Hello everyone. Because when I set mh_replic=25000, the results are good, but when I set mh_replic=250000 or greater, my results are particularly terrible, and a certain shock explains most of the fluctuations
You should check the convergence diagnostics and particularly the trace_plots. But usually, more draws are more reliable.
Dear professor，I found the parameter that made one of my shock huge, but its trace_plot was like this, what was going on.
Your posterior may be bimodal.
Dear professor, What I did was just change the Taylor rule from pegging output growth rates(The results are very good) to pegging steady-state output (my mentor’s suggestion), and then I get this results. How should I do to deal with this problem.
When I set mh_replic=1000000,the trace_plots are the same.
Did you have a look at the trace plot of the posterior? What happens there?
No. When you use generate_trace_plots, the first plot is for the posterior density.
This clearly shows you that your chain drifted to a region of higher posterior density, i.e. the initial point was not the true mode. Ideally, you restart estimation and mode-finding from the new mode.
Dear professor, what can I do to solve this problem. What I did was just change the Taylor rule from pegging output growth rates (The results are very good) to pegging output gap.
As I said, read out the new mode and restart mode-finding from there.
Dear Prof. Pfeifer,
You mentioned that “But usually, more draws are more reliable.” You used the word “usually”. Does this mean there could be some exceptions? A relatively small value could lead to a better result? For example, the results from 1million draws could be better than that from 1.1 million drawa.
Thank you,
Alex
It might be that you get stuck at a local mode and more draws do not help the MCMC traverse to regions of higher density. In that case, more draws are a waste of time.