I have been asked to forward the following query to the list; please reply
to Miles Cox directly.
Emmanuel Pothos
-----------------------------------------------------------------------------------------------------------------
Dear Allstaters,
I need to use multiple regression analysis to establish the best set of
predictors (from 12) of a dependent variable. That is, I would choose the
set of predictors that yield the highest F ratio in the regression model,
by successively eliminating predictors with low standardised beta
coefficents that seem not to contribute to the model. Because this
approach would clearly involve exploratory model building (rather than
hypothesis testing), I need to test the model's generalisability.
I believe that there is a technique called the bootstrap method, which
would allow me to test generalisability from a single sample. For
example, I would divide the sample into four or five parts, build a model
on each part, and then test it on the other three or four parts taken
together. For example, I might build a model based on 20% of the
subjects, test it on the other 80%, build another model on a different 20%
and test it on the other 80%, etc. At the end, I would pool all the
tests, in effect reconstituting the full sample, and yet not having ever
tested on the same subjects on which I explored for the model(s).
However, I am not sure how best to go about splitting the sample into the
parts, nor how to combine the results in the end. Can anyone provide a
reference on this topic, or any other information about how to do it?
Thank you very much.
-----------------------------------------------------------------------------
W. Miles Cox, Ph.D., Office: +44-1248-383774 | /\
Professor of Psychology, Messages: +44-1248-382201 | /
\/\
School of Psychology, Facsimile: +44-1248-382599 | /\/ \
\
University of Wales, Internet: | /
======\=\
Bangor LL57 2AS, [log in to unmask] | B A N G
O R
United Kingdom. http://www.psychology.bangor.ac.uk
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|