| Author |
 Topic  |
|
|
dg773
4 Posts |
 Posted - 11 Jan 2008 : 10:51:17
|
Hi -
The documentation explains how MulReg can run into problems when many independent variables are collinear, or close to it, and how SplitMulReg avoids some of those problems. I'm trying to decide between the classes and wonder a.) just how close to perfect correlation you have to get in MulReg before the equations become "ill conditioned" (as the documentation ominously puts it  , and b.) Conversely, under what conditions (if any) SplitMulReg might perform worse predictively than MulReg.
Also, can you point me to the academic literature behind the SplitMulReg algorithm?
Many thanks in advance!
David |
|
|
quinncurtis
1164 Posts |
Posted - 11 Jan 2008 : 11:35:21
|
Sorry, but we do not have any exact answers for you. Our manual merely repeats the phrases used in the literature to describe the performance of the general algorithms we use. In general it would relate to the precision of the numeric type on the host machine (8-byte reals in this case) and the size of the source data arrays.
There is a selected bibliograpy at the end of the QCMatpack manual: http://www.quinn-curtis.com/QCMatPackNetManual.pdf. A couple of the referenced books are:
Afifi, A. A. and Clark, Virginia, Computer Aided Multivariate Analysis, Lifetime Learning Publications, 1984.
Kleinbaum, David G. et al., Applied Regression Analysis and Other Multivariable Methods, PWS-Kent Publishing Company, 1988. |
 |
|
| |
Topic |
|
|
|
SplitMulReg vs. MulReg
