I work at the intersection of statistics and machine learning, developing methods for causal inference, experimentation, and measurement in large-scale behavioral systems.
Examines the benefits of blocked assignment in online A/B tests with highly skewed populations. Blocking targets the average treatment effect with fixed weights and enables targeted within-block winsorization to control noise while preserving tail-end treatment effects.
An empirical-Bayes approach to reducing selection bias when experiments are chosen for launch based on their observed effects, with experiment-specific local shrinkage and calibrated uncertainty.
A Bayesian framework for interim analysis in online experiments, designed to support resource-constrained decision-making without compromising the fidelity of the experiment.
A framework for treatment-effect learning when sequential treatment assignments create carry-over effects across sessions, combining T-learners with the g-formula.
A general chi-squared testing framework for statistical inference when sensitive group membership is protected with differential privacy. The approach unifies several common hypothesis tests and accounts for privacy noise when comparing outcomes across groups.
Studies how misreporting of sensitive binary outcomes affects the bias, variance, and statistical power of randomized trials. Develops a method for choosing sample sizes that guarantees adequate power under worst-case assumptions about misreporting.
A causal approach to estimating position bias by using variation from existing A/B tests as instrumental variables, avoiding the need for additional ranking experiments.
Introduces a deviation-from-equal-representation statistic and an alerting framework for finding potentially important representation effects across large-scale experimentation systems.
A local regression adjustment to random-forest methods that improves estimation for smooth signals, with applications to heterogeneous treatment effects.