Variance and Standard Deviation

The variance of a set of measurements is the average of the squared deviations of each observation from the mean. Formally, for a population of size N with mean μ (mu), the population variance is:

Squaring serves two purposes. First, it makes all deviations positive, so they cannot cancel out (a raw sum of deviations from the mean is always zero — that is a defining algebraic property of the mean). Second, squaring weights large deviations disproportionately, which is often biologically sensible: an individual far from the mean phenotype represents more "unusual" variation than the sum of several individuals close to the mean would suggest. The standard deviation, σ (sigma), is the square root of the variance:

Taking the square root undoes the squaring of units, returning the measure to the original scale — millimeters instead of millimeters², beats per minute instead of (beats per minute)². This is the crucial practical reason SD, not variance, is the number biologists actually report and interpret: it is directly comparable to the mean and to the raw data. In practice, biologists almost never have access to an entire population — a full census of every finch on every Galápagos island, every neuron in a species' brain, every red blood cell in a patient's body. Instead they work with a sample of size n drawn from that population, and they use the sample to estimate the population parameters (see Parameters versus Statistics). This distinction — μ and σ² for populations, versus $\bar{x}$ (x-bar) and s² for samples — is not mere notation. It changes the formula itself, because estimating variance from a sample average with the sample's own mean systematically underestimates the true population variance. Correcting for that bias is the single most important technical wrinkle in this entire topic.

Step 1: Center the data. Compute the mean of the data set, $\bar{x} = \frac{\sum x_i}{n}$, using the arithmetic mean as covered in The Mean, Median, and Mode. This is the reference point from which spread is measured.

Step 2: Compute deviations. For every observation, calculate $(x_i - \bar{x})$, the signed distance from the mean. Some will be positive (above the mean), some negative (below).

Step 3: Square each deviation. Compute $(x_i - \bar{x})^2$ for every observation. Squaring removes the sign so that above-mean and below-mean deviations both contribute positively to the total spread, and it inflates the influence of outliers relative to typical points.

Step 4: Sum the squared deviations. This total, $\sum (x_i - \bar{x})^2$, is called the sum of squares (SS). It is the raw currency of variability that reappears throughout statistics — in ANOVA, regression, and beyond.

Step 5: Divide by the appropriate denominator. This is where population and sample formulas diverge:

Step 6: Take the square root. $s = \sqrt{s^2}$ (sample SD) or $\sigma = \sqrt{\sigma^2}$ (population SD), converting back to the original units.

Why n − 1? The sample mean $\bar{x}$ is calculated from the same data whose spread is being measured, and $\bar{x}$ is mathematically guaranteed to be the point that minimizes the sum of squared deviations for that particular sample. Using the true population mean μ (which is unknown) would generally give a larger sum of squares. Because $\bar{x}$ is "tuned" to the sample, using it slightly understates the true spread around μ. Dividing by a smaller number, n − 1 instead of n, inflates the estimate just enough to compensate on average. This is Bessel's correction, and it makes $s^2$ an unbiased estimator of σ²: averaged over many repeated samples, $s^2$ neither systematically over- nor under-estimates the population variance, whereas dividing by n would systematically underestimate it. This loss of one degree of freedom (n − 1 "degrees of freedom") occurs because one piece of information — the mean itself — has already been extracted from the data and used up in the calculation. The effect of the correction is largest for small samples (dividing by 4 instead of 5 matters a lot) and vanishes as n grows large (dividing by 999 versus 1000 barely matters).

Step 7: Interpret via the empirical rule. For data that are approximately normally distributed (a bell-shaped curve — see the Gaussian discussion under measurement error), the SD has a direct probabilistic interpretation:

This "68–95–99.7 rule" is only exact for a true normal distribution, but it gives useful intuition for any roughly symmetric, unimodal data set — a quick way to sanity-check whether an observation is unusually extreme.

Unlike a discovery such as the mean itself — which is ancient — the formal concept of variance was defined and named by one specific act of statistical invention, and the n − 1 correction was independently justified and empirically vindicated in classic ways.

Karl Pearson (1893–1894) coined "standard deviation." In lectures and papers on the mathematical theory of evolution, Pearson introduced the term "standard deviation" for the root-mean-square deviation from the mean, building on earlier work by Gauss and by Pearson's own predecessors on the "probable error." Pearson was explicitly motivated by Francis Galton's biometric program — quantifying variation in human and animal traits to study heredity and natural selection. Before Pearson's clean formalization, biologists and astronomers used related but less standardized measures ("mean error," "probable error"), making comparisons across studies difficult. Pearson's SD gave the field a single, mathematically well-behaved yardstick.

Ronald Fisher (1920s) rigorously established degrees of freedom and the n − 1 correction. Fisher, working on the statistical analysis of agricultural field trials at Rothamsted Experimental Station, proved formally that dividing the sum of squared deviations from the sample mean by n gives a biased (systematically too-small) estimate of the population variance, and that dividing by n − 1 removes this bias exactly. His reasoning: the sample mean is itself estimated from the data, using up one "degree of freedom," so only n − 1 independent pieces of information remain to estimate spread. This is not merely a rule of thumb — it can be proven algebraically by expanding $E[\sum(x_i - \bar{x})^2]$ and showing it equals $(n-1)\sigma^2$ in expectation, a result confirmed repeatedly by simulation: if you draw thousands of random samples of size n from a known population, average their "divide by n" variances, the average falls short of the true σ²; average their "divide by n − 1" variances, and the average converges exactly on σ².

Empirical demonstration by resampling. The cleanest modern "experiment" proving the bias is a simulation, and it is worth describing because it is exactly reproducible and routinely used in statistics instruction: take a population with a known variance (for instance, generate 10 million values from a normal distribution with σ² = 4). Repeatedly draw samples of size n = 5, and for each sample compute both $\frac{SS}{n}$ and $\frac{SS}{n-1}$. Averaged over, say, 100,000 such samples, $\frac{SS}{n}$ converges to approximately 3.2 (biased low, roughly $\frac{n-1}{n}\sigma^2$), while $\frac{SS}{n-1}$ converges to almost exactly 4.0. This computational demonstration — impossible before electronic computing but trivial today — is the direct, repeatable confirmation of Fisher's 1920s algebraic proof, and it is why every statistical software package (R's `var()`, Excel's `VAR.S`) defaults to n − 1 for sample data.

Biological application that made variance indispensable: R. A. Fisher's analysis of variance (ANOVA). Fisher's work at Rothamsted on comparing crop yields across fertilizer treatments required partitioning total variability into components — variation between treatment groups versus variation within groups. This partitioning is only possible because sums of squares (built from the same squared-deviation logic as variance) can be added and subtracted meaningfully, unlike measures based on absolute deviations. ANOVA's success in agricultural and later biomedical experimentation is, historically, the strongest practical vindication of building spread measures on squared deviations rather than alternatives like mean absolute deviation.

Example 1 — a genetics-adjacent physiological measurement: litter sizes in a mouse colony. A researcher records the number of pups in seven litters from a mutant mouse line suspected of having reduced, but more erratic, litter sizes:

Step 1 — mean:

Step 2/3 — deviations and squares:

Step 4 — sum of squares:

Step 5 — this is a sample of litters (not every litter this mouse line will ever produce), so divide by n − 1 = 6:

Step 6 — standard deviation:

Interpretation: mean litter size is about 6.9 pups, but the typical litter deviates from that mean by roughly 2.8 pups — a large relative spread, consistent with the "erratic" litter sizes the researcher suspected, driven heavily by that one litter of 12 (which alone contributes more than half the total sum of squares — a useful illustration of how squaring amplifies the influence of extreme values).

Example 2 — clinical scenario: comparing variability of blood pressure control between two drugs. A trial measures systolic blood pressure reduction (mmHg) in patients on two antihypertensive drugs, n = 6 patients each: Drug A: 8, 10, 9, 11, 10, 12 Drug B: 2, 18, 4, 16, 3, 17 Both drugs might have similar means. Check:

Identical means — a clinician looking only at average effect would call these drugs equivalent. Now compute spread. Drug A deviations squared: $(-2)^2, 0^2, (-1)^2, 1^2, 0^2, 2^2 = 4,0,1,1,0,4$; SS = 10

Drug B deviations squared: $(-8)^2, 8^2, (-6)^2, 6^2, (-7)^2, 7^2 = 64,64,36,36,49,49$; SS = 298

Interpretation: Drug A reduces blood pressure by 10 mmHg with tight consistency (SD ≈ 1.4 mmHg) — nearly every patient responds predictably. Drug B has the same average effect but wildly inconsistent responses (SD ≈ 7.7 mmHg) — some patients barely respond (2–4 mmHg reduction) while others respond dramatically (16–18 mmHg reduction). A clinician choosing based on the mean alone would miss this crucial difference: Drug A is the more predictable choice for an individual patient, even though "on average" the drugs look identical. This is precisely why variance and SD, not just the mean, are reported in clinical trial results.

Clinical trials and drug regulation. Regulatory agencies require not just the mean effect of a treatment but its variability, since a drug with a large SD in response may need dose adjustment per patient or reveal a subgroup of non-responders — exactly the logic of the blood pressure example above, played out routinely in Phase II/III trial reporting.

Genomics and quality control. Gene expression measurements (RNA-seq, microarrays) are normalized and filtered using variance: genes with near-zero variance across samples are uninformative and are routinely dropped before differential expression analysis or clustering, while variance-stabilizing transformations correct for the fact that raw expression variance often scales with mean expression.

Conservation biology. Population viability analyses use the variance in population growth rate (not just its mean) to predict extinction risk — high-variance populations are far more vulnerable to stochastic extinction even with an identical average growth rate, directly paralleling the Drug A/Drug B logic above.

Agriculture and animal breeding. Selective breeding programs track both the mean and the variance (heritable component) of traits like milk yield or growth rate; a trait with low variance offers little for selection to act on, connecting directly back to Fisher's original agricultural motivation for developing these tools.

Laboratory quality control. Repeated measurements of a control sample (e.g., a calibration standard on a spectrophotometer or blood analyzer) are tracked via their SD to monitor instrument precision; a sudden rise in SD signals a malfunctioning instrument even if the mean reading looks unchanged.

Public health surveillance. Age-adjusted rates and their variability across regions or time periods rely on variance calculations to distinguish a true trend (e.g., rising disease incidence) from noise expected under random sampling — the basis for confidence intervals reported in epidemiological bulletins.

Ask the tutor: If two samples have very different sizes, how should I decide whether their difference in standard deviations is a real biological difference or just sampling noise?