Every calibration certificate you’ll ever read ends with a line like this:
Measurement uncertainty: ±0.002 mm (k = 2)
That number is the expanded uncertainty. It’s the figure the whole calibration exists to produce, and it’s the one people most often misread.
Expanded uncertainty is the combined standard uncertainty multiplied by a coverage factor, giving an interval with a stated level of confidence normally about 95%.
The rest of this page explains where it comes from, why k is almost always 2, when 2 isn’t enough, and the part that matters commercially how expanded uncertainty decides whether a part passes or fails.
The Formula for Expanded Uncertainty
U = k × u꜀
Where:
- U is the expanded uncertainty
- k is the coverage factor
- u꜀ is the combined standard uncertainty
That’s the whole formula. The difficulty is never the arithmetic it’s knowing what u꜀ should be and what k should be.
Note the capital letter. Upper-case U is always expanded. Lower-case u is always standard. Mixing them up changes the answer by a factor of two, and it’s the most common notation error in metrology.
Standard, Combined and Expanded Uncertainty
Three terms, three stages of the same calculation. Getting them straight is most of the battle.
| Term | Symbol | What it is | Confidence |
|---|---|---|---|
| Standard uncertainty | u | One source, as one standard deviation | ~68% |
| Combined standard uncertainty | u꜀ | All sources merged in quadrature | ~68% |
| Expanded uncertainty | U | Combined value × coverage factor | ~95% at k=2 |
The sequence:
- Each source becomes a standard uncertainty a certificate value divided by its k, a resolution divided by 2√3, a set of readings reduced to s ÷ √n.
- Those merge into the combined standard uncertainty: u꜀ = √(u₁² + u₂² + …)
- That gets multiplied by k to become the expanded uncertainty you actually report.
Why bother with the last step? Because 68% confidence is a weak basis for a decision. It means roughly one measurement in three falls outside your stated band. Nobody wants to accept or reject parts on those odds.
What Is the Coverage Factor (k)?
The coverage factor is the multiplier that widens your interval to a useful confidence level.
| k | Approximate confidence | Typical use |
|---|---|---|
| 1 | 68.27% | Internal working figures only |
| 2 | 95.45% | The near-universal default |
| 2.58 | 99% | Occasionally specified |
| 3 | 99.73% | Safety-critical measurements |
The relationship comes from the normal distribution. If your combined uncertainty behaves like a normal distribution and with several contributing sources it usually does, by the central limit theorem then roughly 95% of the probability sits within two standard deviations of the mean.
So k = 2 isn’t a convention someone invented. It’s the property of a bell curve.
Why k = 2 Is the Default
Three reasons, and it’s worth knowing all of them.
It’s mathematically appropriate. With enough contributing sources, the combined distribution approaches normal, and 2σ captures about 95%.
It’s internationally agreed. The GUM, ILAC, ISO/IEC 17025 and virtually every national accreditation body use k = 2 as the standard reporting basis. That makes certificates comparable across laboratories and countries which is the entire point of having a standard.
It’s a sensible risk balance. k = 1 is too narrow to make decisions on. k = 3 is so wide it would fail parts that are almost certainly fine. Two sits where most people are comfortable.
One thing to be clear about: k = 2 gives approximately 95.45%, not exactly 95%. For exactly 95% you’d use k = 1.96. Most laboratories state 95.45% or write “approximately 95%”. If a certificate says 95% at k = 2, it’s rounding.
When k = 2 Is Not Enough
Here’s what most explanations skip, and it matters.
The 95% claim assumes your combined standard uncertainty is well characterised. If it’s dominated by a Type A term based on very few readings, your estimate of that term is itself uncertain and k = 2 no longer delivers 95% confidence.
The fix is to calculate the effective degrees of freedom using the Welch–Satterthwaite formula:
ν_eff = u꜀⁴ ÷ Σ (uᵢ⁴ ÷ νᵢ)
Where ν = n − 1 for Type A sources, and ν = ∞ for well-characterised Type B sources (which therefore contribute nothing to the denominator).
Then read k from the t-distribution:
| ν_eff | k |
|---|---|
| 1 | 13.97 |
| 2 | 4.53 |
| 3 | 3.31 |
| 4 | 2.87 |
| 5 | 2.65 |
| 6 | 2.52 |
| 8 | 2.37 |
| 10 | 2.28 |
| 20 | 2.13 |
| 50 | 2.05 |
| ∞ | 2.00 |
How much difference does this make?
A concrete case. Suppose repeatability dominates your budget and you took only three readings:
- u_A = 0.0866, based on 3 readings, so ν_A = 2
- One small Type B source, u = 0.02, ν = ∞
- u꜀ = √(0.0866² + 0.02²) = 0.0889
Welch–Satterthwaite gives ν_eff ≈ 2. From the table, k = 4.53.
With k = 2: U = 0.18 With k = 4.53: U = 0.40
More than double. Using k = 2 here would understate your uncertainty by a factor of 2.3, and every conformity decision made from it would be overconfident.
When you can safely ignore this: if Type B sources dominate your budget, ν_eff comes out large and k = 2 is fine. That covers most real calibrations. But if repeatability is your biggest term and you took fewer than about six readings, run the calculation.
How to Calculate Expanded Uncertainty: Worked Example
Calibrating a torque wrench at 50 N·m.
Type A – six readings: 49.8, 50.2, 50.1, 49.9, 50.0, 50.1 N·m
Mean = 300.1 ÷ 6 = 50.02 N·m Sample standard deviation, s = 0.147 N·m u_A = 0.147 ÷ √6 = 0.060 N·m (ν = 5)
Type B sources:
| Source | Value | Distribution | Divisor | Standard uncertainty |
|---|---|---|---|---|
| Repeatability | 0.147 N·m | Normal | √6 | 0.060 N·m |
| Reference transducer certificate | 0.10 N·m (k=2) | Normal | 2 | 0.050 N·m |
| Resolution (0.1 N·m) | 0.05 N·m | Rectangular | √3 | 0.029 N·m |
| Temperature effect | 0.05 N·m | Rectangular | √3 | 0.029 N·m |
Combined standard uncertainty:
u꜀ = √(0.060² + 0.050² + 0.029² + 0.029²) u꜀ = √0.00777 = 0.088 N·m
Check the degrees of freedom:
ν_eff = (0.088)⁴ ÷ [(0.060)⁴ ÷ 5] ≈ 23
From the table, ν_eff = 23 gives k ≈ 2.11 close enough to 2 that it makes no practical difference. Six readings was enough.
Expanded uncertainty:
U = 2 × 0.088 = 0.176 ≈ 0.2 N·m
Result: 50.0 ± 0.2 N·m (k = 2, approximately 95% confidence)
How to Read Expanded Uncertainty on a Calibration Certificate
A certificate line contains more information than most people extract from it.
Reported value: 25.0004 mm Expanded uncertainty: ±0.0008 mm (k = 2)
What that actually tells you:
The combined standard uncertainty was 0.0004 mm. Divide U by k to recover it. You need this figure if you’re going to use the instrument in a budget of your own you must never carry an expanded value into another budget without converting it back.
The true value is 95% likely to lie between 24.9996 and 25.0012 mm.
It says nothing about whether the item passed. Uncertainty and tolerance are separate things. Conformity is a decision made using both.
It’s specific to those conditions. The stated uncertainty applies at that measured value, at that temperature, with that method. Extrapolating it to a different size or condition isn’t valid.
Three things to check on any certificate you receive: that a k value is stated (without it the number is ambiguous), that a confidence level is stated, and that the uncertainty is plausible for the equipment. An uncertainty far smaller than you’d expect is a warning sign, not a bargain.
Expanded Uncertainty vs Tolerance vs Accuracy
These get conflated constantly, and the confusion has real cost.
| Term | What it describes | Belongs to |
|---|---|---|
| Expanded uncertainty | Doubt about the measurement result | The measurement |
| Tolerance | Permitted variation in the characteristic | The part, from the drawing |
| Accuracy | Closeness to the true value | The instrument or result |
Expanded uncertainty is not tolerance. Tolerance is what the design allows the part to be. Uncertainty is how well you know what the part actually is. One comes from the designer, the other from your measurement process.
The relationship between them determines whether you can make a decision at all. If a part has a ±0.05 mm tolerance and your expanded uncertainty is ±0.04 mm, your doubt consumes most of the permitted range and reliable pass/fail calls become impossible.
Expanded uncertainty is not accuracy either. A manufacturer’s accuracy specification is a limit on the instrument. Expanded uncertainty is a calculated property of a specific measurement, incorporating that specification along with everything else.
How Expanded Uncertainty Affects Pass/Fail Decisions
This is where it stops being theory.
A shaft must be 25.00 mm ±0.01 mm. You measure 25.009 mm with an expanded uncertainty of ±0.004 mm.
Is it in tolerance?
The measured value sits inside the limits. But the true diameter could be anywhere from 25.005 to 25.013 mm and the upper part of that range is out of tolerance. You cannot honestly declare a pass.
ISO 14253-1 addresses this directly. Two approaches:
Simple acceptance. Accept if the measured value is within the specification limits, and let uncertainty sit where it falls. Fast, but it accepts some genuinely non-conforming parts.
Guarded acceptance. Shrink the acceptance limits by the expanded uncertainty:
Acceptance limit = Specification limit − U
For our shaft: 25.010 − 0.004 = 25.006 mm. Anything above that is rejected. Our 25.009 mm part fails guarded acceptance.
This is called a guard band, and it moves the risk of a wrong decision from your customer to you. Which approach applies is a decision rule, and ISO/IEC 17025 requires laboratories to agree it with the customer in advance and state it on the report.
Test Uncertainty Ratio (TUR)
TUR compares your tolerance with your measurement uncertainty:
TUR = Tolerance ÷ Expanded uncertainty
For the torque wrench above, with a ±2 N·m tolerance and U = 0.2 N·m:
TUR = 2 ÷ 0.2 = 10:1
| TUR | Verdict |
|---|---|
| 10:1 or better | Excellent; guard banding rarely needed |
| 4:1 | The widely accepted minimum |
| 2:1 | Marginal; guard banding essential |
| Below 1:1 | The measurement cannot support the decision |
4:1 is the traditional benchmark. Below it, your uncertainty consumes enough of the tolerance that guard banding becomes necessary and below 1:1, you’re not really measuring the characteristic at all.
This is also the practical answer to “how small does my uncertainty need to be?” Small enough to give you a TUR of at least 4:1. Beyond that you’re spending money for no decision-making benefit.
What Is CMC?
Calibration and Measurement Capability is the smallest expanded uncertainty a laboratory can achieve under normal working conditions, using its best equipment on a near-ideal item.
It appears in the laboratory’s accreditation scope, and it matters for one reason: a lab’s reported uncertainty for a real job can never be smaller than its CMC for that measurement.
Real jobs involve real items with form error, surface roughness and thermal behaviour that the ideal artefact used to establish the CMC didn’t have. So your certificate uncertainty will typically be larger than the CMC, sometimes considerably.
If you’re choosing a calibration supplier, compare CMCs rather than marketing claims. And if you receive a certificate quoting an uncertainty smaller than the lab’s accredited CMC, question it that shouldn’t happen.
Expanded Uncertainty in ISO/IEC 17025
For accredited laboratories, reporting expanded uncertainty isn’t optional.
ISO/IEC 17025:2017 requires laboratories to evaluate measurement uncertainty and report it. Calibration certificates must always include it. Testing reports must include it where it’s relevant to the validity of the result, where a customer asks, or where it affects a conformity statement.
The standard also requires:
- A stated coverage factor and confidence level alongside the number
- A documented decision rule where conformity statements are made, agreed with the customer
- Metrological traceability, an unbroken chain of calibrations to the SI, each link carrying its own uncertainty
That last point is where expanded uncertainty accumulates through a traceability chain. Your uncertainty always includes the uncertainty of the standard you calibrated against, which included the uncertainty of the standard it was calibrated against, and so on back to the national institute.
Common Misunderstandings
“±0.002 mm means the instrument is accurate to 0.002 mm.” No. It means the true value of that specific measurement is 95% likely to lie within 0.002 mm of the reported value. It isn’t a general specification for the instrument.
“The k value doesn’t really matter.” It changes the number by a factor of two or more. A certificate without a stated k is incomplete.
“Expanded uncertainty tells me whether the part passed.” It doesn’t. Conformity requires comparing the result and its uncertainty against the tolerance, using an agreed decision rule.
“I can put a certificate value straight into my uncertainty budget.” Only after dividing by its k. Carrying an expanded value into a budget as though it were standard doubles that contribution.
“k = 2 always means 95%.” It means 95.45%. Exactly 95% would be k = 1.96. Minor, but worth knowing when a specification states one or the other.
“A smaller expanded uncertainty is always better.” Only up to a point. Once your TUR passes about 4:1, further reduction buys nothing for the decision you’re making.
“k = 2 is safe regardless of how many readings I took.” Not if repeatability dominates and n is small. With three readings, k can legitimately be 4.53.
Frequently Asked Questions
What is expanded uncertainty?
The combined standard uncertainty multiplied by a coverage factor, giving an interval expected to contain the true value with a stated confidence normally about 95% at k = 2.
What is the formula for expanded uncertainty?
U = k × u꜀, where k is the coverage factor and u꜀ is the combined standard uncertainty.
What does k = 2 mean?
The combined standard uncertainty was multiplied by 2, widening the interval from about 68% to about 95.45% confidence.
What is the difference between standard and expanded uncertainty?
Standard uncertainty is one standard deviation, roughly 68% confidence. Expanded uncertainty multiplies it by a coverage factor to reach roughly 95%.
Why is k = 2 used instead of 1.96?
Because 2 is simpler, internationally agreed, and gives 95.45% near enough to 95% for practical purposes while keeping certificates comparable between laboratories.
When should I use a coverage factor other than 2?
Use k = 3 for safety-critical work. Use a larger factor derived from the t-distribution when a Type A term based on few readings dominates your budget.
How do I convert expanded uncertainty back to standard uncertainty?
Divide by the stated k. A value of ±0.008 mm at k = 2 corresponds to a standard uncertainty of 0.004 mm. You must do this before using a certificate value in your own budget.
Is expanded uncertainty the same as tolerance?
No. Tolerance is the variation permitted in the part by its design. Expanded uncertainty is the doubt in your measurement of it.
What is a good test uncertainty ratio?
4:1 is the widely accepted minimum, meaning your expanded uncertainty is no more than a quarter of the tolerance. 10:1 is comfortable.
What is CMC?
Calibration and Measurement Capability the smallest expanded uncertainty a laboratory can achieve under ideal conditions, listed in its accreditation scope. Reported uncertainties for real items are usually larger.
