Thanks to Kit Baum, the packages -pweibull- and -qweibull-
have been revised on SSC. These programs provide diagnostic plots
for comparison of the distribution of a single
variable with a two-parameter Weibull distribution.
. ssc inst pweibull
. ssc inst qweibull
will install in an up-to-date Stata with internet
access. The option ", replace" will be necessary
if you previously installed these packages.
This follows my recent revision of -pexp- and -qexp-,
announced on 31 October.
In revising -pweibull- and -qweibull- from programs
written in 1998, what was done was
0 To internalise the parameter estimation.
(More on this below: a detail of interest
way beyond this problem is the usefulness
of -nlcom- for reparameterisation problems.)
1 To revise the graphics to Stata 8.
2 To support fweights and iweights.
3 To allow generation of variables to hold
what was plotted.
4 To allow fitting of the whole distribution,
but a graph showing only part of the data, e.g.
one tail. (This last arose out of a suggestion
on a different problem by Stephen Jenkins.)
In addition, these programs allow specification of
parameter values (scale and shape parameters) as well
as fitting the distribution on the fly.
As implied, -pweibull- and -qweibull- now require Stata 8.
The original versions, now renamed -pweibull4- and
-qweibull4-, remain in the package for any interested
users of Stata 4 through 7. However, they are
otherwise frozen as was and were not revised
to incorporate 0 to 4 above.
Nick
[email protected]
note on parameter estimation for Weibull distribution
=====================================================
To plot against what we would expect
if the distribution were Weibull, we need parameter
values, usually estimated from the data.
The Stata solution that has been possible
for some time is to set up the problem
as one for the official Stata command -weibull-,
which will yield maximum likelihood estimates.
Note that no covariates are involved in this problem,
nor are any required by -weibull-. A small difficulty
then often ensuing is that the parameterisation
used by -weibull- is chosen, it seems, for its suitability
for survival modelling. At least one other
parameterisation is very common in the literature,
even if, cosmetically, there is little agreement on the
notation to be used. To convert from -weibull-'s
parameterisation to this other one, a small wrapper
program -wbull- is available on SSC, and it remains
a solution for users of Stata 5 through 7.
As of Stata 8,
* -weibull- has gone undocumented, and users
are directed instead to -streg-. This strikes
me as a small step backwards, as you can be
interested in Weibull distributions without
(consciously) engaging in survival (or reliability) analysis.
To use -streg- you have to set up your problem as a survival
problem even if it isn't. This is a very minor matter,
however, while -weibull- remains accessible.
* -nlcom- offers a much better and more
general way of reparameterising the problem,
for which many thanks.
Anyway, -pweibull- and -qweibull- for Stata 8 now
internalise parameter estimation. You
don't need to estimate parameters separately.
Conversely, you could even use -pweibull-
or -qweibull- just as a way of estimating Weibull
parameters, and ignore the graph, although naturally
I wouldn't advise doing that.
*
* 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/
