| Author | Message | ||
     E. Perez (epf) New member Username: epf Post Number: 3 Registered: 6-2009 |
Many thanks to you all, your comments are really helpful. For a start, I will try Daniel´s suggestion given that the main software I have access to is WinISI, but I will definitely try Tony´s and Christian´s MATLAB´s CVA analysis. Cheers, Estefanía Pérez | ||
| Christian Hansen (chha) Advanced Member Username: chha Post Number: 25 Registered: 6-2010 |
Hello Perez, A little elaboration on Tonys comment. If you use Matlab you can freely download an easy-to-use algorithm to do CVA on spectral data (more variables than samples) from this adress http://models.kvl.dk/algorithms It is called "Extended Canonical Variates Analysis" (ECVA). It use PLS to deal with the problem (co-linearity) mentioned by Tony. Not on the actual data but on the covariance matrix between classes. The original reference is: N�rgaard, L., Bro, R., Westad, F., & Engelsen, S. B. (2006). A modification of canonical variates analysis to handle highly collinear multivariate data. Journal of Chemometrics, 20(8�10), 425�435. | ||
| Daniel Alomar (dalomar) New member Username: dalomar Post Number: 4 Registered: 2-2009 |
Hi E. P�rez, My experience is restricted to work with a NIRSystems monochromator and WinISI software. In this context, use of a discriminant equation is simpler and straightforward. Plus, you don't need to feed a reference value. Instead, you group samples in different files (A and B) according to constituent. Hope this helps... Daniel | ||
| Tony Davies (td) Moderator Username: td Post Number: 241 Registered: 1-2001 |
Hello Perez. Many people do use regression analysis to do discriminant analysis. It does work but I prefer to use the right tool for the job! The method I have most experience with is Canonical Variates Analysis (CVA). It does have other names. With CVA the model is less constrained except for one important parameter. The number of input variables must be less than the number of samples. You probably have 700 or more variables but not 700 samples! The way round the problem is to use Principal Component Analysis (PCA) to reduce the number of variables; then use the first 20 PCs as the input variables to CVA. Software may be a problem but if you have access to MATLAB I have PCA/CVA available in MATLAB code (written for me by Tom Fearn, who is currently at the Chambersburg conference along with most other lucky people who might have responded to your question!). Best wishes, Tony | ||
| E. Perez (epf) New member Username: epf Post Number: 2 Registered: 6-2009 |
Dear all, I am developing a calibration for one constituent which, in the population of samples under study, can only have two values, say "A" OR "B". Subsequently, I intend to use the equation to predict if unknown samples (of the same nature) are "A" or "B" for that constituent. Initially I have developed a global equation with reasonable R square (>0.70) and predictions, but I have seen that in similar studies authors developed a discriminant equation for this purpose. Could anyone help with this, please? Which would be the best method for calibration development in this case? Should I also develop a discriminant equation? Many thanks in advance! |