Re: st: Stata and Bayesian capabilities?

[n j cox <n.j.cox@durham.ac.uk>]

We need a sharp distinction here between what StataCorp will (or 
according to some, should) do and what users might do. Well, users
can program what they like and they are not obviously going Bayesian
within the Stata community. Admittedly, it would take a brave and bold 
Stata programmer to start cloning WinBUGS, but I don't detect much
interest even, and if few users care, why should StataCorp?

My own guess is that StataCorp are most unlikely to go Bayesian in any
serious way in any foreseeable future. The lack of interest that is
evident (or, if you prefer, not evident) arises from a mix of academic 
and marketing reasons.

StataCorp developers are more academic than most academics in pursuing 
stuff they find interesting and ignoring stuff they don't, and the good 
folks at StataCorp aren't (detectably) that interested. What the good 
folks a few kilometres [miles] away at Texas A and M are interested in 
is neither here nor there given that. This kind of development just 
cannot be done by talking to, or even commissioning, a few outside experts.

Also, the fact that Bayesian statistics are on a roll within 
mathematical statistics is not that crucial. Some (small scale) 
statistical advances can be implemented in Stata as soon as a paper is 
published. With other (large scale) developments the prudent thing is to 
wait ten or twenty years to see what, if anything, emerges as, long 
term, part of the landscape. Otherwise, StataCorp wastes everybody's 
time (and money) implementing stuff that was bang up-to-date last year, 
but already very dated next year.

I am not clear that Bayesian statistics is that stable. In fact, in some 
 ways it is in retreat, not a story that is prominent, but bear with 
me. A few decades ago all the emphasis was on identifying your 
personalistic (subjective) probabilities, writing down your own loss 
functions, stating your coherent bets, whatever, but where has all that 
gone? Down the drain of discarded dogmas. These days, when people are 
talking about Bayesian statistics, more and more it seems that they are 
talking about a branch of numerical analysis. More importantly, when the 
textbooks start telling more or less the same story, then we will have 
some stability. At present, most of the supposed textbooks look more 
like research monographs to me.

Anyway, for a variety of technical and sociological reasons, I can't see 
Bayesian analyses taking over applied statistical science. One banal but 
crucial detail is that very few applied researchers have the time and 
inclination to get into the rather scary literature. The statistics 
curriculum is too challenging for most of its takers; what happens if it 
is rewritten Bayesian-style?

Alternatively, that's all wrong and my crystal ball is just a bucket of
mud. In that case, why should StataCorp waste everyone's time on adding
modules when it is years behind the alternatives? The nice answer on 
this list is that people like using Stata a lot and don't much want to 
learn other packages. Well, that is understandable, but it is not 
necessarily going to change anything. Some users would like Stata to 
write mailers and word processors on the same grounds!

Nick
n.j.cox@durham.ac.uk

Stas Kolenikov

I would argue it shouldn't, as it involves too many user decisions.
MCMC is a powerful yet fragile technique, and one really needs to know
enough strong theory to apply it properly. WinBUGS is a nice
educational tool, but its Gibbs sampling engine is becoming very
restrictive very quickly once you start looking into any serious
analysis that might involve hierarchical models, improper priors,
non-standard links, stepping between the models, etc. Any serious
Bayesian analysis usually requires doing everything at the low level,
oftentimes as low as C. And programming-wise, Stata would have to have
a capacity of holding two data sets in memory at the same time: the
original data, and the sampled values -- a similar problem is with the
bootstrap estimates that have to be an extraneous data set.

The same story goes conceptually with multiple imputation (which
intrinsically is a Bayesian technique): while there is a
user-contributed module, it is up to the end user to say something
like "Oh yeah, I know it works perfectly well for my application
because [reference to three to five Annals-JASA-Biometrika papers
showing why it works perfectly in this situation], and I also know
that -ice- does it exactly along those guidelines, as [SJ paper]
explains". I understand Stata Corp.'s steering away from it: the users
know better what kind of tools they need, and Stata wants to implement
the tools for which there is clear enough understanding in the
literature with sufficient degree of standardization (shall we make it
another option for -regress-?), and for which a graduate level
statistician can be hired to provide user support :)).

If SAS is having workshop at UNC Biostat... then it must be Joe
Ibrahim working with them -- the set of routines implemented clearly
shows the bias towards the survival models he does. And he certainly
is one of the big names in Bayesian biostatistics. There is a whole
Bayesian department at Duke that SAS does not seem to be (able to?)
use. I don't really know if TAMU has a comparable quality faculty to
use as a local resource.

M.Delprato

> I think it should be a good idea to include it. For the moment being 
I use
> Matlab for Bayesian analysis but I guess with the new matrix langauage of
> Stata, Mata, and good simulations routines, Stata could be expanded 
in that
> direction.

S J

> > SAS is making some moves towards adding Bayesian
> > capabilities:
> >
> > http://support.sas.com/rnd/app/da/bayesproc.html?ETS=6718&PID=299503
> >
> > What about Stata?

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/