code & data

Research code, datasets, and replication materials from the lab and collaborators. Links to the corresponding papers are on the research page.

Code and replication materials

Causal inference and policy learning

• Batch-adaptive annotations for causal inference — batch-adaptive annotation allocation and AIPW estimation with expert and imputed labels; includes notebooks for simulated and RetailHero data (Nwankwo, Goldkind, Zhou, AISTATS 2026; maintained by Ezinne Nwankwo).
• Empirical Gateaux derivatives for causal inference — supplementary code for computerized influence functions and finite-difference bias correction (Jordan, Wang, Zhou, NeurIPS 2022; zip download).
• Structured Difference-of-Q — orthogonalized R-learner-style estimation of Q-function contrasts for offline RL (Cao & Zhou, AISTATS 2026).
• Stateful offline contextual policy evaluation and learning — supplementary code for sequential policy evaluation and learning with exogenous arrivals and resource states (Zhou, AISTATS 2022; zip download).
• Interval estimation of individual-level causal effects — sensitivity-analysis bounds on CATE under unobserved confounding (Kallus, Mao, Zhou, AISTATS 2019).
• Confounding-robust policy improvement — minimax-regret policy learning under unobserved confounding (Kallus & Zhou, NeurIPS 2018 / Management Science 2021).
• Confounding-robust infinite-horizon OPE — partially identified bounds for off-policy evaluation in infinite-horizon RL (Kallus & Zhou, NeurIPS 2020).
• Continuous policy learning — off-policy evaluation and optimization with continuous treatments (Kallus & Zhou, AISTATS 2018).

Algorithmic fairness

• xAUC — fairness metrics for bipartite ranking with risk scores (Kallus & Zhou, NeurIPS 2019).
• Fairness with unobserved protected class — partial identification of disparity measures via data combination (Kallus, Mao, Zhou, Management Science 2020).
• Disparate impact of personalized interventions — auditing personalized interventions with partially identified ROC/xROC curves (Kallus & Zhou, NeurIPS 2019).
• Optimal and fair encouragement policy evaluation and learning — supplementary code for optimal encouragement policies with fairness constraints and non-adherence (Zhou, NeurIPS 2023; zip download).

Decision-aware machine learning

• Task-loss reweighted MSE — reweighing prediction loss by decision regret for contextual stochastic optimization (Lawless & Zhou, 2022).
• Off-policy evaluation with policy-dependent optimization response — supplementary code for causal policy evaluation when outcomes feed a downstream linear optimization problem (Wang, Jordan, Zhou, NeurIPS 2022; zip download).

Data

• NYC bail reform synthetic controls — replication code and a constructed dataset of publicly reported crime data from 27 large municipalities, used to evaluate the impact of New York’s bail reform (Zhou et al., Statistics and Public Policy).