Towards Bridging Causal Inference and Algorithmic Decision-Making

Abstract: The goal in causal inference is to estimate counterfactual outcomes of units (e.g. patients, customers, subpopulations) under different interventions (e.g. medical treatments, discounts, socioeconomic policies). However, the end goal in practice is often to use these counterfactual estimates to make a decision which optimizes some downstream objective (e.g., maximizing life expectancy or revenue, minimizing unemployment). To bridge counterfactual estimation and decision-making, there are additional challenges one must take into account. We study two such challenges: (i) interventions are applied adaptively using some learning algorithm, (ii) units are strategic in what data they share about themselves. Specifically, we focus on the setting of panel data, where a learner observes repeated, noisy measurements of units over time. This talk is based on the following papers: https://arxiv.org/pdf/2307.01357 (NeurIPS 2023) and https://arxiv.org/pdf/2312.16307 (preprint).