Metrology Monday! #95 A Discussion on Conformity Assessment, Decision Rules and Measurement Decision Risk – ISO 14253-1 and NCSL RP-10 decision rules

Last week I shared my personal favorite guardband method, so it is only fair that I also introduce my least favorite.

This guardband method is generally referred to as ISO 14253-1 or the ILAC G8 method.? ISO 14253-1 is an ISO standard is associated with Geometrical product specifications and was originally developed to solve disagreements between suppliers and customers for dimensional parts.? Originally this standard stated that a part could not be called good unless its measured value was the tolerance minus the manufacturer’s uncertainty, and it could not be rejected unless the part measured greater than the tolerance plus the customer’s uncertainty. ?This left a pretty wide zone of “I don’t know.”?

In the first version of ILAC G8, this was the only decision rule that was thoroughly explained.? The present version of ILAC G8 describes many types of decision rules, but this name still is commonly used for this method.? The decision rule is defined as:

Where the Test Limit is computed as the Specification Limit minus the expanded Uncertainty of measurement.? While it is very easy to employ, and the resulting False Accept risk is very low, like all guardband-based decision rules, it is not a binary rule.? If the measurement happens to be less than the test limit, it is a pass or in-tolerance. If it is greater than the specification limit, it is fail or out-of-tolerance.? If the measurement lies between the test limit and the specification limit it is an indeterminate result, meaning that while it is not a failing measurement, we cannot state with sufficient False Accept risk that the measurement is good.? This guardband method is used when evaluating risk using conditional probability.? If the TUR is low, as for example 2:1, you have an indeterminate zone that is 50% of the specification limit.

We have evaluated this method using our joint probability tools, assuming a 95% End of Period Reliability and a 95% level of confidence in the measurement uncertainty and published the risk curves on the graph above with the associated arrows pointing at them.? While the False Accept risk is very small, approximately 0.02% at a 4:1 TUR and about 0.0325% at a 2:1 TUR, the False Reject risk is about 10% at 4:1 and goes up to 30% at 2:1 (because the guardband is .5 of the specification limit, as noted last week).? This may be a good rule for your company, if you have a very low tolerance for False Accept risk.? For most customers, 1 to 2% False accept risk is perfectly fine, so this rule is overly conservative and drives False Reject to a very high value.? There was a time when European accreditation bodies would only allow this decision rule to be used by its accredited laboratories.? The reason that I really don’t like this rule was because it was required by people who really did not understand the consequences of False Accept and False Reject from the rule.? Now, accreditation bodes are more open to using other decision rules if the risk is documented and understood.

Years ago, the National Conference of Standards Laboratories (now known as NCSL International) utilities committee came up with their own decision rule that was a hybrid of the 4:1 rule and the ISO 14253-1 rule.? NCSL International named the rule after the Recommended Practice (RP-10) that it was published in.

The NCSL RP-10 rule is shown as above.? Note that if there is a 4:1 TUR, the uncertainty would be .25 of the specification limit, so for this case the Test Limit equals the Specification Limit, so no guardband would be applied.? If the TUR was less than 4:1, it would create a guardband that was not as aggressive as the ISO 14253-1 rule, but it was still easy to compute.? However, this also creates a nonsensical situation for TURs that are greater than 4:1 because for those cases, the Test Limit actually exceeds the Specification Limit.? For these cases the False Accept risk would be much greater than 2%, which defeats the whole purpose of this decision rule.? What the utility committee members would do to avoid this is to estimate that the TUR is no greater than 4:1, no matter what it really was.? If your TUR was 80:1, they would still plug in 4:1.?

While I noted above that it is of medium difficulty to compute guardbands, I did not include a risk evaluation because of the issues above a 4:1 TUR.

The risk curves associated from a 4:1 TUR to a 1:1 TUR for the NCSL RP-10 method are shown above with the arrows.? It is interesting to note that while the False Accept risk is less than 1% (equal to the risk of a 4:1 TUR) at 4:1, the False Accept risk drops down to about 0.3% at a 2:1 TUR.? False Reject once again starts at the same point as the 4:1 TUR, about 1.5%, but it increases to about 15% at a 2:1 TUR.

While I am closing for this week, I still have several more decision rules to discuss. I hope that you can appreciate that different decision rules have different levels of risk. #MetrologyMonday #FlukeMetrology??