2020-01-03 | Xiaojie Mao: Partial identification for unmeasured confounding with applications in causal inference and algorithmic fairness
Unmeasured confounding is one major challenge to reliable causal inference and decision making from observational data. The existence of unmeasured confounders can render the parameter of interest unidentifiable, i.e., it is impossible to learn the parameter even with infinite amount of observed data. As a result, the usual point estimators are generally biased, which may lead to misleading causal conclusions, invalidate evaluation of decision rules, and generate harmful personalized decisions. In this talk, I will present partial identification analysis, a general approach to deal with unmeasured confounding. Partial identification analysis aims to learn all possible values of the parameters of interest under reasonable confounding. I will demonstrate this technique in two applications: (1) estimating the conditional average treatment effect that is important for deriving personalized decision rule; (2) evaluating the outcome disparity of decisions (e.g., loan application approval, recidivism risk assessment algorithm, etc.) with respect to some protected attributes (e.g., race and ethnicity) when the protected attributes cannot be observed directly and must be estimated on an auxiliary dataset.
This talk is based on the following papers:
Nathan Kallus, Xiaojie Mao, and Angela Zhou (2019). "Interval Estimation of Individual-Level Causal Effects Under Unobserved Confounding." The 22nd International Conference on Artificial Intelligence and Statistics.
Jiahao Chen, Nathan Kallus, Xiaojie Mao, Geoffry Svacha, & Madeleine Udell (2019). Fairness under unawareness: Assessing disparity when protected class is unobserved. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 339-348). ACM.
Nathan Kallus, Xiaojie Mao, Angela Zhou (2019). Assessing Algorithmic Fairness with Unobserved Protected Class Using Data Combination. Major revision in Management Science Special Issue on Data-Driven Prescriptive Analytics.
Xiaojie Mao (毛小介) is a PhD candidate in Department of Statistics and Data Science at Cornell University, and is currently based in Cornell Tech campus in New York City. He researches at the intersection of causal inference and statistical machine learning. He is particularly interested in developing flexible and robust statistical methods for data-driven decision-making that involves causal reasoning. See https://xiaojiemao.github.io/ for more information.