A finishing time of 23.5 minutes and one of 51 minutes. A price of $452k and one of $1,050k. A height in centimetres and a weight in kilograms. These pairs cannot be compared — not because the arithmetic is hard, but because the numbers are measured on different rulers. Standardizing fixes that with two operations: subtract the mean, then divide by the standard deviation. What comes out is a z-score, a pure number with no units at all, and pure numbers can be compared. Move the sliders below and watch both values land on a single shared scale.
Two distributions, two different rulers
A
B
The same two values on one ruler
within 1 SDwithin 2 SDvalue Avalue B
From a z-score to a percentile
Z-scoresmove a slider to begin
—
—
×1.00
Send A to a chosen z
Report the percentile as
z on A—
z on B—
A — at or below—
B — at or below—
A − its mean—raw units
B − its mean—raw units
gap in z—standard deviations
Read z = (x − mean) ÷ SD as two separate moves and it stops being a formula to memorize. Subtracting the mean recenters the distribution so that its middle sits at zero; what is left, x − mean, is the deviation — how far this value is from typical, and on which side. Dividing by the standard deviation rescales it so that one SD becomes one unit of the new scale. And here is the part worth pausing on: the numerator is in minutes and the denominator is in minutes, so the units cancel. Dollars over dollars, kilograms over kilograms — whatever you feed in, z comes out as a bare number. That cancellation is the entire trick. It is why a race time, an asking price and a body weight can be laid side by side on one axis, and why the same z means the same thing no matter what was measured. The move also runs backwards: x = mean + z × SD turns a standing back into a raw value, which is exactly what the five buttons above do to A.
Five ways this goes wrong. First, a z-score is not a percentile. Converting one to the other — the third chart — runs the number through a bell curve, so that percentage is only as good as the assumption that the data are roughly bell-shaped. Standardizing itself assumes nothing at all; only the percentile step does. Real race times, incomes and prices are usually right-skewed, so treat those percentages as a model’s estimate, not a count. Second, standardizing does not change the shape. It slides and stretches the axis, nothing more. A skewed distribution standardized is a skewed distribution with new labels — it does not become normal, and no amount of dividing will make it so. Third, a negative z is not automatically bad. In the race scenario the more negative z is the better run. Whether high or low is desirable is a fact about the world, not about the arithmetic, and z will never tell you which. Fourth, |z| > 2 is a rough flag, not a verdict. It marks a value as unusual relative to its own group; it does not make it an error, an outlier by any formal rule, or a problem. In a bell-shaped group about 1 value in 20 clears that bar routinely. Fifth, z is only as good as the mean and SD you divide by. Standardize against the wrong reference group and the arithmetic stays flawless while the answer becomes meaningless.