This paper sets up a Gibbs sampler for a three state Markov switching model with nonconstant transition probabilities. The step from two to three states is accomplished by the use of a multinomial probit model for the latent variable process. The algorithm is then applied to Swiss GDP data in order to analyze the business cycle. The results suggest Markov switching between three different regimes. Furthermore, evidence for duration dependence in recessions is found, i.e., the longer a recession lasts the more likely it is to end.