🚀 go-pugleaf

In article <1992Dec15.083258.1474@lth.se> andersh@maths.lth.se (Anders Holtsberg) writes:
>
>gary mcclelland mcclella@yertle.colorado.edu:
>>Stepwise regression is a tool of the devil. Anyone smart enough
>>to be using Mathematica ought to be able to decide which
>>questions he or she wants to ask of the data rather than letting
>>a demonstrably suboptimal stepwise algorithm decide which
>>questions to ask.  One day there may be good AI programs for
>>doing statistical analysis, but stepwise regression won't be
>>among them.
>
>You mean what is bad? Let's say we want to make predictions.
>Is the idea of using a subset of the predictors that is bad or
>do you mean the stepwise way to pick them? If the latter: do
>you know any better way (except trying all combinations)?
>

Efroymson's stepwise algorithms are bad for any of
a number of reasons.  First, the statistic labelled
"F" is generally a pseudo-F statistic, whose value
CANNOT be reasonably compared to the central F
distribution.  Second, the method tends to converge
to wrong models, and often the model selected
omits important regressors.  In fact, the algorithms
may never even consider the appropriate model (
whatever THAT is.)

There are better subset regression algorithms
available.  The most efficient known algorithm
is Furnival and Wilson's Leaps and Bounds, implemented
in BMDP 9R and in the new SAS PROC REG (v. 6.07).
Ron Hocking and Lynn LaMotte did a lot of work in
this area during the 70s.  Hocking's review papers
in Biometrics (around '80 or so) and Tech (around '85?)
are the best starting place for self study in this
area.

As a statistical consultant I dread it when a
collaborator comes into my office and says "I've
been running PROC STEPWISE and I've got this
nice regression model. . ."  The only reason
not to run all possible subsets is that it
takes too much resources.  Smart algorithms
cut way down on the resources needed to get
the effect of all possible subsets.  You ought
to use them.

Dennis

---
Dennis L. Clason                          dclason@nmsu.edu
Dept of Experimental Statistics             ESTATX08@NMSUVM1.NMSU.EDU
New Mexico State University
Las Cruces, NM 88001