How is the F statistic in a one-way ANOVA built, and what are its two degrees of freedom?

## Partitioning the variability Start with N observations spread over k groups — say 40 plots of land, 10 under each of four fertilizers. Compute the **grand mean** over all 40 yields and each **group mean** over its own 10. Every observation's distance from the grand mean can be split in two: how far its group mean sits from the grand mean, plus how far the observation sits from its own group mean. Squaring and summing over all observations, the cross terms vanish and you get the exact identity: ``` SS_total = SS_between + SS_within ``` with - `SS_between = sum over groups of n_i * (group_mean_i - grand_mean)^2` — the **explained** part, larger when the group means are pulled apart. Note the weight `n_i`: a big group whose mean is off-centre contributes more. - `SS_within = sum over all observations of (value - its group mean)^2` — the **unexplained** or error part, the scatter that survives after each group is centred on its own mean. ## From sums of squares to mean squares Sums of squares are not comparable on their own — `SS_within` is built from far more terms than `SS_between`, so it is bigger almost by construction. Dividing each by its degrees of freedom turns them into variances on a common footing: - `MS_between = SS_between / (k-1)`. The numerator degrees of freedom are **k-1** because the k group means are free to vary around the grand mean, but once you know the grand mean and k-1 of them the last is determined. - `MS_within = SS_within / (N-k)`. The denominator degrees of freedom are **N-k** because one mean is estimated inside each of the k groups, spending one degree of freedom apiece out of N observations. The degrees of freedom partition just as the sums of squares do: `(k-1) + (N-k) = N-1`, which is the total degrees of freedom around the grand mean. In the four-fertilizer example with 40 plots, that is 3 and 36, summing to 39. ## The statistic and why it centres on 1 ``` F = MS_between / MS_within ``` `MS_within` is always an unbiased estimate of the common error variance sigma-squared, whether or not the means differ — centring each group on its own mean removes any group effect. `MS_between` estimates sigma-squared **only when the null of equal means is true**; otherwise its expectation is sigma-squared plus a positive term driven by how far apart the true means are. So under the null the ratio of two estimates of the same quantity hovers around 1, and under the alternative the numerator inflates while the denominator does not, driving F upward. (Strictly, the mean of an F distribution is `df2 / (df2 - 2)`, which is a little above 1 and approaches 1 as the error degrees of freedom grow. "About 1" is the right interview-grade statement.) ## Reading it as signal over noise F = 1.1 means the group means are no further apart than the within-group scatter would produce by chance. F = 9 with df (3, 36) means the between-group spread is nine times the noise level — far more separation than chance explains. Whether a given F is large enough is decided by the F(k-1, N-k) distribution, which is right-skewed and defined only on non-negative values. Only the **upper tail** is used: a large F is evidence against equal means, while a tiny F says the means are unusually close, which is not evidence for the alternative. ## What moves F - **Bigger true differences between the means** raise `SS_between` and so raise F. - **More observations per group at fixed means** also raise F, because `SS_between` carries the `n_i` weight while `MS_within` keeps estimating the same sigma-squared. This is just power: the same effect becomes detectable with more data. - **Noisier measurements** raise `MS_within` and shrink F. Reducing measurement noise or blocking out a nuisance source of variability is often a cheaper route to a detectable result than collecting more units. - **More groups with nothing going on** raise k-1 without raising `SS_between` proportionally, diluting F. ## Common slips worth avoiding Using `N-1` as the denominator degrees of freedom is the most frequent error — that is the *total*, not the error. Treating F > 1 as automatically significant is another: with df (3, 36) the 5% critical value sits well above 2.8, and small samples make the F distribution's right tail long. Finally, keep sums of squares and mean squares distinct; comparing raw `SS_between` to `SS_within` without dividing by degrees of freedom compares apples to oranges.