Primary Submission Category: Causal Inference in Networks
Staggered Rollout Designs with Clustering
Authors: Mayleen Cortez-Rodriguez, Matthew Eichhorn, Christina Lee Yu,
Presenting Author: Mayleen Cortez-Rodriguez*
Many approaches for estimating causal effects under interference rely on total knowledge of the underlying causal network, which is often unrealistic in practice. Recent work has shown that even with no network knowledge, one can still obtain unbiased estimates for causal effects by leveraging a staggered rollout experimental design and polynomial interpolation (PI). This approach can have high variance due to extrapolating a polynomial far from the support of the data. Additionally, it disregards potentially useful information or covariate data that may be available about the graph.
In this work, we investigate PI estimators under a two-stage experimental design wherein a graph clustering in the first stage selects a subpopulation on which a staggered rollout design is implemented in the second stage. Limiting the experiment to this subpopulation allows for a larger experimental budget in the second stage of the experiment, reducing the extrapolation error in high degree models. However, this approach can increase sampling error as the chosen subset may not be representative of the entire population. We provide experiments that illustrate the impact of homophily and clustering quality on this tradeoff between extrapolation and sampling error. We also explore the robustness of PI estimators under model misspecification. These experiments help us understand when clustering is a good choice for this style of estimator and design.