Degrees of Freedom

Degrees of freedom (df) counts how many values in a calculation are genuinely free to vary after the constraints you’ve already imposed. It sounds abstract; one example makes it concrete.

The intuition

Five numbers must average 8. Pick any four freely — say 4, 6, 8, 10 — and the fifth is forced (it must be 12). Fixing the mean used up one number’s freedom: five values, one constraint, df = n − 1 = 4. That is exactly why sample standard deviation divides by n − 1: deviations measured around a fitted mean carry only n − 1 independent pieces of information.

Where df shows up

t-tests: with σ unknown, the t-distribution replaces the normal — and its shape depends on df. Small df → fat tails → bigger critical values (df = 4 needs t ≈ 2.78 for 95%, versus the z table’s 1.96). As df grows, t melts into the standard normal — by df ≈ 30 the z table is already a good approximation, which is the real meaning of the "n ≥ 30" rule of thumb in z-testing. Chi-square tests: df = (rows − 1) × (columns − 1) plays the same role for categorical data. Regression: every fitted parameter spends a degree of freedom — fit too many and none remain to estimate the noise.

The through-line: df is the honest count of information left over for estimating uncertainty. Fewer degrees of freedom, wider intervals, humbler conclusions.