The Overall F Test for a Multiple Regression Model

Why Test a Regression Model as a Whole?

In multiple regression, each coefficient describes the relationship between the response and one predictor while holding the other predictors fixed. An individual coefficient test asks whether one particular slope is zero. The overall F test asks a broader question: do the non-intercept coefficients, considered together, provide evidence that the response’s conditional mean changes with at least one predictor?

This is useful when a model contains several predictors and the first question is whether there is evidence of any linear signal in the set. A significant overall test does not identify which predictor is associated with the response; that requires examining the individual coefficient tests and their context.

Definition: For a multiple regression model with an intercept and \(m\) non-intercept coefficients, the overall F test tests \(H_0:\beta_1=\beta_2=\cdots=\beta_m=0\) against \(H_a:\) at least one of \(\beta_1,\ldots,\beta_m\) is not zero. The test statistic is the mean square for regression divided by the mean square error.

Here \(m\) counts the non-intercept coefficients in the fitted model. A categorical variable represented by several indicator variables contributes several coefficients to this count. The null hypothesis says that the conditional mean does not depend linearly on any of the included predictors; it does not say that every observed response is identical.

Model, Assumptions, and the F Statistic

For observations \(i=1,\ldots,n\), write the multiple regression model as

Y_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_mx_{im}+\varepsilon_i.

The usual exact overall F test assumes that, conditional on the predictor values, the errors are independent and normally distributed with mean zero and a common variance \(\sigma^2\). The mean model must be appropriately linear in its coefficients, and the full design matrix, including the intercept column, must have full column rank with \(n>m+1\) so that the residual degrees of freedom are positive. These conditions imply that the fitted model is estimable and that the stated F distribution applies under the null. In practice, the study design and residual diagnostics are used to judge whether the assumptions are reasonable; they cannot be proved merely by calculating the statistic.

Let \(\hat{Y}_i\) be the fitted response and \(\bar{x}\) denote the sample mean of the response values. The regression sum of squares measures variation of the fitted values around the response mean; the error sum of squares measures variation of observations around their fitted values. With an intercept, the total sum of squares separates into these two parts:

SSR=\sum_{i=1}^{n}(\hat{Y}_i-\bar{x})^2,\qquad SSE=\sum_{i=1}^{n}(Y_i-\hat{Y}_i)^2,\qquad SST=SSR+SSE.

The regression degrees of freedom are \(m\), one for each non-intercept coefficient. The residual degrees of freedom are \(n-m-1\), because the model estimates \(m\) slopes and one intercept. Dividing each sum of squares by its degrees of freedom gives a mean square.

MSR=\frac{SSR}{m},\qquad MSE=\frac{SSE}{n-m-1},\qquad F=\frac{MSR}{MSE}.

The statistic is large when the variation attributed to the fitted predictors, per regression degree of freedom, is large relative to the residual variation per residual degree of freedom. The test is upper-tailed: evidence against \(H_0\) comes from large values of \(F\). Under \(H_0\) and the assumptions above, \(F\) follows an F distribution with \(m\) numerator degrees of freedom and \(n-m-1\) denominator degrees of freedom.

Why the Statistic Has an F Distribution

The distribution follows from how the model divides the response into fitted and residual components. This justification also explains why the test uses two degrees of freedom counts.

Separate fitted and residual variation.
With an intercept, the fitted deviations from the response mean and the residuals are orthogonal, so \(SST=SSR+SSE\).

Use the null model.
Under \(H_0\), the slopes are zero. After removing the intercept, the fitted predictor component has \(m\) independent dimensions. Normal errors therefore give \(SSR/\sigma^2\) a chi-square distribution with \(m\) degrees of freedom.

Account for the residual dimensions.
Estimating the intercept and \(m\) slopes leaves \(n-m-1\) residual dimensions. Thus \(SSE/\sigma^2\) has a chi-square distribution with \(n-m-1\) degrees of freedom, independent of \(SSR/\sigma^2\).

Form the ratio.
Dividing each independent chi-square quantity by its degrees of freedom and taking their ratio gives \(F=(SSR/m)/(SSE/(n-m-1))\), which has the stated F distribution under \(H_0\).

The independence and chi-square statements in this argument rely on the normal-error model and the regression projection geometry. If the assumptions are seriously violated, the usual reported F-test p-value may not have its advertised calibration.

Worked Examples

Worked Example: Calculating the Overall Test From Sums of Squares

Suppose a hypothetical regression uses study hours and class attendance to predict exam score for \(n=24\) students. Its ANOVA summary gives \(SSR=40\) and \(SSE=210\), measured in squared exam-score points. Test the overall model at \(\alpha=0.05\).

There are \(m=2\) non-intercept coefficients, so the hypotheses are \(H_0:\beta_1=\beta_2=0\) versus \(H_a:\) at least one of these slopes is nonzero. The degrees of freedom are \(2\) for regression and \(24-2-1=21\) for error.

Calculate the mean squares and statistic.

MSR=\frac{40}{2}=20,\qquad MSE=\frac{210}{21}=10,\qquad F=\frac{20}{10}=2.000.

For an \(F\) distribution with numerator degrees of freedom 2 and denominator degrees of freedom \(v\), the upper-tail probability at \(F=f\) is \(\left(\frac{v}{v+2f}\right)^{v/2}\). Here the p-value is

P(F_{2,21}\ge 2.000) =\left(\frac{21}{21+2(2.000)}\right)^{21/2} =(0.84)^{10.5}\approx 0.1603.

Since \(0.1603>0.05\), we fail to reject \(H_0\). These data do not provide sufficient evidence at the 5% level that study hours and attendance, considered together, are linearly related to exam score. This is not proof that both population slopes equal zero.

For this hypothetical analysis, suppose the students were independently sampled, the predictors were not perfectly collinear, and the residual plots and normal probability plot showed no substantial departures from linearity, constant variance, independence, or approximate normality. Those checks make the standard test more defensible, though they cannot guarantee that the model is correct.

Worked Example: Interpreting a Significant Overall Test

A hypothetical analysis predicts a clinic patient’s recovery time, in days, from age and a baseline severity score. For \(n=24\) patients, the software reports \(SSR=72\), \(SSE=168\), \(F=4.50\), and \(p=0.0236\). Assume the usual regression conditions are reasonable based on the study design and residual checks.

There are \(m=2\) slopes and \(n-m-1=21\) residual degrees of freedom. The reported statistic can be checked from the sums of squares:

MSR=\frac{72}{2}=36,\qquad MSE=\frac{168}{21}=8,\qquad F=\frac{36}{8}=4.50.

The reported p-value is consistent with this statistic:

P(F_{2,21}\ge 4.50) =\left(\frac{21}{21+2(4.50)}\right)^{21/2} =\left(\frac{21}{30}\right)^{10.5} \approx 0.0236.

Because \(0.0236<0.05\), reject \(H_0\) at the 5% significance level. There is evidence that at least one of the age and baseline severity slopes is nonzero in the specified model for mean recovery time. The overall result alone does not show which slope is nonzero, how large either association is, or whether either predictor causes recovery time to change.

Worked Example: Reading the F Test in Software Output

A hypothetical study models a follow-up symptom score using baseline symptom score and treatment assignment, with \(n=34\) participants. The software’s overall row reports \(F=3.90\) with regression degrees of freedom 2 and residual degrees of freedom 31. Interpret this result at \(\alpha=0.05\).

The residual degrees of freedom agree with the sample size and model: \(34-2-1=31\). For the reported F statistic, the upper-tail probability is approximately

P(F_{2,31}\ge 3.90) =\left(\frac{31}{31+2(3.90)}\right)^{31/2} =\left(\frac{31}{38.8}\right)^{15.5} \approx 0.0308.

Software typically rounds this to \(p=0.031\). Since \(0.031<0.05\), reject the null that both non-intercept coefficients are zero. The model gives evidence that at least one of baseline symptom score and treatment assignment is associated with the mean follow-up symptom score, conditional on the other predictor in this model. This test does not by itself establish that treatment assignment has an effect; its individual coefficient and the study design matter for that question.

Reading an ANOVA Summary and Reporting the Result

Regression software often presents the overall F test in an analysis-of-variance (ANOVA) summary. The row labelled “Regression,” “Model,” or “Regression/Model” usually supplies \(SSR\), \(m\), \(MSR\), and \(F\). The “Residual” or “Error” row supplies \(SSE\), \(n-m-1\), and \(MSE\). The p-value is the probability, assuming \(H_0\) and the model conditions, of obtaining an F statistic at least as large as the observed one.

Software entry	Meaning for the overall test
Regression/model df	Number \(m\) of non-intercept coefficients being tested
Residual/error df	\(n-m-1\), the denominator degrees of freedom
F	Observed ratio \(MSR/MSE\)
Significance F or model p-value	Upper-tail probability under the null F distribution

A clear report names the response and predictors, gives the statistic with both degrees of freedom and the p-value, and states the conclusion about the set of slopes. For example: “The overall regression of recovery time on age and baseline severity was statistically significant, \(F(2,21)=4.50\), \(p=0.0236\); there is evidence that at least one slope differs from zero.” A non-significant result should be reported as insufficient evidence against the joint null, not as proof that the model has no relationship whatsoever.

Common Mistakes and Why the Test Matters

Claiming every predictor is significant. Rejecting the joint null means at least one tested slope is nonzero; it does not mean all are. Look at individual coefficient estimates and tests to address specific predictors.
Calling the p-value the probability that the null is true. The p-value is calculated assuming the null and model conditions; it is the probability of an F statistic at least this large, not the probability that \(H_0\) is correct.
Using the wrong degrees of freedom. The numerator degrees of freedom are \(m\), while the denominator degrees of freedom are \(n-m-1\). Counting only the predictors by name can be wrong when a categorical predictor uses several coefficients.
Interpreting significance as causation or practical importance. The test detects evidence of a linear association in the model, not a causal effect or a large, useful predictive gain. Design, coefficient estimates, uncertainty, and subject-matter context remain important.
Assuming a large F value is meaningful without checking the model. The reference distribution relies on regression assumptions. A p-value cannot repair a poorly specified mean model or dependence in the observations.

The overall F test is a compact test of a precise joint claim: whether all non-intercept coefficients are zero. It provides a useful first assessment of the fitted model while leaving the identities, sizes, and practical meaning of individual relationships to other parts of regression interpretation.

Check Your Understanding

Answer each question using the definitions and procedures above.

A model has \(n=40\) observations and five non-intercept coefficients. What are the numerator and residual degrees of freedom for its overall F test?
In context, explain the difference between the overall null hypothesis and the alternative hypothesis.
A software output gives \(F(3,26)=1.10\) and \(p=0.367\). What conclusion is appropriate at \(\alpha=0.05\), and what conclusion would be too strong?
Why does a statistically significant overall F test not imply that every predictor has a nonzero coefficient?
Name two model assumptions needed for the usual exact F test and describe what each means in a regression setting.