Causal Fairness for Outcome Control

As society transitions towards an AI-based decision-making infrastructure, an ever-increasing number of decisions once under control of humans are now delegated to automated systems. Even though such developments make various parts of society more efficient, a large body of evidence suggests that a great deal of care needs to be taken to make such automated decision-making systems fair and equitable, namely, taking into account sensitive attributes such as gender, race, and religion. In this paper, we study a specific decision-making task called outcome control in which an automated system aims to optimize an outcome variable $Y$ while being fair and equitable. The interest in such a setting ranges from interventions related to criminal justice and welfare, all the way to clinical decision-making and public health. In this paper, we first analyze through causal lenses the notion of benefit, which captures how much a specific individual would benefit from a positive decision, counterfactually speaking, when contrasted with an alternative, negative one. We introduce the notion of benefit fairness, which can be seen as the minimal fairness requirement in decision-making, and develop an algorithm for satisfying it. We then note that the benefit itself may be influenced by the protected attribute, and propose causal tools which can be used to analyze this. Finally, if some of the variations of the protected attribute in the benefit are considered as discriminatory, the notion of benefit fairness may need to be strengthened, which leads us to articulating a notion of causal benefit fairness. Using this notion, we develop a new optimization procedure capable of maximizing $Y$ while ascertaining causal fairness in the decision process.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer BFhw7/10 · confidence 4/52023-07-03

Summary

The authors analyze a new dimension of decision making: outcome control and benefit from decisions. Specifically, this work uses causal analysis to introduce new definitions of fairness on benefits and outcome control. The authors also propose new algorithms to study fairness and design fair algorithms for outcome fair decisions.

Strengths

* The paper proposes new fairness notions for outcome control, that is an important aspect of many decision making settings * The authors use a simple example to walk through all the different aspects and notions, that greatly helps the readability of the paper. * Through the synthetic use case, the authors show when and how benefit fairness is important, and how causal path analysis could also become critical for decision making while accounting for fairer outcomes.

Weaknesses

* While the appendix provides additional results on a real-world COVID dataset, the main paper has limited results on various different settings. * While the authors assume causal knowledge, the paper does not discuss limitations explicitly for when computing counterfactuals and especially path-specific counterfactuals might become challenging in practice. * Section 3 should have better explained which counterfactuals we compute to look at the path-specific factors and how the 3 different components are found out by decomposition.

Questions

* Two works [1, 2] seem somewhat related to the authors’ notions and framework. How does the framework of decision outcome control relate to the thinking of fairness of “treatments” in [1]? Similarly, how does the notion of “benefit” the authors introduce relate to the notion of “harm” in [2]?
 [1] Madras, David, et al. "Fairness through causal awareness: Learning causal latent-variable models for biased data." Proceedings of the conference on fairness, accountability, and transparency. 2019.
[2] Richens, Jonathan, Rory Beard, and Daniel H. Thompson. "Counterfactual harm." Advances in Neural Information Processing Systems 35 (2022): 36350-36365.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

* While the paper’s primary focus is on introducing a new theoretical framework of fairness in decision making related to outcome control and benefit, more analysis and results on different experimental settings might have been an added benefit.

Reviewer H5h57/10 · confidence 4/52023-07-05

Summary

The paper proposes a new fairness notion called benefit fairness that considers fairness in the outcomes of the decisions. In order to avoid the unidentifiability issue of the principle fairness, this paper proposes to condition on the conditional average treatment effect (CATE). The paper provides an algorithm that is proven to be optimal and satisfy benefit fairness. The paper in addition provides an algorithm to satisfy both benefit fairness and counterfactual benefit fairness.

Strengths

The proposed new fairness notion is quite interesting. It is different from traditional fairness notions, yet it is well-motivated and -explained using a real-world example. The proposed algorithms come with theoretical analysis that shows their optimality and fairness satisfaction. A budget is involved in the problem formulation that makes the algorithms more generally applicable.

Weaknesses

The connection between benefit fairness and traditional fairness notions other than principle fairness, such as demographic parity, equal opportunity, etc., is not clear. Are they consistent or do they conflict to some extent? A rigorous study may be out of the scope of this paper but some discussions will be helpful. The measure of benefit fairness depends on the observation of the outcome Y. Can we generally assume that Y is available when we construct the decision policy? In the running example, Y is observed after a 2-year period. How can we estimate Y when we construct the decision policy before the 2-year period? Some statements and notations are not very clear. For example, in Definition 3, $y_C$ is not defined and the definition of the pathway is not clear. My understanding is that the paper attempts to formulate path-specific counterfactual fairness, but the notations are quite confusing. In addition, $\Delta_{C}$ is not defined either. The two terms CBF and BFC may easily confuse the readers.

Questions

Although I didn’t check the correctness of the proof, Theorem 2 seems quite strong. Are there any assumptions about the policy or the causal model in Theorem 2? My understanding of benefit fairness is that, it considers the causal effect of the link X->Y. If we remove this link, then $\Delta(x_0,z,w)$ is always equal to $\Delta(x_1,z,w)$, and benefit fairness degrades to the conditional demographic parity. Thus, the meaning of benefit fairness is to make the causal effect of the link X->D match the causal effect of the link X->Y? Is this understanding correct? If my above understanding is correct, does that mean that CBF cannot be satisfied if the link X->Y exists, because CBF generally means that the decision should not depend on X?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

None.

Reviewer pGG26/10 · confidence 4/52023-07-05

Summary

This paper studies the problem where a decision maker must allocate a treatment $D$ to optimize an outcome variable $Y$ while ensuring that the decision is fair, formulating fairness as the protected attribute $X$ not having an influence on $D$. It uses a clinical decision-making process (a running example on Cancer Surgery, Fig. 2(b)) to showcase its proposed method. The paper proposes the notion of benefit based on the potential outcomes (PO) framework. $Y_{d_0}$ denotes the outcome of a patient that didn’t undergo surgery while $Y_{d_1}$ denotes the outcome of a patient that did undergo surgery. Under PO, only one of the two potential outcomes are observed per patient. Benefit is defined as $Y_{d_1} - Y_{d_0}$. Hence, a decision maker will derive a $D^*$ that maximizes overall benefit: i.e., those patients for which $Y_{d_0}=0$ (died without surgery) and $Y_{d_1}$ (survived with surgery). Through the running example, the paper shows how a decision maker (under imperfect information) can be unfair via $D$ despite optimizing for the benefit. It then introduces a set of algorithms to reach fair benefit.

Strengths

[S1] The topic is relevant and novel within (causal) fairness. Under a heterogenous population of individuals, it is reasonable to expect that not all individuals will benefit from a treatment (captured by $D$ and its causal effect on $Y$ in the paper). The notion of benefit becomes relevant as (i) it formalizes which individuals need $D$ and which do not; (ii) it raises the potential issue of what happens when those that benefit under $D$ don’t embody societal representational fairness goals. [S2] The use of the running example was very useful for understanding the paper’s goal. [S3] Fig. 3 illustrates well the problem of allocating a treatment without knowing the type of patient w.r.t. to $D$. [S4] Def. 2 (Benefit Fairness), putting aside the estimation issues, is a nice extension to other group-level fairness definitions (like Equal Opportunity). It states that individuals across the protected attribute $X$ that have the same (potential) benefit $\delta$ to gain from treatment should have the same probability of getting treated.

Weaknesses

[W1] PO limitations should be discussed further: Given the use of the PO framework, there’s a lack of treatment on how the benefit ($\Delta = Y_{d_1} - Y_{d_0}$) can be estimated in a consistent and unbiased way, at least, at the individual level. For each individual, we can only observe $Y_{d_0}$ or $Y_{d_1}$: for each individual, the one observed is the factual and the other one the counterfactual. However, the paper lacks a meaningful discussion on how we can estimate the counterfactuals in the non-oracle setting, which is the one of practical interest. Even in the running example, isn’t there a risk that the clinicians are using a biased $\Delta$? If so, how would that affect the proposed algorithms? The limitations should be discussed explicitly. In particular, the derivation and justification of Eq. (16), which is central to the algorithm(s) is not clear. Is it a statement on the same tuple $(x, z, w)$: if so, how can the same individual be benefited and harmed at the same time by the treatment? Or is it a “match” between two individuals? Also, what is $z$ in the running example (this random variable is only presented and used in Fig. 1)? Counterfactual generation/estimation is not exploited in the example. For instance, if we have access to the SCM model, couldn’t we generate the counterfactual distribution via Pearl’s abduction, action, and prediction steps for $Y$. We could then identify the “doomed” and “safe”? This is important as Algorithm 1 can only operate meaningfully if it can identify the borders of the types of individuals (w.r.t. the notion of benefit). Line 1 in Algorithm 1 needs to be discussed further. [W2] Stronger Section 2: This section could be tighter. For instance, what’s the role of the budget $b$ from Def. 1? Even in the oracle setting, it is possible that we can’t treat all 100 patients if the budget is low enough, no? Conversely, if the budget is high enough and $\Delta \geq 0$, we could treat all patients, no? Further, as a follow up to W1, if there is the pair \{$Y_{d_0}$, $Y_{d_1}$\} for each individual then there is also a $\Delta$ parameter. Now, under the oracle, we can see through the future and allocate $D$ based on $\Delta$. In the non-oracle setting, there’s a brief discussion about using $W$ as a proxy that leads to unfairness: how does that translate into estimating $\Delta$? Otherwise, if we can estimate it, then why are we even considering the proxy $W$? What I’m hinting at here – do correct me if I’m wrong – is that Def. 2 needs to have some measure of uncertainty to highlight that the non-oracle allocation will have some error w.r.t. to the oracle allocation (for a fixed $b$). Otherwise, it seems like we are still in the oracle setting. [W3] Limited evaluation: The running example is essentially the only use case. Although it shows the proposed methodology, it would’ve helped to, e.g., test the algorithms under different parameters for the same synthetic data. Similarly, could the algorithms handle other variables on top $W$ or under a second protected attribute like race? The evaluation, even for the single use case, could’ve been pushed further to show the robustness of the approach. [W4] Some definitions are unclear or not fully explained. For instance, In Def. 1, the role of the budget $b$ is not explained. It would’ve also helped to formulate it under the PO framework explicitly. In Def. 2, please define $\Delta$ and $\delta$ within the definition itself. These are presented later. In Def. 3, is it at the individual level? Under the SCM presented (Fig. 2, but also Fig. 2), how are the conditionals the same under the interventions captured by the causal pathways. Don’t we need to update for downstream effects under $\mathcal{C}$?

Questions

See my questions in the Weaknesses section. I’m willing to increase my score if these questions are addressed.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Further discussion on $\Delta$ is missing. It’s a considerable limitation for the application of this definition and it’s not evaluated here theoretically.

Reviewer 9Jxk7/10 · confidence 3/52023-07-10

Summary

The paper focuses on outcome control from a causal perspective of the decision-maker. The authors introduce benefit fairness taking the perspective of the decision maker and provide theoretical guarantee that the algorithmic result is optimal and satisfies benefit fairness. To support the decision maker in indicating potential discrimination early on, algorithm 2 evaluates the difference in benefit to demographic groups. And the paper defines causal benefit fairness with an accompanying algorithm and a guarantee of optimality.

Strengths

S1 - In fairness research, the need for more causal perspectives has been a growing discussion. This research is significant in contributing to an underdeveloped area. S2 - The research problem is well developed and simple guarantees provided with proofs in the appendix.

Weaknesses

Exposition clarity needs minor improvement. For instance, what is the relationship between the first and second paragraph of the introduction? What do the authors see as the connection between outcome control and historical biases?

Questions

The authors do not actually provide a clear definition of “outcome control” as used in this paper. Based on the content, I assumed they are using the high level definition put forth by [Procedural Justice in Algorithmic Fairness, CSCW 2019], which defines outcome control as enabling the correctability and possible recourse against an individual decision. But this paper is not cited in the present paper. The authors reference [Causal Fairness Analysis, 2022], which is an ICML tutorial where section 5.3.3 contains exact wording and sentences from the present paper lines 35-38. The tutorial includes more insight clarifying that outcome control setting requires that the institution and the individual have the same utility function. It would be helpful to clarify this in the main text of the present paper. What is the definition of “outcome control” being used in this paper?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

It is not clear how generalizable this framework is given the demonstration on clinical examples.

Reviewer pGG22023-08-10

Dear authors, thank you for the very detailed answers. You really did go over all of my concerns. Under these proposed changes/clarifications, on top of the general comments (G1 and G2), I see a stronger paper. I will update my score accordingly.

Authorsrebuttal2023-08-11

We would like to thank the reviewer for the constructive review process, in which we were able to improve our paper (specifically regarding identifiability results for counterfactuals). We also thank the reviewer for acknowledging our response, and for adjusting the grade as well.

Reviewer BFhw2023-08-10

Thank you to the authors for addressing all my questions and doubts and adding 2 suggested fairness papers in their related work. I do think that the real-world empirical analysis is a strong case for your framework and can strengthen the theoretical work even more. This is especially true since it directly shows a tangible application to a causal method, which is generally hard to show! I do think the work is very interesting and adds a new view to fairness, so I maintain my acceptance suggestion.

Authorsrebuttal2023-08-11

We again thank the reviewer for the suggestions during the review process, and also for acknowledging our answers. We are also encouraged by the fact that the reviewer sees the work as adding something novel and interesting to the literature, thank you!

Reviewer H5h52023-08-10

Thank you for your responses. Your explanations about the CBF definition are quite helpful. I have a separate question. I consider the treatment benefit $\Delta$ as some sort of stratification of the population. From this perspective, how do you suggest we perform the conditioning on $\Delta=\delta$ in practice, especially when each individual has a different treatment effect? For example, do you suggest we bin the treatment effects and then condition on bins?

Authorsrebuttal2023-08-11

We thank the reviewer for acknowledging our responses. Regarding $\Delta$ in practical applications: exactly, our approach would then be to use a fixed number of bins, on which we condition. The bins can correspond to quantiles of the $\Delta$ distribution, for example. The larger the sample size, the more bins we could use in practice, of course, and in the infinite sample limit this would correspond to conditioning on fixed $\Delta = \delta$ values. Actually, the approach with bins was also used in the experiment in Appendix E, where we conditioned on the percentiles of the $\Delta$ distribution (which was possible due to quite large sample size). Please let us know if there are any further questions!

Reviewer 9Jxk2023-08-16

Thank you for the detailed response.

© 2026 NYSGPT2525 LLC