| |
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
RE: st: Are you a Bayesians?
I didn't notice Nick's response to this (sorry for the lengthy delay).
I agree generally with his analysis. The Bayesian take on things is
a real change in paradigm for most researchers, some of whom find
existing frequentist techniques challenging enough as they are. A
new learning curve is a barrier for many. On the other hand, ideas
from the Bayesian side of things have a lot to offer. The world
isn't necessarily universally normally distributed as many models
assume (although recent frequentist approaches, such as XTGEE, allow
different distributions for the dependent variable), leaving aside
the need for priors. I agree that the gauntlet lies before those who
straddle both worlds to some extent and who can come up with
applications influenced by the Bayesian paradigm. I hope to be among
those, eventually. The -bic- and -bicdrop1- commands (for example)
from ssc are example of using Bayesian concepts in a frequentist way
that is easy and powerful for use in conjunction with conventional
models. I agree that Stata's role is to implement accepted
techniques, not invent new ones or to try to implement rarely used
esoteric algorithms.
- Paul
At 05:53 AM 08/03/2006, you wrote:
One main thrust of Bayesian work seems to have been the insistence
that each problem ideally requires its own model, which commands fairly
easy assent as an abstract principle. This would be made easier by MCMC
engines, etc. but the extent to which it can be automated is
questionable. I don't think this is mainly because analysts need to reach
into their subconscious to pull out prior distributions: it is mainly
because of the need to customise a model according to the
structure of the scientific or substantive problem.
So, ultimately each researcher needs to write their own
"program", which I put in quotation marks because
in Stata that need not necessarily mean a program
in Stata's sense. That's why WinBugs and R are languages widely
used for this: WinBugs is designed for the purpose,
and R is designed mainly for statistical people willing to write
their own programs. Or so I perceive.
I have to guess that most researchers using statistics are
most unlikely to want to write their own program. Also,
the prevailing mindset, as shown by many, many posts on this
list, is that there is a "correct" analysis that can be
obtained by plugging your data into a pre-existing program.
Just tell me what it is, please!
While Bayesian stuff seems to be growing on an exponential,
I predict that exponential will turn into a sigmoid,
given the likely mass unwillingness of people applying
statistics to adopt it. The intellectual arguments are
secondary here.
That all said, the crunch really is this.
1. There is no detectable interest on the part of StataCorp in
providing the tools. If this is true, they probably won't say so,
or say much more than that there is no present intent to do
Bayesian stuff in a major way. StataCorp prefer positive
statements to zero or negative ones.
2. Regardless of 1, StataCorp do not like doing token efforts
or playing with something. (When they have done something
that ended up as a token effort, usually by accident, they
have regretted it bitterly.) So if StataCorp go Bayesian, they
will go Bayesian in a massive way, and that's a long-term project.
3. Regardless of 1 and 2, user-programmers could do a lot more
than they have done already, but there is very little interest.
My main guess here is, as above and as mentioned previously
on this thread, interested people just do it in other
software.
4. It is always nice when people say, "Well, I use X for Y, but
I would rather use Stata." However, when other programs are years
ahead of Stata, it is not clear why Stata should play catch-up.
I don't write as anti-Bayes or non-Bayes. I have, in a minor way,
implemented Bayesian ideas in Stata for one problem. (See -cij-
on SSC, most of which was adopted in official -ci-.) I have seen work
in which the frequentist answer was a heap of garbage and the Bayesian
solution neat and elegant and scientifically much more acceptable.
I have also seen Bayesian projects that seemed to take up many, many times
more effort than a frequentist solution that got most of the way.
Nick
[email protected]
Paul Millar
> In the past I have used R or Winbugs for Bayesian problems. I agree
> Stata could be better equipped for this approach. In fact, I don't
> think Bayesian approaches will, despite their power compared to
> frequentist techniques, get into the mainstream until people develop
> routines for packages like Stata that make it easy for the researcher
> to take advantage of.
>
> >If you are a Bayesian using stata, please respond with raised voice.
> >
> >Most of my work is frequentist in nature, but I apply Bayesian
> >techniques for some of my more onerous problems. As was mentioned in
> >the fall, "Stata is not much of a vehicle for doing Bayesian
> >things." Should this change?
> >
> >The paucity of interest in Bayesian techniques, or its appearance,
> >may represent an area of development for stata. Bayesians, if you
> >are out there, I personally would like to how you manage. Maybe
> >stata and its users will develop greater tools if we can show that
> >there is a market.
*
* 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/
*
* 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/