Nonparametric "rich covariates" without saturation

Federico Martellosio; J.M.C. Santos Silva; Ludgero Glorias

REVIEW 3 major objections 3 minor 48 references

Nonparametric "rich covariates" without saturation

T0 review · 3 major / 3 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 Two nonparametric fixes restore a causal interpretation for linear IV with covariates.

desk verdict Worth a serious referee, but the variance formulas in Theorems 2 and 4 need correction before the inference claims are usable. read the letter →

arxiv 2505.21213 v2 pith:7GZP63DS submitted 2025-05-27 econ.EM

Ludgero Glorias , Federico Martellosio , J.M.C. Santos Silva This is my paper

classification econ.EM MSC 62G0862G2062P20

keywords instrumentalvariablesrichcovariateslocalaveragetreatmenteffectsemiparametricestimationkernelregressionneuralnetworksinstrumentresidualcontrolfunction

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

Linear instrumental variables estimators carry a clean causal interpretation only when the rich-covariates condition holds, meaning the instrument's conditional expectation given the covariates is a linear function of the controls included in the regression. This paper shows how to make that condition hold by construction, without saturating the model or assuming the instrument is randomly assigned, by first estimating $\zeta_0(c)=E[z\mid c]$ nonparametrically. Either using the residual $z-\hat\zeta(c)$ as the instrument or adding $\hat\zeta(c)$ as a regressor forces the rich-covariates condition to be satisfied, and both estimators share the same probability limit. When the first-step estimate is a kernel regression satisfying the paper's conditions, the treatment coefficient is $\sqrt{n}$-consistent and asymptotically normal, with the limit equal to a positively weighted average of conditional complier treatment effects. Simulations and an empirical application indicate that a neural-network first step works well in higher dimensions and can avoid a cross-fitting failure that arises with sparse binary controls.

What carries the argument

The load-bearing object is the conditional expectation of the instrument, $\zeta_0(c)=E[z\mid c]$. The theory estimates it with a higher-order Nadaraya-Watson kernel; the simulations use a shallow ReLU neural network. The two constructions then force rich covariates exactly: $z-\hat\zeta(c)$ is mean-independent of every function of $c$, so its conditional expectation is the constant zero, and adding $\hat\zeta(c)$ as a regressor makes the instrument's conditional expectation linear in the regressors by definition. The asymptotic mechanism is undersmoothing—choosing the bandwidth so that $nh^{2d}/(\ln n)^2\to\infty$ and $nh^{2m}\to 0$, which requires $m>d$—so that first-step bias disappears faster than $\sqrt{n}$ while variance remains controlled, in the standard two-step semiparametric framework.

What would settle it

Run Monte Carlo replications of the kernel-first-step estimator under the paper's assumptions, form 95% confidence intervals using the variance estimator of Theorem 2, and check empirical coverage at a sample size like $n=8000$. If coverage is systematically below 95%, the $\sqrt{n}$-consistency claim fails; if the same experiment with a neural-network first step on a high-dimensional, non-smooth $\zeta_0$ shows bias that does not shrink at the $\sqrt{n}$ rate, the practical recommendation for neural networks is also in doubt.

Watch

Extended reading notes

Core claim

The paper's central claim is that two feasible two-step linear IV estimators—one that instruments the endogenous binary treatment with $z-\hat\zeta(c)$, and one that adds $\hat\zeta(c)$ to the regressor list—satisfy the rich-covariates condition asymptotically without parametric assumptions on $E[z\mid c]$. Theorems 1 and 3 show that, under the stated kernel, bandwidth, smoothness, integrability, and identification assumptions, both estimators are $\sqrt{n}$-consistent and asymptotically normal, with the variance estimators of Theorems 2 and 4 consistent. The probability limit of the treatment coefficient is $\alpha_{\mathrm{rich}} = E[\omega(c)\,EC(y(1)-y(0)\mid c)]$, with overlap weights $\omega(c)=\mathrm{Cov}(z,t\mid c)/E[\mathrm{Cov}(z,t\mid c)]$, so the estimand is a positively weighted average of conditional complier effects and is weakly causal under monotonicity. The formal results cover only the kernel first step; the neural-network implementation is supported by simulation evidence, with its asymptotic theory explicitly left for future work.

Load-bearing premise

The $\sqrt{n}$-consistency theorems require an undersmoothed higher-order kernel bandwidth and a true conditional expectation $\zeta_0$ that is $m$-times continuously differentiable with $m$ exceeding the number of covariates $d$; the paper's recommended neural-network first step is not covered by those theorems, so its guarantees in that implementation rest on simulations.

Editorial extensions

If this is right

Applied researchers can enforce rich covariates without saturation by running one nonparametric first-step regression and then using either $z-\hat\zeta(c)$ as the instrument or $\hat\zeta(c)$ as an extra regressor.
Under the kernel conditions, the resulting treatment-coefficient estimates are $\sqrt{n}$-consistent and asymptotically normal, so the usual t-statistics and confidence intervals are available.
Both estimators identify the same positively weighted complier average, $\alpha_{\mathrm{rich}}$, so the causal interpretation does not depend on which construction is used.
In the simulations the instrument-residual version is less sensitive to the number of controls and to the first-step estimator than the control-function version, and both improve on a parametric instrument-residual alternative when the parametric model is misspecified.
In the empirical application the proposed estimators reverse the sign of the LIVE estimate and line up with the saturated-model estimate, suggesting that enforcing rich covariates can materially change conclusions.

Reading between the lines

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

Extending the asymptotic theory to the neural-network first step would turn the paper's practical recommendation into a theorem; that is the natural next step and would close the gap between what is proved and what is recommended.
Because the control-function estimator's asymptotic variance depends on the coefficient on $\hat\zeta(c)$, the choice of other regressors changes efficiency through an additional channel; this may explain the instrument-residual estimator's better performance in the simulations.
The sign reversal in the empirical application implies that published linear-IV results with many sparse binary controls may be sensitive to the rich-covariates assumption; re-examining such applications with these estimators is a direct way to test how much the assumption matters.
The framework should extend to multi-valued or continuous instruments and to overidentified models using existing estimand results, which would widen the applicability beyond binary treatment and binary instrument.

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.

Desk Editor's Note

read the letter

Candid take: this paper has a real, useful idea and one substantive flaw in the inference formulas. The nonparametric-first-step versions of the instrument-residual and control-function LIVE estimators are a genuine step beyond Lee (2021) and Borusyak and Hull (2021), which rely on parametric or simulation-based first steps. The kernel asymptotics via Newey and McFadden are plausible, and the simulations plus the Dube–Harish illustration make the practical case. The critique of DML cross-fitting with sparse binary controls is also an important observation; the empirical demonstration that cross-fitting fails when a dummy is nonzero in only three clusters is credible. Credit where due: the paper is clearly written, the estimand derivation is clean, and the limitations (neural-network first step without theorem, curse of dimensionality for kernels) are flagged in the text rather than hidden.

The soft spot is real and central to inference. Theorem 2's Ω̂ has a leading minus sign, which makes it negative semidefinite. More substantively, φ̂ replaces E[ε|c] with ε̂_i. The influence function in Theorem 1 has first component z*(ε − E[ε|c]); the estimator as written uses z*ε̂ + z*ε̂ = 2z*ε̂, and never estimates E[ε|c]. Unless E[ε|c] = 0—which the paper explicitly does not assume—the proposed variance estimator is inconsistent. The same structure appears in Theorem 4. The empirical section sidesteps the issue by using cluster-robust IV standard errors that ignore the first step, but that means the asymptotic inference claim in Section 4 is not operational as written. This is fixable: add a nonparametric estimator of E[ε|c] and correct the sign, but it needs to be done before practitioners can trust the standard errors.

The neural-network recommendation is another gap, but the paper honestly says that theory is left for future work. The kernel version does enough to establish √n-consistency as a feasibility result. The simulation design is clever—d increases while the DGP is fixed—and the results are informative, including the deterioration of the kernel version as d grows.

Bottom line: applied econometricians and methodologists working on IV with heterogeneous effects will get real value from this paper. The identification and point-estimation contributions are solid; the variance formulas need correction and re-proof. A careful author can fix this without changing the paper's core message, and the paper deserves a serious referee rather than a desk reject. I would bring it to a reading group and would cite the identification and cross-fitting critique, but I would not use the current standard-error formulas.

Referee Report

3 major / 3 minor

Summary. The paper proposes two two-step linear IV estimators that enforce the rich-covariates condition without a saturated model or unconditional randomization: the instrument-residual estimator, which uses z − ζ̂(c) as the instrument, and the control-function estimator, which adds ζ̂(c) as a regressor. It shows that both estimators target αrich = E[ω(c)EC(y(1)−y(0)|c)], proves √n-consistency and asymptotic normality for a kernel first step under Assumptions 1–6, and reports simulations with kernel and neural-network first steps and an application to Dube and Harish (2020). The point-estimator proofs are plausibly derived from Newey and McFadden (1994), but the variance estimators in Theorems 2 and 4 are not valid as written.

Significance. The paper addresses an important practical gap: it offers alternatives to saturated specifications for ensuring rich covariates, with an estimand that has a clear causal interpretation under strong monotonicity, and it connects to existing Lee/Kim-Lee and DML estimators. The simulation design is thoughtful, and the honest discussion of the neural-network first step (explicitly not covered by the theorems) is a strength. The proof strategy via Newey and McFadden (1994) is appropriate and the identification argument appears sound. The blocking issue is inferential: the proposed variance estimators contain a sign error and an infeasible or inconsistent plug-in for the adjustment term, so the paper's claim of a consistent variance estimator is not yet supported.

major comments (3)

[Section 4.2, Theorem 2] The displayed estimator is Ω̂ := −(1/n)∑_{i=1}^n τ̂_i τ̂_i′, with a leading minus sign. This makes Ω̂ negative semidefinite, so it cannot be a consistent estimator of the covariance matrix Ω, which is positive semidefinite. The same minus sign appears in the control-function variance estimator in Section 4.3. If this is a typographical error, it must still be corrected before the formula is usable.
[Section 4.2–4.3, Theorems 2 and 4] The plug-in for the first-step adjustment term is inconsistent with the influence function. Theorem 1 defines φ=(−z*(ζ0)E[ε|c],0_k')′, so the first component of q(ζ0)ε+φ is z*(ζ0)(ε−E[ε|c]). The estimator in Theorem 2 sets φ̂_i=(z*_i(ζ̂(c_i))ε̂_i,0_k')′, which uses the regression residual ε̂_i in place of E[ε_i|c_i]. Since E[ε|c] is not assumed to be zero (the whole setting allows misspecification), τ̂_i has first component 2z*_iε̂_i rather than z*_i(ε̂_i−E[ε_i|c_i]); this changes the variance being estimated. No estimator of E[ε|c] is supplied, and the proof of Theorem 2 verifies Lipschitz conditions for the infeasible quantity −E[ε(β)|c]z*(ζ), not for the estimator actually implemented. Theorem 4 has the same problem, with ε̂_i^* used in place of E[ε*|c]. The consistency claim for the variance estimator is therefore not established.
[Section 6, footnote 26] The empirical illustration reports cluster-robust IV standard errors that 'do not account for the fact that the estimator relies on a preliminary first step.' This means the application does not use the Theorem 2/Theorem 4 variance estimators, so the paper provides no practical demonstration of the proposed inference procedure. The simulations in Section 5 also do not report coverage rates or rejection rates. Unless a feasible first-step-robust variance estimator is supplied and implemented, the paper's asymptotic-inference claim remains non-operational.

minor comments (3)

[Section 5] The bandwidth formula for the Nadaraya-Watson first step is chosen by experimentation, and the simulation conclusions for the kernel version may depend on this calibration; a sentence on sensitivity would help.
[Abstract and Section 7] The abstract and conclusions recommend neural networks for the first step, while Section 3.3 and Section 7 correctly state that the theorems cover only the kernel first step; this scope caveat should be made more prominent in the abstract so that readers do not take the neural-network version as theorem-backed.
[Table 2 and surrounding text] The row labels are informative, but the text should clarify that the 'DML neural net no cross-fitting' estimate is a comparison method rather than one of the proposed estimators, and the footnote on standard errors should state explicitly that none of the reported standard errors correspond to the variance formulas in Theorems 2 and 4.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the rich-covariates condition is enforced by construction, the estimand is benchmarked externally, and the kernel asymptotics rely on standard Newey–McFadden theorems with no self-citation chain.

full rationale

The paper's two estimators are constructed so that the rich-covariates condition holds algebraically at the population level: the instrument residual z − ζ0(c) is mean-independent of any function of c, and adding ζ0(c) as a regressor makes L[z | r, ζ0(c)] equal to ζ0(c). This is an explicit construction, not a fitted parameter later relabeled as a prediction. The target estimand αrich is taken from Lee (2021) and Blandhol et al. (2025), with the equivalence shown in Appendix A; it is not derived from the paper's own fitted values. In the application, the proposed estimators are compared against the externally computed saturated-model estimate (−0.509) reported by Blandhol et al. (2022), and no parameter is fit to that benchmark. Theorems 1 and 3 are proven by verifying the conditions of Newey and McFadden (1994), an independent reference theorem, with explicit derivative and bound calculations; no uniqueness theorem from the authors' own prior work is invoked. The reference list contains no self-citations that are load-bearing. The skeptical observation about the negative sign and plug-in error in the variance estimator in Theorems 2 and 4 is a substantive mathematical correctness concern, not a circularity concern: even if the variance estimator is inconsistent as written, the √n-consistency claim does not reduce to its own inputs. Similarly, footnote 26's use of standard errors that ignore the first step is an implementation disclosure, not a circular derivation. Finally, the paper explicitly limits its asymptotic theory to the kernel first step and states that extension to neural networks is left for future work, so no unproven claim is smuggled in through self-citation or ansatz. Overall, the derivation chain is self-contained against external benchmarks and contains no circular step.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claims rest on standard semiparametric assumptions, monotonicity, and conditional exogeneity, plus the unproven accuracy of the neural network first step when used in the recommended implementation.

free parameters (2)

Kernel bandwidth constants in simulations = (1.1 + 0.725d) σ̂ n^{-1/(2d+1)}
Tuning rule chosen after experimentation for the Nadaraya-Watson first step; it affects finite-sample performance but is not fitted to the target estimand.
Neural network architecture = Single hidden layer, 100 ReLU nodes, Adam optimizer, default pystacked options
Chosen for simulations and application; no theoretical optimality guarantee and no asymptotic results are provided for this first-step estimator.

assumptions (6)

domain assumption Conditional instrument exogeneity: (y(0), y(1), t(0), t(1)) ⊥⊥ z | c
Assumed in Section 2; required for any causal interpretation of the IV estimand.
domain assumption Strong monotonicity: sign of Cov(z,t|c) is the same for all c
Assumed in footnote 5; needed for αrich to be a positively weighted average of conditional complier effects.
domain assumption Compact support of c with density bounded away from zero and infinity (Assumption 3)
Standard for kernel regression; used throughout the proofs in Appendix C.
domain assumption ζ0 is m-times continuously differentiable with m > d (Assumption 4)
Needed for bias decay under undersmoothing; this is a strong smoothness condition in high dimensions.
domain assumption Higher-order kernel with vanishing moments up to m-1 and bandwidth conditions (Assumptions 1 and 2)
Required for √n-consistency; higher-order kernels can take negative values and are rarely used in applied work.
standard math Moment existence and nonsingularity (Assumptions 5, 6, and 6*)
Regularity conditions for asymptotic normality and for the identification of the moment condition.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Nonparametric "rich covariates" without saturation." pith.science (2026). https://pith.science/paper/7GZP63DS

@misc{pith2026250521213,
  author       = {Pith},
  title        = {Pith review of: Nonparametric "rich covariates" without saturation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GZP63DS}},
  note         = {Machine review of arXiv:2505.21213}
}

read the original abstract

We consider two nonparametric approaches to ensure that linear instrumental variables estimators satisfy the rich-covariates condition emphasized by Blandhol et al. (2025), even when the instrument is not unconditionally randomly assigned and the model is not saturated. Both approaches start with a nonparametric estimate of the expectation of the instrument conditional on the covariates, and ensure that the rich-covariates condition is satisfied either by using as the instrument the difference between the original instrument and its estimated conditional expectation, or by adding the estimated conditional expectation to the set of regressors. We derive asymptotic properties when the first step uses kernel regression, and assess finite-sample performance in simulations where we also use neural networks in the first step. Finally, we present an empirical illustration that highlights some significant advantages of the proposed methods.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 45 canonical work pages

[1]

(2003): Semiparametric Instrumental Variable Estimation of Treatment Response Models, Journal of Econometrics, 113, 231--263

Abadie, A. (2003): Semiparametric Instrumental Variable Estimation of Treatment Response Models, Journal of Econometrics, 113, 231--263

work page 2003
[2]

Ahrens, A., C. B. Hansen, and M. E. Schaffer (2023): pystacked: Stacking Generalization and Machine Learning in Stata , The Stata Journal, 23, 909--931

work page 2023
[3]

Ahrens, A., C. B. Hansen, M. E. Schaffer, and T. Wiemann (2024): ddml: Double/Debiased Machine Learning in Stata , The Stata Journal, 24, 3--45

work page 2024
[4]

Angrist, J. D., K. Graddy, and G. W. Imbens (2000): The Interpretation of Instrumental Variables Estimators in Simultaneous Equations Models with an Application to the Demand for Fish, The Review of Economic Studies, 67, 499--527

work page 2000
[5]

Angrist, J. D. and G. W. Imbens (1995): Two-Stage Least Squares Estimation of Average Causal Effects in Models with Variable Treatment Intensity, Journal of the American Statistical Association, 90, 431--442

work page 1995
[6]

Angrist, J. D. and J.-S. Pischke (2009): Mostly Harmless Econometrics: An Empiricist's Companion, Princeton (NJ): Princeton University Press

work page 2009
[7]

--- -.1pt --- -.1pt --- (2010): The Credibility Revolution in Empirical Economics: How Better Research Design is Taking the Con Out of Econometrics, Journal of Economic Perspectives, 24, 3--30

work page 2010
[8]

(2017): Breaking the Curse of Dimensionality with Convex Neural Networks, Journal of Machine Learning Research, 18, 1--53

Bach, F. (2017): Breaking the Curse of Dimensionality with Convex Neural Networks, Journal of Machine Learning Research, 18, 1--53

work page 2017

Show all 48 references

[9]

(1994): Approximation and Estimation Bounds for Artificial Neural Networks, Machine Learning, 14, 115--133

Barron, A. (1994): Approximation and Estimation Bounds for Artificial Neural Networks, Machine Learning, 14, 115--133

1994
[10]

Bauer, B. and M. Kohler (2019): On Deep Learning as a Remedy for the Curse of Dimensionality in Nonparametric Regression, Annals of Statistics, 47, 2261--2285

2019
[11]

Bonney, M

Blandhol, C., J. Bonney, M. Mogstad, and A. Torgovitsky (2022): When is TSLS Actually LATE ? University of Chicago, Becker Friedman Institute for Economics Working Paper No. 2022-16

2022
[12]

--- -.1pt --- -.1pt --- (2025): When is TSLS Actually LATE ? NBER Working Paper No. w29709

2025
[13]

Borusyak, K. and P. Hull (2021): Non-Random Exposure to Exogenous Shocks: Theory and Applications, mimeo

2021
[14]

--- -.1pt --- -.1pt --- (2023): Nonrandom Exposure to Exogenous Shocks, Econometrica, 91, 2155--2185

2023
[15]

Kohler, S

Braun, A., M. Kohler, S. Langer, and H. Walk (2024): Convergence Rates for Shallow Neural Networks Learned by Gradient Descent, Bernoulli, 30, 475--502

2024
[16]

Chen, X. and H. White (1999): Improved Rates and Asymptotic Normality for Non- parametric Neural Network Estimators, IEEE Transactions on Information Theory, 45, 682--691

1999
[17]

Chetverikov, M

Chernozhukov, V., D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, and J. Robins (2018): Double/Debiased Machine Learning for Treatment and Structural Parameters, The Econometrics Journal, 21, C1--C68

2018
[18]

Dube, O. and S. Harish (2020): Queens, Journal of Political Economy, 128, 2579--2652

2020
[19]

Farrell, M. H., T. Liang, and S. Misra (2021): Deep Neural Networks for Estimation and Inference, Econometrica, 89, 181--213

2021
[20]

Liu, and J

Gao, Q., L. Liu, and J. S. Racine (2015): A Partially Linear Kernel Estimator for Categorical Data, Econometric Reviews, 34, 959--978

2015
[21]

Hansen, B. E. (2005): Exact Mean Integrated Squared Error of Higher Order Kernel Estimators, Econometric Theory, 21, 1031–1057

2005
[22]

Stinchcombe, and H

Hornik, K., M. Stinchcombe, and H. White (1989): Multilayer Feedforward Networks are Universal Aapproximators, Neural Networks, 2, 359--366

1989
[23]

Imbens, G. W. and J. D. Angrist (1994): Identification and Estimation of Local Average Treatment Effects, Econometrica, 62, 467--475

1994
[24]

and M.-J

Kim, B. and M.-J. Lee (2024): Instrument-Residual Estimator for Multi-valued Instruments Under Full Monotonicity, Statistics & Probability Letters, 213, 110187

2024
[25]

(1980): Existence of Moments of k-Class Estimators, Econometrica, 48, 241--249

Kinal, T. (1980): Existence of Moments of k-Class Estimators, Econometrica, 48, 241--249

1980
[26]

Kingma, D. P. and J. Ba (2015): Adam: A Method for Stochastic Optimization, arXiv:1412.6980v9

2015 arXiv
[27]

Kohler, M. and S. Langer (2021): On the Rate of Convergence of Fully Connected Deep Neural Network Regression Estimates, The Annals of Statistics, 49, 2231--2249

2021
[28]

(2013): Estimation in an Instrumental Variables Model With Treatment Effect Heterogeneity, Working Paper 2013-2, Princeton University

Kolesár, M. (2013): Estimation in an Instrumental Variables Model With Treatment Effect Heterogeneity, Working Paper 2013-2, Princeton University

2013
[29]

and M.-J

Lee, G. and M.-J. Lee (2025): Double-Debiasing Instrumental Variable Estimator with Machine-Learned Plug-In for Any Outcome and Heterogeneity, mimeo

2025
[30]

Lee, M.-J. (2021): Instrument Residual Estimator for any Response Variable with Endogenous Binary Treatment, Journal of the Royal Statistical Society Series B: Statistical Methodology, 83, 612--635

2021
[31]

Lee, M.-J. and C. Han (2024): Ordinary Least Squares and Instrumental-Variables Estimators for any Outcome and Heterogeneity, The Stata Journal, 24, 72--92

2024
[32]

Mogstad, M. and A. Torgovitsky (2024): Instrumental Variables with Unobserved Heterogeneity in Treatment Effects, in Handbook of Labor Economics, ed. by C. Dustmann and T. Lemieux, Amsterdam: Elsevier, vol. 5, 1--114

2024
[33]

Torgovitsky, and C

Mogstad, M., A. Torgovitsky, and C. Walters (2021): The Causal Interpretation of Two-Stage Least Squares with Multiple Instrumental Variables, American Economic Review, 111, 3663--3698

2021
[34]

Murphy, K. M. and R. H. Topel (1985): Estimation and Inference in Two-Step Econometric Models, Journal of Business & Economic Statistics, 3, 370--379

1985
[35]

Nadaraya, E. A. (1964): On Estimating Regression, Theory of Probability and Its Applications, 9, 141--142

1964
[36]

Newey, W. K. (1994): The Asymptotic Variance of Semiparametric Estimators, Econometrica, 62, 1349--1382

1994
[37]

--- -.1pt --- -.1pt --- (1997): Convergence Rates and Asymptotic Normality for Series Estimators, Journal of Econometrics, 79, 147--168

1997
[38]

Newey, W. K., F. Hsieh, and J. M. Robins (2004): Twicing Kernels and a Small Bias Property of Semiparametric Estimators, Econometrica, 72, 947--962

2004
[39]

Newey, W. K. and D. L. McFadden (1994): Large Sample Estimation and Hypothesis Testing, in Handbook of Econometrics, ed. by R. F. Engle and D. L. McFadden, Amsterdam: Elsevier, vol. 4, 2111--2245

1994
[40]

(1984): Econometric Issues in the Analysis of Regressions with Generated Regressors, International Economic Review, 25, 221--247

Pagan, A. (1984): Econometric Issues in the Analysis of Regressions with Generated Regressors, International Economic Review, 25, 221--247

1984
[41]

Varoquaux, A

Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay (2011): Scikit-learn: Machine Learning in P ython, Journal of Machine...

2011
[42]

Ramsey, J. B. (1969): Tests for Specification Errors in Classical Linear Least-Squares Regression Analysis, Journal of the Royal Statistical Society Series B: Statistical Methodology, 31, 350--371

1969
[43]

Robins, J. M., S. D. Mark, and W. K. Newey (1992): Estimating Exposure Effects by Modelling the Expectation of Exposure Conditional on Confounders, Biometrics, 48, 479--495

1992
[44]

(1988): Root-N-Consistent Semiparametric Regression, Econometrica, 56, 931--954

Robinson, P. (1988): Root-N-Consistent Semiparametric Regression, Econometrica, 56, 931--954

1988
[45]

(2020): Nonparametric Regression Using Deep Neural Networks with ReLU Activation Function, The Annals of Statistics, 48, 1875--1897

Schmidt-Hieber, J. (2020): Nonparametric Regression Using Deep Neural Networks with ReLU Activation Function, The Annals of Statistics, 48, 1875--1897

2020
[46]

Scott, D. W. (2015): Multivariate density estimation: theory, practice, and visualization, New York (NY): John Wiley & Sons

2015
[48]

--- -.1pt --- -.1pt --- (2024): When Should We (Not) Interpret Linear IV Estimands as LATE ? arXiv:2011.06695v7

2024 arXiv
[49]

(1964): Smooth Regression Analysis, Sankhy \=a : The Indian Journal of Statistics, Series A , 26, 359--372

Watson, G. (1964): Smooth Regression Analysis, Sankhy \=a : The Indian Journal of Statistics, Series A , 26, 359--372

1964

Pith tools

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

[1] [1]

(2003): Semiparametric Instrumental Variable Estimation of Treatment Response Models, Journal of Econometrics, 113, 231--263

Abadie, A. (2003): Semiparametric Instrumental Variable Estimation of Treatment Response Models, Journal of Econometrics, 113, 231--263

work page 2003

[2] [2]

Ahrens, A., C. B. Hansen, and M. E. Schaffer (2023): pystacked: Stacking Generalization and Machine Learning in Stata , The Stata Journal, 23, 909--931

work page 2023

[3] [3]

Ahrens, A., C. B. Hansen, M. E. Schaffer, and T. Wiemann (2024): ddml: Double/Debiased Machine Learning in Stata , The Stata Journal, 24, 3--45

work page 2024

[4] [4]

Angrist, J. D., K. Graddy, and G. W. Imbens (2000): The Interpretation of Instrumental Variables Estimators in Simultaneous Equations Models with an Application to the Demand for Fish, The Review of Economic Studies, 67, 499--527

work page 2000

[5] [5]

Angrist, J. D. and G. W. Imbens (1995): Two-Stage Least Squares Estimation of Average Causal Effects in Models with Variable Treatment Intensity, Journal of the American Statistical Association, 90, 431--442

work page 1995

[6] [6]

Angrist, J. D. and J.-S. Pischke (2009): Mostly Harmless Econometrics: An Empiricist's Companion, Princeton (NJ): Princeton University Press

work page 2009

[7] [7]

--- -.1pt --- -.1pt --- (2010): The Credibility Revolution in Empirical Economics: How Better Research Design is Taking the Con Out of Econometrics, Journal of Economic Perspectives, 24, 3--30

work page 2010

[8] [8]

(2017): Breaking the Curse of Dimensionality with Convex Neural Networks, Journal of Machine Learning Research, 18, 1--53

Bach, F. (2017): Breaking the Curse of Dimensionality with Convex Neural Networks, Journal of Machine Learning Research, 18, 1--53

work page 2017