Written and maintained by CASRAI Editorial Board
Last updated
A t-statistic is the estimate you care about — a mean difference, a regression coefficient, a correlation — expressed as a multiple of its own standard error. It is a signal-to-noise ratio, not a size. That single fact resolves most of what confuses people about a printed t-value: why a t of 8 can come from a trivial real-world effect, why a t of 2 can come from a substantial one, and why “significant” and “large” are answering two different questions. This guide covers the t-statistic itself — the one construction that shows up in a one-sample t-test, a two-group comparison, a paired design, and a coefficient row in a regression table — not the full mechanics of running or reporting any single one of those tests. For the complete t-test output block (p-value, confidence interval, effect size, which line to report), see CASRAI’s guide to interpreting t-test results.
The one formula behind every t-statistic
Every t-statistic, regardless of where it appears, is built the same way:
t = (estimate − hypothesized value) / SE(estimate)
The numerator is how far your estimate sits from the value the null hypothesis claims (almost always zero — no difference, no slope, no correlation). The denominator is the standard error of that estimate: a measure of how much the estimate would bounce around if you repeated the study, driven by sample size and variability in the data. Divide one by the other and you get a number in standard-error units — not in the outcome’s original units, and not a percentage or a probability.
The same formula specializes differently depending on what you ran:
| Context | What the “estimate” is | What the t-statistic tests |
|---|---|---|
| One-sample t-test | Sample mean minus a fixed reference value | Whether the population mean differs from that reference value |
| Independent-samples t-test | Difference between two group means | Whether the two population means differ. See CASRAI’s guide to the t-test for which of the three designs applies to your data. |
| Paired t-test | Mean of the within-pair differences | Whether the population mean difference is zero |
| Regression coefficient | An estimated slope, b | Whether the population slope is zero — i.e., whether that predictor has a relationship with the outcome once the other predictors in the model are held constant |
| Correlation coefficient, r | Sample correlation, transformed as r√(n−2) / √(1−r²) | Whether the population correlation is zero |
What varies across rows is the estimate and its standard error. What never varies is the interpretation of the ratio itself: distance from the null, measured in standard-error units.
Why a bigger t-statistic doesn’t mean a bigger effect
Because the standard error sits in the denominator, and standard error shrinks as sample size grows, the same real-world effect produces a bigger t-statistic in a larger sample — and, just as consequentially, a genuinely trivial effect can still clear a large t-statistic if the sample is large enough to shrink its standard error down far enough. The t-statistic cannot, on its own, tell these two situations apart. This is the same point the American Statistical Association’s 2016 statement on p-values makes about the p-value that a t-statistic feeds into: “smaller p-values do not necessarily imply the presence of larger or more important effects” (Wasserstein & Lazar, The American Statistician, 70(2), 2016) — and the logic transfers directly to t, since a smaller p and a larger |t| are two views of the same underlying distance from the null.
A worked comparison makes the mechanism concrete. Both cases below are simulated regression datasets (a single predictor X and outcome Y), computed with ordinary least squares — not real study data, built specifically to isolate this one effect:
| Case A | Case B | |
|---|---|---|
| Sample size (n) | 12 | 300 |
| Estimated slope, b | 3.90 per unit of X | 0.019 per unit of X |
| Standard error of b | 0.48 | 0.0017 |
| Degrees of freedom | 10 | 298 |
| t-statistic | t = 8.11 | t = 10.99 |
| What the slope means in real terms | Across the full 12-unit range of X, the model predicts roughly a 43-point rise in Y | Across the full 300-unit range of X, the model predicts roughly a 5.6-point rise in Y |
Case B’s t-statistic is larger — not because its effect is bigger, but because its standard error is roughly 280 times smaller, a direct consequence of the much larger sample. Per unit of X, Case B’s real-world effect is about 200 times smaller than Case A’s. Reading “t = 10.99 beats t = 8.11” as “Case B’s relationship matters more” gets the comparison backward. The t-statistic answers “how confidently can I distinguish this from zero,” which is a question about precision, not about size.
t and degrees of freedom: the same t doesn’t always mean the same thing
A t-statistic is only interpretable alongside its degrees of freedom (df), because df determines which t-distribution the statistic is compared against — and that comparison is what eventually produces a p-value. With few degrees of freedom the t-distribution has heavier tails than a normal distribution, so a larger t is needed to be considered unusual: at df = 10, a two-tailed test needs roughly t = 2.23 to cross the conventional .05 threshold. As df grows, the t-distribution converges toward the standard normal, and the threshold tightens toward roughly 1.96. See CASRAI’s guide to degrees of freedom for how df is calculated for each design — roughly n − 1 for a one-sample or paired test, n₁ + n₂ − 2 for an independent-samples test (or a non-integer value under Welch’s correction), and n − k − 1 for a regression with k predictors.
Reading a t-statistic in a regression coefficient table
Statistical software prints a t-statistic on every coefficient row of a regression table — one per predictor, each testing its own null hypothesis that the population slope for that predictor is zero, holding the others constant. A few things specific to this context:
- Each predictor’s t is conditional on the rest of the model. Adding or removing another predictor changes every other predictor’s standard error (and sometimes its coefficient), which changes every other predictor’s t — a coefficient’s t-statistic is not a fixed property of that variable in isolation.
- A small, precisely estimated coefficient can out-t a large, noisy one. This is the mechanism from the worked example above, applied directly to a regression table: don’t rank predictors by t-statistic and call the top one “most important” without also checking the coefficient’s own size, ideally in standardized units.
- Inflated standard errors deflate t, independent of the real relationship. Multicollinearity among predictors is a common cause — see CASRAI’s guide to multicollinearity and VIF for how to check whether a small t-statistic reflects a genuinely weak relationship or an inflated standard error.
For the mechanics of producing and reading a full regression table in a specific package, see CASRAI’s guides to regression analysis and running a regression in Stata.
Reading a t-statistic in a t-test
In a t-test, the t-statistic is one of several numbers — alongside df, the p-value, a confidence interval, and an effect size — that only mean something read together. This page has focused on the t-statistic itself because that piece generalizes across every context above; the rest of the output block (which confidence-interval traps to watch for, when to report the Welch’s-correction row instead of the standard one, how to turn the whole block into a reporting sentence) is specific to the t-test and covered in full in CASRAI’s guide to interpreting t-test results.
What the t-statistic does not tell you
- It does not tell you the size of the effect. A t-statistic and an effect size (Cohen’s d for a mean difference, a standardized regression coefficient for a slope) answer different questions, and a t-test’s or regression’s output answers both only if you read both numbers. See CASRAI’s guides to calculating Cohen’s d and effect size in research.
- It does not tell you practical or clinical importance. A large t-statistic from a huge sample can attach to a difference too small to matter in practice; a modest t-statistic from a small, well-designed study can attach to a difference that matters a great deal. See CASRAI’s comparison of statistical significance vs. clinical significance.
- Its sign tells you direction, nothing more. A negative t simply means the estimate fell below the hypothesized value rather than above it — it carries no information about magnitude beyond the absolute value.
A quick way to read any t-statistic
- Check the sign for direction — above or below the null value.
- Check it against its degrees of freedom (or just read the accompanying p-value) to see how unusual it is under the null — not how big.
- Stop there only if all you need is a yes/no on “distinguishable from the null.” To say anything about size, go to the estimate itself in its real units, and to a standardized effect size if you need to compare across studies or variables measured on different scales.
Frequently asked questions
What counts as a “good” or “large” t-statistic?
There’s no universal cutoff — what counts as large depends entirely on degrees of freedom. A t of 2.5 clears the conventional significance threshold easily at df = 200 but barely clears it at df = 5. Compare a t-statistic to its own degrees of freedom (or just read the p-value it produces), not to a fixed number.
Does a bigger t-statistic always mean a bigger effect?
No. Because the t-statistic is the estimate divided by its standard error, and standard error shrinks as sample size grows, a large t-statistic can come from a large sample and a trivial effect just as easily as from a genuinely large one. Check the estimate’s own units and a standardized effect size before concluding a larger t-statistic reflects a larger effect.
What does a negative t-statistic mean?
Only that the estimate fell below the hypothesized (usually zero) value rather than above it. The sign gives direction; the absolute value, read against degrees of freedom, gives a sense of how unusual the result is under the null. Neither the sign nor the magnitude of t describes effect size.
How is the t-statistic for a regression coefficient different from a t-test’s t-statistic?
It isn’t, structurally — both are an estimate divided by its standard error, tested against zero. What differs is what’s being estimated (a group-mean difference versus a predictor’s slope) and how the standard error is calculated, which for a regression coefficient depends on the other predictors in the model. Degrees of freedom also differ: n − k − 1 for a regression with k predictors, versus a design-specific formula for a t-test.
Why do two different predictors in the same regression have very different t-statistics?
Because each predictor’s coefficient and standard error are estimated separately (conditional on the other predictors in the model), and t-statistics are not comparable in any absolute sense across predictors measured on different scales. A predictor with a small, precisely estimated coefficient can have a larger t-statistic than one with a large, noisily estimated coefficient.
