The coefficient of variation (CV) is a measure of relative statistical dispersion: the standard deviation expressed as a proportion (usually a percentage) of the mean. Where standard deviation reports spread in the original units of the data, CV strips the units away, which is what makes it useful for comparing variability across variables measured on different scales, in different units, or with very different means — something a raw standard deviation cannot do on its own.
Formula and basic interpretation
CV = SD ÷ mean (multiply by 100 to express as a percentage)
For a sample, that is CV = s / x̄ × 100%; for a population, CV = σ / μ × 100%. A CV of 15% means the standard deviation is 15% of the mean — in practical terms, the typical spread of the data is small relative to its average size. A CV is only defined when the mean is nonzero, and its usefulness depends heavily on the mean not being close to zero, which the limitations section below covers in detail.
CV is one of several measures of dispersion alongside range, interquartile range, variance, standard deviation, and mean absolute deviation. It is the only one of that group that is unitless, which is both its main advantage and the source of its main limitations.
Why relative dispersion matters
Standard deviation alone answers “how spread out is this dataset?” but it cannot answer “which of these two datasets is more variable, relative to what is being measured?” when the two datasets are on different scales or have very different means. A standard deviation of 5 mmHg on blood pressure readings and a standard deviation of 5 kg on body weight readings are not comparable numbers — they are different units entirely. Even within the same unit, a standard deviation of 5 means something different for a variable averaging 20 than for one averaging 2,000.
CV solves this by normalizing spread to the mean, producing a dimensionless figure that can be compared directly: across variables measured in different units, across instruments or methods with different typical outputs, or across groups whose means differ substantially even though the underlying measurement is the same.
Worked examples
Same standard deviation, very different CV
Two illustrative datasets, each with SD = 4, but very different means:
| Dataset | Mean | SD | CV |
|---|---|---|---|
| A | 20 | 4 | 4 ÷ 20 = 20% |
| B | 200 | 4 | 4 ÷ 200 = 2% |
Both datasets have identical absolute spread, but Dataset A is ten times more variable relative to its own scale than Dataset B. Reporting only the shared SD of 4 would hide that difference entirely; the CV exposes it in one number.
Same CV, very different standard deviation
The reverse also happens. Two illustrative datasets, each with a CV of 10%:
| Dataset | Mean | SD | CV |
|---|---|---|---|
| C | 50 | 5 | 5 ÷ 50 = 10% |
| D | 500 | 50 | 50 ÷ 500 = 10% |
Dataset D has a standard deviation ten times larger than Dataset C, yet the two are equally variable relative to their own means. This is the specific comparison CV is designed to make possible: a fair read of “how variable is this, proportionally,” independent of the scale the variable happens to be measured on.
Critical limitations — read before using CV
CV is convenient, but it is also one of the more frequently misapplied descriptive statistics. Four limitations matter enough to check before reporting it:
- Unstable and often meaningless when the mean is near zero. Because the mean sits in the denominator, a mean close to zero makes the CV extremely large, extremely sensitive to small changes in the mean, or undefined outright at exactly zero. A variable that legitimately takes values near or on both sides of zero (a temperature anomaly, a profit/loss figure, a change score) is a poor candidate for CV regardless of how meaningful its standard deviation is.
- Invalid for interval scales with an arbitrary zero. CV is only meaningful for ratio-scale data, where zero represents a true absence of the quantity. Temperature in Celsius or Fahrenheit is the standard teaching example of why: 0°C is an arbitrary reference point (the freezing point of water), not “no temperature,” so the CV of a set of Celsius readings changes if the same data is converted to Fahrenheit — and can even flip sign or explode toward infinity if the readings straddle 0°C. A ratio-scale equivalent, like temperature in Kelvin, has a true zero and does not have this problem. The general rule: before reporting a CV, confirm the variable has a real zero point that means “none of this quantity,” not just a conventional reference value.
- Requires a true ratio scale with a meaningful zero. This follows from the point above but is worth stating on its own: mass, length, concentration, reaction time, and count data are ratio-scale and safe for CV. Standardized test scores, IQ scores, dates, and most Likert-type scales are not.
- Sensitive to outliers through both terms. An extreme value pulls the mean and inflates the standard deviation simultaneously, so a single outlier can distort CV more than it distorts either component alone. As with any mean/SD-based statistic, check the shape of the distribution — a histogram is the fastest way — before relying on CV for skewed or outlier-prone data.
Where CV is standard practice: assay precision and method validation
Despite those limitations, CV is genuinely standard practice in one area central to research operations: reporting the precision of an analytical assay or measurement method, where the measured quantity (concentration, count, signal intensity) is reliably ratio-scale and comfortably away from zero.
- Intra-assay CV (repeatability) — the CV calculated across replicate measurements of the same sample run together, within a single assay run. It captures the precision of the measurement process itself under the most favorable conditions (same operator, same reagents, same day).
- Inter-assay CV (intermediate precision / reproducibility) — the CV calculated across replicate measurements of the same sample run on different days, batches, operators, or instrument calibrations. It is typically larger than intra-assay CV because it captures additional real-world sources of variability that a single run does not.
Laboratories and method-validation protocols commonly treat lower intra- and inter-assay CVs as an informal marker of good precision — numbers such as an intra-assay CV under roughly 5% and an inter-assay CV under roughly 10–15% are widely cited as conventional benchmarks of an acceptable assay in many bioanalytical and immunoassay contexts. These are field conventions and internal quality benchmarks, not universal regulatory thresholds — acceptable CV ranges vary by analyte, assay type, concentration near the limit of detection, and the specific validation protocol or guidance a lab or sponsor has adopted, and different assay platforms and disciplines set their own working thresholds rather than following one fixed rule. Always confirm the specific acceptance criteria that apply to a given method or regulatory context rather than assuming a single percentage applies universally.
CV also underpins general analytical QC and repeatability reporting more broadly — tracking a control sample’s CV over time (a run chart or Levey-Jennings-style chart) is a standard way to monitor whether an assay’s precision is drifting, independent of any single acceptance threshold. This overlaps with, but is distinct from, agreement statistics like the intraclass correlation coefficient, which assess consistency between raters or instruments rather than the relative spread of repeated measurements from one source.
CV vs. SD vs. variance vs. IQR
| Statistic | Units | Comparable across different scales? | Sensitive to outliers? | Typical use |
|---|---|---|---|---|
| Standard deviation (SD) | Same as the data | No | Yes | Default spread statistic reported alongside a mean |
| Variance | Squared units | No | Yes (more so than SD) | Building block for ANOVA, regression, and other inferential methods |
| Coefficient of variation (CV) | Unitless (%) | Yes, for ratio-scale data | Yes (through both mean and SD) | Comparing relative variability across scales; assay/method precision |
| Interquartile range (IQR) | Same as the data | No | No (robust) | Skewed data or data with outliers, paired with the median |
See CASRAI’s full guide to descriptive statistics for how these fit alongside central tendency and shape measures, and the dedicated guide to interquartile range and the five-number summary for the outlier-robust alternative.
Geometric CV for log-normal data
Many measurements common in bioanalysis and pharmacokinetics — drug concentrations, titers, and other quantities that are naturally multiplicative rather than additive in how they vary — are better approximated by a log-normal distribution than a normal one. Applying the ordinary (arithmetic) CV formula to log-normal data can be misleading, because the arithmetic mean and SD are themselves distorted by the distribution’s skew. The geometric CV (GCV) is the standard adjustment: it is derived from the standard deviation of the log-transformed data and is typically calculated as GCV = √(es² − 1) × 100%, where s² is the variance of the natural-log-transformed values. Geometric CV is the conventional way precision and variability are reported in pharmacokinetic and bioequivalence work specifically because the underlying data is log-normal by nature, not because it is a universal replacement for the ordinary CV.
Relationship to signal-to-noise
CV is closely related to, and is functionally the inverse of, the signal-to-noise ratio (SNR) used in measurement and instrumentation contexts: SNR is commonly defined as mean ÷ SD, the exact reciprocal of the CV ratio (before the CV’s conventional multiplication by 100). A low CV corresponds to a high signal-to-noise ratio — a strong, stable signal relative to its own variability — and a high CV corresponds to a low signal-to-noise ratio, where noise (variability) makes up a larger share of the measured value. Reporting CV and describing a measurement’s signal-to-noise characteristics are, in this sense, two framings of the same underlying ratio.
Calculating CV in R, Python, and Excel
None of the three has a single built-in CV function — in each case it is calculated directly from the mean and standard deviation:
- R:
sd(x) / mean(x) * 100. Therasterand several other add-on packages provide acv()convenience function, but the one-line base-R calculation is just as common. - Python:
x.std() / x.mean() * 100on a pandas Series or NumPy array (useddof=1innp.std()for the sample standard deviation, matching R’s default). SciPy also providesscipy.stats.variation(x), which returns the CV directly as a proportion (multiply by 100 for a percentage). - Excel:
=STDEV.S(range)/AVERAGE(range), formatted as a percentage, or multiplied by 100 for an explicit percentage value.
Frequently asked questions
What is a “good” coefficient of variation?
There is no single universal threshold — what counts as a low or high CV depends entirely on the field and the type of measurement. In assay precision and method validation, an intra-assay CV under roughly 5% and an inter-assay CV under roughly 10–15% are commonly treated as informal benchmarks of good precision in many contexts, but these are conventions adopted by individual labs, platforms, or validation protocols, not fixed regulatory limits, and other fields set entirely different working thresholds.
Can CV be negative?
Standard deviation is never negative, so CV is negative only when the mean itself is negative, which produces a CV that is difficult to interpret meaningfully. This is one more symptom of the near-zero-mean problem covered above — CV is best reserved for ratio-scale variables with a mean that is comfortably positive.
Why can’t I use CV on Celsius temperature data?
Celsius (like Fahrenheit) is an interval scale with an arbitrary zero point — 0°C does not mean “no temperature,” it means “the freezing point of water.” Because CV depends on where zero falls, converting the same temperatures to Fahrenheit changes the CV, which makes the statistic meaningless as a measure of the data’s actual variability. Kelvin, which has a true zero (absolute zero), does not have this problem.
Is CV the same as relative standard deviation (RSD)?
Yes — relative standard deviation and coefficient of variation are the same statistic under different names, and “%RSD” is a common way to see it labeled specifically in analytical chemistry and laboratory reporting.
How is CV different from a confidence interval or standard error?
CV describes the spread of individual observations relative to their mean; standard error and confidence intervals describe the precision of an estimate (such as a sample mean) as an approximation of a population value. A dataset can have a large CV (individual values vary widely) while still producing a precise estimate of the mean if the sample size is large enough to shrink the standard error — the two statistics answer different questions and are not interchangeable.
