| Author | Message | ||
     Malcolm Ray Brown (bro609) Junior Member Username: bro609 Post Number: 9 Registered: 7-2007 |
Thanks Howard and Mark for comments and reference to the previous related post, which I have read. These have been valuable. The take home message that I infer are to proceed with extreme caution, and the absolute need to properly validate any prediction using the (gold standard) lab reference analysis. Malcolm | ||
| Tony Davies (td) Moderator Username: td Post Number: 235 Registered: 1-2001 |
Malcolm (and Howard), The previous discussion is in General, All others, "NIR to calibrate NIR". Pierre's comments are very appropriate in your case. Best wishes, Tony | ||
| Tony Davies (td) Moderator Username: td Post Number: 234 Registered: 1-2001 |
Hello Malcolm, The danger of this approach is that it can become circular but in this case it is worth a try. BUT you will have to test any model on real lab analysis and base any confidence on the SEP. You may well get quite a good SEC. Good luck! Tony | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 336 Registered: 9-2001 |
Malcolm - The topic was discussed on this forum several months ago. Unfortunately, I don't recall the topic heading it was under, but perhaps if you do a search you can find it - my memory is that it was a chain of 6-10 messages in a thread that was mainly about a different topic. Maybe Ian can help find it again. As I recall there was some dispute over it, but I believe it is a BAD IDEA. My belief is based on a computer experiment I did once, where I took real NIR data and assigned random numbers for the "constituent" values, and then was able to predict those random numbers from a different set of NIR data. My conclusion from the experiment was that NIR data could predict other NIR data, regardless of the relationship to actual constituent compositions. Howard \o/ /_\ | ||
| Malcolm Ray Brown (bro609) Junior Member Username: bro609 Post Number: 8 Registered: 7-2007 |
This may be a dumb question, and I’m prepared to be “shot down in flames”…My question relates to the validity of using NIR-predicted data as reference data, for the prediction of associated samples, also scanned by NIR (but in a different manner). It might be easier to understand if I give the specific scenario. From previous work, we have developed a “good” chemometric model (r2 about 0.9 with PLS1) to predict fat in fish FILLETS – using a standard lab wet chemistry analysis for the reference data, and NIR scans from several regions of the flesh side of the FILLET (averaged) as the reference spectra. But in the future we want to estimate fat in large numbers of fish from a breeding program, using either live (anaesthetised) or whole (unfilleted) fish, by scanning through the SKIN surface. I acknowledge the best approach would be to build a new model (based on SKIN-scans and lab wet chemistry for fat reference data) – but would another valid (but less precise) approach be to use fat predictions from NIR scans of the FILLET – and then, by modelling with NIR scans through the SKIN of the same fish, build a model to allow you in future to get estimation solely on scans of the SKIN ? I can see there may be concerns because with the latter approach, both the skin-scans and reference data are based on NIR spectral properties of the sample (so not truly independent measures; albeit scans from different sections of fish) – but I would be appreciative of any feedback… however critical ! Thanks, Malcolm |