Discovering What Mattered: Answering Reverse Causal Questions by Detecting Unknown Treatment Assignment and Timing as Breaks in Panel Models

Link to Paper

Abstract:

Effective policy design requires knowing which interventions have made a measurable difference, yet outcomes such as emissions or growth often change without clear attribution. Rather than evaluating known policies, we start from the outcome itself and detect structural breaks in panel data, attributing them to potential causes. Formally, we develop an econometric framework that identifies unknown treatment timing and assignment as structural breaks in fixed-effects panel models. We show that a dummy-saturated Two-Way Fixed Effects model nests the standard staggered difference-in-differences design as a special case and that machine learning methods – such as indicator saturation and the adaptive LASSO – can recover treatment without prior knowledge. Established properties of indicator saturation allow control of the False Discovery Rate of discovered interventions. The approach provides a theoretical foundation for recent climate econometric applications and is demonstrated by identifying the Swedish carbon tax and detecting the economic impact of ETA terrorism. The methods are available in the open-source R package getspanel.

Pretis, Felix and Schwarz, Moritz, Discovering What Mattered: Detecting Unknown Treatment as Breaks in Panel Models (February 21, 2026). Available at SSRN: https://ssrn.com/abstract=4022745 or http://dx.doi.org/10.2139/ssrn.4022745