Key takeaways
- Z-scores are unitless. They measure distance in standard deviations, so inches, points, and dollars become comparable.
- Sign gives direction, size gives rarity. z = −2.0 is below the mean and about as unusual as z = +2.0.
- Percentiles need the normal shape. The z-to-percentile conversion assumes a normal distribution, not just any data.
- |z| > 3 is rare under normality. About 0.27% of values fall that far out, which is the basis of the three-sigma outlier flag.
- A z-score is not automatically an outlier. In small or skewed samples, unusual values are common and the normal cutoff misleads.
What a z-score measures
A z-score, also called a standard score, expresses a value as a signed multiple of the standard deviation. If a class averages 72 points with a standard deviation of 8, a score of 86 is 14 points above the mean, which is 14 ÷ 8 = 1.75 standard deviations. Its z-score is +1.75.
The transformation removes the original units, so 1.75 standard deviations above average means the same relative position whether the variable is exam points, heights, or measurement errors. A z-score of 0 is the mean, and a full set of z-scores has standard deviation 1.
The formula: population and sample
For a full population with mean μ and standard deviation σ:
z = (x − μ) ÷ σ
When the mean and standard deviation come from a sample, the same arithmetic uses x̄ and s: z = (x − x̄) ÷ s. Many textbooks use n − 1 in the sample standard deviation, which makes s slightly larger and z slightly smaller than the population version. The difference is negligible for large samples and noticeable for small ones, so state which standard deviation you used.
Subtracting the mean centers the distribution on zero; dividing by σ makes one unit equal one standard deviation. Neither operation changes the shape, so skewed data stays skewed.
Worked example: one exam, two students
An exam has a mean of 72 and a standard deviation of 8. Student A scores 86: z = (86 − 72) ÷ 8 = 1.75. Student B scores 60: z = (60 − 72) ÷ 8 = −1.50. Student A is 1.75 standard deviations above average, and Student B is 1.5 below.
The raw gap is 26 points, but the relative gap is 3.25 standard deviations. If the class is roughly normal, Student A sits near the 96th percentile and Student B near the 7th: the tails are 4.0% and 6.7%, because the normal curve thins as it moves from the center.
Percentiles from the normal curve
Under a standard normal distribution, the area to the left of a z-score is its percentile. The cumulative distribution function Φ gives that area: Φ(1.75) ≈ 0.9599, so a z of 1.75 is about the 96th percentile. Φ(−1.50) ≈ 0.0668, the 6.7th percentile. Common landmarks are Φ(0) = 0.50, Φ(1) ≈ 0.8413, Φ(1.96) ≈ 0.9750, and Φ(2) ≈ 0.9772.
For a range, subtract the two areas. The share of values between z = −1 and z = +1 is 0.8413 − 0.1587 = 0.6827, or 68.27%. Between z = 0 and z = 1.5 the share is 0.9332 − 0.5000 = 0.4332, or 43.3%. Tables and spreadsheet functions report the same left-tail area; reading a two-tailed table as one-tailed is the usual error.
The 68-95-99.7 rule
For a normal distribution, about 68% of values lie within one standard deviation of the mean, 95% within two, and 99.7% within three. In z terms: roughly 68% of values have |z| ≤ 1, 95% have |z| ≤ 2, and 99.7% have |z| ≤ 3.
The rule is a quick sanity check, not a law. It applies to normal data only. For any distribution with a finite variance, Chebyshev’s inequality guarantees at least 75% of values within two standard deviations and at least 88.9% within three, which is a much weaker but universal statement. If a dataset claims 99% of values within one standard deviation, either the distribution is not normal or the standard deviation was computed wrongly.
Outliers: three sigma and the IQR rule
A common flag treats |z| > 3 as an outlier, because only about 0.27% of normal values fall that far out. The threshold is a convention, not a verdict: with 1,000 normal observations you expect roughly 2.7 values beyond three sigma, so a few extreme z-scores are normal in large samples.
The interquartile range rule is distribution-free and often preferred for skewed data: flag values below Q1 − 1.5 × IQR or above Q3 + 1.5 × IQR. The two rules can disagree, and that disagreement is informative. In a skewed income dataset, z-based screening flags too many high earners while the IQR rule does not. Report which rule you used.
Comparing scores on different scales
Standardization is how admissions offices compare applicants across test versions. Suppose the SAT has a mean of 1050 and a standard deviation of 150, and the ACT has a mean of 21 and a standard deviation of 5. A 1300 SAT gives z = (1300 − 1050) ÷ 150 = 1.67. A 28 ACT gives z = (28 − 21) ÷ 5 = 1.40. The SAT score is stronger relative to its test, even though 28 “looks” higher than 1300 in isolation.
The comparison is only as good as the reference group: national norms and a single school’s scores answer different questions, so always ask which mean and standard deviation produced the number.
Z-scores for sample means
Averages have their own distribution. For a sample of size n drawn from a population with mean μ and standard deviation σ, the sample mean x̄ has standard error σ ÷ √n, so its z-score is:
z = (x̄ − μ) ÷ (σ ÷ √n)
With μ = 100, σ = 12, n = 36, and an observed mean of 104, the standard error is 12 ÷ 6 = 2 and z = 4 ÷ 2 = 2.0. That observed average is two standard errors above the population mean — unusual, but not impossible. This is the calculation behind a one-sample z-test, and it is why larger samples make the same difference look more extreme.
Common mistakes
- Dividing by the variance instead of the standard deviation, which inflates every z-score.
- Mixing a sample mean with a population standard deviation without saying so, or using n instead of n − 1 in the sample standard deviation.
- Reading a negative z-score as “bad” when it only means “below the mean.”
- Converting z to a percentile for data that is not approximately normal.
- Treating every |z| > 2 as an error or outlier rather than as a value worth checking.
- Rounding z to one decimal before looking up a percentile, which can move the result by several percentile points near the tails.
Practical caveats
A z-score needs a meaningful mean and standard deviation. For a skewed variable such as household income, the mean is pulled upward by a few large values, and standardizing does not fix the shape — it only re-labels the axis. In that setting, percentiles from the raw distribution, or the IQR rule, describe the data more honestly than a normal-curve conversion.
Growth charts are a good example of z-scores used carefully: child height and weight are compared with reference distributions that are themselves skewed, so the charts use a modified z-score rather than a plain normal table. The same caution applies to any “percentile” claim: z-score and percentile coincide only when the normal model fits.
Frequently asked questions
Can a z-score be negative?
Yes. A negative z-score means the value is below the mean. It is not an error and not necessarily a bad result; z = −0.5 is simply half a standard deviation below average.
What is considered a high z-score?
There is no universal cutoff. In a normal distribution, |z| > 2 is uncommon (about 5% of values) and |z| > 3 is rare (about 0.27%). In quality control, 3 is a common flag; in admissions, 1.5 may be notable.
Is a z-score the same as a percentile?
No. A percentile is a position in a distribution; a z-score is a distance in standard deviations. They correspond only under a normal distribution.
What is the difference between a z-score and a t-score?
A t-score is a linear rescaling of a z-score: t = 50 + 10z. It is used in some psychological and educational tests so that scores are positive whole numbers, with a mean of 50 and a standard deviation of 10.
Can I calculate a z-score without the standard deviation?
No. The standard deviation sets the scale, so it is required. If you only know the mean, you can say whether a value is above or below average but not how unusual it is.
Why does standardization set the mean to 0 and the standard deviation to 1?
Subtracting the mean shifts the center to zero, and dividing by the standard deviation rescales the spread to one. Both are linear operations, so the shape of the distribution is unchanged.
Further reading
Keep exploring
Decision path