|
Krouwer Consulting
|
|
|
Bias advice for GUM – It’s uncertain! I have critiqued GUM (1) and a recent Letter reminds me of another concern with GUM (2). This Letter contains advice about bias which is taken from GUM: “if a bias is considered small compared with the overall uncertainty, it simply may be neglected”. The spirit of this advice is straightforward – if a bias would affect values after the seventh decimal point, ignore it. But life is not always that simple. Take linear drift for example, which is a “protocol dependent” bias (3) and depends on at least two quantities:
Assume that the calibration algorithm is incorrect because it leaves out temperature in the calculation of the response. Temperature is left out because it is supposed to be constant with respect to the temperature when the calibrator was assayed but assume that the temperature is drifting. The bias that results is a function of when the sample is assayed, and the rate of temperature drift. This brings up another concept which is the probability of a large enough error due to drift. That is, a large temperature drift and a sample assayed at the end of a calibration cycle (both of which would cause an error that is not negligible) might occur about 5% of the time, right at the border of many uncertainty intervals. Thus, can one ignore rare (e.g., those that occur much less often than 5%) large biases? So far, all of this is for one bias. What happens if one has more than one bias, whereby each bias considered by itself is “negligible”? So I find some “uncertainty” in the “Guide to the Expression of Uncertainty in Measurement”. I note also in passing that the Guide is really two guides: the Guide (pp 1-28) and the appendices - annexes in the ISO world (pp 29-90). So is there an alternative to GUM? Yes, calculate empirically based uncertainty intervals (4-6). References
|