Hi:
Frank Harrell identified
several problems with using stepwise regression to select variables.
See section
4.3 of:
Regression
Modeling Strategies With Applications to Linear Models, Logistic Regression,
and Survival Analysis
Series: Springer
Series in Statistics
Harrell,
Frank E. Jr.
1st ed. 2001.
Corr. 2nd printing, 2006, XXIII, 568 p. 141 illus., Hardcover
ISBN:
978-0-387-95232-1
HTH,
Shawn Boles, Ph.D.
Senior Research Associate
Oregon Research Institute
1715 Franklin Blvd.
Eugene,
Oregon 97403-1983
USA
Phone (541) 484-2123 ext 2225
Fax: (541) 484-1108
From:
s-news-owner@lists.biostat.wustl.edu
[mailto:s-news-owner@lists.biostat.wustl.edu] On
Behalf Of duo wan
Sent: Wednesday, November 28, 2007
9:56 AM
To: s-news@lists.biostat.wustl.edu
Subject: [S] use of stepwise cox
regression in comparing relative "importance" of variables
I am reading some clinical literatures for our research
paper. They adopted stepwise algorithm when building a final prognostic models.
However, they used this algorithm for:
1) Comparing
the relative importance of two highly correlated variables (different
indicators for the same thing). Say the one kept in the model is more
important.
2) Claiming
that the variables with higher chi-squares are more important.
I really do not think two highly correlated variables can be compared
with stepwise regression. First, usually one of two highly correlated variables
has to go with stepwise regression. It is hard to tell the magnitude of the
differences between variables from stepwise. Also the left one can be more
important that some other variables kept in model in terms of higher
proportions of explained variations.
Is there any good statistical paper about misusing of stepwise in
ranking the order of significant of variables or relative importance of
variables.
Be a better sports nut! Let your teams follow you with Yahoo Mobile. Try
it now.