How Do You Evaluate Measurement Uncertainty Using the GUM Method?

Every measurement—whether in a lab, on a production floor, or in field calibration—is incomplete without a statement of its uncertainty. But how do you quantify it rigorously and consistently? The answer lies in the Guide to the Expression of Uncertainty in Measurement (GUM), published as ISO/IEC Guide 98-3 (and JCGM 100). This article explains the core method, step by step, so you can build a defensible uncertainty budget.

What It Is

Measurement uncertainty is a parameter that characterizes the dispersion of values that could reasonably be attributed to a measurand (the quantity being measured). It is not the same as an error—error is a single difference from a true value, while uncertainty is a range. The GUM provides a unified, internationally accepted framework for evaluating and expressing this range, making results comparable across laboratories and countries.

The GUM method is built on the "law of propagation of uncertainty": you model your measurement as a function of input quantities, estimate the uncertainty of each input, and combine them to get a standard uncertainty for the result.

How It Works: The GUM Steps

The GUM approach (per JCGM 100) follows four main steps:

1. Define the measurand and the measurement model.
Write the relationship between the output estimate \( y \) and input estimates \( x_i \):
\[
y = f(x_1, x_2, \ldots, x_N)
\]
This is your mathematical model.

2. Evaluate the standard uncertainty of each input.
Each input uncertainty is classified by how it is evaluated:

• Type A evaluation: Based on statistical analysis of repeated observations. The standard uncertainty is the experimental standard deviation of the mean:

\[
u_A = \frac{s}{\sqrt{n}}
\]
where \( s \) is the sample standard deviation and \( n \) is the number of readings.

• Type B evaluation: Based on scientific judgment, calibration certificates, manufacturer's specs, or other non-statistical information. For example, if a certificate states an expanded uncertainty \( U \) with coverage factor \( k \), then:

\[
u_B = \frac{U}{k}
\]
If a specification gives a rectangular distribution with half-width \( a \), then:
\[
u_B = \frac{a}{\sqrt{3}}
\]

3. Compute the combined standard uncertainty \( u_c(y) \).
For uncorrelated inputs, combine them in quadrature using sensitivity coefficients \( c_i = \partial f/\partial x_i \):
\[
u_c(y) = \sqrt{\sum_{i=1}^{N} \left( c_i \, u(x_i) \right)^2}
\]
If the model is a simple sum or difference, the coefficients are ±1. If it is a product or quotient, you can work in relative uncertainties.

4. Determine the expanded uncertainty \( U \).
Multiply the combined standard uncertainty by a coverage factor \( k \) to obtain an interval with a high level of confidence:
\[
U = k \cdot u_c(y)
\]
The GUM commonly uses \( k = 2 \), which corresponds to approximately 95% confidence for a normal distribution. (For higher confidence, \( k = 3 \) gives about 99.7%.)

A Worked Illustrative Example

Example data (illustrative only). Suppose you measure the length of a block with a calibrated ruler. The model is:
\[
L = L_{\text{read}} + \delta_{\text{cal}}
\]
where \( L_{\text{read}} \) is the reading and \( \delta_{\text{cal}} \) is a calibration correction.

• Type A: You take 5 readings: 20.02, 20.01, 20.03, 20.02, 20.02 mm. The mean is 20.02 mm, and the sample standard deviation \( s \) is 0.0071 mm. Thus:

\[
u_A = \frac{0.0071}{\sqrt{5}} \approx 0.0032 \text{ mm}
\]

• Type B: The calibration certificate states the ruler has an expanded uncertainty of 0.01 mm at \( k = 2 \). Therefore:

\[
u_B = \frac{0.01}{2} = 0.005 \text{ mm}
\]

• Combined standard uncertainty (assuming no correlation, sensitivity coefficients = 1):

\[
u_c = \sqrt{0.0032^2 + 0.005^2} \approx \sqrt{0.000010 + 0.000025} \approx 0.0059 \text{ mm}
\]

• Expanded uncertainty at \( k = 2 \) (≈95% confidence):

\[
U = 2 \times 0.0059 \approx 0.012 \text{ mm}
\]

Reported result: \( L = 20.020 \pm 0.012 \) mm (k=2, approximately 95% confidence).

Common Pitfalls

• Forgetting correlation: If input quantities are correlated (e.g., same calibration standard), the simple root-sum-square formula is invalid. Use the full GUM formula with covariance terms.
• Mixing up Type A and Type B: The distinction is about evaluation method, not about "random" vs. "systematic" error. Both contribute to uncertainty.
• Using the wrong distribution for Type B: A rectangular distribution (divide by √3) is only valid when the spec is a hard limit with no central tendency. If the manufacturer gives a standard deviation or a confidence interval, use that directly.
• Reporting without a coverage factor: An uncertainty value without \( k \) or a confidence level is meaningless. Always state both.

---

Mastering uncertainty evaluation is essential for credible quality control and calibration. To simplify the arithmetic and avoid calculation errors, try the free, GUM-compliant calculator at: https://www.6sq.com/tools/gum/ — it walks you through Type A and Type B inputs, combines them automatically, and gives you the expanded uncertainty in seconds.