Pith. sign in

REVIEW 2 major objections 5 minor 97 references

A hypothesis test can tell, from historical data, whether personalizing interventions beats the best single action for everyone.

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 · grok-4.5

2026-07-13 05:32 UTC pith:BKROJHYR

load-bearing objection Solid multi-action personalization test with real efficiency claims and cleaner CIs than the baselines; the uniqueness/gap assumption is the main limit and is already flagged. the 2 major comments →

arxiv 2607.08951 v1 pith:BKROJHYR submitted 2026-07-09 stat.ME stat.APstat.ML

A Statistical Test for the Benefits of Personalizing Interventions

Zhaoqi Li , Emma Brunskill This is my paper
classification stat.ME stat.APstat.ML MSC 62G1062P1062P25
keywords personalization effecthypothesis testdoubly robust estimationcross-fittingsemiparametric efficiencypolicy learningoffline evaluationheterogeneous treatment effects
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.

Decision-makers often want to know whether assigning different interventions to different people will raise average outcomes enough to justify the extra cost and complexity of personalization. This paper supplies a statistical test for that question: given offline data of features, actions, and outcomes, it estimates the personalization effect—the gap between the best policy in a chosen class and the single best fixed intervention—and tests whether that gap is positive. The estimator reuses data carefully across folds and random partitions so that the resulting z-statistic has controlled Type-I error and, under stronger conditions, the smallest possible asymptotic variance among regular estimators. On job-training, depression-treatment, MOOC, and joke-recommendation data the procedure yields stable confidence intervals that are at least as tight as prior alternatives and that correctly flag when personalization helps and when it does not. The result is a practical tool that lets practitioners quantify the benefit of personalization before committing to more elaborate policies.

Core claim

The K-fold Personalization Test (KPT) produces an asymptotically normal, semiparametrically efficient estimator of the personalization effect ψ and a valid one-sided test of H0: ψ = 0 versus Ha: ψ > 0, provided the best single intervention is unique and can be learned at a fast rate.

What carries the argument

A doubly-robust (AIPW-style) personalization-effect score that treats both the learned personalized policy and the learned best single action as nuisance parameters, evaluated by K-fold cross-fitting and aggregated over repeated random partitions of the data.

Load-bearing premise

The single best intervention for the whole population must be unique and learnable faster than the usual square-root rate; if two interventions are essentially tied at the population level the Type-I guarantee no longer holds.

What would settle it

Construct a synthetic data-generating process with two equally good population-level interventions and a known positive personalization effect; if the KPT z-statistic still rejects H0 at the nominal rate while its coverage of the true effect collapses, the uniqueness assumption is necessary and the claim fails.

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

If this is right

  • Practitioners can obtain a confidence interval for the expected gain of personalization before investing in more complex delivery systems.
  • Any existing best-arm or contextual-policy learner can be plugged into the same pipeline without redesigning the test.
  • When the test fails to reject, stakeholders receive quantitative evidence that a single intervention is competitive, reducing pressure to personalize.
  • The same estimator immediately yields a policy-value comparison between any two learned policies, one of which may be non-personalized.

Where Pith is reading between the lines

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

  • The same cross-fit AIPW construction could be reused to test whether a continuous or high-cardinality action space still yields a personalization gain once discretized.
  • If the uniqueness gap is small, a two-stage procedure that first tests for zero average treatment effect and then runs KPT only on the residual heterogeneity could restore validity without sacrificing power.
  • Stakeholders who must choose among several policy classes could treat the KPT z-statistic itself as a model-selection criterion, automatically preferring the class that most clearly exceeds the best fixed action.

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 proposes the K-fold Personalization Test (KPT), a cross-fit AIPW-style estimator and z-test for the personalization effect ψ = V(π*) − max_a U(a). Under Assumptions 1–10 (including uniqueness of the best arm and consistency of the learned policy), the z-statistic has asymptotic Type-I error zero under the null; under the stronger Assumptions 1–8 and 10–11 (fast policy rates), the estimator is asymptotically normal and attains the semiparametric efficiency bound for ψ. The method handles multi-action settings and multi-dimensional covariates, reuses data via K-fold cross-fitting and repeated random partitions, and is compared on four real/semi-synthetic datasets (Job Corps, Nefazodone, MOOC, Jester) and fully synthetic environments against SRP, PAPD, and a TrainEval baseline. Empirically KPT yields more stable and often tighter confidence intervals and better-calibrated significance decisions.

Significance. If the claims hold, the paper supplies a practical, multi-action, multi-covariate test for whether personalization improves expected outcomes over the best single intervention—an estimand that is necessary for justifying the cost and fragility of personalized policies but is not answered by HTE tests alone. Strengths include: (i) explicit Type-I and efficiency theorems in the SM with pathwise differentiability and influence-function arguments; (ii) a reusable AIPW construction that treats both the policy and the best arm as nuisance parameters; (iii) reproducible code and multi-domain experiments showing stability gains over single-split baselines; and (iv) an honest discussion of the gap/uniqueness limitation (SM §4.8) together with empirical checks when the gap is zero. The contribution is of clear interest to biostatistics, econometrics, and offline policy learning.

major comments (2)
  1. Assumption 6 (fast best-arm learner: P(â_k ≠ a*) = o(n_k^{-1/2})) is load-bearing for both Theorem 3.1 / Corollary 3.2 (Type I) and the efficiency argument (Lemma 3.11 + Propositions 3.18–3.20). The main text should state this more prominently (not only in SM §3 and §4.8) and give practical diagnostics—e.g., estimated action gap, best-arm stability across splits (already reported in Table S1)—so readers can assess whether the guarantee applies to their data. The zero-ATE simulation and MOOC results are helpful but do not replace a clear statement of the formal regime.
  2. Efficiency (Lemma 3.11) further requires Assumption 11 (low-regret policy learning at o(n^{-1/2})), typically justified by margin conditions. In the synthetic experiments (Figs. S3–S4) policy accuracy is often well below 100% when power is already high; the paper should either (a) report empirical regret rates or (b) clarify that the efficiency claim is asymptotic under margin conditions that may not hold in the finite-sample regimes shown, so that readers do not over-interpret the efficiency result as explaining the observed power.
minor comments (5)
  1. Figure 3 caption and column labels: several typos (“columb”, “parition”, “paritions”) and inconsistent capitalization of method names; fix for production.
  2. Main-text description of fold allocation (K=6, half for policy/best-arm learning) is useful; a short sensitivity note or pointer to SM on how results change with K or the policy-learning fraction would help practitioners.
  3. Notation: the personalization effect is written both as ψ and as Ψ in places (e.g., Algorithm 2 variance); unify.
  4. Related-work discussion of qualitative-interaction tests and conditional-moment-inequality methods is thorough; a one-sentence clarification that those methods typically target binary treatment and single covariates would further highlight the multi-action contribution.
  5. SM Algorithm 3 (PAPD baseline) and the centering of Y should be briefly mentioned in the main-text comparison paragraph so that the PAPD results are fully reproducible from the main text alone.

Circularity Check

0 steps flagged

No circularity: estimand ψ is defined independently of the AIPW/cross-fit estimator; Type-I and efficiency proofs reduce to standard regularity + gap assumptions, not to fitted inputs or self-citation chains.

full rationale

The personalization effect ψ is introduced in Eq. (1)/ (S1) purely as the difference between the value of the optimal policy in a fixed class Π and the value of the best single action; this definition does not mention the estimator. The KPT estimator (Alg. 1–2, Eq. (2)/(S17)) is a standard cross-fit AIPW score that treats the learned policy and best arm as additional nuisance parameters; under H0 the score is shown (Thm. 3.1, Cor. 3.2, Lem. 3.3) to be o_p(1) once the best-arm indicator vanishes at rate o(n^{-1/2}) (Ass. 6) and the usual product-rate conditions on outcome/propensity models hold. Efficiency (Lem. 3.11 + Props. 3.18–3.20) likewise follows from pathwise differentiability of the same fixed functional under uniqueness of π* and a*, again without feeding the target back into the definition. Empirical sections compare the resulting CI and z-statistic against independent baselines (SRP, PAPD, TrainEval) and against fully synthetic data whose true ψ is known by construction; none of these comparisons close a definitional loop. Minor citations to the authors’ earlier policy-learning papers are used only as black-box oracles that satisfy the stated rate assumptions; they are not invoked as uniqueness theorems that force the present claims. Consequently the derivation chain is free of the six circularity patterns.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

Standard causal identification assumptions plus uniqueness/gap conditions on the best arm and fast rates for policy learning; the free parameters are algorithmic choices (folds, splits, policy class) that affect finite-sample behavior but not the asymptotic claims.

free parameters (3)
  • number of folds K = 6
    Fixed at K=6 in experiments; theory holds for any fixed K≥3 but finite-sample power depends on the split.
  • number of repeated splits S = 100
    Chosen S=100 (or 200 in sensitivity); asymptotic results hold for fixed S, but stability improves with larger S.
  • policy class Π and learner hyper-parameters
    DR trees/forests, linear threshold policies, lasso; depth, min-leaf, regularization chosen by hand and shown to affect significance in sensitivity analyses.
axioms (5)
  • domain assumption SUTVA, unconfoundedness, strong overlap, bounded rewards (Assumptions 1–4)
    Standard causal identification conditions required for the AIPW score to be unbiased for ψ.
  • domain assumption Unique best single intervention with positive gap (Assumption 6 / MG)
    Needed for Type-I control; fails when ATE≈0 or multiple arms are tied.
  • standard math Consistent nuisance estimators and product-rate condition (Assumptions 5,7)
    Classic DML-style rates that make the remainder o_p(n^{-1/2}).
  • domain assumption Low-regret / margin condition for policy learning (Assumption 11 / MA)
    Required only for efficiency, not for Type-I; may fail with limited data or flat reward surfaces.
  • domain assumption Unique optimal policy almost everywhere (Assumption 10)
    Rules out large sets of non-responders where multiple actions are optimal.
invented entities (2)
  • personalization effect ψ independent evidence
    purpose: Scalar target: value of best policy in Π minus value of best single action.
    Defined in Eq. (1)/S1; not previously given an efficient multi-action estimator.
  • K-fold Personalization Test (KPT) estimator independent evidence
    purpose: Cross-fit AIPW estimator of ψ that treats both policy and best arm as nuisances.
    Algorithms 1–2; the paper’s main methodological contribution.

pith-pipeline@v1.1.0-grok45 · 62709 in / 2749 out tokens · 38287 ms · 2026-07-13T05:32:44.696104+00:00 · methodology

0 comments
read the original abstract

From medicine to marketing to social sciences, the promise of tailoring interventions to individuals is undeniable. However, practical applications force weighing personalization's potential benefits with its possible increased cost and fragility. We introduce a statistical hypothesis test that evaluates, given historical data, evidence that a personalized intervention policy's performance will surpass deploying the best single intervention. The test maintains strict type-I error control while achieving asymptotic normality with the minimal possible variance under specified conditions. Results on diverse datasets from job training, depression treatment, education and recommendation systems demonstrate the test's versatility and its superior performance over alternatives. This test can support decision-makers throughout the intervention sciences by providing a simple and powerful quantification of the potential benefits of personalization.

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

97 extracted references · 3 linked inside Pith

  1. [1]

    Machine Learning , pages=

    Did we personalize? assessing personalization by an online reinforcement learning algorithm using resampling , author=. Machine Learning , pages=. 2024 , publisher=

    2024

  2. [2]

    A Review of Off-Policy Evaluation in Reinforcement Learning , author=

  3. [3]

    Statistical science: a review journal of the Institute of Mathematical Statistics , volume=

    Replicable bandits for digital health interventions , author=. Statistical science: a review journal of the Institute of Mathematical Statistics , volume=

  4. [4]

    Journal of Machine Learning Research , volume=

    Breaking the Curse of Nonregularity with Subagging---Inference of the Mean Outcome under Optimal Treatment Regimes , author=. Journal of Machine Learning Research , volume=

  5. [5]

    1993 , publisher=

    Efficient and adaptive estimation for semiparametric models , author=. 1993 , publisher=

    1993

  6. [6]

    New England Journal of Medicine , volume=

    A comparison of nefazodone, the cognitive behavioral-analysis system of psychotherapy, and their combination for the treatment of chronic depression , author=. New England Journal of Medicine , volume=. 2000 , publisher=

    2000

  7. [7]

    American Economic Review , volume=

    Can higher prices stimulate product use? Evidence from a field experiment in Zambia , author=. American Economic Review , volume=. 2010 , publisher=

    2010

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    What's the harm? sharp bounds on the fraction negatively affected by treatment , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    Operations Research , volume=

    Offline multi-action policy learning: Generalization and optimization , author=. Operations Research , volume=. 2023 , publisher=

    2023

  10. [10]

    Proceedings of the National Academy of Sciences , volume=

    Eight-minute self-regulation intervention raises educational attainment at scale in individualist but not collectivist cultures , author=. Proceedings of the National Academy of Sciences , volume=. 2017 , publisher=

    2017

  11. [11]

    Journal of the American Statistical Association , volume=

    Estimation and inference of heterogeneous treatment effects using random forests , author=. Journal of the American Statistical Association , volume=. 2018 , publisher=

    2018

  12. [12]

    Proceedings of the National Academy of Sciences , volume=

    Metalearners for estimating heterogeneous treatment effects using machine learning , author=. Proceedings of the National Academy of Sciences , volume=. 2019 , publisher=

    2019

  13. [13]

    Proceedings of the National Academy of Sciences , volume=

    Confidence intervals for policy evaluation in adaptive experiments , author=. Proceedings of the National Academy of Sciences , volume=. 2021 , publisher=

    2021

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Inference for batched bandits , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Neural computation , volume=

    Combined 5 2 cv F test for comparing supervised classification learning algorithms , author=. Neural computation , volume=. 1999 , publisher=

    1999

  16. [16]

    Proceedings of the National Academy of Sciences , volume=

    Combining exchangeable p-values , author=. Proceedings of the National Academy of Sciences , volume=. 2025 , publisher=

    2025

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Post-contextual-bandit inference , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    2000 , publisher=

    Asymptotic statistics , author=. 2000 , publisher=

    2000

  19. [19]

    Journal of the American Statistical Association , volume=

    Cauchy combination test: a powerful test with analytic p-value calculation under arbitrary dependency structures , author=. Journal of the American Statistical Association , volume=. 2020 , publisher=

    2020

  20. [20]

    Proceedings of the National Academy of Sciences , volume=

    Recursive partitioning for heterogeneous causal effects , author=. Proceedings of the National Academy of Sciences , volume=. 2016 , publisher=

    2016

  21. [21]

    Statistical Science , volume=

    Doubly robust policy evaluation and optimization , author=. Statistical Science , volume=. 2014 , publisher=

    2014

  22. [22]

    The Journal of Machine Learning Research , volume=

    Batch learning from logged bandit feedback through counterfactual risk minimization , author=. The Journal of Machine Learning Research , volume=

  23. [23]

    Journal of the American Statistical Association , volume=

    Evaluating treatment prioritization rules via rank-weighted average treatment effects , author=. Journal of the American Statistical Association , volume=. 2024 , publisher=

    2024

  24. [24]

    Econometrica , volume=

    Underinvestment in a profitable technology: The case of seasonal migration in Bangladesh , author=. Econometrica , volume=. 2014 , publisher=

    2014

  25. [25]

    American Economic Review , volume=

    Price subsidies, diagnostic tests, and targeting of malaria treatment: evidence from a randomized controlled trial , author=. American Economic Review , volume=. 2015 , publisher=

    2015

  26. [26]

    Estimating Optimal Policy Value in Linear Contextual Bandits Beyond Gaussianity , author=

  27. [27]

    The Quarterly Journal of Economics , volume=

    Inference on winners , author=. The Quarterly Journal of Economics , volume=. 2024 , publisher=

    2024

  28. [28]

    Advances in Neural Information Processing Systems , volume=

    Double Q-learning , author=. Advances in Neural Information Processing Systems , volume=

  29. [29]

    American Economic Review , volume=

    Rational overoptimism (and other biases) , author=. American Economic Review , volume=. 2004 , publisher=

    2004

  30. [30]

    Management Science , volume=

    The optimizer’s curse: Skepticism and postdecision surprise in decision analysis , author=. Management Science , volume=. 2006 , publisher=

    2006

  31. [31]

    Journal of the American Statistical Association , volume=

    P-values for high-dimensional regression , author=. Journal of the American Statistical Association , volume=. 2009 , publisher=

    2009

  32. [32]

    arXiv preprint arXiv:2404.14644 , year=

    Identifying sparse treatment effects , author=. arXiv preprint arXiv:2404.14644 , year=

    Pith/arXiv arXiv

  33. [33]

    Annals of Statistics , volume=

    Performance guarantees for individualized treatment rules , author=. Annals of Statistics , volume=. 2011 , publisher=

    2011

  34. [34]

    Econometrica , volume=

    Who should be treated? empirical welfare maximization methods for treatment choice , author=. Econometrica , volume=. 2018 , publisher=

    2018

  35. [35]

    Annals of Statistics , volume=

    Statistical inference for the mean outcome under a possibly non-unique optimal treatment strategy , author=. Annals of Statistics , volume=. 2016 , publisher=

    2016

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Action-gap phenomenon in reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    Management Science , volume=

    Policy learning with adaptively collected data , author=. Management Science , volume=. 2024 , publisher=

    2024

  38. [38]

    Econometrica , volume=

    Policy learning with observational data , author=. Econometrica , volume=. 2021 , publisher=

    2021

  39. [39]

    Transactions on Machine Learning Research , issn=

    Estimating Optimal Policy Value in Linear Contextual Bandits Beyond Gaussianity , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  40. [40]

    International Conference on Machine Learning , pages=

    Model selection in batch policy optimization , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  41. [41]

    Annals of Statistics , volume=

    Targeted sequential design for targeted learning inference of the optimal treatment rule and its mean reward , author=. Annals of Statistics , volume=

  42. [42]

    Annales de l'IHP Probabilites et statistiques , volume=

    Performance guarantees for policy learning , author=. Annales de l'IHP Probabilites et statistiques , volume=

  43. [43]

    Econometrica , volume=

    Statistical treatment rules for heterogeneous populations , author=. Econometrica , volume=. 2004 , publisher=

    2004

  44. [44]

    arXiv preprint arXiv:2212.06355 , year=

    A review of off-policy evaluation in reinforcement learning , author=. arXiv preprint arXiv:2212.06355 , year=

    Pith/arXiv arXiv

  45. [45]

    The Econometrics Journal , pages=

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

    2018

  46. [46]

    2024 , note =

    Causal Inference: A Statistical Learning Approach , author=. 2024 , note =

    2024

  47. [47]

    Econometrica: Journal of the Econometric Society , pages=

    Efficiency bounds for semiparametric regression , author=. Econometrica: Journal of the Econometric Society , pages=. 1992 , publisher=

    1992

  48. [48]

    International conference on machine learning , pages=

    Doubly robust off-policy value evaluation for reinforcement learning , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  49. [49]

    Journal of Machine Learning Research , volume=

    Double reinforcement learning for efficient off-policy evaluation in markov decision processes , author=. Journal of Machine Learning Research , volume=

  50. [50]

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

    Statistical inference of the value function for reinforcement learning in infinite-horizon settings , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2022 , publisher=

    2022

  51. [51]

    Advances in Neural Information Processing Systems , volume=

    Instance-optimal pac algorithms for contextual bandits , author=. Advances in Neural Information Processing Systems , volume=

  52. [52]

    International Conference on Algorithmic Learning Theory , pages=

    Importance-weighted offline learning done right , author=. International Conference on Algorithmic Learning Theory , pages=

  53. [53]

    Econometrica , volume=

    Inference based on conditional moment inequalities , author=. Econometrica , volume=. 2013 , publisher=

    2013

  54. [54]

    Econometrica , volume=

    Intersection bounds: Estimation and inference , author=. Econometrica , volume=. 2013 , publisher=

    2013

  55. [55]

    The American Economic Review , pages=

    Evaluating the econometric evaluations of training programs with experimental data , author=. The American Economic Review , pages=. 1986 , publisher=

    1986

  56. [56]

    Journal of Business & Economic Statistics , volume=

    Conditional stochastic dominance testing , author=. Journal of Business & Economic Statistics , volume=. 2013 , publisher=

    2013

  57. [57]

    The Econometrics Journal , volume=

    Consistent tests for conditional treatment effects , author=. The Econometrics Journal , volume=. 2017 , publisher=

    2017

  58. [58]

    Journal of Econometrics , volume=

    Testing functional inequalities , author=. Journal of Econometrics , volume=. 2013 , publisher=

    2013

  59. [59]

    Studies in the economics of uncertainty: In honor of Josef Hadar , pages=

    Testing for stochastic dominance , author=. Studies in the economics of uncertainty: In honor of Josef Hadar , pages=. 1989 , publisher=

    1989

  60. [60]

    Econometrica , volume=

    Consistent tests for stochastic dominance , author=. Econometrica , volume=. 2003 , publisher=

    2003

  61. [61]

    The Econometrics Journal , volume=

    Nonparametric tests of conditional treatment effects with an application to single-sex schooling on academic achievements , author=. The Econometrics Journal , volume=. 2015 , publisher=

    2015

  62. [62]

    Journal of the American Statistical Association , year=

    A sparse random projection-based test for overall qualitative treatment effects , author=. Journal of the American Statistical Association , year=

  63. [63]

    Statistics in Medicine , volume=

    Qualitative interaction trees: a tool to identify qualitative treatment--subgroup interactions , author=. Statistics in Medicine , volume=. 2014 , publisher=

    2014

  64. [64]

    Statistics in Medicine , volume=

    A comparison of the power of two tests for qualitative interactions , author=. Statistics in Medicine , volume=. 1993 , publisher=

    1993

  65. [65]

    Epidemiologic methods , volume=

    A tutorial on interaction , author=. Epidemiologic methods , volume=. 2014 , publisher=

    2014

  66. [66]

    The Annals of Applied Statistics , pages=

    Estimating treatment effect heterogeneity in randomized program evaluation , author=. The Annals of Applied Statistics , pages=. 2013 , publisher=

    2013

  67. [67]

    Statistics in Medicine , volume=

    Test for qualitative interaction of clinical significance , author=. Statistics in Medicine , volume=. 1997 , publisher=

    1997

  68. [68]

    Biometrics , pages=

    Testing for qualitative interactions between treatment effects and patient subsets , author=. Biometrics , pages=. 1985 , publisher=

    1985

  69. [69]

    Journal of Biopharmaceutical Statistics , volume=

    Variable selection for qualitative interactions in personalized medicine while controlling the family-wise error rate , author=. Journal of Biopharmaceutical Statistics , volume=. 2011 , publisher=

    2011

  70. [70]

    International Conference on Machine Learning , pages=

    Counterfactual risk minimization: Learning from logged bandit feedback , author=. International Conference on Machine Learning , pages=

  71. [71]

    International Conference on Artificial Intelligence and Statistics , pages=

    Sublinear optimal policy value estimation in contextual bandits , author=. International Conference on Artificial Intelligence and Statistics , pages=

  72. [72]

    American Economic Review , volume=

    Does job corps work? Impact findings from the national job corps study , author=. American Economic Review , volume=. 2008 , publisher=

    2008

  73. [73]

    Biometrika , volume=

    Quasi-oracle estimation of heterogeneous treatment effects , author=. Biometrika , volume=. 2021 , publisher=

    2021

  74. [74]

    The Econometrics Journal , volume=

    Debiased machine learning of conditional average treatment effects and other causal functions , author=. The Econometrics Journal , volume=. 2021 , publisher=

    2021

  75. [75]

    Advances in Neural Information Processing Systems , volume=

    Sequential experimental design for transductive linear bandits , author=. Advances in Neural Information Processing Systems , volume=

  76. [76]

    Advances in Neural Information Processing Systems , volume=

    Design of experiments for stochastic contextual linear bandits , author=. Advances in Neural Information Processing Systems , volume=

  77. [77]

    Artificial Intelligence in Marketing , pages=

    AI and personalization , author=. Artificial Intelligence in Marketing , pages=. 2023 , publisher=

    2023

  78. [78]

    Journal of the American Statistical Association , volume=

    Estimating individualized treatment rules using outcome weighted learning , author=. Journal of the American Statistical Association , volume=. 2012 , publisher=

    2012

  79. [79]

    Proceedings of the National Academy of Sciences , volume=

    Scaling up behavioral science interventions in online education , author=. Proceedings of the National Academy of Sciences , volume=. 2020 , publisher=

    2020

  80. [80]

    International Conference on Machine Learning , pages=

    Optimal and adaptive off-policy evaluation in contextual bandits , author=. International Conference on Machine Learning , pages=. 2017 , organization=

    2017

Showing first 80 references.