Invited Session in Identification of DSGE Models
We show that weak identification is a serious concern in empirically relevant DSGE models with many nominal, real and financial frictions, as the likelihood/posterior functions of these models are typically not well behaved. To this end, we offer (1) a formal Bayesian approach by using a well-established indicator based on the posterior precision to detect weak identification and (2) a set of applied tips on how to cope with the estimation difficulties. We focus particularly on which posterior sampling method (RW-MH, TaRB-MH, Slice) to use and how to fine-tune it. We find that the slice sampler performs very well in terms of number of draws and computational time, and thus should be used when performing this strength of identification analysis before taking a model to data.