Statistics › Section M · The F distribution and ANOVA › Topic M.2
Facts about the F distribution
ANOVA's verdicts are read from the F distribution - a family of right-leaning curves built from ratios of variances. Meet the family before the two-variance test puts it to final use: two degrees of freedom, values near 1 when nothing is happening, and a long right tail when something is.
A distribution of ratios
The F distribution describes what happens when two independent variance-style estimates are divided. In ANOVA, \(MS_{\text{between}}\) and \(MS_{\text{within}}\) are both built from squared deviations averaged over their own degrees of freedom; if the null is true, they estimate the same noise, and their ratio wanders around 1. The F distribution is the curve such ratios follow - the reference curve for every question of the form "is this variance estimate bigger than that one?" Named in honour of R. A. Fisher, it is the last reference family this course meets, after the normal, the t and the chi-square.
Two parameters, always
Every F distribution is named by a PAIR of degrees of freedom - numerator first, denominator second: \(F(d_1, d_2)\), written \((d_1, d_2)\). The pair changes everything, because the numerator's estimate is noisier at small \(d_1\) and the denominator's wobble shrinks as \(d_2\) grows. ANOVA's teaching-methods F of 4.93 meant something only as 4.93 ON (2, 27); the same number on (2, 200) would be more extreme (\(p \approx 0.008\)), and on (2, 4) far less (\(p \approx 0.08\)). The pair travels with the ratio, always.
Shape: leaning right from zero
F values are never negative - a ratio of squared quantities - so the curve starts at zero and stretches a long right tail: ordinary ratios cluster near 1, and huge ratios are rare but possible. For \(d_1 > 2\) it rises to a peak a little below 1; for \(d_1 = 1\) or 2 it is highest at zero and falls away. The curve is right-skewed at every df, with the skew easing as the degrees of freedom grow: at large \(d_1\) and \(d_2\) the curve concentrates tightly around 1, because both estimates become precise and their ratio stops wandering. Small-numerator F's (like ANOVA's \(d_1 = 2\)) stay visibly skewed - the tail that ANOVA's verdicts are read from.
The mean
For \(d_2 > 2\), the F distribution's mean is
- set by the DENOMINATOR alone. At \((7, 6)\): mean \(6/4 = 1.5\). At \((2, 27)\): mean \(27/25 = 1.08\), close to 1 and pulled above it by the skew. At \((10, 100)\): mean \(100/98 \approx 1.02\), nearly 1 - the large-\(d_2\) curve concentrating, as promised. The formula gives the same instant reading habit as chi-square's mean = df: locate the statistic against the curve's centre before any table.
Worked example: reading F against its centre
The reading habit assembles fast now. First attach the df pair. Then locate the centre: \(d_2/(d_2-2)\), near 1 for any decent denominator df. Then judge the distance: an F of 2 at \((5, 40)\) sits modestly right of centre (\(p \approx 0.10\), plausible noise); an F of 4.93 at \((2, 27)\) sits far out the tail - the methods' verdict, p ≈ 0.015; an F of 1.1 at \((3, 200)\) is dead centre, nothing to see. Software supplies exact tail areas; the centre-and-sketch habit catches absurdities before they reach a report.
Where F appears
Two jobs in this course, and many beyond. ANOVA (the previous topic): \(F = MS_{\text{between}}/MS_{\text{within}}\), testing equality of several means. The two-variance test (next topic): \(F = s_1^2/s_2^2\), testing equality of two spreads. Beyond this course: regression's overall significance, model comparisons, quality control. The family reunites them all: whenever a question becomes "is this variance bigger than that one?", the answer is an F ratio read against a named curve.
Worked example: one statistic, three curves
A study reports \(F = 3.0\). Three colleagues read it differently. a) On \((2, 27)\) - an ANOVA of three groups with 30 observations: centre \(27/25 = 1.08\), and 3.0 sits well out: \(p \approx 0.067\) - suggestive, not conclusive. b) On \((10, 60)\) - eleven groups, larger study: the bigger dfs tighten the curve, and the same 3.0 gives \(p \approx 0.004\) - clearly significant. c) On \((20, 200)\): the curve concentrates near 1, and 3.0 is overwhelming - \(p < 0.001\). Check: one statistic, three verdicts, from suggestive to overwhelming - the df pair is not a footnote but half the result, and larger studies give the same ratio more weight.
Common mistakes to avoid
First, quoting F without its df pair - the unreadable result. Second, reading F symmetrically: ANOVA's test, and the larger-on-top two-variance test, consult only the right tail, and values below the centre are simply ordinary. Third, forgetting the mean is \(d_2/(d_2-2)\), slightly above 1 - "F near 1" means near that centre, not exactly 1. Fourth, swapping the pair: \(F(2, 27)\) and \(F(27, 2)\) are different curves, and the ratio's direction matters (numerator on top, always). Fifth, expecting an F table to work without both dfs - it is indexed by the pair. Sixth, inventing an F of 0.5 as "half the evidence": F below 1 is not half of anything; it is a ratio whose numerator lost.
Key takeaways
The F distribution is the curve of variance-estimate ratios: never negative, right-skewed, clustered near 1, named by the df pair \((d_1, d_2)\) - numerator first. The mean is \(d_2/(d_2-2)\) for \(d_2 > 2\): 1.5 at (7, 6), 1.08 at (2, 27), nearly 1 at large \(d_2\), where the curve concentrates. Larger df shrink the skew; small \(d_1\) keeps the tail long. Reading habit: pair, centre, distance - the same F = 3.0 runs suggestive at (2, 27), significant at (10, 60), overwhelming at (20, 200). ANOVA and the next topic's two-variance test are the family's two jobs here.
Flashcards
Flashcard
Numerator estimate over denominator estimate
Test your understanding
Question 1 of 5
Every F distribution is fully named by:
Question 2 of 5
The mean of an F(d1, d2) distribution (d2 > 2) is:
Question 3 of 5
The mean of F(7, 6) is:
Question 4 of 5
An F statistic of 4.93 at (2, 27) sits:
Question 5 of 5
F values below 1 mean:
