Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Debiased Ill-Posed Regression

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper shows that debiasing the projected error via its influence function makes the estimator's bias second-order in the nuisances, so slow operator estimation costs only its squared error and root-n inference on linear functionals…

desk verdict Interesting influence-function debiasing idea for ill-posed regression, but the central rate theorem's proof has a gap that needs fixing before the quartic-rate claim is accepted. read the letter →

arxiv 2505.20787 v1 pith:4GTKEJFR submitted 2025-05-27 stat.ME econ.EMstat.ML

AmirEmad Ghassami , James M. Robins , Andrea Rotnitzky This is my paper · ORCID
classification stat.MEecon.EMstat.ML MSC 62G0562G2062G0865J22
keywords ill-posedregressionconditionalmomentrestrictioninfluencefunctiondebiasingsecond-orderbiasiterativeTikhonovregularizationhyper-parameterselectionnonparametricinstrumentalvariableproximalcausalinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper studies estimation of a function $h_0$ that solves a conditional moment restriction, an ill-posed inverse problem covering nonparametric instrumental variables and proximal causal inference. The standard approach minimizes an estimated projected mean-squared error; the paper shows this is fragile because the operator estimate enters linearly. It constructs a new objective by adding the influence-function debiasing term of the projected error, which makes the bias second-order in the nuisance estimators. The central result is a finite-sample bound in which the cost of operator-estimation error is the fourth power of $\|T - \widehat T\|$ rather than its square, with a robustness guarantee: misspecifying one of the two nuisance functions costs no more than reverting to the undebiased rate. It then provides a cross-validated choice of the regularization parameter and proves that, under an additional smoothness condition, the resulting estimator is accurate enough for root-n inference on linear functionals such as counterfactual means.

What carries the argument

The central object is the influence function of the projected error functional $\psi(h)=E[(E[g_1(V)h(V_h)|V_q]-g_0(V))^2]$, derived in Theorem 2. The estimator augments the squared projected residual with the debiasing cross term $2\{(\widehat T h)(V_q)-\widehat r(V_q)\}\{g_1(V)h(V_h)-(\widehat T h)(V_q)\}$, which cancels the first-order nuisance bias. The debiased loss is minimized with two-step iterative Tikhonov regularization, and the analysis combines the second-order bias identity of Proposition 2, localized Rademacher complexity bounds for empirical-process terms, the $\beta$-source condition for regularization bias, and the $\alpha$-error condition that converts projected-error bounds into source-error bounds by interpolation in a common basis.

What would settle it

Run a simulation of the nonparametric instrumental variable model with a compact operator whose singular system is known, so the truth is computable. Make the operator estimator converge at a controlled slow rate, say $\|T-\widehat T\|\asymp n^{-1/4}$, and let $\widehat r$ converge at a different rate; the theorem predicts the debiased source error contains $\|T-\widehat T\|^4$ while the non-debiased estimator contains $\|T-\widehat T\|^2$. If the observed scaling of $\|\widehat h-h_0\|^2$ with $\|T-\widehat T\|$ at the optimal $\lambda$ is quadratic rather than quartic, the central second-order-bias claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the $L^2$-minimal solution of the integral equation $T h = r_0$ can be estimated by minimizing a debiased version of the projected error, and that the debiasing fully exploits the structure of the influence function. Concretely, Theorem 4 asserts that with two-step iterative Tikhonov regularization and under the $\beta$-source condition, the source error of the debiased estimator is bounded by $(1/\lambda^2)\max\{\delta_n^2, \|T - \widehat T\|^4, \|T - \widehat T\|^2\|\widehat r - r_0\|^2\} + \lambda^{\min\{4,\beta\}}$, and choosing $\lambda$ optimally yields source error of order $\Delta_n^{\min\{3,\beta\}/\min\{5,\beta+2\}}$. This is the quartic-in-operator-error improvement over the quadratic dependence of the non-debiased estimator. For hyper-parameter selection, the paper claims that a cross-validated choice of $\lambda$ over a fine grid attains the oracle projected-error rate up to terms that vanish as candidate functions converge, and that under the $\alpha$-error condition the source error is bounded by a power of the projected error. Finally, for the regular parameter class whose influence functions are products of the two nuisance solutions, the paper claims root-n consistent and asymptotically normal estimation with nonparametric nuisance rates, including for counterfactual means under proximal causal inference.

Load-bearing premise

The cross-validated source-error guarantee and the root-n functional results rest on Assumption 4, which requires every candidate error to be uniformly smooth in one common basis with a single unknown constant $\alpha$ that cannot be verified before one already knows the error converges.

Editorial extensions

If this is right

  • With the debiased estimator, the source-error bound depends on $\|\widehat T-T\|^4$ instead of $\|\widehat T-T\|^2$, so a slowly converging operator estimator no longer dominates the rate.
  • If the debiasing nuisance $\widehat r$ is misspecified, the error bound degrades at most to that of the non-debiased estimator, making the procedure robust to one misspecified nuisance.
  • Cross-validated hyper-parameter selection over a fine grid preserves the oracle rates whenever the oracle variance terms dominate; any rate loss comes only from approximation terms that vanish as candidate functions converge.
  • For linear functionals of solutions, root-n consistency and asymptotic normality hold with nonparametric convergence rates for all nuisance estimators, provided the product of their projected errors is $o_p(n^{-1/2})$.
  • In the proximal causal inference illustration, this yields root-n inference for counterfactual means using only nonparametric bridge-function estimators, where the undebiased alternative would demand faster-than-parametric rates.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to test whether the quartic dependence on $\|\widehat T-T\|$ persists when the debiasing nuisance $\widehat r$ is estimated on the same fold, since the analysis assumes separate folds for candidates and nuisances.
  • The two-layer debiasing template of debiasing the loss for the nuisance and then debiasing the final functional could transfer to other inverse problems, such as density deconvolution or imaging problems where the operator is estimated at slow nonparametric rates.
  • Because a finite $\alpha$ in the $\alpha$-error condition can only be certified after convergence is known, practitioners should treat the cross-validated source-error rate as a heuristic; the oracle-$\lambda$ results remain the guaranteed ones.
  • Increasing the number of Tikhonov iterations would replace $\lambda^{\min\{4,\beta\}}$ by $\lambda^\beta$ at the price of an exponential constant, so benchmarking two-step versus many-step debiasing on simulated data would clarify whether the qualification-limited term binds in realistic sample sizes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper studies estimation of the L2-minimal solution h0 of a conditional moment restriction T h = r0, covering nonparametric IV and proximal causal inference as leading examples. It proposes an influence-function-debiased estimator h_hat^IF_{lambda,2} that minimizes a debiased version of the projected squared error plus an iterated Tikhonov penalty. The central claims are: (i) the debiasing induces a second-order bias in the nuisance estimates (Proposition 2); (ii) under a beta-source condition the source error depends on max{delta_n^2, ||T - T_hat||^4, ||T - T_hat||^2 ||r_hat - r0||^2}, replacing the quadratic dependence on ||T - T_hat|| of the non-debiased estimator (Theorem 4 and Remark 1); (iii) a cross-validation procedure for the hyper-parameter has a projected-error oracle bound (Theorem 6) and, under an alpha-error condition, a source-error bound (Theorem 7); and (iv) these rates imply root-n consistency for linear functionals in the parameter class of Ghassami et al. (2022), with an application to proximal causal inference (Corollaries 2-3). The paper contains detailed proofs of the bias calculations and of the rate algebra.

Significance. If the rate claims survive a repair of the localization argument, the paper makes a substantial contribution: it proposes a debiased estimator for ill-posed conditional moment models with second-order bias in both the operator and the r-nuisance, and it quantifies a genuine robustness improvement over Li et al. (2024) in the dependence on the operator estimation error. The influence-function derivation (Theorem 2) and the bias-structure calculation (Proposition 2) are clean and appear correct. The treatment of hyper-parameter selection is also valuable, as the rate loss from not knowing the optimal lambda is rarely analyzed in inverse problems. The root-n conditions in Corollary 2 are concrete and checkable given the rates. However, the central rate theorem currently rests on an unjustified Lipschitz claim, and the cross-validated root-n claim rests on an unverifiable alpha-error condition; both need to be addressed before the main results can be accepted.

major comments (3)
  1. [Sec. 3.1 and Appendix A, proof of Theorem 4 (also Theorem 3)] The claimed L2-Lipschitz bound for the loss function l(V;h) is not a consequence of Assumption 3. In the displayed verification, terms such as ||2 g1(V)(T_hat h1)(Vq)(h1 - h2)(Vh)||_2 and ||2 g1(V) h2(Vh)(T_hat(h1 - h2))(Vq)||_2 are bounded by a constant times ||h1 - h2||_2 only if the multipliers g1, T_hat h1, and h2 are uniformly bounded in L-infinity; Assumption 3(ii) provides only L2(P0) boundedness, and products of L2 functions are not controlled in L2. The subsequent inequality also drops the term ||T_hat - T|| ||h1 - h2||_2 and invokes "T is a contraction", which requires a bound such as |g1| <= 1 that is not stated as a general assumption in Theorem 4. Because Lemma 2 is applied to this l, the empirical-process step leading to the (1/lambda^2) max{delta_n^2, ||T - T_hat||^4, ||T - T_hat||^2 ||r_hat - r0||^2} bound is not justified. The same issue affects the proof of Theorem 3, where Lemma 2 is invoked for the same loss. This is load-bearing for Theorem 4 and, through it, for Corollaries 2 and 3.
  2. [Sec. 3.2 and Appendix A, proof of Theorem 6] The variance bound in the proof of Theorem 6 is derived by expanding E[(l(h*,eta_hat) - l(h*,eta0) - l(h,eta_hat) + l(h,eta0))^2] into products such as 2 g1(V) h*(Vh)((T_hat - T)h*)(Vq) and 2 g0(V)((T - T_hat)(h* - h))(Vq), and then asserting the expectation is O(||T - T_hat||^2 ||h - h*||_2^2 + ||r0 - r_hat||_2^2 ||h - h*||_2^2). Bounding the second moments of these products again requires L-infinity bounds on g0, g1, h*, T_hat h*, and r_hat, which Assumption 3 does not provide. Without these bounds, the Bernstein step that yields the log(2M/zeta)/n term is unsupported. Since Theorem 6 is the basis for the hyper-parameter selection guarantee and for Corollary 3, this is a second load-bearing gap that must be repaired, either by strengthening Assumption 3 or by a different argument.
  3. [Sec. 3.2, Assumption 4 and Remark 8] The alpha-error condition is not verifiable from primitive conditions. Remark 8 concedes that a finite alpha is guaranteed only once source-error convergence is already known, so the condition cannot be checked in advance. Theorem 7 and Corollary 3 therefore depend on an assumption whose verification is essentially the source-error convergence that the theorem is meant to establish. The paper should either prove, for its examples, that Assumption 4 holds under the beta-source condition used in Theorems 4-5, or state explicitly that Corollary 3 is a conditional result and does not provide verifiable root-n inference under cross-validated hyper-parameter selection. As written, the cross-validated source-error claim collapses if Assumption 4 fails, leaving only the oracle-lambda results (Theorems 3-5, Corollary 2).
minor comments (6)
  1. [Abstract and Introduction] There are several typographical errors: "in the filed of causal inference" should be "in the field of causal inference", "Correspondance" should be "Correspondence" in the footnote, and "underscore the importance" should be "underscores the importance" in the abstract.
  2. [Proposition 2] In the final display of Proposition 2, the term "||r - r_hat0||_2" should be "||r0 - r_hat||_2" or equivalently "||r_hat - r0||_2" for consistency with the rest of the paper.
  3. [Appendix B, Lemma 2] Lemma 2 states that the loss is Lipschitz with respect to the L2(P0) norm, but the proof of Theorem 4 verifies a bound on the L2 norm of the difference l(V;h1) - l(V;h2). The paper should clarify which Lipschitz notion the lemma requires, since the pointwise and L2 versions lead to different sufficient conditions.
  4. [Sec. 3.2] The sentence "We assume that the debiasing succeed in the sense that delta_n^2 = Delta_n and delta_{M,n}^2 = Delta_{M,n}" introduces a strong assumption about the candidate set and the nuisance estimates; this should be stated as an explicit assumption in Theorems 6 and Corollary 1 rather than appearing as an informal remark, because the subsequent oracle bound depends on it.
  5. [Example 1] The Cauchy-Schwarz step in Example 1 bounds sum_i (mu_i^(n))^alpha <e_n, phi_i>^2 by n^{-alpha/10} (sum_i i^{-6alpha})^{1/2} (sum_i <e_n, phi_i>^4)^{1/2}; the final bound should involve ||e_n||_2^2 rather than ||e_n||_2, so the displayed conclusion "||e_n||_2 <= n^{alpha/10 - 1/3}" does not follow as written and the numerical value alpha = 10/3 should be re-derived.
  6. [Sec. 3.1, definitions] The notation X_1 <= X_2 is defined, but in several displays the paper writes "<=+" or "≲+" without explanation; please define these symbols or avoid them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the debiased estimator is derived from an influence-function calculation and explicit bias algebra, with all rate bounds conditional on stated regularity conditions.

full rationale

The paper's central derivation is a pathwise influence-function calculation (Theorem 2) followed by a bias decomposition (Proposition 2) that is algebraic in the nuisance errors, and finite-sample bounds obtained from localized Rademacher complexity and Bernstein-type concentration (Lemmas 2-3, Theorems 3-6). No parameter is fitted to a target and then renamed a prediction; the debiasing term is constructed from the influence function of the projected error, and its second-order bias is verified, not assumed. Assumption 4 (the alpha-error condition) is a stated regularity condition: Theorem 7 is a direct interpolation consequence of its two inequalities, and Remark 8 explicitly concedes that finite alpha may only be verifiable once source-error convergence is known. That is an uncheckable-assumption weakness, not a circular reduction, because the theorems are conditional on Assumption 4 and do not purport to derive it. Self-citations (Ghassami et al. 2022; Robins et al. 2008; Rotnitzky et al. 2021) supply background model classes and mixed-bias structure, but the rate and root-n claims are proven in the present paper rather than imported. The skeptic's concern about the Lipschitz verification in the proof of Theorem 4 requiring sup-norm boundedness is a correctness gap in the proof, not evidence that the conclusion is equivalent to the assumptions by construction. I therefore find no significant circularity.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

No constants are fitted to data in this paper; the regularization parameter lambda is selected by cross-validation and the b_n, B_n grid sequences are described analytically (Section 3.2). The paper's honest content is the influence-function derivation and concentration analysis; its load-bearing costs are the stated regularity assumptions. The most expensive single assumption is the alpha-error condition (Assumption 4), which is tailored to the proof of Theorem 7 and is acknowledged to be barely verifiable. No new entities (particles, forces, mediators) are introduced.

assumptions (7)
  • domain assumption Assumption 1: r0 belongs to R(T), so a solution to Th = r0 exists.
    Stated in Section 2. Without existence of a population solution, the estimation problem is not well-posed at the target.
  • domain assumption Assumption 2: beta-source condition h0 = (T*T)^(beta/2) w0 with ||w0||2 <= B.
    Standard regularity condition for inverse problems controlling the regularization bias; invoked in Theorems 4 and 5 via Lemma 1.
  • domain assumption Assumption 3: h0 is in the user-specified class H, and H and R are uniformly bounded in the L2(P0) norm.
    Realizability and complexity control. The proofs also appear to need sup-norm boundedness of product terms and of g1 for the Lipschitz step; see red flag on unflagged assumption.
  • domain assumption The population regularized solutions h*_{lambda,1} and h*_{lambda,2} (Equation 5) belong to H.
    Stated in Theorems 4 and 5; requires the population competitor to be inside the class for the basic inequality argument.
  • ad hoc to paper Assumption 4: alpha-error condition with common basis phi_i and sequence mu_i^(n) non-increasing in both i and n, plus smoothness and projected-error lower bounds on all candidate errors.
    Introduced to convert projected-error bounds into source-error bounds for the cross-validated estimator (Theorem 7). Remark 8 concedes a finite alpha is only guaranteed once source-error convergence is known, so the condition is hard to verify in advance.
  • standard math Influence-function calculus for the nonparametric model (Robins et al. 1994; van der Laan and Robins 2003) is valid for the projected error functional.
    Theorem 2 computes pathwise derivatives of the projected error under regular parametric submodels; standard semiparametric theory.
  • standard math Localized Rademacher complexity machinery, especially the localized concentration inequality of Lemma 2 (Foster and Syrgkanis 2019), applies to the four function classes used in Theorems 3-5.
    Used to control the empirical-process terms; requires uniform boundedness of the classes, which links back to Assumption 3(ii).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Debiased Ill-Posed Regression." pith.science (2026). https://pith.science/paper/4GTKEJFR

@misc{pith2026250520787,
  author       = {Pith},
  title        = {Pith review of: Debiased Ill-Posed Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4GTKEJFR}},
  note         = {Machine review of arXiv:2505.20787}
}
read the original abstract

In various statistical settings, the goal is to estimate a function which is restricted by the statistical model only through a conditional moment restriction. Prominent examples include the nonparametric instrumental variable framework for estimating the structural function of the outcome variable, and the proximal causal inference framework for estimating the bridge functions. A common strategy in the literature is to find the minimizer of the projected mean squared error. However, this approach can be sensitive to misspecification or slow convergence rate of the estimators of the involved nuisance components. In this work, we propose a debiased estimation strategy based on the influence function of a modification of the projected error and demonstrate its finite-sample convergence rate. Our proposed estimator possesses a second-order bias with respect to the involved nuisance functions and a desirable robustness property with respect to the misspecification of one of the nuisance functions. The proposed estimator involves a hyper-parameter, for which the optimal value depends on potentially unknown features of the underlying data-generating process. Hence, we further propose a hyper-parameter selection approach based on cross-validation and derive an error bound for the resulting estimator. This analysis highlights the potential rate loss due to hyper-parameter selection and underscore the importance and advantages of incorporating debiasing in this setting. We also study the application of our approach to the estimation of regular parameters in a specific parameter class, which are linear functionals of the solutions to the conditional moment restrictions and provide sufficient conditions for achieving root-n consistency using our debiased estimator.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 18 canonical work pages

  1. [1]

    Bennett, A., Kallus, N., Mao, X., Newey, W., Syrgkanis, V., and Uehara, M. (2023). Source condition double robust inference on functionals of inverse problems. arXiv preprint arXiv:2307.13793

    arXiv 2023

  2. [2]

    Carrasco, M., Florens, J.-P., and Renault, E. (2007). Linear inverse problems in structural econometrics estimation based on spectral decomposition and regularization. Handbook of econometrics , 6:5633--5751

    2007

  3. [3]

    Cavalier, L. (2011). Inverse problems in statistics. In Inverse Problems and High-Dimensional Estimation: Stats in the Ch \^a teau Summer School, August 31-September 4, 2009 , pages 3--96. Springer

    2011

  4. [4]

    and Pouzo, D

    Chen, X. and Pouzo, D. (2012). Estimation of nonparametric conditional moment models with possibly nonsmooth generalized residuals. Econometrica , 80(1):277--321

    2012

  5. [5]

    and Reiss, M

    Chen, X. and Reiss, M. (2011). On rate optimality for ill-posed inverse problems in econometrics. Econometric Theory , 27(3):497--521

    work page 2011

  6. [6]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters

    2018

  7. [7]

    Cui, Y., Pu, H., Shi, X., Miao, W., and Tchetgen Tchetgen, E. (2023). Semiparametric proximal causal inference. Journal of the American Statistical Association , pages 1--12

    work page 2023

  8. [8]

    Darolles, S., Fan, Y., Florens, J.-P., and Renault, E. (2011). Nonparametric instrumental regression. Econometrica , 79(5):1541--1565

    2011

Show all 33 references
  1. [9]

    W., Hanke, M., and Neubauer, A

    Engl, H. W., Hanke, M., and Neubauer, A. (1996). Regularization of inverse problems , volume 375. Springer Science & Business Media

    1996

  2. [10]

    Florens, J.-P., Johannes, J., and Van Bellegem, S. (2011). Identification and estimation by penalization in nonparametric instrumental regression. Econometric Theory , 27(3):472--496

    2011

  3. [11]

    Foster, D. J. and Syrgkanis, V. (2019). Orthogonal statistical learning. arXiv preprint arXiv:1901.09036

    2019 arXiv

  4. [12]

    Ghassami, A., Ying, A., Shpitser, I., and Tchetgen Tchetgen, E. (2022). Minimax kernel machine learning for a class of doubly robust functionals with application to proximal causal inference. In International conference on artificial intelligence and statistics , pages 7210--7...

    2022

  5. [13]

    and Horowitz, J

    Hall, P. and Horowitz, J. L. (2005). Nonparametric methods for inference in the presence of instrumental variables

    2005

  6. [14]

    Hartford, J., Lewis, G., Leyton-Brown, K., and Taddy, M. (2017). Deep iv: A flexible approach for counterfactual prediction. In International Conference on Machine Learning , pages 1414--1423. PMLR

    2017

  7. [15]

    Hern \'a n, M. A. and Robins, J. M. (2020). Causal inference: what if . Boca Raton: Chapman & Hall/CRC

    2020

  8. [16]

    Kallus, N., Mao, X., and Uehara, M. (2021). Causal inference under unmeasured confounding with negative controls: A minimax learning approach. arXiv preprint arXiv:2103.14029

    2021 arXiv

  9. [17]

    Kress, R. (2013). Linear Integral Equations . Applied Mathematical Sciences. Springer New York

    2013

  10. [18]

    Li, Z., Lan, H., Syrgkanis, V., Wang, M., and Uehara, M. (2024). Regularized deepiv with model selection. arXiv preprint arXiv:2403.04236

    2024 arXiv

  11. [19]

    Miao, W., Geng, Z., and Tchetgen Tchetgen, E. J. (2018). Identifying causal effects with proxy variables of an unmeasured confounder. Biometrika , 105(4):987--993

    2018

  12. [20]

    Newey, W. K. and Powell, J. L. (2003). Instrumental variable estimation of nonparametric models. Econometrica , 71(5):1565--1578

    2003

  13. [21]

    Neyman, J. (1923). On the application of probability theory to agricultural experiments. essay on principles. Ann. Agricultural Sciences , pages 1--51

    1923

  14. [22]

    Robins, J., Li, L., Tchetgen Tchetgen, E., van der Vaart, A., et al. (2008). Higher order influence functions and minimax estimation of nonlinear functionals. In Probability and statistics: essays in honor of David A. Freedman , pages 335--421. Institute of Mathematical Statistics

    2008

  15. [23]

    M., Rotnitzky, A., and Zhao, L

    Robins, J. M., Rotnitzky, A., and Zhao, L. P. (1994). Estimation of regression coefficients when some regressors are not always observed. Journal of the American statistical Association , 89(427):846--866

    1994

  16. [24]

    Rotnitzky, A., Smucler, E., and Robins, J. M. (2021). Characterization of parameters with a mixed bias property. Biometrika , 108(1):231--238

    2021

  17. [25]

    Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology , 66(5):688

    1974

  18. [26]

    Tautenhahn, U. (1996). Error estimates for regularization methods in hilbert scales. SIAM Journal on Numerical Analysis , 33(6):2120--2130

    1996

  19. [27]

    J., Ying, A., Cui, Y., Shi, X., and Miao, W

    Tchetgen Tchetgen, E. J., Ying, A., Cui, Y., Shi, X., and Miao, W. (2024). An introduction to proximal causal inference. Statistical Science , 39(3):375--390

    2024

  20. [28]

    van der Laan, M. J. and Dudoit, S. (2003). Unified cross-validation methodology for selection among estimators and a general cross-validated adaptive epsilon-net estimator: Finite sample oracle inequalities and examples

    2003

  21. [29]

    van der Laan, M. J. and Robins, J. M. (2003). Unified methods for censored longitudinal data and causality . Springer

    2003

  22. [30]

    Van der Laan, M. J. and Rose, S. (2011). Targeted learning: causal inference for observational and experimental data . Springer Science & Business Media

    2011

  23. [31]

    and Wellner, J

    van der Vaart, A. and Wellner, J. A. (2023). Weak convergence and empirical processes: With applications to statistics. Springer

    2023

  24. [32]

    W., Dudoit, S., and van der Laan, M

    van der Vaart, A. W., Dudoit, S., and van der Laan, M. J. (2006). Oracle inequalities for multi-fold cross validation. Statistics & Decisions , 24(3):351--371

    2006

  25. [33]

    Wainwright, M. J. (2019). High-dimensional statistics: A non-asymptotic viewpoint , volume 48. Cambridge University Press

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.