Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Who With Whom? Learning Optimal Matching Policies

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A planner can learn a near-optimal one-to-one matching policy from conditionally randomized training data by solving an entropy-regularized optimal transport problem, with a welfare regret bound that vanishes as the training sample grows.

desk verdict Entropic OT for matching with estimated costs is a real gap and the paper mostly fills it, but a false mean-value step in the main proof needs repair before the regret bound can be taken as stated. read the letter →

arxiv 2507.13567 v1 pith:U5NMPKBN submitted 2025-07-17 econ.EM

classification econ.EM MSC 62P2091B6890C08
keywords policylearningtwo-sidedmatchingentropy-regularizedoptimaltransportSinkhornalgorithmregretboundjobsearchassistancecostfunctionestimationwelfaremaximization
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

This paper develops a way for a planner to learn which individuals should be matched with which from historical data, when the two sides each have many observable characteristics and matching is one-to-one. The proposal is to estimate the average cost of each possible match from training data and then solve an entropy-regularized optimal transport problem to choose the assignment. The paper's central theoretical result is a finite-sample welfare regret bound: under bounded costs, the expected loss from using the estimated policy instead of the true optimal one is at most an exponential-in-regularization term times the $L^1$ plus squared $L^2$ estimation error of the cost function, plus a regularization bias of $\log n/\eta$. If the cost function is estimated at a parametric rate and $\eta$ is chosen proportional to $\log N$, the regret converges to zero. In simulations calibrated to French administrative data, the method could improve six-month job-finding rates by about one percentage point relative to random assignment, roughly the same order as costly counseling programs.

What carries the argument

The central object is the entropy-regularized empirical optimal transport plan, defined as the minimizer of estimated average match cost plus $(1/\eta)\mathrm{KL}(\pi_n\|\mu_n\otimes\nu_n)$ over couplings with the empirical marginal distributions of the two sides. The KL penalty pushes the solution away from the extreme-point permutations of the linear program, smoothing the policy and limiting its sensitivity to errors in the estimated cost function, while $\eta$ controls how close the plan is to random matching. The dual formulation reduces the decision variables from $n^2$ to $2n$, making the Sinkhorn algorithm applicable, and the plan is converted into a distribution over permutations via the Birkhoff–von Neumann procedure. The proof machinery decomposes regret into a plug-in error term governed by the $L^1$ and squared $L^2$ errors of the cost estimator and a regularization-bias term of $\log n/\eta$.

What would settle it

A randomized field trial that implements the estimated policy in some agencies and keeps random matching in others, and that finds no positive or about-one-percentage-point gain in six-month job-finding rates, would contradict the paper's welfare prediction.

Watch

Extended reading notes

Core claim

The paper claims that a welfare-optimal one-to-one matching policy can be learned even though the match cost function is unknown, by plugging an estimated cost function $\hat{c}$ into an entropy-regularized optimal transport program and applying the Sinkhorn algorithm. Theorem 1 gives, under the single regularity condition that costs are bounded in $[0,\bar c]$, the non-asymptotic bound $$$e^{{2\eta\bar c}}$\left(\|\hat{c}-c\|_{$L^{1}$(\mu\otimes\nu)} + \|\hat{c}-c\|^2_{$L^{2}$(\mu\otimes\nu)}\right) + \frac{\log n}{\eta}$$ for expected regret relative to the oracle unregularized optimal policy. The first term is the cost of estimation error and sampling uncertainty, the second is the regularization bias, and the choice of $\eta$ trades the two off. Expected regret therefore converges to zero when the integrated $L^1$ error of the cost estimator decays and $\eta$ grows like $\log N$. The calibrated job-seeker and caseworker simulations indicate that the learned policy can capture most of the welfare gap between random matching and the oracle optimum, with about a one-percentage-point gain in job-finding rates under realistic training sample sizes and complementarities.

Load-bearing premise

The load-bearing premise is that, in the historical training data, match assignment is as good as random once observed characteristics are accounted for, so the estimated cost function is unbiased.

Editorial extensions

If this is right

  • A public employment service that currently assigns caseworkers randomly can instead learn an assignment from its administrative records and raise average job-finding rates without training new personnel.
  • The same recipe applies whenever a planner controls a one-to-one match and has data from conditionally randomized matches: doctors to patients, teachers to classrooms, attorneys to defendants, and auditors to taxpayers.
  • The $\log n/\eta$ regularization-bias term means some stochasticity in assignments is statistically beneficial when the training sample is small, so the planner should not automatically pick the unregularized optimal transport solution.
  • For matching markets of a few hundred to a few thousand individuals, the Sinkhorn implementation is computationally feasible, unlike the full $n^2$-variable linear program.
  • The regret converges to zero only if the cost estimator's integrated $L^1$ error decays with training sample size; nonparametric estimators give slower rates, but the bound can still vanish with an appropriate choice of $\eta$.

Reading between the lines

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

  • A data-driven choice of $\eta$ is the natural next step: because the regret bound splits into an estimation term and a bias term, a planner could select $\eta$ by cross-validating the estimated policy's average outcome on a hold-out sample, a step the paper does not implement.
  • The method's logic carries over to many-to-one assignment, where each caseworker or doctor has a capacity larger than one, by replacing the uniform marginal constraint on that side with capacity-weighted margins and solving the same regularized transport program.
  • If historical matching was influenced by caseworkers' private information about job seekers, the exogeneity condition fails and the learned policy could systematically reward the types that were historically over-served; the paper's guarantee would then not apply, and a randomized pilot would be required before rollout.
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

2 major / 6 minor

Summary. This paper studies policy learning for two-sided one-to-one matching. A planner observes the characteristics of the individuals to be matched and must choose a matching policy under capacity constraints, while the match cost function is unknown and must be estimated from a training sample. The authors propose solving an entropy-regularized optimal transport problem with the estimated cost function, using the Sinkhorn algorithm for tractability, and they derive a non-asymptotic regret bound relative to the oracle unregularized optimal matching policy. The main theorem (Theorem 1, Eq. (31)) claims the regret is bounded by e^{2ηc̄}(∥ĉ−c∥_{L1} + ∥ĉ−c∥²_{L2}) + log n/η, exhibiting a bias-variance trade-off in the regularization parameter η. The paper also presents two simulation studies, one simple and one calibrated to French job-search data, suggesting that learned matching policies could improve job-finding rates by about one percentage point at essentially no reallocation cost.

Significance. If the main theorem were correct, the paper would make a useful contribution to an understudied area: empirical policy learning for two-sided matching with estimated costs, where the number of 'treatments' grows with the sample and capacity constraints bind. The entropy-regularized formulation is computationally attractive, and a non-asymptotic bound that separates cost-estimation error from regularization bias is exactly the kind of guarantee the literature needs. The calibrated simulation is thoughtful and directly relevant to an important policy problem. However, the central proof contains a false step (Eq. (38)) that invalidates the squared-L2 term in the advertised bound; the likely repair replaces that term with an additional L1 error, which still yields convergence but changes the stated result. The paper also relies on an untestable conditional exogeneity assumption and, as the authors note, has not yet applied the method to real administrative data. For these reasons, the contribution is promising but needs substantial revision before the theoretical claims can be accepted.

major comments (2)
  1. [Appendix A.2, Eq. (38)] The displayed equality is false. The mean value theorem applied to F(t)=Φ_n((1−t)c+tĉ, f̂_n, ĝ_n) gives F(1)−F(0)=F′(t̃) for some t̃∈[0,1], so the integrand is (ĉ−c) exp{−η[(1−t̃)c+t̃ĉ]+η f̂_n+η ĝ_n}, which is linear in (ĉ−c), not quadratic. For n=1 with ĉ−c≡δ, the left-hand side of (38) is of order δ while the right-hand side is of order δ², so inequality (26) is false as stated. Consequently the squared-L2 term in Proposition 1 and in Theorem 1 is not established. A correct bound is |Φ_n(ĉ, f̂_n, ĝ_n)−Φ_n(c, f̂_n, ĝ_n)| ≤ e^{2ηc̄}∥ĉ−c∥_{L1(μn⊗νn)}, which yields ReROT(π̂ROT_n) ≤ 2e^{2ηc̄}∥ĉ−c∥_{L1(μn⊗νn)}. With this repair, the bias-variance trade-off and the convergence conclusion for η∝log N remain qualitatively unchanged, but the theorem and Proposition 1 must be revised accordingly, or a genuinely different argument for the L2² term must be supplied.
  2. [Section 3.1, Eq. (5)] The conditional exogeneity assumption Wℓ⊥ϵℓ|Xℓ is untestable from the training data and is likely to be violated in observational administrative data, where caseworker assignment may depend on unobserved job-seeker characteristics. The paper does not analyze the bias in the estimated cost function or in the resulting regret bound when this assumption fails. Because the motivating application is to French administrative data (and the authors state in Section 6 that the real-data application is still in progress), this is a substantive limitation for interpreting the regret bound as a welfare guarantee. I ask the authors to at least discuss the direction and potential magnitude of the bias, and ideally to provide a sensitivity analysis or a partial-identification extension.
minor comments (6)
  1. [Appendix A.3, Proposition 2] The statement that π*_n is 'supported only on n points with equal probability masses 1/n' is not true for a non-extreme optimal solution of the transport LP. However, an extreme-point optimal solution always exists, and for any feasible coupling KL(π||μn⊗νn)≤log n, so the bound in (30) is valid. Please rephrase the proof to avoid the incorrect assertion about the support of π*_n.
  2. [Section 4, after Theorem 1] The symbol ᾱ in the sentence 'η is chosen to be proportional to ᾱ/(2c̄) log(N) with ᾱ<α' is undefined; it should be a separate parameter (e.g., α̃) with its relation to α stated explicitly.
  3. [Section 5.2] The calibrated parameters (a,b,c,d) of the logistic job-finding function are not reported; providing them would make the simulation reproducible and allow readers to judge the implied degree of complementarity beyond the summary gap γ.
  4. [Figures 5 and 6] Figures 5 and 6 appear to be identical (same captions and same panels for the same values of N and γ). If they are intended to differ, the figures or captions should be corrected.
  5. [References] The reference 'Highbee, S., O. A. Katta, and G. Pouliot' appears to contain a spelling error; please verify the author's name.
  6. [Section 5.1] The rescaled welfare measure is defined using (μn⊗νn)(c) in the text but the notation is not fully consistent across equations; clarify the notation so that the empirical and population measures are distinguished.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regret bound is derived forward from explicit assumptions, and self-citations are not load-bearing.

full rationale

The central derivation, Theorem 1, follows from clearly stated ingredients: Assumption 1 (bounded costs), the dual formulation of entropy-regularized optimal transport, Lemma 1's bounds on the dual potentials and density, Proposition 1's bounds on the regularized-regret components, and Proposition 2's bound on the regularization bias via the KL divergence of a permutation distribution. None of these steps defines a quantity in terms of the target regret or fits a parameter to match the bound. The cost-estimation error terms in the bound are inputs to the guarantee, not outputs that are relabeled as predictions. The paper's self-citations appear only in the literature review and in passing references to prior treatment-choice work; they do not carry any step of the proof. The questionable mean-value expansion in Eq. (38) is a potential correctness concern about the proof, but it is not a circularity: the claimed bound is not assumed as its own input, and the proof would need repair rather than reduction to a fit. The numerical simulations are calibrated to external estimates from Dromundo and Haramboure (2022), and their reported gains are computed from the resulting DGP rather than by fitting the reported outcome. Therefore no circular step is exhibited, and the paper is not circular in the sense assessed here.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central theoretical claim rests on exogeneity of training matches, no-interference, bounded costs, and standard transportation theorems; none of these are circular. The numerical headline rests on hand-chosen regularization values and calibrated logistic DGP parameters that are not reported. No new entities are introduced.

free parameters (2)
  • regularization parameter 1/η = 0, 0.002, 0.01, 0.05 in simulations
    Chosen by hand; Section 6 notes there is no data-driven tuning method, so the numerical results depend on this choice.
  • logistic DGP parameters a, b, c, d in Section 5.2 = not reported; solved from four calibration equations
    Calibrated to match job-finding rates and complementarity gaps from Dromundo and Haramboure (2022); the simulated one percentage point gain depends on these fitted constants.
assumptions (5)
  • domain assumption Training matches satisfy conditional exogeneity Wℓ ⊥ ϵℓ | Xℓ
    Identifies c(x,w) from training data; Section 3.1, Eq.(5).
  • domain assumption No-interference: Yi(j) ⊥ (Xn,Wn) | Xi,Wj
    Makes expected total cost additive over pairs; Section 2.2, Eq.(1).
  • domain assumption Cost function and estimator are nonnegative and bounded above by c̄
    Assumption 1; used for Lemma 1 and the exponential constant in the regret bound.
  • standard math Birkhoff-von Neumann theorem
    Reduces the permutation assignment problem to the transportation polytope; Section 2.2.
  • domain assumption Individuals to be matched are iid draws from population distributions µ and ν
    Used in Theorem 1 to replace empirical norms with population norms in the expected regret.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Who With Whom? Learning Optimal Matching Policies." pith.science (2026). https://pith.science/paper/U5NMPKBN

@misc{pith2026250713567,
  author       = {Pith},
  title        = {Pith review of: Who With Whom? Learning Optimal Matching Policies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U5NMPKBN}},
  note         = {Machine review of arXiv:2507.13567}
}
read the original abstract

There are many economic contexts where the productivity and welfare performance of institutions and policies depend on who matches with whom. Examples include caseworkers and job seekers in job search assistance programs, medical doctors and patients, teachers and students, attorneys and defendants, and tax auditors and taxpayers, among others. Although reallocating individuals through a change in matching policy can be less costly than training personnel or introducing a new program, methods for learning optimal matching policies and their statistical performance are less studied than methods for other policy interventions. This paper develops a method to learn welfare optimal matching policies for two-sided matching problems in which a planner matches individuals based on the rich set of observable characteristics of the two sides. We formulate the learning problem as an empirical optimal transport problem with a match cost function estimated from training data, and propose estimating an optimal matching policy by maximizing the entropy regularized empirical welfare criterion. We derive a welfare regret bound for the estimated policy and characterize its convergence. We apply our proposal to the problem of matching caseworkers and job seekers in a job search assistance program, and assess its welfare performance in a simulation study calibrated with French administrative data.

Figures

Figures reproduced from arXiv: 2507.13567 by the authors.

Figure 1
Figure 1. Estimated cost functions ˆc(x, w) for different training sample sizes, and the true cost function c(x, w) 17 [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. Relative welfare gain of the ROT plan learnt using the true cost [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Relative welfare gain for ROT plan learnt using estimated cost ˆc [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Distribution of the simulated job seekers’s characteristic [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Estimated cost functions ˆc(x, w) for different training sample sizes, and the true cost function c(x, w) (γ = 0.06) 23 [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: Estimated cost functions ˆc(x, w) for different training sample sizes, and the true cost function c(x, w) (γ = 0.06) Welfare evaluation of the different transport plans We compare the average (over training samples) welfare generated by the following matching policies:…
Figure 8
Figure 8. Figure 8: This is because we set the job finding probability function to be a logistic function. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 7
Figure 7. Figure 7: Absolute and relative gains from oracle ROT plans [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 8
Figure 8. Figure 8: Expected (over training set samples) joint distribution of ( [PITH_FULL_IMAGE:figures/full_fig_p028_8.png]
Figure 9
Figure 9. Figure 9: Joint distribution of (X, W) implied by the random vs. various feasible transport plans, for complementarities set at γ = 0.06 (cost function estimated with N = 5, 000, 000) [PITH_FULL_IMAGE:figures/full_fig_p029_9.png]
Figure 10
Figure 10. Figure 10: Absolute and relative gains from oracle ROT plans [PITH_FULL_IMAGE:figures/full_fig_p030_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Optimal Transport under Group Fairness Constraints

    stat.ML 2026-01 conditional novelty 6.0 of 10

    Group-fairness targets are added as constraints to entropic optimal transport, with a modified Sinkhorn algorithm and two relaxations (penalty and cost learning) that come with sample-complexity bounds.

Reference graph

Works this paper leans on

55 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Adjaho, C. and T. Christensen (2022): Externally valid treatment choice, arXiv preprint arXiv:2205.05561

  2. [2]

    Fang, and H

    Ai, C., Y. Fang, and H. Xie (2024): Data-driven policy learning for continuous treatments, arXiv preprint arXiv:2402.02535

  3. [3]

    Athey, S. and S. Wager (2021): Policy Learning With Observational Data, Econometrica, 89, 133--161

  4. [4]

    Cr \'e pon, and M

    Behaghel, L., B. Cr \'e pon, and M. Gurgand (2014): Private and public provision of counseling to job seekers: Evidence from a large controlled experiment, American Economic Journal: Applied Economics, 6, 142--174

  5. [5]

    (2009): Inferring optimal peer assignment from experimental data, Journal of the American Statistical Association, 104, 486--500

    Bhattacharya, D. (2009): Inferring optimal peer assignment from experimental data, Journal of the American Statistical Association, 104, 486--500

  6. [6]

    Bhattacharya, D. and P. Dupas (2012): Inferring welfare maximizing treatment assignment under budget constraints, Journal of Econometrics, 167, 168--196

  7. [7]

    (1946): Tres Observaciones Sobre el Algebra Lineal, Universidad Nacionale de Tucum\' a n Revista Series

    Birkhoff, G. (1946): Tres Observaciones Sobre el Algebra Lineal, Universidad Nacionale de Tucum\' a n Revista Series

  8. [8]

    Galichon, M

    Chernozhukov, V., A. Galichon, M. Hallin, and M. Henry (2017): Monge-Kantorovich depth, quantiles, ranks and signs, Annals of Statistics, 45, 223--256

Show all 55 references
  1. [9]

    Christensen, T., H. R. Moon, and F. Schorfheide (2022): Optimal decision rules when payoffs are partially identified, arXiv preprint arXiv:2204.11748

  2. [10]

    (2013): Sinkhorn Distances: Lightspeed Computation of Optimal Transport, in Advances in Neural Information Processing Systems, ed

    Cuturi, M. (2013): Sinkhorn Distances: Lightspeed Computation of Optimal Transport, in Advances in Neural Information Processing Systems, ed. by C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger, Curran Associates, Inc., vol. 26

  3. [11]

    (2024): Defying Distance? The Provision of Services in the Digital Age, Working Paper

    Dahlstrand, A. (2024): Defying Distance? The Provision of Services in the Digital Age, Working Paper

  4. [12]

    Dehejia (2005): Program evaluation as a decision problem, Journal of Econometrics, 125, 141--173

  5. [13]

    Gaillac, and A

    D'Haultfoeuille, X., C. Gaillac, and A. Maurel (2025): Partially Linear Models under Data Combination, The Review of Economic Studies, 92, 238--267

  6. [14]

    Dromundo, S. and A. Haramboure (2022): Do Job Counselors Matter? Measuring Counselor Value-Added in Job Search, Working Paper

  7. [15]

    Dupuy, A. and A. Galichon (2014): Personality Traits and the Marriage Market, Journal of Political Economy, 122, 1271--1319

  8. [16]

    Heckman, and L

    Ekeland, I., J. Heckman, and L. Nesheim (2004): Identification and estimation of hedonic models, Journal of Political Economy, 112, 60--109

  9. [17]

    (2016): Optimal Transport Methods in Economics, Princeton University Press

    Galichon, A. (2016): Optimal Transport Methods in Economics, Princeton University Press

  10. [18]

    Galichon, A. and M. Henry (2011): Set identification in models with multiple equilibria, Review of Economic Studies, 78, 1264--1298

  11. [19]

    Galichon, A. and B. Salani\' e (2022): Cupid's Invisible Hand: Social Surplus and Identification in Matching Models, Review of Economic Studies, 89, 2600--2629

  12. [20]

    Chizat, F

    Genevay, A., L. Chizat, F. Bach, M. Cuturi, and G. Peyré (2019): Sample Complexity of Sinkhorn Divergences, Proceedings of the 22nd International Conference on Ar- tificial Intelligence and Statistics (AISTATS)

  13. [21]

    Graham, B. S. (2011): Econometric methods for the analysis of assignment problems in the presence of complementarity and social spillovers, Handbook of Social Economics, 1, 965--1052

  14. [22]

    Graham, B. S., G. W. Imbens, and G. Ridder (2014): Complementarity and aggregate implications of assortative matching: A nonparametric analysis, Quantitative Economics, 5, 29--66

  15. [23]

    --- -.1pt --- -.1pt --- (2020): Identification and efficiency bounds for the average match function under conditionally exogenous matching, Journal of Business & Economic Statistics, 38, 303--316

  16. [24]

    Graham, B. S., G. Ridder, P. Thiemann, and G. Zamarro (2023): Teacher-to-classroom assignment and student achievement, Journal of Business & Economic Statistics, 41, 1328--1340

  17. [25]

    (2023): Distributional synthetic controls, Econometrica, 91, 1105--1117

    Gunsilius, F. (2023): Distributional synthetic controls, Econometrica, 91, 1105--1117

  18. [26]

    Highbee, S., O. A. Katta, and G. Pouliot (2025): Distributionally robust optimal transport for program evaluation, Work in progress

  19. [27]

    Hirano, K. and J. R. Porter (2009): Asymptotics for statistical treatment rules, Econometrica, 77, 1683--1701

  20. [28]

    Mordant, C

    Hundrieser, S., G. Mordant, C. A. Weitkamp, and A. Munk (2024): Empirical optimal transport under estimated costs: Distributional limits and statistical applications, Stochastic Processes and their Applications, 178, 104462

  21. [29]

    Ishihara, T. and T. Kitagawa (2021): Evidence aggregation for treatment choice, arXiv preprint arXiv:2108.06473

  22. [30]

    Lei, and A

    Ji, W., L. Lei, and A. Spector (2023): Model-agnostic covariate-assisted inference on partially identified causal effects, arXiv preprint arXiv:2310.08115

  23. [31]

    Kallus, N. and A. Zhou (2018): Policy Evaluation and Optimization with Continuous Treatments, Proceedings of the 21st International Conference on Artificial Intelligence and Statistics (AISTATS), 84

  24. [32]

    (2022): Distributionally robust policy learning with wasserstein distance, arXiv preprint arXiv:2205.04637

    Kido, D. (2022): Distributionally robust policy learning with wasserstein distance, arXiv preprint arXiv:2205.04637

  25. [33]

    Lee, and C

    Kitagawa, T., S. Lee, and C. Qiu (2022): Treatment choice with nonlinear regret, arXiv preprint arXiv:2205.08586

  26. [34]

    --- -.1pt --- -.1pt --- (2025): Leave no one undermined: Policy learning with regret aversion, arXiv preprint arXiv:2506.16430

  27. [35]

    Sakaguchi, and A

    Kitagawa, T., S. Sakaguchi, and A. Tetenov (2021): Constrained classification and policy learning, arXiv preprint arXiv:2106.12886

  28. [36]

    Kitagawa, T. and A. Tetenov (2018): Who should be treated? empirical welfare maximization methods for treatment choice, Econometrica, 86, 591--616

  29. [37]

    --- -.1pt --- -.1pt --- (2021): Equality-Minded Treatment Choice, Journal of Business Economics and Statistics, 39, 561--574

  30. [38]

    Manski, C. F. (2004): Statistical treatment rules for heterogeneous populations, Econometrica, 72, 1221--1246

  31. [39]

    Masten, M. A. (2023): Minimax-regret treatment rules with many treatments, The Japanese Economic Review, 74, 501--537

  32. [40]

    Mbakop, E. and M. Tabord-Meehan (2021): Model selection for treatment choice: Penalized welfare maximization, Econometrica, 89, 825--848

  33. [41]

    Montiel Olea, J. L., C. Qiu, and J. Stoye (2023): Decision Theory for Treatment Choice Problems with Partial Identification, arXiv preprint arXiv:2312.17623

  34. [42]

    (2025): Should Top Surgeons Practice at Top Hospitals? Sorting and Complementarities in Healthcare, Working Paper

    Mourot, P. (2025): Should Top Surgeons Practice at Top Hospitals? Sorting and Complementarities in Healthcare, Working Paper

  35. [43]

    Peyre, G. and M. Cuturi (2019): Computational Optimal Transport, Foundations and Trends in Machine Learning, 11, 355--607

  36. [44]

    Rigollet, P. and A. J. Stromme (2025): On the sample complexity of entropic optimal transport, The Annals of Statistics, 53, 61--90

  37. [45]

    (2021): Estimation of Optimal Dynamic Treatment Assignment Rules under Policy Constraints, arXiv preprint arXiv:2106.05031

    Sakaguchi, S. (2021): Estimation of Optimal Dynamic Treatment Assignment Rules under Policy Constraints, arXiv preprint arXiv:2106.05031

  38. [46]

    Schennach, S. and V. Starck (2022): Optimally-transported Generalized Method of Moments, cemmap working paper, CWP17/22

  39. [47]

    Spurr, S. J. (1987): How the market solves an assignment problem: The matching of lawyers with legal claims, Journal of Labor Economics, 5, 502--532

  40. [48]

    (2009): Minimax regret treatment choice with finite samples, Journal of Econometrics, 151, 70--81

    Stoye, J. (2009): Minimax regret treatment choice with finite samples, Journal of Econometrics, 151, 70--81

  41. [49]

    --- -.1pt --- -.1pt --- (2012): Minimax regret treatment choice with covariates or with limited validity of experiments, Journal of Econometrics, 166, 138--156

  42. [50]

    (2021): Empirical welfare maximization with constraints, arXiv preprint arXiv:2103.15298

    Sun, L. (2021): Empirical welfare maximization with constraints, arXiv preprint arXiv:2103.15298

  43. [51]

    Sunada, K. and K. Izumi (2025): Optimal treatment assignment rules under capacity constraints, arXiv preprint arXiv:2506.12225

  44. [52]

    (2012): Statistical treatment choice based on asymmetric minimax regret criteria, Journal of Econometrics, 166, 157--165

    Tetenov, A. (2012): Statistical treatment choice based on asymmetric minimax regret criteria, Journal of Econometrics, 166, 157--165

  45. [53]

    (2025): Policy targeting under network interference, Review of Economic Studies, 92, 1257--1292

    Viviano, D. (2025): Policy targeting under network interference, Review of Economic Studies, 92, 1257--1292

  46. [54]

    Bessone, A

    Weigel, J., P. Bessone, A. Bergeron, G. Tourek, and J. Kabeya Kabeya (2025): Supermodular Bureaucrats: Evidence from Randomly Assigned Tax Collectors in the RDC, American Economic Review (Forthcoming)

  47. [55]

    (2021): Optimal decision rules under partial identification, arXiv preprint arXiv:2111.04926

    Yata, K. (2021): Optimal decision rules under partial identification, arXiv preprint arXiv:2111.04926

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.