Z-score

Z-score (or standard score) measures how many standard deviations a value sits above or below the mean. Formula: z = (x − mean) ÷ standard_deviation.

A value with z = 0 is the mean. z = 1 is one standard deviation above the mean. z = −2 is two standard deviations below. For a roughly normal distribution, |z| > 2 covers about 5% of values; |z| > 3 covers about 0.3%.

The point of converting to z-scores: it puts different scales on the same yardstick. A 75 on a math test where the class mean is 70 with SD 10 gives z = 0.5. A 68 on a French test where the mean is 65 with SD 5 gives z = 0.6. The French score is relatively better, which raw scores can’t tell you.

Z-scores feature in standardised tests (SAT, GMAT), clinical reference ranges (bone density, growth curves), quality control (process capability indices), and any context where you need to compare across populations with different means and spreads.

Our statistics calculator computes mean and standard deviation; the z-score is a one-line subtraction-then-division on the result.

From z-score to percentile, and back: for a normal distribution the cumulative distribution function (Φ) converts z-scores directly to percentiles. z = 0 → 50th percentile (the median), z = 1 → ~84th, z = 1.645 → 95th, z = 1.96 → 97.5th (the standard two-sided 95% confidence limit), z = 2.576 → 99.5th. These cutoffs are the backbone of hypothesis testing — a z-score of 1.96 is the value above which a two-sided test rejects the null at α = 0.05. For non-normal data the z-score still measures “deviations from the mean in SD units” but the corresponding percentile depends on the actual distribution and is not interchangeable with the normal tables.

Robust z-scores for messy data: outliers inflate the standard deviation, which deflates every z-score in the dataset and hides the very outliers you were trying to surface. The modified z-score (Iglewicz & Hoaglin, 1993) uses the median and the median absolute deviation (MAD) instead of mean and SD: z_modified = 0.6745 · (x − median) / MAD. The 0.6745 constant scales MAD so that for normal data, the modified z-score matches the conventional one. The convention is to flag |z_modified| > 3.5 as an outlier — robust against contamination in a way the standard z-score is not. Related: percentile, sample standard deviation, standard deviation explained.

Worked example

A two-year-old child measures 89 cm tall. WHO growth-chart data for two-year-old girls has mean 86.4 cm, SD 3.3 cm. Z-score: z = (89 − 86.4) / 3.3 = 0.79. Looking that z up in the standard normal table gives a percentile of ~78 — taller than 78% of two-year-olds, well within normal range (paediatricians flag for evaluation around z < −2 or z > 2, corresponding to roughly <2.3rd or >97.7th percentile). The same child’s 14 kg weight against mean 12.3 kg, SD 1.4 kg: z = (14 − 12.3) / 1.4 = 1.21 → ~89th percentile. Plotting height-z (0.79) against weight-z (1.21) shows the child is heavier-than-expected for their height — a clinical signal more useful than either raw measurement alone.

When and why it matters

Z-scores let you compare apples to oranges by stripping away scale. They drive the WHO and CDC growth charts, FICO and other credit-score normalisation, the Cooper Test in fitness assessment, Six Sigma process capability indices (a 6σ process means the nearest spec limit is at z = 6 — about two defects per billion), and almost every hypothesis test in statistics (Student’s t-test, ANOVA, regression coefficient significance all reduce to a z-like ratio of effect to standard error). The pitfall to know: z-scores assume the underlying distribution is at least approximately normal. For heavily skewed data (income, response time, file size) a z of 3 is not the 99.87th percentile — robust z-scores or non-parametric methods are needed instead. Reference: WHO Child Growth Standards.