Pith. sign in

REVIEW 3 major objections 3 minor 69 references

Quantile-Optimal Policy Learning under Unmeasured Confounding

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

Pith's one-line read This paper proves that pessimistic policies can achieve near-$\tilde{O}(n^{-1/2})$ quantile regret under unmeasured confounding when an instrument or a negative control is observed.

desk verdict First pessimism framework for quantile-optimal policy learning under unmeasured confounding, but the advertised 'mild' coverage assumption is an operator invertibility condition that can fail, and the fallback to single-policy concentrability is unproven. read the letter →

arxiv 2506.07140 v1 pith:OJZVTSSC submitted 2025-06-08 stat.ML cs.LGecon.EM

classification stat.MLcs.LGecon.EM MSC 62G0562G2062G08
keywords quantile-optimalpolicylearningunmeasuredconfoundinginstrumentalvariablesnegativecontrolsofflinedecisionmakingpessimismconditionalmomentrestrictionminimaxestimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that quantile-optimal policies---policies that maximize the $\alpha$-quantile rather than the mean of the reward distribution---can be learned from offline data even when unobserved confounders bias the action--reward relationship, as long as an instrumental variable or a negative control pair is observed. To do this, it identifies the structural quantile function as the solution of a nonlinear conditional moment restriction, estimates it by a minimax procedure, and then applies pessimism to remain safe under limited coverage of the offline dataset. The central guarantee is a quantile regret of order $\tilde{O}(n^{-1/2})$ for the solution-set pessimistic policy, matching the rate of expected-reward methods despite the nonlinearity of the quantile objective. A regularized variant has regret $O(n^{-(1/2)(1-\varepsilon_R)})$ for an arbitrarily small $\varepsilon_R$, trading a little rate for computational tractability.

What carries the argument

The central object is the nonlinear operator $(T h)(x,z) = \mathbb{E}[\mathbf{1}\{Y \le h(A,X)\}-\alpha \mid X=x,Z=z]$, whose zero set contains the structural quantile function $h^\ast_{\alpha}$. The argument replaces the squared conditional expectation $\|T h\|_2^2$ by a Fenchel-dual minimax loss computed from data, constructs a solution set of $h$'s whose empirical loss is within $e_n$ of the minimum, and proves that with high probability $h^\ast_{\alpha}$ lies in the set and every $h$ in the set has $\|T h\|_2 = O(\sqrt{e_n}+\eta_n)$. The bridge from $\|T h\|_2$ to regret is the pathwise derivative of $T$ at $h^\ast_{\alpha}$, which defines a pseudo-metric $\|h-h^\ast_{\alpha}\|_{ps}$; a local-curvature assumption bounds this pseudo-metric by $c_0\|T h\|_2$, and a change-of-measure assumption converts the pseudo-metric into expected regret under the optimal policy.

What would settle it

Generate offline data from a confounded model in which the treatment effect is multiplicative rather than additive, e.g. $Y = \exp(h^\ast_{\alpha}(A,X)+\epsilon)$ with $\mathbb{P}(\epsilon\le 0 \mid X,Z)=\alpha$, so the conditional $\alpha$-quantile of $Y$ is $e^{h^\ast_{\alpha}}$ rather than $h^\ast_{\alpha}$. If the moment restriction is solved in this model, the estimated $h$ will target the wrong policy and the claimed quantile regret will not vanish; observing that failure would refute the identification step.

Watch

Extended reading notes

Core claim

The paper's central claim is that, under the additive structural quantile model $Y = h^\ast_{\alpha}(A,X)+\epsilon$ with $\mathbb{P}(\epsilon\le 0 \mid X,Z)=\alpha$, the causal quantile function $h^\ast_{\alpha}$ is identified by the conditional moment restriction $\mathbb{E}[\mathbf{1}\{Y \le h(A,X)\}-\alpha \mid X,Z]=0$, and that solving this restriction via minimax estimation is enough to build policies that are near-optimal in quantile regret. The proof route is: any $h$ in the constructed solution set has conditional moment error $\|T h\|_2 = O(\sqrt{e_n}+\eta_n)$; a local curvature assumption converts that error into a bound on the pseudo-metric $\|h-h^\ast_{\alpha}\|_{ps}$; and a change-of-measure assumption converts the pseudo-metric error into regret under the optimal interventional policy. With threshold $e_n=O(\eta_n^2)$ and $\eta_n=\tilde{O}(n^{-1/2})$ for linear function classes, the solution-set policy has regret $c_0\|b\|_2\cdot\tilde{O}(n^{-1/2})$, and the regularized policy has regret $O(\eta_n^{1-\varepsilon_R})$. The same machinery is extended to negative controls by solving two coupled conditional moment restrictions.

Load-bearing premise

The load-bearing premise is Assumption 3.1: the observed reward decomposes as $Y = h^\ast_{\alpha}(A,X)+\epsilon$ with the $\alpha$-quantile of $\epsilon$ equal to zero conditional on $(X,Z)$, so that the conditional moment restriction identifies the causal quantile function; if the outcome's quantile structure is not additive in this way, the identified function need not be the structural quantile function.

Editorial extensions

If this is right

  • If the paper's guarantees hold, offline quantile-optimal policy learning under unmeasured confounding is no harder, in rate, than expected-reward policy learning: both achieve $\tilde{O}(n^{-1/2})$ regret in the solution-set variant.
  • Pessimism removes the need for global coverage of the state-action space; it is enough that the offline data cover the distribution induced by the oracle optimal policy, which the change-of-measure assumption formalizes.
  • Both instrumental variables and negative controls can serve as the identification device, giving practitioners two routes depending on which auxiliary variables are available.
  • The regularized algorithm gives a practical, gradient-friendly objective whose regret is $O(n^{-(1/2)(1-\varepsilon_R)})$, an arbitrarily small rate loss relative to the solution-set version.
  • The analysis framework extends beyond quantiles to any objective identified by a nonlinear conditional moment restriction, including spectral-risk-measure objectives.

Reading between the lines

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

  • If the local-curvature analysis is as general as the paper suggests, the same minimax-plus-pessimism template should transfer to other nonsmooth functionals of the reward distribution, such as conditional value-at-risk or expectile-based objectives.
  • The requirement of continuous rewards is a technical boundary: a Gaussian smoothing step, which the paper itself sketches, would make the method applicable to discrete outcomes, at the cost of a bias term worth quantifying.
  • The change-of-measure function $b(X,Z)$ is a sensitivity knob: computing an estimate of $\|b\|_2$ from data would give practitioners a diagnostic for whether the offline dataset can support quantile-optimal policy learning at all.
  • Because the theory only needs coverage of the optimal policy's distribution, the method should also work in settings where exploration is targeted or logged by a different behavioural policy, provided the change-of-measure condition is satisfied.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper studies offline quantile-optimal policy learning when there are unmeasured confounders, using instrumental variables or negative controls. It identifies the structural quantile function through a conditional moment restriction derived from an additive structural quantile model (Assumption 3.1), estimates it with a minimax nonlinear functional integral equation approach, and constructs two pessimistic policies: a solution-set version (Algorithm 1) and a regularized version (Algorithm 2). The main theoretical results are Theorem 4.5 (uncertainty quantification for the solution set), Theorem 4.8 (consistency), Theorem 4.11 (regret O(sqrt(ε_n)+η_n) for the solution-set policy), Theorem 4.15 (regret O(η_n^{1-ε_R}) for the regularized policy), and Theorem I.7 (η_n = tilde O(n^{-1/2}) for linear function classes). Negative-control counterparts appear in Sections C and D, and simulations compare the regularized pessimistic policy with a greedy baseline.

Significance. The problem formulation is timely, and the proof machinery is substantial: the Fenchel-dual minimax loss, bracketing-number concentration arguments, local expansion around a nonlinear conditional-moment operator, and the pessimism regret decomposition are nontrivial. If the rate claim were established under a genuinely mild coverage condition, the paper would be an important first step for quantile offline policy learning under confounding. However, the paper's own Assumption 4.10 is not a mild coverage condition; it is a positivity/range condition on a density-weighted conditional expectation operator. The stress-test counterexample lands: under Assumption 3.1 one can have f_ε(0)=0, in which case Eq. (4.4) has no square-integrable solution b, so the claim that Assumption 4.10 is weaker than single-policy concentrability is false. The derivation under the stated (strengthened) conditions appears coherent and the appendices are detailed, so the central result is salvageable by adding an explicit positivity condition and revising the abstract and the Appendix A discussion.

major comments (3)
  1. [§4.3, Assumption 4.10; Appendix A] Equation (4.4) is not a coverage condition: the conditional density p_{Y|A,X,Z}(h*_α(A,X)) appears as a multiplicative weight inside the expectation, so Assumption 4.10 also requires a positivity/range condition on the weighted operator b ↦ E[b(X,Z)p_{Y|A,X,Z}(h*_α(A,X)) | A,X]. Concretely, take ε independent of (A,X,Z) with CDF F(t)=0.5+0.5t^3 on [-1,1] and set h*_α=0, Y=ε. Then Assumptions 3.1 and 4.7 hold with α=0.5, but f_ε(0)=0, so the left side of (4.4) is identically zero for every b; even when the target policy is the behavior policy and rp=p, the right side equals 1, so no b exists. The Appendix A assertion that the tabular rank condition rank[P(Z|A,X)] ≥ |A||X| suffices is therefore incomplete, and the claim that Theorems 4.11 and 4.15 remain valid under the usual single-policy concentrability is unsupported by the proof, which uses b in an essential way through the Cauchy-Schwarz step.
  2. [§4.2, Assumption 4.9 and Lemma G.1] The pseudo-metric in (4.3) is degenerate in exactly the same class of models. Lemma G.1 shows that ||h−h*||_ps equals sqrt(E[(E[p_{Y|A,X,Z}(h*)(h−h*)|X,Z])^2]), so when p_{Y|A,X,Z}(h*)=0 the pseudo-metric vanishes identically. In that case Assumption 4.9 is vacuous: the bound ||h−h*||_ps ≤ c0||Th||_2 carries no information, and the local expansion cannot convert the RMSE bound of Theorem 4.5(ii) into a bound on the interventional regret term in Corollary 4.6. A uniform lower bound on p_{Y|A,X,Z}(h*)—equivalently, under Assumption 3.1, a positivity condition on the conditional density of ε at zero—is needed for the local-curvature argument to be non-degenerate. This positivity condition should be stated explicitly and included in the statements of Theorems 4.11 and 4.15.
  3. [Abstract and §4.4, Theorem 4.15] The abstract says that the policies learned by 'these methods' are tilde O(n^{-1/2}) quantile-optimal, but Theorem 4.15 gives regret O(η_n^{1−ε_R}) for the regularized algorithm for any fixed ε_R ∈ (0,1). With η_n = tilde O(n^{-1/2}) from Theorem I.7, this is tilde O(n^{-(1−ε_R)/2}), which is not tilde O(n^{-1/2}) for any fixed ε_R>0. Letting ε_R depend on n and tend to zero is not covered by the proof, since λ_n=η_n^{−(1+ε_R)} and the consistency argument in Theorem 4.14 treat ε_R as fixed. The abstract and the discussion following Theorem 4.15 should either restrict the n^{-1/2} claim to the solution-set algorithm or provide a sharper analysis of the regularized algorithm.
minor comments (3)
  1. [§4.3 and §4.4] The events EXt{...} and ERϵ are used in the proofs before being defined in the main text; please define these high-probability events explicitly before Theorems 4.11 and 4.15.
  2. [§5.1] The sentence 'As shown in (3.10), we generate Y according to the structural model' should refer to Eq. (3.1), not Eq. (3.10).
  3. [Abstract and Introduction] The abstract and introduction should state that the identification argument rests on the additive structural quantile model in Assumption 3.1; without that model, the conditional moment restriction (3.2) need not identify the interventional quantile function, so describing the conditions as 'mild' overstates the scope.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the quantile-regret guarantee follows from stated assumptions and self-contained proofs rather than from fitted constants or load-bearing self-citations.

full rationale

The derivation chain is self-contained. Assumption 3.1 defines the structural quantile model, which directly yields the conditional moment restriction T^IV h* = 0; estimation then minimizes the Fenchel-dual empirical loss (3.5), and Theorem 4.5 bounds the population RMSE ||T h||_2 on the solution set using the concentration event (4.1) and Condition 4.4. The regret bound in Theorem 4.11 is obtained by the algebraic pessimism decomposition (4.2), Assumption 4.10's change of measure, and Assumption 4.9's local-curvature inequality; none of these quantities is fitted to data or renamed as a prediction. The n^{-1/2} rate is supplied by Theorem I.7, an empirical-process calculation for linear classes with proofs in Appendix I. Self-citations (Chen et al. 2023 for Theorem 4.5 parallels and covering-number lemmas I.5/J.3; Chen and Pouzo 2012 and Chen et al. 2014 for identification regularity) support standard technical steps and are not the origin of the main quantile-regret conclusion; the main theorem is proved in the appendix. Appendix A's claim that Assumption 4.10 is a weaker substitute for single-policy concentrability is a substantive correctness concern (e.g., the density p_Y(h*) can vanish, making (4.4) unsolvable), but that is an assumption-validity issue, not circular reasoning. No prediction reduces by construction to an input, and no fitted parameter is relabeled as a result.

Assumptions & free parameters 3 free parameters · 9 assumptions · 0 invented entities

The main result rests on a standard set of nonparametric IV/NC identifiability assumptions plus a local curvature condition that is not quantified. The strongest assumptions are the additive structural quantile model (3.1) and the change-of-measure condition (4.10); both are domain assumptions. No novel physical entities or forces are introduced, and no numerical constants are fitted to data in the theoretical derivations.

free parameters (3)
  • solution set threshold epsilon_n = O(eta_n^2)
    Hyperparameter controlling the width of the pessimistic solution set; chosen theoretically as O(eta_n^2) to ensure h* is in S and regret is O(eta_n). Not fitted to data in the paper.
  • regularization weight lambda_n = eta_n^{-(1+epsilon_R)}
    Hyperparameter for the regularized algorithm; set from theory to balance bias and uncertainty. In experiments, lambda_n = sqrt(n) is used.
  • function class dimensions H_n, J_n = O(log n)
    Dimensions of the linear feature spaces H and Theta in the theory; chosen to let eta_n = O_tilde(n^{-1/2}).
assumptions (9)
  • domain assumption Assumption 3.1: Y = h*_alpha(A,X) + epsilon and P(epsilon <= 0 | X, Z) = alpha.
    This additive structural quantile model is the identifying restriction that all later conditional moment conditions rely on. It is a nontrivial functional form assumption, not implied by the causal graph alone.
  • domain assumption Assumption 4.1: h* is in H and is the unique solution to T h = 0.
    Global identifiability and realizability are needed for the solution set to concentrate on h*. The paper notes that careful sufficient conditions are rare; Wong (2022) is cited as the only study with explicit global identification conditions.
  • domain assumption Assumption 4.2: For all h in H, inf_theta ||theta - T h||_2 = O(n^{-1/2}).
    The test function class must be rich enough to approximate T h, so that the minimax loss can represent the conditional RMSE.
  • domain assumption Assumption 4.7: The conditional density of Y given (A,X,Z) is continuous and bounded.
    Needed to ensure T is continuous and the pathwise derivative exists. This excludes discrete outcomes; the paper suggests smoothing for discrete Y.
  • domain assumption Assumption 4.9: ||h - h*||_ps <= c0 ||T h||_2 near h*.
    This local curvature condition links the pseudo-metric to the RMSE, and is what turns RMSE bounds into regret bounds. If c0 is large or unbounded, the 1/sqrt(n) rate does not follow.
  • domain assumption Assumption 4.10: There exists b(X,Z) such that E[b(X,Z) p_Y|A,X,Z(h*(A,X)) | A=a, X=x] = rp(x) pi*(a|x) / p(x,a).
    Change-of-measure condition that requires the offline data to cover the oracle policy. The abstract calls this mild, but it is a nontrivial completeness-type condition.
  • domain assumption Condition 4.4: A uniform concentration event holds with rate eta_n.
    Requires the empirical loss to concentrate uniformly over the function class Q; verified for linear classes in Appendix I using bracketing numbers.
  • domain assumption Assumption 4.13: inf_h L_n(h) = O(eta_n^2).
    For the regularized algorithm, this sample criterion is assumed rather than verified. It is satisfied when empirical loss can be minimized to a negligible level, but the paper does not prove it for the proposed classes.
  • domain assumption Assumption C.1: Negative control conditions with latent unconfoundedness and completeness of E.
    For the NC extension, requires V independent of A given U,X and completeness of E to identify the bridge functions. Standard in proximal causal inference but strong.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantile-Optimal Policy Learning under Unmeasured Confounding." pith.science (2026). https://pith.science/paper/OJZVTSSC

@misc{pith2026250607140,
  author       = {Pith},
  title        = {Pith review of: Quantile-Optimal Policy Learning under Unmeasured Confounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJZVTSSC}},
  note         = {Machine review of arXiv:2506.07140}
}
abstract

We study quantile-optimal policy learning where the goal is to find a policy whose reward distribution has the largest $\alpha$-quantile for some $\alpha \in (0, 1)$. We focus on the offline setting whose generating process involves unobserved confounders. Such a problem suffers from three main challenges: (i) nonlinearity of the quantile objective as a functional of the reward distribution, (ii) unobserved confounding issue, and (iii) insufficient coverage of the offline dataset. To address these challenges, we propose a suite of causal-assisted policy learning methods that provably enjoy strong theoretical guarantees under mild conditions. In particular, to address (i) and (ii), using causal inference tools such as instrumental variables and negative controls, we propose to estimate the quantile objectives by solving nonlinear functional integral equations. Then we adopt a minimax estimation approach with nonparametric models to solve these integral equations, and propose to construct conservative policy estimates that address (iii). The final policy is the one that maximizes these pessimistic estimates. In addition, we propose a novel regularized policy learning method that is more amenable to computation. Finally, we prove that the policies learned by these methods are $\tilde{\mathscr{O}}(n^{-1/2})$ quantile-optimal under a mild coverage assumption on the offline dataset. Here, $\tilde{\mathscr{O}}(\cdot)$ omits poly-logarithmic factors. To the best of our knowledge, we propose the first sample-efficient policy learning algorithms for estimating the quantile-optimal policy when there exist unmeasured confounding.

Figures

Figures reproduced from arXiv: 2506.07140 by the authors.

Figure 1
Figure 1. (a) A DAG illustrating the causal relationship between random variables during ODCP [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. (a) A DAG illustrating the causal relationship between random variables of the ODCP [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Average regret of the greedy and pessimistic approaches for different sample sizes [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Average regret of the greedy and pessimistic approaches for different sample sizes [PITH_FULL_IMAGE:figures/full_fig_p028_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 60 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := ...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    , Angrist, J

    Abadie, A. , Angrist, J. and Imbens, G. (2002). Instrumental variables estimates of the effect of subsidized training on the quantiles of trainee earnings. Econometrica, 70 91--117

  4. [4]

    and Chen, X

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

  5. [5]

    and Chen, X

    Ai, C. and Chen, X. (2012). The semiparametric efficiency bound for models of sequential moment restrictions containing unknown functions. Journal of Econometrics, 170 442--457

  6. [6]

    Angrist, J. D. , Imbens, G. W. and Rubin, D. B. (1996). Identification of causal effects using instrumental variables. Journal of the American statistical Association, 91 444--455

  7. [7]

    , Cheng, J

    Baiocchi, M. , Cheng, J. and Small, D. S. (2014). Instrumental variable methods for causal inference. Statistics in medicine, 33 2297--2340

  8. [8]

    , Gelada, C

    Buckman, J. , Gelada, C. and Bellemare, M. G. (2020). The importance of pessimism in fixed-dataset policy optimization. arXiv preprint arXiv:2009.06799

Show all 69 references
  1. [9]

    , Mannor, S

    Cassel, A. , Mannor, S. and Zeevi, A. (2023). A general framework for bandit problems beyond cumulative objectives. Mathematics of Operations Research

  2. [10]

    Chamberlain, G. (1992). Efficiency bounds for semiparametric regression. Econometrica: Journal of the Econometric Society 567--596

  3. [11]

    , Wang, Y

    Chen, S. , Wang, Y. , Wang, Z. and Yang, Z. (2023). A unified framework of policy learning for contextual bandit with confounding bias and missing observations. arXiv preprint arXiv:2303.11187

  4. [12]

    , Chernozhukov, V

    Chen, X. , Chernozhukov, V. , Lee, S. and Newey, W. K. (2014). Local identification of nonparametric and semiparametric models. Econometrica, 82 785--809

  5. [13]

    , Linton, O

    Chen, X. , Linton, O. and Van Keilegom, I. (2003). Estimation of semiparametric models when the criterion function is not smooth. Econometrica, 71 1591--1608

  6. [14]

    and Pouzo, D

    Chen, X. and Pouzo, D. (2009). On nonlinear ill-posed inverse problems with applications to pricing of defaultable bonds and option pricing. Science in China Series A: Mathematics, 52 1157--1168

  7. [15]

    and Pouzo, D

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

  8. [16]

    and Pouzo, D

    Chen, X. and Pouzo, D. (2015). Sieve wald and qlr inferences on semi/nonparametric conditional moment models. Econometrica, 83 1013--1079

  9. [17]

    and Hansen, C

    Chernozhukov, V. and Hansen, C. (2005). An iv model of quantile treatment effects. Econometrica, 73 245--261

  10. [18]

    and Hansen, C

    Chernozhukov, V. and Hansen, C. (2008). Instrumental variable quantile regression: A robust inference approach. Journal of Econometrics, 142 379--398

  11. [19]

    Dai, B. , He, N. , Pan, Y. , Boots, B. and Song, L. (2017). Learning from conditional distributions via dual embeddings. In Artificial Intelligence and Statistics. PMLR

  12. [20]

    , Lewis, G

    Dikkala, N. , Lewis, G. , Mackey, L. and Syrgkanis, V. (2020). Minimax estimation of conditional moment models. Advances in Neural Information Processing Systems, 33 12248--12262

  13. [21]

    Dong, J. , Mo, W. , Qi, Z. , Shi, C. , Fang, E. X. and Tarokh, V. (2023). Pasta: pessimistic assortment optimization. In International Conference on Machine Learning. PMLR

  14. [22]

    and Blake, D

    Dowd, K. and Blake, D. (2006). After var: the theory, estimation, and insurance applications of quantile-based risk measures. Journal of Risk and Insurance, 73 193--229

  15. [23]

    Fang, E. X. , Wang, Z. and Wang, L. (2023). Fairness-oriented learning for optimal individualized treatment rules. Journal of the American Statistical Association, 118 1733--1746

  16. [24]

    and Scaillet, O

    Gagliardini, P. and Scaillet, O. (2012). Nonparametric instrumental variable estimation of structural quantile effects. Econometrica, 80 1533--1562

  17. [25]

    , Lewis, G

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

  18. [26]

    Hern \'a n, M. A. and Robins, J. M. (2016). Using big data to emulate a target trial when a randomized trial is not available. American journal of epidemiology, 183 758--764

  19. [27]

    Horowitz, J. L. and Lee, S. (2007). Nonparametric instrumental variables estimation of a quantile regression model. Econometrica, 75 1191--1208

  20. [28]

    , Shang, Z

    Hu, T. , Shang, Z. and Cheng, G. (2020). Sharp rate of convergence for deep neural network classifiers under the teacher-student setting. arXiv preprint arXiv:2001.06892

  21. [29]

    , Yang, Z

    Jin, Y. , Yang, Z. and Wang, Z. (2021). Is pessimism provably efficient for offline rl? In International Conference on Machine Learning. PMLR

  22. [30]

    , Mao, X

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

  23. [31]

    , Rajeswaran, A

    Kidambi, R. , Rajeswaran, A. , Netrapalli, P. and Joachims, T. (2020). Morel: Model-based offline reinforcement learning. Advances in neural information processing systems, 33 21810--21823

  24. [32]

    , Zhou, A

    Kumar, A. , Zhou, A. , Tucker, G. and Levine, S. (2020). Conservative q-learning for offline reinforcement learning. Advances in neural information processing systems, 33 1179--1191

  25. [33]

    Lee, B. K. , Lessler, J. and Stuart, E. A. (2010). Improving propensity score weighting using machine learning. Statistics in medicine, 29 337--346

  26. [34]

    , Jeon, W

    Lee, J. , Jeon, W. , Lee, B. , Pineau, J. and Kim, K.-E. (2021). Optidice: Offline policy optimization via stationary distribution correction estimation. In International Conference on Machine Learning. PMLR

  27. [35]

    Lee, S. (2003). Efficient semiparametric estimation of a partially linear quantile regression model. Econometric theory, 19 1--31

  28. [36]

    , Kumar, A

    Levine, S. , Kumar, A. , Tucker, G. and Fu, J. (2020). Offline reinforcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643

  29. [37]

    Li, G. , Ma, C. and Srebro, N. (2022). Pessimism for offline linear contextual bandits using \_p confidence sets. Advances in Neural Information Processing Systems, 35 20974--20987

  30. [38]

    , Chu, W

    Li, L. , Chu, W. , Langford, J. , Moon, T. and Wang, X. (2012). An unbiased offline evaluation of contextual bandit algorithms with generalized linear models. In Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2. JMLR Workshop and Conference Proceedings

  31. [39]

    Linn, K. A. , Laber, E. B. and Stefanski, L. A. (2017). Interactive q-learning for quantiles. Journal of the American Statistical Association, 112 638--649

  32. [40]

    , Tchetgen, E

    Lipsitch, M. , Tchetgen, E. T. and Cohen, T. (2010). Negative controls: a tool for detecting confounding and bias in observational studies. Epidemiology (Cambridge, Mass.), 21 383

  33. [41]

    , Ding, L

    Liu, M. , Ding, L. , Yu, D. , Liu, W. , Kong, L. and Jiang, B. (2022). Conformalized fairness via quantile regression. Advances in Neural Information Processing Systems, 35 11561--11572

  34. [42]

    , Chen, Z

    Liu, Q. , Chen, Z. and Wong, W. H. (2024). An encoding generative modeling approach to dimension reduction and covariate adjustment in causal inference with observational studies. Proceedings of the National Academy of Sciences, 121 e2322376121

  35. [43]

    , Min, Y

    Lu, M. , Min, Y. , Wang, Z. and Yang, Z. (2022). Pessimism in the face of confounders: Provably efficient offline reinforcement learning in partially observable markov decision processes. arXiv preprint arXiv:2205.13589

  36. [44]

    , Giguere, S

    Metevier, B. , Giguere, S. , Brockman, S. , Kobren, A. , Brun, Y. , Brunskill, E. and Thomas, P. S. (2019). Offline contextual bandits with high probability fairness guarantees. Advances in neural information processing systems, 32

  37. [45]

    Miao, R. , Qi, Z. , Shi, C. and Lin, L. (2023). Personalized pricing with invalid instrumental variables: Identification, estimation, and policy learning. arXiv preprint arXiv:2302.12670

  38. [46]

    , Geng, Z

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

  39. [47]

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

  40. [48]

    Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82 669--688

  41. [49]

    , Jagannathan, K

    Prashanth, L. , Jagannathan, K. and Kolla, R. K. (2020). Concentration bounds for cvar estimation: The cases of light-tailed and heavy-tailed distributions. In Proceedings of the 37th International Conference on Machine Learning

  42. [50]

    Prudencio, R. F. , Maximo, M. R. and Colombini, E. L. (2023). A survey on offline reinforcement learning: Taxonomy, review, and open problems. IEEE Transactions on Neural Networks and Learning Systems

  43. [51]

    , Zhu, B

    Rashidinejad, P. , Zhu, B. , Ma, C. , Jiao, J. and Russell, S. (2021). Bridging offline reinforcement learning and imitation learning: A tale of pessimism. Advances in Neural Information Processing Systems, 34 11702--11716

  44. [52]

    , Zhu, H

    Rashidinejad, P. , Zhu, H. , Yang, K. , Russell, S. and Jiao, J. (2022). Optimal conservative offline rl with general function approximation via augmented lagrangian. arXiv preprint arXiv:2211.00716

  45. [53]

    Rosenbaum, P. R. and Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70 41--55

  46. [54]

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

  47. [55]

    Tchetgen, E. J. T. , Ying, A. , Cui, Y. , Shi, X. and Miao, W. (2020). An introduction to proximal causal learning. arXiv preprint arXiv:2009.10982

  48. [56]

    , Imaizumi, M

    Uehara, M. , Imaizumi, M. , Jiang, N. , Kallus, N. , Sun, W. and Xie, T. (2021). Finite sample analysis of minimax offline reinforcement learning: Completeness, fast rates and first-order efficiency. arXiv preprint arXiv:2102.02981

  49. [57]

    van de Geer, S. (2009). Empirical Processes in M-Estimation. Cambridge Series in Statistical and Probabilistic Mathematics, Cambridge University Press. ://books.google.com/books?id=0VEcQAAACAAJ

  50. [58]

    and Wellner, J

    van der vaart, A. and Wellner, J. (2013). Weak Convergence and Empirical Processes: With Applications to Statistics. Springer Series in Statistics, Springer New York. ://books.google.com/books?id=zdDkBwAAQBAJ

  51. [59]

    Van der Vaart, A. W. (2000). Asymptotic statistics, vol. 3. Cambridge university press

  52. [60]

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

  53. [61]

    , Yang, Z

    Wang, L. , Yang, Z. and Wang, Z. (2021). Provably efficient causal reinforcement learning with confounded observational data. Advances in Neural Information Processing Systems, 34 21164--21175

  54. [62]

    Wang, Y. , Fu, H. and Zeng, D. (2018). Learning optimal personalized treatment rules in consideration of benefit and risk: with an application to treating type 2 diabetes patients with insulin therapies. Journal of the American Statistical Association, 113 1--13

  55. [63]

    Wong, W. H. (2022). An equation for the identification of average causal effect in nonlinear models. Statistica Sinica, 32 539--545

  56. [64]

    Wu, T. Z. , Yu, K. and Yu, Y. (2010). Single-index quantile regression. Journal of Multivariate Analysis, 101 1607--1621

  57. [65]

    , Cheng, C.-A

    Xie, T. , Cheng, C.-A. , Jiang, N. , Mineiro, P. and Agarwal, A. (2021). Bellman-consistent pessimism for offline reinforcement learning. Advances in neural information processing systems, 34 6683--6694

  58. [66]

    , Lafferty, J

    Yang, D. , Lafferty, J. and Pollard, D. (2019). Fair quantile regression. arXiv preprint arXiv:1907.08646

  59. [67]

    , Duan, Y

    Yin, M. , Duan, Y. , Wang, M. and Wang, Y.-X. (2022). Near-optimal offline reinforcement learning with linear representation: Leveraging variance information with pessimism. arXiv preprint arXiv:2203.05804

  60. [68]

    , Thomas, G

    Yu, T. , Thomas, G. , Yu, L. , Ermon, S. , Zou, J. Y. , Levine, S. , Finn, C. and Ma, T. (2020). Mopo: Model-based offline policy optimization. Advances in Neural Information Processing Systems, 33 14129--14142

  61. [69]

    and Tan, V

    Zhu, Q. and Tan, V. (2020). Thompson sampling algorithms for mean-variance bandits. In International Conference on Machine Learning. PMLR

Pith tools

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