Re: stepwise regression by GENSTAT

On Sun, 17 Aug 2008 06:33:29 -0700 (PDT), hubert.ruypers@xxxxxxxxx
wrote:

On 17 aug, 01:47, RichUlrich <rich.ulr...@xxxxxxxxxxx> wrote:
On Sat, 16 Aug 2008 07:40:57 -0700 (PDT), Old Mac User

<chendrixst...@xxxxxxxxx> wrote:
On Aug 16, 3:20 am, hubert.ruyp...@xxxxxxxxx wrote:
I am using GENSTAT for performing a multiple linear regression. To
select the most suitable explanatory variables this programm provides
a "stepwise regression method". However I do not understand how this
method leads to a result.

According to my handbook it drops the variable which gives the lowest
mean square (Residual) in an ANOVA set-up.

What it drops is the variable that makes the least improvement
in the mean square (Residual).   That is the variable that has the
least effect, at that point.  If your handbook says otherwise, you
need a new handbook.


Is "the least improvement in the mean square (Residual)" not the same
as "the variable that results in the highest value of the mean square
(Residual)?

Do you notice that what you wrote this time is "highest", where
before you wrote "lowest"?



If your handbook makes a general recommendation, that stepwise
methods are desirable, you need a new handbook -- there was a
spell in the 1970s, when computers were first available to them,
that some practitioners generated a fad for Stepwise; but wiser
statisticians (and their own experiences) shortly convinced them
that there was no magic in it.  You can Google groups
< stepwise group:sci.stat.*  author:ulrich > for comments and
threads on the subject.

My handbook considers only stepwise regression as a method to select
the explanatory variables (Mc Conway, Jones and Taylor, Statistical
Modelling using Genstat").

What's the date on that handbook?

There are specific cases where stepwise is okay, usually when
ample crossvalidation is available, or where the only interest
is in having a smaller model (and you have no doubts about the
components). Data-mining and economical prediction.
The handbook could mention that.

If it doesn't caution against using Stepwise generally, it is out
of date, you should not depend on reference that is 30 years
out of date.

I occasionally review for a research journal whose "statistical
guidelines" include, effectively, a ban on stepwise. Reading
the "ban" bothers me a little, because I wonder if other journals
may adopt the same guidelines without enough thought,
but the ban is okay for the papers it publishes.




And here I am losing the logic. I was convinced that a lower mean
square (Residual) resulted in a higher variance ratio (which is the
m.s.(Regression) divided by the m.s.(Residual) and thus in a lower F-
probability.

Can some-one help me out ?

Hubert

I respectfully submit that if you do not understand multiple
regression
and it's related offshoots, you should get some assistance from a
practicing statistician.  Building models from data has many facets,
it can be tricky, and the risk of getting silly results is high.  OMU

Right.

Who am I to contest this. However some knowledge of the statistical
possibility's (and limitations) is never wrong (I think)

As I suggested before -- Google "groups", using
< stepwise group:sci.stat.*  author:ulrich >
for comments and threads on the subject.

--
Rich Ulrich
.

Relevant Pages

  • Re: stepwise regression by GENSTAT
    ... mean square in an ANOVA set-up. ... need a new handbook. ... statisticians shortly convinced them ... My handbook considers only stepwise regression as a method to select ...
    (sci.stat.math)
  • Re: stepwise regression by GENSTAT
    ... a "stepwise regression method". ... According to my handbook it drops the variable which gives the lowest ... mean square in an ANOVA set-up. ... If your handbook makes a general recommendation, that stepwise ...
    (sci.stat.math)
  • Re: stepwise regression by GENSTAT
    ... a "stepwise regression method". ... mean square in an ANOVA set-up. ... random variation in the residuals. ... in statistics by the Open University. ...
    (sci.stat.math)