{"id":"929377a7-9417-4c96-a805-8fe0f4ccfdac","arxiv_id":"2412.20840","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"This paper identifies and estimates the population-wide average treatment effect in a sharp regression discontinuity design using a separate auxiliary dataset, via outcome and treatment bridge functions.","lead":"Regression discontinuity designs normally reveal treatment effects only at a cutoff. This paper shows how an extra dataset containing a related variable can extend the estimate to the entire population.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The existence of the treatment bridge function (Assumption 4) is not verified and may fail for continuous U in sharp RD, which would put the simulations and real-data analysis outside the scope of Theorem 1.","rationale":"The reader's weakest-assumption analysis focused on Assumption 1 (latent confounding), which is indeed strong and untestable. However, the more load-bearing and more concrete gap is Assumption 4, the existence of the treatment bridge function. Without it, Theorem 1's identification formulae are not guaranteed to hold even if Assumption 1 is satisfied. Unlike latent confounding, Assumption 4 depends only on the observed auxiliary distribution (U,X) and can in principle be checked numerically, yet the paper provides no such check in its simulations or real-data section. In fact, the continuous-U simulation setting gives no obvious guarantee that the required integral equation is solvable, so the simulation evidence may not even fall under the theorem's assumptions. Since the central mathematical claim is conditional on Assumption 4, the appropriate verdict remains conditional: the paper should verify or relax this assumption before the method is used in practice.","tokens_in":25352,"tokens_out":22632,"duration_ms":224083,"concrete_test":"For the auxiliary data in Setting 2 and in the real-data application, estimate the treatment-bridge loss L(lambda) = min_f max_h' { E_a[f(X,W)h'(U,W) - h'(U,0) - h'(U,1)] - lambda E_a[h'^2] } for a decreasing sequence lambda -> 0 with large n2. Under Assumption 4 the population minimax value at lambda=0 is 0; if the estimated loss does not approach 0 and the resulting f is unstable, no treatment bridge function exists, so the simulation and application are outside Theorem 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1 is internally sound, but its force depends critically on Assumption 4: there exists f0 in L2(X,W) with E[f0(X,W) | U,W] = 1/p(W|U). In a sharp RD design, W is deterministic in X, so for each w this is a Fredholm integral equation of the first kind on a half-line. For w=1, it reads E[f0(X,1) | U, X>=c] = 1/P(X>=c|U). Existence is not implied by latent positivity (Assumption 3); it requires the right-hand side to lie in the range of a conditional-expectation operator, which is a strong completeness/Picard condition (Proposition 2). The paper never verifies this condition in either simulation setting or in the vitamin A application, where U=retinoic acid is continuous. In Setting 2, U~Uniform(0,1) and X|U~N(U-0.5,1), so the equation becomes integral_0^infinity f(x) phi(x-u+0.5) dx = 1/Phi(u-0.5) for u in [0,1]; solvability is not automatic and may well fail. If f0 does not exist, identification formulae (4) and (5), and the doubly robust argument, are vacuous; the finite-sample 'good performance' may then reflect regularization artifacts rather than identification of tau_w.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for identifying the global average treatment effect in a sharp regression discontinuity design when treatment assignment is deterministic in the running variable and the usual positivity assumption fails. The key idea is to introduce a latent auxiliary variable U such that the potential outcomes are independent of the running variable and treatment given U (Assumption 1), and to suppose that an auxiliary dataset containing (U, X) is available jointly with the main RD dataset containing (X, W, Y). Under Assumptions 1-4, which additionally include exchangeability, latent positivity, and existence of a treatment bridge function, Theorem 1 identifies τw = E[Y(w)] through three equivalent formulae involving an outcome bridge function h0 and a treatment bridge function f0. The paper then develops minimax estimators for the bridge functions, proposes outcome-regression, inverse-probability-weighted, and doubly robust estimators of τw, and establishes convergence rates, consistency, and asymptotic normality under high-level conditions. Simulations and an application to vitamin A supplementation and autism severity illustrate the methods.","tokens_in":25689,"tokens_out":4357,"duration_ms":44668,"significance":"If the assumptions hold, the paper provides a practically valuable way to extrapolate RD causal effects away from the cutoff using a separate auxiliary sample, a setting that previous work addressed only under stronger data availability. The identification argument (Theorem 1) is cleanly derived and represents a useful extension of proximal causal inference ideas to regression discontinuity designs. The three estimators, especially the doubly robust one, are a natural and useful contribution, and the asymptotic analysis follows the minimax-learning template. The main weakness is that the key existence assumption for the treatment bridge function is not verified in the continuous-U simulation settings or in the real data application, so the empirical evidence and the practical guidance are not yet underpinned by the identification theorem. The paper's central theoretical result is sound conditional on its assumptions, but the applicability to the reported data is not established.","major_comments":[{"comment":"Assumption 4 (existence of the treatment bridge function f0 satisfying Eq. (2)) is the load-bearing condition for Theorem 1, but it is never verified in the continuous-U settings used in the simulations or in the real data application. The only existence example in Section 7.8 is for binary U, while Setting 2 (Table 1) has U ~ Uniform(0,1) and X|U ~ N(U-0.5,1), and the vitamin A application has U = retinoic acid, which is continuous. For sharp RD with continuous U, Eq. (2) with w=1 reads E[f0(X,1) | U, X>=c] = 1/P(X>=c|U), a Fredholm integral equation of the first kind. Latent positivity (Assumption 3) does not imply solvability; Proposition 2's Picard condition (condition (3)) is not checked for any of these settings. As a result, the simulation results in Tables 4-5 and the real-data estimates in Table 6 may be produced by the regularization and choice of function spaces even when no f0 exists, and Theorem 1 would not apply. Please either verify the existence condition for the simulation data-generating processes, provide a continuous-U existence example, or restrict the empirical claims to settings where the existence is guaranteed.","section":"Section 5.2, Assumption 1"},{"comment":"The real-data conclusion in Section 5.2 depends on Assumption 1, which is asserted from biological reasoning about retinoic acid without any sensitivity analysis. The statement that serum retinol and SRS score are \"likely to be independent conditional on retinoic acid level\" is a plausibility argument, not a check. If U does not fully capture the association between X and the potential outcomes, the bridge-function equations identify quantities different from the target τw, and the reported estimates are biased. Since the main dataset has only 149 observations and the auxiliary dataset is separate, a sensitivity analysis (e.g., allowing a residual direct effect of X on Y of varying strength) would be needed to support the application; otherwise the results should be framed as conditional on an untestable assumption.","section":"Section 4.2, Theorem 6"},{"comment":"The proof of the asymptotic decomposition for the doubly robust estimator is incomplete at a load-bearing point. In Section 7.7, the claim that the empirical-process remainder terms (ˆEm−E)[(ˆf−f0)I(W=w)Y] and (ˆEa−E)[(1−ˆf I(W=w))ˆh−(1−f0I(W=w))h0] are oP(n^{-1/2}) is justified by invoking the continuity of a Gaussian process at zero, citing Krätschmer and Urusov (2023). However, the argument that the sup over an L2-ball of the empirical process converges to zero under the Donsker and covering-number conditions is only sketched; the statement \"converges to zero based on Corollary 1.2\" is a substantial step. Since the asymptotic normality result and the variance formula depend on this remainder rate, a complete proof or a precise theorem reference with the verification of its conditions should be provided.","section":"Section 4.2, Theorem 6"}],"minor_comments":[{"comment":"The treatment rule is defined as W = I(X ≥ c), but in the vitamin A example the treatment is assigned when serum retinol concentration is below 1.05 µmol/L, i.e., W = I(X < c). The convention should be stated consistently, or the example should be aligned with the definition.","section":"Section 4.1, Theorem 5"},{"comment":"In Theorem 5, the conclusion says \"we have ˆτ h w is consistent\" but it should refer to ˆτ f w; this appears to be a typographical error that could confuse readers.","section":"Section 5.1, Tables"},{"comment":"Table 6 has a column header \"Treatment Control\" that is ambiguous; the entries are estimates of τ1 and τ0, so the header should be \"Treatment (w=1)\" and \"Control (w=0)\". Also, the caption should explicitly state that the intervals are 95% confidence intervals.","section":"Section 7.2"},{"comment":"In the proof of Proposition 1, the notation h(U, X) appears where the definitions elsewhere use h(U, W); this inconsistency should be fixed to avoid confusion about the argument of the bridge function.","section":"Section 5.1, Table 1"},{"comment":"In Table 1, \"U nif orm(0, 1)\" contains a typo and should read \"Uniform(0,1)\".","section":"Section 3, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The paper's identification theorem is a solid contribution, but the empirical sections need to address the existence of the treatment bridge function in continuous-U settings and the untestable latent confounding assumption. The authors may also want to consider whether the real-data application is overinterpreted given the small sample size and the lack of sensitivity checks. The proof of Theorem 6 should be completed or made precise before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper carefully. The main claim—identification of the global ATE in sharp RD using an auxiliary (U,X) sample via bridge functions—is a genuine extension of Angrist-Rokkanen and proximal causal inference. The proof of Theorem 1 is clean, and the three estimators plus the doubly robust variant are a natural and useful translation of the classical causal inference trinity to this data structure. The real-data application is honest; they report a non-significant effect rather than overselling.\n\nThe soft spot is exactly the one the stress-test flags. Assumption 4—existence of a treatment bridge function—is the load-bearing premise. The paper states it as an assumption and gives a Proposition 2 with Picard-type sufficient conditions, but never checks those conditions in the simulations or the real data. For the continuous-U simulation (Setting 2), the equation is a Fredholm integral equation of the first kind; solvability is not automatic and may fail. If it fails, the identification formulas are vacuous, and the finite-sample performance may be regularization artifacts. This is not a minor technicality; it is the difference between estimating an identified parameter and estimating a projection artifact. The paper needs to either verify the condition (e.g., via completeness and Picard checks) or add a sensitivity analysis to show the estimates are stable when the bridge function is only approximately solvable.\n\nOther weaknesses are secondary. Assumption 1 (latent confounding) is strong and untestable; a sensitivity analysis would help, especially for the vitamin A example where it is asserted from biological knowledge. The asymptotic proofs are sketches in places, e.g., the o_P(n^{-1/2}) remainder in Theorem 6 leans on an external result and some steps are compressed. No code or data is included, so the simulation results are not reproducible, and there are no baseline comparisons to existing RD extrapolation methods.\n\nOverall, this is a serious paper with an important idea, but the existence condition is not just a loose end; it is the hinge on which the whole identification argument turns. I would send it to review, but with the expectation that a referee pushes hard on Assumption 4. The paper is worth a reading group discussion precisely because the bridge-function existence question is subtle and often glossed over.","headline":"Genuinely novel data-fusion identification for RD, but the treatment bridge function existence assumption is never verified and may fail for continuous U, making the simulations potentially vacuous.","tokens_in":26211,"tokens_out":2350,"would_cite":false,"duration_ms":23942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62G05","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A sharp regression discontinuity design can identify whole-population average treatment effects when an auxiliary variable makes the running variable a surrogate and an auxiliary sample records both.","keywords":["average treatment effect","causal inference","data fusion","extrapolation","regression discontinuity design","bridge functions","doubly robust estimation","auxiliary data"],"falsifier":"Simulate data in which $X$ has a direct effect on $Y$ that bypasses $U$ and $W$, then run the paper's estimator: the estimated $\\tau_w$ will deviate from the true $E[Y(w)]$, exposing the latent-confounding assumption. In an observed-data setting where $U$ is actually available in the main sample, one can compute the standard fully adjusted doubly robust estimate and compare it with the proposed estimator that uses only the auxiliary $(U,X)$ sample; systematic disagreement beyond sampling error would indicate Assumption 1 or Assumption 2 is violated.","tokens_in":25167,"feed_emoji":"🎯","tokens_out":9937,"duration_ms":90237,"temperature":0.7,"pith_summary":"Regression discontinuity designs normally identify causal effects only at the threshold, because treatment is a deterministic function of the running variable and positivity fails away from it. This paper claims that the whole-population average treatment effect becomes identifiable if there exists an auxiliary variable $U$ such that, given $U$, the running variable $X$ carries no further association with the potential outcomes, and if an independent auxiliary sample records $(U,X)$ alongside the usual RD data. Under latent positivity and the existence of two bridge functions, the paper proves three equivalent identification formulas for $\\tau_w = E[Y(w)]$ and gives estimators that mirror outcome regression, inverse probability weighting, and doubly robust estimation. If the claim is right, a researcher who can collect a separate sample of the running variable and its surrogate can estimate global average treatment effects from a regression discontinuity design, not just local effects at the cutoff.","feed_headline":"Auxiliary data makes regression-discontinuity ATE identifiable","feed_subtitle":"Bridge functions let researchers estimate whole-population effects from RD data plus a second sample of X and U.","key_machinery":"The load-bearing objects are the outcome bridge function $h_0(U,W)$ and the treatment bridge function $f_0(X,W)$, defined by the conditional moment equations $E[h_0(U,W) \\mid X,W] = E[Y \\mid X,W]$ and $E[f_0(X,W) \\mid U,W] = 1/p_{W\\mid U}(W \\mid U)$. The treatment bridge function plays the role of an inverse propensity score on the latent $U$ scale, and the outcome bridge function plays the role of an outcome regression on the observed $X$ scale; chaining the two equations is what turns an expectation computable in the auxiliary sample into the global $\\tau_w$. Proposition 1 recasts both defining equations as minimax population risks, and the paper estimates $h_0$ and $f_0$ by empirical minimax optimization over neural-network classes, then plugs them into three estimators: an outcome-regression analogue, an inverse-probability-weighted analogue, and a doubly robust analogue. The doubly robust estimator combines both bridge functions so that consistency survives if one of them is misspecified.","core_discovery":"The central discovery is that the non-positivity of sharp RD can be converted into an unmeasured-confounding problem and then solved with a second, independent sample of the auxiliary variable and the running variable. Theorem 1 shows that, under latent confounding ($(X,W) \\perp\\!\\!\\perp (Y(0),Y(1)) \\mid U$), exchangeability of $(U,X)$ across the main and auxiliary samples, latent positivity $0 < p_{W\\mid U}(w \\mid U) < 1$, and existence of a treatment bridge function $f_0$ satisfying $E[f_0(X,W) \\mid U,W] = 1/p_{W\\mid U}(W \\mid U)$, the average potential outcome is identified as $$\\tau_w = E[h_0(U,w)] = E[Y f_0(X,W)\\mathbb{I}(W=w)] = E[Y f_0(X,W)\\mathbb{I}(W=w) + h_0(U,w)(1 - f_0(X,W)\\mathbb{I}(W=w))],$$ where $h_0$ is the outcome bridge function solving $E[h_0(U,W) \\mid X,W] = E[Y \\mid X,W]$. The identification does not require uniqueness of the bridge functions and does not require observing $U$ in the main regression discontinuity sample.","pith_inferences":["A sensitivity analysis that reports how large a direct $X \\to Y$ effect must be to move the estimated ATE by a given amount would make the latent-confounding assumption actionable; the paper does not provide one.","Because the treatment bridge function is guaranteed only under a completeness condition and a summability condition on the singular values of a conditional-expectation operator (Proposition 2), candidates for $U$ could be screened from the auxiliary sample by estimating how rapidly those singular values decay.","The same identification logic transfers to non-positivity problems outside RD: whenever a covariate's support prevents overlap between treated and control groups, an auxiliary sample recording the covariate and the non-overlapping variable may supply the missing joint distribution."],"forward_implications":["Global average treatment effects, not just local effects at the cutoff, become identified and estimable in sharp RD whenever a suitable auxiliary variable and a separate $(U,X)$ sample are available.","The three estimators are consistent under growing sample sizes, with rates governed by localized Rademacher complexity; the doubly robust estimator is asymptotically normal with asymptotic variance split into a main-sample component and an auxiliary-sample component.","The doubly robust estimator is consistent if either the outcome bridge function or the treatment bridge function is consistently estimated, so misspecifying one of the two nuisance functions does not by itself destroy the inference.","A practitioner can take an existing RD study, collect only a new sample of the running variable and the auxiliary variable, and estimate the whole-population average effect without re-running the treatment study.","In the vitamin A and autism spectrum disorder application, the doubly robust estimate of the average SRS score is 90.87 under supplementation versus 98.02 under control, with a 95% bootstrap confidence interval for the average effect of (−12.48, 0.93), numerically favoring supplementation but not reaching statistical significance."],"supporting_citations":[{"why":"Introduced the surrogacy idea that potential outcomes are mean-independent of the running variable given covariates, which this paper adapts to an unobserved auxiliary variable.","marker":"Angrist and Rokkanen (2015)"},{"why":"Gave the noise-induced randomization framework with an unobserved latent running variable and conservative intervals, the gap the auxiliary-sample identification fills.","marker":"Eckles et al. (2020)"},{"why":"Established that sharp RD identifies only local treatment effects at the cutoff without extra assumptions, the limitation being overcome.","marker":"Hahn et al. (2001)"},{"why":"Supplies the minimax learning formulation and critical-radius analysis used to estimate bridge functions.","marker":"Kallus et al. (2022)"},{"why":"Provides the semiparametric proximal causal inference and doubly robust estimation template that the three estimators parallel.","marker":"Cui et al. (2023)"},{"why":"Defines bridge functions and their identifying conditional moment equations, which Definition 1 adopts.","marker":"Tchetgen Tchetgen et al. (2024)"}],"fun_headline_variants":["Bridge functions extend RD to average causal effect","Auxiliary data solves RD non-positivity for ATE","RD ATE identified with second sample of X and U","Auxiliary data unlocks whole-population RD effects","New bridge method identifies average causal effect in RD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 1: after conditioning on the auxiliary variable $U$, the running variable $X$ has no remaining association with the potential outcomes $Y(0),Y(1)$. If $U$ does not fully capture the link between $X$ and $Y$, the bridge functions solve different equations and the estimated average treatment effect is biased.","fun_headline_variants_meta":{"raw":{"variants":["Bridge functions extend RD to average causal effect","Auxiliary data solves RD non-positivity for ATE","RD ATE identified with second sample of X and U","Auxiliary data unlocks whole-population RD effects","New bridge method identifies average causal effect in RD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000154,"raw_usage":{"total_tokens":1256,"prompt_tokens":1035,"completion_tokens":221,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":144}},"tokens_in":651,"tokens_out":221,"duration_ms":3097,"temperature":1.0,"reasoning_tokens":144,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:08:43.110368+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data in which $X$ has a direct effect on $Y$ that bypasses $U$ and $W$, then run the paper's estimator: the estimated $\\tau_w$ will deviate from the true $E[Y(w)]$, exposing the latent-confounding assumption. In an observed-data setting where $U$ is actually available in the main sample, one can compute the standard fully adjusted doubly robust estimate and compare it with the proposed estimator that uses only the auxiliary $(U,X)$ sample; systematic disagreement beyond sampling error would indicate Assumption 1 or Assumption 2 is violated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced the surrogacy idea that potential outcomes are mean-independent of the running variable given covariates, which this paper adapts to an unobserved auxiliary variable."},{"cited_title":"Noise-Induced Randomization in Regression Discontinuity Designs","cited_arxiv_id":"2004.09458","evidence_quote":"Gave the noise-induced randomization framework with an unobserved latent running variable and conservative intervals, the gap the auxiliary-sample identification fills."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the minimax learning formulation and critical-radius analysis used to estimate bridge functions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the semiparametric proximal causal inference and doubly robust estimation template that the three estimators parallel."}],"review_version":1}