T-test calculator

What is this test?

A t-test is a family of procedures for comparing an observed mean or mean difference against a null hypothesis. The null hypothesis usually says that the relevant difference is zero. The test divides the estimated difference by its standard error, producing a t statistic. Large positive or negative t statistics are less expected when the null hypothesis and model assumptions are reasonable. Statoma supports four common versions: one-sample, paired, independent two-sample with equal variance, and Welch's two-sample test. The right version depends on how the data were collected, not on which result looks strongest.

The t distribution matters because the population standard deviation is usually unknown and must be estimated from the sample. That extra uncertainty is reflected through degrees of freedom. With small samples, the t distribution has heavier tails than a normal distribution. As sample size grows, the t distribution becomes closer to the normal curve. This is why the same t statistic can have different p-values under different degrees of freedom.

The test is most useful when the statistical question is already framed as a comparison of means. It is not a general-purpose test for every numeric outcome. If the outcome is highly skewed, dominated by a few extreme observations, censored, ordinal, or measured on a scale where the mean is not meaningful, another summary or model may be better. A t-test can be robust, especially with balanced moderate samples, but robustness is not a substitute for understanding how the data were produced.

When to use it

• Use a one-sample t-test when one sample mean is compared with a fixed reference value, target, benchmark, or historical mean.
• Use a paired t-test when measurements come in matched pairs, such as before-and-after observations on the same person or matched units in a study design.
• Use the independent equal-variance test only when the two groups are independent and it is reasonable to treat their population variances as equal.
• Use Welch's t-test for two independent groups when sample sizes or standard deviations differ. It is often a better default than the equal-variance version.
• Use a two-sided alternative when differences in either direction matter. Use a one-sided alternative only when the direction was specified before looking at the data.

The most important decision is the design decision. A paired test asks whether the average within-pair difference is zero, so it removes between-person or between-unit baseline variation from the comparison. An independent test compares two separate groups, so it must account for variation within each group. Welch's test keeps that structure but does not force both groups to share one pooled variance estimate. Choosing the mode after seeing the answer weakens the analysis because the test no longer matches the original question.

How it works

All versions use the same basic structure: estimate a difference, divide by the standard error of that estimate, then compare the statistic with a t distribution. For a one-sample t-test, the difference is the sample mean minus the hypothesized mean.

In this formula, the sample mean is the observed average, the hypothesized mean is the null value, the sample standard deviation estimates spread, and the sample size controls the standard error. For a paired test, the same formula is applied to the list of paired differences. For two independent groups, the estimate is the difference between the two group means. Welch's version uses a standard error based on each group's own variance and a fractional degrees-of-freedom approximation.

The p-value is then calculated from the selected tail area of the t distribution. A two-sided test doubles the smaller tail area because evidence in either direction counts against the null hypothesis. A greater-than test uses the right tail, and a less-than test uses the left tail. The tail choice should follow from the research question before calculation. It is part of the test definition, not a formatting option for the final result.

Worked example

Suppose a researcher measures whether a training program changes a score. The sample mean after training is 82.4, the reference mean is 80, the sample standard deviation is 6, and the sample size is 36. The estimated difference is 2.4 points. The standard error is 6 divided by the square root of 36, which equals 1. The t statistic is therefore 2.4. The degrees of freedom are 35. With a two-sided alternative, the p-value is the probability of seeing a t statistic at least 2.4 units from zero in either direction under a t distribution with 35 degrees of freedom.

The interpretation is not that the null hypothesis has a specific probability of being true. Instead, the result says that this sample would be relatively uncommon if the true mean equaled the reference value and the model assumptions were adequate. The estimated difference still needs practical interpretation. A 2.4 point difference may be important in one field and trivial in another. The method gives statistical evidence; the study context supplies meaning.

The same arithmetic would not justify the same conclusion in every study. If the 36 observations came from a convenience sample, a repeated-measures setup treated as independent, or a process with strong outliers, the numerical t-test result would be incomplete. A careful report would state the test version, the alternative hypothesis, the estimated difference, the t statistic, degrees of freedom, and the p-value, then explain whether the assumptions are plausible enough for the conclusion being drawn.

Common mistakes