If you have ever compared two machines, two shifts, or two suppliers and found yourself saying “well, their standard deviation is bigger, but they’re also making a bigger part, so is that really worse?”, you were reaching for the Coefficient of Variation, even if you didn’t know its name yet. The Coefficient of Variation (CV) is the tool that lets you compare variability fairly across datasets that don’t share the same scale, the same units, or the same average.
Standard deviation on its own only tells you how spread out a dataset is in its own units. It can’t tell you whether that spread is a big deal or a small deal relative to the size of what you’re measuring. The Coefficient of Variation fixes that by expressing variability as a percentage of the mean, which makes it one of the most useful, and most underused, measures of dispersion in a Six Sigma practitioner’s toolkit.
What Is the Coefficient of Variation?
The Coefficient of Variation is the standard deviation of a dataset divided by its mean, expressed as a percentage. It is also called relative standard deviation (RSD), which is a more descriptive name: it tells you how large the standard deviation is relative to the average value, rather than in absolute terms.
Because it’s a ratio of two quantities measured in the same units, the units themselves cancel out. A CV is always a pure percentage, with no unit attached, regardless of whether the underlying data was measured in millimeters, minutes, dollars, or defects per unit. That’s what makes it so useful for comparing variability across datasets that would otherwise be impossible to compare directly.
Where the Coefficient of Variation Comes From
The Coefficient of Variation isn’t a modern Six Sigma invention, it dates back to the statistician Karl Pearson in the mid-1890s. Pearson was studying anthropometric data (physical measurements of people) and ran into exactly the same problem Six Sigma practitioners run into today: he needed to compare the variability of groups that had meaningfully different average sizes, in his case, comparing variability between groups with different average body measurements. A raw standard deviation couldn’t answer that question fairly, so he devised a dimensionless ratio, the standard deviation divided by the mean, that could. More than a century later, it’s the exact same tool you’d reach for to compare a small-parts line against a large-parts line, or a lab instrument’s precision against a different lab’s instrument measuring on a different scale.
Coefficient of Variation Formula
There are two versions of the formula, depending on whether you’re working with a full population or a sample, the same distinction that applies to standard deviation generally.
| Scenario | Formula | Symbols |
|---|---|---|
| Population | CV = (σ / μ) × 100% | σ = population standard deviation, μ = population mean |
| Sample | CV = (s / x̄) × 100% | s = sample standard deviation, x̄ = sample mean |
In practice, almost every CV you calculate on the job will use the sample formula, since you’re almost always working from a sample of a process rather than every unit that process has ever produced or ever will.
Why Not Just Use Standard Deviation?
Here’s the problem CV solves. Suppose you run two production lines.
- Line A fills small bottles. Mean fill weight = 50 grams, standard deviation = 2 grams.
- Line B fills large drums. Mean fill weight = 5,000 grams, standard deviation = 40 grams.
Line B’s standard deviation (40 grams) is 20 times larger than Line A’s (2 grams). At a glance, that looks like Line B has a much bigger variability problem. But look at the CV instead:
- Line A: CV = 2 / 50 × 100% = 4%
- Line B: CV = 40 / 5,000 × 100% = 0.8%
Relative to what each line is actually producing, Line B is five times more consistent than Line A, the exact opposite conclusion you’d draw from the raw standard deviations. This is the core reason CV exists: standard deviation answers “how spread out is the data,” while CV answers the more useful question, “how spread out is the data, relative to what’s typical for this process.”
Worked Example: Calculating the Coefficient of Variation
Suppose a Green Belt is measuring cycle time, in minutes, for a customer service call center. A sample of 8 calls gives the following handle times:
4.2, 5.1, 3.8, 6.0, 4.5, 5.5, 4.0, 4.9 (minutes)
Step 1: Calculate the mean. Add the 8 values and divide by the count: (4.2 + 5.1 + 3.8 + 6.0 + 4.5 + 5.5 + 4.0 + 4.9) / 8 = 38.0 / 8 = 4.75 minutes.
Step 2: Calculate each deviation from the mean, then square it. This is the same groundwork you’d do to find the standard deviation on its own:
| Call Time (x) | Deviation (x − mean) | Squared Deviation |
|---|---|---|
| 4.2 | −0.55 | 0.3025 |
| 5.1 | 0.35 | 0.1225 |
| 3.8 | −0.95 | 0.9025 |
| 6.0 | 1.25 | 1.5625 |
| 4.5 | −0.25 | 0.0625 |
| 5.5 | 0.75 | 0.5625 |
| 4.0 | −0.75 | 0.5625 |
| 4.9 | 0.15 | 0.0225 |
Step 3: Sum the squared deviations and divide by n − 1 (we’re working from a sample, so we use n − 1, not n). The squared deviations sum to 4.1. Dividing by (8 − 1) = 7 gives a sample variance of 4.1 / 7 ≈ 0.586.
Step 4: Take the square root to get the sample standard deviation. √0.586 ≈ 0.765 minutes.
Step 5: Divide by the mean and convert to a percentage. CV = (0.765 / 4.75) × 100% ≈ 16.1%.
That 16.1% is now a portable number. It can be compared directly to the CV of a completely different call center, a different queue, or even a different metric entirely, like order-processing time measured in seconds, something a raw standard deviation of “0.765 minutes” could never do on its own.
Interpreting a Coefficient of Variation
A lower CV means the data clusters more tightly around the mean, relative to the size of the mean; a higher CV means more relative spread. Several rough rule-of-thumb bands show up across different industries and textbooks, but it’s important to treat them as context-dependent heuristics, not universal statistical law. There is no single, official cutoff where “moderate” becomes “high.”
| CV Range | Typical Interpretation |
|---|---|
| Under ~10% | Low relative variability (common target in lab/measurement settings) |
| ~10% to 20% | Moderate relative variability (often acceptable in field/process settings) |
| ~20% to 30% | Elevated; context-dependent, investigate the source of variation |
| Above ~30% | High relative variability by most conventions |
What counts as acceptable in a precision laboratory measurement (often under 10%) would be considered tight for a messier, human-driven service process (where 15-20% might be a realistic, perfectly healthy target). Always set your own threshold based on the process, the customer’s requirements, and what’s actually achievable, rather than importing a number from an unrelated industry.
Coefficient of Variation vs. Standard Deviation vs. Variance
These three measures of spread are closely related, but they answer different questions:
- Variance is the average of the squared deviations from the mean. It’s mathematically foundational (it’s what you calculate first to get standard deviation), but its units are squared, which makes it hard to interpret directly. If your data is in minutes, variance is in minutes squared.
- Standard deviation is the square root of variance, which brings the units back to the original scale. It’s the most commonly reported measure of spread, and it’s the right choice when you’re only looking at one dataset, or comparing datasets that share the same units and a similar mean.
- Coefficient of Variation is standard deviation expressed as a percentage of the mean. It’s the right choice specifically when you need to compare variability across datasets with different units, different scales, or meaningfully different means.
In short: use variance when you need it as a stepping stone to another calculation, use standard deviation as your everyday measure of spread, and reach for CV the moment you need to compare apples to oranges.
When Six Sigma Practitioners Actually Use CV
CV shows up in a handful of specific, practical situations on a Six Sigma project:
- Comparing variability across product lines with different target sizes. A process making 2-inch bolts and a process making 12-inch bolts will naturally have different absolute tolerances and different standard deviations; CV lets you ask “which process is relatively more consistent” without that size difference distorting the answer.
- Comparing the same metric across very different volumes. A high-volume product line will often show a larger raw standard deviation than a low-volume line purely because of its scale; CV normalizes that away.
- Benchmarking measurement systems. Some Gage R&R and lab-precision studies report %RSD (CV) alongside, or instead of, raw standard deviation, specifically because it’s portable across instruments measuring different magnitudes.
- Supplier and vendor comparisons. If two suppliers report defect rates, cycle times, or dimensional data with different average outputs, CV gives you an apples-to-apples read on which one is actually more consistent, not just which one has a smaller number on paper.
Calculating CV in Excel or Minitab
You rarely need to do the squared-deviation arithmetic by hand once you’re working with real data. In Excel or Google Sheets, with your data in cells A2:A9, the whole calculation is one formula:
=STDEV.S(A2:A9)/AVERAGE(A2:A9)*100
Use STDEV.S for a sample (the far more common case) or STDEV.P if you genuinely have the entire population, not a sample of it. In Minitab, CV is available directly as an option inside Stat > Basic Statistics > Display Descriptive Statistics, listed in the output alongside the mean and standard deviation, so you don’t need to build the formula manually at all.
Common Mistakes and Limitations
CV is a simple calculation, but it’s easy to misuse. Watch for these pitfalls:
- Don’t use CV when the mean is at or near zero. Because the mean is in the denominator, a mean close to zero produces a CV that blows up toward infinity or becomes wildly unstable, even for a perfectly normal, well-behaved dataset. Profit margins, temperature changes, and other measures that can be positive, negative, or zero are poor candidates for CV.
- CV only makes sense for ratio-scale data. A ratio scale has a true, meaningful zero, like length, weight, time, or money. Interval-scale data, like temperature in Celsius or Fahrenheit, does not have a true zero (0°C isn’t “no temperature”), so a CV calculated on Celsius data will change if you switch to Fahrenheit, even though nothing about the underlying variability has changed. That makes it meaningless for interval-scale measurements.
- Don’t compare a CV calculated on raw data to one calculated on already-transformed data (like percentages, index scores, or log-transformed values) without checking that the comparison still makes sense; transforming the data changes what the mean and standard deviation actually represent.
- Small sample sizes make CV unstable. Like any statistic built from standard deviation, CV calculated from a handful of data points can swing widely from sample to sample. Treat a CV from fewer than about 10 data points as a rough signal, not a precise number.
Case Study: Using CV to Settle a Debate Between Two Shifts
A packaging plant runs two shifts on the same filling line. The day shift and night shift fill different SKUs: the day shift mostly runs a large 2-liter container, and the night shift mostly runs a smaller 500ml container. After a few weeks of complaints that the night shift was “less consistent,” a Black Belt pulled fill-weight data from both shifts.
| Shift | Mean Fill Weight | Standard Deviation | CV |
|---|---|---|---|
| Day (2L containers) | 2,000 g | 18 g | 0.9% |
| Night (500ml containers) | 500 g | 12 g | 2.4% |
On raw standard deviation alone, night shift’s 12 grams actually looks better than day shift’s 18 grams. But once the Black Belt converted both to CV, the real picture emerged: night shift’s variability, relative to the much smaller amount it’s filling, was more than two and a half times larger than day shift’s. The complaint about night shift wasn’t unfounded, it was just being argued with the wrong number.
With the real gap quantified, the team could move into root cause analysis on the night shift’s filling equipment specifically, rather than chasing a problem that the raw numbers had been hiding. This is the practical value of CV on a live project: it stops you from comparing processes that look similar on the surface but aren’t actually on the same footing.
How CV Thresholds Differ Across Industries
One of the most common mistakes practitioners make with CV is borrowing a threshold from a field that has nothing to do with their own process. The acceptable range genuinely does shift by industry and by what’s being measured, because the underlying risk and achievable precision are different in each case.
- Analytical and clinical laboratories often hold themselves to CV under 5-10% for repeated measurements of the same sample, because instrument precision at that level is both achievable and necessary; a blood test with 25% relative variability on a repeat draw would be clinically useless.
- Manufacturing and process industries commonly treat CV under 10% as tight control, 10-20% as workable, and above that as a signal to investigate, though the real target should always come from the customer’s tolerance, not a generic rule.
- Agricultural and field research routinely accepts CV up to 20%, and sometimes higher, because the sources of variation (weather, soil, biological variability) are much harder to control than in a closed manufacturing process.
- Service and transactional processes (call handle time, order processing, approval cycle time) often run CVs of 20-40% simply because human-driven work has more natural variation than machine-driven work, and that’s not automatically a defect, it’s a baseline to improve from, not a failure to be alarmed by.
- Finance and investment risk analysis uses CV to compare the risk-adjusted return of different assets: an investment with a higher CV of returns is considered riskier per unit of expected return, regardless of which currency or scale it’s measured in.
The pattern across every one of these fields is the same: CV is only useful as a comparison tool once you’ve picked the right peer group. Comparing your service process’s CV against a laboratory’s CV tells you nothing useful, they’re not solving the same kind of variability problem.
How CV Relates to Process Capability (Cp and Cpk)
Process capability indices like Cp and Cpk answer a related but different question than CV does. Cp and Cpk compare your process’s spread to a fixed specification width set by the customer or engineering requirement; CV compares your process’s spread to its own mean, with no specification involved at all.
That distinction matters because a process can have an excellent (low) CV and still have poor process capability, if its specification limits are tight relative to its natural variation, and vice versa: a process with a high CV can still be perfectly capable if its specification limits are wide. CV describes how variable a process is relative to itself; Cp and Cpk describe how variable a process is relative to what the customer actually needs. Use CV when you’re comparing processes to each other; use Cp and Cpk when you’re comparing a single process to its own specification.
Frequently Asked Questions
Can the Coefficient of Variation be over 100%?
Yes. A CV over 100% simply means the standard deviation is larger than the mean itself, which happens with highly skewed data, data with a lot of near-zero or zero values, or small means with genuinely large spread. It’s a valid result, not an error, but it’s also a strong signal that the data may not be well-described by a mean and standard deviation at all, and a distribution check is worth doing before drawing conclusions.
Can the Coefficient of Variation be negative?
Standard deviation is always zero or positive, so the sign of CV follows the sign of the mean. A negative mean will produce a negative CV, which is one more reason CV is only meaningful for ratio-scale data (like time, length, weight, or cost) that can’t legitimately be negative or doesn’t cross zero in the first place.
What’s a “good” CV for a Six Sigma project?
There isn’t a single number that applies across every project, and claiming otherwise would be oversimplifying a genuinely context-dependent question. The honest approach is to establish a baseline CV for your own process, then track whether your improvement project moves that number down over time, or compare it against a legitimately similar process (same industry, same type of measurement) rather than an arbitrary universal benchmark.
Is CV the same thing as %RSD?
Yes. Relative standard deviation (RSD), usually reported as %RSD, is the same calculation as the Coefficient of Variation. %RSD is simply the more common name for it in laboratory and analytical chemistry contexts, while “Coefficient of Variation” is the more common name in statistics, manufacturing, and Six Sigma contexts. The formula and the interpretation are identical.
Does my data need to be normally distributed to use CV?
No, CV doesn’t require a normal distribution the way some other statistical tools do. It’s purely a descriptive ratio of two sample statistics (standard deviation and mean), not a test with distributional assumptions behind it. That said, CV is most informative and most stable when the data is reasonably symmetric; on a heavily skewed distribution, the mean itself can be a poor summary of “what’s typical,” which in turn makes a CV built on that mean less meaningful, even though the arithmetic still works. If you suspect skew, it’s worth checking the distribution shape first and considering whether the median and a percentile-based spread measure might describe the data more honestly.
Quick Reference
- What it is: standard deviation expressed as a percentage of the mean (also called relative standard deviation).
- Formula: CV = (standard deviation / mean) × 100%.
- Use it when: comparing variability across datasets with different units, scales, or means.
- Don’t use it when: the mean is near zero, the data is interval-scale (no true zero), or the sample size is very small.
- Lower is more consistent; higher is more variable, relative to the mean. There’s no universal cutoff for “good” or “bad,” set thresholds based on your own process and customer requirements.
For more on the building blocks behind this calculation, see Dispersion and Standard Deviation. For how variability measures like this feed into capability analysis, see Process Capability (Cp & Cpk).