| Author | Message | ||
     David Russell (russell) Intermediate Member Username: russell Post Number: 31 Registered: 2-2001 |
I recall hearing Ken Beebe speak on PDS in the early days and noting that in some instances PDS works less well than no correction | ||
| Richard Kramer (kramer) Junior Member Username: kramer Post Number: 8 Registered: 1-2001 |
Bias and slope are tried and true! However, please remember to revalidate adequately. | ||
| Benoit Igne (benoit) Junior Member Username: benoit Post Number: 7 Registered: 11-2006 |
Dear Tony, The case you described is exactly what we have here except that instruments are not too old. And the best standardization methods we have is slope and bias post regression correction. Using PDS, that generate new coefficients, I must add noise when correcting the slave's spectra. But it is always interesting to try. Thank you, Benoit | ||
| Tony Davies (td) Moderator Username: td Post Number: 142 Registered: 1-2001 |
Hello Benoit, There is a saying �If it ain�t broke. It don�t need fixing� or in �proper� English �If something is not broken then you do not have to mend it�. If you do not need standardisation why are you doing it? There should be a statistical test in your method to check that standardisation is an improvement. If there is no improvement then it should not be applied. The answer is �No� I have never seen this but a similar problem was common in the old days of filter instruments and MLR calibrations. These tended to drift so people used a skew/bias adjustment to correct the calibration. This was OK but they needed to be educated into doing the statistical test. I think the cause is �noise� the PDS is adding another set of coefficients. Every time you increase the number of coefficients you pay the price of added noise. If there is nothing to correct they will just add noise and hence your less good results. Best wishes, Tony | ||
| Benoit Igne (benoit) Junior Member Username: benoit Post Number: 6 Registered: 11-2006 |
Hello, I am doing some standardization between two similar instruments (transmittance) and I face a problem. When I do not use any standardization technique and I predict slave samples with my master calibration, I get very good results (similar to what I obtain with the same samples on my master). But when I use PDS to correct slave spectra, I get worse results. My standardization set is good (cover the range, good chemistry, not outliers) and I used a window 1:1, 1:3, 1:5. Have you ever met this case where PDS increase the prediction error? Thank you for your help, Benoit |