Problem with geweke's modified harmonic mean estimator

i get the following message after doing the MCMC replications.
Warning: Log of zero.

In metropolis at 822
In dynare_estimation at 778
In Version_25 at 448
In dynare at 26

have you guys had this problem with the harmonic mean estimator?


Yes, it happens when you start Metropolis with a bad mode estimate and when the standard error of some parameter is estimated to be zero. This can be because the optimizer get stock againt a boundary.

It this case, there is no movement in that direction during the Metropolis iterations and you get this “log of zero” message. You can also get it if you only run a very small number of Metropolis iterations (mh_replic).

If it is a mode problem, look at the plot of the posterior densities. If you ran a large number of Metropolis iterations and that the computed posteriror mode (green vertical line) is different from the mode of the simulated posterior distribution (black line), you can try to restart optimization, using the readings of the mode on the black line as initial values for the optimization (estimated_params_init)