Procedural Fairness Through Addressing Social Determinants of Opportunity
We approach procedural fairness by explicitly considering influences on individuals from social determinants of opportunity.
摘要
评审与讨论
This work explores incoporating the concept of social determinants of opportunity, variables that relate to an individual's academic success causally and potentially implicitly. The authors deviate from past work by modeling implicit relationships between these variables rather than simplified relationships and futher explore adding previously omitted variables. Then, framing academic preparedness as an optimization problem, the authors find a correlation between race and social determinants of opportunity, using GPA as an estimation of academic preparedness and analyzing data from the University of California's admissions data. As an analysis of existing data, this work proposes modeling protected characteristics and studying the influence of contexts and environments on the individual for fairness analysis.
优点
The authors seek to model fairness in college admissions by disentangling variables that implicitly model each other, which provides an interesting framework for considering the intersectionality of factors that influence an individual. When applied to college admissions and academic preparedness, the authors provide a convincing argument for abstracting out social determinants of opportunity and studying the underlying framework and its impact on the individual. Further, the authors demonstrate various applications of their framework in Section 4 in studying historical admissions systems.
缺点
While there are limitations in studying observational data, this paper could have benefitted from a further analysis of the University of California dataset; the authors do acknowledge the limitations of summary statistics but further discussion of the dataset and analysis on other datasets could have provided more support to the experimental section of this paper. The authors briefly discuss their experimental findings but the three separate graphs in Figure 3 could have benefitted from further discussion, particularly in relation to each other and how the correlation between race and social discriminators of opportunity correlates with understanding academic preparedness in a region. Potentially interleaving the methods and providing experimental results for the modeling of past admissions systems could have provided more tractable examples of how this framework compares to prior work.
问题
How does modeling University of California admissions data via the presented framework differ experimentally from past methods? What are possible extensions of this framework in developing more holistic admissions systems? A theoretical analysis of what this kind of admissions system could look like could provide further insight into the extensions of this model.
We are very grateful for your constructive and insightful comments, and for the time and effort devoted! We have provided a revised manuscript, where we use blue font to indicate added/revised material. Below please see our responses to specific points in the review comment:
C1: "While there are limitations in studying observational data, this paper could have benefitted from a further analysis of the University of California dataset"
A1: Thanks for the insightful comment. Following your suggestion, we have included an additional section Appendix C.1 to present description of the UC data and further analysis on the results, together with a side note Re: C1 by Reviewer h9Xa on page 24 to help locate the material.
C2: "Analysis on other datasets could have provided more support to the experimental section of this paper"
A2: Thanks for the constructive suggestion. We have included additional data analyses on the US Census data in Appendix C.2, along with a side note Re: C2 by Reviewer h9Xa on page 26.
C3: "Figure 3 could have benefitted from further discussion, particularly in relation to each other and how the correlation between race and social [determinants] of opportunity correlates with understanding academic preparedness in a region"
A3: Thanks for the constructive comment. We follow the suggestion and have add further discussions in Appendix C.1.3 (with a side note Re: C3 by Reviewer h9Xa on page 25) and also provided the pointer in main text (footnote 6).
C4: "Potentially interleaving the methods and providing experimental results for the modeling of past admissions systems could have provided more tractable examples of how this framework compares to prior work."
A4: Thank you for the thoughtful comment. While we present theoretical analyses followed by experimental results (instead of interchangeably), we provide accompanying illustrative figures (Figure 2) for the theoretical analysis of each kind of policies (Theorems 4.5 -- 4.7), to explicate the implications of different strategies.
Prior works in algorithmic fairness typically drop the information that is relevant to social determinants of opportunity (among other potential issues, as presented in Section 3.1). When such information is dropped, it is not immediate (if possible) for previous works to provide similar nuanced analyses.
Q5: "How does modeling University of California admissions data via the presented framework differ experimentally from past methods?"
A5: Thanks for the question. Our framework differs experimentally from previous methods in both the handling of data and the question intended to answer.
-
Since social determinants of opportunity correlate with protected features, we do not drop variables that seem to be unrelated to the decision-making, but in fact capture influence from the contextual environment to the individual (e.g., the address). We discuss this point in Section 3.1 (side note
Re: Q5.1 by Reviewer h9Xaon page 4). -
Apart from data handling, our framework also differs from previous approaches in the questions we intended to address. Specifically, we aim to characterize and address social determinants of opportunity when achieving procedural fairness, which cannot be simplified into merely considering the protected features at the individual level. In light of your question, we have included the above discussion in more detail in Appendix A.2, together with a side note
Re: Q5.2 by Reviewer h9Xaon page 19.
Q6: "What are possible extensions of this framework in developing more holistic admissions systems?"
A6: Thanks for the question about the next steps and how to go further. We use the college admission as a concrete example, but our advocacy for addressing social determinants of opportunity to achieve procedural fairness has various implications to a broader scope (which goes beyond admission decision-making).
For example, from the college perspective, outreach programs in community can be helpful especially when the educational resource is scarce in the area. From the social policy perspective, investments to improve living and environmental conditions can positively affect people's overall health and better avail them to pursue diverse life plans.
Thank you for providing additional details. The disussion about Figures 3/4 in the appendix is clarifying and I would encourage parts of the explanation to be moved to the main text.
As another reviewer mentioned, Appendix C2 seems to be a separate analysis of census data that doesn't relate to the model presented in the paper. I was more so curious how well the presented model extrapolates to other datasets, though I understand how the analysis seeks to frame using social determinants of opportunity more broadly.
Thank you for the feedback, and for the constructive suggestion, many of which help us further improve our manuscript. We will follow your suggestion and move part of the material to main text.
The Appendix C.2 is actually very related to our proposed framework. We followed your suggestion (C2 of the original comment) and considered real-world data where the region information (e.g., the PUMA in census data) is not latent and readily available. The goal of these analyses is to demonstrate on real-world data that the region information, which is often dropped by previous approaches, actually contains rich implications of various contextual influences in different regions. Therefore, Appendix C.2 provides further practical evidence that supports our advocacy for explicit considerations of social determinants of opportunity.
Please let us know if there is any remaining question or concern. Thanks again for the constructive suggestions.
This paper develops a model of the interactions between ethnicity, academic preparedness, and "social determinants of opportunity", which capture socio-geographic influences on academic preparedness. The model is used to study different college admissions policies both in theory, and applied to a dataset of UC Berkeley admissions.
优点
The paper is clearly arranged and easy to follow. This paper is also laudable for trying to raise the salience of geographic and community influences on opportunity over a reductive focus on ethnicity.
缺点
UPDATE: After the discussion period I stand by these weakness of the paper.
Ultimately, the theoretical results in this paper do not provide novel insight, and the empirical results aren't very interesting or plausible. The model is used to show that:
- Quota-based affirmative action harms disadvantaged members of majority groups.
- "Plus factors" for being from an underrepresented group benefit advantaged members of that group more than disadvantaged members.
- "Top-percent" policies that are blind to ethnicity reallocate opportunity to regions with less of it.
All three of these findings are well-known, and have been part of the debate around these policies for decades. The model doesn't provide extra insights into these policies.
When the model is deployed on real data, it also doesn't provide insights. It seems as though the admissions data from Berkeley is too censored to study the impacts of social determinants of opportunity, since it doesn't include anything about an applicant's geography beyond whether they are in-state. The regions inferred by the model don’t make a lot of sense given what we know about California’s ethnic geography (eg they don't show any signs of the racial segregation induced by California's restrictive housing policies).
问题
Are the regions in the experiment latent? Region 1 and 3 are almost identical, calling into question the identifiability of the model. Also, if regions don’t correspond to geographies or social networks then how can they capture social determinants of opportunity?
Notes:
“Specifically, by definition of causality, this edge asserts that there is a difference in the distribution of education status, when we “intervene” on individual’s race while keeping all other things unchanged”
“all other things” meaning every other variable in their causal graph, not literally every other possible thing. Since these graphs typically only use a handful of features, I don’t think this edge is an endorsement of racial essentialism - it just summarizes dozens of effects that the model is too coarse to model explicitly.
“If a certain edge or path in the causal model does not reflect an actual real-world causal process, subsequent causal fairness analyses based on causal effects may not provide informative conclusions.”
This is certainly true, but to my knowledge not a single causal graph in history has ever actually described a real-world causal process where humans were involved. It’s extremely difficult to establish single treatment effects, let alone a network of them.
Thank you for the thoughtful questions and comments, and for the time devoted! We have provided a revised manuscript, where we use blue font to indicate added/revised material. Below please see our responses to specific points in the review comments:
C1: "All three of these findings are well-known, and have been part of the debate around these policies for decades. The model doesn't provide extra insights into these policies."
A1: Thanks for carefully consider our theoretical results. We respond in threefold:
-
To begin with, to the best of our knowledge, our modeling that explicitly considers the contextual influence on individuals is novel, and there is no overlap with previous algorithmic fairness literature. Technically speaking, previous causal fairness approaches do not have the capacity to directly produce the same set of "well-known" arguments, because of the issues we discussed in Section 3.1, e.g., the region is typically not on the radar.
-
Furthermore, apart from the phenomena summarized by these findings, our causal modeling aims to characterize the reason behind. Social determinants of opportunity and their nontrivial role in the pursuit of procedural fairness, as you kindly commented, go beyond a reductive focus on ethnicity.
-
In addition, algorithmic fairness is naturally an inter-disciplinary research. We do not view "findings being well-known and part of debate over policies" as a shortcoming. The quantitative analyses facilitated by our causal modeling very much align with domain knowledge in debates from related disciplines. This actually indicates that our model is able to present important findings in a precise and transparent way.
Q2: "Are the regions in the experiment latent? Region 1 and 3 are almost identical, calling into question the identifiability of the model."
A2: Thanks for the insightful question. Yes, the regions are latent in our empirical results, and the optimization problem can indeed be under-constrained when the publicly available data contains only summary statistics (which is intentional for legal and ethical reasons).
However, this is not a technical barrier to apply our framework in practice. When the practitioner has access to the whole data set (e.g., the university's internal research or audit), the region is no longer a latent variable (e.g., directly obtained from address of home or the attended school).
In the revised manuscript, we have added extensive analyses on the US census data, where the region information is readily available. The related material can be found in Appendix C.2, along with a side note Re: Q2 by Reviewer 8F2j on page 26.
Q3: "If regions don’t correspond to geographies or social networks then how can they capture social determinants of opportunity?"
A3: By definition, geographical regions correspond to geographical locations, and the pattern of social networks (among people that live and/or operate on the location) differ across regions. Please kindly let us know if we accidentally misunderstood your question.
Meanwhile, we would like to clarify that we are using region as a surrogate for social determinants of opportunity, and we demonstrate that our model facilitates more nuanced analyses than using the protected feature to enclose all these related correlations (which is the more-or-less the default modeling choice in the algorithmic fairness literature). If there exist better measurements of social determinants of opportunity, our framework naturally incorporates them into the analyses.
(continuing)
(continued)
C4: "I don’t think this edge is an endorsement of racial essentialism - it just summarizes dozens of effects that the model is too coarse to model explicitly."
A4: There might be some misunderstandings. We did not claim that the edge is an endorsement of racial essentialism. Instead, we are worried about the unintentional alignment in implications with racial essentialism, as an (unintentional) outcome of the seemingly neutral technical choice.
Furthermore, as you kindly pointed out, such edge summarizes dozens of effects that previous frameworks are too coarse to model explicitly. If the modeling itself is too coarse, the precision and comprehensiveness needed for analyzing the legal/societal implication and policy intervention will face additional challenges. This is exactly part of motivation behind our framework to address this issue and make the modeling more fine-grained.
C5: "[The authors' claim] is certainly true, but to [reviewer’s] knowledge not a single causal graph in history has ever actually described a real-world causal process where humans were involved. It’s extremely difficult to establish single treatment effects, let alone a network of them."
A5: Thanks for sharing your insights.
We agree that causal graph has its limitations. However, recent advances in causal discovery and causal representation learning suggest that under mild assumptions, we may be able to discover causal relations among both observed and latent causal variables with identifiability guarantees (see, e.g., Xie et al. 2020, Schölkopf et al. 2021, Huang et al. 2022, Dong et al. 2024, Zhang et al. 2024). As the discovery methods identify more and more latent causal factors that are essential and necessary, we can get closer and closer to the underlying true causal process.
At the same time, we expect and sincerely hope that ideas and methods regarding how to achieve fairness can be developed in parallel with causal discovery and causal representation learning, such that these research efforts can inspire and benefit each other.
References
Dong, X., Huang, B., Ng, I., Song, X., Zheng, Y., Jin, S., Legaspi, R., Spirtes, P. & Zhang, K. (2024). A versatile causal discovery framework to allow causally-related hidden variables. International Conference on Learning Representations.
Huang, B., Low, C. J. H., Xie, F., Glymour, C., & Zhang, K. (2022). Latent hierarchical causal structure discovery with rank constraints. Advances in Neural Information Processing Systems, 35, 5549-5561.
Schölkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., & Bengio, Y. (2021). Toward causal representation learning. Proceedings of the IEEE, 109(5), 612-634.
Xie, F., Cai, R., Huang, B., Glymour, C., Hao, Z., & Zhang, K. (2020). Generalized independent noise condition for estimating latent variable causal graphs. Advances in Neural Information Processing Systems, 33, 14891-14902.
Zhang, K., Xie, S., Ng, I., & Zheng, Y. (2024). Causal representation learning from multiple distributions: A general setting. International Conference on Machine Learning.
Thanks for the response.
To clarify Q3, I was asking how, given that social determinants of opportunity are mostly defined in terms of geography and the social networks people are in, we can say that the latent regions in the experiment (as opposed to regions defined by known geographies or social networks) capture social determinants of opportunity.
At the moment I stand by my review/score based on: a) the findings of the model being well-known, and b) the application of the model to real data being uncompelling (due to the lack of granular data).
Thanks for clarifying Q3. In the additional experimental analyses on the US census data, the PUMAs are regions defined by known geographies or social network. The material can be found in Appendix C.2 of our revised manuscript.
We believe there are still some misunderstandings. Please kindly allow us to clarify:
a) "the findings of the model being well-known"
Re: a) We understand that similar arguments exist in debates, e.g., in legal cases. However, we would like to note that
-
Our work of precisely capture contextual influences via causal modeling is novel, there is no overlap with previous algorithmic fairness literature
-
Due to the issues of previous modeling (Section 3.1), previous causal fairness approaches do not produce the nuanced analyses facilitated by our framework
-
The algorithmic fairness literature is naturally multi-disciplinary, many of the ideas often echo wisdom from other discipline (philosophy, legislation, sociology, etc.). The fact that our model provides quantitative and transparent findings, that are aligned with arguments in other disciplines, demonstrate the value of our approach, especially these findings are not produced by previous causal fairness approaches.
b) "the application of the model to real data being uncompelling (due to the lack of granular data)"
Re: b) In our revised manuscript, we have followed your original suggestion and provided extensive analyses of the US Census data, where granular data is available. We have added the material in Appendix C.2, along with a side note on page 26 to help locate the content.
Please kindly let us know if our revised manuscript and the clarification help address your concern. Thanks again for your time and careful review.
It looks like the census data in the Appendix is just a discussion of how different PUMAs differ? Is the model in this paper applied at all to this data?
And unfortunately I don't see the causal modeling in this paper to be a major contribution. In practice it's an additional binary variable (poor region vs rich) and an additional conditional independence statement. Conceptually it's nice to consider sociogeographic factors in this sort of analysis but in practice this is a simple extension of previous work.
Thank Reviewer 8F2j for engaging in further conversation.
The additional analyses on US Census data are for the purpose of demonstrating how different regions instantiate very different contextual environments, and calling for attention to the social determinants of opportunity. We view our approach as a necessary and important first step to raise the salience of geographic and community influences on opportunity over a reductive focus on ethnicity, as you kindly pointed out.
We respectfully disagree with the comment that our approach is "an additional binary variable and an conditional independence statement [..., and] in practice is a simple extension of previous work":
-
As demonstrated in the census data analysis, the region is not just a binary variable, but a surrogate to social determinants of opportunity of various contextual environments (we explicitly mentioned that it can go beyond binary cases, e.g., at line 457). Our framework naturally incorporate better measurements of social determinants of opportunity, if they are available.
-
We identify and aim to address the potential issues of previous causal fairness approaches, namely, the recapitulation of stereotypes, the limited scope of individual-level variables, and omitting relevant variables. In practice, addressing these issues involves intentionally looking for or developing better measurements of social determinants of opportunity, instead of causal effects originating from the protected feature.
Thanks again for sharing your thoughts. Please let us know if there is remaining question or concern on any specific point.
This paper discusses the consideration of "social determinants of opportunity" such as geographical locations for algorithmic fairness.
优点
The discussion of the actual effects of different approaches to achieve "fairness" is discussed, which is often not considered enough in our field.
缺点
First of all, I am not sure if this conference is a good fit for this paper/topic since it is from my perspective hardly at all concerned with "learning representation". It is more a general societal consideration about how fairness could be achieved.
While the claim of the paper is to discuss the "social determinants of opportunity" in general, the discussion focusses very much on a single use case, i.e., university admissions.
The paper is written in a very US-centric way, specifically considering the legal situation.
The case considerations in Section 4 often (e.g., Section 4.3.) come to conclusions that are quite trivial. E.g., that taking the top-x % per region increases the share of "weaker" regions was literally my first thought. The accompanying formulas appear to just make a trivial insight more sophisticated.
The authors should more critically reflect on their approach. For example, (i) even if "academic preparedness" is caused by certain external factors, isn't academic preparedness still a key factor to a succesful university curriculum? If someone is not well prepared for university, they should not be admitted - that should be at the core of all admission procedures (ii) the legal implications of adjusting for "social determinants of opportunity" should be considered, specifically if this correlates with sensitive attributes such as race. (iii) trying to form groups again beyond sensitive attributes - again - introduces new sources of unfairness. For example "poor" students growing up in "rich" regions. Also, if this kind of admission procedure would gain traction, it would also be possible to trick procedures, e.g. for "rich" people renting temporarily to appear to be from a "poor" region.
The assumptions in the paper appear to be somewhat arbitrary. E.g., why assume the gamma parameterization in 4.3., and how is this justified?
The writing of the paper should also be improved. For example, what is the purpose of Section 2.1. For the contents of the paper?
问题
What are the authors thoughts of the critical reflection of the approach (see weaknesses)? Why is this paper a good fit for specifically ICLR?
(continued)
C5.1: Critical Reflection - "Even if 'academic preparedness' is caused by certain external factors, isn't academic preparedness still a key factor to a successful university curriculum?"
A5.1: Yes, you are totally right. This is exactly why our framework presents no objection to the importance of academic preparedness itself. Instead, we aim to address the issue of attributing discrimination only through the simplified relationship between race and the academic preparedness (as in previous causal fairness approaches). We argue that social determinants of opportunity, while not being individual-level attributes, correlate with race and should be considered explicitly to achieve procedural fairness.
C5.2: Critical Reflection - "The legal implications of adjusting for 'social determinants of opportunity' should be considered, specifically if this correlates with sensitive attributes such as race."
A5.2: Thanks for the constructive comment. We totally agree, and this is actually part of the reason behind our special attention to the clear trajectory of legal cases.
In terms of legal implication of "adjusting for social determinants of opportunity", we wholeheartedly agree that this question should be addressed, and our framework of precisely and explicitly modeling the relationship (which is often not considered enough in our field, as you kindly pointed out) would serve as an important and necessary first step.
C5.3: Critical Reflection - "Trying to form groups again beyond sensitive attributes - again - introduces new sources of unfairness."
A5.3: Thanks for the thoughtful question. There might be some misunderstandings and please allow us to clarify.
We are not trying to form new groups beyond sensitive attributes and treat them as if they were from a different group. Instead, we aim to explicitly model and address the different boosts/impediments to opportunities faced by different demographic groups, when they are part of various contextual environments. In other words, the influence from contextual environments are not individual-level attributes attached to the person, and they will change accordingly if an individual is subject to a different context.
C5.4: Critical Reflection - "[what if] 'poor' students growing up in 'rich' regions ... [or] 'rich' people renting temporarily to appear to be from a 'poor' region."
A5.4: Thanks for trying to go further and consider potential adversarial behaviors. While the profile can appear to be different from the truth (e.g., the temporary rental at a different region), the underlying mechanism cannot be easily faked. If the student is attending a specific school, the school's influence on the student is not altered by where the student (temporarily) lives. Furthermore, the potential adversarial behavior can be modeled in dynamic settings as an extension of our framework, which is a natural direction for further research.
Q6: "The assumptions in the paper appear to be somewhat arbitrary. E.g., why assume the gamma parameterization in 4.3., and how is this justified?"
A6: Thanks for carefully considering the assumptions in theoretical analyses. The assumptions are not arbitrary.
According to educational research (see, e.g., Arthur et al., 2019), the distribution of student scores is roughly bell-shaped but is often not perfectly Gaussian. The distribution tends to skew towards the low-score end, and the support is often bounded (e.g., falls in [Min, Max]). Therefore in Assumption 4.3, we use Gamma distributions to parameterize the score distribution. They are versatile to model the skewness and long-tail behaviors, while at the same time facilitate closed-form theoretical analyses. We have included this discussion in Appendix C.1.2, along with a side note Re: Q6 by Reviewer g5W4 on page 25.
Q7: "What is the purpose of Section 2.1 for the contents of the paper?"
A7: Thanks for the careful reading and thoughtful question. Section 2.1 provides a brief introduction of causal modeling with a directed acyclic graph (DAG). Since in our framework we use a DAG to represent causal process, which is also utilized in previous causal fairness literature. For completeness, we introduce the DAG representation of causality together with our notation conventions in Section 2.1.
References
Arthurs, N., Stenhaug, B., Karayev, S., & Piech, C. Grades Are Not Normal: Improving Exam Score Models Using the Logit-Normal Distribution. Proceedings of the 12th International Conference on Educational Data Mining, 2019.
Sowell, Thomas. Affirmative Action Around the World: An Empirical Study. Vol. 67. Yale University Press, 2004.
Thanks for the thoughtful and detailed comments, as well as the time and effort devoted! We have provided a revised manuscript, where we use blue font to indicate added/revised material. Below please also see our responses to specific comments and questions:
Q1: "Why is this paper a good fit for specifically ICLR?"
A1: Thanks for the question. According to ICLR 2025 Call for Papers, the non-exhaustive list of relevant topics include "societal considerations including fairness, safety, privacy." Our paper is about procedural fairness ("a societal consideration about how fairness could be achieved" as you kindly summarized), which is very relevant to ICLR.
C2: "While the claim of the paper is to discuss the 'social determinants of opportunity' in general, the discussion focuses very much on a single use case, i.e., university admissions."
A2: Thanks for asking about the scope of the discussion. We strive to balance between a broad discussion and a case study. We believe a concrete empirical setting would be helpful to demonstrate the nuanced analyses our framework facilitates, which can be applied to more general practical scenarios other than university admission.
In light of your comment, in Appendix A.4, we have added discussions on the role played by social determinants of opportunity in various scenarios, including health, education, and employment, along with a side note Re: C2 by Reviewer g5W4 on page 20 to help locate the related material.
C3: "The paper is written in a very US-centric way, specifically considering the legal situation."
A3: Thanks for the comment. We would like to respond in twofold: (1) why we pay special attention to US legal cases, and (2) the implication of our framework is not only limited to US.
-
We pay special attention to US legal cases in part because of the clear trajectory of jurisprudence, many of which reached the US Supreme Court. Since algorithmic fairness is naturally a cross-disciplinary topic, we aim to provide a new perspective through the explicit causal modeling of social determinants of opportunity for procedural fairness.
-
The implication of our framework is not only limited to US. As pointed out in previous literature (see, e.g., Sowell, 2004), quotas and group preferences (although under a variety of names) have existed in various countries with different histories and traditions. Incorporating social determinants of opportunity in procedural fairness analysis can potentially help address different scenarios that share very similar characteristics.
In light of your comment, we have included the above discussion in Appendix A.4, along with side node Re: C3 by Reviewer g5W4 on page 20.
C4: "The case considerations in Section 4 often (e.g., Section 4.3.) come to conclusions that are quite trivial ... The accompanying formulas appear to just make a trivial insight more sophisticated."
A4: Thanks for sharing your thoughts that our findings are not that surprising. We do not view not-being-surprising as a shortcoming, especially when previous algorithmic fairness approaches (e.g., causal fairness) do not produce such conclusions, because of the issues discussed in Section 3.1.
Furthermore, in addition to the unintended consequences of certain policies, our framework also facilitates quantitative and causal analyses that aim to uncover the reason behind these phenomena. Instead of "what might or might not happen", our framework precisely quantifies what will definitely happen under the scenario. The fact that our findings align with domain knowledge actually demonstrates the value of our approach towards procedural fairness.
(continuing)
Thank Reviewer g5W4 for the detailed and thoughtful comments. We have prepared point-by-point responses, and a revised manuscript together with color-coded side notes to help locate the related material.
As the reviewer-author discussion phase quickly approaching an end, we are very eager to know if the clarifications and additional materials help address the questions and comments, and especially, potential misunderstandings.
Thanks again for the time and effort devoted. We are eagerly looking forward to your feedback.
Sincerely,
Authors of Submission 4488
Dear Reviewer g5W4,
Thanks again for your thoughtful and detailed comments.
As we replied in our point-by-point responses, our work is very relevant to ICLR as per the call-of-paper. We have also provided a revised manuscript with further discussions and additional experiments, together with color-coded sidenotes to help locate relevant materials. As the discussion phase quickly coming to an end, we will be very grateful for an opportunity to engage in a conversation. We look forward to your feedback, and we are eager to understand if the original questions and concerns are resolved.
Yours sincerely,
Authors of Submission 4488
The paper addresses procedural fairness by modeling social determinants of opportunity and their causal influence on individual attributes, such as academic preparedness. Using college admissions as a central example, the authors explore various admissions policies and their potential impacts on fairness. They argue that accounting for social determinants of opportunity can mitigate procedural unfairness, which is often overlooked by traditional fairness approaches that focus solely on protected attributes like race or gender.
Overall, the reviewers agree that the research question is important and acknowledge the authors' efforts in addressing it. However, concerns were raised about the (potentially) limited empirical contributions, reliance on a US-centric legal framework, and the lack of generalizable findings beyond college admissions. While the authors provided responses to the reviewers' concerns, including discussions of the modeling contribution, insights from the modeling framework, and the generalizability of the findings, the reviewers’ opinions remain largely unchanged. Therefore, we recommend rejecting the paper in its current form but hope the authors find the reviewer comments helpful for future revisions.
审稿人讨论附加意见
The authors provided responses to the reviewers' concerns, including discussions of the modeling contribution, insights from the modeling framework, and the generalizability of the findings, the reviewers were not entirely convinced and their opinions remain largely unchanged.
Reject