Is there a limit for mh_jscale?

Estimation ML/Bayesian estimation
1

Hello everyone,

I am wondering if there is a range for the mh_jscale parameter. Normally, I have to use small values in order to get an acceptance ratio of around 30%. Nevertheless, I am estimating a model in which I have to fix a mh_jscale of 10.
I know that the efficient variance of the MH random walk depends on the model I am estimating but is it not too high? is the MH going to reach the tails of the distribution?
Have I misunderstood some technical concepts or should I worry about some mistake in my model?

(I am considering two blocks and the ratios are close to each other)

Thank you in advance for any comment.

2

That depends on the curvature of the proposal density matrix. As long as you get a decent acceptance rate, you should be fine. That being said, I would always check the mixing via trace_plots later on.

3

Dear Professor jpfeifer,thanks for your reply!

I have checked the trace_plots and the model do not converge at all. I guess this is due to the finded mode. I am using mode_compute = 6 and even with this option I cannot get a good estimation of the modes.

I know that theoretically does not matter from which mode I start the MCMC that it will converge to the actual value. Nevertheless, I have run 2 000 000 replication and it is not even close.

What could be the problem or how could I get better estimation of the mode? (My data is stationary and zero mean)

I would really appreciate your help since I am stuck with this problem.

Thank you in advance.

4

You are not handling parameter dependence correctly. You need to fix this and then check identification.

5

Thank you very much for the correction.

I have modified the mod file and now the mode_check plots are much better. Nevertheless, when I run the metropolis hasting the parameter r_pinf (response to inflation) goes to its lower bound and I do not understand why. Could you please help me?

Thank you very much in advance for any helpful comment

6

What do you mean with

?
You uploaded a trace plot, is that for this parameter? Looking at that plot, there is a second mode. Have you checked whether that one has a higher posterior density? You can do that by looking at the trace plot of the posterior.