Core inference
T-Statistic Calculator
The t-statistic is a signal-to-noise ratio: how far your estimate sits from the value under the null, measured in standard errors. It is the number at the heart of every t-test — t = (effect) ∕ (standard error). Enter summary statistics for a one-sample, two-sample (Welch or pooled) or paired comparison and this returns t, its degrees of freedom, the standard error and a two-tailed p-value, with the t-curve shaded so you can see how extreme your value is.
Enter summary statistics, not raw data. A paired test is just a one-sample test on the differences against zero. For the full test with confidence intervals and effect sizes, use the t-test calculator.
Result
In plain English
A t-statistic turns a difference into a question of scale: is the effect you measured big relative to the uncertainty in measuring it? The numerator is the effect — how far the sample mean is from the null value, or how far apart two means are. The denominator is the standard error — how much that estimate would wobble from sample to sample. Dividing one by the other gives a unitless ratio: a t of 2 means the effect is twice as large as its own noise.
- t-statistic
- Effect divided by standard error: t = (estimate − null) ∕ SE. A signal-to-noise ratio for a mean or a difference of means. Its sign just records direction.
- standard error (SE)
- The standard deviation of the estimate itself, shrinking with the square root of the sample size. For one mean it is s ∕ √n; for a difference it combines the two groups’ errors.
- degrees of freedom (df)
- Sets which t-curve to compare against. n − 1 for one sample, n₁ + n₂ − 2 when variances are pooled, and a fractional Welch value when they are not. Small df means fatter tails and a higher bar.
- why “Student’s” t
- William Gosset, publishing as “Student” while at Guinness, derived it for small samples where the population SD must be estimated from the data — the extra uncertainty that makes t, not z, the honest yardstick.
- t versus z
- z assumes the true standard deviation is known; t estimates it from the sample, paying for that with heavier tails. As n grows the estimate sharpens and the t-curve converges on the normal.
- not the effect size
- A large t can come from a trivial effect measured very precisely (huge n). t answers “is it distinguishable from noise?”, not “is it big or important?” — that is what Cohen’s d and confidence intervals are for.
Frequently asked
How do you calculate a t-statistic?
Take the effect and divide by its standard error. For one sample, t = (x̄ − μ₀) ∕ (s ∕ √n): the gap between the sample mean and the hypothesised value, scaled by the standard error s ∕ √n. For two independent samples, t = (x̄₁ − x̄₂) ∕ SE, where SE combines both groups’ variability — √(s₁²/n₁ + s₂²/n₂) for the Welch version, or a pooled estimate when you assume equal variances. A paired comparison reduces to the one-sample formula applied to the within-pair differences, tested against zero. The result is then read against a t-distribution with the appropriate degrees of freedom.
What is a “good” or significant t value?
There is no universal threshold — it depends on the degrees of freedom and your chosen significance level. As a rough guide, with moderate-to-large samples a |t| above about 2 corresponds to two-tailed significance at the 5% level, because the t-curve approaches the normal where ±1.96 is the cut-off. But with only a handful of observations the bar is higher (with 4 df you need |t| ≈ 2.78), and with thousands it is barely 1.96. That is why a t-statistic is meaningless without its df: always report it as t(df) = value, and let the p-value or critical value make the call.
What is the difference between a t-statistic and a t-test?
The t-statistic is the number; the t-test is the whole procedure. The test states a hypothesis, computes the t-statistic, compares it to a t-distribution to get a p-value or critical value, and reaches a decision — usually adding a confidence interval and an effect size along the way. This page computes the statistic and its two-tailed p-value so you can see what it measures; the t-test calculator runs the full procedure with intervals and Cohen’s d.
Can a t-statistic be negative?
Yes, and the sign carries no judgement — it simply records direction. A negative t means the first mean is below the second (or the sample mean is below the hypothesised value); a positive t means the reverse. For a two-tailed test only the magnitude |t| matters, so swapping the two groups flips the sign but leaves the p-value unchanged. The sign only becomes meaningful when your hypothesis is directional (one-tailed), where you care specifically whether the effect runs the predicted way.