Variance

Variance measures how spread out a dataset's values are around their mean, computed as the average of each value's squared deviation from the mean. A variance of zero means every value in the dataset is identical; a larger variance means values are, on average, farther from the mean — and the squaring step, rather than simply averaging the raw deviations, is the mathematical core of the measure, included specifically because deviations above and below the mean always sum to exactly zero by the mean's own definition, meaning a simple average of raw deviations would always equal zero regardless of how spread out the data actually was, telling nothing about spread at all.

The squaring step that makes variance mathematically well-behaved also makes it awkward to interpret directly, because squaring a deviation also squares its units — the variance of a dataset measured in dollars is expressed in squared dollars, a unit with no intuitive real-world meaning, which is precisely the practical reason standard deviation (the square root of variance) exists: taking the square root returns the measure to the original unit, producing a number that can be interpreted directly as "how far, on average, values typically sit from the mean," in dollars rather than squared dollars. Variance itself remains central to statistical theory specifically because its mathematical properties (it can be decomposed, combined across independent variables by simple addition, and connects directly to probability distributions) make it more tractable for further calculation than standard deviation is, even though standard deviation is the figure more commonly reported for human interpretation.

Population variance and sample variance are computed with a subtly different formula, and the difference matters: population variance divides the sum of squared deviations by n (the number of observations), while sample variance conventionally divides by n minus 1 — a correction known as Bessel's correction, which exists specifically because using a sample's own mean to calculate deviations (rather than the true, usually unknown population mean) systematically underestimates the true population variance, and dividing by n minus 1 instead of n corrects for that bias, producing an unbiased estimate of the population variance from sample data. This correction becomes less consequential as sample size grows (the difference between dividing by n versus n minus 1 shrinks toward negligible), but it's a standard, deliberately applied convention rather than an arbitrary stylistic choice, and most statistical software defaults to the n minus 1 version specifically because sample data, not full population data, is what's actually being analyzed in the overwhelming majority of real applications.

Variance's squaring mechanism has a direct and consequential side effect: large deviations from the mean are weighted disproportionately more heavily than small ones, since squaring a value twice as large produces a result four times as large, not just twice — which is exactly why variance (and standard deviation, inheriting the same property through its square-root relationship) is more sensitive to outliers than a measure like the mean absolute deviation (which simply averages the absolute value of deviations without squaring them), a tradeoff variance accepts deliberately in exchange for the mathematical properties that make it foundational to regression, hypothesis testing, and most of the rest of classical inferential statistics.

Read variance as the foundational measure of spread that underlies most of classical statistics' mathematical machinery, even though it's rarely the number actually reported to a human audience: its squaring step is what makes deviations above and below the mean both contribute positively rather than canceling out, its awkward squared units are exactly why standard deviation exists as its more interpretable counterpart, and the population-versus-sample distinction (via Bessel's correction) is a deliberate, necessary adjustment for the fact that real-world variance calculations are almost always performed on a sample rather than a full, completely known population.