How To Calculate Standard Deviation

Two data sets can share a mean and look completely different. Ten values clustered tightly around fifty and ten values scattered from five to ninety-five both average to around fifty, and no average will distinguish them. Standard deviation fills exactly that gap: it is a single number describing how far values typically sit from the centre, expressed in the same units as the measurements themselves.

The construction has three moves, and each one solves a specific problem. First subtract the mean, giving signed distances. Then square them. Then average and take the square root at the end. The squaring exists because summing signed deviations always gives exactly zero by definition — the positive and negative distances cancel — so some way of removing the signs is required before averaging means anything.

Squaring is not the only option and it is worth knowing why it won. Taking absolute values would be more literal and would give the mean absolute deviation, but absolute values are algebraically awkward: the absolute value function has a kink at zero that breaks the clean calculus underlying almost every statistical procedure. Squares are smooth, differentiable, and lead naturally to the variance-additivity results that make statistical inference possible.

The cost is interpretability, and it is the reason the square root comes back at the end. Squared deviations are measured in squared units — dollars squared, seconds squared — which nobody can reason about. Dividing by the count gives the variance, still in squared units, and taking the square root returns values to the original measurement scale. Variance is the mathematically convenient quantity; standard deviation is the human-readable one.

Then there is the divisor, which is where genuine subtlety lives. Dividing by n describes the data you actually have, treating it as the entire population of interest. Dividing by n-1 estimates the spread of the larger population you drew the sample from. The intuition: any sample mean is pulled slightly toward the sample's own values, making it closer to them than the true population mean would be. Squared deviations around the sample mean therefore come out systematically too small, and replacing n with n-1 removes almost exactly that bias.

The size of the correction is anything but academic at small counts. Dividing by 5 instead of 4 here gives a variance of 29.36 rather than 36.70 — a twenty percent difference, which is why sample versus population is not a rounding question. As n grows the two converge: at n = 100 the difference is one percent, and with large samples the choice barely matters. Below about thirty observations it matters a great deal.

Reading the number requires knowing the shape. If the distribution is roughly bell-shaped, about 68% of observations fall within one standard deviation of the mean and about 95% within two — the familiar empirical rule. For our data that means roughly 68% of values sit between 10.74 and 22.86. That interpretation depends heavily on symmetry, and for skewed or heavy-tailed distributions the rule can be badly wrong even though the calculation is identical.

This sensitivity explains standard deviation's main weakness: it treats a value far above the mean and one equally far below as identical contributions, which is misleading for one-sided risks. Investment analysts distinguish downside deviation for exactly this reason, and any situation where losses and gains are not symmetric deserves the same scrutiny.

Finally, the units are the feature that makes it practical. A standard deviation of 6.06 for data measured in minutes means about six minutes of typical spread, directly comparable to the mean of 16.8 minutes. The coefficient of variation — standard deviation divided by the mean, here 36.06% — removes units entirely and allows comparison between quantities that could never otherwise be put side by side.