What are degrees of freedom?

The intuition: how many values are still free

Degrees of freedom, written $df$, answer a counting question. You start with $n$ data values, all free to be anything. Then you compute something from them, and that computed value locks in a relationship the data now have to satisfy. Every relationship you impose costs you one free value.

Picture five test scores whose mean you have already calculated as 12. The first four scores can still be anything: 9, 14, 10, 15 are all allowed. But the fifth is no longer free. The five scores must total $5 \times 12 = 60$, and the first four total 48, so the fifth is forced to be 12. Four values were free, one was determined, so $df = 4$.

That is the whole idea. $df$ is the number of independent pieces of information left in the data after you have spent some of them on estimates. Because it counts how much information is available, it also controls how much the resulting statistic bounces around, which is why it selects a curve rather than just labeling one.

Why the sample standard deviation divides by n - 1

The sample standard deviation is built from deviations around the sample mean:

Here $\bar{x}$ (x-bar) is the sample mean. Those deviations carry only $n - 1$ independent pieces of information, because they always sum to exactly zero. Learn any $n - 1$ of them and the last one is fixed. In the five-score example the deviations are $-3, 2, -2, 3, 0$, and the fifth had to be 0 once the first four were known.

So dividing the sum of squares by $n$ would be dividing by more pieces of information than you have. It also produces an estimate that is too small on average, since the deviations are measured around $\bar{x}$, the one center that makes the sum of squares as small as it can possibly be. Dividing by $n - 1$ corrects both problems at once and makes $s^2$ an unbiased estimator of the population variance $\sigma^2$ (sigma-squared).

If you knew the population mean $\mu$ (mu) and measured deviations around it instead, you would divide by $n$, because nothing was estimated from the data. The full argument is in why you divide by n - 1.

Why a one-sample t test uses n - 1

The one-sample t statistic is

where $\mu_0$ (mu-nought) is the hypothesized mean, and the $df$ comes straight from the $s$ in the denominator. Since $s$ carries $n - 1$ independent deviations, the t curve you compare against is the one indexed by $n - 1$. A sample of 25 gives $df = 24$, no matter what the numbers are.

The curve matters because $s$ is an estimate. A z-score has a fixed number in its denominator, so it varies only through the numerator. A t statistic has an estimate in the denominator too, and a denominator that happens to come out small makes $t$ large all on its own. That extra source of variation is what fattens the tails of the t distribution.

How fat depends on $df$. Few degrees of freedom mean $s$ is unreliable and the tails are heavy; many degrees of freedom mean $s$ is close to $\sigma$ and the tails thin out. The same $n - 1$ applies to a paired t procedure, where $n$ is the number of pairs, since you compute one standard deviation from one column of differences. See the one-sample t-test walkthrough and when to use a paired t-test.

Why a two-way chi-square test uses (r - 1)(c - 1)

For a chi-square test of independence or homogeneity on a table with $r$ rows and $c$ columns, the degrees of freedom are

The same counting argument produces it. The expected counts are built from the row totals and the column totals, so those totals are treated as fixed. Ask how many cells you can fill in freely before the rest are forced.

Take a 2 by 3 table with row totals 60 and 40, column totals 50, 30, and 20, and grand total 100. Put 25 in the first cell and 20 in the second. The third cell of row 1 is now forced to $60 - 25 - 20 = 15$. Row 2 is forced cell by cell from the column totals: $50 - 25 = 25$, $30 - 20 = 10$, and $20 - 15 = 5$, which sum to 40 as required. Two cells were free, and $(2-1)(3-1) = 2$.

In general you can fill the first $r - 1$ rows and the first $c - 1$ columns freely, a block of $(r-1)(c-1)$ cells, and the margins determine everything else. A 3 by 4 table gives $(3-1)(4-1) = 6$. Notice that $df$ here depends only on the shape of the table, never on the sample size, which is the opposite of the t case. The procedure itself is in chi-square tests explained.

A chi-square goodness-of-fit test on $k$ categories uses $df = k - 1$ by the same logic, since the counts have to add to the sample total. That test was removed from the AP exam in the Fall 2026 redesign, but it is still standard college content.

Why a two-sample t gives a decimal, and why that is not a count

A two-sample t procedure for a difference of means breaks the pattern. Its standard error is

and it pools nothing: each group estimates its own spread. The exact distribution of the resulting statistic is not a t distribution at all. The Welch-Satterthwaite formula finds the t curve that comes closest, and the $df$ it returns is a tuning value for that curve, not a count of anything.

Because it is a tuning value, it has no reason to be a whole number. With $n_1 = 12$, $s_1 = 4.0$, $n_2 = 9$, and $s_2 = 7.0$, the two variance terms are $16/12 = 1.3333$ and $49/9 = 5.4444$, and the formula returns $df = 11.88$. That is a real t curve, since the t density is defined for any positive $df$.

Be honest about the difference. Every other $df$ on this page is an exact count of free values, and it is exactly right. The Welch value is an approximation chosen to make the p-value close to correct, not a count of anything. It does obey a window, $\min(n_1 - 1, n_2 - 1) \le df \le n_1 + n_2 - 2$, so the example's 11.88 sits between 8 and 19 as it must. That window is a quick check on your data entry, and the details are in why df is a decimal on a two-sample t-test.

How t approaches z as df grows

The clearest way to feel what $df$ does is to watch the two-sided 0.05 critical value $t^*$ (t-star) shrink as $df$ climbs. Every row below is the same tail area, read from the t-table, and the last row is the standard normal value $z^*$ (z-star).

The collapse is dramatic at the start and slow at the end. Going from $df = 1$ to $df = 5$ drops $t^*$ by more than 10 points, while going from $df = 30$ to $df = 100$ shaves off only 0.058. By $df = 100$ the gap to 1.960 is about 0.024, roughly 1.2%.

Read the pattern through the estimate. At $df = 1$ you have a single independent deviation informing $s$, so $s$ can land far from $\sigma$ and the interval has to be enormous to keep its promised coverage. At $df = 100$, $s$ is a steady stand-in for $\sigma$ and t behaves almost like z. The two never actually meet, which is why t stays the correct choice for a mean at any sample size.

Every df rule in this course, in one table

Two habits keep you out of trouble. Write the $df$ next to every test statistic you report, because $t = 2.2$ means different things at $df = 3$ and $df = 300$. And if a table forces you onto a whole-number row, round the $df$ down, never up, which keeps the tails heavier and the p-value slightly larger rather than smaller than the truth.