Why Test a Block of Coefficients?
A regression model can include several predictors whose coefficients answer related questions. Sometimes the question is not whether any one coefficient differs from zero on its own, but whether a specified group of coefficients is jointly zero after accounting for the other predictors. A partial F-test answers that joint question by comparing a reduced model with a full model that includes the entire block.
For example, suppose a model includes two measurements, \(x_1\) and \(x_2\), alongside other predictors. The block test asks whether the data provide evidence that the coefficients of both \(x_1\) and \(x_2\) are not simultaneously zero, conditional on the other predictors in the model. It does not test each coefficient separately, and its conclusion need not match the conclusions from separate coefficient tests.
Models, Hypotheses, and the Test Statistic
Consider a full linear regression model with \(n\) observations and \(p_F\) estimated coefficients, including the intercept if the model has one. Suppose the block contains \(q\) coefficients. The full model includes all predictors, while the reduced model is formed by setting the \(q\) block coefficients to zero and retaining the other terms. The models must use the same outcome and the same observations.
Write \(SSE_F\) for the full model's residual sum of squares and \(SSE_R\) for the reduced model's residual sum of squares. Since the reduced model is nested within the full model, fitting the full model cannot increase the residual sum of squares, so \(SSE_R-SSE_F\) is nonnegative. The partial F statistic measures the average reduction in residual sum of squares per restriction relative to the full model's residual variance estimate.
The numerator degrees of freedom are \(q\), the number of restrictions. The denominator degrees of freedom are \(n-p_F\), the full model's residual degrees of freedom. A large value of \(F\) means that the full model's reduction in residual sum of squares is large relative to the residual variation remaining in that model. The p-value is the upper-tail probability \(P(F_{q,n-p_F}\geq F_{\text{observed}})\), calculated under the null hypothesis.
Assumptions for the Test
For the usual exact F reference distribution, the regression errors are assumed to be independent, normally distributed with mean zero and a common variance, conditional on the predictor values. The regression mean is correctly represented by the specified linear model, and the design matrix for the full model has full column rank, so the coefficients are estimable. The full and reduced models must be nested and fitted to the same response observations. These conditions make the comparison and its degrees of freedom well-defined.
The observed data and model checks can help assess whether these assumptions are reasonable, but a residual sum of squares alone cannot verify them. The examples below treat the conditions as plausible for the stated hypothetical settings; in an applied analysis, the model form and residual behavior should be examined as part of the analysis.
Why the Partial F Statistic Has This Form
The numerator compares the full model's fit improvement with the number of restrictions; the denominator estimates the error variance using the full model. The following justification explains why their ratio has an F distribution under the joint null.
With the predictor values fixed and errors independent normal with common variance \(\sigma^2\), least-squares residual sums of squares correspond to squared lengths of projections of the response onto residual subspaces. Under \(H_0\), the component of the response removed by adding the \(q\) block coefficients contains only error variation. Its sum of squares is \(SSE_R-SSE_F\), and, divided by \(\sigma^2\), it has a chi-square distribution with \(q\) degrees of freedom. The full model's residual sum of squares, divided by \(\sigma^2\), has a chi-square distribution with \(n-p_F\) degrees of freedom. These two quantities are independent because they come from orthogonal components of the normal error vector.
Consequently, dividing each chi-square quantity by its degrees of freedom gives two independent estimates on the same variance scale. Their ratio has an \(F_{q,n-p_F}\) distribution, which is exactly the partial F statistic above. This also shows why the denominator uses the full model's residual sum of squares and why the reference degrees of freedom depend on the number of coefficients in that full model.
Worked Applications
Worked Example: Testing Two Predictors Together
Hypothetical data. A researcher models weekly study hours for 35 students using an intercept, a measure of prior preparation, and two additional predictors: weekly work hours and commute time. The question is whether work hours and commute time are jointly associated with study hours after accounting for prior preparation. The reduced model keeps prior preparation but omits the two predictors in the tested block.
Assume the full model has four estimated coefficients: an intercept and three slopes. Thus \(p_F=4\), and the block size is \(q=2\). Suppose the reduced model has \(SSE_R=152\) study-hour-squared units and the full model has \(SSE_F=124\) study-hour-squared units. The full model has \(n-p_F=35-4=31\) residual degrees of freedom. The data are hypothetical, and we assume the linear model and error conditions are reasonable for this illustration.
The joint hypotheses are \(H_0:\beta_{\text{work}}=\beta_{\text{commute}}=0\) and \(H_a:\) at least one of these coefficients is not zero. Substituting into the partial F formula gives
The test has \(2\) and \(31\) degrees of freedom. Its upper-tail p-value is approximately \(0.0426\). At \(\alpha=0.05\), reject \(H_0\): the data provide evidence that work hours and commute time are not both unrelated to study hours in the specified model. This result does not establish that both coefficients are individually nonzero, nor does it establish a causal effect.
Worked Example: A Joint Result Need Not Be an Individual Result
Hypothetical data. Suppose a model of a test score for 25 students includes an intercept, two background predictors, and a block of two study-strategy measurements. The block question is whether the two study-strategy coefficients are jointly zero after accounting for the background predictors. The reduced model retains the intercept and the two background terms.
The full model has five estimated coefficients, so \(p_F=5\), and the block has \(q=2\) restrictions. Suppose \(SSE_R=120\) score-squared units and \(SSE_F=80\) score-squared units. The full-model residual degrees of freedom are \(25-5=20\). Assume the same observations were used for both fits and that the model and error assumptions are reasonable for this hypothetical illustration.
The hypotheses are \(H_0:\beta_1=\beta_2=0\) for the two study-strategy terms, against the alternative that at least one is nonzero. The calculation is
For \(F_{2,20}\), the upper-tail probability at \(5.00\) is
At the 0.05 significance level, the joint test rejects the null: the data provide evidence that at least one study-strategy coefficient differs from zero, conditional on the background predictors. The joint test does not say which coefficient is responsible. Individual tests address separate hypotheses, and their results can differ from the block test because they ask different questions. In particular, do not infer that both coefficients are significant from a significant joint test.
Worked Example: A Block That Does Not Improve Fit Enough
Hypothetical data. A model predicts a laboratory measurement from two established predictors and a block of three additional predictors. There are 30 observations. The reduced model contains an intercept and the two established predictors; the full model adds all three terms in the tested block. Suppose \(SSE_R=276\) measurement-squared units and \(SSE_F=240\) measurement-squared units. The full model has six estimated coefficients, including the intercept, so its residual degrees of freedom are \(30-6=24\). Assume the same observations and a plausible linear model with independent, normally distributed errors of constant variance.
Here \(q=3\). The test statistic is
The reference distribution is \(F_{3,24}\). The 5% upper-tail critical value is approximately \(3.009\), and \(1.20<3.009\). Therefore, at \(\alpha=0.05\), do not reject the joint null hypothesis. These data do not provide sufficient evidence that at least one of the three added coefficients differs from zero, conditional on the two established predictors. This conclusion is not proof that all three population coefficients are exactly zero; it reflects the evidence available in this sample and model.
How to Report and Interpret a Block Test
A clear report identifies the block, the predictors retained in the reduced model, the statistic and its degrees of freedom, the p-value, and the contextual conclusion. For example: “After adjusting for prior preparation, work hours and commute time were jointly associated with weekly study hours, \(F_{2,31}=3.50\), \(p=0.0426\).” The wording describes evidence of an association within the specified regression model; it does not claim causation.
The p-value is the probability, assuming the joint null hypothesis and the model assumptions, of obtaining an F statistic at least as large as the observed statistic. It is not the probability that the null hypothesis is true. A non-significant result is not evidence that the null is certainly true, and a significant result does not identify which member of the block has a nonzero coefficient.
Common Mistakes
- Replacing the block test with separate tests. Testing each coefficient individually asks a different set of questions. The partial F-test evaluates one joint null about the entire specified block.
- Counting the wrong degrees of freedom. The numerator degrees of freedom equal the number of restrictions, \(q\); the denominator degrees of freedom equal \(n-p_F\), using the full model.
- Comparing models fitted to different data. The residual-sum-of-squares comparison is meaningful for nested models using the same outcome and the same observations.
- Claiming every coefficient is nonzero after rejecting the joint null. Rejection means at least one coefficient in the block is not zero; it does not identify which one or imply all are nonzero.
- Treating failure to reject as proof of no effect. The correct conclusion is that the data do not provide sufficient evidence against the joint null at the chosen significance level.
Check Your Understanding
Answer each question using the definitions and test logic in this tutorial.
- A block contains four coefficients. What are the numerator degrees of freedom for its partial F-test?
- A full model has 42 observations and estimates seven coefficients, including the intercept. What are the denominator degrees of freedom?
- What does rejecting the joint null for a three-coefficient block establish, and what does it not establish?
- Why must the reduced and full models use the same observations for the residual-sum-of-squares comparison?
- A block test has \(p=0.12\) at \(\alpha=0.05\). State an appropriate conclusion without claiming the coefficients are zero.