A0779
Title: Global-deep synthetic control: A heterogeneous treatment effect estimator for shared market
Authors: Klaus Ackermann - Monash University (Australia) [presenting]
Priscila Grecov - Monash University (Australia)
Christoph Bergmeir - Monash University (Australia)
Abstract: Estimating heterogeneous treatment effects in time series is crucial in many real-world scenarios. Ensuring the no-interference assumption for control units can be challenging when measuring treatment effects via counterfactual and synthetic control methods in shared market environments. To address spillovers that bias results, global-deep synthetic control is introduced, which uses globally trained deep learning models to estimate treatment effects across multiple treated instances. Global-Deepsc handles violations of the non-interference assumption by forecasting counterfactual outcomes using only pre-intervention data. This approach reduces reliance on post-intervention disturbances, thus mitigating spillover interference and enhancing estimation accuracy. Moreover, the framework includes data-driven clustering and partitioning to eliminate control units affected by spillovers and identify variables causing heterogeneous impacts. Experiments on synthetic and real-world datasets demonstrate the method outperforms prominent structural causal models in treatment effect estimation and statistical significance tests. An experimental study analyzing sales promotion effects on supermarket sales found that younger, lower-income individuals should be the primary targets of marketing strategies, as identified by conditional treatment effect analysis.
