Bayesian estimation: Autocorrelation of MC


I’m wondering how DYNARE constructs the moments of posterior distributions. I understand that the first “mh_drop” percent of the chain is dropped. But, are all the other accepted candidates used to compute parameters posterior moments?

I’m asking because, apparently, RWM algorithm generates a MC of autocorrelated accepted parameters. So, some authors (e.g. Canova’s book, 2007) suggested to discard some accepted draws from the MC (e.g. take one value every k positions in the chain, or pick randomly one every N) to reduce autocorrelation. Does DYNARE do it?

Thanks in advance and congratulations for the toolbox. In my estimation DYNARE worked properly and efficiently

Hi Bene

Dynare uses all the other accepted candidates (the ones not dropped) to compute the moments. The autocorrelation in the draws is not a really relevant problem if it is there, but if you want to check it using the values picked every k positions in the chain you can do it by your own. The chains are stored in the NAMEOFTHEMODFILE_mhNUMBER_blckNUMBER.mat files.