mode compute=4or 6or 9

[ZBCPA]

Dear Johannes,

Could I ask one question which confuses me for a long long time, even though there have been already many similar questions on this forum?

When I start to find mode, (set mh_replic=0), sometimes no matter I use mode_compute=9, the result always shows"(minus) the hessian matrix at the "mode" is not positive definite!".
In this case, I will run mode_compute=6 to get a "nice" mode check plot. Then I use this mode file to run mode_compute=9 again, if hessian matrix is positive definite, then I run MCMC ,like 200000 draws with mode_compute=9;
If hessian matrix is still not positive definite, I just run MCMC with mode_compute=6 to get final results.( 9 is better than 6 to run MCMC if hessian matrix is positive definite)

Are these steps reasonable? or Could you give me some advice? By the way,is mode_compute=9 always better than 4 since the former one is globally finding mode, while the later one is locally.

Many thanks,
Huan

[jpfeifer]

The issue in the end always is to find the global mode. What you describe is an iterative procedure to find it and it sounds reasonable. However, you should keep track of whether there is actual improvement in the posterior density across runs (and particularly the MCMC)

If you are unsure where the mode is and you have rather diffuse priors that do not smooth out the likelihood by much, mode_compute=4 is not advisable, because it tends to get stuck at local modes