Variable selection in panel models with breaks

Abstract

We develop a Bayesian approach that performs variable selection in panel regression models affected by breaks. Our approach enables deactivation of pervasive regressors and activation of weak regressors for short periods (regimes). We establish theoretical results on the concentration properties of the posterior as well as the rate of convergence for estimating the break dates. Our methodology is demonstrated in simulations and in an empirical application to firms’ choice of capital structure. We find that ignoring breaks can lead to overestimating the number of relevant regressors, but also a failure to activate regressors that are informative only in short-lived regimes.