25 sept. 2013

1-VR , R squared or coefficient of Determination

One of the most important statistics when we develop a model, to check the performance is the, R square (RSQ) or Coefficient of determination. This video (from the Kahn Academy), describe its calculation and reminds me the statistic 1 - VR from Win ISI.
The statistic called 1-VR (one minus variance ratio)  describes the variance explaind during the cross validation process. 
Beeing:
VR=(SECV*SECV) / (SD*SD)
where:

SECV = Standard Error Cross Validation
SD = Standard Deviation of Reference Values

So
1-VR = 1 - ((SECV*SECV) / (SD*SD))



17 sept. 2013

Minimizing Squared Error to Regression Line (Kahn Academy)

In all this series of videos from the Kahn Academy, we can see  how complex is the calculation of the slope and intercept, compared with the way we are use to do it with a scientific calculator or using the formulas from Excel or any other program. Anyway the first video is really interesting and sure you want to continue to the end to see the results of all these calculations.
 
 
 
 
 

13 sept. 2013

Sort and Select samples in Win ISI

Sometimes it is necessary to sort our sample sets for a better selection of the Calibration and
Validation Sets, so they will be almost similar distributed. Let´s see how to do it in Win ISI.
1)We are going to sort this CAL file by Protein ascending constituent value.
2) Now we are going to select every 5th sample for a validation set and the rest for a calibration set

9 sept. 2013

The importance of well trained operators for NIR success

(Imagine) We have develop a good calibration taking care of a lot of issues, like the performance of the instrument: precision and accuracy, the variability of the samples and source of variance of our calibration set, the instrument drifts using a check sample, the lab conditions: temperature, humidity,...., the way to present the sample to the instrument: grounded, well homogenized, diluted, clean without impurities, chopped, dry,....

Now this model is in routine and different operators are going to prepare and scan the samples in routine. According to the spectra and results, certain samples will be send to the laboratory for reference analysis in order to increase our database and to improve the calibration. At this point we must be careful, and we have to train the operators in order to present the sample as better as possible.

We have to explain them the importance of a good sampling, how to grind and homogenize the sample correctly, how to pack it, etc. This includes the importance to clean dust from the instrument, to run the diagnostics and check cell periodically, the cleaning of the cups,....

It occurs very often that because of very busy operators, lack of personnel, carelessness, boredom, and other things the results, correlation of the spectra with the constituents, ...., etc is quite poor and we won´t get the expected results from the NIR, thinking that the problem is the model. So we are going to add all these variables to aour calibration set, and the results will decrease in value.

So take all this into account and keep the operators as well trained as possible and tell him the importance of their work in the success of the NIR.

1 sept. 2013

Shootout 2002 files

If somebody is interested to play with the 2002 shootout, the files can be still downloaded from this link:
http://www.idrc-chambersburg.org/ss20022012.html.

I will write some post about these files, so you can follow better the posts. There are different literature about this shootout 2002, as this one from Eigenvector:
Development of an NIR Method for Determining the Active Ingredient in Pharmaceutical Tablets

NSAS Files are imported into  Win ISI or Vision.
JCAMP can be imported into other Chemometric software´s.

If you want to dowloaded in a Matlab format, visit:
http://www.eigenvector.com/data/tablets/

The post: Calibrating and validating with VISION (a quick view) has a video showing the files loaded in Vision.