The Student's t-Distribution

One more distribution needs introducing before we get to hypothesis testing: the Student's t-distribution. It looks like the normal but with heavier tails, and it's parameterized by degrees of freedom (related to sample size).

Why does it exist? Here's the situation: when you run a hypothesis test, you often need to know the standard deviation of a population — but you don't. You only have a sample standard deviation. And that sample standard deviation has its own uncertainty, especially when your sample is small. Estimating variance from a small sample is less reliable, and that extra uncertainty has to go somewhere.

The t-distribution accounts for it. Its heavier tails represent the increased probability of extreme values when your variance estimate is imprecise. When sample sizes are large (roughly n > 30), the t-distribution converges to the normal — the extra uncertainty becomes negligible. When samples are small, the heavier tails make your tests appropriately more conservative.

t-Distribution Explorer

The t-distribution (violet) overlaid with the standard normal (dashed gray). As degrees of freedom increase — reflecting a larger sample — the heavier tails shrink and the curves converge.

Degrees of freedom (df = n − 1)5
160
Threshold — test statistic |t*|2.0
0.54.0
−5±2.0+5
t-distributiondf = 5
10.19%P(|X| > 2.0)
Standard normal
4.55%P(|Z| > 2.0)
t assigns 5.64 pp more probability to the tails — using normal here would overstate significance.

Heavy tails — noticeably wider than normal

5
Degrees of freedom
1.667
Variance (df / df−2)
5.64 pp
Tail excess vs. normal
t-distribution (df = 5)Standard normalThreshold ±2.0

Adjust degrees of freedom and threshold to see how the t-distribution's heavier tails shrink toward the normal — and how much more tail probability a t-test assigns compared to a z-test at the same cutoff.

When to Use t vs. Normal

In practice: always use the t-distribution for tests about means unless you somehow know the population standard deviation (which you almost never do). For large samples, the t and normal give nearly identical results. For small samples, the t is more honest about uncertainty.

Checkpoint

You're running a test with n=8 samples. You compute a test statistic and need to determine whether it's extreme enough to reject the null hypothesis. Should you compare it to the normal distribution or the t-distribution — and why?