The general procedure follows the approach of the GUM guide and consists of five steps: define what is being measured and its mathematical model, identify the sources of uncertainty, quantify each one, combine them and expand the result. It is not a fixed recipe, because the sources depend on the instrument and the method, but the logic is always the same.
Step 1: measurement model
Write the relationship between the quantity being determined and the variables it depends on. In a typical calibration, the instrument error is the reading minus the value of the standard, plus any applicable corrections for temperature, resolution or other effects.
Step 2: sources of uncertainty
- Uncertainty of the standard, taken from its certificate.
- Repeatability of the instrument readings.
- Resolution of the instrument and the standard.
- Temperature difference or other environmental effects.
- Stability of the standard between calibrations.
- Effects of the setup or connection, as applicable.
Step 3: quantify
Sources evaluated statistically, such as repeatability, are treated with the standard deviation of the mean (Type A evaluation). The others are evaluated with prior information, such as certificates or specifications, assuming a suitable probability distribution, for example rectangular when only the limits are known (Type B evaluation). All are converted to standard uncertainty.
Step 4: combine
Each standard uncertainty is multiplied by its sensitivity coefficient, which indicates how much it influences the result, and they are combined in quadrature if the sources are independent. If there is correlation, it must be treated explicitly.
Step 5: expand
The combined uncertainty is multiplied by a coverage factor chosen for the desired confidence level. If the effective degrees of freedom are few, the factor must be corrected using the t-distribution.
A qualitative example
If the standard contributes little and repeatability dominates, improving the standard will not reduce the total uncertainty; repeatability is what must be addressed. That is why it pays to look at the full uncertainty budget and not just the final result.