| Author | Message | ||
     Marion Cuny (marion) New member Username: marion Post Number: 4 Registered: 6-2009 |
Hi Alisha, I was going to advise you as Mark to do an experimental design and in your case a mixture design. what you need to know before starting is what type of model you are expecting: *linear *with interaction *with square terms This will determine the type of mixture design you need: * axial design (linear) * lattice design (linear and interaction) with a lattice degree that you can adapt to get more samples * centroid design (linear+interaction+square terms) There is always a need to do a center sample in your case 25/25/25/25 and replicate it (3 times in general) to assess the experimental error. If you suspect any noise (there are always noises in spectroscopy) you can duplicate your design to get a better coverage of the noise. Basically the more the model is complicated and the more noise/interference you suspect the more samples you will need. To suggest you a good book for Mixture design analysis I will give you the reference of J. A. Cornell Experiments with mixtures - Designs, Models and the analysis of mixture data. Marion | ||
| Atul Karande (atul) New member Username: atul Post Number: 2 Registered: 4-2008 |
Dear all.. Recently we have published one paper in IJP, which addresses the same concern raised by Alisha for building the multicomponent calibration model.. @alisha may it would be of help to you particularly for adding randomness to calibration samples and to have relevant and representative calibration samples.. Here are the details of the paper Liew CV, Karande AD, Heng PW. In-line quantification of drug and excipients in cohesive powder blends by near infrared spectroscopy. Int J Pharm. 2010 Feb 15; 386(1-2):138-148. Epub 2009 Nov 13. Cheers!!! Atul | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 303 Registered: 9-2001 |
Alisha - I agree with both Fred and Gabi. Their advice can be made more specific, however, depending on the control that you have over your samples. What Fred and Gabi said is good when you have no control over your samples, and must select samples more-or-less at random, from your process or other source of samples. Then, truly, "the more the better". But from the way you worded your message, I get an impression you can make samples to whatever specifications you need. In this case, you can use your control over the sample-making procedure to ensure that all the requirements that we like to see in a sample set are met. If you can, in fact, make up samples to order, then you can apply the principles of what are known as "statistical experimental design". This can give you the maximum "bang" for the minimum number of samples. For example, if you have four different components and they can all be controlled independently, then you can use a design where each component is included at the maximum level and at the minimum level, in all possible combinations. These are called "factorial designs". With four different components that would mean 2 ^ 4 = 16 samples. There are experimental designs that can achieve the same goal with fewer samples. These are called "fractional factorial designs". If, as Gabi said, these four components make up 100% of the sample so that increasing some means that others have to decrease, then there are other designs, called "mixture designs" that can specify the samples to be made, that include the restriction. It would probably be a good idea, with any design, to include more samples than the bare minimum. There are ways to optimize the extra samples to include. Before starting to use any of these experimental designs, it would probably be a good idea to read a good book on statistics, with emphasis on the chapters dealing with experimental designs. Howard Mark \o/ /_\ | ||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 29 Registered: 5-2009 |
Hi Alisha, I tend to agree with Fred, but want to make another point - when building a matrix of samples it is very important to verify that there will not be an interdependnce is such way that when one component goes up, one of the other components always goes down - the increase in one component must be "matched" by decrease in other components in a random way - otherwise you will not get reliable regressions that will hold in real life situations. If you know what small minute components may exist in the real product, and their range of variations it may be useful to add some of them in such way as to include these "disturbances" into the model already. Gabi | ||
| W. Fred McClure (mcclure) Senior Member Username: mcclure Post Number: 49 Registered: 2-2001 |
Alisha, Sorry, I cut my response too short. If you had a "perfect" mixutre for each sample you probably could get by with fewer samples. Remember that there are more variables than the components (moisture, temperature, time, packing density, cup-cup variations, instrument noise, etc. These are the reasons for having as many samples as practical (possible). Best with the test, Fred. | ||
| W. Fred McClure (mcclure) Senior Member Username: mcclure Post Number: 48 Registered: 2-2001 |
Alisha - The rule of Thumb is "the more samples the better." However, statistics experts tell us that there should be a minimum least 30 for any type of analysis. I don't think you will get a sound and reliable calibration with 4 samples - Fred | ||
| Alisha (agnosus) New member Username: agnosus Post Number: 2 Registered: 1-2009 |
Hello, I want to build a NIR calibration model for a simple four-component mixture. I know the range of each component in future samples. Now what is the minimum number of samples required build a model? Is it correct to say that for each independent source of variation we must have at least one independent sample, so in this case we are going to need at least 4 samples with uncorrelated levels of these components? Will appreciate your inputs. Cheers, Alisha |