| Author | Message | ||
     Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 63 Registered: 3-2002 |
Hello Nebojsa, I think the method proposed by Tony will lead to overoptimistic performances. It is more for a feasibility study than for the routine use of the models. The validations sets have to be independent of the calibration. The simplest way to set up the sets independently is to sort the spectra by date and make blocks of days, months or years. Pierre | ||
| Nebojsa Todorovic (nebojsa) Junior Member Username: nebojsa Post Number: 8 Registered: 10-2010 |
Thank you Tony, Best wishes Nebojsa | ||
| Tony Davies (td) Moderator Username: td Post Number: 259 Registered: 1-2001 |
Hello Nebojsa, There has been much written about sample selection. Assuming you have one set of n analysed samples: the most common practice is to sort them by the analyte you want to calibrate and then separate them into two sets by numbering them from 1- n and putting the odd numbered samples in the calibration set and the even in the prediction set. One potential problem is with duplicate samples, you most put them in the same set. Pre-treatment is part of the calibration process so you do this on the calibration samples and treat the prediction set in the same way. In most software this will be done automatically. Best wishes, Tony | ||
| Nebojsa Todorovic (nebojsa) Junior Member Username: nebojsa Post Number: 7 Registered: 10-2010 |
Hello all, Recently I met with the problem of selection of samples for calibration and prediction set. Is there a rule for that? I know for SD, mean...is there anything else? I know that one possibility to divide the samples into two sets (cal and pred) is to calculate a PCA and select the samples. However, there is a problem whether PCA calculate before or after the pre-treatment of the spectra? Thank you a lot in advance Nebojsa |