How To Calculate A Z-Score

Every measurement comes with its own scale, and comparing across scales is where intuition fails. An exam score of 85 means one thing in a class averaging 75 and something quite different in a class averaging 82. Dividing away both the location and the spread leaves a pure number — how far from typical this value sits, expressed in the units of that population's own variability. That number is the z-score, and it travels well.

The calculation itself is two steps. Subtracting the mean recentres the distribution so that the average becomes zero; dividing by the standard deviation rescales so that one unit equals one standard deviation of spread. Nothing about the shape changes in this transformation — an asymmetric distribution stays asymmetric — which is why z-scores carry no information about normality no matter how often they are described as if they did.

Reading the sign is straightforward but worth stating plainly. Positive means above the mean, negative below, and the magnitude answers how many standard deviations' worth. A z of 2 is a value two SDs above average; a z of -1.3 sits that far below. The absolute value answers a different and often more useful question: how unusual is this, irrespective of direction.

A common misuse is treating z as though it automatically yields a percentile. Converting a z-score to a percentage of the population below it requires knowing the shape of the distribution — for a normal one, a table or an error function does the job, and jumping to "z = 1 means the 84th percentile" silently assumes exactly that bell shape. On skewed data such as household income or hospital stays, the same z-score corresponds to a very different percentile.

The empirical rule gives rough expectations when the distribution is at least approximately bell-shaped: roughly 68 percent of values fall within one standard deviation, 95 percent within two, and 99.7 within three. Those figures are properties of the normal curve, not of z-scores as such. Using them as a general guide is fine; quoting them as precise statements about your particular data is not.

Outlier thresholds built on z-scores inherit the same sensitivity problem. A familiar rule flags anything beyond three standard deviations, but the mean and standard deviation that go into the calculation are themselves pulled by extreme values. A single wild observation inflates the SD and thereby reduces every z-score, which can hide the very point you were looking for. Robust alternatives based on the median and interquartile range behave better here.

The standard deviation chosen must match the population the value belongs to. Mixing a sample SD into a claim about a population, or using last year's spread with this year's mean, produces a number with no clear interpretation. This is the error that most often survives review, because the arithmetic is correct even though the comparison is not.

There is also a subtlety about which SD to divide by. If your data is the entire population, divide by the population SD with n in the formula. If it is a sample estimating an unknown population spread, the unbiased version dividing by n - 1 is usually intended. The two differ materially in small samples — with five observations the difference in the SD is already more than 11 percent — and choosing wrongly shifts every z-score by that same factor.

None of this requires the original units to be meaningful on their own. Because they cancel, z-scores let you ask whether a commute is more unusual than a temperature reading, or whether a fund's return was more extreme than a player's scoring season. That comparability is the entire reason the statistic exists, and it is worth protecting by being explicit about which mean and which standard deviation produced it.