F-statistic Calculator

Analysis of variance asks whether group means differ more than chance explains, and it answers with a single ratio: the F-statistic, the variance between groups divided by the variance within them. This calculator computes F from your ANOVA ingredients — the sums of squares and their degrees of freedom, or the mean squares directly — and completes the verdict with the exact p-value from the F distribution. Signal in the numerator, noise in the denominator, and a distribution to judge the quotient: the whole logic of ANOVA in one computation. If you hold an ANOVA table from any software, its last two columns can be audited here in under a minute.

A Ratio of Two Variances

ANOVA's insight is that "do the means differ?" can be asked as "is the spread of the means large against the spread within the groups?" The total variability of all observations splits exactly: SST = SSB + SSW. Divide each part by its degrees of freedom and the pieces become comparable mean squares — MSB, an estimate of variance inflated by any real group differences, and MSW, the pure noise estimate from inside the groups. Their ratio is F. Under the null (all group means equal), MSB and MSW estimate the same quantity and F hovers near 1; real differences swell only the numerator, driving F upward, and the F distribution — indexed by the pair (dfB, dfW), the pair this hub's degrees-of-freedom page exists to supply — prices how surprising the quotient is. The design's elegance is its indifference to which means differ: F is an omnibus alarm, deliberately silent about where the fire is.

Worked Example: Four Groups, One Alarm

The defaults: SSB = 120 on dfB = 3 (four groups), SSW = 200 on dfW = 36 (forty subjects in all). Mean squares: MSB = 120/3 = 40; MSW = 200/36 ≈ 5.556. The ratio: F(3, 36) = 7.20, with an exact p-value of 0.0007 — under the null, a quotient this large appears about seven times in ten thousand, and the alarm sounds. The effect size rides along: η² = 120/(120+200) = 0.375, group membership explaining 37.5% of the total variance — a substantial effect by the usual benchmarks, reported beside the p because (as everywhere on this hub) size and certainty are separate testimonies. What the example has not established is which of the four groups differ from which — six pairwise questions that F refuses on principle, and which belong to post-hoc procedures (Tukey, Scheffé, Bonferroni-corrected t-tests) whose multiplicity discipline exists because six uncorrected questions at 5% apiece manufacture a false alarm about a quarter of the time even when nothing is there.

F's Extended Family

The F-statistic recurs across statistics in costumes worth recognising. In regression, the overall F tests all slopes jointly (explained against unexplained mean squares) — and for a single predictor it is exactly the slope's t, squared: F(1, df) = t², an identity the hub's regression pages quietly rely on. In variance testing, the ratio of two sample variances is F-distributed, the classical (if fragile) test for equal spreads. In model comparison generally, nested models are judged by the extra-sum-of-squares F — the quadratic-versus-line bend test of the regression family is this page's statistic wearing a lab coat. And the distribution itself is a ratio of scaled chi-squares, which is why its two df both matter and why its right tail, though long, is the only tail that signifies: F near 0 is not evidence against the null in ANOVA's question, merely numerator silence — one of the few statistics whose small values are uninformative by construction.

Assumptions: The Fine Print With Teeth

Classical ANOVA prices its p under three assumptions, and the calculator's honesty requires naming them. Normality of the observations within each group — mercifully robust at moderate sizes by the central limit's grace. Homogeneity of variance: MSW is a pooled noise estimate, meaningful only if the groups share a common variance; with unequal spreads (especially with unequal group sizes) the F can mislead in either direction, and Welch's ANOVA adaptation is the standard remedy. Independence: repeated measures stuffed into a one-way ANOVA is the design error that no post-hoc virtue repairs — paired and repeated designs have their own F constructions with corrected denominators. Violations do not make the computed F "wrong" as arithmetic; they uncalibrate its p, the same distinction this hub draws everywhere between a statistic and its interpretation. Robustness checks (residual plots, Levene's test, the nonparametric Kruskal–Wallis alternative) are the working statistician's due diligence, and they precede the celebration of any p with three zeros.

Two Groups: Where F Is t, Squared

The identity deserves one numeric portrait. Take the paired-design t from elsewhere on this hub — t = 2.32 on df = 10 — and pose the same comparison as a two-group ANOVA: the F-statistic on (1, 10) degrees of freedom is 2.32² ≈ 5.38, and its p-value is the t-test's 0.043 to the digit, because with one numerator df the F distribution is exactly the square of t. The equivalence is a useful audit (two instruments, one verdict, or someone has erred) and the reason ANOVA can be introduced honestly as "the t-test, generalised past two groups without surrendering the error rate" — the generalisation costs the direction information a t carries in its sign, which is the first entry in the ledger of what omnibus tests trade away.

Reading an ANOVA Table in the Wild

Published ANOVA tables all wear the same five columns — Source, SS, df, MS, F, p — and the defaults let you rehearse the reading on known numbers: the Between row (120, 3, 40), the Within row (200, 36, 5.556), their quotient in the F column, and the p beside it. Audit any table you meet with the three identities this page embodies: MS must equal SS ÷ df in every row; the SS and df columns must add to their totals (320 and 39 here — forty observations, fully accounted); and F must equal the MS ratio. Tables that fail these checks are misprints or worse, and they occur in print more often than the profession enjoys admitting — the calculator doubles as the audit instrument, since any row's numbers can be re-entered and re-derived in seconds.

Related Tools

More from the Statistics Calculators hub: the Degrees of Freedom Calculator for the (dfB, dfW) pair, the Critical Value Calculator for z and t fences, and the t-statistic Calculator — whose square, on one numerator df, is this page's F.

Frequently Asked Questions

What does F = 7.2 actually mean?

That the between-group variance estimate is 7.2 times the within-group one. Under the null the ratio centres near 1 (its exact expectation is dfW/(dfW − 2), close to 1 for healthy dfW), so 7.2 is a large excursion — and the p-value converts "large" into a probability under the null's distribution for these specific degrees of freedom. F is a signal-to-noise ratio with a known null behaviour; both halves of that description are needed to read it.

Why is the p-value so small when η² is what I care about?

Because they answer different questions: p prices the evidence against "no difference at all" given this design's size; η² prices the difference's magnitude. With vast samples a trivial η² earns a tiny p; with small samples a large η² can miss significance entirely — the worked example, happily, has both substantial. Report the pair, always: significance without size is a headline without a story.

Can F be less than 1, and what then?

Yes — whenever the group means happen to sit closer together than within-group noise alone would predict, MSB < MSW and F < 1. It is unremarkable (under the null, values below 1 occur roughly half the time) and never significant for ANOVA's question: the test is right-tailed by design. An F suspiciously near zero across many studies is its own forensic curiosity (data too regular to be honest), but a single small F is just a quiet numerator.

What do I do after a significant F?

Ask where the difference lives, with multiplicity discipline: planned contrasts if the comparisons were designed in advance, Tukey's HSD for all pairwise comparisons, Scheffé for arbitrary contrasts — each controls the familywise error that raw t-tests across all pairs would inflate. The omnibus F's significance licenses the search; it does not conduct it. And if F is not significant, the etiquette is to stop: post-hoc fishing after a silent omnibus is the garden of forking paths with a gate left open.

How are the degrees of freedom for F determined?

dfB = k − 1 for k groups; dfW = N − k for N total observations — the calculator's inputs, and the pair the degrees-of-freedom page computes from the design directly. They are reported with every F (F(3, 36)) because the same F value carries different p-values under different pairs: F = 7.2 at (3, 36) gives p = 0.0007, while at (3, 6) it gives a far less dramatic verdict — the df are part of the evidence, not the formatting.

Is ANOVA just multiple t-tests done at once?

It is the disciplined alternative to them. Six raw t-tests among four groups each spend a 5% false-alarm budget, compounding to roughly 26% overall; ANOVA spends the budget once on the omnibus question, then licences corrected follow-ups. Mathematically, with two groups ANOVA is the t-test (F = t², identical p) — the machinery generalises the pair comparison without surrendering the error control that makes comparisons believable.