You use the Student t-distribution whenever you’re working with a sample mean but don’t know the population’s true standard deviation, which is nearly every real-world scenario. If you’re estimating the standard deviation from your sample data (using the sample standard deviation instead of a known population value), the t-distribution accounts for the extra uncertainty that estimate introduces. The more familiar normal (z) distribution only applies when the population standard deviation is already known, a rare luxury outside of textbook problems.
The Core Decision: Known vs. Unknown Standard Deviation
The single biggest factor is whether you know the population standard deviation or are estimating it from your data. When you plug in a sample standard deviation instead of the true population value, you introduce a small error. That error matters most with small samples, and the t-distribution corrects for it by having thicker tails than the normal distribution. Those thicker tails translate to wider confidence intervals and higher thresholds for statistical significance, which protects you from being overconfident in results drawn from limited data.
In practice, you almost never know the true population standard deviation. That makes the t-distribution the default choice for most hypothesis tests and confidence intervals involving means. The z-distribution is reserved for cases where the population standard deviation is genuinely known (certain standardized tests, industrial processes with long historical records) or where you’re working with proportions rather than means.
How Sample Size Shapes the Distribution
The t-distribution’s shape depends on a value called degrees of freedom, calculated as your sample size minus one (n − 1). With a small sample of, say, 10 observations, you have 9 degrees of freedom, and the distribution’s tails are noticeably higher and heavier than those of the normal distribution. That extra area in the tails reflects greater uncertainty: with fewer data points, your sample standard deviation is a rougher estimate of the true value.
As the sample size grows, the t-distribution gradually narrows and starts to look identical to the normal distribution. By the time you reach about 30 degrees of freedom, the two curves are nearly indistinguishable. This is where the well-known “n = 30 rule” comes from, though it’s a guideline rather than a hard cutoff. At sample sizes of 100 or more, the difference between a t-value and a z-value is negligible for all practical purposes, but using the t-distribution at any sample size is never wrong. It simply converges to the same answer the z-distribution would give.
Three Common Types of T-Tests
One-Sample T-Test
Use this when you have a single group of measurements and want to compare its mean to a specific known value. For example, a manufacturer might test whether the average weight of cereal boxes coming off a production line differs from the target of 500 grams. You have one sample and one reference number.
Paired (Dependent) Samples T-Test
Use this when you measure the same subjects twice, typically before and after some treatment or intervention. Blood pressure readings taken on the same patients before and after starting a medication would call for a paired t-test. Each “before” measurement has a natural partner in the “after” column, and the test analyzes the differences within each pair. The key requirement is that both measurements come from the same individual or matched unit.
Independent (Unpaired) Samples T-Test
Use this when you’re comparing the means of two separate groups with no natural pairing between them. A clinical trial comparing a treatment group of 40 patients against a control group of 40 different patients is a classic example. The only connection between the groups is that you measured the same variable (say, symptom severity) in both. Unlike the paired test, no observation in one group corresponds to a specific observation in the other.
Building Confidence Intervals With T-Values
One of the most common uses of the t-distribution is constructing a confidence interval for a population mean. The formula follows a simple structure: take your sample mean, then add and subtract a margin of error. That margin of error equals the critical t-value (based on your confidence level and degrees of freedom) multiplied by the sample standard deviation divided by the square root of your sample size.
Written out, the interval looks like: sample mean ± t* × (sample standard deviation / √n). The t* value is larger than the equivalent z-value would be, especially at small sample sizes, which produces a wider interval. That wider interval honestly reflects the fact that you’re less certain about the population mean when you have fewer observations and an estimated standard deviation. As your sample grows, t* shrinks toward the z-value and the intervals narrow accordingly.
The Normality Assumption
The t-distribution formally assumes that the underlying data come from a population that is approximately normally distributed. For small samples (under 30), this assumption matters. If your data are heavily skewed or contain extreme outliers, the t-test’s results can be unreliable, and a nonparametric alternative may be more appropriate. You can check normality visually with a Q-Q plot or formally with a test like the Shapiro-Wilk test, which can be run with as few as three observations.
The good news is that the t-test is surprisingly robust against mild violations of normality. Research has shown that even for highly skewed distributions like the exponential, t-tests perform well at modest sample sizes in terms of both false-positive rates and statistical power. When sample sizes are equal between two groups, the test is especially resilient. The practical takeaway: you don’t need perfectly bell-shaped data, but you should be cautious with very small samples that are clearly non-normal.
When Variances Differ Between Groups
The standard independent-samples t-test assumes both groups have roughly equal variability. When that assumption breaks down, a variant called Welch’s t-test adjusts the degrees of freedom to compensate. A common rule of thumb is to use Welch’s version whenever the ratio of the two sample variances exceeds 4:1 in either direction. If one group’s variance is more than four times the other’s, the standard pooled t-test can give misleading results.
Many statisticians now recommend Welch’s t-test as the default for two-group comparisons, since it performs just as well as the standard version when variances are equal and protects you when they aren’t. Most statistical software packages offer it as an option, and some use it as the default.
Quick Reference: T-Distribution vs. Z-Distribution
- Population standard deviation known: use the z (normal) distribution
- Population standard deviation unknown (estimated from sample): use the t-distribution
- Small sample size (under 30): the t-distribution is essential because its thicker tails account for greater estimation uncertainty
- Large sample size (30+): the t-distribution still works but gives results virtually identical to the z-distribution
- Comparing means (one group, two groups, or paired measurements): use the appropriate t-test
- Working with proportions rather than means: use the z-distribution instead
If you’re ever unsure, defaulting to the t-distribution is the safer choice. It never gives you a wrong answer in situations where the z-distribution would be correct; it simply gives you the same answer. But using the z-distribution when you should be using the t can make you overstate your confidence, especially with smaller samples.