Pith. sign in

REVIEW 2 major objections 5 minor 93 references

This paper proves that the convergence rate of a model with pre-trained embeddings is driven by source-task error and a transferability condition, and that valid DML inference requires all parts of the rate below n^{-1/4}.

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 18:09 UTC pith:JF2ONQ5J

load-bearing objection A rigorous, carefully written theory paper, but the headline claim is conditional on a transferability condition that likely fails in the paper's own motivating settings. the 2 major comments →

arxiv 2607.17378 v1 pith:JF2ONQ5J submitted 2026-07-19 econ.EM math.STstat.MLstat.TH

Econometrics with Pre-Trained Embeddings for Unstructured Data

Yuya Shimizu This is my paper
classification econ.EM math.STstat.MLstat.TH MSC 62G0562G20
keywords transfer learningdouble machine learningpre-trained embeddingsunstructured datatask diversityconvergence ratesridge regressionpartial identification
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.

This paper tries to establish when economists can trust the common shortcut of taking embeddings from a pre-trained image or text model and using them as covariates in a target economic analysis. It argues that two conditions are sufficient: the source and target tasks share an approximately true embedding, and the source task is diverse enough that any embedding nearly reproducing the source predictions also nearly reproduces the target predictions. Under these conditions, the target model's L2 error is bounded by three terms stacked in a single convergence rate, and the paper gives lower-level VC-dimension and ridge versions of that rate. If the combined rate is o(n^{-1/4}), double machine learning (DML) — estimating a low-dimensional parameter while controlling for complex nuisance functions — yields valid inference. The paper's own simulation shows what happens when the condition fails: an out-of-span target nuisance shifts the DML estimate to 1.197 instead of the true 1.

Core claim

Put in one sentence, the paper proves a decomposition: the target-model error satisfies ||bf_qh - a*_ta|| = O_P(rta_n + delta_so,m / nu_n + epsilon_ta,n). The first term is the target-stage estimation rate given the embedding; the second is the source stage's estimation error divided by the transferability strength nu_n; the third is the approximation error of the shared embedding. The source-side term is the paper's main addition to existing work: the pre-trained model was fit on an unobserved source sample, and its estimation error propagates to the target unless the target directions are well covered by source tasks. In the linear case, transferability is exactly the row-span condition be

What carries the argument

nu_n-transferability / task diversity. For radius rho, define the source near-identified set of embeddings h that can reproduce the source model's predictions within rho, and define the target near-identified set analogously. Transferability requires the source near-identified sets to be contained in the target near-identified sets after dividing the radius by nu_n; this converts source convergence delta_so,m into a target-side error delta_so,m / nu_n. Together with the approximately-true-embedding definition, this set-inclusion condition carries the proof. For linear source heads, the condition reduces to the target slope lying in the row span of the source coefficient matrix, and the paper

Load-bearing premise

The load-bearing premise is transferability: every embedding that nearly reproduces the source task's predictions must also nearly reproduce the target task's predictions; if the target contains a direction the source cannot see, the source-error term cannot be tamed.

What would settle it

Construct a partially linear DML setting in which the target nuisance coefficients are orthogonal to the row space of the source-task coefficient matrix, keep all other conditions favorable, and let both sample sizes grow; if the estimator stays centered and intervals attain nominal coverage, the transferability condition is not necessary. The paper's out-of-span design, which shifts the estimate to 1.197, is the predicted failure.

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

If this is right

  • If all rate components are o_P(n^{-1/4}), DML estimators in partially linear models, demand elasticity, imputation, and average treatment effects with unstructured confounders are asymptotically normal.
  • Source-task estimation error cannot be ignored: treating the pre-trained embedding as fixed truth leaves inference uncontrolled no matter how large the target sample is.
  • For linear source heads, transferring to every possible target requires at least K source tasks, because arbitrary target slopes must lie in the span of the T source-task slopes.
  • Nonidentification of the embedding up to invertible linear transformations makes LASSO-style sparsity fragile; ridge regression is the paper's recommended robust target model.
  • Later hidden layers, especially the last one, are better embedding choices than earlier layers because transferability is easier to justify there.

Where Pith is reading between the lines

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

  • A cheap diagnostic follows from the theory: compare downstream estimates across source models with different numbers of source classes; estimates that shift with T signal source tasks that fail to cover target-relevant directions.
  • The proof's triangle-inequality structure extends to multimodal embeddings, with source errors added across modalities, so one badly identified modality can dominate the final rate.
  • A testable implication is interval coverage: when the target slope is deliberately chosen outside the row span of Bso, DML confidence intervals should under-cover even with a large source sample; the paper's own out-of-span simulation is one example.

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 / 5 minor

Summary. The paper develops a convergence-rate theory for double machine learning when nuisance functions are estimated from embeddings delivered by a pre-trained source model. The main result (Theorem 1) bounds the L2(Pta) error of the target model by rta,n + δso,m/νn + ϵta,n, where δso,m is the source estimation rate, νn is a transferability coefficient, and ϵta,n is the target approximation error. Theorems 6–8 give primitive conditions via VC-dimension bounds and a ridge analysis, and Theorems 3–5 give sufficient and necessary characterizations of the transferability condition. The paper also shows that LASSO is fragile under the embedding identification problem and recommends ridge or shallow neural networks. A Monte Carlo simulation illustrates that a target direction outside the source-identified span shifts the DML estimator, while an in-span design is centered at the truth.

Significance. If the results hold, this is a useful formalization of a widely used workflow in empirical economics. The paper makes the transferability condition explicit and gives both sufficient and necessary characterizations, recognizing the identification problem of embedding functions up to linear transformations. The explicit inclusion of source-task estimation error, the ridge analysis in Theorem 8, and the negative simulation result are genuine contributions. The paper is honest about the restrictiveness of the assumptions, including the density-ratio condition and the failure of transferability in simple nonlinear-head examples. The theory is internally consistent and the proofs in Supplement F appear to support the stated rates.

major comments (2)
  1. [Abstract; Section 1; Examples 6–7; Example 10] The motivating applications are VGG19/ResNet50 and BERT, with source output dimensions T=1000 or T=2 and embedding dimensions K=4096/2048 or K=1024. For these, the row-span condition in Example 9 requires T≥K for universal transferability, and Example 10 shows that nonlinear source heads can break transferability even in very simple cases. The paper acknowledges this in Example 10 and the simulation in Section I, but the abstract and introduction present the workflow as generally justifiable. I recommend adding an explicit caveat that for many standard pre-trained models the transferability condition is restrictive and may fail, and that the theory provides a formal characterization rather than a blanket justification. This is a scope/emphasis issue rather than a technical error.
  2. [Theorem 7; Remark 7; Lemma 5] The source-side rate δso,m degrades with T under multinomial logistic loss because τm ≍ T^{-1} (Remark 7). For ImageNet-scale problems (T=1000, K=4096, m≈1.28×10^6), the term δso,m is likely to be far above n^{-1/4} for realistic target sample sizes. The paper notes the dependence on T only indirectly. Please add a short paragraph quantifying this and explaining the practical implications: the o(n^{-1/4}) condition needed for DML will typically require either small T/K regimes or additional structure that restores transferability.
minor comments (5)
  1. [Section 2.1, Eq. (3)] The notation ∥·∥P,2 for the target distribution is used before P is defined in the transfer-learning section. Please define P at first use in Section 2.
  2. [Section 3.2, Theorem 1] The phrase 'for every sequence of estimated embedding functions qh∈Hm satisfying (12)' is technically strong. It might be clearer to state that the bound holds uniformly over such sequences, or to explicitly note that the bound is pointwise in qh, as you do in Remark 4.
  3. [Section 4.2.1, Eq. (23)] The projection qhproj is defined using qBso, but the notation qBso appears before the estimation problem (6) is introduced in full. Consider a brief reminder of the source-head notation.
  4. [Section I, Table 2] The standard deviations of the prediction errors are enormous relative to the means (e.g., 464.783 vs 10.579). This suggests heavy-tailed or occasionally disastrous fits. A brief comment on robustness or trimming would help.
  5. [References] Several citations are to unpublished manuscripts or preprints with no DOI. For a journal version, consider adding a note that these are available from the authors or arXiv.

Circularity Check

0 steps flagged

No significant circularity: the central rate is a transparent conditional bound under explicit transferability and convergence-rate assumptions, and those assumptions are given independent content by primitive sufficient/necessary results.

full rationale

The paper's main bound (Theorem 1) is indeed little more than a triangle-inequality decomposition once Assumptions 1 and 2 are granted: the δso,m/νn term is exactly the content of νn-transferability (Definition 3). This is, however, a stated assumption rather than a fitted or predicted quantity; the paper does not hide the dependence, and it supplies substantive lower-level content for the assumption. Theorem 3 and Example 9/Corollary 1 give primitive sufficient conditions for transferability (e.g., via a Lipschitz link Γ or the row-span condition β'_ta = b'Bso,m), and Theorems 4–5 show necessity in linear and general settings. Theorem 7 derives δso,m from VC-dimension bounds and Lemma 5, and Theorem 6 derives rta,n from the target class complexity; these are standard empirical-process rates and are not calibrated to target data. There are no author self-citations used as load-bearing evidence, and no fitted constant is renamed as a prediction. The Monte Carlo is explicitly illustrative and is not part of the proof chain. Accordingly, the derivation is self-contained conditional on transparent assumptions; any concern that the transferability condition is strong or likely violated for particular architectures belongs to correctness/applicability, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 7 axioms · 0 invented entities

The central claim rests on explicit transferability, common-support, and approximation-error-decay assumptions. These are stated as Assumptions 1-6 rather than hidden, but they are strong and generally not verifiable from target data alone. The mathematical background—VC dimension bounds, localized empirical processes, and matrix inequalities—is standard.

axioms (7)
  • standard math True source and target models exist and are square-integrable; population minimizers exist.
    Used in Section 3 to define a*_so and a*_ta and to discuss uniqueness under logit/least-squares losses.
  • domain assumption Approximation errors ϵso,m and ϵta,n converge to zero as m,n grow.
    Definition 1 and the consistency claim require the shared embedding to approximate both tasks; the paper states this is necessary for the DML rate.
  • domain assumption νn-transferability / task diversity: Hso,m(ρso,m) ⊆ Hta,m(ρso,m/νn).
    Assumption 2 is the load-bearing condition. In linear models it reduces to the target slope lying in the row span of the source head (Example 9).
  • domain assumption Density ratio pso/pta bounded below by w_n² (Assumption 3).
    Used to convert source L2 nearness to target L2 nearness in Theorem 3 and the ridge analysis. The paper acknowledges this is strong for unstructured high-dimensional Z.
  • domain assumption Target loss excess risk is quadratically equivalent to L2 distance and the loss is Lipschitz (Assumption 4).
    Standard in DML/statistical learning; used to convert empirical risk bounds into L2 convergence rates in Theorem 6.
  • domain assumption Source loss curvature, Lipschitz loss, and Lipschitz source heads (Assumption 5).
    Used in Theorem 7 to derive source-side rates via localized empirical process bounds and entropy composition.
  • domain assumption Ridge target-slope representation β'_ta = b'Bso and source-head singular value bounds (Assumption 6).
    This is the transferability condition specialized to ridge; it confines the target slope to the row span of the source head and restricts the source head's conditioning.

pith-pipeline@v1.3.0-alltime-deepseek · 65080 in / 16584 out tokens · 181952 ms · 2026-08-01T18:09:13.496461+00:00 · methodology

0 comments
read the original abstract

Unstructured data, such as images and text, are increasingly used in empirical economics. Since training machine-learning models on unstructured data is costly, economists often use off-the-shelf pre-trained deep learning models developed by computer scientists to extract embeddings, which are then used as covariates in target economic analyses. Despite the popularity of this practice, its theoretical foundations remain limited. There are two main difficulties. First, the pre-trained model is usually trained on a different dataset and for a different task. Consequently, it is unclear when such a model can be used reliably for the target task. Second, the embedding function is subject to an identification problem, which makes it difficult to analyze the estimation error of the embedding function and its effect on the target task. In this paper, we provide sufficient conditions to overcome these difficulties and derive the convergence rate of machine learning models with pre-trained embeddings. We illustrate the theory through double machine learning applications for estimating parameters of interest, such as partially linear regression with unstructured controls, price elasticity in demand estimation considering the product quality measured by images and text, missing data imputation with unstructured data, and the average treatment effect with unstructured confounders.

Figures

Figures reproduced from arXiv: 2607.17378 by Yuya Shimizu.

Figure 1
Figure 1. Figure 1: Histogram of θb for the in-span and out-of-span designs. Under these conditions, we show that the convergence rate of transfer learning can be faster than n −1/4 , which is sufficient for valid inference in the double machine learning (DML) framework of Chernozhukov, Chetverikov, Demirer, Duflo, Hansen, Newey, and Robins (2018) for a size-n target sample. Our convergence rate accounts for randomness in bot… view at source ↗
Figure 2
Figure 2. Figure 2: Sample Splitting with Transfer Learning only to estimate nuisance parameters, the sample-splitting condition holds once the target sample is split between nuisance estimation and parameter estimation ( [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

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

93 extracted references · 10 linked inside Pith

  1. [1]

    2018 , publisher=

    Deep learning with Python , author=. 2018 , publisher=

    2018

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    On the theory of transfer learning: The importance of task diversity , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Optimistic rates for multi-task representation learning , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Science , volume=

    Prediction-powered inference , author=. Science , volume=. 2023 , publisher=

    2023

  5. [5]

    arXiv preprint arXiv:2505.00282 , year=

    A Unifying Framework for Robust and Efficient Inference with Unstructured Data , author=. arXiv preprint arXiv:2505.00282 , year=

    arXiv

  6. [6]

    arXiv preprint arXiv:2501.09731 , year=

    Predictions as surrogates: Revisiting surrogate outcomes in the age of ai , author=. arXiv preprint arXiv:2501.09731 , year=

    Pith/arXiv arXiv

  7. [7]

    Proceedings of the National Academy of Sciences , volume=

    Cross-prediction-powered inference , author=. Proceedings of the National Academy of Sciences , volume=. 2024 , publisher=

    2024

  8. [8]

    Inference for Regression with Variables Generated by

    Battaglia, Laura and Christensen, Timothy and Hansen, Stephen and Sacher, Szymon , journal=. Inference for Regression with Variables Generated by

  9. [9]

    Debiasing

    Zhang, Jingwen and Xue, Wendao and Yu, Yifan and Tan, Yong , journal=. Debiasing. 2026 , publisher=

    2026

  10. [10]

    arXiv preprint arXiv:2411.10959 , year=

    Program Evaluation with Remotely Sensed Outcomes , author=. arXiv preprint arXiv:2411.10959 , year=

    Pith/arXiv arXiv

  11. [11]

    The Econometrics Journal , pages=

    Double/debiased machine learning for treatment and structural parameters , author=. The Econometrics Journal , pages=. 2018 , publisher=

    2018

  12. [12]

    Econometrica , volume=

    Locally robust semiparametric estimation , author=. Econometrica , volume=. 2022 , publisher=

    2022

  13. [13]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Transfer learning for high-dimensional linear regression: Prediction, estimation and minimax optimality , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2022 , publisher=

    2022

  14. [14]

    Journal of the American Statistical Association , volume=

    Estimation and inference for high-dimensional generalized linear models with knowledge transfer , author=. Journal of the American Statistical Association , volume=. 2024 , publisher=

    2024

  15. [15]

    Science , volume=

    Combining satellite imagery and machine learning to predict poverty , author=. Science , volume=. 2016 , publisher=

    2016

  16. [16]

    arXiv preprint arXiv:2503.20711 , year=

    Demand estimation with text and image data , author=. arXiv preprint arXiv:2503.20711 , year=

    arXiv

  17. [17]

    Journal of Econometrics , volume =

    Hedonic prices and quality adjusted price indices powered by AI , author =. Journal of Econometrics , volume =. 2025 , issn =

    2025

  18. [18]

    Econometrica , volume=

    Deep neural networks for estimation and inference , author=. Econometrica , volume=. 2021 , publisher=

    2021

  19. [19]

    2023 , publisher=

    Weak Convergence and Empirical Processes: With Applications to Statistics , author=. 2023 , publisher=

    2023

  20. [20]

    1998 , publisher=

    Asymptotic statistics , author=. 1998 , publisher=

    1998

  21. [21]

    2019 , publisher=

    High-dimensional statistics: A non-asymptotic viewpoint , author=. 2019 , publisher=

    2019

  22. [22]

    Journal of Machine Learning Research , volume=

    Nearly-tight VC-dimension and pseudodimension bounds for piecewise linear neural networks , author=. Journal of Machine Learning Research , volume=

  23. [23]

    Econometrica , volume=

    Program evaluation and causal inference with high-dimensional data , author=. Econometrica , volume=. 2017 , publisher=

    2017

  24. [24]

    arXiv preprint arXiv:2506.14329 , year=

    Adjustment for confounding using pre-trained representations , author=. arXiv preprint arXiv:2506.14329 , year=

    arXiv

  25. [25]

    arXiv preprint arXiv:2402.01785 , year=

    Doublemldeep: Estimation of causal effects with multimodal data , author=. arXiv preprint arXiv:2402.01785 , year=

    Pith/arXiv arXiv

  26. [26]

    Unpublished manuscript , year=

    Are Patent Examiners Gender Neutral , author=. Unpublished manuscript , year=

  27. [27]

    Proceedings of the National Academy of Sciences , volume=

    Estimating wage disparities using foundation models , author=. Proceedings of the National Academy of Sciences , volume=. 2025 , publisher=

    2025

  28. [28]

    International Conference on Machine Learning , pages=

    Approximation and estimation ability of transformers for sequence-to-sequence functions with infinite dimensional input , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  29. [29]

    arXiv preprint arXiv:2410.15800 , year=

    On the VC dimension of deep group convolutional neural networks , author=. arXiv preprint arXiv:2410.15800 , year=

    Pith/arXiv arXiv

  30. [30]

    Econometrica , volume=

    Automobile Prices in Market Equilibrium , author=. Econometrica , volume=

  31. [31]

    The RAND Journal of Economics , pages=

    Estimating discrete-choice models of product differentiation , author=. The RAND Journal of Economics , pages=. 1994 , publisher=

    1994

  32. [32]

    Journal of Econometrics , volume=

    Semi-nonparametric estimation of random coefficients logit model for aggregate demand , author=. Journal of Econometrics , volume=. 2023 , publisher=

    2023

  33. [33]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  34. [34]

    arXiv preprint arXiv:1908.10084 , year=

    Sentence-bert: Sentence embeddings using siamese bert-networks , author=. arXiv preprint arXiv:1908.10084 , year=

    Pith/arXiv arXiv 1908

  35. [35]

    Nature Communications , volume=

    A generalizable and accessible approach to machine learning with global satellite imagery , author=. Nature Communications , volume=. 2021 , publisher=

    2021

  36. [36]

    2023 , publisher=

    Deep learning: Foundations and concepts , author=. 2023 , publisher=

    2023

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    3rd International Conference on Learning Representations,

    Very deep convolutional networks for large-scale image recognition , author =. 3rd International Conference on Learning Representations,

  39. [39]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  40. [40]

    Advances in Neural Information Processing Systems , volume=

    Representation learning beyond linear prediction functions , author=. Advances in Neural Information Processing Systems , volume=

  41. [41]

    1999 , publisher=

    Geometry of sets and measures in Euclidean spaces: fractals and rectifiability , author=. 1999 , publisher=

    1999

  42. [42]

    International Conference on Machine Learning , pages=

    Hidden symmetries of ReLU networks , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  43. [43]

    2002 , publisher=

    Real analysis and probability , author=. 2002 , publisher=

    2002

  44. [44]

    1999 , publisher=

    Real Analysis: Modern Techniques and Their Applications , author=. 1999 , publisher=

    1999

  45. [45]

    2008 , publisher=

    Matrix Algebra From a Statistician's Perspective , author=. 2008 , publisher=

    2008

  46. [46]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  47. [47]

    arXiv preprint arXiv:2501.00382 , year=

    Adventures in demand analysis using AI , author=. arXiv preprint arXiv:2501.00382 , year=

    arXiv

  48. [48]

    arXiv preprint arXiv:2311.02299 , year=

    The fragility of sparsity , author=. arXiv preprint arXiv:2311.02299 , year=

    Pith/arXiv arXiv

  49. [49]

    Econometrica , volume=

    Inferential theory for factor models of large dimensions , author=. Econometrica , volume=. 2003 , publisher=

    2003

  50. [50]

    Foundations and Trends

    Spectral methods for data science: A statistical perspective , author=. Foundations and Trends. 2021 , publisher=

    2021

  51. [51]

    Neural Networks , volume=

    Error bounds for approximations with deep ReLU networks , author=. Neural Networks , volume=. 2017 , publisher=

    2017

  52. [52]

    arXiv preprint arXiv:2601.05374 , year=

    From Unstructured Data to Demand Counterfactuals: Theory and Practice , author=. arXiv preprint arXiv:2601.05374 , year=

    Pith/arXiv arXiv

  53. [53]

    Automatic Locally Robust

    Escanciano, Juan Carlos and P. Automatic Locally Robust. arXiv preprint arXiv:2301.10643v4 , year=

    arXiv

  54. [54]

    arXiv preprint arXiv:2501.16120 , year=

    Copyright and Competition: Estimating Supply and Demand with Unstructured Data , author=. arXiv preprint arXiv:2501.16120 , year=

    arXiv

  55. [55]

    arXiv preprint arXiv:2202.10054 , year=

    Fine-tuning can distort pretrained features and underperform out-of-distribution , author=. arXiv preprint arXiv:2202.10054 , year=

    Pith/arXiv arXiv

  56. [56]

    2009 , publisher=

    The elements of statistical learning: data mining, inference, and prediction , author=. 2009 , publisher=

    2009

  57. [57]

    Annals of Statistics , volume=

    Dimension free ridge regression , author=. Annals of Statistics , volume=. 2024 , publisher=

    2024

  58. [58]

    Advances in Neural Information Processing Systems , volume=

    Debiased machine learning without sample-splitting for stable estimators , author=. Advances in Neural Information Processing Systems , volume=

  59. [59]

    Vershynin, Roman , title =

  60. [60]

    Annals of Statistics , volume=

    Surprises in high-dimensional ridgeless least squares interpolation , author=. Annals of Statistics , volume=

  61. [61]

    The 22nd International Conference on Artificial Intelligence and Statistics , pages=

    Support and invertibility in domain-invariant representations , author=. The 22nd International Conference on Artificial Intelligence and Statistics , pages=. 2019 , organization=

    2019

  62. [62]

    and Sun, Ji-guang , publisher =

    Stewart, Gilbert W. and Sun, Ji-guang , publisher =. Matrix Perturbation Theory , year =

  63. [63]

    1994 , publisher=

    Topics in matrix analysis , author=. 1994 , publisher=

    1994

  64. [64]

    1991 , publisher=

    Probability in Banach Spaces: isoperimetry and processes , author=. 1991 , publisher=

    1991

  65. [65]

    2008 , publisher=

    Introduction to empirical processes and semiparametric inference , author=. 2008 , publisher=

    2008

  66. [66]

    arXiv preprint arXiv:2511.04957 , year=

    Training and testing with multiple splits: A central limit theorem for split-sample estimators , author=. arXiv preprint arXiv:2511.04957 , year=

    arXiv

  67. [67]

    Annual Review of Economics , volume=

    Large language models: An applied econometric framework , author=. Annual Review of Economics , volume=

  68. [68]

    International Economic Review , pages=

    Econometric issues in the analysis of regressions with generated regressors , author=. International Economic Review , pages=. 1984 , publisher=

    1984

  69. [69]

    Econometrica , volume=

    Asymptotic variance of semiparametric estimators with generated regressors , author=. Econometrica , volume=. 2013 , publisher=

    2013

  70. [70]

    Econometrica , pages=

    Sample selection bias as a specification error , author=. Econometrica , pages=. 1979 , publisher=

    1979

  71. [71]

    Annals of Statistics , volume=

    Deep neural networks for nonparametric interaction models with diverging dimension , author=. Annals of Statistics , volume=. 2024 , publisher=

    2024

  72. [72]

    Journal of Machine Learning Research , volume=

    Dropout: a simple way to prevent neural networks from overfitting , author=. Journal of Machine Learning Research , volume=. 2014 , publisher=

    2014

  73. [73]

    International Conference on Machine Learning , pages=

    The implicit and explicit regularization effects of dropout , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  74. [74]

    Annals of Statistics , volume=

    On the rate of convergence of fully connected deep neural network regression estimates , author=. Annals of Statistics , volume=. 2021 , publisher=

    2021

  75. [75]

    Annals of Statistics , volume=

    Deep nonparametric regression on approximate manifolds: Nonasymptotic error bounds with polynomial prefactors , author=. Annals of Statistics , volume=. 2023 , publisher=

    2023

  76. [76]

    arXiv preprint arXiv:2503.00725 , year=

    Causal inference on outcomes learned from text , author=. arXiv preprint arXiv:2503.00725 , year=

    Pith/arXiv arXiv

  77. [77]

    Political Analysis , volume=

    Machine learning predictions as regression covariates , author=. Political Analysis , volume=. 2021 , publisher=

    2021

  78. [78]

    Advances in Neural Information Processing Systems , volume=

    Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models , author=. Advances in Neural Information Processing Systems , volume=

  79. [79]

    arXiv preprint arXiv:2505.00526 , year=

    Pre-Training Estimators for Structural Models: Application to Consumer Search , author=. arXiv preprint arXiv:2505.00526 , year=

    arXiv

  80. [80]

    arXiv preprint arXiv:2511.01680 , year=

    Making Interpretable Discoveries from Unstructured Data: A High-Dimensional Multiple Hypothesis Testing Approach , author=. arXiv preprint arXiv:2511.01680 , year=

    Pith/arXiv arXiv

Showing first 80 references.