Primary Submission Category: Synthetic Control Method
The Perils of Nonstationary Data in Synthetic Control Applications
Authors: Hongyu Mou, Yiqing Xu, Ziyi Liu, Yifan Sun,
Presenting Author: Hongyu Mou*
The synthetic control method (SCM) is widely used to estimate treatment effects in comparative case studies. However, inference with SCM remains challenging. Existing inferential approaches include randomization inference, which requires random assignment of the treatment, or conformal inference, which demands stationary or cointegrated error term time-series. While justifying random assignment is often difficult, our replication of thirteen SCM applications in economics and political science shows that the stationarity requirement is frequently unmet. Rosenbaum-type sensitivity analyses or analyses based on stationarized data indicate that many existing SCM findings are either spurious or underpowered. To obtain valid inference, we recommend a programmatic diagnostic procedure for future SCM applications.