Bayesian Indicator-Saturated Regression

Abstract

Structural break detection has emerged as an important tool for assessing the effects of policies in settings where conventional policy evaluation methods might not be applicable.In this paper, we introduce a unified Bayesian framework for detecting structural breaks with unknown timing and arbitrary sequence in longitudinal data. The proposed setup builds on a indicator-saturated regression design and uses a spike-and-slab prior for selection among indicators. We establish that a non-local prior as the slab component is a necessary condition to provide model selection consistency in this model class. Simulation results show that the method outperforms comparable frequentist approaches, particularly in environments with a high probability of structural breaks. We illustrate the proposed framework by analysing climate policies in the European road transport sector.