{"id":"94067760-5952-408b-842f-3eaa4072226f","arxiv_id":"1908.08596","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Researchers can bound an adjusted slope coefficient using user-specified ranges for unmeasured confounding, via an exact finite algorithm that computes a 'confounding interval'.","lead":"This paper shows how to turn a researcher's guesses about unmeasured confounders into an interval for the adjusted regression slope, instead of just a warning label. The method computes the interval exactly with a small algorithm, and the authors apply it to a study of PBDE exposure and child IQ.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Proposition 2.2 contains an algebraic error in the Lagrange-multiplier step for the (3c)+(4) active case, so the finite set S may not contain the true extrema of βx|w.","rationale":"The reader's weakest assumption concerned the positivity of RwxRwy and boundary cases with zero lower bounds. That is a valid but localized issue about well-definedness of (4). My stress-test identifies a more fundamental problem: the proof of the main enumeration theorem contains a demonstrable algebraic error in the only case that can produce interior extrema. The incorrect derivative leads to an unjustified Rwx=Rwy condition, so the candidate set S may be incomplete. If the theorem is false, the core contribution of the paper fails; a numerical check can settle this directly without relying on extrapolation from the flawed proof. I therefore disagree with the reader's choice of weakest assumption and recommend withholding a verdict until the proof is corrected or a counterexample is ruled out.","tokens_in":14784,"tokens_out":46210,"duration_ms":368381,"concrete_test":"For ρxy=0.5, set b_rho = l_hatx_haty = u_hatx_haty = 0.5 and box bounds l2x=l2y=0.1, u2x=u2y=0.9. Using the correct Lagrange equations (with ∂g/∂Rwx = b_rho Rwy ∓ Rwx (1−Rwy²)^{1/2}(1−Rwx²)^{-1/2} and ∂g/∂Rwy = b_rho Rwx ∓ Rwy (1−Rwx²)^{1/2}(1−Rwy²)^{-1/2}), solve the system g=ρxy and H_u F_v − H_v F_u = 0 numerically for (Rwx,Rwy) in [sqrt(0.1),sqrt(0.9)]². If a solution exists with Rwx ≠ Rwy, compare βx|w at that point with the endpoints of the interval produced by the published code for the same inputs. Any discrepancy confirms that S omits the true extremum and Proposition 2.2 is false.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim of the paper is Proposition 2.2, which asserts that the min and max of βx|w over the feasible set Ω are attained on the finite set S. The proof in Appendix C attempts to characterize the case where exactly two constraints, (3c) and (4), are active. In that case, constraint (4) is written as g(Rwx,Rwy)=b_rho Rwx Rwy ± sqrt(1−Rwx²) sqrt(1−Rwy²)=ρxy, and the objective restricted to this surface is h(Rwx,Rwy)=±(σy/σx) sqrt(1−Rwy²)/sqrt(1−Rwx²). The paper then computes the ratio ∂g/∂Rwx / ∂g/∂Rwy incorrectly: the published expression contains a factor of 2 and (1+Rwy²)^{1/2}, whereas the correct derivatives are ∂g/∂Rwx = b_rho Rwy ∓ Rwx (1−Rwy²)^{1/2}(1−Rwx²)^{-1/2} and ∂g/∂Rwy = b_rho Rwx ∓ Rwy (1−Rwx²)^{1/2}(1−Rwy²)^{-1/2}. Because of this error, the conclusion that stationary points on this surface require Rwx=Rwy is unsupported. Indeed, the points in lines (7b) and (7c) do not satisfy the correct Lagrange condition except when ρxy=b_rho. The proof therefore does not establish that all interior extrema on the two-active surface are included in S. If such an interior extremum exists inside the box, the algorithm will return an interval that is too narrow, violating equation (6). This is a more direct threat to the central claim than the boundary positivity issue noted by the reader.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a 'confounding interval' for the adjusted slope βx|w in linear regression with unmeasured confounders. It uses three parameters—R²wx, R²wy, and the correlation ρhat between fitted values—and propagates user-specified bounds through an algebraic identity (Proposition 1.1) and a realizability condition (Proposition 2.1). The main algorithmic claim (Proposition 2.2) is that the endpoints of the interval are attained on a finite set S of at most 88 candidate triples, so exact enumeration is possible. A case study on PBDE exposure and IQ illustrates the method.","tokens_in":15206,"tokens_out":32600,"duration_ms":299180,"significance":"If Proposition 2.2 were correct, the paper would provide a useful, computationally exact sensitivity analysis with interpretable confounding parameters, and it would improve on earlier sign-reversal conditions. The identity in Proposition 1.1 is cleanly proved with projections, and the paper ships Python and R implementations. However, the proof of the central enumeration contains a load-bearing algebraic error, and the method as stated mishandles the zero lower-bound cases used in the application. The contribution therefore needs substantial correction before its claims can be relied upon.","major_comments":[{"comment":"The Lagrange-multiplier step for the case where (3c) and (4) are active is not correct. With g(Rwx,Rwy)=b_rho RwxRwy ± sqrt(1−Rwx²)sqrt(1−Rwy²), the printed derivative contains a spurious factor of 2 and a spurious (1+Rwy²)^{1/2}; the correct derivative is ∂g/∂Rwx = b_rho Rwy ∓ Rwx sqrt(1−Rwy²)/sqrt(1−Rwx²), and similarly for ∂g/∂Rwy. The printed ratio therefore does not imply Rwx=Rwy, and the candidates (7b) and (7c) are not the only stationary points on this surface. A concrete check: take ρxy=0.85, σy/σx=1, l²x=l²y=0.1, u²x=u²y=0.9, and l_rhohat=u_rhohat=0.8. The point (R²wx,R²wy,ρhat)≈(0.659,0.326,0.8) satisfies the correct Lagrange equations and (4), is feasible, and is not any of the forms (7a)–(7g); its βx|w is about 1.406, while the feasible candidates of the listed forms include (0.5,0.1,0.8) with value about 1.342. Thus the stated enumeration can return an interval that is too narrow, contradicting equation (6).","section":"Appendix C / Proposition 2.2"},{"comment":"The feasibility condition (4) divides by RwxRwy, so it is undefined when R²wx=0 or R²wy=0. The main text nevertheless allows l²x=0 and l²y=0 in (3a) and (3b), and the Section 3 example sets l²y=0 and reports [−36.60,17.71]. Proposition 2.1's converse also nominally covers R²wx=0 or R²wy=0. The paper must either require 0<l²x and 0<l²y, or treat the zero cases separately, for example by a limiting argument or by a different parametrization of (4). As written, the example's interval is not supported by the stated theory.","section":"Section 2, constraints (3a)–(3c) and (4)"}],"minor_comments":[{"comment":"The formula in (7g) appears to contain a typo: for the case (3b)+(3c)+(4), the quadratic should be written in terms of b²y, not b²x, in the argument of q²±; as printed, the formula does not match the derivation that precedes it.","section":"Appendix C and Proposition 2.2, line (7g)"},{"comment":"There is a notation inconsistency in (7e): S is a set of triples (R²wx,R²wy,ρhat), so the first two entries should be the squared bounds b²x and b²y, but the formula writes bx and by without squares.","section":"Proposition 2.2, line (7e)"},{"comment":"The caption reports u²y=50%, whereas the text of Section 3 specifies 0≤R²wy≤0.2; one of these is a typo and should be corrected.","section":"Figure 2 caption"},{"comment":"The word 'Propositon' appears instead of 'Proposition'; this is a minor typo but should be fixed.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the Lagrange-derivative error in Appendix C is decisive for the exactness claim of Proposition 2.2, and the zero lower-bound issue affects the worked example. I would ask the authors to redo the enumeration carefully, correct the formulas in (7e) and (7g), and add a numerical stress test comparing the S-based interval with a dense grid or a general-purpose optimizer over Ω."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I want you to know about this paper before it gets too much traction: it proposes a fast, exact-looking method for bounding an adjusted slope under unmeasured confounding, but the proof of its central theorem contains a wrong Lagrange multiplier step. The algorithm may still work in practice, but the exactness guarantee in equation (6) is not established.\n\nThe paper does several things well. Equation (2) is from Knaeble and Dutter (2017), so the real novelty is the optimization layer: Proposition 2.1 characterizes which (R^2_wx, R^2_wy, rho_hatx_haty) tuples are realizable, and Proposition 2.2 asserts that the extrema of beta_{x|w} over the feasible set are attained on a finite set S of at most 88 points. The exposition is clear, and they provide Python and R code.\n\nThe soft spots. First, a minor boundary issue that the authors never state: constraint (4) divides by R_wx R_wy. The main text allows l2x=0 or l2y=0, and the worked example sets l2y=0. This should be handled explicitly.\n\nMore seriously, the proof of Proposition 2.2 in Appendix C has an algebraic error in the case where (3c) and (4) are both active. For g(R_wx,R_wy)= b R_wx R_wy ± sqrt(1-R_wx^2) sqrt(1-R_wy^2), the partial derivatives are b R_wy ∓ R_wx sqrt(1-R_wy^2)/sqrt(1-R_wx^2) and b R_wx ∓ R_wy sqrt(1-R_wx^2)/sqrt(1-R_wy^2). The expression printed in the appendix contains an extra factor of 2 and uses (1+R_wy^2)^{1/2} where it should be (1-R_wy^2)^{1/2}. Once you use the correct derivatives, the condition ∇h=λ∇g no longer forces R_wx=R_wy, so the candidate points in (7b) and (7c) are not justified. That means the finite set S may omit interior extrema on that surface, and the interval returned by the algorithm could be too narrow. This is a load-bearing flaw, not a cosmetic typo.\n\nThe paper deserves a serious referee: the idea is useful, the code is concrete, and the error is likely repairable. But I would not accept it as is, and I would not cite the exactness claim until the proof is fixed. If you want a reading group case study on why referees check Lagrange multipliers, this is a good one.","headline":"Good idea with a clean implementation, but the central exactness proof has an algebraic error in the Lagrange step and the finite-set enumeration is not yet justified.","tokens_in":15719,"tokens_out":7801,"would_cite":false,"duration_ms":70115,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J05","62F35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A confounding interval for an adjusted regression slope can be computed exactly by checking at most 88 candidate parameter tuples, so unmeasured confounding becomes quantifiable rather than a caveat.","keywords":["unmeasured confounding","confounding interval","sensitivity analysis","coefficient of determination","partial correlation","regression slope","causal inference","model uncertainty"],"falsifier":"Run a numerical search: for fixed $\\rho_{xy}$ and $\\sigma_y/\\sigma_x$, generate many random feasible tuples on dense grids inside $\\Omega$ for several bound sets, evaluate $\\beta_{x|w}$, and compare the extremes against the 88-point enumeration; any sampled value outside $[l,u]$ would disprove Proposition 2.2. To test the bounds' validity rather than the enumeration, simulate data with known unmeasured $w$, compute the true $\\beta_{x|w}$, and construct bounds that exclude the true tuple; the method will then return an interval missing the true value, confirming that the diagnostic power comes from the bounds.","tokens_in":1816,"feed_emoji":"📊","tokens_out":2574,"duration_ms":88756,"temperature":0.7,"pith_summary":"Regression analyses of observational data usually report a confidence interval for the slope but add a warning that unmeasured confounders could break a causal reading. This paper replaces that warning with a numerical answer: if a researcher can state plausible bounds on three statistics describing the unmeasured confounders, the adjusted slope is guaranteed to lie in an interval that can be computed exactly. The central result is that the interval is the exact image of the feasible set under the adjusted-slope formula, and that image can be found by checking no more than 88 candidate parameter tuples. The authors demonstrate the method on a study of prenatal PBDE exposure and childhood IQ, where a confounding interval like $[-36.60, -5.25]$ is far more informative than a generic caution.","feed_headline":"Exact confounding intervals from checking 88 candidates","feed_subtitle":"Specify three plausible bounds on unmeasured confounders and get an exact interval for the adjusted slope.","key_machinery":"The load-bearing object is the rational map $\\beta_{x|w}$ defined by equation (2), together with the feasible set $\\Omega$ carved out by three box constraints and the realizability interval (4). The mechanism that carries the argument is constraint enumeration: because $\\beta_{x|w}$ is monotone in $\\rho_{\\hat{x}\\hat{y}}$ and in the confounding-strength coordinates except in special cases, an optimum must occur with at least two constraints active, and enumerating the combinations of active bounds, solving quadratics where needed and applying the Lagrange condition where the realizability curve is active, produces at most 88 candidate points. This turns a nonconvex continuous optimization problem into finite enumeration.","core_discovery":"The paper's central claim is Proposition 2.2: for any nonempty feasible set $\\Omega$ defined by box constraints on $(R^2_{wx}, R^2_{wy}, \\rho_{\\hat{x}\\hat{y}})$ together with the realizability constraint (4), the minimum and maximum of the adjusted slope $\\beta_{x|w}$ over $\\Omega$ are attained on a finite subset $S$ with $|S| \\le 88$. Therefore the confounding interval $[l,u] = \\beta_{x|w}(\\Omega)$ can be computed exactly by evaluating $\\beta_{x|w}$ at the candidate points in $S$. The candidates are generated by solving the stationarity conditions that arise when subsets of the eight inequality constraints are active, including quadratic formulas and Lagrange multiplier conditions; points where no constraints are active cannot be optimal because the objective is monotone in at least one free coordinate. The formula being propagated is $\\beta_{x|w} = \\frac{\\sigma_y}{\\sigma_x}\\frac{\\rho_{xy} - R_{wx}R_{wy}\\rho_{\\hat{x}\\hat{y}}}{1-R^2_{wx}}$, and the realizability constraint ensures that every tuple considered can actually arise from real data.","pith_inferences":["Not in the paper: the same finite-candidate machinery applies to any function of $(R^2_{wx}, R^2_{wy}, \\rho_{\\hat{x}\\hat{y}})$ that is monotone in at least one coordinate on $\\Omega$, such as the partial correlation in equation (9), so the enumeration could be reused for other sensitivity targets.","Not in the paper: if the bounds are estimated from measured subsets of $w$ rather than being subject-matter certainties, a natural next step is to propagate sampling error in the bounds into a conservative widening of the interval, perhaps by bootstrapping the bound estimates.","Not in the paper: the realizability constraint in (4) divides by $R_{wx}R_{wy}$, so the case where a lower bound is zero, as in the worked example's $l^2_y = 0$, requires a limiting argument that the text does not supply."],"forward_implications":["For any user-specified bounds that make the feasible set nonempty, the algorithm returns the exact endpoints of the confounding interval rather than a Monte Carlo or grid approximation.","The interval $[l,u]$ is the full set of adjusted slopes compatible with the bounds: a true slope outside it implies that at least one stated bound was wrong.","Researchers can report a confounding interval alongside the usual confidence interval, separating sampling uncertainty from uncertainty about unmeasured attributes.","Tighter subject-matter bounds on the correlation between fitted treatment and fitted outcome can directly shrink the interval, as the diet example shows.","Because the number of candidate points is bounded by 88 independent of the number of unmeasured confounders, the computation stays trivial as the confounding set grows."],"supporting_citations":[{"why":"Supplies the adjusted-slope identity used as Proposition 1.1 and the earlier sign-reversal condition that this paper extends to an interval.","marker":"Knaeble and Dutter (2017)"},{"why":"Supplies the Lagrange multiplier optimality condition used in the proof of Proposition 2.2 to characterize extrema on the realizability surface.","marker":"Nocedal and Wright (2006)"},{"why":"Offers an alternative parameterization of confounding impact and an index for bounding confounder effects that motivates the factored form used here.","marker":"Frank (2000)"},{"why":"Establishes reference points for speculating about omitted confounders, the practice this paper replaces with a computed interval.","marker":"Hosman et al. (2010)"},{"why":"Provides the PBDE-and-IQ case study data, including the unadjusted slope, standard deviations, and adjusted estimate used to demonstrate the method.","marker":"Eskenazi et al. (2013)"},{"why":"Supplies the propensity-score and ignorable-treatment-assignment framework that justifies causal interpretation when an admissible confounding set $w$ exists.","marker":"Rosenbaum and Rubin (1983)"}],"fun_headline_variants":["A finite check for exact unmeasured-confounding intervals","88 points pin down the confounding interval exactly","Exact adjustment for unmeasured confounders via 88 candidates","Confidence intervals miss confounding; this interval covers it exactly","Propagate three bounds, get exact confounding intervals"],"cache_read_input_tokens":17664,"weakest_assumption_plain":"The entire calculation inherits the researcher's bounds on the three confounding statistics, and the feasibility formula silently assumes the product of the two confounding strengths is positive, so a zero lower bound (as in the worked example) needs a limiting interpretation not supplied in the text.","fun_headline_variants_meta":{"raw":{"variants":["A finite check for exact unmeasured-confounding intervals","88 points pin down the confounding interval exactly","Exact adjustment for unmeasured confounders via 88 candidates","Confidence intervals miss confounding; this interval covers it exactly","Propagate three bounds, get exact confounding intervals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1765,"prompt_tokens":989,"completion_tokens":776,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":698}},"tokens_in":605,"tokens_out":776,"duration_ms":7298,"temperature":1.0,"reasoning_tokens":698,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:34:34.037060+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a numerical search: for fixed $\\rho_{xy}$ and $\\sigma_y/\\sigma_x$, generate many random feasible tuples on dense grids inside $\\Omega$ for several bound sets, evaluate $\\beta_{x|w}$, and compare the extremes against the 88-point enumeration; any sampled value outside $[l,u]$ would disprove Proposition 2.2. To test the bounds' validity rather than the enumeration, simulate data with known unmeasured $w$, compute the true $\\beta_{x|w}$, and construct bounds that exclude the true tuple; the method will then return an interval missing the true value, confirming that the diagnostic power comes from the bounds.","supporting_citations":[{"cited_title":"and Wright, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Lagrange multiplier optimality condition used in the proof of Proposition 2.2 to characterize extrema on the realizability surface."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Offers an alternative parameterization of confounding impact and an index for bounding confounder effects that motivates the factored form used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes reference points for speculating about omitted confounders, the practice this paper replaces with a computed interval."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PBDE-and-IQ case study data, including the unadjusted slope, standard deviations, and adjusted estimate used to demonstrate the method."},{"cited_title":"and Rubin, D","cited_arxiv_id":null,"evidence_quote":"Supplies the propensity-score and ignorable-treatment-assignment framework that justifies causal interpretation when an admissible confounding set $w$ exists."}],"review_version":1}