Author |
Message |
Lee Streeter (lees)
New member Username: lees
Post Number: 3 Registered: 6-2008
| Posted on Tuesday, June 24, 2008 - 8:41 pm: | |
Hello Howard Thank you, I'll be sure to try the other list. Lee |
Howard Mark (hlmark)
Senior Member Username: hlmark
Post Number: 198 Registered: 9-2001
| Posted on Tuesday, June 24, 2008 - 7:17 pm: | |
Lee - it occurs to me that this may not be the best place to ask this question. You might be better off putting it on the discussion group of the International Chemometrics Society. Try sending a message to [email protected] <[email protected]> Probably you'll get a response saying that you're not registered on the list, but I think the response will also tell you how to register. Once you've registered, you can ask your question. \o/ /_\ |
Lee Streeter (lees)
New member Username: lees
Post Number: 2 Registered: 6-2008
| Posted on Tuesday, June 24, 2008 - 3:11 pm: | |
I have considered the issue as to whether it's enough to just obtain coefficients that solve the problem at hand. The problem I have with that line of thinking is if new independent data is obtained then if there is new confounding variation, then "good enough" coefficients can quickly become invalid. "True" coefficients (assuming linearity) will respond correctly with the important information to which will be added the response of the new confounding information. Of course if linearity does not hold then non of what I say is true, but that is beyond the scope of what I am trying to ask. |
Gustavo Figueira de Paula (gustavo)
New member Username: gustavo
Post Number: 5 Registered: 6-2008
| Posted on Tuesday, June 24, 2008 - 6:09 am: | |
Hi Lee, I�ll extend your question asking in a subtle different way: When using PLS, what matters are the "true" coefficients or the "enough" ones (to classify correctly)? Lee, think on that. |
Lee Streeter (lees)
New member Username: lees
Post Number: 1 Registered: 6-2008
| Posted on Monday, June 23, 2008 - 5:18 pm: | |
My apologies if this topic has been hashed out in the past, but I'm a bit unclear on the following. I understand how PLS computes weight vectors and corresponding scores to model both the X and y block data. From what I understand PLS then computes a coefficient vector as a linear combination of the weight vectors. While I understand that weights define a subspace where the important X and y information resides (and hopefully the random noise doesn't), it's not clear to me why it is reasonable to assume that the "true" coefficients reside in this subspace. Can anyone help? |