Bias and Measurement Uncertainty

Many instruments provide results with a bias, meaning a deviation from the true value, which can either overestimate or underestimate the value being measured.

This systematic drift is sometimes difficult to detect, especially when it is small compared to the result. In such cases, the operator generally does not take it into account. However, in trace analysis, for example, its impact may have significant consequences, and it will therefore need to be included either in the result or in the uncertainty.

It is this second option, uncertainty, that we will try to clarify.

Some common-sense statements:

- The final uncertainty shall be greater than or equal to the uncertainty that would have been reported if the bias had been corrected.
- Underestimating the uncertainty indicates an invalid uncertainty statement; similarly, overestimating the uncertainty indicates a poorly constructed uncertainty budget.
- The method shall be limited to the method proposed by the Guide (GUM) when the bias correction is applied.
- For any coverage factor and any amplification of the bias, the confidence level of the expanded uncertainty should be at least the level obtained when the bias is corrected. For example, if the distribution of values that could reasonably be attributed to the measurand is Gaussian, then k = 2 should imply at least 95% confidence.
- The method shall be transferable so that the uncertainty and bias of a result can be used as components in another uncertainty statement.

Review of a few key concepts:

Measurement results shall be corrected for bias, and the uncertainty associated with the bias correction should be included as a contribution to the combined standard uncertainty.
However, when correction for measurement bias is not possible, the bias should be explicitly taken into account in the uncertainty statement.
A complete uncertainty statement includes the combined standard uncertainty, calculated as if the measurement result had been corrected for bias; an explicit statement of the bias value; and an expanded uncertainty that includes a term accounting for the bias.
The usual method of using the expanded uncertainty U, for a measurement result y that has an unknown “true” value of the measurand Y, is to produce an uncertainty interval, with a confidence level determined by the coverage factor k, such that:

Therefore, the measurement result is often expressed as:y ± U.

In cases where the result is corrected for bias δ, uA similar uncertainty interval can be constructed for the corrected measurement result ycor = (y-δ) given by:

ycor – U≤Y≤ycor + U

This is equivalent to the uncertainty interval of: Y– (U + δ)≤Y≤Y + (U – δ)

Accordingly, the measurement result could be stated as:

Unfortunately, this may lead to the possibility that one of the uncertainty bounds could become negative, for example, if the bias is positive and δ>U then the upper limit of the uncertainty will be negative.

This may create confusion for operators, particularly when making a statement of conformity. Therefore, an additional condition must be imposed: the uncertainty limits shall be greater than or equal to zero for all values of δ.

Comparison of Methods for Combining Bias and Uncertainty:

If we compare the proposed method for handling uncorrected bias with two other techniques that have been proposed to address this issue:

The first method treats the uncorrected bias as another source of uncertainty and simply combines it with the combined standard uncertainty. This method is denoted RSSuc.

The second method adds the bias to the expanded uncertainty. This method is denoted RSSU.

In contrast, the third proposed method combines the bias algebraically with the expanded uncertainty, except when the bias is large. This method is referred to as SUMU.

The three methods are presented below:

- Method RSSuc : Y = y ± U (RSSuc) with:

Often used consensually in standards, but not statistically relevant

- Method RSSU : Y = y ± U (RSSU) with:

- Method SUMU : Y = y ± U with:

with δ>0 and if k.uc-δ>0; 0 if k.uc-δ≤0

withδ<0 and if k.uc+δ>0; 0 if k.uc+δ≤0

Significant differences can be observed between the three methods.

The three graphs below show the actual statistical confidence of the three methods as a function of the uncorrected bias for coverage factors k = 1, 2, and 3.

The distributions are assumed to be Gaussian (normal) in all cases.

For example, in the case where k = 2, we would ideally like to maintain 95% confidence for all values of the uncorrected bias.

The SUMU method maintains this level of confidence until the ratio of the bias to the combined standard uncertainty becomes greater than the coverage factor.

For large bias values, the SUMU method produces uncertainty intervals that are slightly more conservative, meaning larger than necessary to provide valid 95% confidence levels.

The RSSuc method, on the other hand, may produce uncertainties that are much larger than necessary.

For example, with k = 2 and a bias twice as large as the combined standard uncertainty (δ/uc=2), the confidence level of the actual interval reaches nearly 100%, rather than the nominal 95%.

Although this overestimation of uncertainty is not necessarily catastrophic, it may lead to a greater reduction in the specification zone.

Conversely, the RSSU method significantly underestimates the true uncertainty.

For example, with a coverage factor k = 2 and an uncorrected bias twice as large as the combined standard uncertainty (δ/uc = 2), the uncertainty interval is underestimated, since the actual confidence achieved is below 80%, which is far from the nominal 95% confidence level.

The scale on the left-hand y-axis defines the full width of the uncertainty interval divided by the combined standard uncertainty. Therefore, the ratio would be equal to 2k, where k is the coverage factor, if the bias had been corrected.

As can be seen, the SUMU method always produces the smallest expanded uncertainty interval compared with the other methods, for all bias values and coverage factors.

An example of how the size of the expanded uncertainty could have an impact is shown on the right-hand y-axis.

This axis describes the percentage of the specification zone that is used by the expanded uncertainty interval for the rather atypical case where the ratio of the specification zone to the expanded uncertainty interval, if the bias had been corrected, is 4:1.

The SUMU method uses the smallest percentage of the specification zone compared with the other two methods.

For example, for k = 2 and an uncorrected bias equal to four times the combined standard uncertainty (δ/uc=4), the SUMU method would use 37.5% of the specification zone — compared with 25% if the bias had been corrected — whereas the RSSU and RSSuc methods use 56% and more than 100%, respectively.

The figures illustrate that, of the three methods described, the SUMU method provides a significant advantage when the uncorrected bias is included in the uncertainty.

This method always maintains an actual confidence level higher than the nominal confidence level corresponding to the coverage factor used.

Although these examples are based on a Gaussian distribution, the SUMU method maintains this relationship between k and the confidence level for any type of distribution, because the resulting interval always contains at least the same interval as if the result had been corrected.

The SUMU method minimizes the percentage of the specification range used by the expanded uncertainty interval.

In addition, this method avoids negative expanded uncertainties, which could be a source of confusion for the operator when determining the conformity zone.

The following examples are established using the consensus method: RSSuc.

Example 1 including a Type A evaluation of the bias uncertainty:

Consider a measurement result, Y, that has a bias δ1.

Assuming that δ1 is evaluated directly by repeated measurements of a reference standard with a combined standard uncertainty uref and as the mean difference between the calibrated value of the standard and the n measurements.

Let s denote the experimental standard deviation of the n measurements; the estimated Type A uncertainty is then s/√n.

uw represents the intra-laboratory uncertainty obtained in the laboratory for the standard.

In this example δ1<0 and kuc1+δ1> 0.

Worked Example No. 1

Suppose that a measuring device is designed to inspect parts with a length of 100 mm and has been verified using a reference gauge block with a combined standard uncertainty of Uref = 1,5 µm.

5 measurements were recorded, with an experimental standard deviation of 3.0 µm and a mean result 4.0 µm below the reference value: the bias is negative.

Within-laboratory uncertainty: uw=2,5 µm.

Therefore:


with a 95% level of confidence (K=2)

Example 2 – Type B bias uncertainty:

For certain measurements, the bias may be estimated rather than directly measured. For example, length measurements are often not corrected to the standard temperature of 20°C.

Therefore, the uncorrected thermal expansion represents a measurement bias. Suppose the temperature of a metal bar varies between 20°C and 30°C, with an estimated average of 25°C.

The estimated magnitude of the bias is given by δ2 (δ2> 0) which represents the length deviation due to uncorrected thermal expansion over the 5°C temperature difference.
Temperature variability can be described by a rectangular distribution over 10°C, i.e., by a standard uncertainty of 2.9°C (10/2√3) which, when multiplied by the appropriate coefficient of thermal expansion, gives the corresponding standard uncertainty utemp.

uw represents the within-laboratory uncertainty obtained by the laboratory for the standard.

Inthis example δ2> 0 et kuc2-δ2> 0.

Worked Example No. 2

Suppose a measuring machine measures the length of 100 mm bars with a standard uncertainty uref = 1,5 µm, and the instrument scale has a thermal expansion coefficient of 9 µm/m/°C, and the bar being measured has a thermal expansion coefficient of 22 µm/m/°C.

Differential thermal expansion is 22 µm/m/°C - 9 µm/m/°C = 13 µm/m/°C, corresponding to a mean bias of:

13 µm/m/°C x 5 °C x 0,100 m = 6,5 µm

The standard uncertainty associated with temperature variation (modeled as a rectangular distribution) is given by:

utemp = 13 µm/m/°C x 0,100 m x 2,9°C = 3,8 µm

within-laboratory uncertainty: uw=2,5 µm.

Therefore:

with a 95% level of confidence (K=2)