www.cmm.bristol.ac.uk/forum

As a first step in my analyses, I am running relatively simple logistic regression models using MCMC estimation.

My simplest model has one outcome, N=4249 at level 1, N=232 at level 2, N=99 at level 3.
Even for this model, Nhat keeps increasing and convergence does not seem to be able to be reached especially for RP2 and RP3. Last night I even run my analyses with 1.000.000 iterations (and also a burn-in of 1.000.000), but now Nhat is at 5.000.000 for some outcome variables.
Basically, it keeps increasing (went from the standard settings to larger and larger and larger number of iterations, but Nhat keeps increasing as I run with more iterations).

Similar stuff happens for slightly more complex models (with 3 fixed effects and 2 interactions, same outcome variables).

Is it juts a matter of increasing the burnin/N iterations more and more, or should I do something different?

Hi Gerine,
This is an interesting email and sounds like you may have specific features of your dataset that are causing these convergence difficulties. I'd happily take a look if it helps? A 3-level model isn't what I'd define as 'relatively simple' and it may be that you do not have much if any variability at the higher levels causing poor mixing in the variance parameters. You could look at my MCMC guide later chapters on techniques like parameter expansion and the paper Browne et al. 2009 in JRSS Series A might be helpful.
Best wishes,
Bill.

Hi Bill,

If you could take a look that would be great. Just PB me with your email address.
I will try to fiddle with the parameter expansion a bit as well.
I have to note that I already use orth=1, as this seemed to help in some other models.

My email address is on the CMM website team page.
Best wishes,
Bill.