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 $\mathcal{C}$?
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.