Pith. sign in

REVIEW 2 major objections 4 minor 56 references

For semiparametric estimation with black-box nuisance learners, TAME attains the unimprovable error rate n^{-1/2} + δ_appr_μ·δ_appr_π + (δ_stoc_n)^2, removing the mixed term that double machine learning pays.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 07:08 UTC pith:GP2TEX65

load-bearing objection The upper bound is a real advance on DML rates, but the matching lower bound is only asserted by analogy, so the 'unimprovable' claim is not yet supported. the 2 major comments →

arxiv 2607.21541 v1 pith:GP2TEX65 submitted 2026-07-23 math.ST stat.MEstat.MLstat.TH

Optimal use of a black-box learner in semiparametric estimation

Yihong Gu This is my paper
classification math.ST stat.MEstat.MLstat.TH MSC 62G0562G20
keywords semiparametric estimationblack-box learnersdouble machine learningpartial linear modeldebiasing weightsadversarial moment calibrationminimax optimalitystructure-agnostic
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Suppose a black-box learner is used to estimate both nuisance functions in a partial linear model. This paper asks how the learner's approximation errors and stochastic error should propagate to the error of the target linear coefficient. It proposes an estimator, TAME-1, whose error is at most the n^{-1/2} parametric rate plus the product of the two approximation errors plus the squared stochastic error — and proves a matching lower bound showing no estimator can do better with only those three budgets. The key improvement is removing the extra term 'max(approximation errors) × stochastic error' that the standard double-machine-learning benchmark incurs, at no additional assumption. A generic procedure, TAME, extends the same idea to two learners, general linear functionals, and average treatment effects.

Core claim

The paper's central claim is that the structure-agnostic optimal rate for estimating the linear coefficient is n^{-1/2}+δ_appr_μδ_appr_π+(δ_stoc_n)^2, attainable for any black-box class with the stated error budgets and provably unimprovable even when the class is a sparse linear model with an orthonormal dictionary. The estimator achieves this by selecting one of two branches: when the outcome function is easier, it fine-tunes the treatment-side debiasing weights on the inference sample; when the treatment function is easier, it fine-tunes the outcome-side weights using a cross-sample objective. In both cases the weights are kept close to the black-box-induced weights while approximately sa

What carries the argument

TAME-1 (and its generic version TAME) is a minimax weight-editing program: on the inference sample it solves argmin_a λ·(1/n)Σ(a_i−ê_a_i)^2 + sup over the class's difference set of an adversarial moment functional, where ê_a is the debiasing weight vector that double machine learning would use. The first term (local editing) preserves the beneficial product structure of the approximation errors; the second term (adversarial moment calibration) forces the higher-order stochastic error to appear only squared. The outcome-harder branch uses a cross-sample construction so that the edited weights stay independent of the treatment residuals, preserving conditional mean-zero structure.

Load-bearing premise

The procedure requires knowing the learner's error budgets (the two approximation errors and the stochastic error) and the ordering between the approximation errors; if these are unknown or mis-specified, the branch choice and λ are not guaranteed to yield the advertised rate.

What would settle it

Find a hypothesis class with known budgets for which, with non-negligible probability, the oracle weights do not satisfy the adversarial moment constraint (i.e., the feasible set is empty), and show the error then contains a leftover mixed term such as max(δ_appr_μ,δ_appr_π)·δ_stoc_n. Conversely, the paper's lower bound predicts that in the sparse-linear class with the same budgets no estimator can beat n^{-1/2}+δ_appr_μδ_appr_π+(δ_stoc_n)^2; a counterexample would refute the no-gap claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Asymptotic normality of the coefficient estimate is possible even if one nuisance function is only consistently estimable, provided the other is well estimated and the easier learner is under-smoothed.
  • The lower bound shows that being agnostic to the nuisance structure costs nothing relative to the worst-case structure-aware model with the same error budgets.
  • TAME can be combined with any initial black-box estimates; when nuisance difficulties are imbalanced it strictly improves on double machine learning's guarantee.
  • A two-learner variant (TAME-2) improves further, and the model-selection rule that emerges is to under-smooth the easier nuisance and balance the harder one.
  • The O-branch extends to general linear functionals and the R-branch to average treatment effect estimation, with analogous oracle-type inequalities.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A practical extension the paper does not develop: an adaptive version that estimates the error budgets from the data; if the budget estimates are biased the rate may degrade, so a sensitivity analysis of the branch choice would be valuable.
  • The transductive weight-editing idea could be applied to other debiased estimators beyond the ones listed, e.g., quantile or survival functionals, where the same mixed-error term appears.
  • The under-smoothing principle for TAME-2 suggests a concrete testable rule in neural-network benchmarks: choose the easier nuisance's network complexity above its bias-variance balance point and the harder one at the balance point.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper studies structure-agnostic estimation of the linear coefficient in the partial linear model when the nuisance functions are estimated by a generic black-box hypothesis class. It characterizes the class by three budgets — approximation errors for the outcome and treatment regressions and a stochastic-error radius — and proposes TAME-1, a debiased estimator claimed to attain |θ̂−θ0| ≲ n^{-1/2} + δ_appr_μ δ_appr_π + (δ_stoc_n)^2, removing the DML mixed term (δ_appr_μ∨δ_appr_π)δ_stoc_n. The paper also develops a two-learner version TAME-2, neural-network rate comparisons, and extensions to general linear functionals and ATE estimation. The upper-bound proof of Theorem 2.1 is detailed; the matching lower bound, however, is stated without proof.

Significance. If the matching lower bound can be established, the paper is a substantial contribution: it would show that the DML first-order stochastic nuisance errors are removable in a fully structure-agnostic setting, recover known sparse-linear rates as special cases, and provide a no-gap result between structure-agnostic and worst-case structure-aware estimation. The Section 5 proof of the upper bound is careful and appears internally consistent, and the extensions to two learners and to ATE estimation are nontrivial. The main weakness is that the paper's headline optimality claim currently rests on an unproved proposition; once that proof is supplied, the contribution is likely to be of high interest to the semiparametric and machine-learning communities.

major comments (2)
  1. [Section 2.5 (Proposition 2.3)] Proposition 2.3 is the sole basis for the 'matching lower bound' and the no-gap claim, but no proof is given anywhere in the manuscript. Section 5 proves only Theorem 2.1, and the text states the result is 'analogous to Proposition 3.1 in Gu et al. (2026)' with the two lower-bound components following 'similarly' from Balakrishnan et al. (2023) and Gu et al. (2026). This is not a routine combination: the product misspecification bound and the sparse-linear slog(p)/n bound must be realized simultaneously in a single experiment under the same orthonormal dictionary, boundedness constraints, and constant conditional variance E[u^2|X]≡v. If the two constructions only yield a maximum rather than a sum, the advertised additive rate is not established. Because 'unimprovable' is the headline contribution, this missing proof is load-bearing and must be supplied or replaced by a precise reference
  2. [Section 2.2 (Eq. (2.3), Corollary 2.2)] The estimator is not adaptive: it requires knowledge of the ordering between δ_appr_μ and δ_appr_π, and λ must be tuned from the full error budgets. Corollary 2.2's choice λ=(δ_stoc_n/(δ_max+δ_stoc_n))^2 depends on these budgets, and the branch choice in (2.3) depends on the ordering. The manuscript is transparent about this in Section 2.2, but the abstract's claim that the improvement over DML is obtained 'at no extra cost or assumption' overstates the situation, since DML does not require the analyst to know these budgets or the ordering in order to implement the estimator. No adaptive procedure or sensitivity analysis is provided. The authors should either develop a data-driven choice or explicitly qualify the practical scope of the advertised rate.
minor comments (4)
  1. [Abstract and Section 2] Notation is inconsistent: the abstract uses δ_s and δ_a, while the body uses δ_stoc_n and δ_appr_μ, δ_appr_π. Please unify the notation throughout.
  2. [Section 5, Proposition 5.1] Proposition 5.1 is a restatement of Proposition S2.3 of Gu et al. (2026), and the proof is deferred to that paper. For a standalone journal submission, either include the proof or state the precise lemma with a full reference so the reader can verify the initial-estimate bounds without consulting a separate preprint.
  3. [Theorem 2.1 and Corollary 2.2] The probability bounds differ: Theorem 2.1 states 1−2e^{−t}−2e^{−n(δ_stoc_n)^2}, while Corollary 2.2 states 1−3e^{−t}−2e^{−n(δ_stoc_n)^2}. Please reconcile the constants.
  4. [Corollary 3.2, Eq. (3.7)] The expression (δ_appr_μ δ_π + δ_μ^2) ∧ (δ_appr_π δ_μ + δ_π^2) is potentially ambiguous. Add explicit parentheses: (δ_appr_μ δ_π + δ_μ^2) ∧ (δ_appr_π δ_μ + δ_π^2).

Circularity Check

0 steps flagged

No circular reduction found; the central upper bound is self-contained, but the matching lower bound is stated without proof via analogy to prior work (rigor gap, not circularity).

full rationale

I walked the claimed derivation chain. The main upper bound (Corollary 2.2) follows from Theorem 2.1, whose proof in Section 5 is essentially self-contained; the only imported technical piece is Proposition 5.1, explicitly a restatement of Proposition S2.3 in the authors' own Gu et al. (2026). That initial-estimate bound is a standard empirical-process result and does not already contain the target rate, so citing it is not a circular reduction. The matching lower bound, Proposition 2.3, is asserted with the sentence 'The following proposition is analogous to Proposition 3.1 in Gu et al. (2026)' and is not proved in the manuscript. This is a serious omitted-proof/rigor issue for the headline optimality claim, but it is not circularity: there is no displayed derivation in which the lower bound reduces to its inputs by construction, nor is a fitted parameter renamed as a prediction. Similarly, the TAME-2 comparison in Corollary 3.2 is a direct application of the oracle-type inequality rather than a renaming of the DML result. The requirement in Section 2.2 that the estimator know the ordering of the error budgets and tune lambda accordingly is an explicit oracle assumption, not a hidden fit of the target quantity. Because the central derivation does not assume the result it claims to establish, the score reflects only minor self-citation in technical lemmas and the lower-bound analogy, not load-bearing circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 6 axioms · 0 invented entities

The central contribution is a new statistical estimator and rate theory. It relies on standard learning-theoretic machinery plus a few domain assumptions that are clearly stated. The only ad hoc element is the oracle knowledge of error budgets, which is explicitly assumed. No new physical or mathematical entities are introduced.

free parameters (1)
  • λ (tuning parameter) = λ = (δ_stoc_n / (δ_appr_max + δ_stoc_n))^2 in Corollary 2.2; oracle choices in Theorems 3.1 and 4.1
    The estimator rate depends on λ through the trade-off between local editing and adversarial moment calibration. The theory assumes λ is set using the (known) error budgets.
axioms (6)
  • domain assumption Boundedness and nondegeneracy of the model (Condition 2.1): |β0|, ||μ0||∞, ||π0||∞, |ε|, |u| ≤ c, and E[u^2] ≥ 1/c.
    Needed for the lower truncation in S2-μ/S2-π, concentration inequalities, and identifiability. Used throughout Section 2 and extensions.
  • domain assumption The black-box class G has stochastic error δ_stoc_n defined via localized Rademacher complexity (Condition 2.2).
    This is the core formalization of 'black-box learnability'. It is used to control the adversarial moment constraints and the initial estimation error.
  • ad hoc to paper The error budgets (δ_appr_μ, δ_appr_π, δ_stoc_n) and the ordering of δ_appr_μ vs δ_appr_π are known to the analyst.
    The estimator chooses between the μ- and π-branches and sets λ based on these budgets. This information is not automatically available in practice and is a key limitation.
  • domain assumption For the generic TAME procedure, the reference class G0 must contain the estimated easier nuisance: bμ ∈ G0 (O-branch) or bπ ∈ G0 (R-branch), as in Condition 3.1.
    The adversarial moment calibration makes sense only if the easier nuisance lies in the calibration class. This restricts 'any initial black-box estimates' to those compatible with G0.
  • domain assumption Initial estimates are independent of the inference split D1.
    Required for conditional mean-zero arguments in the proof (e.g., Lemma 5.3 and Lemma 5.5). This is standard sample-splitting.
  • standard math Standard empirical process inequalities and concentration bounds for local Rademacher complexities.
    Used in Propositions 5.1–5.4 and the proofs of the oracle-type inequalities without full derivations.

pith-pipeline@v1.3.0-alltime-deepseek · 3036 in / 9115 out tokens · 206379 ms · 2026-08-01T07:08:18.072930+00:00 · methodology

0 comments
read the original abstract

Consider the partial linear model $Y = \mu_0(X) + \beta_0 \cdot T + \varepsilon$ and $T = \pi_0(X) + u$ in the structure-agnostic setting, where we are blind to the structure $\mu_0$ and $\pi_0$ and estimate the nuisances by a black-box hypothesis class. The learnability of the class is characterized by the estimation error $\delta_s$ in the absence of model misspecification and its $L_2$ mis-specification error $\delta_{a, \mu}$ and $\delta_{a, \pi}$ for $\mu_0$ and $\pi_0$, respectively. We propose a novel estimator of the target linear coefficient $\theta_0 = \beta_0$ with error rate \[ \frac{1}{\sqrt{n}} + \delta_{a, \mu} \cdot \delta_{a, \pi} + [\delta_s]^2. \] A matching lower bound is also established, implying that this rate is unimprovable. Compared with the product rate yielded by double machine learning (DML), our estimator removes the suboptimal term $\max(\delta_{a, \mu}, \delta_{a, \pi})\cdot \delta_s$ at no extra cost or assumption. Building on the underlying insights, which are neither tailored to the one-learner setting nor the partial linear model, we propose Transductive Adversarial Moment-calibrated Editing (TAME), which locally edits debiasing weights induced by black-box regression estimates on the inference sample through adversarial conditional moment calibration. TAME can be combined with any initial black-box estimates and can strictly improve on DML guarantees when the nuisance difficulties are imbalanced. We discuss how to fully exploit the advantages introduced by TAME, including the gains from using two learners, the resulting under-smoothing principle for model selection, and extensions to other linear functional estimation problems.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 7 linked inside Pith

  1. [1]

    & Chen, X

    Ai, C. & Chen, X. (2003). Efficient estimation of models with conditional moment restrictions containing unknown functions. Econometrica , 71(6), 1795--1843

    2003

  2. [2]

    W., & Wager, S

    Athey, S., Imbens, G. W., & Wager, S. (2018). Approximate residual balancing: debiased inference of average treatment effects in high dimensions. Journal of the Royal Statistical Society Series B: Statistical Methodology , 80(4), 597--623

    2018

  3. [3]

    H., & Wasserman, L

    Balakrishnan, S., Kennedy, E. H., & Wasserman, L. (2023). The fundamental limits of structure-agnostic functional estimation. arXiv preprint arXiv:2305.04116

    Pith/arXiv arXiv 2023

  4. [4]

    Bartlett, P., Bousquet, O., & Mendelson, S. (2005). Local rademacher complexities. Annals of Statistics , 33(4), 1497--1537

    2005

  5. [5]

    & Kohler, M

    Bauer, B. & Kohler, M. (2019). On deep learning as a remedy for the curse of dimensionality in nonparametric regression. The Annals of Statistics , 47(4), 2261--2285

    2019

  6. [6]

    J., Ritov, Y., & Tsybakov, A

    Bickel, P. J., Ritov, Y., & Tsybakov, A. B. (2009). Simultaneous analysis of Lasso and Dantzig selector . The Annals of Statistics , 37(4), 1705 -- 1732

    2009

  7. [7]

    H., Dukes, O., & Balakrishnan, S

    Bonvini, M., Kennedy, E. H., Dukes, O., & Balakrishnan, S. (2024). Doubly-robust inference and optimality in structure-agnostic models with smoothness. arXiv preprint arXiv:2405.08525

    Pith/arXiv arXiv 2024

  8. [8]

    K., & Zhu, Y

    Bradic, J., Chernozhukov, V., Newey, W. K., & Zhu, Y. (2019). Minimax semiparametric learning with approximate sparsity. arXiv preprint arXiv:1912.12213

    Pith/arXiv arXiv 2019

  9. [9]

    Bradic, J., Fan, J., & Zhu, Y. (2022). Testability of high-dimensional linear models with nonsparse structures. Annals of statistics , 50(2), 615

    2022

  10. [10]

    Bruns-Smith, D., Dukes, O., Feller, A., & Ogburn, E. L. (2025). Augmented balancing weights as linear regression. Journal of the Royal Statistical Society Series B: Statistical Methodology , (pp.\ qkaf019)

    2025

  11. [11]

    Cai, T. T. & Guo, Z. (2017). Confidence intervals for high-dimensional linear regression: Minimax rates and adaptivity . The Annals of Statistics , 45(2), 615 -- 646

    2017

  12. [12]

    & Tao, T

    Candes, E. & Tao, T. (2007). The dantzig selector: Statistical estimation when p is much larger than n. The Annals of Statistics , (pp.\ 2313--2351)

    2007

  13. [13]

    J., Fan, J., Gijbels, I., & Wand, M

    Carroll, R. J., Fan, J., Gijbels, I., & Wand, M. P. (1997). Generalized partially linear single-index models. Journal of the American Statistical Association , 92(438), 477--489

    1997

  14. [14]

    & Wainwright, M

    Celentano, M. & Wainwright, M. J. (2023). Challenges of the inconsistency regime: Novel debiasing methods for missing data models. arXiv preprint arXiv:2309.01362

    Pith/arXiv arXiv 2023

  15. [15]

    Chen, X. (2007). Large sample sieve estimation of semi-nonparametric models. Handbook of econometrics , 6, 5549--5632

    2007

  16. [16]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., & Newey, W. (2017). Double/debiased/neyman machine learning of treatment effects. American Economic Review , 107(5), 261--265

    2017

  17. [17]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal , 21(1), C1--C68

    2018

  18. [18]

    K., & Singh, R

    Chernozhukov, V., Newey, W. K., & Singh, R. (2022a). Automatic debiased machine learning of causal and structural effects. Econometrica , 90(3), 967--1027

  19. [19]

    K., & Singh, R

    Chernozhukov, V., Newey, W. K., & Singh, R. (2022b). Debiased machine learning of global and local parameters using regularized riesz representers. The Econometrics Journal , 25(3), 576--601

  20. [20]

    K., Singh, R., & Syrgkanis, V

    Chernozhukov, V., Newey, W. K., Singh, R., & Syrgkanis, V. (2026). Adversarial estimation of riesz representers. Journal of the American Statistical Association , (pp.\ 1--12)

    2026

  21. [21]

    Donald, S. G. & Newey, W. K. (1994). Series estimation of semilinear models. Journal of Multivariate Analysis , 50(1), 30--40

    1994

  22. [22]

    F., Granger, C

    Engle, R. F., Granger, C. W., Rice, J., & Weiss, A. (1986). Semiparametric estimates of the relation between weather and electricity sales. Journal of the American statistical Association , 81(394), 310--320

    1986

  23. [23]

    Fan, J. & Gu, Y. (2024). Factor augmented sparse throughput deep relu neural networks for high dimensional regression. Journal of the American Statistical Association , 119(548), 2680--2694

    2024

  24. [24]

    Fan, J., Gu, Y., & Zhou, W.-X. (2024). How do noise tails impact on deep ReLU networks? The Annals of Statistics , 52(4), 1845 -- 1871

    2024

  25. [25]

    & Huang, T

    Fan, J. & Huang, T. (2005). Profile likelihood inferences on semiparametric varying-coefficient partially linear models. Bernoulli , 11(6), 1031--1057

    2005

  26. [26]

    Fan, J., Imai, K., Lee, I., Liu, H., Ning, Y., & Yang, X. (2022). Optimal covariate balancing conditions in propensity score estimation. Journal of Business & Economic Statistics , 41(1), 97--110

    2022

  27. [27]

    Foster, D. J. & Syrgkanis, V. (2023). Orthogonal statistical learning. The Annals of Statistics , 51(3), 879--908

    2023

  28. [28]

    Gu, Y., Fang, C., B \"u hlmann, P., & Fan, J. (2025). Causality pursuit from heterogeneous environments via neural adversarial invariance learning . The Annals of Statistics , 53(5), 2230 -- 2257

    2025

  29. [29]

    Gu, Y., Yin, Q., Cai, T., & Fan, J. (2026). Optimally taming biases in black-box models for efficient semiparametric estimation. arXiv preprint arXiv:2606.06368

    Pith/arXiv arXiv 2026

  30. [30]

    Hellerstein, J. K. & Imbens, G. W. (1999). Imposing moment restrictions from auxiliary data by weighting. Review of Economics and Statistics , 81(1), 1--14

    1999

  31. [31]

    Hirshberg, D. A. & Wager, S. (2021). Augmented minimax linear estimation. The Annals of Statistics , 49(6), 3206--3227

    2021

  32. [32]

    & Ratkovic, M

    Imai, K. & Ratkovic, M. (2014). Covariate balancing propensity score. Journal of the Royal Statistical Society Series B: Statistical Methodology , 76(1), 243--263

    2014

  33. [33]

    & Montanari, A

    Javanmard, A. & Montanari, A. (2014). Confidence intervals and hypothesis testing for high-dimensional regression. Journal of Machine Learning Research , 15(82), 2869--2909

    2014

  34. [34]

    & Montanari, A

    Javanmard, A. & Montanari, A. (2018). Debiasing the lasso: Optimal sample size for Gaussian designs . The Annals of Statistics , 46(6A), 2593 -- 2622

    2018

  35. [35]

    Jin, J., Mackey, L., & Syrgkanis, V. (2025). It's hard to be normal: The impact of noise on structure-agnostic estimation. arXiv preprint arXiv:2507.02275

    arXiv 2025

  36. [36]

    & Langer, S

    Kohler, M. & Langer, S. (2021). On the rate of convergence of fully connected deep neural network regression estimates. The Annals of Statistics , 49(4), 2231--2249

    2021

  37. [37]

    Koltchinskii, V. (2006). Local Rademacher complexities and oracle inequalities in risk minimization . The Annals of Statistics , 34(6), 2593 -- 2656

    2006

  38. [38]

    Kong, J. (2025). On the asymptotics of the minimax linear estimator. arXiv preprint arXiv:2510.16661

    arXiv 2025

  39. [39]

    K., & Robins, J

    Liu, L., Mukherjee, R., Newey, W. K., & Robins, J. M. (2017). Semiparametric efficient empirical higher order influence function estimators. arXiv preprint arXiv:1705.07577

    arXiv 2017

  40. [40]

    & N \'e d \'e lec, \'E

    Massart, P. & N \'e d \'e lec, \'E . (2006). Risk bounds for statistical learning. Ann. Statist. , 34(1), 2326--2366

    2006

  41. [41]

    Ning, Y., Sida, P., & Imai, K. (2020). Robust estimation of causal effects via a high-dimensional covariate balancing propensity score. Biometrika , 107(3), 533--554

    2020

  42. [42]

    Rice, J. (1986). Convergence rates for partially splined models. Statistics & probability letters , 4(4), 203--208

    1986

  43. [43]

    Robins, J., Li, L., 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 , volume 2 (pp.\ 335--422). Institute of Mathematical Statistics

    2008

  44. [44]

    T., & van der Vaart, A

    Robins, J., Li, L., Tchetgen, E. T., & van der Vaart, A. (2016). Technical report: Higher order influence functions and minimax estimation of nonlinear functionals. arXiv preprint arXiv:1601.05820

    Pith/arXiv arXiv 2016

  45. [45]

    M., Rotnitzky, A., & Zhao, L

    Robins, J. M., Rotnitzky, A., & 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

  46. [46]

    Schmidt-Hieber, J. (2020). Nonparametric regression using deep neural networks with relu activation function (with discussion). The Annals of Statistics , 48(4), 1875--1921

    2020

  47. [47]

    Tan, Z. (2020). Model-assisted inference for treatment effects using regularized calibrated estimation with high-dimensional data. Annals of Statistics , 48(2), 811--837

    2020

  48. [48]

    Van de Geer, S., B \"u hlmann, P., Ritov, Y., & Dezeure, R. (2014). On asymptotically optimal confidence regions and tests for high-dimensional models. The Annals of Statistics , 42(3), 1166--1202

    2014

  49. [49]

    van der Laan, L., Luedtke, A., & Carone, M. (2024). Doubly robust inference via calibration. arXiv preprint arXiv:2411.02771

    Pith/arXiv arXiv 2024

  50. [50]

    Van der Laan, M. J. (2014). Targeted estimation of nuisance parameters to obtain valid statistical inference. The international journal of biostatistics , 10(1), 29--57

    2014

  51. [51]

    & Chervonenkis, A

    Vapnik, V. & Chervonenkis, A. Y. (1971). On the uniform convergence of relative frequencies of events to their probabilities. Theory of Probability & Its Applications , 16(2), 264--280

    1971

  52. [52]

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

    2019

  53. [53]

    & Shah, R

    Wang, Y. & Shah, R. D. (2024). Debiased inverse propensity score weighting for estimation of average treatment effects with high-dimensional confounders . The Annals of Statistics , 52(5), 1978 -- 2003

    2024

  54. [54]

    & Zhang, S

    Zhang, C.-H. & Zhang, S. S. (2014). Confidence intervals for low dimensional parameters in high dimensional linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 76(1), 217--242

    2014

  55. [55]

    Zhao, Q. (2019). Covariate balancing propensity score by tailored loss functions . The Annals of Statistics , 47(2), 965 -- 993

    2019

  56. [56]

    Zubizarreta, J. R. (2015). Stable weights that balance covariates for estimation with incomplete outcome data. Journal of the American Statistical Association , 110(511), 910--922

    2015