Why Test a Block of Coefficients Jointly?
A model may contain several fixed-effect coefficients whose scientific meaning is best assessed together. The question might be whether an entire vector of slopes is zero, or whether a vector equals specified nonzero values. Testing each coefficient separately does not answer exactly the same question: a joint test accounts for the estimated covariance among the coefficients and evaluates the block as a whole.
For a generalized linear mixed model (GLMM), let \(\boldsymbol\beta\) denote the vector of fixed-effect coefficients, and let \(\hat{\boldsymbol\beta}\) be its estimate. Inference about these coefficients uses their estimated covariance matrix, written \(\widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})\). This is the covariance of the estimated fixed effects; it is not the random-effects covariance matrix.
A common special case selects a block of \(q\) coefficients and tests that all equal zero. Then \(C\) is a selector matrix, \(\mathbf c=\mathbf 0\), and the test uses the estimated coefficients and covariance submatrix for that block. The off-diagonal entries of that submatrix matter whenever the estimates are correlated.
Assumptions and Construction
The reference to a chi-square distribution is an approximation, not an exact finite-sample result. It relies on the fitted GLMM providing an approximately multivariate normal estimator of the fixed effects, with a covariance estimate that adequately represents its sampling uncertainty. In practice, this requires a sufficiently well-behaved, identifiable model, reliable estimation, and enough independent information—often meaning enough independent clusters—for the approximation to be credible. The response distribution, link, random-effects structure, and dependence assumptions must also be appropriate for the data.
The covariance matrix used in the calculation must correspond to the same coefficient parameterization as \(\hat{\boldsymbol\beta}\). For a selected block, take the rows and columns for the selected coefficients from the full fixed-effect covariance matrix. Do not use only the diagonal standard errors when the off-diagonal covariances are available.
Specify the coefficient vector or linear restrictions being tested, including the hypothesized values.
Use \(C\widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})C^{\mathsf T}\), which is the covariance matrix of the estimated restrictions.
Compute the quadratic form \(W\), use \(q\) independent restrictions for the reference chi-square distribution, and find the upper-tail p-value.
Why the Quadratic Form Has \(q\) Degrees of Freedom
The justification follows from the approximate sampling distribution of the estimated restrictions. Under \(H_0\), define \(\hat{\boldsymbol\theta}=C\hat{\boldsymbol\beta}-\mathbf c\) and \(S=C\widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})C^{\mathsf T}\). If the normal approximation is adequate, then \(\hat{\boldsymbol\theta}\) is approximately \(N_q(\mathbf 0,S)\). Suppose \(S\) is positive definite and write its Cholesky decomposition as \(S=LL^{\mathsf T}\). The standardized vector \(\mathbf z=L^{-1}\hat{\boldsymbol\theta}\) is then approximately \(N_q(\mathbf 0,I)\). Consequently,
The last step uses the definition of a chi-square random variable as the sum of squares of \(q\) independent standard normal variables. In an actual GLMM, \(S\) is estimated, so the chi-square calibration is generally approximate. If the restrictions are redundant, \(S\) may be singular; the stated inverse and degrees of freedom are then not valid as written.
Worked Example: Testing Two Coefficients Together
Suppose a hypothetical fitted GLMM estimates two fixed-effect coefficients as \(\hat{\boldsymbol\beta}_B=(0.30,0.20)^{\mathsf T}\). The estimated covariance submatrix for these coefficients is $$ \hat V_B= \begin{pmatrix} 0.0625 & -0.0120\\ -0.0120 & 0.0400 \end{pmatrix}. $$ Test \(H_0:\boldsymbol\beta_B=(0,0)^{\mathsf T}\) against the alternative that at least one component differs from zero. Assume the fit and covariance estimate are reliable enough for the approximate Wald calibration.
Here \(q=2\). The determinant is $$ \det(\hat V_B)=(0.0625)(0.0400)-(-0.0120)^2=0.002500-0.000144=0.002356. $$ For a two-by-two covariance matrix, the quadratic form can be written $$ W= \frac{0.30^2(0.0400)-2(0.30)(0.20)(-0.0120)+0.20^2(0.0625)} {0.002356}. $$
The numerator is \(0.003600+0.001440+0.002500=0.007540\). Therefore, $$ W=\frac{0.007540}{0.002356}\approx 3.200. $$ For a \(\chi^2_2\) distribution, the upper-tail probability is \(P(\chi^2_2\ge W)=\exp(-W/2)\). Thus \(p=\exp(-3.200/2)\approx 0.202\). At \(\alpha=0.05\), this test does not reject the joint null. The data do not provide sufficient evidence at that level that at least one of these two coefficients differs from zero; this result does not establish that both coefficients are zero.
Worked Example: Testing a Selected Block From a Larger Model
Consider a hypothetical GLMM with fixed effects \((\beta_0,\beta_1,\beta_2)\). Suppose the estimated coefficient vector is \((0.10,0.60,-0.20)^{\mathsf T}\), and the full estimated covariance matrix is $$ \widehat{\operatorname{Var}}(\hat{\boldsymbol\beta})= \begin{pmatrix} 0.09 & 0.01 & 0.00\\ 0.01 & 0.09 & 0.018\\ 0.00 & 0.018 & 0.04 \end{pmatrix}. $$ The target is the pair \((\beta_1,\beta_2)\), with null values \((0,0)\). Assume the model has converged satisfactorily, these parameters are identifiable, and the large-sample approximation is reasonable.
Selecting rows and columns 2 and 3 gives $$ \hat{\boldsymbol\beta}_B= \begin{pmatrix}0.60\\-0.20\end{pmatrix}, \qquad \hat V_B= \begin{pmatrix}0.09&0.018\\0.018&0.04\end{pmatrix}. $$ The determinant is \(0.09(0.04)-0.018^2=0.0036-0.000324=0.003276\). The numerator of the quadratic form is $$ 0.60^2(0.04)-2(0.60)(-0.20)(0.018)+(-0.20)^2(0.09) =0.0144+0.00432+0.0036=0.02232. $$ Hence \(W=0.02232/0.003276\approx6.813\), with \(q=2\) degrees of freedom. The p-value is \(\exp(-6.813/2)\approx0.0331\). At the 0.05 level, the joint null is rejected: there is evidence that at least one of \(\beta_1,\beta_2\) differs from zero. The joint test alone does not determine which coefficient is responsible.
Worked Example: Testing Specified Nonzero Values
Suppose a hypothetical GLMM estimates a two-coefficient block as \((1.12,0.38)^{\mathsf T}\), with covariance submatrix $$ \hat V_B= \begin{pmatrix}0.01&0\\0&0.04\end{pmatrix}. $$ The scientific null specifies \((\beta_1,\beta_2)=(1.00,0.50)\), rather than zero. Assume the estimates are approximately normal and the covariance matrix is reliable.
The estimated departure from the null is \((0.12,-0.12)^{\mathsf T}\). Since this covariance is diagonal, $$ W=\frac{0.12^2}{0.01}+\frac{(-0.12)^2}{0.04} =1.44+0.36=1.80. $$ There are two independent restrictions, so \(df=2\), and \(p=\exp(-1.80/2)=\exp(-0.90)\approx0.407\). At \(\alpha=0.05\), the test does not reject the specified joint values. The p-value is not the probability that those values are true; it describes how unusual a statistic at least this large would be under the null and the reference approximation.
Interpretation, Alternatives, and Failure Modes
A joint test asks whether the full set of restrictions is compatible with the data, accounting for the joint uncertainty in their estimates. Its p-value is not obtained by combining individual p-values, and a joint rejection does not imply that every coefficient in the block is nonzero. Conversely, separate tests can each be nonsignificant while a joint test is significant, because the directions of uncertainty and the covariance between estimates affect the joint statistic.
| Approach | What it evaluates | Important distinction |
|---|---|---|
| Joint Wald test | A specified set of restrictions using estimated coefficients and their covariance | Convenient from one fitted model; relies on a local normal approximation |
| Separate Wald tests | One restriction at a time | Does not directly test the whole block and does not use its full joint covariance in one statistic |
| Likelihood-ratio test | Difference in fit between models under the null and alternative | Uses refitted models rather than only the alternative-model coefficient estimate and covariance |
Before trusting a joint Wald result, check the model’s convergence and warnings, the estimates and standard errors, and whether the covariance submatrix is finite and invertible. Very large standard errors, nearly redundant coefficients, sparse information, or separation can signal that normal-based inference is unstable. A covariance submatrix that is singular cannot be inverted; do not quietly replace it with diagonal entries or report an ordinary \(q\)-df test. Reconsider whether the restrictions are independent and whether the fitted model can estimate the requested parameters.
Wald tests can also be sensitive in finite samples, especially when estimates are far from the null or the likelihood is asymmetric. A likelihood-ratio test is a common alternative when the required models can be fit appropriately. Neither approach repairs a misspecified GLMM or creates information absent from the data. A non-rejection is not proof of equality, and a rejection identifies evidence against the joint restrictions rather than a causal effect.
Check Your Understanding
For each question, identify the restrictions and use the stated covariance information. Solutions follow each problem.
- A two-coefficient block has estimate \((0.20,-0.10)^{\mathsf T}\) and covariance \(0.04I\). For the null that both coefficients are zero, calculate \(W\), its degrees of freedom, and the approximate p-value. Solution: \(W=0.20^2/0.04+(-0.10)^2/0.04=1.25\), with \(df=2\). Thus \(p=\exp(-1.25/2)\approx0.535\); the joint null is not rejected at 0.05.
- A test specifies three independent restrictions on fixed effects. What is the reference degrees of freedom? Solution: \(q=3\), so the reference distribution is \(\chi^2_3\), regardless of how many other fixed effects are in the model.
- Why is a test of two coefficients not generally calculated by adding their two squared estimate-to-standard-error ratios? Solution: Those ratios omit covariance between estimates. The joint statistic uses the inverse of the full covariance submatrix; the sum-of-squares shortcut applies only when that covariance is diagonal.
- A joint test rejects the null that three coefficients are all zero. What can be concluded? Solution: There is evidence against the joint restrictions, so at least one restriction is inconsistent with the data at the chosen test level. The joint result alone does not show that all three coefficients differ from zero or identify which one does.