I'll introduce a two-stage rollout design to mitigate the high variance from extrapolation in interpolation-based estimators for the total treatment effect under unknown network interference. Then, I'll quantify the bias and variance of this estimator as a function of a clustering used in the two-stage design and illustrate their trade-off through a series of experiments.
Aug 3, 2025
Interference effects, where the treatment of one individual has an effect on the outcome of another, are pervasive in real-world settings but violate assumptions of many classical causal estimators. While the Horvitz-Thompson estimator can account for interference, it has prohibitively high variance. In this talk, we'll survey recent approaches to improve on this variance guarantee by imposing additional structural assumptions on the potential outcomes model or the interference network. Then, I'll introduce a class of estimators, pseudoinverse estimators, that can be adapted to any experimental design and have strong bias and variance guarantees. For the setting of cluster randomized designs, I'll quantify the bias and variance as functions of the selected clustering.
May 26, 2025
Interference effects, where the treatment of one individual has an effect on the outcome of another, are pervasive in real-world settings but violate assumptions of many classical causal estimators. While the Horvitz-Thompson estimator can account for interference, it has prohibitively high variance. In this talk, we'll survey recent approaches to improve on this variance guarantee by imposing additional structural assumptions on the potential outcomes model or the interference network. Then, I'll introduce a class of estimators, pseudoinverse estimators, that can be adapted to any experimental design and have strong bias and variance guarantees. Finally, I'll show how theoretical bounds on the performance of the pseudoinverse estimator can provide practical advice when selecting an experimental design.
Jun 17, 2024
Poster
May 16, 2024
Poster
May 15, 2024
Recent work on causal inference under interference falls under two approaches, using structural assumptions on the interference effects to select a good randomized design or using structural assumptions on the potential outcomes to select a good estimator. In this work, we quantify the gains that can be made when these approaches are considered together, in particular by studying pseudoinverse estimators under cluster randomized designs.
Oct 18, 2023
We consider the problem of online allocation, where irrevocable decisions whether to allocate to agents must be made before all agents have been observed, in settings with priority and quota constraints, By leveraging structure from an offline variant of this problem, we develop a policy for which the sum of the efficiency loss (number of unallocated resource units) and priority loss (number of higher-priority agents who are blocked from an allocation by lower-priority agents) is constant with respect to the input size.
Oct 16, 2023
Poster
Oct 2, 2023
Poster
May 25, 2023
Poster
Dec 2, 2022