| Author | Message | ||
     (Unregistered Guest) Unregistered guest |
>>As a rule, the first stage in multivariate modeling using projection methods is to subtract the average from each variable. This "conventional wisdom" is far from universally true, and very often just plain wrong. Mean centering can have several undesireable effects by giving more leverage to some of the noisiest (lowest information content) samples while giving less leverage to some of the best samples in a calibration set. Calibrations developed with mean centering are often significantly less robust than calibrations which use the origin of the native data space as the origin of the regression space. Unfortunately, the subtleties of this issue are seldom discussed in the literature. Malinowski mentions it in his book <Factor Analysis in Chemistry>. | ||
| Suresh Kumar .BV (Unregistered Guest) Unregistered guest |
Dear Liam, As a rule, the first stage in multivariate modeling using projection methods is to subtract the average from each variable. This operation, called mean-centering, ensures that all results will be interpretable in terms of variation around the mean. For all practical purposes center the data is recommended. An alternative to mean-centering is to keep the origin (0-value for all variables) as model center. This is only advisable in the special case of a regression model where you would know in advance that the linear relationship between X and Y is supposed to go through the origin. | ||
| Tony Davies (Td) Moderator Username: Td Post Number: 122 Registered: 1-2001 |
Dear Liam, The correct answer to your question was given by Solomon Abebe in another thread. Click on "Last Week" and you will find it in "I need help". It sounds as if several people need a chemometrics book. Of course I must not advertise here (!) but you will find it on another page of this website. Click bottom left "NIR Publications"; Its the one with four authors! Best wishes, Tony | ||
| David von Boisman (David_von_boisman) New member Username: David_von_boisman Post Number: 2 Registered: 2-2006 |
Hi, "centered" in this sense means a file where you have taken away "bad" spectra based on a GH assesment. ISIScan usually does that in the "develop equations" section - if you set "cutoff by H" to 3. If you set it to a higher value it will not trow away "bad" spectra - which makes sense if you had already cleaned up your cal files using "create a score file from a spectra file". A few samples may or may not be excluded depending on the settings, and that will case small changes to the statistics. However, its important to NOT do both cleaning of files using scores AND further GH cleaning in the calibration procedure as this will also remove important information. hope this helps. | ||
| (Unregistered Guest) Unregistered guest |
Hi David, Thanks for that...I am using a FOSS 6500 - my impression is that the spectra file is "centered" when I first make a score file...but I seem to get slightly differnet results if I go straight to the calibration without making the score file first. So when I use the word Centred, I guess I mean making a score file first - but I can't understand why that should change things. LK | ||
| David von Boisman (David_von_boisman) New member Username: David_von_boisman Post Number: 1 Registered: 2-2006 |
Hi, I have been working with ISI software for 7 years and I belive that "centered" refer to some of the math treatments typically applied to reflectande spectra. Typically you would use SNV/DT (wich includes a "centering" function) and 1,4,4,1 derivatives - though some people use other math treatments. What instrument platform do you use? | ||
| Liam Kaye (Unregistered Guest) Unregistered guest |
I'm using ISI software to develop PLS calibrations and using cross validation. I have been told that I should always perform the calibrations on a file that has been "centered" but I can't find anything that explains why that is. The equations developed on a centered vs an uncentered poulation seem to be a little differnet. Apologies - this is a question born of ignorance but any assistance would be appreciated. The ISI manual is not very helpful. |