How normal probabilities are worked out
First turn x into a z-score, the number of standard deviations it is from the mean. Then the probability is the area under the standard normal curve, written Φ(z).
P(X < x) = Φ(z) · P(X > x) = 1 − Φ(z) · P(a < X < b) = Φ(z_b) − Φ(z_a)
Inverse: x = μ + z × σ, where Φ(z) = the percentile you want
Φ has no simple closed form, so it's computed numerically: with a power series near the middle and a continued fraction in the tails, so tiny tail areas like P(Z > 6) stay accurate. Percentiles use Peter Acklam's rational approximation, refined with Halley's method. Results were checked against Python's statistics.NormalDist.
Worked example
Many IQ tests are scaled to have a mean of 100 and a standard deviation of 15. For a score of 115, z = (115 − 100) ÷ 15 = 1. So P(X < 115) = Φ(1) = 0.841345, about 84.13%, and P(X > 115) = 0.158655.
Between 85 and 115 (z from −1 to 1): 0.841345 − 0.158655 = 0.682689, about 68.27%. Above 130 (z = 2): 0.022750, about 2.28%.
The other way: the 95th percentile has z = 1.644854, so x = 100 + 1.644854 × 15 = 124.6728.
z-values for common confidence levels
| Middle area | z |
|---|---|
| 80% | ±1.281552 |
| 90% | ±1.644854 |
| 95% | ±1.959964 |
| 98% | ±2.326348 |
| 99% | ±2.575829 |
| 99.9% | ±3.290527 |
The 68-95-99.7 rule is a rounded version of the same numbers: about 68% of values are within 1 SD of the mean, 95% within 2 SD and 99.7% within 3 SD.
Getting a realistic number
- Real data is rarely exactly normal, especially far out in the tails. A tiny tail probability tells you a value is rare, not exactly how rare.
- Use the population mean and SD if you know them. With a small sample, estimates of the SD are rough.
- For whole-number data, like test scores, a continuity correction helps: for P(X ≤ 120), use 120.5 as x.
- With a normal distribution, P(X < x) and P(X ≤ x) are the same, because any single exact value has probability 0.