maybe all of you know that the default optimizer used in the first stage of the estimation, when dynare maximizes the loglik to find the posterior modes, has a random component (am I right?)…sometimes it happens that running the code more than once, I end up with different modes, different s.d., different Laplace approximation and so on… These values are then used to initialize the M-H. I know that the results from the M-H should be independent from the starting point, but I would like to ask if someone has noted the same strange behaviour of the that optimizer and if one should care about it or not…
yes i have noticed the same behavior. it is true that the sims optimizer has a random component. i try to do MCMC also to check if the mode that the optimizer found is indeed the right one.
thanks Reuben…I try as well to do the MCMC but apart from the fact that the results are awful in my case, the results of the optimizer are each time very different and this affects the MCMC (for instance in a running dynare starts the MCMC procedure, in the following it is not able to find the initial mh_scale value and I have to put 0.00001, or for instance in a running the optimizer finds the maximum after 6 iterations, and in another one at the 800th iteration it still has not found the maximum!!!). I was wondering if we can interpret the behavior of the optimizer as a sign of problems with the model (misspecification? bad priors?)
why dont do you just do the MCMC , concatenate the draws , and examine how the logpo2 , ie the posterior value, evolves?
if it is higher than the mode value that the optimizer gives, then you should give the parameter values corresponding to the mcmc mode