The Overall F Test for Regression

Why Test the Regression Model as a Whole?

A multiple regression model can include several predictors, and a question about the model as a whole is different from a question about one coefficient. For example, a hypothetical insurer might model annual claim cost using age, vehicle characteristics, and prior claim count. The overall F test asks whether the predictors, considered together, provide evidence of a linear association with expected claim cost. It does not identify which predictor is responsible for that evidence.

The test compares the fitted regression model with an intercept-only model. The intercept-only model predicts the same response mean for every observation; the full model can use the predictors to explain variation around that mean. The comparison is summarized by the amount of variation explained per predictor relative to the residual variation per residual degree of freedom.

Definition: For a regression with $n$ observations and $p$ predictors, the overall F test tests whether all $p$ slope coefficients are zero. With an intercept in the model, its null and alternative hypotheses are $$ H_0:\beta_1=\beta_2=\cdots=\beta_p=0 \qquad\text{versus}\qquad H_A:\text{at least one }\beta_j\ne 0. $$ The test is about linear association in the conditional mean. Rejection is evidence against the intercept-only model, not proof of causation or of a particular predictor’s importance.

Statistic, Degrees of Freedom, and Assumptions

Let $SSE$ be the residual sum of squares from the full model and $SSR$ the regression sum of squares, measured relative to the intercept-only model. With an intercept, $SSR=SST-SSE$, where $SST$ is the total sum of squares about the sample mean. The overall statistic is

F=\frac{SSR/p}{SSE/(n-p-1)}.

The numerator degrees of freedom are $p$, because the full model adds $p$ slope parameters beyond the intercept-only model. The denominator degrees of freedom are $n-p-1$, the residual degrees of freedom in the full model. The statistic is large when the full model explains substantial variation relative to the remaining residual variation.

For the usual exact F test, use the following conditions: the predictor values are fixed or the analysis is conditional on them; the model is linear in its coefficients and includes an intercept; the design matrix has full column rank; $p\ge 1$; $n>p+1$; and the errors are independent, normally distributed with mean zero and common variance $\sigma^2$. The condition $n>p+1$ is essential: it ensures positive residual degrees of freedom, so $SSE/(n-p-1)$ is defined. Under these conditions and $H_0$, the statistic has an $F$ distribution with $p$ and $n-p-1$ degrees of freedom. If the assumptions do not hold, the usual exact reference distribution need not apply.

Interpretation: At significance level $\alpha$, reject $H_0$ when the observed $F$ exceeds the upper-tail critical value $F_{\alpha;p,n-p-1}$, or equivalently when the p-value is at most $\alpha$. Rejection says that the data provide evidence that at least one slope is nonzero. It does not say that all slopes are nonzero.

Why the Statistic Has an F Distribution

The F reference distribution follows from decomposing the response into orthogonal fitted components and residuals. This also explains why the test uses two mean squares rather than comparing $SSR$ and $SSE$ without adjusting for their degrees of freedom.

Represent the fits as projections.
Let $P_X$ be the projection matrix onto the columns of the full design matrix, including the intercept, and let $P_0$ project onto the intercept-only model. Then $SSR=Y^\mathsf{T}(P_X-P_0)Y$ and $SSE=Y^\mathsf{T}(I-P_X)Y$.

Use the null model.
Under $H_0$, the mean response lies in the intercept-only model. With independent normal errors of variance $\sigma^2$, each quadratic form divided by $\sigma^2$ has a chi-square distribution: $SSR/\sigma^2$ has $p$ degrees of freedom, and $SSE/\sigma^2$ has $n-p-1$ degrees of freedom.

Use orthogonality and form the ratio.
The added-predictor fitted space and the residual space are orthogonal. The two chi-square quantities are therefore independent. Dividing each by its degrees of freedom and taking their ratio gives $F=(SSR/p)/(SSE/(n-p-1))$, with an $F_{p,n-p-1}$ distribution under $H_0$.

This argument requires positive degrees of freedom for both parts of the ratio. Full column rank gives the stated ranks $p$ and $n-p-1$, but it does not by itself guarantee positive residual degrees of freedom. The additional condition $n>p+1$ is therefore part of the test’s assumptions.

Worked Calculations

Worked Example: Testing a Two-Predictor Claim-Cost Model

A hypothetical analyst fits an intercept and two predictors to $n=12$ policy records. Suppose $SSR=48$ and $SSE=36$, and assume the conditions for the usual exact test hold. At $\alpha=0.05$, test whether both slopes are zero. Use the supplied critical value $F_{0.05;2,9}=4.256$.

Method. Apply the overall F statistic. There are $p=2$ predictors, so the numerator degrees of freedom are 2 and the residual degrees of freedom are $12-2-1=9$, which is positive.

Calculate the mean squares. The regression mean square is $SSR/p=48/2=24$. The residual mean square is $SSE/(n-p-1)=36/9=4$.

Calculate and compare. Thus $F=24/4=6.000$. Since $6.000>4.256$, reject $H_0$ at the 5% level. The data provide evidence that at least one of the two slopes is nonzero. For these degrees of freedom, the upper-tail p-value is approximately $0.0221$, which is also below $0.05$.

Check and trap avoided. The residual degrees of freedom are $9$, not $12-2=10$, because one degree of freedom is used for the intercept. The conclusion is about at least one slope, not necessarily both.

Worked Example: Recovering the Test From R-Squared

A hypothetical model of monthly service cost uses four predictors and an intercept for $n=25$ accounts. Its $R^2$ is $0.48$. Assume the exact-test conditions hold. Test the overall null at $\alpha=0.05$, using $F_{0.05;4,20}=2.87$.

Method. With an intercept, $R^2=SSR/SST$, so $SSR=R^2SST$ and $SSE=(1-R^2)SST$. The common factor $SST$ cancels in the ratio: $$ F=\frac{R^2/p}{(1-R^2)/(n-p-1)}. $$

Substitute the values. The numerator is $0.48/4=0.12$. The residual degrees of freedom are $25-4-1=20$, and the denominator is $(1-0.48)/20=0.52/20=0.026$.

Calculate and decide. Therefore $F=0.12/0.026=4.6154$, or $4.615$ to three decimal places. Since $4.615>2.87$, reject the overall null at the 5% level. There is evidence that at least one slope is nonzero.

Check and trap avoided. The denominator uses the unexplained fraction $1-R^2$ divided by 20 residual degrees of freedom. Dividing it by $n$, or omitting the intercept when calculating degrees of freedom, would give the wrong statistic.

Worked Example: A Small F Statistic

For a hypothetical model of processing time, $n=15$, $p=2$, $SSR=10$, and $SSE=50$. Assume the exact-test conditions hold. At $\alpha=0.05$, use $F_{0.05;2,12}=3.885$ to test whether both slopes are zero.

Method. Calculate the regression and residual mean squares, then compare their ratio with the supplied upper-tail critical value.

Calculate the degrees of freedom and mean squares. The numerator degrees of freedom are $2$, and the residual degrees of freedom are $15-2-1=12$. The regression mean square is $10/2=5$, while the residual mean square is $50/12=4.1667$.

Calculate and decide. The statistic is $F=5/4.1667=1.200$, to three decimal places. Since $1.200<3.885$, do not reject $H_0$ at the 5% level. These data do not provide sufficient evidence of a linear association from the predictors as a group.

Check and trap avoided. A failure to reject does not prove that every slope is exactly zero; it means the observed evidence is not strong enough to reject the joint null at the chosen level. Here $SSR$ is positive, but the relevant comparison is its mean square against the residual mean square.

What Rejection Does—and Does Not—Establish

The overall test is useful when the inferential question is whether a collection of predictors contributes any linear explanatory signal. Its joint null is also distinct from the individual coefficient tests covered in the earlier material on regression inference. An overall rejection can occur even when no single coefficient test is individually compelling, particularly when predictors are related to one another. Conversely, a non-rejection does not establish that the true slopes are all zero.

The test assesses the specified linear mean model. A rejection does not establish that the model form is appropriate, that predictions will be accurate outside the observed predictor range, or that a predictor causes the response. It also does not measure practical importance: a statistically detectable association may be too small to matter for a decision. Those judgments require context and additional model assessment.

A common exam trap is to write “all predictors are significant” after rejecting the overall null. The alternative hypothesis is a union: at least one slope differs from zero. The overall test alone does not identify which slope or slopes; that requires separate, appropriately framed inference. Another trap is to compute the formula when residual degrees of freedom are zero. If $n=p+1$, then $n-p-1=0$, so $SSE/(n-p-1)$ and the usual overall F statistic are undefined. Full column rank does not repair this problem.

Key takeaway: For a regression with an intercept, $p\ge1$ predictors, full column rank, and $n>p+1$, the overall statistic is $F=(SSR/p)/(SSE/(n-p-1))$. Under the usual independent, normal, constant-variance error assumptions, it has an $F_{p,n-p-1}$ distribution under the null that all slopes are zero. Rejection supports a linear association for at least one slope, not a claim that every predictor matters.

Check Your Understanding

For each problem, use the stated critical value and assume the exact-test conditions hold unless the problem asks you to assess whether the test is defined.

A hypothetical model has $n=18$, $p=2$, $SSR=30$, and $SSE=60$. Find the overall F statistic and its degrees of freedom. (Answer: $F=3.750$, with degrees of freedom $2$ and $15$.)
A hypothetical account model has $n=25$, $p=4$, and $R^2=0.48$. Using $F_{0.05;4,20}=2.87$, calculate the statistic and state the decision at the 5% level. (Answer: $F=4.615$; reject the null that all four slopes are zero.)
A model has $n=14$, $p=3$, $SSR=24$, and $SSE=40$. Using $F_{0.05;3,10}=3.71$, calculate the statistic and state the decision. (Answer: $F=2.000$; do not reject at the 5% level.)
A hypothetical model has $n=30$, $p=5$, and $R^2=0.50$. Using $F_{0.05;5,24}=2.62$, calculate the overall statistic and interpret rejection. (Answer: $F=4.800$; reject at the 5% level, giving evidence that at least one slope is nonzero.)
A regression has $n=8$ observations and $p=7$ predictors plus an intercept. Even if the design matrix has full column rank, can the usual overall F statistic be calculated? Explain using the residual degrees of freedom. (Answer: No; $n-p-1=8-7-1=0$, so the residual mean square and usual F statistic are undefined.)

What You'll Learn