An important point has to be considered performing a calibration of an NIR-system. The major method error of the NIR-method is due to the error of the reference method, because the accuracy of the NIR-method is the sum of the reference error and the error of the NIR system. While the repeatibility (repeated measurements on the same sample with same device and same people) of the NIR-system is realy small the main error source is most often the reference method. And if you would like to compare the accuracies you should not compare the repeatability of the reference method with the SEP of the NIR system. The better comparisson is to compare reproducability (same sample - in case of destroying measurement nearly same sample, different devices from same type, different location and different people) with NIR-accuracy because this is also a different device. Additional it is very important for the accuracy of the NIR-method not to mix the reference methods, e.g. not to mix drying oven and moisture analyzer.

Recommendation: Check reproducability of the reference methods and take into account this error number for the evaluation of the NIR-method.

To check the lab performance you could do the following:

You have to check the precision (repeatability) and reproducability of the reference method.

1) A sample has been measured with NIR take and divide into 5 parts and divide each resulting sample again into 2 equal parts.

2) performe reference method on a set of 5 in house or your standard reference laboratory (important vary the name of the 5 samples randomly) so that no inference possible to the origin of each 5 samples) => Repeatability

3) send the second 5 pieces to a 2nd refernce laboratory . Compare the average value mean value obtained in 2) or compare the error sum of individual differences => Reproducability

4) Compare Repeatability and Reproducability with the NIR error (SEP). Presumption NIR difference to the average values ​​of the two 5-measurement is smaller (maybe the same) as the standard deviation of the two reference multiple determinations. => NIR is a method that works within the error limits valid reference.

As Grimm said, split your original sample into two or three lots and send it to the different labs for chemical analysis. I also prefer repeated measurement of a sample in a lab to get a better result. Then compare result from all labs to check lab error. I prefer using median value of all the lab rather than mean as a reference value to use in NIR calibration.

Follow ISO 12099

Dear Yi,

Let us start with unlike chemical analysis methods, NIR measurements are "Matrix" dependent. i.e, you can not ignore other components in the soil such as moisture content, etc. and therefore, you must make sure what are all the components first. If you are correlating carbon or nitrogen content of soil to chemical analysis data make sure that you remove the mositure first and then scan your sample and use the same sample for chemical analysis and obtain the percentages based on that dry weight. Once you obtain the chemical analysis results then you could create your Quantitative analysis model by correlating the chemical analysis data to your NIR spectral data e.g., nitrogen content or carbon content. However, if the contents are in the part per milion level I think you not going to be very successful but if the values are more like percent level that by all means continue. Make sure your spectral data are very consistent because light scattering could cause probelm. The physical effect of soil must be minimized, i.e., if you run your samples through a sieve collect the same mesh size for all your samples. Thisway the physical size of sample is uniform. When you are satisfied with your measurements to compare two different methods we have used Bland and Altman comparison where SD for two methods are compared to each other (see my publication and you could download the paper in th Obesity 03/2011; 19(7):1497-502. If you are still intereted in comparing the chemcial analysis errors, I would suggest first determine what repeatability and reproducibilty in the chemcial analysis are before you compare that to your RMSEP. In our fatty acid analysis using Fourier transform near infrared spectroscopy we have been able to obtain results matching our RMSECV with those of GC analysis.

I hope that is helpful,

Hormoz

The calculation of SEL for lab results and RPD for the ratio of lab SD to SE of NIR spectral model might help to interprete the performances between the two.

You can also plot a trend line of your independent validation set between NIR predicted results and Lab results to see if these lines following the same trend and have any offset in between. The caution is you need to make sure that your lab result is accurate and with good repeatability which can be indicated by SEL.

There is also a statement of criteria that the NIR model should produce SEP lesser than 2.0SEL or 1.5SEL at least. If the SEP is larger than your lab SEL then your model needs more calibration development and/or more samples.

Hopefully this information is helpful to you.

Jittima

Statistical design ..Chemometrics may actually sometime reflect good calibration model in Nir spectroscopy. If you compare any standard reference method in context your nir result.Then you can go to F .test for precession of the method. But you can seen model error in context of accuracy,  you may go to cross validation.