What a z-score actually tells you
A z-score answers one question: how many standard deviations away from the average is this value? It is the great equaliser of statistics — a height of 180 cm, a test score of 1150 and a delivery time of 42 minutes all become comparable once they are expressed as z. The formula is one subtraction and one division: z = (x − μ) ÷ σ.
Reading the number
- z = 0 means the value is exactly average; z = +1 means one standard deviation above.
- Percentile comes from the normal curve: z = 1 → 84.13% of the population is below, 15.87% above. z = 1.96 → 97.5% below, which is why 1.96 is the classic 95% two-tailed cut-off.
- z = 2 → 97.72% below (about 1 in 44 above), and z = 3 → 99.87% (about 1 in 741).
- The 68–95–99.7 rule is the quick version: ±1σ holds 68.3% of a normal population, ±2σ holds 95.4% and ±3σ holds 99.7%.
Why standardising is useful
Two students score 85 in different exams. If exam A averages 70 with σ = 10 and exam B averages 60 with σ = 15, the z-scores are 1.5 and 1.67 — the second result is actually the stronger performance, even though the raw marks look the other way round. The same logic compares growth measurements of children of different ages, quality-control samples against a process mean, and any measurement where "is this unusual?" matters more than the raw number.
Percentiles and the reverse calculation
Working backwards is just as common: "what score do I need for the 90th percentile?" Take z = 1.282 (the 90th percentile of a normal distribution) and convert back with x = μ + zσ. On a test with mean 100 and σ = 15, that is 100 + 1.282 × 15 = 119.2. This is exactly how manufacturers set specification limits and how admissions offices map percentile targets onto score scales.
When not to use a z-score
Three cautions. Skew: if the data is heavily skewed (income, waiting times), the normal percentile is wrong — use the empirical percentile from the data instead. Small samples: with fewer than about 30 observations, the standard deviation is itself uncertain, and a t-distribution (which has fatter tails) gives more honest probabilities. Non-normal distributions: for a bounded scale like a 1–5 satisfaction rating, most of the population sits in two or three categories and z-scores add little.
Standard-deviation-only statistics
If you only have a summary — mean 100, SD 15 — you can still answer questions about proportions, which is what the table below is for. Read the row for your z and take the cumulative figure. For a value between two z-scores, subtract the smaller cumulative proportion from the larger; for a range straddling the mean, add the two distances from 0.5.
Frequently asked questions
What is a z-score in simple terms?
The number of standard deviations a value sits from the mean. Someone scoring 115 on a test with mean 100 and standard deviation 15 has a z-score of +1, which puts them above about 84% of test takers.
How do I convert a z-score to a percentile?
Use the normal curve: Φ(1) = 84.13%, Φ(1.645) = 95%, Φ(1.96) = 97.5%, Φ(2) = 97.72%. The table above lists the cumulative proportion for every z from −4.0 to 4.0 in steps of 0.1, and the calculator does it for any z you enter.
What z-score is the 90th percentile?
z = 1.282. To turn that into a raw score, use x = μ + zσ. On a scale with mean 100 and σ = 15, the 90th percentile is 100 + 1.282 × 15 = 119.2, and the 95th percentile is 100 + 1.645 × 15 = 124.7.
Why is 1.96 used so often in statistics?
Because Φ(1.96) = 0.975: 2.5% of a normal distribution lies above it and 2.5% below −1.96, leaving exactly 95% in the middle. That is the basis of the familiar 95% confidence interval, mean ± 1.96 standard errors.
When should I use a t-score instead of a z-score?
When the sample is small (roughly under 30) or the population standard deviation is unknown and estimated from the sample. The t-distribution has fatter tails, so it reports slightly wider intervals and larger p-values — the honest answer when your estimate of σ is itself uncertain.
Can a z-score be negative, and is that bad?
Yes, negative simply means the value is below the mean — a z of −0.5 is perfectly ordinary, sitting above about 31% of the population. Only the absolute size matters for unusualness: |z| above 2 is uncommon, and above 3 deserves a check of the data before you call it a discovery.