Bayes in Stata: how many MCMC chains?

[Guest]

Hi

Reading help files and also when looking at Bayesian trace plots provided by Stata, it appears that Stata only uses one MCMC chain. Is this a correct understanding of Stata's Baysian analysis? We should use two ore more chains (and see how they overlap/converge), so the next question would be: if Stata uses only one chain, can this be overruled by asking for two chains?

[Guest]

For anyone interested, I asked Stata Technical Support for advice, this is the answer I received:

Yes, Stata uses only one chain in -bayesmh-, and there is currently no
option to specify otherwise.

[Joseph Coveney]

Couldn't you run the chains in tandem, saving the results (discarding burn-in and adaptation samples) to separate files, and then append or merge the files for plotting the traces in parallel, computing Gelman-Rubin diagnostics or whatever?

Somewhere (where will depend upon whether you're appending or merging), you'll need to do something like the following (details of the code will also depend upon whether you're appending or merging)

quietly expand _frequency
sort _index
quietly replace _index = _n

in order to get the chains to align.

[Guest]

Indeed, Joseph, you can improvise to have two chains, even though Stata currently does not offer that as part of the standard input (or as the default, which I think Stata should have, with the option to run more than two chains). Apart from telling me that my concern would be forwarded to the developers, the answer from Stata Technical Support also included this:

---------

In the meantime, you can produce trace plots for two chains by doing
something like the following:

    webuse oxygen
    bayesmh change age group, likelihood(normal({var})) prior({change:}, flat) prior({var}, jeffreys) saving(bayes1)
    bayesmh change age group, likelihood(normal({var})) prior({change:}, flat) prior({var}, jeffreys) saving(bayes2)
    clear
    append using bayes1 bayes2, gen(chain)
    describe eq*, varl
    xtset chain _index
    twoway tsline eq1_p1 if chain==1 || tsline eq1_p1 if chain==2

--------

Your suggestion for comparing the chains is interesting.

[Joseph Coveney]

If you're doing just trace (line) plots, then you won't need to expand. The line plot continues level until the next new y-value, and the change will be made at the corresponding value of x (value of _index).