{"id":"ca90724b-74e4-4f80-a55a-788a293ea3f4","arxiv_id":"2511.16029","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A validified possibilistic posterior yields finite-sample confidence sets for the treatment effect whenever the user's exogeneity-violation set contains the true violation.","lead":"This paper proposes a possibility-theory method for instrumental variable regression that lets users specify a set of plausible exogeneity violations, then returns uncertainty intervals for the treatment effect with a claimed finite-sample coverage guarantee when the specified set is correct. It is a new sensitivity-analysis tool for causal inference, especially useful when only one possibly invalid instrument is available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop. 2's proof hinges on an unproved monotonicity claim: enlarging the violation set A must make the validified posterior π_w(β|A) pointwise no smaller than π_w(β|{α0}). This is not established and is not obviously true, so the central finite-sample coverage guarantee is currently unsupported.","rationale":"The reader's verdict is CONDITIONAL, and the reader's rationale explicitly mentions the unproved monotonicity in Prop. 2 as a concern. However, the reader's labeled 'weakest_assumption' focuses on the untestable choice of A and on the plug-in approximations, whereas the more fundamental issue is that the proof of the finite-sample guarantee itself has a gap. I agree with the CONDITIONAL verdict because the gap may be fixable and the simulations suggest the method can work, but the theorem as currently stated is not established. The proposed concrete test would settle whether the monotonicity assertion is true; if it fails, the theoretical guarantee is invalid and the verdict would need to be REJECT or UNVERDICTED for the theoretical claim, even if the practical method remains useful. Since the reader already called for conditions, I do not change the verdict, but I sharpen the specific condition that needs to be met.","tokens_in":13486,"tokens_out":7463,"duration_ms":67975,"concrete_test":"Run a Monte Carlo check of the monotonicity claim in the single-instrument Gaussian model of §4.1 with true parameters β=1, α0=0.5, n=100. For a fixed observed dataset w (or across many w), compute π_w(β|A) and π_w(β|{α0}) for a grid of β values using M=10^4 independent draws of W from P_β, with A={0.5} and A=[0,0.5]. If for any β and w the estimated π_w(β|A) is less than π_w(β|{α0}) by more than Monte Carlo error, the asserted monotonicity fails and Prop. 2's proof is invalid. Alternatively, attempt an analytic counterexample using a simplified scalar W distribution where f_A(W) and f_{α0}(W) have known forms, and check whether the inequality π_A ≥ π_{α0} can be reversed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim is Proposition 2: for any δ∈[0,1], sup_β P_β(π_W(β|A)≤δ)≤δ whenever A contains the true α0. The proof in Appendix A.3 is two steps: (i) for A={α0}, π_W(β|{α0}) is a probability integral transform and is stochastically no smaller than Uniform[0,1]; (ii) for any A⊇{α0}, 'the validified posterior becomes no more informative, i.e., π_w(β|A) ≥ π_w(β|{α0})', so the same bound holds. Step (ii) is asserted without proof. It is not a direct consequence of the definition of π_w: π_w(β|A) = P_β(f_A(W) ≤ f_A(w)), where f_A(W) = f(β|α∈A, W) is the normalized posterior possibility. While the unnormalized numerator g_A(W)=sup_{α∈A} f_S(α,β|W) is pointwise nondecreasing in A, the denominator also changes, and the normalized f_A(W) need not be pointwise nondecreasing. Moreover, even if f_A(W) ≥ f_{α0}(W) pointwise, the inequality π_A ≥ π_{α0} does not follow: both the test statistic and the threshold change with A. The proof therefore relies on a non-obvious monotonicity property that is neither derived nor cited. Without it, Proposition 2—and hence Corollary 1 and the finite-sample coverage guarantee highlighted in the abstract—is unproven. The numerical simulations in §4.1 show nominal coverage for correctly chosen A, but that is empirical evidence, not a substitute for the theorem. This concern is distinct from the practical misspecification issue of choosing A; even with perfect knowledge of α0, the size of the validified intervals may not be controlled if step (ii) fails.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a possibilistic approach to instrumental variable regression when instruments may be invalid. The model is a linear structural equation with a direct-effect parameter α; the analyst specifies a set A of possible α values. The authors derive a closed-form expression for the posterior possibility of the treatment effect β conditional on α∈A (Prop. 1), then apply the 'validification' transform of Martin and Liu to define a validified posterior possibility π_w(β|A). Proposition 2 claims a finite-sample coverage guarantee for the upper-level sets of π_w whenever A contains the true α; Corollary 1 states the resulting confidence sets. Simulations with one and five instruments compare coverage with TSLS, PGMM, BudgetIV, gIVBMA, and CIIV, and two real-data examples illustrate the method.","tokens_in":13961,"tokens_out":11743,"duration_ms":116274,"significance":"If the coverage theorem is correct, the paper is a valuable contribution: it provides a principled way to handle a single potentially invalid instrument, yields intervals that are valid in finite samples under a user-specified violation set, and has computational benefits because the possibilistic posterior has closed form. The authors are transparent about the need for the violation set to contain the true α and about the trade-off between validity and informativeness. The paper includes reproducible code and a thorough comparison with existing methods. However, the central finite-sample guarantee (Prop. 2) is not established by the proof as written; the decisive monotonicity step is asserted without justification. Consequently, the contribution is conditional on a rigorous proof or a weaker statement.","major_comments":[{"comment":"The proof of Prop. 2 hinges on the assertion that π_w(β|A) ≥ π_w(β|{α0}) for A⊇{α0}. This is not proven. The function f(β|α∈A,W) is a ratio of suprema over α, so both numerator and denominator change with A; pointwise monotonicity of the ratio is not immediate. Even if the ratio were pointwise increasing, the event in (4), {f_A(W) ≤ f_A(w)}, can be invariant under monotone transformations (e.g., if f_A = h(f_{α0}) with h increasing), so the inequality π_A ≥ π_{α0} does not follow. Since Corollary 1 and the finite-sample coverage claim in the abstract rest on this step, the proof is incomplete.","section":"Appendix A.3"},{"comment":"The MC approximation of π_w requires sampling from P_β, but P_β is not fully specified: the sampling distribution of W depends on α, γ2, and Σ in addition to β. In practice one must plug in estimates, and the paper does not show that the finite-sample guarantee of Prop. 2 survives this plug-in step. The χ² approximation is asymptotic and cannot inherit the finite-sample guarantee either. The simulations appear to use the true nuisance parameters for the MC samples, so the empirical coverage does not validate the practical implementation. Please clarify how P_β is constructed and whether the reported coverage applies to the feasible version.","section":"Section 3.3"}],"minor_comments":[{"comment":"The notation P_β is used without specifying the full parameter vector; the distribution of W also depends on α, γ2, and Σ. Please state the dependence explicitly or define a profile/plug-in distribution.","section":"Section 2, Eq. (4)"},{"comment":"The description of the MC approximation does not report the number of Monte Carlo samples M used. Please specify.","section":"Section 4.1"},{"comment":"Figure A.1 is helpful but is not referenced in the main text; consider adding a cross-reference.","section":"Appendix A.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central contribution depends on Proposition 2. If the monotonicity assertion cannot be justified, the finite-sample coverage guarantee is lost, and the paper becomes a heuristic method with simulations. I recommend sending the revised proof back to a referee familiar with inferential models. The authors should either prove the monotonicity property or weaken the claims to approximate/valid asymptotic inference. The manuscript is otherwise carefully written and the empirical work is publishable, but the theoretical claim is essential."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something new and useful—finite-sample sensitivity analysis for IV with possibly invalid instruments, built on possibility theory—but the theorem that carries the finite-sample guarantee is not actually proven. The gap is in the proof of Prop 2, where the authors assert that enlarging the violation set A can only make the validified posterior more diffuse, without showing it. That monotonicity claim is doing all the work, and it's not obvious from the definitions.\n\nWhat's new and good: The projection-based posterior of Prop 1 is clean and works in closed form, including the single-instrument case. The sensitivity-analysis framing—how much invalidity would overturn the conclusion—is genuinely useful. The simulations look solid: when A is correctly specified, coverage is near 95%, and the comparisons to TSLS, PGMM, and BudgetIV are fair as far as I can tell. The applications are reasonable and the code is on GitHub.\n\nThe soft spots:\n\n(i) The proof of Prop 2 in Appendix A.3 is two sentences. It relies on π_w(β|A) ≥ π_w(β|{α0}) for A ⊇ {α0}, which is asserted but not derived. Both the test statistic and the threshold change with A; the normalized posterior possibility is not obviously pointwise monotone in A. Without that, Corollary 1 and the abstract's finite-sample guarantee are unsupported. This is not cosmetic—it's the core theoretical claim.\n\n(ii) The MC and χ² implementations use plug-in estimates of the sampling distribution. The guarantee is only proven for the exactly calibrated transform under the true P_β. The paper doesn't show that estimation preserves the finite-sample coverage. Simulations suggest it works, but the theorem doesn't cover it.\n\n(iii) Coverage is conditional on A containing the true α0, which is untestable. The paper acknowledges this and frames it as sensitivity analysis; that's the right framing, so I don't hold it against them.\n\nNone of this kills the method. The idea is sound, and simulation evidence suggests the validified intervals behave as intended when A is chosen sensibly. But as it stands, the key proposition is unproven. This is fixable in principle: either prove the monotonicity step, cite an existing result if one applies, or state a weaker form of the guarantee. I'd want that addressed before relying on the finite-sample claim.\n\nThe paper is for applied micro-econometricians and the imprecise-probability/possibility-theory community. It deserves serious peer review—the idea is new and the construction is clever—but the referee should be asked to scrutinize the proof of Prop 2 carefully.","headline":"New possibilistic IV sensitivity method with a clean construction, but the finite-sample coverage guarantee in Prop 2 is not actually proven.","tokens_in":14431,"tokens_out":3150,"would_cite":true,"duration_ms":29831,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a possibilistic instrumental-variable method keeps finite-sample coverage of causal-effect intervals even when the only available instrument is invalid, as long as the analyst's violation set contains the true exogene","keywords":["instrumental variables","possibility theory","invalid instruments","exogeneity violations","finite-sample coverage","partial identification","sensitivity analysis","posterior possibility"],"falsifier":"Take a single-instrument design with true α = 0.05, set A = [−0.1, 0.1], and compute the Monte Carlo validified 95% interval over thousands of datasets at n = 50. If empirical coverage falls materially below 0.95, the finite-sample guarantee does not survive the plug-in/Monte Carlo implementation; if it stays at or above 0.95, the practical method matches the theory.","tokens_in":13380,"feed_emoji":"","tokens_out":5235,"duration_ms":57938,"temperature":0.7,"pith_summary":"The paper is trying to establish a method for instrumental-variable regression that does not require the analyst to bet on whether the instrument is valid. The core idea is to perform posterior inference on the treatment effect conditionally on a user-specified set of possible exogeneity violations, using possibility theory rather than standard probability. The key theoretical result is a finite-sample coverage guarantee: if the violation set contains the true violation, the calibrated ('validified') uncertainty intervals for the treatment effect are valid confidence sets at any level, even with a single potentially invalid instrument. A sympathetic reader would care because validity of instruments is often uncertain, and existing approaches either require many instruments or lose coverage when instruments fail. The paper also gives practical approximations and demonstrates on two real datasets that qualitative conclusions can survive plausible violations.","feed_headline":"One invalid instrument still gives valid effect bounds","feed_subtitle":"Finite-sample coverage holds even with a single invalid instrument, if the analyst's violation set contains the true violation.","key_machinery":"The central object is the conditional posterior possibility f(β | α∈A, W): a curve over treatment-effect values β obtained by maximising the structural posterior possibility over all α in the violation set A and over the covariance Σ. For each β, the best α is the projection of t(β) = γ̂1 − βγ̂2 onto A under the Z'Z metric; values of β whose implied α can lie inside A get possibility 1, forming the partial-identification plateau. The validification transform then converts this raw posterior into π_w(β|A) = P_β(f(β|α∈A,W) ≤ f(β|α∈A,W=w)), which is the probability that a resampled dataset produces a posterior no larger than the observed one. Because this is a probability integral transform, it","core_discovery":"The paper's central claim is Proposition 2: after applying the validification transform to the conditional posterior possibility, for any δ in [0,1] one has sup_β P_β(π_W(β|A) ≤ δ) ≤ δ whenever A contains the true value of the exogeneity-violation vector α. In other words, the upper level sets of the validified posterior possibility are valid 100(1−δ)% confidence sets for the treatment effect β in finite samples. The posterior itself is built by mapping the reduced-form estimates to structural parameters through a projection of t(β) = γ̂1 − βγ̂2 onto the violation set A under the metric induced by Z'Z; values of β whose implied α can stay inside A receive the highest possibility, forming a p","pith_inferences":["A direct but unstated consequence of the validification argument is that the same construction could yield simultaneous confidence bands or family-wise error control over multiple hypotheses about β, provided A is fixed before seeing the data.","The method's usefulness hinges on choosing A; a natural extension the paper leaves open is data-dependent selection of A with split-sample or Bonferroni-type corrections to preserve the finite-sample guarantee.","In multi-instrument settings, the projection-based mode corresponds to a 'least-violation' point estimate of β; this could be developed into an identification-robust test statistic, though the paper stops at interval estimation.","The paper's own simulations show the χ² approximation can undercover when the true α sits on a corner of the hypercube; a practical reading is to treat χ² as a screening device and use the Monte Carlo version for final reports."],"forward_implications":["If the violation set A contains the true violation vector, the reported 100(1−δ)% uncertainty intervals cover the true treatment effect with at least nominal frequency in finite samples, regardless of whether any instrument is actually valid.","The method yields a genuine posterior possibility function (possibly diffuse) for every violation set, so sensitivity analyses do not require a binary choice between valid and invalid instruments.","Widening A trades validity against informativeness: intervals remain valid but become more conservative, and a sufficiently large A makes inference completely uninformative.","A Monte Carlo approximation and a cheaper χ² approximation are provided; the paper argues the Monte Carlo version is preferable when β is not point-identified because the Gaussian approximation mishandles tails.","In the real-data examples, allowing plausible violation bounds leaves qualitative conclusions intact, while sufficiently large violation sets can erase significance."],"fun_headline_variants":["Invalid IV? Valid bounds still proven","One bad instrument, valid effect bounds","Finite-sample cover for invalid instruments","Possibilistic IV: valid despite violation","Invalid IV, calibrated intervals guaranteed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the user-specified violation set A actually containing the true exogeneity violation α; this is untestable from the data, and the paper's own simulations show coverage collapsing when the set is wrong.","fun_headline_variants_meta":{"raw":{"variants":["Invalid IV? Valid bounds still proven","One bad instrument, valid effect bounds","Finite-sample cover for invalid instruments","Possibilistic IV: valid despite violation","Invalid IV, calibrated intervals guaranteed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1073,"prompt_tokens":669,"completion_tokens":404,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":413,"completion_tokens_details":{"reasoning_tokens":353}},"tokens_in":413,"tokens_out":404,"duration_ms":5098,"temperature":1.0,"reasoning_tokens":353,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:12:47.807982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single-instrument design with true α = 0.05, set A = [−0.1, 0.1], and compute the Monte Carlo validified 95% interval over thousands of datasets at n = 50. If empirical coverage falls materially below 0.95, the finite-sample guarantee does not survive the plug-in/Monte Carlo implementation; if it stays at or above 0.95, the practical method matches the theory.","supporting_citations":[],"review_version":1}