Statistical Inference with Strategic Agents: Accounting for Incentives and Information Asymmetry

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

This project aims to advance the understanding of statistical analysis within a broader environment with multiple stakeholders. In particular, the results of statistical hypothesis tests are often used to make yes/no decisions (e.g., whether to release a drug candidate to the public) that affect multiple parties. The decision materially impacts the party who developed the drug and public health generally, but in different ways, and this may impact how the stakeholders interact (e.g. when they choose to release data). We seek to develop statistical protocols that are robust to the behavior of different stakeholders who may have different aims. Progress on this front will lead to more reliable conclusions from data that are collected by multiple parties, contributing to the trustworthiness of the scientific enterprise. Moreover, the project is interdisciplinary in nature, bringing together ideas from statistics and also economics and social science. Fostering such connectivity is helpful for both fields, and such links are also useful educationally. The project also provides research training opportunities for students and postdocs. The technical goal of this proposal is to study statistical inference in settings that involve both incentive structures because actions and statistical decisions affect the behavior of other parties, and information asymmetry, as one party may have private information not available to others. In order to bring sharp focus to the core issues, this project tackles this challenge within a particular model of interaction known as the principal-agent model. As opposed to standard game-theoretic analyses of such interactions, the PIs study a statistical version of the principal-agent model, wherein the statistician plays the role of the principal. The principal has the goal of carrying out some form of statistical inference. In order to do so, the principal can interact with one or more agents that can provide statistically relevant information (e.g., a drug to test in a clinical trial, a feature set, datasets of varying and uncertain quality). Viewed as a two-person game, the principal moves first by specifying a statistical protocol, along with some kind of payment structure associated with it. The agent then makes its decision, which the PIs model it in terms of expected utility maximization. Our primary goal is to develop methodology and theory for hypothesis testing in this interactive setting. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.