One of the stylized facts of unemployment is that shifts in its mean rate between decades and half-decades account for most of its variance. In this paper, we use a statistical analysis based on Markov switching regression models to identify the dates of infrequent changes in the mean of the unemployment rate series of fifteen countries. We find that in most countries, unemployment persistence is much reduced once the (infrequently) changing mean rate, induced by large shocks to unemployment, has been removed. We conclude that the observed persistence in unemployment appears to be consistent with multiple equilibria models and models with an endogeneous natural rate.