Causal Inference beneath Incentives: An Annotated Studying Checklist – Machine Studying Weblog | ML@CMU

Causal inference is the method of figuring out whether or not and the way a trigger results in an impact, sometimes utilizing statistical strategies to tell apart correlation from causation. Studying causal relationships from knowledge is a crucial job throughout all kinds of domains starting from healthcare and drug growth, to internet marketing and e-commerce. In consequence, there was a lot work within the literature on economics, statistics, laptop science, and public coverage on designing algorithms and methodologies for causal inference.

Whereas many of the focus has been on questions that are statistical in nature, one should additionally take game-theoretic incentives into consideration when doing causal inference about strategic people who’ve a desire over the therapy they obtain. For instance, it could be arduous to deduce causal relationships in randomized management trials when there may be non-compliance by individuals within the examine (i.e. when individuals don’t adhere to the therapy they’re assigned). Extra usually, causal studying could also be tough every time people are free to self-select their very own remedies and there may be enough heterogeneity between people with totally different preferences. Even when compliance might be enforced, people might strategize by modifying the attributes they current to the causal inference course of so as to be assigned a extra fascinating therapy.

This annotated studying listing is meant to function a short abstract of labor on causal inference within the presence of strategic brokers. Whereas this listing isn’t complete, we hope that it is going to be a helpful place to begin for members of the machine studying neighborhood to study extra about this thrilling analysis space on the intersection of causal inference and recreation concept.

The studying listing is organized as follows: (1, 3) examine non-compliance in randomized trials, (2-4) give attention to instrumental variable strategies, (4-6) contemplate incentive misalignment between the person working the causal inference process and the topics of the process, (7,8) examine cross-unit interference, and (9,10) are about artificial management strategies.

1. [Robins 1998]: This paper gives an summary of strategies to right for non-compliance in randomized trials (i.e., non-adherence by trial individuals to the therapy project protocol).
2. [Angrist et al. 1996]: This seminal paper outlines the idea of instrumental variables (IVs) and describes how they can be utilized to estimate causal results. An IV is a variable that impacts the therapy variable however is unrelated to the end result variable besides via its impact on the therapy. IV strategies leverage the truth that variation in IVs is unbiased of any confounding so as to estimate the causal impact of the therapy.
3. [Ngo et al. 2021]: Not like prior work on non-compliance in scientific trials, this work leverages instruments from data design to disclose details about the effectiveness of the remedies in such a approach that individuals turn out to be incentivized to adjust to the therapy suggestions over time.
4. [Harris et al. 2022]: This paper research the issue of creating selections a few inhabitants of strategic brokers. The authors make the novel remark that the evaluation rule deployed by the principal is a sound instrument, which permits them to use commonplace strategies for instrumental variable regression to study causal relationships within the presence of strategic conduct.
5. [Miller et al. 2020]: This paper considers the issue of strategic classification, the place a principal (i.e. resolution maker) makes selections a few inhabitants of strategic brokers. Given data of the principal’s deployed evaluation rule, the brokers might strategically modify their observable options so as to obtain a extra fascinating evaluation (e.g., a greater rate of interest on a mortgage). The authors are the primary to point out that designing good incentives for agent enchancment (i.e. encouraging strategizing in a approach which really advantages the agent) is at the very least as arduous as orienting edges within the corresponding causal graph.
6. [Wang et al. 2023]: Incentive misalignment between sufferers and suppliers might happen when common handled outcomes are used as high quality metrics. Such misalignment is usually undesirable in healthcare domains, as it could result in decreased affected person welfare. To mitigate this subject, this work proposes another high quality metric, the entire therapy impact, which accounts for counterfactual untreated outcomes. The authors present that rewarding the entire therapy impact maximizes complete affected person welfare.
7. [Wager and Xu 2021]: Motivated by purposes resembling ride-sharing and tuition subsidies, this work research settings by which interventions on one unit (e.g. an individual or product) might have an affect on others (i.e., cross-unit interference). The authors give attention to the issue of setting supply-side funds in a centralized market. They use a mean-field modeling-based method to mannequin the cross-unit interference, and design a category of experimentation schemes which permit them to optimize funds with out disturbing the market equilibrium.
8. [Li et al. 2023]: Like [Wager and Xu 2021], this paper research the consequences of cross-unit interference, though the interference thought of right here comes from congestion in a service system. In consequence, the interference thought of right here is dynamic, in distinction to the static interference thought of within the earlier entry.
9. [Abadie and Gardeazabal 2003]: That is the primary paper on artificial management strategies (SCMs), a preferred method for estimating counterfactuals from longitudinal knowledge. Within the SCM setup, there’s a pre-intervention time interval throughout which all models are beneath management, adopted by a post-intervention time interval when all models endure precisely one intervention (both the therapy or management). Given a take a look at unit (who was given the therapy) and a set of donor models (who remained beneath management), SCMs use the pre-treatment knowledge to study a relationship (often linear or convex) between the take a look at and donor models. This relationship is then extrapolated to the post-intervention time interval so as to estimate the counterfactual trajectory for the take a look at unit beneath management.
10. [Ngo et al. 2023]: A typical assumption within the literature on SCMs is that of “overlap”: the outcomes for the take a look at unit might be written as a mixture (e.g., linear or convex) of the donor models. This work sheds mild on this typically ignored assumption and exhibits that (i) when models choose their very own remedies and (ii) there may be enough heterogeneity between models preferring totally different remedies, then overlap doesn’t maintain. Like [Ngo et al. 2021], the authors use instruments from data design and multi-armed bandits to incentivize models to discover totally different remedies in a approach which ensures that the overlap situation will steadily turn out to be happy over time.

[Editor’s note: this article is cross-posted in SIGecom Exchanges 22.1.]

References:

1. Abadie, A. and Gardeazabal, J. 2003. The economic costs of conflict: A case study of the basque country. American economic review 93, 1, 113–132.
2. Angrist, J. D., Imbens, G. W., and Rubin, D. B. 1996. Identification of causal effects using instrumental variables. Journal of the American statistical Association 91, 434, 444–455.
3. Harris, K., Ngo, D. D. T., Stapleton, L., Heidari, H., and Wu, S. 2022. Strategic instrumental variable regression: Recovering causal relationships from strategic responses. In International Conference on Machine Learning. PMLR, 8502–8522.
4. Li, S., Johari, R., Kuang, X., and Wager, S. 2023. Experimenting under stochastic congestion. arXiv preprint arXiv:2302.12093.
5. Miller, J., Milli, S., and Hardt, M. 2020. Strategic classification is causal modeling in disguise. In International Conference on Machine Learning. PMLR, 6917–6926.
6. Ngo, D., Harris, K., Agarwal, A., Syrgkanis, V., and Wu, Z. S. 2023. Incentive-aware synthetic control: Accurate counterfactual estimation via incentivized exploration. arXiv preprint arXiv:2312.16307.
7. Ngo, D. D. T., Stapleton, L., Syrgkanis, V., and Wu, S. 2021. Incentivizing compliance with algorithmic instruments. In International Conference on Machine Learning. PMLR, 8045–8055.
8. Robins, J. M. 1998. Correction for non-compliance in equivalence trials. Statistics in medicine 17, 3, 269–302.
9. Wager, S. and Xu, K. 2021. Experimenting in equilibrium. Management Science 67, 11, 6694–6715.
10. Wang, S., Bates, S., Aronow, P., and Jordan, M. I. 2023. Operationalizing counterfactual metrics: Incentives, ranking, and information asymmetry. arXiv preprint arXiv:2305.14595.