Statistical Inference for Bandit Data

This lecture covers statistical inference for bandit data, focusing on personalized treatment actions in sequential decision-making studies. It discusses contextual bandit problems, adaptive weighting in least squares, and the challenges of standard estimators on bandit data. The instructor presents simulations with continuous and binary rewards, confidence regions, and the role of adaptive weights. The lecture concludes with a discussion on model misspecification, causal inference, and regret minimization. The content is based on the paper 'Statistical Inference with M-Estimators on Adaptively Collected Data' by Kelly Wang Zhang, Lucas Janson, and Susan Murphy, presented at NeurIPS 2021.