Is it always wider-tailed than normal?

A central t distribution with finite degrees of freedom has heavier tails than standard normal.

Can degrees of freedom be non-integer?

Yes, including in some approximate inference procedures.

Does shifting a t distribution make it noncentral?

No. Location-shifting and noncentral t distributions are distinct constructions.

AI Glossary / Statistics and probability / Intermediate

Student's t distribution

Definition

Student's t distribution is a family of symmetric probability distributions indexed by degrees of freedom. It resembles the standard normal distribution but has heavier tails at finite degrees of freedom. It arises when a normally distributed quantity is standardized using an independent estimate of its variability. A familiar application is inference about a normal population mean when the population standard deviation is unknown and estimated from the sample.

What it means

The extra tail weight reflects uncertainty introduced by estimating variability. With few degrees of freedom, unusually large standardized values are more plausible than under a standard normal reference distribution. As degrees of freedom increase, the t distribution approaches the standard normal.

The distribution and a t-test are related but different concepts. The distribution is a mathematical object; a test is a procedure using a statistic and an appropriate reference distribution. Student's t distribution also appears in regression inference and can be used as a model for heavy-tailed observations. Its relevance is not limited to a rule that every small dataset should receive a t-test.

How it works

One construction is T = Z / sqrt(V / nu), where Z is standard normal, V has a chi-square distribution with nu degrees of freedom, and the two are independent. The resulting T has a central t distribution with nu degrees of freedom.

For n independent observations from a normal population, the statistic T = (sample mean − population mean) / (sample standard deviation / sqrt(n)) has n − 1 degrees of freedom. The denominator uses the sample standard deviation calculated with n − 1 in the variance formula.

Use the relevant t quantile to construct a confidence interval or evaluate a test statistic. Degrees of freedom depend on the procedure; they are not always n − 1. Regression and unequal-variance comparisons use different calculations.

Assess the data-generating and sampling assumptions before interpreting the result. Normality gives the exact one-sample result; approximations may be useful in other settings but can be unreliable with severe skewness, outliers, or dependence. Changing location or scale in a t model is also different from using a noncentral t distribution.

Three practical examples

1. An analyst estimates a mean measurement from a small independent normal sample, using a t interval because variability is estimated from the data.

2. A regression report uses an appropriate t reference distribution for a coefficient under the fitted model's assumptions.

3. A modeling team considers a t error distribution to allow heavier tails than a normal error model and checks whether it fits the application.

Think of it this way

Think of measuring distance with a ruler whose scale is itself estimated. The uncertainty is greater than if the ruler were perfectly known. The heavier tails allow for that additional uncertainty in the relevant standardized statistic, without claiming that every raw dataset must look like a t distribution.

Worked example

Consider an original one-sample teaching example. Nine independent measurements are assumed to come from a normal population. Their sample mean is 12 units and sample standard deviation is 3 units. To examine a null mean of 10, the estimated standard error is 3 / sqrt(9) = 1 unit.

The observed t statistic is (12 − 10) / 1 = 2, with 9 − 1 = 8 degrees of freedom. For a prespecified two-sided 5% test, the positive critical value is approximately 2.306. Since 2 is smaller than 2.306, this rule does not reject the null.

The corresponding 95% confidence interval is 12 ± 2.306 × 1, or approximately 9.694 to 14.306 units. The null value 10 lies inside the interval. The interval's confidence level describes the long-run coverage of the procedure under its assumptions, not a post-data probability assigned to this particular fixed population mean.

Using 1.96 instead would give the narrower interval 10.04 to 13.96 and a different threshold decision. That substitution ignores the finite-degree-of-freedom reference needed here. Non-rejection still does not prove that the true mean equals 10 or that all practically relevant effects are absent.

Illustrative case study

In a fictional company, Hazel Instruments summarizes a small calibration sample using a normal critical value despite estimating variability from the same measurements. A reviewer recomputes the interval with the appropriate t reference under the stated normal-sample model. The interval widens enough to change the threshold conclusion. The team reports the estimated difference and uncertainty, then checks independence and measurement quality before deciding whether more data are needed. The corrected calculation improves the report without turning a statistical threshold into a complete operational decision.

Common misconceptions and mistakes

The t distribution is not automatically appropriate simply because a sample is small. Its exact use depends on the statistic and assumptions.

Degrees of freedom are not a universal synonym for sample size minus one. The relevant procedure determines them.

A t distribution is symmetric, but its mean is undefined when degrees of freedom are at most one; its variance is finite only above two. Heavy tails therefore have substantive mathematical consequences. Finally, the t distribution approaching normal as degrees of freedom grow does not repair dependent observations, selection bias, or a misspecified model.

Business use

Understanding the t distribution helps businesses interpret small-sample estimates, regression tables, and uncertainty intervals. It discourages treating every standardized difference as if variability were known exactly.

Reports should state the estimate, standard error, degrees of freedom, interval method, and material assumptions where they affect the conclusion. For AI evaluation, repeated prompts from the same user or task can violate simple independence assumptions. Choosing a t critical value alone does not account for that dependence; the design and variance estimate need to match the evaluation structure.

Check your understanding

Answer: Ten independent normal observations are used for a one-sample mean t statistic.

What are its degrees of freedom?

Answer: Nine. This answer follows from that specific procedure, not from a universal rule for all t-based methods.

Related terms

Normal distribution
Degrees of freedom
Standard error
T-test
Confidence interval
chi-square distribution
noncentral t distribution
heavy tails

Learn next

Study one-sample and paired t procedures, then regression degrees of freedom and methods for dependent or non-normal data.

Degrees of freedom

Sources

SciPy primary distribution documentation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.t.html . NIST critical-value table: https://www.itl.nist.gov/div898/handbook/eda/section3/eda3672.htm . The eight-degree-of-freedom critical value was checked against the table. The numerical sample summaries and fictional case are original teaching illustrations.

Previous: Stt Next: Studentized range statistic ( q )