Definition of bias in ISO 12099

ptillmann's picture

(if there is still someone awake in this forum.)

In ISO 12099:2016 the definition of bias has changed (compared to ISO 12099:2007):

It was bias = sum (NIRS - ref) / n   (2007),

it is bias = sum (ref - NIRS) / n     (2016).

What seems to be a small change, is in fact truning the world upside down (I don't assume.our Australian hosts for NIR 2019 caused it.)

In my understanding a NIRS method with a bias of +1 unit, results in NIRS values 1 unit above the targeted reference level. But the new formular changes this. My understanding is supported by third party / third industry literature.

Anybody has noticed this change?
Anybody knows why it was changed?




hlmark's picture

I don't know anything about the ISO standards or why the definiiton of "bias" was changed, but here's some common sense: The concept oof "bias" comes from classical wet chemistry, when "bias" was caused by the effect of systematic variables whose nature could be determined and values measured. Then NIR came along. All of a sudden the picture changed. Bias was no longer caused by systematic variables of known provenence, but by random variables with unknown causes. The only thing that could be done was to treat it empirically, like everything else about NIR. One of the last steps in development of an NIR method was to measure the bias of a method and then make a "bias correction" on future samples. Once this is done, if done propoerly, the bias simply disappears (or is so small as to be statistically non-significant); it has been computed away, and it doesn't matter what it was beforehand. Purists worry about which way it was computed, but that has no substantive meaning or significance; it's like counting how many angels can dance on the head of a pin.

hlmark's picture

More thoughts: The key word in the question is "reference". Wet chemistry is based on the existence of reference samples whose properties are considered "known" and can therefore be used as absolute checks on the accuracy of the analytical method when determining the "bias". Again, in NIR we geenrally have no such absolute references, and therefore have no way to know whether a given measured value has a positive or negative "bias" to start with.

ianm's picture

Yes I noticed it, I have changed every instance I see the 2016 formulation to the 2007 formulation.

However, the 2016 formulation does calculate the number that should be entered in the bias screen in order to fix the problem. With the 2007 formulation you have to enter the negative of the number that computes.

This is like sample ID and moisture basis, the simplest hard problems we have.

Posted on behalf of Charlie Hurburgh by Ian Michael