{"id":"f2825b59-c581-4cc0-bd3c-5be2e070e8fc","arxiv_id":"2506.22754","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A cross-fitted, doubly robust estimator estimates continuous-treatment causal effects on metric-space-valued outcomes using Hilbert space embeddings and Fréchet means.","lead":"This paper develops a doubly robust estimator for the causal effect of a continuous treatment on outcomes that live in non-Euclidean spaces, such as age-at-death distributions. The method embeds these outcomes into a Hilbert space, estimates the Fréchet mean dose-response curve, and is tested on simulations and U.S. county mortality data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 is not supported as stated: no bandwidth or nuisance-rate conditions are given, and Eq. (7) centers the residual differently from the derived influence function, so the normal approximation lacks a verifiable proof.","rationale":"The reader's conditional verdict is appropriate, and my concern reinforces it, but I locate the risk differently from the reader's weakest_assumption. The embedding assumptions (Assumptions 2–3) are explicitly acknowledged as limitations in Section 7 and are satisfied in the distributional application, so they are domain restrictions rather than hidden flaws. The more load-bearing issue is internal to Theorem 2, which is the basis for the pointwise confidence bands in Section 6. The estimator in Eq. (7) uses gamma(t,X_i) in the residual, while the efficient influence function in Theorem 1 and Corollary 2 uses E(V_T|X,T), evaluated at T_i; reconciling these requires a bias calculation that is not shown. The stated bias term B_t appears to be missing the standard 1/2 factor in the curvature term, and the theorem states no bandwidth or nuisance-rate conditions. Because the supplement with proofs and code is not included in this version, the claimed sqrt(nh) normal approximation cannot be verified from the manuscript as posted. This is not a refutation of the methodology; adding standard rate conditions and correcting or clarifying the bias derivation would likely repair the theorem. Thus the existing conditional verdict is unchanged.","tokens_in":20769,"tokens_out":23928,"duration_ms":261083,"concrete_test":"Re-derive the exact leading bias of Eq. (7) in the scalar Hilbert case H=R with T|X ~ Uniform(0,1), V_t = γ(t,X)+ε, γ(t,X)=a t^2, using oracle γ and f and any kernel k with second moment μ_2. The leading bias of hat_theta_t,CF against a t^2 should be h^2 a μ_2 + o(h^2); compare this with Theorem 2's stated h^2 B_t = h^2 μ_2 (2a) when f'=0. If the factor of 2 persists, the centering in Theorem 2 is incorrect as stated. Separately, check the supplement for an explicit bandwidth condition such as nh→∞ with nh^5=O(1), which is absent from Assumptions 1–8; without it, h=n^{-1} is admissible and the stated CLT cannot hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 2 in Section 4. For it to hold, the cross-fitted estimator (7) must equal the influence-function estimator of Theorem 1 plus a controlled remainder, but two things block verification. First, Assumptions 1–8 contain no condition on the bandwidth h beyond an informal 'h goes to 0', and no product-rate conditions such as (nh)^{1/2} ργ ρf → 0 or (nh)^{1/2}(α_n^2+ν_n^2) → 0 are imposed. Without such conditions, the cross-fitting and kernel remainders need not vanish at the sqrt(nh) rate, so the CLT cannot follow from the stated assumptions. Second, Eq. (7) and the asymptotic representation in Theorem 2 center the residual at γ(t,X_i), whereas the efficient influence function in Theorem 1/Corollary 2 centers at E(V_T|X,T) evaluated at T_i. These are not the same object, and the implied O(h^2) bias in Theorem 2 is not derived; as written, B_t omits the standard factor 1/2 multiplying the curvature term. The supplement containing the proofs is not included in arXiv v1, so neither the missing rate conditions nor the bias algebra can be checked. If the rates are absent or B_t is off by a constant, the claimed normal approximation and the confidence bands built on it do not follow.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a doubly robust estimator for the Fréchet exposure-dose function β_t = E⊕(Y_t) when the outcome Y is metric-space-valued, the treatment T is continuous, and unconfoundedness given X is assumed. The identification strategy embeds Y into a Hilbert space through a known isometry ρ, targets E(V_t) with V_t = ρ(Y_t), and then maps back via ρ^{-1}. Three estimators are presented: IPW, a doubly robust estimator, and a cross-fitted doubly debiased estimator. The central theoretical claim is Theorem 2, which asserts a sqrt(nh)-rate normal approximation with an h^2 bias correction for the cross-fitted estimator under Assumptions 1–8. The paper also constructs confidence regions using the HulC method and applies the method to county-level PM2.5 exposure and age-at-death distributions.","tokens_in":21053,"tokens_out":11022,"duration_ms":112310,"significance":"If the main theorem and the efficient-influence-function derivations are correct, this would be the first doubly robust method for continuous-treatment causal inference with random object outcomes, and the Hilbert-embedding device is a natural and broadly applicable route for distributional data, SPD matrices, and Riemannian manifold-valued outcomes. The paper is explicit about the scope limitation to metric spaces admitting a known, convex, closed Hilbert embedding, and it acknowledges excluded cases such as phylogenetic tree space. However, the central asymptotic result is not verifiable from the manuscript as submitted: the proof is deferred to an unavailable supplement, the theorem statement omits the necessary rate conditions, and there are internal inconsistencies between the derived influence function and the estimator actually analyzed. The promise of the framework justifies a major revision rather than rejection, but the inferential claims currently rest on unverified algebra.","major_comments":[{"comment":"The statement of Theorem 2 is not supported by Assumptions 1–8 as written. No rate condition on the bandwidth is stated beyond the informal 'h → 0' in Section 3.1; in particular, there is no condition such as h → 0, nh → ∞, and nh^5 → 0 that would make the h^2 B_t bias term negligible relative to (nh)^{-1/2}. There are also no product-rate conditions linking the nuisance errors ργ, ρf, α_n, ν_n to (nh)^{1/2}, which are standard for a cross-fitted kernel-based DML estimator; without conditions such as (nh)^{1/2}(ργ + ρf) → 0 and (nh)^{1/2}(α_n^2 + ν_n^2) → 0, the claimed o_P(1) remainder cannot follow. Because the proof is in an unavailable supplement, these missing conditions cannot be checked, and the confidence bands used in Sections 5 and 6 depend directly on this theorem.","section":"Theorem 2, Section 4"},{"comment":"The estimator in (7) centers the residual at γ_l(t, X_i) in both the offset term and the kernel term, whereas the efficient influence function in Theorem 1 and Corollary 2 centers the kernel residual at E(V_T|X,T) evaluated at T_i, i.e., at γ(T_i, X_i), with the offset γ(t, X_i). The estimator in (6) uses yet another centering, writing the conditional expectation as E(V_T|X,T). These are not the same object, and replacing γ(T_i, X_i) by γ(t, X_i) or moving the GPS denominator between f(t|X_i) and f(T_i|X_i) introduces additional O(h) and O(h^2) terms that are not derived anywhere in the paper. The first-order representation in Theorem 2 therefore requires an unverified bias correction before the sqrt(nh) CLT can be accepted.","section":"Equations (6)-(7) and Corollary 2"},{"comment":"The bias term in Theorem 2 is written as B_t = (∫u^2 k(u)du)[E(d^2γ(t,X)/dt^2) + E(dγ(t,X)/dt · (d f_{T|X}(t|X)/dt)/f_{T|X}(t|X))]. For a kernel-smoothed local moment of the type considered here, the leading bias is conventionally (h^2/2) μ_2 E[γ''(t,X) + 2γ'(t,X) f'(t|X)/f(t|X)] under the centering in Corollary 2. The stated formula omits the factor 1/2 and appears to omit a factor 2 on the log-density derivative. Since the theorem subtracts h^2 B_t in the CLT, an incorrect constant or derivative combination changes the asymptotic distribution and the resulting confidence bands. No derivation of B_t is given in the main text.","section":"Theorem 2, bias term B_t"},{"comment":"Proposition 4 claims that E[hatϑ_{t,DR}] = ρ(β_t) exactly when either nuisance is correctly specified. As written, the estimator involves a bandwidth h that is later sent to zero, and the theorem itself introduces an h^2 B_t bias term. The proposition should therefore be qualified: double robustness can hold only up to the kernel bias, or in a limiting sense as h → 0. If the intended claim is exact unbiasedness for fixed h, it is incompatible with Theorem 2; if it is asymptotic, the statement should say so and the proof should identify the residual terms that vanish.","section":"Proposition 4, Section 3.2"}],"minor_comments":[{"comment":"The text says 'two estimators of the CITROCIN'; the acronym should be CTROCIN. Elsewhere, 'Brochner integral' should be 'Bochner integral', and there are several typos such as 'continouous' and 'SUTV A' in Section 2.","section":"Section 3, opening paragraph"},{"comment":"Condition 2 states that there exist c1 > 0 and c2 < 0 such that f_T(s) < c2, which is impossible for a probability density; this is likely intended to be a positive upper bound. Please correct the statement.","section":"Corollary 1"},{"comment":"The norm notation ∥·∥_{2,P_X} is defined with the integrand weighted by f_{TX}(t,x), which is not the marginal measure P_X; either the notation or the definition should be aligned. Also, the assumption introduces γ_l(t,·) on the left-hand side but the target is γ_l(t,·), and the relation between these functions is unclear.","section":"Assumption 7"},{"comment":"The caption says the data-generating mechanisms include outcome models '(A), (B), and (C)', but only settings (A) and (B) are defined in the text; the caption or the text should be corrected.","section":"Table 1"},{"comment":"Table 2 reports only confidence band widths, not empirical coverage; since the confidence bands are the main practical output of Theorem 2, coverage should be reported at the nominal level. Additionally, Figure 4's caption refers to a cyan region while the text describes a blue region; these should be harmonized.","section":"Table 2 and Figure 4"},{"comment":"The application uses L = 100 cross-fitting folds with n = 2392, giving about 24 observations per fold, while the GPS is estimated with 19 covariates. This is a practical stability concern that should be discussed, for example by also reporting results with a smaller L or repeated sample splitting.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's conditional assessment: the core idea is promising and the efficient-influence-function framework is the right strategy, but the manuscript as submitted does not yet make the central theorem verifiable. The missing rate conditions, the centering inconsistency between Eq. (7) and the influence function, and the unverified bias constant are all fixable within the manuscript's scope, so this is not a rejection. However, I would not recommend acceptance until the supplement is available and the theorem statement is corrected or the estimator is modified to match the derived influence function."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first paper I am aware of that tries doubly robust continuous-treatment causal inference for metric-space-valued outcomes, and the Dirac-measure device for deriving the influence function is genuinely clever. Second, as written, the central theorem is not backed by the stated assumptions, and the estimator in (7) appears not to match the influence function derived in Corollary 2. That makes the asymptotic normality claim unverifiable from the text.\n\nThe setup is clear and the authors know the relevant literature: Lin et al. for distributional outcomes with binary treatment, Kurisu et al. for geodesic spaces, Colangelo and Lee for continuous treatments in Euclidean settings. Embedding the metric space into a Hilbert space and estimating the Frechet mean via the embedded expectation is a sensible approach, and the PM2.5 application is a good motivating example. The HulC-based confidence intervals are a reasonable addition.\n\nNow the soft spots. Equation (7) has gamma_hat(t, X_i) inside the kernel residual, i.e., the outcome regression evaluated at the target t, rather than gamma_hat(T_i, X_i). The efficient influence function in Corollary 2 centers the residual at E(V_T | X, T), evaluated at the observed T_i. With the current centering, the second term does not have zero conditional mean even with correctly specified nuisances; it introduces a bias of order h^2 that is not the standard local constant bias. Either this is a typo, in which case it needs to be fixed and the theorem restated, or the estimator is wrong. Relatedly, Theorem 2 states a CLT under Assumptions 1-8, but those assumptions contain no rate conditions linking h, n, and the nuisance errors, such as sqrt(nh) rho_gamma -> 0. Without such conditions the remainder need not vanish. The bias term B_t also appears to be missing the factor 1/2 that multiplies the second-derivative term in kernel smoothing. All of this sits on proofs that are relegated to an unavailable supplement, and there is no code or data.\n\nThe simulations do not address the causal estimand directly: they report leave-one-out MSE for observed outcomes and confidence band widths, but no coverage and no comparison of estimated rho(beta_t) to a known truth. The abstract promises conformal inference, but the paper only gives the asymptotic bands and the HulC method.\n\nThese are load-bearing issues, not cosmetic ones. The kernel centering mistake in (7) could be a typo, and the supplementary might prove a corrected estimator, but as submitted the central claim cannot be checked. Researchers working on causal inference for object data should read it with caution. A serious referee should see it because the question is timely and the Dirac-measure derivation is novel. My recommendation: send it to review, but instruct reviewers to verify the estimator against the EIF and to demand the missing rate conditions and the supplement. If those do not check out, the paper should not be accepted in anything close to current form.","headline":"A genuinely new setup for continuous-treatment causal inference with object-valued outcomes, but the main theorem is not verifiable as written: the estimator in (7) centers the kernel residual at the wrong argument, and the assumptions lack the rate conditions the CLT needs.","tokens_in":21556,"tokens_out":7144,"would_cite":false,"duration_ms":74734,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62G08","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that continuous-treatment causal effects on distribution-valued outcomes can be estimated by a doubly robust cross-fitted estimator built on Hilbert-space embeddings.","keywords":["causal inference","continuous treatments","random objects","Fréchet regression","doubly robust estimation","Hilbert space embeddings","efficient influence functions","air pollution and mortality"],"falsifier":"In a simulation with a known Fréchet exposure-dose function, such as Gaussian distributions under the Wasserstein metric, compute $\\sqrt{nh}(\\hat\\vartheta_{t,\\mathrm{CF}} - \\rho(\\beta_t) - h^2B_t)$ over many replications and compare its empirical distribution with the normal limit in Theorem 2; coverage far from nominal at the stated rate would falsify the theorem.","tokens_in":20581,"feed_emoji":"📊","tokens_out":10726,"duration_ms":104863,"temperature":0.7,"pith_summary":"This paper establishes a doubly robust, cross-fitted estimator for the causal effect of a continuous treatment on a response that lives in a general metric space, such as a probability distribution, an image, a network, or a matrix. The target is the Fréchet exposure-dose function $\\beta_t = \\arg\\min_y E[d^2_Y(Y_t, y)]$, the metric-space analogue of an average dose-response curve. The key move is to embed the outcome space into a Hilbert space by a known isometry $\\rho$, so that $\\beta_t = \\rho^{-1}(E(\\rho(Y_t)))$, and then estimate the Hilbert-space expectation from observed outcomes at the assigned treatment level. Under unconfoundedness, positivity, and smoothness assumptions, the paper proves $\\sqrt{nh}(\\hat\\vartheta_{t,\\mathrm{CF}} - \\rho(\\beta_t) - h^2 B_t) \\to N(0, \\Sigma_t)$, yielding confidence bands for the dose-response curve. The motivating application is the effect of long-term PM2.5 exposure on county-level age-at-death distributions across the United States.","feed_headline":"Causal dose-response curves come to distribution-valued outcomes","feed_subtitle":"A doubly robust estimator with a normal limit now covers outcomes like age-at-death mortality distributions.","key_machinery":"The load-bearing identity is that a closed, convex Hilbert-space embedding commutes with Fréchet means: $E^\\oplus(Y_t) = \\rho^{-1}(E(\\rho(Y_t)))$. The paper proves the needed property that the expectation of a Hilbert-space-valued random element lies in a closed convex set, so the abstract metric-space problem becomes a Hilbert-space regression problem. Estimation is carried by the efficient influence function for $E(\\int w(s)V_s\\,ds)$; writing the observed outcome as $V_T = \\int V_s\\,d\\delta_T(s)$ separates the random function $V$ from the treatment $T$ and allows the influence function to be derived. The cross-fitted estimator $\\hat\\vartheta_{t;\\mathrm{CF}}$ mimics the limiting form of this influence function with kernel weights $K_h(T_i-t)/\\hat f_{T|X}(t|X_i)$ and an estimated outcome regression $\\hat\\gamma_l(t,X_i)$.","core_discovery":"The central claim is that, when the outcome space $(\\mathcal{Y}, d_Y)$ admits a known continuous injective isometry $\\rho: \\mathcal{Y} \\to \\mathcal{H}$ into a Hilbert space with $\\rho(\\mathcal{Y})$ closed and convex, the Fréchet exposure-dose function $\\beta_t$ can be recovered as $\\rho^{-1}(E(\\rho(Y_t)))$ and estimated at the rate $\\sqrt{nh}$ with a Gaussian limit. The estimator uses only the observed pair $(T, V_T)$ with $V_T = \\rho(Y_T)$, not the entire potential-outcome curves. It combines a kernel-weighted inverse probability term with an outcome-regression term, and the main theorem states that, under Assumptions 1 through 8, the cross-fitted estimator satisfies $\\sqrt{nh}(\\hat\\vartheta_{t,\\mathrm{CF}} - \\rho(\\beta_t) - h^2 B_t) \\xrightarrow{d} N(0,\\Sigma_t)$, with explicit bias $B_t$ and variance $\\Sigma_t$. If correct, this is the first doubly robust method for continuous treatments with random object outcomes, and it also yields finite-sample confidence regions for contrasts between treatment levels via the HulC construction.","pith_inferences":["Editorial inference: the same influence function could be adapted to estimate conditional Fréchet exposure-dose functions within subgroups, which would sharpen the heterogeneity analysis reported in Section 6.","Editorial inference: because the theorem identifies the bias term $h^2B_t$, practitioners could select an under-smoothed bandwidth so that confidence intervals are centered at the estimand rather than at the biased target.","Editorial inference: for outcome spaces that do not admit the required embedding, such as phylogenetic trees with the BHV or hyperbolic metric, an intrinsic analogue could be built from doubly robust estimates of $E[d_Y^2(Y_t,y)]$ for each $y$, as the conclusion suggests.","Editorial inference: a direct simulation study comparing empirical coverage of Theorem 2 confidence bands to nominal levels across sample sizes would tell practitioners when the asymptotic approximation becomes usable."],"forward_implications":["If only the generalized propensity score or only the outcome regression is correct, but not necessarily both, the estimator remains consistent for $\\rho(\\beta_t)$.","Confidence bands for the Fréchet exposure-dose function and for contrasts $\\Delta_{tt'}$ follow from the asymptotic normality in Theorem 2, with a finite-sample alternative supplied by the HulC construction.","The framework covers distributional outcomes under Wasserstein or Fisher–Rao metrics, SPD matrices and networks, compositional data, and Riemannian-manifold-valued outcomes, all of which admit Hilbert-space embeddings.","In the PM2.5 analysis, lower exposures are causally associated with left-shifted age-at-death distributions, and low-income counties with high Black population share show the largest apparent benefits.","For binary treatment the method connects to the existing BTROCIN literature, so the continuous-treatment theory unifies and extends that line of work."],"supporting_citations":[{"why":"Provides the isometric embedding of a negative-type metric space into an RKHS, which underlies Assumptions 2–3.","marker":"Sejdinovic et al. (2012)"},{"why":"Supplies the cross-fitting double/debiased machine-learning framework used to build the estimator in (7).","marker":"Chernozhukov et al. (2018)"},{"why":"Defines Fréchet regression, the tool used to estimate the conditional outcome regression $\\gamma(t,x)$.","marker":"Petersen & Müller (2019)"},{"why":"The binary-treatment doubly robust estimator for distributional outcomes that this paper extends to continuous treatments.","marker":"Lin et al. (2023)"},{"why":"The geodesic-space causal inference framework that CTROCIN seeks to generalize.","marker":"Kurisu et al. (2024)"},{"why":"Provides the HulC finite-sample confidence regions used for contrasts between treatment levels.","marker":"Kuchibhotla et al. (2023)"},{"why":"The PM2.5-mortality study whose subgroup analysis structures the application and heterogeneity comparison.","marker":"Josey et al. (2023)"},{"why":"Supplies the continuous-treatment generalized propensity score methodology and cardinal function adopted in simulations.","marker":"Wu et al. (2024)"},{"why":"Establishes the potential-outcome and ignorability framework behind Assumption 1.","marker":"Rubin (1974)"}],"fun_headline_variants":["Doubly robust causal estimation now handles non-Euclidean outcomes","First doubly robust method for continuous treatments on random objects","Causal effects on distributions with continuous doses, robustly estimated","Nonparametric causal inference for random objects with continuous exposure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the outcome space can be mapped by a known distance-preserving map into a Hilbert space whose image is closed and convex, so that the Fréchet mean can be pulled back from an ordinary expectation; without such a map the estimator has no defined target.","fun_headline_variants_meta":{"raw":{"variants":["Doubly robust causal estimation now handles non-Euclidean outcomes","First doubly robust method for continuous treatments on random objects","Causal effects on distributions with continuous doses, robustly estimated","Nonparametric causal inference for random objects with continuous exposure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3099,"prompt_tokens":1010,"completion_tokens":2089,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":2022}},"tokens_in":626,"tokens_out":2089,"duration_ms":15235,"temperature":1.0,"reasoning_tokens":2022,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:59:38.741363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a simulation with a known Fréchet exposure-dose function, such as Gaussian distributions under the Wasserstein metric, compute $\\sqrt{nh}(\\hat\\vartheta_{t,\\mathrm{CF}} - \\rho(\\beta_t) - h^2B_t)$ over many replications and compare its empirical distribution with the normal limit in Theorem 2; coverage far from nominal at the stated rate would falsify the theorem.","supporting_citations":[],"review_version":1}