How to Interpret the F-Statistic

Q: Can an F-statistic be negative?

No. Both the numerator and denominator are mean squares — sums of squared deviations divided by degrees of freedom — and a sum of squares can never be negative. The smallest an F-statistic can be is 0, which happens only if the explained variance is exactly zero (the group means are identical, or the model’s fitted values are all equal to the outcome’s grand mean).

Q: What does an F-statistic close to 1 mean?

It means the explained variance and the unexplained variance are roughly the same size — the pattern you’re testing (group differences, or a predictor’s relationship with the outcome) is not producing more spread than you’d expect from noise alone. Under the null hypothesis that there’s no real effect, F is expected to average close to 1 (technically, close to dfW/(dfW−2) for the within-groups mean square’s expectation), so values near 1 are exactly what “nothing going on” looks like.

Q: Why does adding more predictors to a regression change the F-statistic even if R² barely moves?

Because F depends on degrees of freedom, not just on R² alone. Adding predictors increases dfR (spending more numerator degrees of freedom) and decreases dfE (leaving fewer denominator degrees of freedom), so even a predictor that adds almost nothing to R² can pull the overall F down — which is exactly why an adjusted R² and each individual predictor’s own significance test matter alongside the omnibus F, not instead of it.

Q: Is a bigger F-statistic always better?

Bigger means stronger evidence against the null hypothesis of “no real explained variance,” not a better-fitting or more useful model. An enormous F on a huge sample can still describe a trivially small, practically unimportant effect; a modest F on a small, well-controlled study can accompany a genuinely large effect. F answers a statistical-significance question, not a practical-importance one — pair it with an effect size before drawing a substantive conclusion.

The F-statistic is a ratio of two variance estimates: how much of the spread in your data is explained by the model or the group differences you tested, divided by how much spread is left over as unexplained noise. A large F means the explained variance dwarfs the unexplained variance — the pattern you found is a strong signal relative to random scatter. A value near 1 means the explained variance is no bigger than what you’d expect from noise alone. This is the same logic a t-statistic uses — a signal-to-noise ratio — just built from squared, pooled variances across possibly more than two groups or predictors instead of a single mean difference. This guide covers what the ratio actually contains, why it always carries two separate degrees-of-freedom values, and works through both a one-way ANOVA and a simple regression by hand so every number in a typical software printout is traceable to its source.

The Ratio Itself: Explained Variance Over Unexplained Variance

Every F-statistic, regardless of which procedure produced it, has the same structure:

F = variance explained by the model or grouping ÷ variance left unexplained

The two pieces are never raw sums of squared deviations (SS) — they’re mean squares (MS), meaning each sum of squares has already been divided by its own degrees of freedom. That division matters: it converts a sum of squares, which mechanically grows with sample size and with how many groups or predictors are in the model, into an average variance per degree of freedom, which is what actually makes the two pieces comparable as a ratio. In one-way ANOVA the label on the two pieces is between-groups variance (does the group you’re in explain some of the spread in scores?) over within-groups variance (how much do scores vary even among people in the same group?). In regression the same ratio is relabeled model (or regression) variance over residual variance — does knowing the predictor values explain some of the spread in the outcome, versus how much spread is left in the residuals after the model’s best fit line is drawn? The arithmetic is identical; only the names of the two variance sources change with the procedure.

Why F Always Comes With Two Degrees-of-Freedom Values

A t-statistic is reported with one degrees-of-freedom number because it’s built from a single variance estimate. An F-statistic is a ratio of two independent variance estimates, and each one has its own degrees of freedom, so F always needs both:

Numerator df (often written df1, or df between/df regression) — the degrees of freedom spent estimating the explained-variance piece. In one-way ANOVA with k groups, this is k−1. In simple regression with p predictors, this is p.
Denominator df (df2, or df within/df residual) — the degrees of freedom left over for the unexplained-variance piece, after the numerator and the grand mean have used theirs up. In one-way ANOVA with N total observations across k groups, this is N−k. In simple regression with N observations and p predictors plus an intercept, this is N−p−1.

Software output conventionally writes this as F(df1, df2) = value — for example F(2, 12) = 15.25. Both numbers are load-bearing for interpretation: df1 tells you how many things were compared (groups, or predictors), and df2 tells you how much data was left to estimate the noise floor those comparisons are being judged against. A large F built on a tiny df2 is far less trustworthy than the same F built on a generous denominator df, because the noise estimate itself is shakier — the same reason a t-test on 4 observations per group is a weaker basis for a claim than the identical mean difference measured on 400.

A Worked One-Way ANOVA Example

Three groups of five students (illustrative data, computed directly for this guide, not a real study) each tried a different study technique before an identical 100-point quiz:

Group	Scores	Group mean
A — reread notes	72, 75, 68, 74, 71	72.00
B — spaced retrieval practice	81, 78, 85, 79, 82	81.00
C — highlight & summarize	77, 73, 79, 76, 75	76.00

The grand mean across all 15 scores is 76.333. The between-groups sum of squares weights each group’s deviation from the grand mean by its size: SSB = 5(72.00−76.333)² + 5(81.00−76.333)² + 5(76.00−76.333)² = 203.333. The within-groups sum of squares adds up every individual score’s deviation from its own group mean: SSW = 80.000. With k=3 groups and N=15 observations:

Source	SS	df	MS
Between groups	203.333	2	101.667
Within groups	80.000	12	6.667
Total	283.333	14	—

F(2, 12) = 101.667 ÷ 6.667 = 15.250, which corresponds to p < .001 — the between-groups variance is over 15 times the within-groups variance, far more than random group assignment would produce on its own. (This p-value was computed directly from the F(2,12) distribution via the regularized incomplete beta function, the same method statistical software uses, not looked up or estimated.) The proportion of total variance the grouping explains — eta-squared, SSB ÷ SST — is 203.333 ÷ 283.333 = 0.718: a large effect by conventional benchmarks, and a reminder that F tells you the ratio is unlikely to be noise, while eta-squared or Cohen’s f tells you how much that difference actually matters in size. F answers “is there a real signal here,” not “how big is it” — those are two different questions with two different statistics.

The Same Ratio in Regression: Model Variance Over Residual Variance

A regression F-test asks the equivalent question about a fitted line or model instead of a set of group means: does knowing the predictor(s) explain more of the variance in the outcome than you’d expect from a model that used no predictors at all (just the outcome’s own mean)? The overall-model F reported at the top of a regression output tests exactly that, and it is built from the identical MS-ratio structure as ANOVA’s F:

SSR (regression, or model, sum of squares) — how far the model’s fitted values fall from the outcome’s grand mean, summed and squared. This is the “explained” piece, playing the same role SSB played in ANOVA.
SSE (residual, or error, sum of squares) — how far each actual observation falls from the model’s fitted value for it. This is the “unexplained” piece, playing the same role SSW played in ANOVA.

MSR = SSR ÷ p (predictors) and MSE = SSE ÷ (N−p−1), and F = MSR ÷ MSE — the same explained-over-unexplained construction, just computed from squared deviations around a fitted line instead of around group means.

A Worked Regression Example

Eight illustrative observations of hours studied (x) against quiz score (y):

Hours (x)	1	2	3	3	4	5	6	7
Score (y)	58	63	65	70	72	78	84	88

Ordinary least squares on this data gives a fitted line of ŷ = 52.489 + 5.100x. Decomposing the total sum of squares (SST = 765.500) around that fit: SSR = 750.911 and SSE = 14.589. With p=1 predictor and N=8 observations, dfR = 1 and dfE = N−p−1 = 6:

Source	SS	df	MS
Regression	750.911	1	750.911
Residual	14.589	6	2.431
Total	765.500	7	—

F(1, 6) = 750.911 ÷ 2.431 = 308.83, p < .0001. Notice R² here is SSR ÷ SST = 0.9809 — and R² and F are not two independent facts about the same model, they’re algebraically the same information in different units: F = [R² ÷ (1−R²)] × (dfE ÷ dfR). Plugging in confirms it exactly: [0.9809 ÷ 0.0191] × (6 ÷ 1) = 308.83, matching the F computed directly from the sums of squares. This is why a regression printout’s overall-model F-test and its R² always move together — a high R² on a reasonable sample size necessarily produces a large F, because they’re the same explained-versus-unexplained comparison expressed two different ways: R² as a proportion of total variance, F as a ratio of the explained and unexplained pieces directly.

Why F Is a Signal-to-Noise Ratio, the Same Way t Is

A t-statistic is, mechanically, a signal (an observed mean difference, or a regression coefficient) divided by an estimate of that signal’s own standard error — noise. F generalizes the identical idea from one comparison to several at once: instead of one signal divided by one noise estimate, it’s a variance attributable to systematic differences (across groups, or across predictors) divided by a variance attributable to everything else, i.e. chance fluctuation. That’s why an F-test with exactly one numerator degree of freedom (one predictor, or two groups) and t-test on the same comparison are mathematically the same test wearing different clothes: F(1, df) = t(df)² exactly, every time — you can confirm this on the standard F(1,10) critical value at α=.05, 4.965, against the standard two-tailed t(10) critical value at α=.05, 2.228: 2.228² = 4.964, matching to rounding. F only earns its own separate identity once there are more than two things being compared at once (three or more group means, or more than one predictor), where a single t no longer applies but the same explained-over-unexplained logic still does.

What the F-Statistic Does Not Tell You

Which specific groups differ. A significant one-way ANOVA F says some pair of group means differs by more than noise would predict — it does not say which pair. That requires a dedicated post hoc procedure run after the omnibus F, such as the ones covered in CASRAI’s one-way ANOVA and two-way ANOVA guides.
How large or practically meaningful the effect is. F (and its p-value) is a function of both effect size and sample size — a trivial effect can still produce a large F given enough data, and a substantial effect can produce a small, non-significant F on a small sample. Report an effect size (eta-squared, Cohen’s f, or R²) alongside F, not instead of it — see CASRAI’s effect size guide.
Which individual predictor matters, in a multi-predictor regression. The overall-model F in multiple regression tests whether the predictors collectively explain more variance than chance; it says nothing about any single predictor’s own contribution. That’s the job of each predictor’s own t-test and coefficient, covered in CASRAI’s multiple regression guide.
Whether the model’s assumptions hold. F from ANOVA or regression assumes independent observations, roughly normal residuals, and homogeneous variance across groups or across the range of fitted values. A large F computed on data that badly violates those assumptions is not a trustworthy signal-to-noise ratio, regardless of its size.

Frequently Asked Questions

Can an F-statistic be negative?

No. Both the numerator and denominator are mean squares — sums of squared deviations divided by degrees of freedom — and a sum of squares can never be negative. The smallest an F-statistic can be is 0, which happens only if the explained variance is exactly zero (the group means are identical, or the model’s fitted values are all equal to the outcome’s grand mean).

What does an F-statistic close to 1 mean?

It means the explained variance and the unexplained variance are roughly the same size — the pattern you’re testing (group differences, or a predictor’s relationship with the outcome) is not producing more spread than you’d expect from noise alone. Under the null hypothesis that there’s no real effect, F is expected to average close to 1 (technically, close to dfW/(dfW−2) for the within-groups mean square’s expectation), so values near 1 are exactly what “nothing going on” looks like.

Why does adding more predictors to a regression change the F-statistic even if R² barely moves?