| Author | Message | ||
     Bruce H. Campbell (Campclan) |
Report on Discussion Group Meeting at the Eastern Analytical Symposium There were roughly four topics discussed. Transfers, linearity in regression equations, outliers, and progress in the software that is claimed to automatically find the best pretreatment and regression situation. Of the roughly 20 people at the session, almost half had attempted and/or completed calibration transfers. One researcher had transferred a large number of calibrations and stressed the need for careful evaluation of the transfer. Another individual said he had success with a wavelet compression technique. The older slope and bias correction had some success, but also I had the impression it also gave problems, especially when trying to transfer between different types of dispersion instruments. I don't remember anyone saying there was an approach that was universal, although the claim for wavelet compression seemed to imply it was the most powerful or useful. I brought up the question about using linear equations for regression, and why that was so rather than some non-linear approach. Although the spectral intensities are supposed to follow Beer's Law, there are a number of other variables that don't have that assurance. Then there is the non-linear response of some spectrophotometers when the absorbance values become about 1.0. Nobody could answer why there weren't some non-linear regressions. I would think there are some. If anyone knows of them and what degree of success resulted, please comment together with some details of the application(s). Identification of outliers continues to be somewhat of vague. One approach often used, Mahalanobis distance, seems to be powerful. I have some concerns, as there are different types of outliers. For example, statisticians differentiate between X and Y outliers, as they have different effects on regression. Further, statisticians have developed linear regression approaches in which they claim up to 50 % of the points can be outliers and yet the regression still fits. The percent of points that are outliers that have a perturbing effect can be quantified for various regressions (PLS has a value and it is 0%, i.e., one point as an outlier can have a deleterious effect, according to what I have read). Gary Ritchie gave a progress report on the software Purdue has. (This is the software that generates calibrations by varying the pretreatment and the regression equations. One example given was a twelve-hour run that generated thousands and thousands of possible calibrations, with the top ten indicated.) There are efforts being made to obtain a release for sale or licensing, etc. No date was given as to when this may be done, other than probably soon. Jim Reeves asked if anyone had success with the Kubelka-Munk treatment. Two responders said they did, with one of them saying the KM treatment of the transmittance data, and then differentiation did the best of the calibration models tested. No reason was given as to why this should have worked, and why the absorbance wasn't as good when given the KM treatment. The above are my recollections of the discussions. If any other attendee has different recollections or additions, please respond. Or, if anyone has any comment, I would be glad to see them. Bruce Campbell | ||
| David Russell (Russell) |
As a member of the EAS Governing Board, I was gratified at the show of support that we received from those in attendance. However, my show duties did not permit me to get away to any of the NIR sessions or the CNIRS meeting. If anyone has any feedback on the 2001 EAS please reply to [email protected] | ||
| hlmark |
If anyone is able to respond to Dave's request, I'd appreciate if they could send me a copy of their comments also. I was all set to go, but had to cancel my plans at the last minute for personal reasons (my daughter made a grandfather out of me just before the conference started - which is one of the better reasons for having to miss the show, but I'm still sorry I couldn't make it). Thanks in advance to anyone who can help us. Howard Mark [email protected] | ||
| Kathryn Lee |
Many people came up to me and told me how much they liked the NY/NJ Section of the Society for Applied Spectroscopy's Gold Medal Award Session, Honoring Karl Norris. Although I was the chairman, Karl was responsible for inviting the other speakers, so I can take no credit! Steve Delwiche opened with a wonderful history of Karl's years at the US Department of Agriculture, and followed with some of Steve's own new work. Bill Fately was ill, (he is recovering, and apparently going into work now), so Gary Ritchie talked about how Karl has provided him with insight on many of his projects. Dave Hopkins gave a talk on his work on "Convolution and Regression in NIR Spectroscopy." Heinz Siesler presented his work on reaction monitoring. And Karl wowed us all with his latest work on NIR analysis of milk. He is still teaching us new things about NIR! | ||
| Elizabeth Schreiber |
Wow! The conference was great and I learned something from every talk. I was sorry to have missed this group's discussion (didn't see a location listed in the program). Bruce's comment on X and Y outliers is interesting. Statisticians do consider both outliers and influential points (which may or may not also be outliers by Mahalanobis distance)in developing regressions. |