Author |
Message |
Klemen Korasa (klemen_korasa)
New member Username: klemen_korasa
Post Number: 4 Registered: 2-2013
| Posted on Saturday, April 20, 2013 - 12:23 pm: | |
Thank you Dave, I will review the chapter. So far we have been using mainly PCA analysis of NIR data obtained during blending. I can say that PCA assessment of blending end point is pretty good. We want to use data analysis with MBSD to evaluate and compare homogeneity of different mixures. |
David W. Hopkins (dhopkins)
Senior Member Username: dhopkins
Post Number: 233 Registered: 10-2002
| Posted on Friday, April 19, 2013 - 4:16 pm: | |
Hi Klemen, Your question led me to review the book by Ciurczak & Drennen, Pharmaceutical and Medical Application of Near-Infrared Spectroscopy (Dekker, 2002). I assume that you are interested monitoring blend uniformity, and their chapter 3 discusses using the SD method of monitoring blend uniformity. I recommend you review that chapter, as there are a number of problems using SD of measurements. They note (page 48), "The analyst must remember the assumptions required in analysis of variance, which include normally distributed errors, independence of errors and equality of variance. Non-normal data can give rise to incorrect conclusions regarding homogeneity when using traditional statistical techniques based on standard deviation." They present other mathematical approaches. They also remind us that the size of a sample as measured by NIR or by reference chemical methods on thieved samples should correspond with the unit dose (page 49). I have a question for those who might have more experience with the measurements. Would the use of first or second derivatives of the spectra give rise to better behaved data, not disturbed by baseline problems that might contribute to non-normality? I think there may be a lot of work ahead of you to implement a stable NIR application. Best wishes, Dave |
Klemen Korasa (klemen_korasa)
New member Username: klemen_korasa
Post Number: 3 Registered: 2-2013
| Posted on Thursday, April 18, 2013 - 3:25 pm: | |
Howard, thank you for your answer. You gave me all the information needed. This will help me a lot at NIR data analysis obtained during blending. |
Howard Mark (hlmark)
Senior Member Username: hlmark
Post Number: 537 Registered: 9-2001
| Posted on Thursday, April 18, 2013 - 11:09 am: | |
Klemen - there's a very complicated formula that would calculate what you'd get from the standard deviation of standard deviations, but I'm pretty sure that's not what you should be doing. If you have a bunch of separate standard deviations, then to get value representing the "overall" standard deviation, which is what the a) method is intended to provide, then the proper calculation would be the root-mean-square of the standard deviations (statisticians call this "pooling" the standard deviations). Note that there's an assumption that each standard deviation was itself computed using the same number of data points, othewise you need a more complicated formula: 1) Square all the standard deviations 2) Add up the squares 3) Divide that sum by the number of standard deviations 4) Compute the square root of that result. This calculation is preferred over the method of the a) choice, but the reasons are a bit too complicated to go into here. \o/ /_\ |
Klemen Korasa (klemen_korasa)
New member Username: klemen_korasa
Post Number: 2 Registered: 2-2013
| Posted on Thursday, April 18, 2013 - 10:11 am: | |
Yes. |
Howard Mark (hlmark)
Senior Member Username: hlmark
Post Number: 536 Registered: 9-2001
| Posted on Thursday, April 18, 2013 - 9:46 am: | |
Klemen - in your description of method b), do you mean that you calculate the standard deviation of the standard deviations? \o/ /_\ |
Klemen Korasa (klemen_korasa)
New member Username: klemen_korasa
Post Number: 1 Registered: 2-2013
| Posted on Thursday, April 18, 2013 - 3:14 am: | |
Hello, I have a question about MBSD (moving block standard deviation). So far I have seen two different approaches for MBSD calculation and I would like to know which one is correct. Both approaches start with calculating a SD at a specific wavenumber for selected number of samples (time points). Thus you get an SD value for every wavenumber. In the next step I have seen two possibilities: a) You calculate an average value for each sample (time point) by averaging SDs along wavenumbers: MBSD = sum of SDs for each sample / number of wavenumbers. b) You calculate MBSD with the same values as in a) (SDs), however, with this approach you do not calculate average value of SDs but standard deviation. Which approach is correct? Thank you for your answers, Klemen |