Demographic Parity-aware Individualized Treatment Rules

(2) Department of Biostatistics; (3) Department of Mathematics; (4) Department of Biostatistics; City University of Hong Kong; Donglin Zeng (4); Statistics; The Hong Kong Polytechnic University; University of Michigan); Wenhai Cui (1); Wen Su (2)

REVIEW 3 major objections 6 minor 28 references

Demographic Parity-aware Individualized Treatment Rules

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper demonstrates that a demographic-parity-aware individualized treatment rule can be estimated as a convex quadratic program using covariance-based fairness proxies, with consistency and risk-bound guarantees, and that it…

desk verdict The convex QP machinery is real, but the proxy constraints do not imply the paper's own demographic parity definition, so the main fairness claim is unsupported. read the letter →

arxiv 2504.19914 v1 pith:HT3P6PCD submitted 2025-04-28 stat.ME

Wenhai Cui (1) , Wen Su (2) , Xiaodong Yan (3) , Donglin Zeng (4) , Xingqiu Zhao (1) ((1) Department of Applied Mathematics , The Hong Kong Polytechnic University , (2) Department of Biostatistics , City University of Hong Kong

show 5 more authors

(3) Department of Mathematics Statistics Xi'an Jiaotong University (4) Department of Biostatistics University of Michigan)

This is my paper · ORCID

classification stat.ME

keywords individualizedtreatmentrulesdemographicparityfairnessproxyconvexquadraticprogrammingoutcomeweightedlearningnonlinearcovarianceriskboundssensitiveattributes

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 an optimal individualized treatment rule can be made fair without solving a hard non-convex problem: replace the exact demographic-parity condition by a computable proxy, the covariance between the decision score and sensitive attributes, or a nonlinear indicator-weighted covariance $\omega(f)$, and the rule becomes the solution of a convex quadratic program. It proves that the resulting estimator is consistent relative to the best rule in the proxy-constrained class and derives a risk bound under a geometric-noise condition. In simulations and in the Next 36 entrepreneurship program data, the method sharply reduces the gap in selection rates across sensitive groups while giving up only part of the policy value of unconstrained outcome-weighted learning. If correct, this gives policymakers a tunable fairness budget $c$ and a practical way to inspect the trade-off between parity and policy value.

What carries the argument

The load-bearing object is the fairness proxy coupled with a convex surrogate loss. The linear proxy is the sample covariance between the sensitive-attribute vector and the decision function; the nonlinear proxy replaces $S$ by the vector of indicators $I(S<s)$ and averages the resulting covariance over $s$, which captures dependence that a plain covariance misses. The estimation machinery first substitutes these proxies for the exact parity equations, then replaces the 0-1 misclassification loss with the outcome-weighted hinge $(1-Af(X,S))_+$, adds a ridge penalty, and passes to the Lagrangian dual. This produces a positive-semidefinite quadratic objective in the dual variables $(\alpha,\gamma,\eta)$, which is exactly what makes the problem convex and solvable, and the same dual structure yields the kernel extension that handles nonlinear decision boundaries.

What would settle it

Build a joint distribution of $(S,f)$ in which $\mathrm{Cov}(S,f)=0$ and $\omega(f)$ is below any given $c$, but the conditional treatment probabilities $P(f>0\mid S=1)$ and $P(f>0\mid S=0)$ differ by a large margin; run the DPA-ITR on data from this distribution at a small $c$ and measure the actual selection-rate gap. If the gap remains large, the proxy does not control demographic parity.

Watch

Extended reading notes

Core claim

The central result is a reformulation: the constrained rule-search problem, maximize $V(D(f))$ subject to equal treatment probabilities across sensitive groups, is converted into two convex quadratic programs, equations (13) and (14), whose dual variables $\gamma,\eta$ enforce the fairness bound and whose objective is the outcome-weighted hinge risk. The fairness constraints are the linear proxy $\mathrm{Cov}(S,f(X,S))$ and the nonlinear proxy $\omega(f)=E\{\Omega(S,f)\}$ built from the indicator functions $I(S<s)$. The paper proves the associated matrices are positive semidefinite (Theorem 4.1), so off-the-shelf quadratic programming applies, and then shows the empirical minimizer attains the $\phi$-risk of the optimal proxy-constrained rule in the limit (Theorem 4.2) and satisfies a finite-sample risk bound under geometric noise (Theorem 4.3). Empirically, the estimated proxy tracks the chosen budget $c$, and the unfairness measure drops from roughly $0.178$ for unconstrained outcome-weighted learning to below $0.02$ in several simulation settings, with the value loss becoming flat as $c$ grows.

Load-bearing premise

The method assumes that having small covariance, or small nonlinear proxy $\omega(f)$, between the decision score and sensitive attributes is an acceptable stand-in for the requirement that every sensitive group receives treatment at the same rate; the paper does not bound how far a proxy-satisfying rule can be from exact demographic parity.

Editorial extensions

If this is right

Tuning the fairness budget $c$ gives a concrete cost-benefit curve: a small reduction in expected reward buys a large reduction in selection-rate disparity across sensitive groups.
Multiple sensitive attributes are handled at once through vector-valued constraints, and nonlinear decision boundaries are handled through kernels, so the method is not limited to binary or linear settings.
Because the dual problems are convex quadratic programs, the estimated rule can be computed by standard solvers, and the estimated proxy stays close to the budget $c$ on the training data.
Under the geometric-noise condition, the excess risk of the estimated rule converges to the proxy-constrained optimum at a rate that approaches $n^{-1/2}$ when the data are well separated.
In the Next 36 application, the nonlinear-proxy policy attains the smallest gender selection gap among methods compared while retaining a substantial part of the value of unconstrained policies.

Reading between the lines

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

Editorial inference: because consistency is stated relative to the proxy-constrained class $\mathcal{F}_c$ rather than the exact parity class, a policymaker who needs true equal treatment rates should run a direct selection-rate audit on the rule before deployment.
Editorial inference: the nonlinear proxy uses only threshold indicators $I(S<s)$, so for continuous sensitive attributes with dependence that changes sign or is non-monotone, a richer function class for the proxy could strengthen the fairness guarantee without changing the quadratic-program structure.
Editorial inference: the paper's empirical rule for the 'most cost-effective' $c$ compares derivatives of fitted value and unfairness curves; this could be formalized as a statistical selection procedure with confidence bands.
Editorial inference: when the data are observational, the value function still relies on unconfoundedness and a correctly specified propensity model; the fairness constraint corrects selection disparity, not confounding.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Fairness claim doesn't survive contact with the proxy constraints. The QP derivations and consistency theorems are competently done, but they concern a proxy-constrained oracle that can be grossly unfair by the paper's own Definition 2.1. That gap is the whole ballgame for a paper titled "Demographic Parity-aware."

What's actually new: casting fair ITR as constrained OWL with covariance or indicator-based mean-covariance constraints, deriving convex duals (13)-(14), proving PSD of the QP, and giving a consistency/rate result relative to the constrained class. The dual derivations look right, and the tuning parameter c gives a useful fairness-value tradeoff. Simulations show the anticipated pattern, the Next36 analysis is a reasonable illustration, and the citation pattern is normal for an OWL extension.

The soft spot is central. The linear proxy Cov(S,f)=0 and the nonlinear proxy omega(f)=0 do not imply P[f>0|S=s] equal across s. The paper acknowledges covariance is not independence, which is why it adds the nonlinear proxy, but omega(f) = E_S Omega(S,f) is an average over thresholds; it too is a mean-independence condition on f, not a sign-independence condition. The stress-test example is valid: take S,X iid Uniform(0,1), q(s)=0.1+0.3s, f=1_{X1-q(S)}. Then E[f|S=s]=0 for every s, so Cov(S,f)=0 and omega(f)=0, yet P[f>0|S=s]=q(s), a 0.3 spread in treatment rates. So even at c=0 the estimated rule can violate Definition 2.1 substantially. Theorems 4.2-4.3 are honest---they target inf over F_c---but they never bound the gap between F_c and the exact parity class. The abstract's "achieve demographic parity" is unsupported.

Other issues are minor by comparison: no code or data, no error bars, no comparison to fair-ITR baselines like Fang et al. or Viviano and Bradic, and the observational application uses estimated propensity scores without formal integration. These are fixable.

Verdict: this is a solid algorithmic contribution in search of an honest fairness guarantee. I would send it to a serious referee, but the revision needs either a proxy-to-DP gap bound under reasonable conditions, a recalibration step that enforces exact parity on a validation set, or a rewritten title and claims positioning this as parity-regularized rather than parity-enforcing. As is, a policymaker who trusts the title could deploy a rule that is still discriminatory.

Referee Report

3 major / 6 minor

Summary. The paper proposes Demographic Parity-aware Individualized Treatment Rules (DPA-ITR). The exact demographic parity constraint P[f(X,S)>0|S=s]=P[f(X,S)>0|S=s'] is replaced by either a linear proxy |Cov(S,f)|≤c or a nonlinear proxy |ω(f)|≤c. The resulting constrained outcome-weighted learning problems are converted into convex quadratic programs via Lagrangian duality, with linear and kernel decision functions. The authors prove convexity of the dual targets, establish consistency and a risk bound for the proxy-constrained estimator, and evaluate the method in four simulation experiments and the Next 36 entrepreneurship data, reporting lower unfairness measures than OWL.

Significance. The computational contribution is solid: the dual derivations are standard, the convexity argument for the quadratic programs is reasonable, and the simulations demonstrate that the method can reduce the estimated proxy and the empirical unfairness measure relative to OWL. The theoretical results follow the existing OWL and kernel learning framework and are clearly stated as relative to the proxy-constrained class. If the proxy-to-exact-DP gap can be closed or the claims reframed, the convex QP formulation is a useful tool for policy learning with covariance or mean-independence constraints. As written, the exact demographic parity claim is not established.

major comments (3)

[Definition 2.1; Sections 3.1-3.2; Eq. (6)] The linear proxy Cov(S,f) and the nonlinear proxy ω(f) are zero-covariance and mean-independence conditions, and neither implies demographic parity as defined in Definition 2.1. For S,X iid Uniform(0,1), q(s)=0.1+0.3s, and f(X,S)=1{X<q(S)}−1{X>1−q(S)}, we have E[f|S=s]=0, hence Cov(S,f)=0 and ω(f)=0, but P[f>0|S=s]=q(s), so the UFM equals 0.3. Thus the constraint set F_c with c=0 contains policies that violate Definition 2.1 by a large margin. No theorem in Section 4 bounds the distance between F_c (or W_c) and the exact demographic-parity class in (5); consequently the phrase 'demographic parity-aware' and contribution (ii) are not supported. The consistency results remain valid only for the proxy-constrained oracle.
[Theorems 4.2 and 4.3; Section 4] The consistency and risk-bound statements target inf_{f∈F_c} R_φ(f) and inf_{f∈F_c}[R_φ(f)−R*_φ], respectively, not the optimal rule under the exact demographic parity constraint (5). The approximation error between F_c and the exact DP class is never bounded, so Theorems 4.2 and 4.3 do not establish that the estimated rule approaches the best exactly fair rule even when c=0. This gap is not a technicality: it is the proxy mismatch described in the first major comment, and it means the theoretical results cannot be read as supporting the paper's fairness claim.
[Equations (11)-(14), Sections 3.4-3.8] The dual problems as stated omit the stationarity condition with respect to the intercept β0. For the linear proxy, differentiating the Lagrangian gives ∑_i α_i A_i = 0; for the nonlinear proxy, the corresponding condition has the form ∑_i α_i A_i + ∑_k (γ_k−η_k) D_k = 0, where D_k = n^{-2}∑_i∑_l [I(S_ik<S_lk)−I(S̄_k<S_lk)] is generally nonzero. Neither condition appears in the constraints listed after (11), (12), (13), or (14). Without these equality constraints, the stated quadratic programs are not equivalent to the primal hinge-loss problems, and the intercept recovered by the KKT margin-point procedure may not correspond to a dual optimum.

minor comments (6)

[Section 5.2] The text says 'Figures 3-6 demonstrates' but the surrounding text refers to Figures 2-4; also the statement that Table 1 'provides validation for the convex property shown in Theorem 1' conflates constraint satisfaction with convexity.
[Section 4] Theorem 4.1 is referred to as 'Theorem 1' in Section 5.2; please unify the numbering.
[Section 6] The application uses a binary reward R∈{-1,1}, while the setup assumes R is nonnegative; the authors should state a rescaling or adjust the assumption.
[Section 5.3] The 'most cost-effective point' rule based on U'(c)−V'(c)>0 is presented without formal justification; it should be labeled a heuristic and accompanied by uncertainty quantification.
[Section 5.1] The tuning procedure describes cross-validation for (λ,σ), but the parameter κ in (10) also needs a selection rule; please clarify.
[Section 3.7] The reference to 'Section 3.2.1' should be to Section 3.4, the linear-proxy linear decision function case.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the consistency and risk bounds target the same proxy-constrained classes used in estimation, and the supporting citations are external published results.

full rationale

I find no circularity in the paper's derivation chain. The dual quadratic programs (10)-(14) are derived algebraically from the hinge surrogate and the proxy constraints, and the consistency result (Theorem 4.2) and risk bound (Theorem 4.3) are stated relative to the same proxy-constrained classes F_c and W_c that the optimization enforces; thus no fitted constant is renamed as a prediction. Tuning parameters and the fairness constraint c are selected by cross-validation or by an explicit empirical selection rule in Section 5.3, not fitted to the theoretical claims. The citations to Zhao et al. (2012), Zhu et al. (2011), and Steinwart and Scovel (2007) are external published results that do not presuppose the paper's conclusion and are not used to force a unique choice of the method. The acknowledged gap between the proxies and exact demographic parity (Definition 2.1) is a validity limitation rather than a circular reduction: the paper proves consistency only for the proxy-constrained oracle, and the reviewer's constructed example shows the proxy can be satisfied while the unfairness measure is positive. That concern belongs to correctness risk, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The main free parameters are tuning knobs (c, λ, σ). The only ad hoc input is the faithfulness of the proxies to exact demographic parity, which is an unquantified assumption rather than an invented entity. No new physical or statistical entities are postulated.

free parameters (3)

Fairness constraint c = grid in simulations (0.02 to 0.16); c = 0.00001 in application
Chosen by the user; paper proposes empirical selection via the most cost-effective point or the four-fifths rule. It controls the fairness-value tradeoff and directly shapes the solution.
Regularization parameter κ (or λ) = selected via two-fold cross-validation over a finite set (values not reported)
Controls the ridge penalty in the surrogate objective; a standard tuning parameter.
Gaussian kernel bandwidth σ = selected via two-fold cross-validation over a finite set (values not reported)
Used only for nonlinear decision functions; standard kernel bandwidth tuning.

assumptions (4)

domain assumption Treatment assignment is randomized with known propensity π = P(A=1), and rewards R are bounded (Section 2.1).
Underlies the inverse-probability-weighted value function in Equation (1). The application extends to observational data but the method is not formally re-derived with estimated propensity scores.
standard math The decision function f belongs to an RKHS with a positive semidefinite kernel; Moore-Aronszajn theorem (Section 3.5).
Required for the representer theorem style dual derivation and for the kernel form of the decision function.
domain assumption For the risk bound, the data distribution satisfies the geometric noise condition with noise exponent q > 0 (Steinwart and Scovel, 2007; invoked in Theorem 4.3).
Standard smoothing assumption used to obtain fast convergence rates for kernel methods; not verified in the simulations.
ad hoc to paper The fairness proxies adequately capture the demographic parity requirement (Sections 3.1 and 3.2).
The central approximation: zero covariance or zero ω(f) is sufficient for independence only in restricted cases; the paper does not quantify the approximation error to exact demographic parity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Demographic Parity-aware Individualized Treatment Rules." pith.science (2026). https://pith.science/paper/HT3P6PCD

@misc{pith2026250419914,
  author       = {Pith},
  title        = {Pith review of: Demographic Parity-aware Individualized Treatment Rules},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HT3P6PCD}},
  note         = {Machine review of arXiv:2504.19914}
}

read the original abstract

There has been growing interest in developing optimal individualized treatment rules (ITRs) in various fields, such as precision medicine, business decision-making, and social welfare distribution. The application of ITRs within a societal context raises substantial concerns regarding potential discrimination over sensitive attributes such as age, gender, or race. To address this concern directly, we introduce the concept of demographic parity in ITRs. However, estimating an optimal ITR that satisfies the demographic parity requires solving a non-convex constrained optimization problem. To overcome these computational challenges, we employ tailored fairness proxies inspired by demographic parity and transform it into a convex quadratic programming problem. Additionally, we establish the consistency of the proposed estimator and the risk bound. The performance of the proposed method is demonstrated through extensive simulation studies and real data analysis.

Figures

Figures reproduced from arXiv: 2504.19914 by the authors.

**Figure 2.** The decision plot in Experiment 1 with p = 3 n = 500; the naive method means estimating the ITR without considering the sensitive attribute S via OWL; The UFM in Figure 2a, 2c,2b are 0.178, 0.109, 0.011, respectively. The estimated ωb(fb) in [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

**Figure 3.** The UFM and the estimated value plot for training sample size [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: The UFM and the estimated value plot for training sample size [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

**Figure 5.** Figure 5: The UFM and the estimated value plot for training sample size [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 6.** Figure 6: The UFM and the estimated value plot for training sample size [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

**Figure 7.** Figure 7: The empirical UFM and value plot in Experiment 1 with [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 24 canonical work pages

[1]

Aronszajn, N. (1950). Theory of reproducing kernels. Transactions of the American mathematical society\/ 68 , 337--404

work page 1950
[2]

Biddle, D. (2017). Adverse impact and test validation: A practitioner's guide to valid and defensible employment testing . Routledge

2017
[3]

Murphy, and V

Chakraborty, B., S. Murphy, and V. Strecher (2010). Inference for non-regular parameters in optimal dynamic treatment regimes. Statistical methods in medical research\/ 19 , 317--343

work page 2010
[4]

Dayan, P. and C. Watkins (1992). Q-learning. Machine learning\/ 8\/ (3), 279--292

work page 1992
[5]

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

work page 2023
[6]

Fowler, R. A., N. Sabur, P. Li, D. N. Juurlink, R. Pinto, M. A. Hladunewich, N. K. Adhikari, W. J. Sibbald, and C. M. Martin (2007). Sex-and age-based differences in the delivery and outcomes of critical care. Cmaj\/ 177\/ (12), 1513--1519

work page 2007
[7]

Ferrari, K

Heidari, H., C. Ferrari, K. Gummadi, and A. Krause (2018). Fairness behind a veil of ignorance: A welfare analysis for automated decision making. 31

work page 2018
[8]

Kim, K. and J. R. Zubizarreta (2023). Fair and robust estimation of heterogeneous treatment effects for policy learning. In International Conference on Machine Learning , pp.\ 16997--17014. PMLR

2023

Show all 28 references

[9]

Lambrecht, A. and C. Tucker (2019). Algorithmic bias? an empirical study of apparent gender-based discrimination in the display of stem career ads. Management Science\/ 65 , 2966--2981

2019
[10]

Liu, Y., Y. Wang, M. R. Kosorok, Y. Zhao, and D. Zeng (2018). Augmented outcome-weighted learning for estimating optimal dynamic treatment regimens. Statistics in medicine\/ 37 , 3776--3788

2018
[11]

Lyons, E. and L. Zhang (2017). The impact of entrepreneurship programs on minorities. American Economic Review\/ 107 , 303--307

2017
[12]

Murphy, S. A. (2003). Optimal dynamic treatment regimes. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 65\/ (2), 331--355

2003
[13]

Penner, L. A., J. F. Dovidio, T. V. West, S. L. Gaertner, T. L. Albrecht, R. K. Dailey, and T. Markova (2010). Aversive racism and medical interactions with black patients: A field study. Journal of Experimental Social Psychology\/ 46 , 436--440

2010
[14]

Qian, M. and S. A. Murphy (2011). Performance guarantees for individualized treatment rules. Annals of statistics\/ 39 , 1180

2011
[15]

Robins, J. M. (2004). Optimal structural nested models for optimal sequential decisions. Proceedings of the Second Seattle Symposium in Biostatistics: analysis of correlated data\/ , 189--326

2004
[16]

Robins, J. M., M. Ángel Hernán, and B. Brumback (2000). Marginal structural models and causal inference in epidemiology. Epidemiology\/ 11 , 550--560

2000
[17]

Rosholm, M. and R. Vejlin (2010). Reducing income transfers to refugee immigrants: Does start-help help you start? Labour Economics\/ 17 , 258--275

2010
[18]

Shi, C., A. Fan, R. Song, and W. Lu (2018). High-dimensional a-learning for optimal dynamic treatment regimes. Annals of statistics\/ 46 , 925

2018
[19]

Shi, C., S. Luo, Y. Le, H. Zhu, and R. Song (2022). Statistically efficient advantage learning for offline reinforcement learning in infinite horizons. Journal of the American Statistical Association\/ , 1--14

2022
[20]

Song, and W

Shi, C., R. Song, and W. Lu (2016). Robust learning for optimal treatment decision with np-dimensionality. Electronic journal of statistics\/ 10 , 2894

2016
[21]

Shi, C., J. Zhu, S. Ye, S. Luo, H. Zhu, and R. Song (2022). Off-policy confidence interval estimation with confounded markov decision process. Journal of the American Statistical Association\/ , 1--12

2022
[22]

Song, R., W. Wang, D. Zeng, and M. R. Kosorok (2015). Penalized q-learning for dynamic treatment regimens. Statistica Sinica\/ 25 , 901

2015
[23]

Steinwart, I. and C. Scovel (2007). Fast rates for support vector machines using Gaussian kernels . The Annals of Statistics\/ 35 , 575 -- 607

2007
[24]

Viviano, D. and J. Bradic (2023). Fair policy targeting. Journal of the American Statistical Association\/ , 1--14

2023
[25]

Fu, and D

Wang, Y., H. Fu, and D. Zeng (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

2018
[26]

Williams, D. R. and R. Wyatt (2015). Racial Bias in Health Care and Health: Challenges and Opportunities . JAMA\/ 314 , 555--556

2015
[27]

Zhao, Y., D. Zeng, A. J. Rush, and M. R. Kosorok (2012). Estimating individualized treatment rules using outcome weighted learning. Journal of the American Statistical Association\/ 107 , 1106--1118

2012
[28]

Zhu, L., L. Li, R. Li, and L. Zhu (2011). Model-free feature screening for ultrahigh-dimensional data. Journal of the American Statistical Association\/ 106 , 1464--1475. PMID: 22754050

2011

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Aronszajn, N. (1950). Theory of reproducing kernels. Transactions of the American mathematical society\/ 68 , 337--404

work page 1950

[2] [2]

Biddle, D. (2017). Adverse impact and test validation: A practitioner's guide to valid and defensible employment testing . Routledge

2017

[3] [3]

Murphy, and V

Chakraborty, B., S. Murphy, and V. Strecher (2010). Inference for non-regular parameters in optimal dynamic treatment regimes. Statistical methods in medical research\/ 19 , 317--343

work page 2010

[4] [4]

Dayan, P. and C. Watkins (1992). Q-learning. Machine learning\/ 8\/ (3), 279--292

work page 1992

[5] [5]

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

work page 2023

[6] [6]

Fowler, R. A., N. Sabur, P. Li, D. N. Juurlink, R. Pinto, M. A. Hladunewich, N. K. Adhikari, W. J. Sibbald, and C. M. Martin (2007). Sex-and age-based differences in the delivery and outcomes of critical care. Cmaj\/ 177\/ (12), 1513--1519

work page 2007

[7] [7]

Ferrari, K

Heidari, H., C. Ferrari, K. Gummadi, and A. Krause (2018). Fairness behind a veil of ignorance: A welfare analysis for automated decision making. 31

work page 2018

[8] [8]

Kim, K. and J. R. Zubizarreta (2023). Fair and robust estimation of heterogeneous treatment effects for policy learning. In International Conference on Machine Learning , pp.\ 16997--17014. PMLR

2023