{"id":"18433bb2-739f-43b3-a662-86d60caf3345","arxiv_id":"2507.15612","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Under the multiplicative IV model, any counterfactual functional of the treated that solves a moment equation is identified and can be estimated efficiently with valid inference.","lead":"This paper shows that a recently proposed multiplicative instrumental variable model can identify and estimate nonlinear counterfactual quantities such as quantile treatment effects on the treated, not just average effects. The authors provide efficient, multiply robust estimators and confidence intervals with asymptotic guarantees, tested in simulations and on the Job Corps dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The identification result is correct under MIV, but Assumption 2 is the sole load-bearing bridge: any non-multiplicative Z–U interaction invalidates the key equality, and the paper neither tests the assumption's observable no-crossing implication nor provides a sensitivity analysis.","rationale":"Reading the proof of Theorem 1 in good faith, I find the algebra correct. The conditional independence U⊥⊥Z|A=1,X in (16) is the key consequence of Assumption 2, and the subsequent factorization of δ_M,A is valid. The central claim is therefore internally consistent under the stated model. What is load-bearing is the multiplicative form itself. Because the proof replaces the U-conditional ratio of propensities by the X-conditional ratio, any interaction between Z and U in the treatment mechanism breaks the identification; the bias is first order, not a smooth perturbation. This is exactly the reader's weakest assumption, so I agree with the conditional verdict. I add two concrete points the reader did not fully develop: (i) Assumption 2 has an observable no-crossing implication P(A=1|Z=1,X)≤P(A=1|Z=0,X) after labeling, which could be checked in the Job Corps application; (ii) a simple simulation with an interaction term can quantify how quickly the MIV estimand diverges from the true counterfactual functional. I do not see an internal inconsistency in Theorem 2's EIF or in the multiple robustness claims; the remaining conditions (bounded moments, relevance, cross-fitting rates) are standard. The paper would be strengthened by a sensitivity analysis, an explicit positivity condition in Theorem 1, and less sweeping language about 'any measurable function.' Because the reader already recommended conditional acceptance on these grounds, I leave the verdict unchanged.","tokens_in":24245,"tokens_out":20591,"duration_ms":225365,"concrete_test":"Simulate with X empty, U~Unif(0,1), Z~Bernoulli(1/2), Y^0=U+N(0,1), and p0(U)=0.2+0.5U. Set P(A=1|Z=0,U)=p0(U) and P(A=1|Z=1,U)=clip(p0(U)(0.7+η(U−0.5)),0,1), so η=0 is exactly MIV and η>0 introduces a Z–U interaction. For q=0.5, compute the true quantile β* = median(Y^0|A=1) by Monte Carlo and the MIV population estimand solving h(β)=0 from (4). Grid η∈{0,0.1,...,0.5}; if |β_MIV(η)−β*| grows linearly in η, the identification equality fails under non-multiplicative misspecification, confirming the concern and motivating a reported sensitivity analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The theorem itself is sound: given Assumption 2, the proof of (16) and the factorization of δ_M,A are valid. The load-bearing step is the line 'by Assumption 2' in the proof of Theorem 1, where P(A=1|Z=1,U,X)/P(A=1|Z=0,U,X) is replaced by P(A=1|Z=1,X)/P(A=1|Z=0,X). This replacement is exactly the statement that the instrument-confounder ratio is constant in U. If the true treatment mechanism is, say, logistic with an interaction term η·Z·U, this ratio varies with U and the observable estimand becomes a weighted average of E[M(Y^0,β)|U,X] with weights p0(U,X)[1−r(U,X)] divided by the corresponding averaged propensity difference. The result is then not E[M(Y^0,β)|A=1,X], and the bias is first order in η rather than a small remainder term. The paper provides no sensitivity analysis for this central assumption and does not state or verify its observable implication that, after labeling, P(A=1|Z=1,X)≤P(A=1|Z=0,X) pointwise. Theorem 1 also omits the positivity/relevance condition δ_A(X)≠0, which Assumption 2 alone does not guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the recently introduced multiplicative IV (MIV) model to inference on general nonlinear counterfactual functionals of the treatment-free outcome among the treated. Under the MIV assumption and standard IV conditions, Theorem 1 identifies any moment equation E[M(Y^{a=0},β)|A=1]=0 as an observable ratio of covariances, and the paper develops efficient influence-function-based estimators with multiple robustness, cross-fitting, and asymptotically valid confidence intervals for the moment value. An inverse-inference procedure then yields confidence sets for the functional itself, including the quantile treatment effect on the treated. The methods are illustrated by simulations and an application to the Job Corps dataset.","tokens_in":24513,"tokens_out":7069,"duration_ms":76039,"significance":"If the results hold, this is a useful expansion of the MIV model beyond the average treatment effect on the treated to quantile and distributional effects. The identification algebra in Theorem 1 is correct under the stated assumptions, and the semiparametric efficiency theory in Theorem 2 and the multiple robustness structure in Corollary 1 are carefully derived and consistent with the literature. The paper provides rigorous proofs, a clearly described cross-fitting procedure, and a simulation study showing tight coverage at moderate sample sizes. The principal limitations are the strong, untestable multiplicative structure of Assumption 2, which is not accompanied by a sensitivity analysis, and a few technical gaps in the theorem statements that affect the interpretation of the identification and inference results.","major_comments":[{"comment":"Theorem 1 divides by δ_A(X) but does not require δ_A(X) ≠ 0 almost surely. Under Assumption 2, δ_A(X) = E[g2(U,X)|X](g1(1,X) - 1), which is non-positive and can vanish when g1(1,X) = 1, so the identified expression in (3) and the estimating equation h(β,P)=0 in (4) may be undefined. The theorem should add a positivity/relevance condition such as δ_A(X) ≠ 0 a.s. and state how this is implied (or not) by the model assumptions; without it, the central identification claim is not well-defined.","section":"Section 2.1, Theorem 1"},{"comment":"The identification rests entirely on the multiplicative factorization in Assumption 2. In the proof, the ratio P(A=1|Z=1,U,X)/P(A=1|Z=0,U,X) is replaced by the observable ratio P(A=1|Z=1,X)/P(A=1|Z=0,X) using exactly this factorization. If the true mechanism is non-multiplicative, for example logistic with an interaction η·Z·U, this replacement fails and the observable estimand (4) becomes a U-weighted average of E[M(Y^{a=0},β)|U,X] rather than the target E[M(Y^{a=0},β)|A=1]; the bias is first order in η. The paper gives no sensitivity analysis or bias bound for departures from Assumption 2, nor does it state the model's testable implication that P(A=1|Z=1,X) ≤ P(A=1|Z=0,X) pointwise. Given that the assumption is untestable, the authors should at least characterize the bias under a non-multiplicative perturbation or provide a sensitivity analysis.","section":"Section 2.1, Assumption 2 and Theorem 1"},{"comment":"The abstract refers to functionals that are 'the unique solution' of a moment equation, but Theorem 1 identifies the moment equation without stating a uniqueness or monotonicity condition for β*. For non-monotone moment functions, or when the counterfactual distribution has flat regions, h(β,P)=0 may have multiple solutions; the confidence set in (11)–(12) will still contain the true β* but may include spurious values, and the claimed inference on the functional is then not well defined. The paper should either impose an explicit identifiability condition (e.g., strict monotonicity of M in β with a continuity/density condition, as suggested in the QTT example) or define β* as a chosen root and discuss the implications for the reported set.","section":"Section 2.3, Corollary 4 and Algorithm 1"}],"minor_comments":[{"comment":"In the final step, both β_L and β_R are defined using 'min'; β_R should be the maximum of the grid values satisfying |θ̂_β| ≤ z_α σ̂_β/√n, otherwise the output is not even the interval of solutions.","section":"Algorithm 1, Step 4"},{"comment":"There are several typographical issues: 'measureable' should be 'measurable' (Section 1.2), 'thatβ' should be 'that β' (Section 2.2), and 'A TT' should be 'ATT' (Section 2.3).","section":"Throughout"},{"comment":"The condition ∥h(β, ˆP) − h(β,P)∥ = o_P(1) is not the usual rate condition for the asymptotic normality result; the authors should state it in terms of the componentwise L2 rates and product-bias rates that appear in Assumption 3, or at least clarify the assumed convergence rate of h(β, ˆP).","section":"Section 2.2, Corollary 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and the efficiency theory is competently executed. The main weakness is the lack of any sensitivity analysis for the multiplicative IV assumption, which is the sole structural bridge identifying the target. The omitted positivity condition in Theorem 1 is a technical gap that must be fixed. I recommend major revision with a request for a formal sensitivity analysis or bias characterization, and for explicit identifiability conditions for the functional β*."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper shows the multiplicative IV model identifies any moment-defined functional of the treatment-free counterfactual among the treated, not just the ATT. The math is right. The novelty is real but modest: it applies a known moment-mapping technique to a model introduced by the same group, and adds efficient estimation and inference. The soft spot is the load-bearing MIV assumption, which is untestable and gets no sensitivity analysis.\n\nWhat's good: Theorem 1 is correct under Assumptions 1 and 2; the proof of U ⊥⊥ Z | A=1, X is clean. The EIF in Theorem 2 and the multiply robust structure in Corollary 1 are consistent with the semiparametric literature. The cross-fitting inference follows Chernozhukov et al. carefully. Simulations show tight coverage at moderate n, and the Job Corps application is a nice touch. The paper is honest about prior work.\n\nSoft spots: (1) Theorem 1 omits the positivity condition δ_A(X) ≠ 0; Assumption 2 alone doesn't guarantee it. (2) The 'any measurable function' claim is too strong; identification is fine for measurable M, but the inference and remainder calculations need boundedness and smoothness conditions. (3) The bigger issue: the multiplicative structure is doing all the work. If the treatment mechanism has a non-multiplicative Z–U interaction, the ratio P(A=1|Z=1,U,X)/P(A=1|Z=0,U,X) is not constant in U, and the observable estimand becomes a weighted average of the counterfactual moment, not the target. The bias is first order in the interaction. The paper neither tests the observable no-crossing implication P(A=1|Z=1,X) ≤ P(A=1|Z=0,X) nor provides a sensitivity analysis. That is worth acknowledging, even if it doesn't sink the paper.\n\nWho it's for: researchers working on IV identification of distributional effects, especially QTT. It's a useful step for the MIV program. It deserves a serious referee. I'd send it out, with the expectation that the authors add the positivity condition, temper the measurable-function claim, and discuss the fragility of the multiplicative assumption.","headline":"Solid extension of the MIV model to general counterfactual moments; the identification is correct given the multiplicative assumption, but the paper should be more upfront that everything hangs on that untestable structure.","tokens_in":25071,"tokens_out":2568,"would_cite":false,"duration_ms":27558,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62G05","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The multiplicative IV model identifies any moment-defined functional of the treatment-free counterfactual among the treated, including quantiles, and supports efficient multiply robust inference.","keywords":["multiplicative instrumental variable model","counterfactual moment equations","quantile treatment effect on the treated","semiparametric efficiency","multiply robust estimation","cross-fitting","test inversion","noncompliance"],"falsifier":"Simulate data satisfying the core IV assumptions but with $P(A=1\\mid Z,U,X)=g_1(Z,X)+g_2(U,X)$, estimate a counterfactual moment such as $E[1\\{Y^{a=0}\\le y\\}\\mid A=1]$ with the proposed estimator in large samples, and compare with the truth; nonvanishing bias would show the multiplicative assumption is carrying the identification. A more direct check is to obtain a proxy $\\tilde U$ and test whether $\\tilde U\\perp\\!\\!\\perp Z\\mid A=1,X$, as the model implies for $U$.","tokens_in":24064,"feed_emoji":"📊","tokens_out":7791,"duration_ms":76679,"temperature":0.7,"pith_summary":"This paper extends the multiplicative instrumental variable model from the average treatment effect on the treated to a broad class of counterfactual targets. It shows that any functional of the treatment-free counterfactual among the treated that solves a moment equation — means, distribution functions, quantiles, and quantile treatment effects — is identified under the MIV assumption. The key step is a ratio identity that rewrites the counterfactual moment as an observable function of the instrument and outcome data. The paper also constructs a semiparametrically efficient, multiply robust estimator and confidence intervals by test inversion, with asymptotic coverage guarantees and simulation support at moderate sample sizes.","feed_headline":"Multiplicative IV model identifies every counterfactual moment","feed_subtitle":"One multiplying-factor assumption turns moment equations about untreated potential outcomes into identified targets with valid intervals","key_machinery":"The load-bearing object is the multiplicative IV assumption itself: $P(A=1\\mid Z,U,X)=g_1(Z,X)g_2(U,X)$ with $g_1(0,X)=1$. This factorization makes the instrument-to-treatment probability ratio independent of $U$, which in turn implies the conditional independence $U\\perp\\!\\!\\perp Z \\mid A=1, X$ used in the proof of Theorem 1. That independence is what converts the unobservable counterfactual moment into the observable ratio of instrument-induced differences $\\delta_{M,A}(\\beta,X)/\\delta_A(X)$, and the same structure supports the efficient influence function, the multiple robustness conditions, and the test-inversion confidence sets.","core_discovery":"Under the multiplicative IV model, the conditional moment of the counterfactual equals an observable ratio of covariances: for any measurable moment function $M$, $$E[M($Y^{{a=0}}$, \\$\\beta$) \\mid A=1] = -E\\left[\\frac{\\delta_{M,A}(\\$\\beta$, X)}{\\delta_A(X)} \\mid A=1\\right],$$ where $\\delta_{M,A}$ and $\\delta_A$ are the instrument-induced differences in $E[M(Y,\\beta)(1-A)\\mid Z,X]$ and in $P(A=1\\mid Z,X)$. Since this holds for every measurable $M$, the solution $\\beta^*$ of the counterfactual moment equation is identified; choosing $M(Y^{a=0},\\beta)=1\\{Y^{a=0}\\ge \\beta\\}-q$ identifies the $q$-th quantile of the treatment-free counterfactual among the treated, and the authors use this to build confidence sets for the quantile treatment effect on the treated. The same framework yields an efficient influence function for the moment, an estimator that remains unbiased under any of three alternative nuisance specifications, and asymptotically valid confidence intervals obtained by inverting the moment test.","pith_inferences":["Editorial extension: if the factorization holds for a non-binary instrument, the same ratio logic may identify moment functionals with the instrument categories collapsed into contrasts, though the paper does not develop this.","Editorial extension: because the model implies $U\\perp\\!\\!\\perp Z\\mid A=1,X$, a researcher with an auxiliary proxy for $U$ could construct a falsification test of the MIV assumption; the paper offers no such test or sensitivity analysis.","Editorial extension: the test-inversion confidence set can be a union of intervals in one dimension, as the paper notes; a smoothed or profiled version could yield connected intervals and shorter coverage."],"forward_implications":["The MIV assumption identifies all moment-defined functionals of the treatment-free counterfactual among the treated, not only its mean, so quantile treatment effects on the treated no longer require monotonicity or a complier subpopulation.","Estimation and inference can use the efficient influence function with cross-fitting; the estimator is multiply robust, remaining unbiased if one of several alternative nuisance models is correct.","Confidence intervals for functionals without closed-form estimates are obtained by inverting the moment test on a grid and carry asymptotic coverage guarantees.","In the Job Corps application, the procedure yields confidence intervals for the median treatment effect on treated log weekly earnings that are positive at conventional levels."],"supporting_citations":[{"why":"introduces the multiplicative IV model and establishes identification of the ATT, which this paper extends to general moment-defined functionals.","marker":"[Liu et al., 2025]"},{"why":"supplies the cross-fitting and Neyman-orthogonal score framework used for the asymptotically valid inferential procedure.","marker":"[Chernozhukov et al., 2018]"},{"why":"provides the semiparametric decomposition used to prove asymptotic normality of the moment estimator.","marker":"[Kennedy, 2016]"},{"why":"supplies the empirical-process lemma bounding the contribution of nuisances estimated on independent folds.","marker":"[Kennedy et al., 2020]"},{"why":"shows the second-order remainder derivation for the efficient influence function of the IV estimand, adapted in the proof of Theorem 2.","marker":"[Levis et al., 2024]"},{"why":"justifies maximizing a p-value over a confidence set for the nuisance quantile when constructing the QTT confidence set.","marker":"[Berger and Boos, 1994]"}],"fun_headline_variants":["Multiplicative IV: inference for any counterfactual moment","Quantile treatment effects identified by multiplicative IV","Multiply robust estimation of counterfactual moments","Beyond the average: multiplicative IV for quantiles","One assumption, many counterfactual inferential targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the untestable assumption that the instrument and the unmeasured confounders influence the probability of treatment by multiplying separate factors; if they combine additively or interact on another scale, the identifying equality can fail.","fun_headline_variants_meta":{"raw":{"variants":["Multiplicative IV: inference for any counterfactual moment","Quantile treatment effects identified by multiplicative IV","Multiply robust estimation of counterfactual moments","Beyond the average: multiplicative IV for quantiles","One assumption, many counterfactual inferential targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000477,"raw_usage":{"total_tokens":2346,"prompt_tokens":911,"completion_tokens":1435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":1362}},"tokens_in":527,"tokens_out":1435,"duration_ms":11561,"temperature":1.0,"reasoning_tokens":1362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:27:56.412114+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data satisfying the core IV assumptions but with $P(A=1\\mid Z,U,X)=g_1(Z,X)+g_2(U,X)$, estimate a counterfactual moment such as $E[1\\{Y^{a=0}\\le y\\}\\mid A=1]$ with the proposed estimator in large samples, and compare with the truth; nonvanishing bias would show the multiplicative assumption is carrying the identification. A more direct check is to obtain a proxy $\\tilde U$ and test whether $\\tilde U\\perp\\!\\!\\perp Z\\mid A=1,X$, as the model implies for $U$.","supporting_citations":[{"cited_title":"Semiparametric theory and empirical processes in causal inference","cited_arxiv_id":null,"evidence_quote":"provides the semiparametric decomposition used to prove asymptotic normality of the moment estimator."}],"review_version":1}