When Hausman Tests Can Mislead

Why a Hausman Result Needs a Diagnostic Check

A Hausman test compares fixed-effects (FE) and random-effects (RE) estimates of the same slopes. Under its usual null hypothesis, both estimators are consistent and RE is efficient; under the alternative, RE may be inconsistent because the unit-specific effect is correlated with the regressors. A small p-value can therefore be evidence against the conventional RE exogeneity assumption. But the test is only as trustworthy as the covariance calculation and error assumptions that produce it.

Three warnings deserve particular attention. The estimated covariance difference may not be positive semidefinite, so the usual quadratic statistic can be negative or otherwise unusable. The error structure may make the conventional covariance estimates invalid, even when robust standard errors appear in the regression output. Or the test may have too little power to detect meaningful correlation. A correlated-random-effects (CRE) test can offer a useful alternative in some settings, but it too tests a specified model rather than every possible form of endogeneity.

The setup matters. The compared estimates must target the same slope parameters, use the same observations and regressor set, and have compatible specifications. The usual covariance-difference formula also relies on the null model’s efficiency relationship. It is not a general recipe for subtracting any two estimated covariance matrices.

Why the Covariance Difference Should Be Nonnegative

Under the null, both estimators are consistent for the same parameter, and RE is efficient among the estimators being compared. The efficiency result gives a useful covariance identity. In population terms, let \(V_{FE}\) and \(V_{RE}\) be the covariance matrices, and suppose that under the null \(\operatorname{Cov}(\hat{\beta}_{FE}-\hat{\beta}_{RE},\hat{\beta}_{RE})=0\). Then

The matrix on the left is a covariance matrix, so it is positive semidefinite: for every vector \(a\), \(a'\operatorname{Var}(\hat{\beta}_{FE}-\hat{\beta}_{RE})a\geq 0\). Consequently, the population covariance difference should also be positive semidefinite under these conditions. In finite samples, estimated matrices may deviate from this ideal because of sampling variation, numerical issues, or a failure of the assumptions. A non-positive difference is a diagnostic warning, not a reason to force the matrix into a form that produces a preferred result.

Misspecified Error Structures and Robust Covariances

The conventional test can also mislead when its covariance estimates rely on an error structure that is wrong. For example, heteroskedasticity across units or serial correlation within a unit can make conventional standard errors unreliable. Using robust standard errors for each estimator is important for inference about those estimators, but it does not automatically make the conventional Hausman subtraction valid.

In general, the variance of the difference between two estimates is

The conventional difference \(V_{FE}-V_{RE}\) follows only when the required null covariance relationship holds. Separately estimated robust covariance matrices do not generally establish that relationship. A robust comparison must account for the covariance between the two estimators—for example, through a suitable joint sandwich calculation—or use an appropriate auxiliary-regression test. A robust standard error for each coefficient, on its own, does not supply the missing cross-covariance.

A robust method also cannot repair a misspecified conditional mean or a failure of the substantive exogeneity assumptions. Robust covariance estimation addresses uncertainty under specified forms of dependence or heteroskedasticity; it does not turn an inconsistent estimator into a consistent one.

Weak Power: A Non-Rejection Is Not Reassurance

Even when the conventional statistic is well-defined and its assumptions are credible, it may have weak power. In particular, FE estimates can be imprecise when much of a regressor’s variation is between units rather than within units. Then the difference between FE and RE may be estimated imprecisely, and the test may fail to detect a meaningful violation of RE exogeneity.

When a Correlated-Random-Effects Test Helps

A correlated-random-effects specification can make the relevant exogeneity restriction explicit. For time-varying regressors \(x_{it}\), a common Mundlak formulation includes each unit’s time average \(\bar{x}_i\) alongside the time-varying regressors:

Here \(a_i\) is a unit-specific component and \(u_{it}\) is the remaining error. In this specified linear formulation, testing \(H_0:\pi=0\) asks whether the included unit means jointly contribute, conditional on the included regressors. If the means capture the relevant correlation between the regressors and the unit-specific component, this test is informative about the conventional RE restriction. The test can be implemented as a joint Wald test using a covariance estimator appropriate to the error structure.

This is not an unrestricted test for every possible dependence between \(x_{it}\) and \(a_i\). Its interpretation depends on the CRE specification, the included means, and the assumptions about the remaining error. For example, nonlinear dependence or omitted relevant features of the regressor history may not be captured by a linear set of means. Time-invariant regressors also require care: their unit means are the regressors themselves, so they cannot be separately identified from the corresponding time-invariant component in this formulation.

Common Pitfalls and a Practical Diagnosis

The most common mistake is to treat a Hausman p-value as an automatic choice between estimators. A reliable decision starts by checking whether the test was validly constructed and whether its result is informative for the question at hand.

• Wrong: “The covariance difference is indefinite, but software returned a number, so its chi-square p-value is usable.” Correction: An indefinite difference violates the usual covariance interpretation; investigate the specification, covariance calculation, and numerical stability before making an inference.
• Wrong: “I used robust standard errors for FE and RE, so their difference is a robust Hausman test.” Correction: Robust marginal covariance estimates do not generally provide the covariance of the difference. Use a test procedure that accounts for the joint covariance or a suitable CRE test.
• Wrong: “A large p-value proves the RE assumption.” Correction: Non-rejection means the test did not find sufficient evidence against its null. Weak power, especially with imprecise within-unit information, can also produce non-rejection.
• Wrong: “A CRE test is an omnibus test for any dependence between regressors and unit effects.” Correction: It tests restrictions implied by the specified included means and model; dependence not captured by that specification may remain undetected.

When a conventional Hausman result is questionable, report the reason rather than presenting a forced decision. Check that the slope vectors and samples match, inspect the covariance difference, and verify that the error assumptions support the covariance calculation. If the concern is heteroskedasticity or within-unit dependence, use an inference procedure designed for that structure. If a CRE formulation is substantively plausible, a joint test of the unit-mean coefficients can provide a useful, explicit diagnostic. In all cases, distinguish evidence against a restriction from proof that an alternative is correct.