Different mode computation with different result

[Grant]

Dear Jpfeifer,

I use Bayesian to estimate some parameters of my model. However, when I use "mode_compute=4" and "mode_compute=5", I get different results through variance decomposition, although the both mode computation command can get good convergence result and identification analysis. Which result should I trust? Does it mean that my model is not robust and may have some fundamental problems?

Regards,

[jpfeifer]

It most probably means you did not get convergence to the same maximum/mode. Use the value of the posterior to see which mode was better.

[Grant]

Dear Jpfeifer,

Thanks for your answer. Could you explain a bit more in detail about how I can use the value of the posterior to check which mode was better? What is the criterion that mode is better?

Regards,

[jpfeifer]

The posterior is like the likelihood. The higher the better. After some mode-finders, Dynare is going to display the objective function at the mode. As the objective is minus the posterior, the lower the better.

[Grant]

Dear Jpfeifer,

Thanks for your explaination. I know that dynare can simulate the posterior distribution at the mode after some mode-finders. But I am still confused. Do you mean that I can compare the log data density? If not, how can I check which mode was better? I mean is there an explicit way to check?

[jpfeifer]

I am not talking about the log data density, but the posterior. If you are working with the unstable version, try using the file at https://github.com/JohannesPfeifer/dynare/commit/edca875d2a815ba7f98d68c64adc7ef18b69aaaf

[Grant]

Dear Jpfeifer,

I am not using the unstable version but the stable version 4.4.3. Do you mean than I can compare the value of "Objective function at mode" shown in the log file? However, I cannot find the corresponding value when I use "mode_compute=6". When I use "mode_compute=6", I can only find "Final value of the log posterior (or the likelihood)". Hope to receive your help. Thanks!

Regards,

[jpfeifer]

Those two things are the same. The objective function is minus the log-posterior. Note that mode_compute=6 also displays MINUS the log posterior.