The t distribution
The t distribution describes how a mean, standardized by an estimated standard error, varies from sample to sample. It looks like the normal curve but with heavier tails, more so with fewer degrees of freedom.
Sources checked on November 21, 2026. Results computed by the site's engine and tested against R and SciPy.
Where it comes from
If you standardize a sample mean using the true population SD, the result follows the standard normal distribution. In practice you use the sample SD, which is itself uncertain. William Gosset, writing as Student in 1908, worked out the distribution of the mean standardized by the sample SD for normal data. It depends on one parameter, the degrees of freedom, which for a one-sample problem is n − 1.
Shape and degrees of freedom
The t distribution is symmetric around zero and bell-shaped. With few degrees of freedom it has much heavier tails than the normal, so large values of t are more common. As df grows it narrows towards the standard normal.
| df | P(T > 2) | P(T > 3) | 97.5th percentile |
|---|---|---|---|
| 1 | 0.1476 | 0.1024 | 12.706 |
| 2 | 0.0918 | 0.0477 | 4.303 |
| 5 | 0.0510 | 0.0150 | 2.571 |
| 10 | 0.0367 | 0.0067 | 2.228 |
| 30 | 0.0273 | 0.0027 | 2.042 |
| 100 | 0.0241 | 0.0017 | 1.984 |
| infinity | 0.0228 | 0.0013 | 1.960 |
With 1 df, a t above 3 happens 0.1024 of the time; under the normal curve it happens only 0.0013 of the time. That is why small samples need larger t values to reach significance.
Mean, variance and special cases
- Mean: 0 for df above 1 (for df = 1 it does not exist).
- Variance: df / (df − 2) for df above 2, always larger than 1. With 5 df it is 1.667; with 30 df, 1.071.
- df = 1 is the Cauchy distribution, whose tails are so heavy that it has no mean.
- As df grows without limit, the t distribution becomes the standard normal.
Where t-tests use it
Every t-test ends by comparing a t statistic with this distribution. The degrees of freedom come from the design: n − 1 for one-sample and paired tests, n1 + n2 − 2 for Student's independent test, and a decimal value from the Welch-Satterthwaite formula for Welch's test. Confidence intervals for means use its percentiles as multipliers. Turn a t into a p-value with the p-value from t calculator, find percentiles with the critical value calculator, or print the t table.
How the values here are computed
Tail areas come from the regularized incomplete beta function: the two-sided p for t with ν df is Iν/(ν+t²)(ν/2, 1/2), as given in NIST's Digital Library of Mathematical Functions. We evaluate it with a continued fraction in a form that avoids subtracting nearly equal numbers, so tiny tail areas keep their accuracy. Percentiles are found by Newton's method on the exact tail area. Our results match R's pt and qt and SciPy's special functions to at least ten significant figures across thousands of test values; the methodology page lists the checks.
Heavy tails in practice
Because the t distribution has heavier tails, a 95% confidence interval from 5 observations uses a multiplier of 2.776 instead of 1.960, making it 42% wider than a z interval would be. With 30 observations the multiplier is 2.045, only 4% wider. This is the practical cost of estimating the SD from the sample.
Sources
- NIST Digital Library of Mathematical Functions, chapter 8 (incomplete gamma and beta functions) checked November 21, 2026
- SciPy documentation: scipy.special.betainc checked November 21, 2026