Strategic Causal Policy Learning with Manipulable Eligibility

arXiv:2611.14572 · stat.AP · Submitted 2026-11-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Strategic Causal Policy Learning with Manipulable Eligibility".

Jane: Here's the gist of the paper. The paper studies how strategic responses to an announced eligibility rule affect treatment allocation and welfare in causal policy learning settings where treatment eligibility is manipulable.

Tom: The single most informative sentence about the paper's subject and finding is: After a policy is announced, individuals may change the covariates used for assignment without changing their underlying treatment effects.

Jane: Formalizing Strategic Causal Policy Learning The authors formalize a model of strategic causal policy learning where treatment is assigned using strategically manipulated covariates but its value is determined by baseline treatment effects.

Tom: They consider pure gaming, where individuals can change the covariates used for treatment assignment without changing their potential outcomes.

Paper summary: Tom: So we've been looking at "Strategic Causal Policy Learning with Manipulable Eligibility" by Suhwan Bong et al., and this paper is about how people try to game the system when they know they can change what eligibility rules look like <ref:2611.14572#pg2>.

Jane: Right, it’s about those strategic responses to an announced rule, where individuals can manipulate the covariates used for assignment without changing their actual treatment effect <ref:2611.14572#pg2>.

Lu: The authors formalize this by defining a strategic treatment gain objective that differs from the standard value we usually calculate based on baseline effects <ref:2611.14572#pg3>.

Meng: For me, what’s interesting is how they break down the welfare loss into two parts: missed beneficial treatment and harmful treatment <ref:2611.14572#pg1>. That decomposition helps us see exactly where the policy is failing.

Lalam: And based on that, we can figure out if the gap between what you ideally want and what actually happens is because of bad assignments or because people just tried to game the rules <ref:2611.14572#pg3>.

Tom: So depending on whether you care most about total welfare, or keeping everyone safe, or balancing fairness across different groups, there’s a specific setting that works best for the allocation rule <ref:2611.14572#pg3>.

Lu: The paper shows how those different goals lead to different results when you look at the thresholds they find, especially when their costs for manipulation aren't uniform across different groups <ref:2611.14572#pg3>.

Meng: It suggests that if we build these learning systems, we need a way to estimate those costs and buffers accurately because the optimal rule changes depending on what’s driving the agent’s decision-making <ref:2611.14572#pg3>.

Jane: And that leads right into how we actually estimate those things—how do we figure out those manipulation buffers when you don't know the exact costs? That’s what we look at next.

Conclusion: Tom: So we're wrapping up our look at "Strategic Causal Policy Learning with Manipulable Eligibility" by Suhwan Bong et al., and this paper is really focused on how people try to game the system when they know they can strategically change what eligibility rules look like.

Jane: Right, it’s about those strategic responses to an announced rule, where individuals can manipulate the covariates used for assignment without changing their actual treatment effect.

Lu: The authors formalize this by defining a strategic treatment gain objective that's different from the standard value we usually calculate based on baseline effects.

Meng: For me, what’s interesting is how they break down the welfare loss into two parts: missed beneficial treatment and harmful treatment. That decomposition helps us see exactly where the policy is failing.

Lalam: And based on that, we can figure out if the gap between what we ideally want and what actually happens is because of bad assignments or because people just tried to game the rules.

Tom: So depending on whether you care most about total welfare, or keeping everyone safe, or balancing fairness across different groups, there’s a specific setting or threshold that works best for the allocation rule.

Jane: It seems like the main point is that for real-world policy learning with human agents, you can't just pick one single perfect allocation method. You have to choose your goal—welfare versus safety versus group fairness—and that choice dictates what kind of strategy you’re going to support in the model.

Lu: The authors show how those different goals lead to different results when you look at the thresholds they find, especially when their costs for manipulation aren't uniform across different groups.

Meng: It suggests that if we build these learning systems, we need a way to estimate those costs and buffers accurately because the optimal rule changes depending on what’s driving the agent's decision-making.

Lalam: What this really means for us is that when we learn how to allocate resources, we have to model not just the treatment effect, but also how much someone might try to game the rules themselves.

Tom: It’s a lot of moving parts, but it gives us a clearer picture of why simple allocation rules fail in strategic environments.

Jane: And that leads right into how we actually estimate these things—how do we figure out those manipulation buffers when we don't know the exact costs? That’s what we look at next.

Harvard University

stat.AP

Submitted: 2026-11-09

Updated: 2026-11-09

Comments: Accepted to the 40th Conference on Neural Information Processing Systems (NeurIPS 2026)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 100/100

The gist: The paper studies how strategic responses to an announced eligibility rule affect treatment allocation and welfare in causal policy learning settings where treatment eligibility is manipulable.

Key concepts

Strategic Causal Policy Learning
This model describes a situation where individuals actively try to manipulate the variables (covariates) used by a policy to determine their treatment assignment, aiming to maximize their personal treatment benefit, even though the actual treatment effect remains fixed.
Treatment-Gain Objective $G_S(oldsymbol{ au})$
This is the objective function individuals try to maximize. It measures the expected value of the treatment received ($ au$) multiplied by a strategic probability distribution $oldsymbol{ heta}$ over their possible covariates $oldsymbol{X}$. It differs from a non-strategic objective because individuals can choose how they present their characteristics.
Manipulation Gap Identity
This identity shows how changing the eligibility rule (the manipulation band) affects the policy's value. It decomposes welfare loss into components of missed beneficial treatment and harmful treatment, helping researchers understand how strategic responses shift the overall outcome compared to a non-strategic allocation.

Terminology

Summary

The paper studies how strategic responses to an announced eligibility rule affect treatment allocation and welfare in causal policy learning settings where treatment eligibility is manipulable. p. 9 The single most informative sentence about the paper's subject and finding is: After a policy is announced, individuals may change the covariates used for assignment without changing their underlying treatment effects.

Formalizing Strategic Causal Policy Learning

The authors formalize a model of strategic causal policy learning where treatment is assigned using strategically manipulated covariates but its value is determined by baseline treatment effects. They consider pure gaming, where individuals can change the covariates used for treatment assignment without changing their potential outcomes. p. 0 The strategic treatment-gain objective is defined as G S (π) = E[τ (X)π(Xπ)]. p. 2 This objective differs from the non-strategic value G N (π) = E[τ (X)π(X)]. p. 3

Decomposing Welfare Loss

The paper introduces an access-burden representation and derives a manipulationgap identity showing how strategic response changes policy value. They decompose welfare regret into missed beneficial treatment and harmful treatment. p. 1 The gap is positive when the manipulation band contains more positive-CATE mass than negative-CATE mass, and negative when harmful entrants dominate. p. 3

Characterizing Thresholds in Linear Models

For linear CATE models with quadratic manipulation costs, they derive closed-form characterizations of strategic treatment assignment. Homogeneous costs admit a single threshold correction that recovers the ideal positive-CATE allocation. p. 1 When manipulation costs are heterogeneous, welfare maximization, pointwise safety, and minimax group welfare regret can favor different thresholds. p. 0

Policy Objectives and Tradeoffs

The paper distinguishes between several policy objectives:

: Welfare:

Maximizing aggregate treatment gains need not eliminate harmful treatment, since individuals with negative CATEs may receive treatment under strategic response. The welfare-maximizing policy solves max π∈Π G S (π). p. 3

: Safety:

An institution may instead wish to limit harmful treatment, defining H(π) = E[(−τ (X))+mπ(X)]. A safety-constrained policy solves max π∈Π G S (π) subject to H(π) ≤ η, where η ≥ 0 is the permitted level of harm. p. 4

: Group welfare regret:

The group welfare regret Fz(π) = B⋆z − B S z (π) measures the difference between ideal and strategic values for each group. One possible fairness objective is to minimize maxz∈Z Fz(π) over the available policies. p. 4

Estimation and Learning Policies

The results of Section 4 characterize policy thresholds in terms of treatment effects and manipulation buffers. When the cost parameters are unknown, behavioral responses to a known, fixed eligibility rule can instead inform the buffer. p. 6 The empirical strategic welfare and group welfare regret are defined using estimated CATEs and estimated buffers. p. 7 The smallest common threshold that is pointwise safe relative to the estimated CATE is bksafe = maxz ρbz. p. 7

Experimental Results on Tradeoffs

Experiments illustrate those threshold tradeoffs and the sensitivity of learned policies to response and CATE estimation. The welfare-maximizing common threshold reduces harmful treatment while retaining most beneficial treatment, achieving the highest aggregate welfare among the common-threshold policies. p. 8 The safe threshold reduces harmful treatment further but misses more beneficial applicants, especially in groups whose smaller buffers make the common cutoff harder to reach. p. 8 The fairness threshold reduces the largest group welfare loss at a modest cost to aggregate welfare. p. 8

Limitations and Future Directions

Several limitations remain, including the focus on pure gaming where manipulation affects eligibility but not potential outcomes. Arbitrary nonlinear costs, constrained actions, and valid uniform CATE confidence bands remain demanding identification and inference problems. p. 9 Extending the framework to nonlinear CATE models, learned cost functions, and institutional constraints on group-specific thresholds is an important direction for future work. p. 9 The strongest closed-form results rely on linear CATEs and squared L2 costs. p. 9

References

Susan Athey and Stefan Wager. Policy learning with observational data. p. 0 Mark Braverman and Sumegha Garg. p. 9 The role of randomness and noise in strategic classification. p. 9 Qianyi Chen, Ying Chen, and Bo Li. p. 9 Practical performative policy learning with strategic agents. p. 9 Yiling Chen, Yang Liu, and Chara Podimata. p. 9 Learning strategy-aware linear classifiers. p. 9 Andrew Estornell, Sanmay Das, Yang Liu, and Yevgeniy Vorobeychik. p. 9 Unfairness despite awareness: group-fair classification with strategic agents. p. 9 Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. p. 0 Strategic classification. p. 0 Lily Hu and Yiling Chen. p. 9 A short-term intervention for long-term fairness in the labor market. p. 2 Nathan Kallus. p. 1 Balanced policy evaluation and learning. p. 3 Amir-Hossein Karimi, Bernhard Schölkopf, and Isabel Valera. p. 9 Algorithmic recourse: from counterfactual explanations to interventions. p. 2 Vijay Keswani and L Elisa Celis. p. 2 Addressing strategic manipulation disparities in fair classification. p. 2 Kwangho Kim and José R Zubizarreta. p. 2 Fair and robust estimation of heterogeneous treatment effects for policy learning. p. 6 Toru Kitagawa and Aleksey Tetenov. p. 0 Who should be treated? p. 2 empirical welfare maximization methods for treatment choice. p. 1 Jon Kleinberg and Manish Raghavan. p. 0 How do classifiers induce agents to invest effort strategically? p. 1 Charles F Manski. p. 0 Statistical treatment rules for heterogeneous populations. p. 1 John Miller, Smitha Milli, and Moritz Hardt. p. 2 Strategic classification is causal modeling in disguise. p. 0 Smitha Milli, John Miller, Anca D Dragan, and Moritz Hardt. p. 2 The social cost of strategic classification. p. 2 Evan Munro. p. 1 Treatment allocation with strategic agents. p. 1 Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt. p. 0 Performative prediction. p. 1 Roshni Sahoo and Stefan Wager. p. 9 Policy learning with competing agents. p. 1 Yonadav Shavit, Benjamin Edelman, and Brian Axelrod. p. 0 Causal strategic linear regression. p. 0 Berk Ustun, Alexander Spangher, and Yang Liu. p. 9 Actionable recourse in linear classification. p. 2 Sandra Wachter, Brent Mittelstadt, and Chris Russell. p. 2 Counterfactual explanations without opening the black box: Automated decisions and the gdpr. p. 0 Baqun Zhang, Anastasios A Tsiatis, Eric B Laber, and Marie Davidian. p. 1 p. 1 A robust method for estimating optimal treatment regimes. p. 1 Baqun Zhang, Anastasios A Tsiatis, Eric B Laber, and Marie Davidian. Robust estimation of optimal dynamic treatment regimes for sequential treatment decisions. p. 1

--- Page 9 ---

Table 1: Plug-in policy performance over 200 Monte Carlo replications with estimated CATEs and known manipulation costs. Policy bk GS Regret Harmful Missed maxz Fz Treat rate Naive k = 0 0. p. 8000 0.163 0.135 135 1278 a) Strategic welfare kmax =0. p. 8773 (a) Strategic welfare b) Worst-group regret kfair =0. p. 8 p. 8625 (b) Worst-group regret c) Harmful treatment ksafe =0. p. 6 p. 6901 (c) Harmful treatment

--- Page 10 ---

Table 2: End-to-end policy performance over 200 Monte Carlo replications with estimated CATEs and manipulation buffers. Policy Regret Harm Missed benefit Uncorrected naive rule 0. p. 9 p. 91278 0.1278 | Welfare, midpoint buffer 0.0118 0.036 | Safety, upper buffer 0.0159 36 |

--- Page 12 ---

Table X: Plug-in policy performance over 200 Monte Carlo replications with estimated CATEs and known manipulation costs. Policy bk GS Regret Harmful Missed maxz Fz Treat rate Naive k = 0. p. 8773 (a) Strategic welfare b) Worst-group regret kfair =0.625 (b) Worst-group regret c) Harmful treatment ksafe =0.901 (c) Harmful treatment

--- Page 11 ---

Table Y: End-to-end policy performance over 200 Monte Carlo replications with estimated CATEs and manipulation buffers. Policy Regret Harm Missed benefit Uncorrected naive rule 0.1278 36 | Welfare, midpoint buffer 0.0159 Safety, upper buffer 36 |

--- Page 8 ---

Figure 1: Population threshold tradeoffs under the three-group Gaussian design (a) Strategic welfare GS(k), (b) Worst-group welfare regret maxz Fz(k), and (c) Aggregate harmful treatment mass H(k).

--- Page 9 ---

Figure 1: Population threshold tradeoffs under the three-group Gaussian design (a) Strategic welfare GS(k), (b) Worst-group welfare regret maxz Fz(k), and (c) Aggregate harmful treatment mass H(k).

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Access Burden Representation for Strategic Treatment Systems: The system can use an access burden representation to show how strategic response changes policy value by characterizing treatment allocation as G S (pi) = E[tau (X)m(X)].

  2. Welfare Maximization with Safety Constraints: The AI can maximize aggregate treatment gains subject to a safety constraint H(pi) <= eta, where H(pi) is defined as E[(−τ (X))+m(X)], allowing the system to limit harmful treatment or exclude it entirely if eta = 0.

  3. Group Welfare Regret Minimization: The system can minimize group welfare regret by solving max z Fz(pi), where Fz(pi) = B⋆z - BSz(pi), allowing the AI to optimize for fairness across different groups based on immutable attributes Z.

  4. Closed-Form Threshold Derivation for Linear CATEs: For linear CATE models with quadratic manipulation costs, the system can derive closed-form characterizations of strategic treatment assignment, specifically finding the group-specific effective baseline cutoff mkz(x) = 1if tauz(x) >= k - ρz.

  5. Homogeneous Cost Optimization: When all groups have the same manipulation cost parameter lambda > 0, the system can implement a single common threshold k = ρ, which implements the ideal CATE allocation in every group and results in G S (ρ) = G⋆ and Fz(ρ) = 0 for all z.

  6. Heterogeneous Cost Tradeoff Analysis: The system can analyze the tradeoffs between objectives under heterogeneous costs by identifying that the welfare threshold depends on treatment-effect distributions as well as manipulation buffers, leading to different welfare, safety, and group welfare regret optimal thresholds.

  7. Estimation Under Uncertainty: When CATE must be learned from data, the AI can use estimated CATEs to determine empirical strategic welfare and group welfare regret by calculating GbS(k) and Fbz(k) using the formula G S (k) = 1/n Sum Xn i=1 τbi1if ai >= k.

  8. Buffer Estimation from Behavioral Responses: The system can use observed successful and unsuccessful responses to identify the buffer for the rule under which the responses were collected, providing lower and upper bounds for the manipulation buffer when costs are unknown.

Sources

Related papers