MCMC, likelihood estimation and identifiability problems in DLM models

In this article we deal with the identification problem within the Dynamic Linear Models family and show that using Bayesian estimation procedures we can deal better with these problems in comparison with the traditional Maximum Likelihood estimation approach. Using a Bayesian approach supported by Markov chain Monte Carlo techniques, we obtain the same results as the Maximum likelihood approach in the case of identifiable models, but in the case of non-identifiable models, we were able to estimate the parameters that are identifiable, as well as to pinpoint the troublesome parameters. Assuming a Bayesian approach, we also discuss the computational aspects, namely the ongoing discussion between single- versus multi-move samplers. Our aim is to give a clear example of the benefits of adopting a Bayesian approach to the estimation of high dimensional statistical models.