Date: Fri, 5 Jan 2001 21:25:18 -0500 From: "Ouellette, Mike" To: 'Greg Gogates' Subject: RE: Uncertainty This is an example where the measurement exceeds the specified limits but not when extended downwards by the measurement uncertainty. There is no universally accepted way to report compliance in this situation. What matters is that you do it in a way that meets your customer requirements and that consistently follows a documented policy and procedure. The issue centres on how you agree to share the risks with your client (i.e., the risk of wrongly accepting an instrument that's actually out of tolerance vs the risk of rejecting one that's actually in tolerance. There are costs associated with each type of risk. This is explained in ILAC Guidance Document G8 "Guidelines on Assessment and Reporting of Compliance with Specification (based on measurements and tests in a laboratory)." ILAC G8 recommends that in the absence of any instruction from your client, that you call it "in spec" if the measurement falls within the limits when extended by HALF the measurement uncertainty. If the measurement falls outside the limits when extended by HALF the measurement uncertainty, then ILAC recommends calling it out of spec. If the measurement falls within the acceptance limits when extended by half the measurement uncertainty, then ILAC G8 proposes that it is not possible to state noncompliance without reducing the level of confidence. The ILAC G8 document can be downloaded from http://www.ilac.org/ . In the example below, subtracting 0.5*uncertainty gives 2.88972 which remains outside the upper acceptance limit of 2.88970. Unless the client requests otherwise, the ILAC document recommends that the device be called out of spec in this example. The ILAC guide also covers the other tricky case when the measurement falls within the acceptance limits but not when extended by the measurement uncertainty. Another valid approach in this situation would be to tighten the limits in attempt to account for the measurement uncertainty. This is called guardbanding. The guardbanding technique recommended by CLAS is to calculate the new limits as follows: NewLimits = SQRT ( OldLimits^2 - MeasurementUncertainty^2 ) . This is explained in CLAS Reference Document 3 published at http://www.nrc.ca/inms/clas/clas003.html . Please keep in mind that we need to update this document to clearly state that the measurement uncertainty is that for the measurement process, not just the uncertainty in the reference standard alone. Mike Ouellette National Research Council Canada Calibration Laboratory Assessment Service (CLAS) Ottawa, Ontario, Canada K1A 0R6 Tel. +1 (613) 993-9619 Fax. +1 (613) 952-1394 Mike.ouellette@nrc.ca http://www.nrc.ca/inms