Interpreting Variance and Standard Deviation

Two datasets can share the same mean and still describe very different situations. A process whose daily output is always close to 500 units and one that swings between 300 and 700 have the same average, but they call for very different planning. Variance and standard deviation are the standard measures of that difference: how far, typically, observations lie from their mean.

Both are among the most widely reported statistics, and among the most frequently misread. This article explains how they are computed, why the sample version divides by n − 1, how to compare spread across datasets on different scales, and which interpretive rules of thumb are valid for which kinds of data.

The variance is the average squared deviation from the mean. The standard deviation is its square root. For a sample of n observations, the sample versions are:

Squaring the deviations does two things. It makes every deviation count positively (raw deviations from the mean always sum to zero), and it gives large deviations disproportionate weight: a deviation of 4 contributes 16, four times as much as a deviation of 2. This is why the variance and standard deviation are sensitive to outliers.

Worked example

Take the eight values 2, 4, 4, 4, 5, 5, 7, 9. Their sum is 40, so x̄ = 5.

The sum of squared deviations is 32. The sample variance is 32 / 7 ≈ 4.571, and the sample standard deviation is √4.571 ≈ 2.138. If these eight values were the entire population of interest, the population variance would be 32 / 8 = 4 and the population standard deviation exactly 2.

Why report both?

Variance is measured in squared units. If the data are in MWh, the variance is in MWh², which has no direct interpretation. The standard deviation is in the original units and can be read alongside the mean: a mean of 5 with a standard deviation of about 2.1. Variance remains useful because it behaves well algebraically. For independent quantities, variances add; standard deviations do not. Two independent components with standard deviations 3 and 4 produce a total with variance 9 + 16 = 25 and standard deviation 5, not 7.

The population variance σ² divides by the number of values N and uses the true mean μ. A sample has to use its own mean x̄ instead, and x̄ is by construction the value that minimizes the sum of squared deviations for that sample. Deviations from x̄ are therefore systematically a little smaller than deviations from μ would be. Dividing by n underestimates σ² on average; dividing by n − 1 corrects this exactly. The correction is named after Friedrich Bessel.

An exhaustive check

Consider a population of three equally likely values {1, 3, 5}. Its mean is 3 and its variance is σ² = (4 + 0 + 4) / 3 = 8/3 ≈ 2.667. Draw samples of size n = 2 with replacement: there are nine equally likely ordered pairs. For a pair (a, b), the sum of squared deviations from the sample mean is (a − b)² / 2.

Two further points follow. First, the correction matters most for small samples: the ratio between the two standard deviations is √(n / (n − 1)), which is about 1.069 for n = 8, 1.017 for n = 30 and 1.0005 for n = 1,000. Second, unbiasedness does not carry over to the square root. In the exhaustive example the sample standard deviations are 0, √2 and √8, whose average is about 1.257, below the population value √(8/3) ≈ 1.633. The sample standard deviation is slightly biased low even with n − 1. In practice this is accepted, and s with n − 1 is the conventional choice.

Insider Magazin reports the sample variance and standard deviation (n − 1) throughout, because analysed data are almost always a sample of a longer process, and the n − 1 version is what most statistical software reports by default.

A standard deviation of 8 minutes is large for a process that averages 10 minutes and small for one that averages 400. To compare variability across groups with very different levels, divide by the mean:

For the worked example, CV = 2.138 / 5 ≈ 0.428, or 42.8%. As an illustration of its use: a channel with mean response time 4 minutes and standard deviation 2 minutes has CV = 50%, while one with mean 40 minutes and standard deviation 8 minutes has CV = 20%. The second has four times the absolute spread but is relatively more consistent.

The CV is meaningful only for ratio-scale data: positive quantities with a true zero, such as durations, counts, amounts or energy. It is not meaningful for temperatures in degrees Celsius, for differences, or for any variable whose mean can be near zero, where the ratio explodes. It inherits the outlier sensitivity of the standard deviation.

A standard deviation is most useful when you can translate it into a statement about how many observations lie near the mean. Two results are commonly used, and they apply under very different conditions.

Chebyshev's inequality: always valid

Chebyshev's inequality holds for every distribution with a finite variance, whatever its shape. At least 75% of values lie within 2 standard deviations of the mean, and at least 88.9% within 3. The bound is deliberately conservative: it must hold even for the most awkward distributions. In the eight-value example, the interval x̄ ± 2s = 5 ± 4.28 contains all eight values, comfortably above the guaranteed 75%.

The 68–95–99.7 rule: only for approximately normal data

For a normal distribution, about 68.3% of values lie within 1 standard deviation of the mean, 95.4% within 2 and 99.7% within 3. This empirical rule is much sharper than Chebyshev, but it relies on the specific shape of the normal curve. Applied to skewed or heavy-tailed data, it gives wrong answers.

The exponential column shows how the normal rule can mislead for right-skewed data such as waiting times. Far more than 68% of values lie within one standard deviation, because the lower side is cut off at zero, and the region beyond 3σ holds about 1.8% of values rather than the 0.27% a normal distribution would leave outside μ ± 3σ. Events that look like "three-sigma rarities" under a normal assumption are roughly seven times as common in this case.

Before using the 68–95–99.7 rule, check the histogram and the skewness. If the data are clearly skewed, either use Chebyshev's bound, describe the spread with quantiles (median and IQR, or the 5th and 95th percentiles), or analyse a transformation such as the logarithm, under which log-normal data become approximately normal.

• Confusing standard deviation with standard error. The standard deviation describes the spread of individual observations. The standard error of the mean, s / √n, describes the uncertainty of the mean itself and shrinks as n grows.
• Mixing n and n − 1. Spreadsheet functions such as STDEV.P and STDEV.S differ in exactly this way. Check which one a figure uses before comparing it with another source.
• Adding standard deviations. Variances of independent components add; standard deviations do not.
• Computing variance with the one-pass formula. The algebraically equivalent expression (Σ xᵢ² − n · x̄²) / (n − 1) suffers from cancellation in floating-point arithmetic when the mean is large relative to the spread. For the values 1,000,000,001, 1,000,000,002 and 1,000,000,003 it returns 0 in double precision, while the correct sample variance is 1. Compute the mean first, then sum squared deviations, as the workspace does.
• Comparing standard deviations across scales. Use the coefficient of variation for positive ratio-scale data, or compare standardized differences such as Cohen's d when comparing two groups.

The overview indicators show the sample standard deviation (n − 1) for the current selection. The distributions summary adds variance, coefficient of variation, IQR and skewness for each category, with box plots for a robust view of spread. The comparison view reports Cohen's d using the pooled standard deviation, as a descriptive effect size rather than a significance test.

Compare the spread of the three channels in the support-response dataset in the interactive demo. Exact definitions are listed in the methodology, and the documentation explains where each statistic appears.