| Author | Message | |||
     Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 81 Registered: 5-2009 |
Hi Dusan, Since I am not familiar with the Piruette, I can't say much about the RV. I think that Unscrambler has something similar, which is called a plot of "Regression Coefficient" which is probably also some type of combined values for all PC's. I have stopped using it long time ago so I don't even remember much about it - and the reason I stopped is primarily because it was useless in terms of understanding relations between the spectra and the regression - which are more evident in the LW. Gabi Levin | |||
| Dusan Kojic (dkojic) Junior Member Username: dkojic Post Number: 10 Registered: 7-2011 |
Dear Gabi, Thank you for your reply, but maybe I wasn't quite accurate in posting my question, since I'm quite familiar with LW and have no problem in understanding them. I am using Pirouette software and, aside from Y-fit and LWs, there's another "metrics", which is a weighted sum of all PCs included in regression model, and it is called Regression Vector (RV). The RV, being derived from PCs, in deed shouldn't be significant in terms of exact numerical values of magnitude and sign of its coefficients. However, I'm confused about correlation with variables (wavelengths). Some people say that large coefficients of RV mean high significance of corresponding variables in the model, while others say the opposite. My understanding (based on concept of weighted average of PCs) is that high values of RV coefficient mean high importance of corresponding variables, but am bewildered by conflicting interpretations I have ran into, while searching the web for answers. BRegards, Dusan | |||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 80 Registered: 5-2009 |
Hi Dusan, I hope this will contribute something. I assume you refer to something we call Loading Weights (LW) for each Principle Component (PC) (Regression Vector?) since we use Unscrambler package. I believe they are same what you refer to. In the attachment I am suing an easy case - a tablet, measured in transmission and the spectra are processed into 2nd derivative. The excipients are same, the weight and size are same, only the dosage in mg is increasing from 1 mg (0.45% by weight) to 10 mg. The active was investigated separately and the transmission spectrum showed a strong peak at 1138 nm. The spectrum of the tablets in 2nd derivative shows clear peak (negative one ofcourse) at 1138 nm. The regression line is quite nice for the small amount of samples used. Only 1 PC was necessary to give us a Cross Validation R value of 0.993. The LW of this PC is also shown and it is clear that the negative "peak" in the LW is due to the 1138 peak of the active. There is, to my mind, clear correlation between the magnitude of the LW peak and the degree of correlation, or R value. As the correlation is more unequivocal - the "larger" the LW peak. Although we use LW frequently when creating regressions as one of the means to validate the specifity of the regression to a given constituent, we don't assign significance to the sign of the LW nor do we assign "quantitative" significance to the magnitude of the LW. In addition, it is important to remember that when a regression requires a larger number of PC's and where the matrix is more complex, the LW will contain information not only from the measured ingredient, but also from other ingredients, since when one is changing, the others also change in %, so there are additional spectral changes that can be correlated to the change in the concentration of the measured species. I hope this helps a little bit. Gabi Levin, Brimrose
| |||
| Dusan Kojic (dkojic) Junior Member Username: dkojic Post Number: 9 Registered: 7-2011 |
First of all, I'd like to wish you all a very Happy New Year ! I have been trying to find some details/explanations on how to interpret regression vectors (PLS or PCR) and have found quite contradictory information. One example of a discussion on this topic can be found here: http://www.namics.nysaes.cornell.edu/news15/list_bvector.html I am aware that one should restrain from direct interpretation of coefficients' magnitude but how about positive/negative values ? Also, are large/small values indicative of high/low correlation or not ? Kind regards, Dusan |
NIR of Tablets _LW