Author |
Message |
Klaas Faber (faber)
Senior Member Username: faber
Post Number: 29 Registered: 9-2003
| Posted on Thursday, March 13, 2008 - 3:01 am: | |
Hi Christian, Why not try one of the following two proposals: H. van der Voet, Comparing the predictive accuracy of models using a simple randomization test. Chemometrics Intell. Lab. Syst. 25 (1994) 313-323 E.V. Thomas, Non-parametric statistical methods for multivariate calibration model selection and comparison, J. Chemometrics, 17 (2003) 653-659 The titles speak more or less for themselves. Both papers have matlab-code. The first proposal is implemented in SAS. Regards, Klaas Faber |
Howard Mark (hlmark)
Senior Member Username: hlmark
Post Number: 104 Registered: 9-2001
| Posted on Tuesday, June 19, 2007 - 11:58 am: | |
Christian - you're getting into some pretty deep statistical waters here. The problem is the last term in your equation, the eij. For different constituents, the eij does not have a constant expected value. You'd need a course in statistics to get a good feel for all the ins and outs of that - or at least a long discussion with a statistician; it's more than can be written down on the discussion group. Or at least read a book on elementary Statistics, to learn why it's not a good idea. Or hire a consultant to explain it. \o/ /_\ |
Christian Mora (cmora)
Junior Member Username: cmora
Post Number: 7 Registered: 2-2007
| Posted on Tuesday, June 19, 2007 - 9:29 am: | |
What if the model is analyzed as a multivariate anova?, i.e. res^2[k=1] ...res^2[k=6] = mu + OBS[i] + SP[j] + OBS*SP[ij] + eij Does it make sense now? CM |
Howard Mark (hlmark)
Senior Member Username: hlmark
Post Number: 102 Registered: 9-2001
| Posted on Monday, June 18, 2007 - 5:38 pm: | |
Christian - no, it doesn't. The short answer for why it doesn't is that in order to combine different estimates, you have to be able to make some assumptions; in this case, that the errors of the different constituents are commensurable. Since (almost by definition) the different constituents will have different units, different ranges and different values of reference lab errors, they are not commensurable, and therefore you will not get any meaningful results from any calculations you do on that data. Howard \o/ /_\ |
Christian Mora (cmora)
Junior Member Username: cmora
Post Number: 6 Registered: 2-2007
| Posted on Monday, June 18, 2007 - 4:43 pm: | |
Dear list members; First of all, thanks for the aswers to my previous post about the bias formula calculation, I think that a good discussion was also generated from it. I have a new question and I would like your comments. I want to compare different calibration models in terms of the error estimates. The structure is as follows: 4 sample preparation methods (SP) x 6 constituents (CONST) x 19 observations in the prediction set (OBS). Using the information on pages 166-168 of the book User-friendly guide to multivariate...I calculated the residuals for each one of the 19 samples in the prediction set and squared them. Then, following the idea of the book: Res^2 = mu + OBS[i] + SP[j] + CONST[k] + SP*CONST[jk] + eijk where, according to the book, OBS would be the sample effect, SP the method effect and I added CONST which is the constituent measured. My question is: can I just extend the model notation presented in the book by, like in this case, adding more term to the model? Second: Does it make sense to include the constituents as factors? again what I have is 4 different sample preparation methods for 6 different constituents and I would like to know which sample prep is (are), across constituents, better in terms of the errors obtained. Any idea will be appreciated Thanks C. Mora |
|