 |
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: RE: mlbeta
Nick and Clive,
thank you very much.
best
viktor
Clive Nicholas wrote:
Nick Cox replied to Viktor Slavtchev:
-mlbeta- is, or was, a user-written program by Sean (Jack) Buckley.
As your report implies, his personal website on which
his files sat seems to have disappeared.
I have a version of his program on my machine from an
earlier downloading. 0s and 1s are ignored by the program
as a direct or indirect consequence of its use of maximum
likelihood and the fact that ln 0 is indeterminate.
I won't post the program here for three reasons. It is not
mine to do so; as I know no location for the author, that
would be posting a program with no known means of support;
and another program exists, -betafit- on SSC, that offers
an alternative. -betafit- resembles -mlbeta- in not
supporting 0s or 1s.
When people have 0s and 1s that often means that they have
one or more such spikes in their distributions, so that
it can be moot whether a beta model is applicable in any case.
It's true that you can no longer download -mlbeta- from Jack Buckley's
old university URL into Stata (I lost my copy of it in a machine
move), but if Viktor desperately wants to get hold of -mlbeta-, you
could always contact Jack at his current address at the National (US)
Institute of Education Sciences ([email protected]) and I'm sure
he'd be more than happy to supply a copy to anyone who wanted one
(he's normally quite good with email reponses). However, I might be
able to snag a copy of the full code as helpfully amended, and
displayed, by Richard Williams and Scott Merryman in response to my
request to run -mlbeta- with a -nocons- option. I'll see if I can find
it for you from the Statalist archives, but it might take a while.
However, you might not need it, for there is an alternative that
Viktor can use to fit his proportional models. It doesn't involve a
user-written program and it might turn out to be better. Typing
. glm y x1 x2, family(binomial) link(logit) robust
fits a 'fractional logit' (FL) model as conceived by Papke and
Wooldridge (1996) (obviously, where -y x1 x2- are your variables). An
advantage that FL models have over beta-distributed (BD) models is
that they can include 0s and 1s in the response variable.
In fitting some FL models to some British opinion-poll data recently,
I found that they comfortably outperformed the same models using the
BD parameterization (via -betafit-) in terms of goodness of fit, in
producing the correct signs on _all_ of the coefficients (although the
BD models didn't do too badly here, either) and in producing a greater
efficiency of those coefficients. Perhaps somebody out there has run a
rather more formal, Monte Carlo analysis on the properties of FL
versus BD models that I'm not aware of.
However, Maarten Buis (from whom I picked up this routine as well as
the above paper!) strikes a cautionary note on FL models, which you
can read at
http://www.stata.com/statalist/archive/2007-07/msg00786.html
I hope this helps.
*
* 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/