"Log data density is -Inf" and Dynare version

Estimation
1

Hi,

I have a model that I estimate well under Dynare 4.4.3 with intuitive posterior means and, most importantly, a definite posterior log data density. When I estimate exactly the same mod file, with the same data, with Dynare 4.5.0 to 4.6.4, I always obtain “Log data density is -Inf” while the estimated parameters and other results are roughly similar.

I read the major user-visible changes from the Release of Dynare 4.5.0 email (2017), but I cannot understand what provokes such a difference.

  1. Does the Dynare team aware of this issue? Can I rely on Dynare 4.4.3 ?

  2. Is there a way to retrieve a definite calculation of the posterior log data density with the last Dynare version?

  3. May I use the last Dynare version (which is assumed to be more robust than the previous ones, right?) and use the log data density (Laplace approx, before MCMC-MH, which is definite) instead of the posterior log data density?

Thank you in advance for your help.

Thanks to all the Dynare team, you are doing a fantastic job!

Best,

Jonathan

2

Can you provide an example code?

3

Unfortunately not :frowning:
From your experience, did you hear about such an issue? Would you recommend relying on the estimation results of Dynare 4.4.3?

I think the other questions are also general enough not to require the model/mod file. What do you think?

4

Let’s try it generically.

  1. Having the Modified Harmonic Mean log data density estimator not work usually means something in the MCMC went wrong, e.g. one parameter getting stuck. A first step would be to compare the mode_check-plots and the trace_plot across the two versions.
  2. Regarding Laplace vs. Modified Harmonic Mean, seehttps://forum.dynare.org/t/the-log-marginal-data-density-in-mode-computer-6/4898/2