treatment effects on Riemannian manifolds. The method is novel and sounds reasonabel. Since previous manifold causal work (e.g., Kurisu et al. 2024) only targets the average effect and just sketches a conditional route without Neyman-orthogonality, pointwise localization, or valid confidence sets, this paper fills a real gap and gives asymptotically normal estimates and studentized confidence sets for manifold-valued causal effects, which before were only estimable as population averages.

2, A sharp negative result with practical consequences showing that separate arm-wise Fréchet smoothing is biased when the arms differ locally is consequential. It means a natural and obvious baseline is wrong in a way that does not vanish with sample size, because the log map is evaluated at an estimated base point, so the smoothing bias is propagated instead of cancelled. This is exactly the kind of insight that changes how practitioners approach manifold-valued causal problems.

Critical weaknesses
  • no real-data experiment, and the motivating application are never exercised: the paper's stated motivation is neuroimaging, shape analysis, and other random-object outcomes. Yet the only experiment is synthetic 2×22\times2 SPD design under one metric. Given that the entire premise is that real outcomes are manifold-valued, the absence of even one real dataset (DTI/SPD covariance from an existing study, a shape dataset, etc.) is a gap.

2, "Known geometry" (a known manifold and a known metric) is the modeling decision that most matters in practice, and it is assumed rather than estimated or stress-tested. The bias-reduction benefit of orthogonality is demonstrated only in a specially constructed finite-support model, so its robustness to nuisance misspecification in realistic settings is unclear.

3, Some baselines' compariation in realted work is not provided.

Questions for authors

1, How does the GO-learner compare against Kurisu et al. (geodesic doubly-robust ATE) and the doubly-robust random-object causal estimator of Bhattacharjee et al. (continuous treatments) on your synthetic design?

Reviewer_dB6eOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

My expertise is in causal inference and orthogonal learning. I have no background in differential geometry and could not independently verify the geometric proofs.

The paper proposes an orthogonal learner for conditional geodesic treatment effects defined through treatment-specific conditional Fréchet means. It establishes convergence guarantees and pointwise inference, evaluated through simulations with positive-definite matrix outcomes.

Critical strengths

The statistical construction appears coherent and connects to standard orthogonal estimation in the Euclidean case. Jointly accounting for uncertainty in both estimated endpoints, while retaining the familiar nuisance-product error structure, is a useful methodological contribution. The simulations carefully examine nuisance misspecification and confidence-set coverage.

Critical weaknesses
  1. Scientific motivation and empirical support. The paper does not develop a concrete causal problem explaining what the proposed tangent-vector effect means, why it is the relevant target, or how it would support a scientific conclusion or decision. The evaluation is restricted to synthetic 2×2 SPD outcomes with two covariates. Although the statistical diagnostics are careful, they do not resolve this motivation gap. Related work, such as Kurisu et al.’s Geodesic Causal Inference, includes real-data applications.

  2. Accessibility. The lack of background for the differential geometry objects present in the paper will make the paper unreadable for causal-inference readers without geometry expertise. A short appendix explaining the geometric objects, their Euclidean counterparts, and a worked example would make the contribution substantially easier to understand and use.

My weak rejection is driven primarily by the first concern. I would increase my score if the authors provide convincing scientific motivation and supporting empirical evidence, together with clearer geometric preliminaries.

Questions for authors
  1. Can you give one concrete causal use case, specifying the treatment, outcome, covariates and metric, and explain what estimating the effect vector and its confidence region would teach us beyond a simpler outcome summary?

  2. Why is the empirical evaluation restricted to synthetic SPD examples? Can you provide a real-data application or a realistic simulation showing how the method helps answer a scientific question?

2 comments

Thank you for taking my comments on board and for the effort you put into the revision. I think the new experiments and clarifications have strengthened the paper, and I’m happy to raise my score to clear acceptance.

Dear Reviewer dB6e

We are very pleased that our responses have addressed your concerns. We would like to express our gratitude again for your valuable and thoughtful comments!

Best regards,

Authors of paper 16716

Reviewer_8gDaOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies estimation and point-wise inference for so-called conditional geodesic treatment effects, what does this mean? Let Y(a) be the potential outcome of treatment a \in {0,1}. Then, we define mu_a(x) = argmin_{m \in M} E[d^(Y(a),m)|X=x] (i.e. the conditional Frechet mean of the potential outcome for treatment a, given X=x). M is a Riemannian manifold and d(.) is the associated metric. Let's define tau(x) = Log_{mu_0(x)}(mu_(x)) (this is the initial velocity vector of the constant speed geodesic that joins mu_0(x) with mu_1(x) so that ||tau_(x)||_M = d(mu_0(x), mu_1(x)).

The goal of the paper is to do point-wise statistical inference for tau(x) at a fixed covariate "x" in the interior of the idealized covariate space (unit Euclidean cube).

The authors proposed "Geodesic Orthogonal Learner" (which I'll call GOL in this review) which has three steps: 1) estimates nuisance functions and preliminary geometric means, 2) locally averages an orthogonal estimating function based on the conditional mean of the "log-vector" outcome (E_[log_m(Y) | A=a, X=x)], 3) return hat{tau(x)} = Log_{hat(mu)_0(x)}(hat{mu}_1(x)).

The authors showed that this estimator has faster convergence properties than state of the art competitors and also provide support for better performance.

Critical strengths

Overall the contribution is an extension of orthogonal estimation for cond. geodesic treatment effects. As mentioned earlier, the paper proposed and developed a concrete estimator for conditional geodesic treatment effects. The estimator account jointly for nuisance estimation errors and uncertainty in both of the conditional Frechet means. They obtain convergence bounds and asymptotically valid point-wise confidence sets.

In terms of evidence of improvement relative to other results in the literature. The authors compare GOL with a natural benchmark, an arm-wise local-linear Frechet regression which estimates mu_0(x) and mu_1(x) separately using local smoothing weights and then estimates tau(x) through the plug-in Log-map. They identified an additional bias term in this baseline which is not a part of the GOL bias expansion. With this, they construct a setting in which GOL achieves a better rate and lower MSE in experiments. This provides strong evidence of improvement relative to this benchmark.

Critical weaknesses

I'm still not fully convinced of the importance of estimating something like tau(x). For example, for covariance-matrix outcomes, I can image that one could use alternatives including the entry-wise conditional mean difference, E(Y(1)|X=x) - E(Y(0)|X=x) or the conditional effect on a scalar summary such as the trace. These are different estimands, not competing estimators of tau(x). A concrete application explaining why the chosen metric and tau(x) answers a scientific question of interest would increase the impact of this paper.

Questions for authors
  1. Can you provide a concrete application where tau(x) under the chosen metric answers a scientific questions that entry-wise methods can't answer or are not adequate to address?
  2. What's the main technical obstruction that prevents your estimation and inference guarantees from following directly from existing orthogonal estimation results for constrained parameters mu_0(x) and mu_1(x) followed by the delta method applied to F(m(0), m(1)) := Lotg_{m(0)})(m(1))?
  3. Why did you use forest-weighted local-constant Frechet estimator rather than the local-linear version available in the cited literature? Would the reported improvement remain with that alternative?

Questions about mathematical correctness:

  1. Page 18, Proposition 12 proof, lines 965–968: The eigenvalues of Γa,hF\Gamma^F_{a,h} are of order hph^p. The stated uniform bounds require h−pΓa,hFh^{-p}\Gamma^F_{a,h}. Please correct the normalization; the common factor cancels in La,h\mathcal L_{a,h}.

  2. Pages 26–27, Proposition 18: The baseline comparison concerns a leading bias–variance expression. Please clarify that this does not establish a two-sided bound for the complete empirical plug-in risk or a minimax lower bound.

  3. Page 27, equation (56) and lines 1420–1421: Expanding the denominator in the exact rough contribution introduces an O(h3/2+3se)O(h^{3/2+3s_e}) term. For se<1/6s_e<1/6, this is not O(h2)O(h^2). Please clarify whether the full rough contribution is retained or include this additional remainder. The leading rate is unaffected.

  4. Pages 27 and 31, numerical implementation: The theory analyzes exact Fréchet minimizers, whereas the implementation uses a fixed 10−710^{-7} gradient tolerance and a 100-iteration limit. Please clarify how numerical optimization error is controlled relative to the statistical error.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.