Joint Wald Tests for Quasi-Estimators

Why Test Several Restrictions Together?

A single coefficient test asks whether one component of a regression parameter equals a specified value. Many scientific questions instead concern several coefficients at once: whether a group of predictors contributes jointly, whether two slopes are both zero, or whether a specified linear combination of coefficients equals a target. These are joint hypotheses. Testing each coefficient separately does not answer the same question, because the estimates can be correlated and the joint null concerns their combined uncertainty.

For a quasi-estimator, the covariance used in the test should reflect the estimating procedure and the variability of the observations. In particular, when the variance model may be misspecified, the sandwich covariance—not a model-based covariance that assumes the variance specification is correct—is the relevant uncertainty estimate. The Wald construction converts the estimated departure from the null into a squared, covariance-standardized distance.

Linear Restrictions and the Wald Statistic

The rank condition means the restrictions are linearly independent. It also ensures that the hypothesis has \(q\), rather than fewer, distinct components. Write \(\hat\beta\) for the quasi-estimator and let \(\hat C\) be its estimated sandwich covariance matrix, including the sample-size scaling so that \(\hat C\) estimates the covariance of \(\hat\beta\) itself. The estimated restriction vector is \(R\hat\beta-r\), and its estimated covariance is \(R\hat C R^\mathsf{T}\).

The covariance scaling is important. If software instead reports \(\hat V\) as an estimate of the asymptotic covariance \(V\) in the limit \(\sqrt n(\hat\beta-\beta)\xrightarrow{d}N(0,V)\), then the covariance estimate for \(\hat\beta\) is \(\hat C=\hat V/n\). The same statistic can then be written

These expressions are equivalent; the factor \(n\) must not be omitted when using an estimate of \(V\) rather than an estimate of \(\operatorname{Cov}(\hat\beta)\).

Why the Reference Distribution Is Chi-Square

The result rests on asymptotic normality of the quasi-estimator and consistency of the sandwich covariance estimate. State these assumptions explicitly: under the null value \(\beta_0\), suppose

where \(V\) is finite and positive definite. Assume \(R\) is fixed, has rank \(q\), and \(RVR^\mathsf{T}\) is positive definite. These assumptions describe the setting in which the estimator has a nonsingular limiting covariance in the tested directions. The asymptotic normality is the quasi-estimator result; the argument below derives the test distribution from it.

Proof. Define \(Z_n=\sqrt n(R\hat\beta-r)\). Since \(R\beta_0=r\), linear transformation of the assumed asymptotic normal limit gives

By assumption, \(C\) is positive definite. Also, \(nR\hat C R^\mathsf{T}\xrightarrow{p}C\), by the covariance consistency and the fact that \(R\) is fixed. The Wald statistic can be expressed as

The inverse is a continuous function at positive-definite \(C\), so \(\left(nR\hat C R^\mathsf{T}\right)^{-1}\xrightarrow{p}C^{-1}\). Slutsky's theorem therefore gives convergence in distribution of \(W\) to \(Z^\mathsf{T}C^{-1}Z\), where \(Z\sim N(0,C)\). Let \(U=C^{-1/2}Z\). Then \(U\sim N(0,I_q)\), and

Thus \(W\xrightarrow{d}\chi^2_q\), as claimed. \(\square\)

This is an asymptotic result. It does not assert that the statistic has an exact chi-square distribution for a fixed sample size. Its degrees of freedom count the independent restrictions, not the number of coefficients in the entire model.

Worked Calculations

Equivalent Ways to Write the Same Restrictions

The numerical form of the restriction equations is not unique. If \(A\) is any invertible \(q\times q\) matrix, then \(R\beta=r\) and \(AR\beta=Ar\) describe the same set of coefficient vectors. The Wald statistic is unchanged by this rewriting.

Proof. Set \(d=R\hat\beta-r\) and \(M=R\hat C R^\mathsf{T}\). For the transformed restrictions, the departure is \(Ad\) and the covariance is \(AMA^\mathsf{T}\). Since \(A\) and \(M\) are invertible,

Consequently, the transformed quadratic form is

This equals the original statistic. \(\square\)

Interpretation and Finite-Sample Cautions

A Wald statistic measures how far the estimated coefficient vector lies from the null restrictions relative to estimated uncertainty in the tested directions. Large values indicate a departure that is large compared with its estimated standard error structure. The sandwich covariance matters because it carries the robust uncertainty information into this comparison. A covariance estimate that ignores dependence among the restricted estimates can misstate the distance.

The chi-square calibration is justified by an asymptotic normal approximation, not by an exact finite-sample identity. In modest samples, especially with many restrictions, the approximation can be poor. Estimated sandwich covariances can also be unstable when there are few independent observational units, strong leverage, or near-collinearity among the tested directions. If \(R\hat C R^\mathsf{T}\) is singular or nearly singular, the usual inverse-based statistic is undefined or numerically unreliable; this may indicate redundant restrictions, insufficient information, or an unstable covariance estimate.

A common mistake is to test each coefficient separately and then conclude that the joint null is rejected if one individual test rejects—or that it is accepted if none does. Neither conclusion is generally justified: individual tests do not account for the joint covariance structure, and their results do not equal the specified \(q\)-restriction Wald test. The correction is to form \(R\hat\beta-r\), calculate its full covariance \(R\hat C R^\mathsf{T}\), and evaluate one joint quadratic form. Also, the chi-square \(p\)-value should be described as an asymptotic approximation, not an exact finite-sample probability.