{"id":"6ec938ee-8673-4931-b935-abcf425b631e","arxiv_id":"2507.17137","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"By modeling the observed outcome with a location-scale model and fitting an induced logistic regression for missingness, the proposed estimator avoids the multiple-root instability of inverse probability weighting for non-ignorable missing data.","lead":"Statistical methods that reweight survey data to account for missing answers can be numerically unstable when missingness depends on the unobserved value. This paper diagnoses the instability as estimation of a moment-generating function and proposes a semiparametric location-scale model with a conditional-likelihood estimator that proves stable and more efficient in simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that the method 'circumvents the estimation of a moment-generating function' is contradicted by the estimator of τ in §3.2, which is built from empirical MGFs of the residuals; the promised stability of the remedy is therefore unproven for the main target parameter.","rationale":"The reader's CONDITIONAL verdict is reasonable, and my stress-test does not overturn it. I focused on the abstract's central claim because the paper's own estimator contradicts it. In Section 3.2, after the stable logistic step for θ, the construction of \\hatτ explicitly uses the empirical MGFs \\hatM_1(\\hatγ) and \\hatM_2(\\hatγ). Since Section 2 identifies MGF estimation as the source of IPW instability, the proposed remedy relocates, rather than removes, that instability in the target estimator. This is an internally verifiable inconsistency, not a matter of consensus. The asymptotic results do not address finite-sample stability of the MGF ratio, and no simulation isolates its contribution. This does not mean the method fails: the logistic step may make the remaining MGF evaluation well-conditioned, and the reported simulations are favorable. But the burden is on the authors to qualify the claim and to demonstrate that the residual MGF stage does not reintroduce instability in harder regimes. The location-scale assumption is also a legitimate concern and remains the reader's weakest assumption; I partially agree with the reader. The proposed concrete test—larger γ, a heavier-tailed residual mixture with finite MGF, and a variance decomposition—would settle whether the MGF stage is benign. If the ratio's contribution remains small, the paper's claim should be rephrased but the method stands; if not, the remedy's central promise fails.","tokens_in":15968,"tokens_out":16333,"duration_ms":170801,"concrete_test":"Run the §4 simulation with the same location-scale model and logistic missingness but with a larger true γ (e.g., γ0=1.5 or 2.0 instead of 0.5) and a residual mixture with larger variance (e.g., 2/3N(-1,1)+1/3N(2,8)). For n=500 and n=2000, record the Monte Carlo variance of \\hatM_2(\\hatγ)/\\hatM_1(\\hatγ) and the share of the total MSE of \\hatτ contributed by this ratio, plus the number of replications with extreme \\hatτ. If this share and the MSE grow sharply with γ0 or residual dispersion, the empirical-MGF stage of the proposed estimator inherits the instability that motivated the paper, and the central claim fails. If the share stays small, the concern is largely semantic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central motivation is that IPW is unstable because its estimating equations require estimating the MGF E(e^{γY}|X,R=1). The abstract concludes that the proposed method 'circumvents the estimation of a moment-generating function and hence overcomes the instability.' However, after estimating θ through the induced logistic model (8), the target mean in §3.2 is estimated by \\hatτ = n^{-1}\\sum_i μ(x_i,\\hatξ) + (1-\\hatη)\\hatM_2(\\hatγ)/\\hatM_1(\\hatγ), where \\hatM_1(t) = \\sum_i r_i e^{t\\hatε_i}/\\sum_i r_i and \\hatM_2(t) = \\sum_i r_i \\hatε_i e^{t\\hatε_i}/\\sum_i r_i. These are exactly empirical MGFs (and the derivative of an MGF) of the residuals. Theorem 2's influence function likewise contains R e^{γ0 ε} terms. Thus the method does not eliminate MGF estimation; it moves it from the propensity-score equations to the final mean functional. For large γ0 or heavy-tailed residual distributions (still with finite MGF, as the model requires), the ratio \\hatM_2/\\hatM_1 can be high-variance and slow to converge, so the 'overcomes instability' claim is not established. The paper gives no simulation or analysis isolating this residual MGF contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies estimation of a response mean tau when Y may be non-ignorably missing and the missingness probability follows a logistic model. The authors argue that inverse probability weighting is unstable because the IPW estimating equations involve estimation of a moment-generating function, and they illustrate multiple roots and non-convergence. As a remedy, they assume that the distribution of Y given X among complete cases follows a semiparametric location-scale model y = mu(x;xi) + epsilon, derive an induced logistic regression for R given X, estimate the propensity parameters by maximum conditional likelihood, and estimate tau by plugging in the estimated mean function plus a residual-based adjustment (1-eta) M2(gamma)/M1(gamma). They provide identifiability conditions, asymptotic normality theorems, simulations comparing several existing estimators, and two real-data analyses.","tokens_in":16274,"tokens_out":10245,"duration_ms":106180,"significance":"The proposed two-step estimator is a useful addition to the non-ignorable missing data toolkit if the location-scale model holds. The paper's strengths include a clear identifiability analysis, a bounded conditional-likelihood score for the propensity parameters, extensive simulations that document non-convergence rates of competitors, and real-data analyses with goodness-of-fit checks. However, the advertised advantage over IPW is not established as stated: the final estimator of tau explicitly uses empirical moment-generating functions of the residuals, so the MGF estimation problem is not circumvented. The strong assumption that epsilon is independent of X is acknowledged but not stress-tested. The theoretical results are deferred to a supplementary file that is not included in the reviewed version. These issues are fixable, but they currently stand between the manuscript and publication.","major_comments":[{"comment":"The claim that the method 'circumvents the estimation of a moment-generating function and hence overcomes the instability of IPW methods' is contradicted by the estimator of tau in Eq. (12), namely \\hat\\tau = n^{-1}\\sum_i \\mu(x_i,\\hat\\xi) + (1-\\hat\\eta)\\hat M_2(\\hat\\gamma)/\\hat M_1(\\hat\\gamma), where \\hat M_1(t)=\\sum_i r_i e^{t\\hat\\varepsilon_i}/\\sum_i r_i and \\hat M_2(t)=\\sum_i r_i \\hat\\varepsilon_i e^{t\\hat\\varepsilon_i}/\\sum_i r_i are empirical moment-generating functions of the residuals. Theorem 2's influence-function vector S2 also contains R e^{\\gamma_0\\varepsilon} and R\\varepsilon e^{\\gamma_0\\varepsilon}. Thus the estimation of an MGF is moved into the mean functional rather than eliminated. The manuscript should either restrict the claim to the propensity-score step or provide a simulation or analysis isolating the behavior of \\hat M_2/\\hat M_1 for large \\gamma_0 or heavy-tailed residual distributions; as written, the 'overcomes instability' conclusion is not supported.","section":"Abstract and Section 3.2, Eq. (12)"},{"comment":"The derivation of the induced logistic model relies on the location-scale assumption that epsilon is independent of X and has constant variance. If the error is heteroscedastic or depends on X, then c(x;\\gamma,\\xi)=\\gamma\\mu(x;\\xi)+\\log M_1(\\gamma) fails, Eq. (8) is misspecified, and \\hat\\tau can be biased. Remark 1 acknowledges that model (7) cannot be verified for unobserved Y; the Breusch-Pagan and USS tests can only check the observed complete cases. Because this assumption is load-bearing for consistency, the paper should add a sensitivity analysis (e.g., a heteroscedastic or correlated-error simulation) or at least an explicit statement of the unverifiable nature of this assumption and its consequences.","section":"Section 3, model (7) and Eq. (8)"},{"comment":"The main text states that all technical details are postponed to a supplementary material file, but no supplementary file is included with the reviewed manuscript. The proofs of Theorems 1 and 2 and the consistency of \\hat\\Sigma and \\hat V are therefore not verifiable from the submitted version. Please include the supplementary material with the revision, or provide the proofs in an appendix.","section":"Section 1 and Section 3.3"},{"comment":"The simulation summary states that RB and MSE are computed only for convergent or non-reliable cases. For competitors with large NCR counts (e.g., Table 3, \\alpha_0=-2.2, mixture error, n=500: \\hat\\tau_{A1} has NCR=575), the reported MSE is conditional on a selected subsample and can substantially understate the actual risk of the method. The paper should report results over all repetitions, with a clearly specified convention for non-convergent repetitions, or explain why the conditional-on-convergence summary is appropriate. As it stands, the relative-performance claim is not fully supported.","section":"Section 4.2, Tables 2-3"}],"minor_comments":[{"comment":"In the full text, the first sentence contains the typo 'u sed' instead of 'used'.","section":"Section 1"},{"comment":"The missingness model is written as pr(D=1|x,y) although the paper uses R for the missingness indicator elsewhere; please unify the notation.","section":"Section 5, ACTG example"},{"comment":"'forth quartile' should be 'fourth quartile'.","section":"Section 5, PPVT example"},{"comment":"The ACTG point estimate is labeled \\hat\\mu while the text defines the estimator as \\hat\\tau; please use consistent notation across the table and text.","section":"Table 5"},{"comment":"Some entries in the reference list, such as Tsiatis (2006), Chen and Liu (2013), and Keziou and Leoni-Aubin (2008), do not appear to be cited in the main text; please reconcile the reference list.","section":"References"},{"comment":"The number of simulation repetitions is not stated in Section 4; please specify it in the simulation setup.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The abstract overstates what is demonstrated: the proposed estimator still uses empirical MGFs in the final mean functional. The editor may also want to confirm that the supplementary material promised in the text actually exists and is included with the submission, since it is not part of the reviewed version. No other concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real idea: under a logistic missingness model plus a location-scale model for the observed outcome, the propensity parameters can be estimated from an induced logistic regression in x, which sidesteps the wildly unstable estimating equations that IPW uses. That portion is new, clearly derived, and backed by simulations showing strong gains over existing methods. The identifiability discussion is useful, and the two real-data examples are honest in their model checks.\n\nWhat I want to flag is the mismatch between the abstract and the actual estimator. The paper says the method 'circumvents the estimation of a moment-generating function.' That's only true for the propensity step. The final target τ is estimated with M_2(γ)/M_1(γ), built from empirical MGFs of residuals. So MGF estimation does not disappear; it moves into the mean functional. The stress-test note is right on this. For large γ or heavy-tailed residual distributions, the ratio can be high-variance, and the paper gives no simulation isolating that contribution. This is not fatal to the method — the simulations look good and the asymptotics are there — but the strong claim in the abstract and Section 2 needs tempering. The authors should either show that residual MGF ratios behave well in the settings they target, or revise the wording.\n\nOther soft spots are minor. The location-scale assumption with constant variance and independent error is untestable on the missing cases, though the paper's goodness-of-fit checks on observed cases are a reasonable defense. The proofs are in the supplement, which is normal for a stats journal and not a problem per se.\n\nOverall this is a serious, careful paper. The central mechanism is the induced logistic model, and that holds up. The overclaim is a wording/scoping issue, not a broken derivation. I'd send it to a careful referee. The referee should focus on the MGF ratio in τ and ask for a version of the simulations that tracks the variance of M_2/M_1.\n\nWho is it for: missing-data methodologists, especially people using IPW under MNAR. Deserves a serious referee.","headline":"A solid semiparametric fix for unstable IPW under non-ignorable missingness, but the abstract overclaims that MGF estimation is circumvented entirely.","tokens_in":16780,"tokens_out":2930,"would_cite":true,"duration_ms":29366,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D10","62F12","62J12"],"pacs":[],"model":"deepseek-v4-flash","headline":"IPW estimators for non-ignorable missing data fail because their score equations estimate a moment-generating function; modeling complete-case outcomes as a location-scale model turns the problem into stable logistic regression.","keywords":["inverse probability weighting","non-ignorable missing data","missing not at random","logistic regression","semiparametric location-scale model","moment-generating function","maximum conditional likelihood","propensity score"],"falsifier":"Generate data from the paper's Example 1 with heteroscedastic errors, e.g. $\\varepsilon\\sim N(0,x_1^2)$, while keeping the logistic missingness model true; if the proposed estimator of $\\tau$ shows substantial bias or if the induced-logistic goodness-of-fit test rejects at a high rate, the location-scale assumption is violated and consistency fails. Alternatively, generate errors from a distribution without a finite moment-generating function, such as a $t$ distribution with 2 degrees of freedom; if the proposed estimator cannot be computed reliably, then the remedy has moved the MGF problem rather than removed it.","tokens_in":15786,"feed_emoji":"📊","tokens_out":11785,"duration_ms":109938,"temperature":0.7,"pith_summary":"Inverse probability weighting is a standard tool for handling data where the chance a value is missing depends on the value itself, but the paper argues that these estimators are structurally fragile. The IPW score equations require estimating a moment-generating function of the unobserved-error distribution, and such estimates are numerically unstable and can have multiple roots, so the resulting estimates of the response mean can be badly biased even with tens of thousands of observations. As a remedy, the paper models the outcome among complete cases as a location-scale model $y=\\mu(x;\\xi)+\\varepsilon$ with an unknown error distribution. Under the usual logistic model for missingness, this turns the missingness indicator into a logistic regression on $\\mu(x;\\xi)$, which can be fit by maximum conditional likelihood with bounded score equations and no moment-generating-function estimation. The paper proves asymptotic normality of the estimators and shows in simulations that the proposed estimator beats existing competitors in bias, MSE, and convergence.","feed_headline":"Logistic fix cures unstable inverse probability weighting","feed_subtitle":"Modeling complete-case outcomes as location plus unknown error makes propensity estimation a stable logistic regression.","key_machinery":"The load-bearing object is the induced logistic model (equation 8): $\\mathrm{pr}(R=1|x)=1/[1+\\exp\\{\\alpha+x_1^\\top\\beta+\\gamma\\mu(x;\\xi)\\}]$. It arises because the location-scale assumption makes the conditional MGF of $Y$ given $X,R=1$ collapse to $\\exp(\\gamma\\mu(x;\\xi))$ times the scalar $M_1(\\gamma)$, so the MGF affects only the intercept. Estimation proceeds in two steps: least squares on complete cases for $\\xi$, then maximum conditional likelihood (a standard logistic regression) for $\\theta=(\\alpha,\\beta,\\gamma)$ with $\\xi$ replaced by $\\hat\\xi$. The final $\\tau$ estimator uses only the ratio $\\hat M_2(\\hat\\gamma)/\\hat M_1(\\hat\\gamma)$ of residual sample moments, not a full MGF curve.","core_discovery":"The paper establishes that, under a logistic model for the missingness probability $\\mathrm{pr}(R=1|x,y)=1/[1+\\exp(\\alpha_0+x_1^\\top\\beta+\\gamma y)]$ and a semiparametric location-scale model $y=\\mu(x;\\xi)+\\varepsilon$ for complete cases, where $\\varepsilon$ is independent of $X$ with mean zero and unknown density, the conditional moment-generating function of $Y$ given $X=x,R=1$ equals $\\exp\\{\\gamma\\mu(x;\\xi)\\}M_1(\\gamma)$. Hence the induced propensity model is $\\mathrm{pr}(R=1|x)=1/[1+\\exp\\{\\alpha+x_1^\\top\\beta+\\gamma\\mu(x;\\xi)\\}]$, with the unknown $M_1(\\gamma)$ absorbed into the intercept $\\alpha$. Estimating $(\\alpha,\\beta,\\gamma)$ by maximizing the conditional likelihood of $R$ given $X$ uses bounded score functions and avoids MGF estimation entirely; the response mean is then $\\tau=E\\{\\mu(X;\\xi)\\}+(1-\\eta)M_2(\\gamma)/M_1(\\gamma)$, estimated by plugging in sample moments of residuals. The paper proves $\\sqrt{n}$-consistency and asymptotic normality of $\\hat\\xi$, $\\hat\\theta$, and $\\hat\\tau$, gives a variance estimator and Wald intervals, and shows in simulations that this estimator has small bias and MSE and far fewer non-convergent or non-reliable cases than IPW, adaptive, and generalized-moment competitors, with the largest gains when the error distribution is non-normal.","pith_inferences":["The $\\hat\\tau$ formula still divides two estimated exponential moments of residuals, so the remedy pushes MGF instability out of the propensity step but does not eliminate it from the mean-estimation step; heavy-tailed errors or large $\\gamma$ could still make $\\hat\\tau$ unstable, and checking whether the error MGF exists or using robust moment estimation would be a natural extension.","The same induced-logistic device could be used with other semiparametric outcome models (e.g., transformation or single-index models) to obtain stable propensity estimates while keeping the outcome flexible.","The identifiability condition that $\\mu(x;\\xi)$ is nonlinear in $x$ is checkable in practice, which suggests a data-driven way to decide whether an instrumental variable is needed rather than assuming one exists.","A direct benchmark extension: comparing the proposed estimator against IPW in settings where the missingness is missing at random rather than non-ignorable would clarify how much of the gain comes from the location-scale model versus from avoiding MGF estimation."],"forward_implications":["With correct location-scale and logistic models, the propensity parameters are identifiable whenever $\\mu(x;\\xi)$ is nonlinear in $x$ or an instrumental variable is present, and the paper supplies the asymptotic covariance for valid Wald intervals.","The proposed estimator avoids the basis-function and kernel choices that make generalized-moment and adaptive competitors non-convergent in small samples or high missingness; in simulations its MSE is at least 50% lower and often much lower than the generalized-moment method.","Users can check the two model assumptions on observed data: the score test for non-constant variance for the location-scale model, and the USS goodness-of-fit test for the induced logistic model.","Because the induced logistic model is a standard logistic regression on constructed covariates, any complete-data logistic tool (variable selection, regularization, diagnostics) becomes applicable to non-ignorable missing-data propensity estimation.","A bootstrap t-interval is recommended for the response mean at moderate sample sizes when the error is non-normal, since the normal-based Wald interval under-covers at $n=500$ in simulations."],"supporting_citations":[{"why":"Establishes that a shadow variable identifies the full-data distribution with non-ignorable missing data, the identification background the paper builds on.","marker":"Miao et al. (2019)"},{"why":"Provides the instrumental-variable condition used to identify propensity parameters and motivates the presence of $x_2$ in the identifiability discussion.","marker":"Wang et al. (2014)"},{"why":"Gives the full-semiparametric-likelihood competitor $\\hat\\tau_P$ and the parametric conditional-outcome modeling approach that the proposed method generalizes by relaxing the error distribution.","marker":"Liu et al. (2022)"},{"why":"Supplies the semiparametric adaptive estimators $\\hat\\tau_{A1}$ and $\\hat\\tau_{A2}$ that serve as benchmarks and the ancillary-variable doubly robust estimation idea.","marker":"Morikawa & Kim (2016)"},{"why":"Provides the generalized-moment estimators $\\hat\\tau_{Gk}$ used as competitors and illustrates the basis-function dependence that the proposed method avoids.","marker":"Ai et al. (2020)"},{"why":"Supplies the USS goodness-of-fit test used to validate the induced logistic model (8) in the real examples.","marker":"le Cessie & van Houwelingen (1995)"},{"why":"Provides the score test for non-constant error variance used to check the location-scale assumption in real data.","marker":"Breusch & Pagan (1979)"},{"why":"Extends the heteroscedasticity diagnostic used alongside Breusch & Pagan to verify constant variance in the complete-case outcome model.","marker":"Cook & Weisberg (1983)"}],"fun_headline_variants":["Conditional likelihood tames unstable IPW for missing data","No moment-generating function: stable fix for IPW instability","Semiparametric remedy overcomes IPW non-convergence and bias","Bounded scores replace MGF estimation in missing-data propensity","Induced logistic model sidesteps IPW's multiple-root problem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The complete-case outcome follows exactly a location-scale model $y=\\mu(x;\\xi)+\\varepsilon$ in which the error $\\varepsilon$ has mean zero, is independent of the covariates, and has constant variance.","fun_headline_variants_meta":{"raw":{"variants":["Conditional likelihood tames unstable IPW for missing data","No moment-generating function: stable fix for IPW instability","Semiparametric remedy overcomes IPW non-convergence and bias","Bounded scores replace MGF estimation in missing-data propensity","Induced logistic model sidesteps IPW's multiple-root problem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1285,"prompt_tokens":1118,"completion_tokens":167,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":79}},"tokens_in":734,"tokens_out":167,"duration_ms":6642,"temperature":1.0,"reasoning_tokens":79,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:56:05.893034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate data from the paper's Example 1 with heteroscedastic errors, e.g. $\\varepsilon\\sim N(0,x_1^2)$, while keeping the logistic missingness model true; if the proposed estimator of $\\tau$ shows substantial bias or if the induced-logistic goodness-of-fit test rejects at a high rate, the location-scale assumption is violated and consistency fails. Alternatively, generate errors from a distribution without a finite moment-generating function, such as a $t$ distribution with 2 degrees of freedom; if the proposed estimator cannot be computed reliably, then the remedy has moved the MGF problem rather than removed it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the full-semiparametric-likelihood competitor $\\hat\\tau_P$ and the parametric conditional-outcome modeling approach that the proposed method generalizes by relaxing the error distribution."},{"cited_title":"Semiparametric Optimal Estimation With Nonignorable Nonresponse Data","cited_arxiv_id":"1612.09207","evidence_quote":"Supplies the semiparametric adaptive estimators $\\hat\\tau_{A1}$ and $\\hat\\tau_{A2}$ that serve as benchmarks and the ancillary-variable doubly robust estimation idea."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the score test for non-constant error variance used to check the location-scale assumption in real data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends the heteroscedasticity diagnostic used alongside Breusch & Pagan to verify constant variance in the complete-case outcome model."}],"review_version":1}