{"id":"f832f2df-bd74-481e-9d19-108a3e026b79","arxiv_id":"1908.02166","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper claims a JPEG-wavelet IV estimator that corrects both selection and endogeneity biases in truncated samples, supported by Monte Carlo evidence but resting on an unproven algorithmic identity.","lead":"This paper proposes a wavelet-based instrumental variable estimator intended to fix the bias that arises when data are endogenously truncated, meaning only a self-selected part of the population is observed. It is worth reading because truncated samples with endogenous covariates are common in applied economics, and the paper claims a correction that avoids normal-distribution assumptions and bandwidth selection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 drops a conditioning on selection when deriving the bias decomposition, so the estimating equations (51)-(52) are not actually implied.","rationale":"The reader's REJECT verdict is appropriate, but the most load-bearing weakness is not the unproved lifting-scheme equivalence in Algorithms 4 and 6. That issue affects computational fidelity: if the transposed-inverse is inexact, the proximal update in (29) is not computed as claimed, but the statistical estimator could in principle still be defined by the matrix version. The conditioning error in Theorem 1 is more fundamental: it invalidates the derivation of the estimating equations themselves, so the estimator's claimed bias correction is not established even in principle. The reader mentions 'statistical theory incomplete' but does not identify the specific dropped outer expectation, which is a clear algebraic error rather than a mere gap. I therefore partially agree with the reader's weakest assumption: the paper should be rejected, but for a more decisive reason that does not depend on the algorithmic implementation details. The concrete check is a direct analytical re-derivation and small numerical evaluation of the contested equality, which would settle whether the decomposition holds.","tokens_in":32722,"tokens_out":5873,"duration_ms":61584,"concrete_test":"Re-derive equations (7)-(10) under a simple parametric DGP where x1 is endogenous and selection depends on w (e.g., the paper's own DGP with bivariate normal disturbances and truncation y2*>0). Compute the left and right sides of equation (9) numerically for a fine grid of x_i values; if E{E[x_i^T beta|x_i]|y2i=1} differs from E[x_i^T beta|x_i] whenever selection depends on w, the decomposition in Theorem 1 is invalid and (10) cannot be used to justify (51)-(52).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central statistical claim that the estimator corrects both endogeneity and truncation bias rests on the decomposition in Theorem 1 leading to equations (10), (14), (51), and (52). In the proof, after applying the tower property, equation (8) correctly keeps outer expectations conditional on y2i=1, but equation (9) drops them: E{E[x_i^T beta|x_i]|y2i=1} is replaced by E[x_i^T beta|x_i], and similarly for the endogeneity term E[xi_1i|x_i]. This replacement is not an identity unless x_i is independent of selection, which is precisely the endogenous truncation case the paper excludes. Consequently, the 'endogeneity bias term' E[xi_1i|x_i] in (10) is a function of the realized endogenous covariate, not a conditional expectation given participation, and (10) is not a valid partially linear model with an additive selection-bias term. The two-step procedure in (51)-(52) therefore lacks a theoretical foundation: even if the transposed-inverse JPEG lifting scheme were exact, the estimator would not be guaranteed to remove the truncation bias. The lifting-scheme issue identified by the reader is secondary to this: it concerns whether the algorithm computes the claimed update, whereas the statistical target itself is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semiparametric instrumental variable estimator for endogenously truncated samples. The method combines a biorthogonal wavelet (JPEG 2000-style) denoising step, an analytic transposed-inverse lifting scheme, and a two-step estimation procedure intended to correct for both endogenous covariates and endogenous truncation. The authors report Monte Carlo simulations with non-normal disturbances, claiming recovery of the true parameters for sample sizes above 2,000 and √n consistency.","tokens_in":33048,"tokens_out":10769,"duration_ms":102541,"significance":"The problem addressed—correcting for both endogeneity of covariates and sample selection in a truncated sample—is important and widely relevant. The paper is ambitious in trying to avoid distributional assumptions and to provide a computationally efficient wavelet implementation. The Monte Carlo design with non-Gaussian, non-symmetric disturbances is a genuine strength. However, the central theoretical decomposition in Theorem 1 is invalid because conditioning on selection is dropped in the derivation of equations (8)–(10). Since the two-step estimator (51)–(52) is built on this decomposition, the statistical target of the estimator is not established. In addition, the claimed exactness of the transposed-inverse lifting scheme is asserted without proof, and the 2,000,000 distribution functions mentioned in the abstract are not substantiated by the described DGP. The empirical evidence cannot compensate for these theoretical gaps.","major_comments":[{"comment":"The passage from equation (8) to equation (9) drops the conditioning on y2i=1 in the first two terms: E{E[x_i^T β | x_i] | y2i=1} is replaced by E[x_i^T β | x_i], and E{E[ξ_1i | x_i] | y2i=1} is replaced by E[ξ_1i | x_i]. This is not an identity; it would require x_i to be independent of selection, which is precisely the endogenous truncation that the paper excludes. Consequently, equation (10) is not a valid partially linear regression model with an additive selection-bias term, and the estimating equations (51)–(52) are not implied by the model. This is a load-bearing error: the statistical target of the estimator is not established.","section":"II-B, Eqs. (8)–(10)"},{"comment":"The paper asserts that Algorithm 4 together with the transposed-inverse filter (Algorithm 6) computes the exact transpose of the inverse JPEG transform Ψ_I^T u in (49), but no proof of equivalence between the lifting operations and the matrix product is supplied. If this equivalence fails for any grid or boundary case, the proximal update in (29) is not a valid step for the objective in (25), so the computational foundation of the estimator, which the abstract calls the main contribution, is unverified.","section":"IV-E, Algorithm 4, Eq. (49)"},{"comment":"The abstract and Section V claim validation with 2,000,000 different distribution functions, but the DGP described in (53) is a single mixture with fixed parameters (μ, σ_a, σ_b, φ, σ_v) = (4, 2.5, 1.5, 2, 1). No mechanism for varying the marginal distributions across observations is provided. The Monte Carlo evidence therefore does not substantiate the stated breadth of validation, and the claim is not reproducible from the manuscript.","section":"V-A, DGP (53)"},{"comment":"The δ consistency measure is used to conclude that the JPEG IV estimator is √n consistent, but δ is only an empirical estimate of the rate at which Monte Carlo standard deviations decline for one parameter (β1) across a few sample sizes. The reported values range from 0.42 to 0.61, which is not an unambiguous confirmation of the √n rate, and no formal theorem establishing consistency or asymptotic normality is proved. The claim of √n consistency is therefore not supported by theory.","section":"V-B, Table 3"}],"minor_comments":[{"comment":"The text contains repeated typos: 'Assumtption 1' and 'Assumtption 2' should read 'Assumption 1' and 'Assumption 2'.","section":"II-C"},{"comment":"The estimation procedures in (51) and (52) do not include the selection-index parameter γ, so the paper never explains how γ is estimated or whether it is assumed known; this is a significant omission in a semiparametric single-index model.","section":"II-G"},{"comment":"Theorem 3 states only conditions (i) and (ii), but its proof uses conditional independence of z and ξ1 given w and s (condition (iii) of Theorem 2) when writing E[z ξ1 | w, s] = E[z | w, s] E[ξ1 | w, s]; the theorem statement should include that assumption.","section":"II-C, Theorem 3"},{"comment":"The sentence 'we select both the thresholding (tuning) parameter as well as the penalty function using a reference-free criterion function' is repeated or incomplete: the criterion in (34) selects λ_j and γ_j, but the mechanics of the two-fold cross-validation over the penalty shape parameter are not described.","section":"III-2"},{"comment":"The text says each observation is generated from a unique mixture of distribution functions, but the DGP in (53) fixes the mixture weights and parameters; the description is internally inconsistent.","section":"V-A"},{"comment":"The δ consistency measure is reported only for β1, not for β2, and the note in the table does not define how the standard deviations σ1 and σ2 are paired across sample sizes.","section":"V-B, Table 3"}],"recommendation":"reject","confidential_remarks":"The manuscript has a fundamental error in the derivation of the main decomposition, and the computational claim is not proven. I recommend rejection. The Monte Carlo results are not sufficient to overcome the theoretical problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's headline result doesn't follow from its own proof. In Theorem 1, equation (8) correctly keeps the outer expectation conditional on y2i=1, then (9) silently drops it. You can't replace E{E[x_i^T beta | x_i] | y2i=1} with E[x_i^T beta | x_i] unless x_i is independent of selection, which is exactly the endogenous-truncation case. Same for E[xi_1i | x_i]. So (10) is not a valid partially linear model with additive selection bias, and the two-step estimator in (51)-(52) is not actually implied. The lifting-scheme gap the reader flags is real but secondary; even a perfect transposed-inverse transform wouldn't save the statistical target.\n\nWhat's genuinely good: the Monte Carlo is substantial — non-normal mixtures, Clayton copula, six sample sizes, 10,000 replications, first-stage results in the appendix. Table 2 shows the JPEG-IV recovering the true coefficients for N≥2000, and the comparison with truncated OLS/IV makes the point that truncation breaks conventional IV. The algorithmic novelty — writing the transposed-inverse of the CDF 9/7 lifting scheme on irregular grids — is a real computational problem, and if the lifting scheme is correct it's useful.\n\nSoft spots, in proportion: the Theorem 1 error is fatal to the theory as stated. The \"2,000,000 distribution functions\" claim is inflated; from the text each observation is drawn from a unique mixture, which is not the same as testing 2M distinct DGPs. The claimed sqrt(n) consistency is inferred from a ratio of Monte Carlo standard deviations, not derived. And the exact transposed-inverse lifting scheme is asserted, not proved; Algorithms 4 and 6 need a verification against the matrix expression in (49), at least for boundary cases.\n\nWho this is for: readers working on wavelet-based semiparametric selection corrections, and anyone interested in lifting schemes for irregular grids. The MC design is a useful template even if the theory fails. My recommendation: I would not accept the paper in its current form. The central statistical claim is unsupported. But it deserves a serious referee: the computational question is nontrivial, and the Monte Carlo is careful enough that a competent referee can separate the salvageable algorithm from the broken theorem. If the authors can fix the conditioning or reframe as a heuristic procedure, there's a publishable core here.","headline":"Load-bearing error in Theorem 1 undercuts the statistical claim, but the Monte Carlo and the wavelet lifting scheme are real enough to warrant referee attention.","tokens_in":33494,"tokens_out":3067,"would_cite":false,"duration_ms":32947,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A wavelet-based IV estimator removes both truncation and endogeneity bias.","keywords":["endogenous truncation","instrumental variable","semiparametric estimation","JPEG wavelet","biorthogonal wavelet","lifting scheme","proximal gradient descent","sample selection"],"falsifier":"Run Algorithm 4 on random irregular grids, including odd-length and boundary cases, and compare its output to the direct matrix product $(\\Psi_I^{(t)})^T u$ from equations (49)-(50); any relative difference above machine precision shows the lifting scheme is not the exact transpose and the proximal update is not valid.","tokens_in":32512,"feed_emoji":"📊","tokens_out":5882,"duration_ms":56482,"temperature":0.7,"pith_summary":"This paper tries to show that the conventional instrumental-variable estimator is no longer valid when the sample is endogenously truncated, and that a semiparametric wavelet-based 'JPEG IV' estimator corrects both the truncation bias and the bias from endogenous covariates. The proof has two parts: a decomposition showing the truncated substantive equation splits into a linear part, an unknown selection-bias function, an endogeneity-bias term, and noise; and a matrix-free lifting-scheme implementation of the transposed inverse JPEG wavelet transform that makes the denoising computation feasible on irregular grids. If the claims hold, empirical researchers working with truncated samples or machine-learning training data with covariate shift would have an estimator that is bandwidth-free, distribution-free, and $\\sqrt{n}$-consistent. Monte Carlo results across 2,000,000 distribution functions are presented as verification, while the conventional IV on truncated data shows bias that persists even at 10,000 observations.","feed_headline":"A matrix-free wavelet IV estimator removes truncation bias","feed_subtitle":"Conventional instruments stay biased in truncated samples; the new estimator recovers sqrt(n) consistency.","key_machinery":"The load-bearing object is the JPEG biorthogonal wavelet transform, the CDF 9/7 filter bank, expressed as a product of auxiliary matrices for shifting, rescaling, and smoothing with interpolation weights for irregular grids. The paper's new step is equation (49), an analytic expression for the transpose of the inverse transform, $(\\Psi_I^{(t)})^T$, obtained by reversing and transposing each lifting step instead of building the full matrix; Algorithms 4 and 6 implement this transposed-inverse filter. This lets the proximal gradient update in equation (29) compute $\\Psi_I^T(u-\\Psi_I\\delta)$ exactly at each iteration, enabling group-wise MCP-thresholded denoising with resolution-dependent penalties, and a reference-free two-fold cross-validation criterion selects the thresholds.","core_discovery":"The central claim is that in an endogenously truncated sample selection model with an endogenous covariate, the substantive equation decomposes into a linear parametric part, a selection-bias term $M_1(w_i^T\\gamma)$, an endogeneity-bias term $E[\\xi_{1i}|x_i]$, and white noise. The conventional IV estimator is inconsistent because the selection indicator makes the instrument $z$ and the disturbance $\\xi_1$ conditionally dependent through the covariates $w$, so $E[z\\xi_1|s]\\neq E[z|s]E[\\xi_1|s]$. The paper proves that removing the bias term from the residual restores orthogonality, and constructs a semiparametric estimator that estimates the unknown bias functions $M_1(\\cdot)$ and $M_2(\\cdot)$ by wavelet denoising. The resulting JPEG IV estimator is claimed to correct both biases simultaneously and to be $\\sqrt{n}$-consistent, with the conventional IV nested as the special case where all detail wavelet coefficients are zero.","pith_inferences":["The matrix-free transposed-inverse lifting construction is not tied to the JPEG 9/7 filter bank; the same reversal-and-transpose logic could in principle be applied to other biorthogonal filter banks, making the proximal-denoising estimation strategy generalizable.","The paper establishes consistency through Monte Carlo convergence rates rather than an asymptotic proof; a formal distributional theory for the two-step estimator with estimated $\\gamma$ and data-adaptive thresholds would be the natural next step.","The reference-free two-fold cross-validation threshold rule could be lifted from this paper and used in other truncated-data settings where the complete distribution is unobservable, such as covariate shift in machine-learning training data.","A direct comparison against kernel-based semiparametric selection estimators on the same data-generating process would quantify the practical gain from avoiding bandwidth selection."],"forward_implications":["Conventional IV should not be used on endogenously truncated samples: the paper's Monte Carlo results show estimates for the endogenous covariate can be roughly a tenth of the true value and remain biased at large sample sizes.","Applying JPEG IV to truncated data reproduces full-sample IV performance once the sample has a few thousand observations.","The estimator achieves the standard $\\sqrt{n}$ convergence rate without requiring a bandwidth parameter or normality assumptions on the disturbances.","Because the denoising is group-wise rather than element-wise, the procedure accounts for dependence among wavelet coefficients of the same resolution level.","The method extends to irregularly spaced data through interpolation weights embedded in the wavelet filters."],"supporting_citations":[{"why":"Supplies the motivation: the existing routine for the inverse transpose transform is approximate, so an analytic exact version is needed.","marker":"[7]"},{"why":"Defines the sample-selection bias framework that the paper extends to truncation and endogeneity.","marker":"[3]"},{"why":"Provides the semiparametric least-squares single-index approach that the two-step estimation adapts.","marker":"[13]"},{"why":"Gives the root-n-consistent semiparametric regression result underlying the bias-function representation.","marker":"[16]"},{"why":"Provides the lifting-scheme factorization that Algorithms 1-4 build on.","marker":"[24]"},{"why":"Defines biorthogonal wavelets and the dual-basis representation used throughout.","marker":"[12]"},{"why":"Supplies the minimax concave penalty used in the regularized least-squares objective.","marker":"[37]"},{"why":"Supplies the two-fold cross-validation reference-free criterion for threshold selection.","marker":"[42]"},{"why":"Supplies the two-step partially linear single-index IV procedure that is modified for the truncated environment.","marker":"[44]"}],"fun_headline_variants":["Wavelet IV estimator fixes truncation and endogeneity","Matrix-free JPEG wavelet IV: unbiased in truncated samples","Semiparametric wavelet IV: no matrices, no normality needed","New IV method: wavelet denoising beats two biases at once"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole computational scheme rests on the assertion that the lifting steps in Algorithms 4 and 6 compute exactly the transpose of the inverse JPEG wavelet transform on irregular grids, but the paper supplies no proof that these steps equal the matrix product in equation (49).","fun_headline_variants_meta":{"raw":{"variants":["Wavelet IV estimator fixes truncation and endogeneity","Matrix-free JPEG wavelet IV: unbiased in truncated samples","Semiparametric wavelet IV: no matrices, no normality needed","New IV method: wavelet denoising beats two biases at once"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1812,"prompt_tokens":901,"completion_tokens":911,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":842}},"tokens_in":517,"tokens_out":911,"duration_ms":9782,"temperature":1.0,"reasoning_tokens":842,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:53:47.026381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 4 on random irregular grids, including odd-length and boundary cases, and compare its output to the direct matrix product $(\\Psi_I^{(t)})^T u$ from equations (49)-(50); any relative difference above machine precision shows the lifting scheme is not the exact transpose and the proximal update is not valid.","supporting_citations":[{"cited_title":"Compression approaches for the regularized solutions of linear systems from large-scale inverse problems,","cited_arxiv_id":null,"evidence_quote":"Supplies the motivation: the existing routine for the inverse transpose transform is approximate, so an analytic exact version is needed."},{"cited_title":"Sample selection bias as a specification error,","cited_arxiv_id":null,"evidence_quote":"Defines the sample-selection bias framework that the paper extends to truncation and endogeneity."},{"cited_title":"Semiparametric least squares (sls) and weighted sls estima- tion of single-index models,","cited_arxiv_id":null,"evidence_quote":"Provides the semiparametric least-squares single-index approach that the two-step estimation adapts."},{"cited_title":"Root-n-consistent semiparametric regression,","cited_arxiv_id":null,"evidence_quote":"Gives the root-n-consistent semiparametric regression result underlying the bias-function representation."},{"cited_title":"Factoring wavelet transforms into lifting steps,","cited_arxiv_id":null,"evidence_quote":"Provides the lifting-scheme factorization that Algorithms 1-4 build on."},{"cited_title":"Biorthogonal bases of compactly supported wavelets,","cited_arxiv_id":null,"evidence_quote":"Defines biorthogonal wavelets and the dual-basis representation used throughout."},{"cited_title":"Nearly unbiased variable selection under minimax concave penalty,","cited_arxiv_id":null,"evidence_quote":"Supplies the minimax concave penalty used in the regularized least-squares objective."},{"cited_title":"Wavelet shrinkage using cross-validation,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-fold cross-validation reference-free criterion for threshold selection."},{"cited_title":"Estimation for partially linear single-index instrumental variables models,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-step partially linear single-index IV procedure that is modified for the truncated environment."}],"review_version":1}