Primary Submission Category: Bayesian Causal Inference
Early Stopping for Time-Varying Treatment Effects
Authors: Sam van Meer, Alberto Abadie,
Presenting Author: Sam van Meer*
When treatment effects evolve over time, early estimates can be misleading about long run program effectiveness. We model the trajectory of treatment effects with a semi-local linear trend model, where treatment estimates are noisy measurements of this latent process. Using filtering methods, we construct sequential confidence intervals for long run summary statistics such as cumulative or discounted treatment effects at each point in time as data accumulates. These intervals remain valid at arbitrary stopping times, allowing researchers to terminate data collection whenever evidence becomes sufficiently strong. We apply the framework to data from ASOS, an online retail platform, and demonstrate that our approach enables rejection of the null hypothesis well before the intended four week experimental horizon while maintaining the type I error guarantee.