You use a t-distribution whenever you’re working with sample data and don’t know the population’s true standard deviation, which is almost every real-world scenario. If you’re estimating a mean, comparing two groups, or building a confidence interval from your own collected data, the t-distribution is the right tool. The normal (z) distribution only applies when you already know the population standard deviation, a rare luxury outside of textbook problems.
The Core Rule: Known vs. Unknown Variability
The distinction comes down to one question: do you know how spread out the entire population is, or are you estimating that spread from your sample? When you collect data and calculate a standard deviation from it, that number carries its own uncertainty. The t-distribution accounts for this extra layer of imprecision by having thicker tails than the normal distribution, meaning it assigns higher probability to extreme values. This makes your confidence intervals wider and your hypothesis tests more conservative, which is appropriate when your information is limited.
In practice, you almost never know the true population standard deviation. Even in fields with decades of published data, the specific population you’re studying may differ. So the t-distribution is the default choice for the vast majority of statistical inference involving means.
How Sample Size Affects the Shape
The t-distribution isn’t a single curve. It’s a family of curves, each defined by its degrees of freedom. For a one-sample test, degrees of freedom equal your sample size minus one (n – 1). With 10 observations, you have 9 degrees of freedom. With 50, you have 49.
At low degrees of freedom, the t-distribution looks noticeably different from the normal distribution: shorter in the middle, fatter in the tails. As your sample grows, the t-distribution gradually narrows and converges toward the normal curve. By around 30 to 60 observations, the two are nearly identical. The BMJ notes that the t-distribution procedure is preferable when you have fewer than 60 observations and certainly when you have 30 or fewer, but there’s no downside to using it with larger samples too. Using t with a large sample gives you essentially the same answer as using z, with the added benefit of not pretending you know something you don’t.
Common Situations That Call for a t-Distribution
Four practical scenarios cover most uses:
- Confidence intervals for a mean. You’ve measured something in a sample and want to estimate the range where the true population mean likely falls. Instead of multiplying the standard error by 1.96 (the familiar number from the normal distribution), you look up a slightly larger multiplier from the t-distribution based on your degrees of freedom. With 17 degrees of freedom, for example, the 95% multiplier is 2.110 rather than 1.96, producing a wider, more honest interval.
- One-sample t-test. You have a sample mean and want to know whether it differs significantly from some known or expected value. For instance, testing whether patients with a particular condition have blood calcium levels that differ from the healthy population average.
- Two-sample t-test. You’re comparing the means of two independent groups to see if the difference is statistically meaningful. This is one of the most common tests in science, from comparing treatment and control groups to evaluating differences between demographic segments.
- Paired t-test. You have two measurements on the same subjects (before and after a treatment, for example) and want to test whether the average change is significantly different from zero.
Choosing Between Student’s and Welch’s t-Test
When comparing two groups, you have two versions of the t-test available. The classic Student’s t-test assumes both groups have equal variance (similar spread in their data). Welch’s t-test drops that assumption and adjusts the degrees of freedom to compensate for unequal variances.
Research in the International Review of Social Psychology makes a strong case that Welch’s version should be your default. When variances genuinely are equal, Welch’s test loses very little statistical power compared to Student’s. But when variances are unequal and sample sizes differ between groups, Student’s t-test can be severely biased, inflating or deflating your false-positive rate in unpredictable ways. The traditional approach of first running a test for equal variances and then choosing between the two t-tests often fails to provide reliable guidance. Using Welch’s t-test by default avoids this problem entirely.
The Normality Assumption
The t-distribution technically assumes your data comes from a normally distributed population. In reality, the t-test is surprisingly tolerant of departures from normality, a property statisticians call robustness. Even with a highly skewed distribution like the exponential, t-tests perform well for modest sample sizes in terms of controlling false-positive rates.
Where the t-test struggles is with outliers. A single extreme value can dominate the calculation of both the mean and the standard deviation, pulling your t-statistic in ways that don’t reflect the bulk of your data. One illustrative example from the statistics literature shows how the value of a t-statistic can swing dramatically as a function of just one observation out of ten, while the other nine remain fixed. If your data is prone to outliers, rank-based methods like the Wilcoxon test are more reliable alternatives.
For most moderately sized datasets without extreme outliers, the normality assumption isn’t something to lose sleep over. The central limit theorem helps here too: as your sample grows, the distribution of the sample mean becomes approximately normal regardless of the underlying data shape.
When You Don’t Need a t-Distribution
A few situations call for something else. If you genuinely know the population standard deviation from extensive prior data, a z-test using the normal distribution is appropriate, though this is rare outside of standardized testing and industrial quality control. If your data is categorical (proportions rather than means), you’d use a z-test for proportions or a chi-square test instead. And if you’re comparing more than two groups at once, you’d move to an analysis of variance (ANOVA), which uses the F-distribution, though ANOVA is built on the same underlying logic as the t-test.
If your data has severe skewness, heavy outliers, or a very small sample from a clearly non-normal distribution, nonparametric tests (which don’t assume any particular distribution shape) are a safer choice. The Wilcoxon rank-sum test for two independent groups and the Wilcoxon signed-rank test for paired data are the most common alternatives.
A Simple Decision Framework
When you’re staring at a dataset and wondering which test to run, walk through these questions in order:
- Are you working with means? If yes, a t-based method is likely appropriate. If you’re working with proportions or counts, look elsewhere.
- Do you know the population standard deviation? If not (and you almost certainly don’t), use the t-distribution.
- Is your data roughly symmetric without extreme outliers? If yes, the t-test will serve you well. If your data is heavily skewed or outlier-prone, consider a nonparametric alternative.
- Are you comparing two groups? Use Welch’s t-test by default. Only switch to Student’s t-test if you have a specific, justified reason to assume equal variances.
The t-distribution was originally developed for small-sample problems, but it’s valid at any sample size. Using it whenever you estimate variability from your data is a safe, defensible practice that modern statistical thinking strongly supports.