We are running a Bayesian estimation code with a medium scale DSGE model. My coauthor and myself are running the same code with the same data, same priors, same guess values in three different machines. Posterior means of parameter estimates are very different. Mean variance decomposition are different. Theoretical moments, VD from stochastic simulation and irf are although similar. We are running dynare 4.5.1. I am using matlab2016b. Coauthor uses 2016a in one machine and b in another.
I am wondering whether co processor speed matters in convergence. This is quite perplexing.
This sounds quite strange. Is the mode found and the value of the posterior at the mode the same?
Never mind. We checked again. We get the same results. We must have shared different files earlier. Sorry about it.
However, a nagging issue still stays. The theoretical variance decomposition vastly differs from the mean variance decomposition based on the posterior mean estimates. I reported this issue earlier. In the theoretical VD following stochsimul one shock implausibly dominates while the posterior mean variance VD looks very reasonable. To investigate this issue further I manually imputed the posterior mean parameter estimates in a stand alone mod file with the same codes but without the estimation loop. The stochastic simulation comes close to the mean variance VD not the theoretical VD. There may be a bug in the programme.
Please provide the files to replicate the issue.