Jim - my old but still usable "tryem" program (in the ssc) finds the set of
k variables that gives the highest R^2 where k is specified by the user. In
general it is hard to justify using this procedure for models with random
errors, but if you are just trying to approximate a function with as few
"wiggles" as possible, this may be worth doing. For a more detailed
justification see my 1994 paper
"Finding the Best Regression Subset by Reduction in Non-Full-Rank Cases",
SIAM (Society for Industrial and Applied Mathematics) Journal of Matrix
Analysis and Applications, Vol. 15, No. 1, 194-204.
Al Feiveson
-----Original Message-----
From: [email protected] [mailto:[email protected]]
Sent: Friday, September 19, 2003 12:07 PM
To: [email protected]
Subject: Re: st: RE: All-possible-regressions procedure
Nick,
Thanks for your response. I'm looking for something analogous to the SAS
command (I forget what it is exactly), which selects the "best" # (you
specify the #) of models using 1 covariate, 2 covariates, etc. The
investigator then explores the resulting models. It's not really an
automatic procedure in the sense of forward or backward selection. Is that
what -allpossible- does?
I know it's not feasible to give me a short course in multiple regression,
but what is your basic philosophy when whittling down potential explanatory
variables when doing an explanatory model (as opposed to a predictive
model)?
Thanks,
Jim
"Nick Cox"
<[email protected]> To:
<[email protected]>
Sent by: cc:
[email protected] Subject: st: RE:
All-possible-regressions procedure
arvard.edu
09/19/2003 09:43 AM
Please respond to
statalist
[email protected]
> Does Stata have an all-possible-regressions procedure, for
> use in model
> building in multiple regression?
Sort of.
I wrote a program -allpossible-
which does this to a limited extent.
It's really a wrapper that runs
lots of regressions or similar
and prints out selected results, but
it has strict limits.
I also wrote a program -selectvars-
which is, indirectly, one of the
tools you need if you want to knit
your own and cycle through subsets
of predictors.
Both are on SSC.
However, I have not written anything
to select "best" regressions in any
sense whatsoever, not only because
that is much more difficult to program,
but also because it's against my religion.
(I am not aware anyone else did either.)
In fact, before someone starts thinking
"gun manufacturer" I will stress that my
own motivation for writing -allpossible-
was the complete opposite of what I take
to be a common motivation for using
such programs. I wanted to show that
a variety of models were, in a class of
problems, almost equally good in terms
of various criteria. The naive opposite
is, naturally, the idea that one can
automate the selection process completely.
I'll add a standard rider. No such program
can cope adequately with the combinatorial
explosion(*) of possibilities. Even
trimming matters down to whether each
predictor is in or out of the model,
then 20 predictors give you 2^20 ~ 10^6
models, and it's pretty hard to keep
track of a million sets of model results.
(20 predictors is perhaps a small model
by many analysts' standards). Even 2^10
is more than I want to compare.
Nick
[email protected]
(*) A nice term. When was it introduced?
*
* 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/
*
* 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/
