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

Much of empirical research focuses on forward causal questions (``Does X cause Y?'') while answering reverse causal questions (``What causes Y?'') can provide invaluable insights but is difficult to implement in practice. Here we operationalise the modelling of reverse causal questions through the detection of unknown treatment assignment and timing as structural breaks in fixed effects panel models. We show that conventional treatment evaluation of known interventions in a two-way fixed effects panel (often interpreted as difference-in-differences) is equivalent to allowing for heterogeneous structural breaks in the treated units' fixed effects. Using machine learning, we can thus detect previously unknown heterogeneous treatment effects as structural breaks in individual fixed effects corresponding to unit-specific treatment which can be subsequently attributed to potential causes. We demonstrate the feasibility of our approach by detecting the impact of ETA terrorism on Spanish regional GDP per capita without prior knowledge of its occurrence. Our proposed method to detect breaks in panel models can be readily implemented using our open-source R-package `gets' with the `getspanel' update or using the (adaptive) LASSO.