Variance and standard deviation: spread, units and denominators

Start with deviations from a stated mean

Take four hypothetical measurements in units of length: 2, 4, 4 and 6. Their arithmetic mean is 4. The deviations are −2, 0, 0 and 2, which sum to zero. Simply averaging signed deviations would therefore hide the spread. Squaring them gives 4, 0, 0 and 4. The NIST discussion of scale measures explains why squared deviations emphasise values farther from the centre. Before interpreting a variance, state whether the numbers are all possible values in a model or a sample drawn from something larger.

Distinguish two legitimate denominators

If the four values form a complete equally weighted distribution, its variance is the squared-deviation total divided by four: 8/4=2. If they are an independent, identically distributed sample used to estimate an unknown population variance, the familiar unbiased sample estimator divides by n−1, giving 8/3. These are different calculations answering different questions. OpenStax’s distributional formula weights deviations by probabilities, while the NIST sample formula uses n−1. Do not switch denominators merely to obtain a preferred result or conceal which quantity was calculated.

Inspect the calculation line by line

The complete-distribution standard deviation is √2, approximately 1.414 length units. The sample standard deviation is √(8/3), approximately 1.633 length units. The corresponding variances have squared length units. Reporting the units makes it harder to confuse a variance of two with a typical displacement of two. Neither statistic says that every measurement lies exactly one standard deviation from the mean.

Keep standard deviation and standard error apart

Standard deviation describes variability of individual observations under the stated model. Standard error describes variability of an estimator across repeated samples. For an independent sample with common finite variance, the standard error of the sample mean is σ/√n, estimated by s/√n when appropriate. More observations can reduce uncertainty in the mean without making the underlying observations less variable. A narrow error bar around an estimated average therefore does not imply that individual experiences, waiting times or measurements all lie close to that average.

Understand how transformations change spread

Adding a fixed constant to every value shifts the mean but leaves variance unchanged. Multiplying every value by a constant c multiplies variance by c² and standard deviation by |c|. The Stanford reference on variance provides the algebraic framework. Converting lengths from metres to centimetres multiplies standard deviation by 100 and variance by 10,000. A comparison across reports is meaningless if one uses metres and the other centimetres without conversion. Keep transformations in the calculation log so units can be traced back to the original measurements.

Recognise sensitivity to extreme observations

A very large value can dominate a squared-deviation total. That is a mathematical property, not automatically proof that the observation should be deleted. First check whether the value is a transcription error, a genuine rare case or part of a different population. The median absolute deviation or interquartile range can answer complementary questions about central spread, but replacing one measure with another changes the interpretation. Report exclusions and their reasons. Removing inconvenient observations without disclosure can make a process appear more stable than the collected evidence supports.

Keep weights visible when observations represent different shares

The four-value example gives each observation equal weight. If the values instead represent categories with different probabilities, a distributional variance uses those probabilities in both its mean and squared-deviation average. Repeating a category label in a presentation table does not establish its population share. Conversely, a category represented by one row can account for many observations. Record the weighting scheme and its source before comparing spread across datasets. Equal-row weighting and population weighting answer different questions, even when both calculations are arithmetically correct.

Spread calculation checklist

• State the variable and units.
• Identify distributional variance or sample estimation.
• Show the mean and squared-deviation total.
• Explain the denominator.
• Keep variance and standard deviation in their proper units.
• Separate observation spread from standard error.
• Document transformations and exclusions.

A compact verification uses the alternative identity E(X²)−E(X)² for a specified distribution. For these four equally weighted values, E(X²)=18 and E(X)²=16, reproducing variance 2. This identity is useful for checking a small example, although numerical software should handle large nearly equal quantities carefully to avoid cancellation error. Retain adequate precision and round only the figures displayed to readers.

Sources and verification

• Measures of scale — NIST/SEMATECH e-Handbook — 2026-10-23
• Mean, expected value and standard deviation — OpenStax, Rice University — 2026-10-23
• Probability and statistics refresher — Stanford University teaching materials — 2026-10-23