Joint Tests Versus Separate Coefficient Tests

Why Test Coefficients Jointly?

A regression model may contain several coefficients whose scientific role is best assessed together. For example, a researcher may ask whether a group of predictors contributes any linear signal after the other model terms are accounted for. That question is not the same as asking whether each coefficient, one at a time, differs from zero. A joint F test and separate t tests use the same fitted model, but they test different null hypotheses and can give different decisions.

This distinction matters especially when coefficient estimates are correlated. A collection of estimates can provide strong evidence that the coefficient vector is not zero even when no single estimate is sufficiently far from zero relative to its own standard error. Conversely, one coefficient can pass its individual t-test cutoff while a joint F statistic does not pass its own cutoff.

Model and Hypotheses

Assume the normal linear model $Y=X\beta+\varepsilon$, where $X$ is a fixed $n\times p$ design matrix of full column rank, $\beta$ is a $p$-vector, and $\varepsilon\sim N_n(0,\sigma^2 I_n)$ with $\sigma^2>0$. Write $\hat{\beta}$ for the least-squares estimator, $s^2=\operatorname{SSE}/\nu$ for the residual mean square, and $\nu=n-p>0$ for its degrees of freedom.

Definition: For a $q\times p$ matrix $R$ of rank $q$ and a specified $q$-vector $r$, the joint general linear hypothesis is $H_0:R\beta=r$, against $H_A:R\beta\ne r$. If $R$ selects two coefficients and $r=0$, this is the hypothesis that both selected coefficients are zero. A separate test of one selected coefficient tests only its own scalar null hypothesis.

For two selected coefficients, the contrast matrix can be taken to select those coordinates. Thus the joint null is $H_0:\beta_1=\beta_2=0$, whereas the separate tests are $H_{01}:\beta_1=0$ and $H_{02}:\beta_2=0$. Rejecting one of the separate nulls is evidence about that coefficient; rejecting the joint null is evidence that the coefficient vector does not satisfy both restrictions. Neither decision, by itself, identifies the best model or establishes a causal interpretation.

Theorem: Under the stated normal linear model, if $R$ has rank $q$, then the statistic $$ F=\frac{(R\hat{\beta}-r)'\{R(X'X)^{-1}R'\}^{-1}(R\hat{\beta}-r)/q}{s^2} $$ has an exact $F_{q,\nu}$ distribution under $H_0:R\beta=r$. The separate statistic for a single coefficient restriction is a $t$ statistic with $\nu$ degrees of freedom.

The matrix $R(X'X)^{-1}R'$ is positive definite because $R$ has full row rank and $X'X$ is positive definite. It accounts for both the individual variances of the restricted estimates and their covariances. The numerator is a squared, covariance-adjusted distance from the restricted value, divided by the number of restrictions.

Why the Joint F Distribution Is Exact

Under the model, $\hat{\beta}\sim N_p(\beta,\sigma^2(X'X)^{-1})$. Let $A=R(X'X)^{-1}R'$, which is positive definite. Under the null, $R\hat{\beta}-r$ has mean zero and covariance matrix $\sigma^2A$. Therefore $$ Z=\frac{1}{\sigma}A^{-1/2}(R\hat{\beta}-r) $$ has a $q$-variate standard normal distribution. Its squared length is $$ Z'Z=\frac{(R\hat{\beta}-r)'A^{-1}(R\hat{\beta}-r)}{\sigma^2}\sim\chi^2_q. $$

To see the denominator distribution and independence, let $P_X=X(X'X)^{-1}X'$, the orthogonal projection onto the columns of $X$. Then $\operatorname{SSE}=Y'(I-P_X)Y=\varepsilon'(I-P_X)\varepsilon$, since $(I-P_X)X\beta=0$. The matrix $I-P_X$ is an orthogonal projection of rank $\nu$. Under Gaussian spherical errors, its quadratic form divided by $\sigma^2$ has a $\chi^2_\nu$ distribution. The estimator $\hat{\beta}$ depends on $P_XY$, while the residual vector depends on $(I-P_X)Y$; these are independent because they are jointly normal and their covariance is zero. Hence $Z'Z$ and $\operatorname{SSE}/\sigma^2$ are independent. By the definition of the F distribution, their scaled ratio is $F_{q,\nu}$. $\square$

How Two Coefficient Tests Relate to the Joint Test

Let the two selected coefficient estimates have standard errors $s\sqrt{c_{11}}$ and $s\sqrt{c_{22}}$, where $C=R(X'X)^{-1}R'$. Define their individual t statistics under the joint null by $t_1=\hat{\beta}_1/(s\sqrt{c_{11}})$ and $t_2=\hat{\beta}_2/(s\sqrt{c_{22}})$, and let $\rho=c_{12}/\sqrt{c_{11}c_{22}}$ be the correlation of the two coefficient estimates. If $q=2$, direct standardization of the quadratic form gives $$ F=\frac{t_1^2-2\rho t_1t_2+t_2^2}{2(1-\rho^2)}. $$

This formula shows why the joint statistic is not generally the average of the two squared t statistics. When the estimates are strongly correlated, the direction in which their values depart from zero matters. The joint statistic adjusts for that correlation, while each separate t statistic standardizes just one estimate by its own standard error.

Derivation: Write $D=\operatorname{diag}(\sqrt{c_{11}},\sqrt{c_{22}})$. Then $C=D\begin{pmatrix}1&\rho\\\rho&1\end{pmatrix}D$, and $R\hat{\beta}=sD(t_1,t_2)'$. The inverse of the middle matrix is $(1-\rho^2)^{-1}\begin{pmatrix}1&-\rho\\-\rho&1\end{pmatrix}$. Substitution into the joint F statistic cancels the factors $s$ and $D$, leaving the displayed expression. The condition $|\rho|<1$ follows from positive definiteness of $C$.

Worked Examples

Worked Example: Both Separate Tests Reject and the Joint Test Rejects

Suppose two coefficient estimates are uncorrelated, the residual degrees of freedom are $\nu=20$, and their observed t statistics are $t_1=2.2$ and $t_2=2.2$. Test at level $0.05$. For a two-sided individual t test, the critical value is approximately $t_{0.975,20}=2.086$, so each separate test rejects its zero-coefficient null.

Here $\rho=0$, so the joint statistic is $$ F=\frac{2.2^2+2.2^2}{2}=\frac{4.84+4.84}{2}=4.84. $$ The 5% critical value is approximately $F_{2,20;0.95}=3.493$. Since $4.84>3.493$, the joint test also rejects the hypothesis that both coefficients are zero. The two procedures reach compatible decisions here, but they do so by comparing different statistics with different reference distributions.

Worked Example: The Joint Test Rejects but Neither Separate Test Does

Suppose $\nu=20$, the two coefficient estimates have correlation $\rho=-0.9$, and the observed t statistics are $t_1=t_2=1.8$. Each separate two-sided test fails to reject at level $0.05$, because $|1.8|<2.086$.

For the joint test, $$ F=\frac{1.8^2-2(-0.9)(1.8)(1.8)+1.8^2}{2(1-(-0.9)^2)} =\frac{3.24+5.832+3.24}{2(0.19)} =\frac{12.312}{0.38} =32.4. $$ Since $32.4>3.493$, the joint test rejects. This is a feasible covariance pattern: the correlation matrix with diagonal entries $1$ and off-diagonal entries $-0.9$ has eigenvalues $0.1$ and $1.9$, both positive. The joint evidence is strong because the estimates move together in a direction that is unusual under the null given their strong negative correlation; neither estimate alone crosses its t-test cutoff.

Worked Example: One Separate Test Rejects but the Joint Test Does Not

Suppose again that $\nu=20$, the two coefficient estimates are uncorrelated, and the observed t statistics are $t_1=2.3$ and $t_2=0$. The first separate test rejects at level $0.05$, since $2.3>2.086$; the second does not.

The joint statistic is $$ F=\frac{2.3^2+0^2}{2}=\frac{5.29}{2}=2.645. $$ Since $2.645<3.493$, the joint test does not reject at level $0.05$. This is not a contradiction: the single t test and the two-restriction F test use different rejection thresholds and answer different questions. The joint test asks whether the evidence for the two-dimensional vector, considered as a whole, is sufficiently strong under its F calibration.

Interpreting Decisions and Avoiding Pitfalls

A joint rejection supports the conclusion that at least one of the specified restrictions fails; it does not show which coefficient is responsible. Separate tests can help describe individual coefficients, but testing many coefficients one by one at level $\alpha$ does not generally keep the probability of at least one false rejection at $\alpha$. The exact familywise error probability depends on the dependence among the tests. If individual claims are intended to have a familywise error guarantee, an appropriate multiplicity procedure is needed.

A joint nonrejection is not proof that all the coefficients are zero. It means that the data did not produce a sufficiently large joint statistic at the chosen level. Likewise, a non-significant separate test does not establish the corresponding coefficient is exactly zero. The conclusions are about evidence relative to specified null hypotheses, not certainty about parameter values.

Wrong reasoning: “Neither coefficient has a significant t test, so the joint F test cannot be significant.” This fails when the coefficient estimates are correlated; their joint covariance-adjusted departure can be large even when both marginal t statistics are below their cutoffs.
Correction: Calculate the joint F statistic using the full covariance matrix for the restrictions. Do not infer its result by counting separate t-test rejections.
Wrong reasoning: “The joint test is significant, so both coefficients must be nonzero.” The joint null says both restrictions hold; rejecting it means at least one restriction is inconsistent with the data, not necessarily both.
Correction: State the conclusion at the level of the hypothesis tested, and treat separate coefficient claims as separate inferential statements.

Key takeaway: A joint F test evaluates a vector of restrictions using its covariance structure; separate t tests evaluate scalar restrictions one at a time. Their rejection decisions need not agree, and each conclusion must be interpreted in terms of its own null hypothesis.

Check Your Understanding

Use the model assumptions and distinctions developed above to answer these questions.

For two coefficients, write the joint null hypothesis and the two separate null hypotheses.
Why does the joint F statistic depend on the covariance between the two coefficient estimates?
If a joint test rejects, what can you conclude about the restrictions? What can you not conclude about each coefficient individually?
With $\nu=20$, uncorrelated coefficient estimates, and t statistics $1.5$ and $1.5$, calculate the joint F statistic and compare it with the 5% critical value $3.493$.
Why does testing each of many coefficients at level $0.05$ not generally guarantee a 5% probability of any false rejection?