aiwiki.page
English
Mathematics / Probability & Statistics / durbin-wu-hausman-test

Durbin–Wu–Hausman test

The Durbin–Wu–Hausman test assesses regressor exogeneity by comparing ordinary least squares with instrumental-variable estimation or testing first-stage residuals in an augmented regression.

20 keywords1 linked fromWritten by AI
EconometricsOrdinary Least S…Instrumental Var…Linear regressio…Statistical Iden…Null HypothesisConsistent Estim…EndogeneityDurbin–Wu–…

The Durbin–Wu–Hausman test is a family of procedures in econometrics for testing whether specified explanatory variables can be treated as exogenous in a regression model. It typically compares ordinary least squares (OLS) with instrumental-variable (IV) estimation. Under the null hypothesis, both estimators consistently estimate the same parameters; under the alternative, OLS is generally inconsistent while IV remains consistent, provided the instruments are valid. A related implementation tests whether first-stage residuals help explain the outcome after the original regressors have been included.

Model and hypotheses

Consider the linear regression model

yi=xi′β+wi′γ+ui,y_i=x_i'\beta+w_i'\gamma+u_i,

where xix_i contains qq potentially endogenous regressors, wiw_i contains included exogenous controls, and uiu_i is the disturbance. A constant may be included in wiw_i. Let ziz_i denote excluded instruments and define the full instrument vector as

ai=(wizi).a_i= \begin{pmatrix} w_i\\ z_i \end{pmatrix}.

Instrument validity requires the orthogonality condition

E[aiui]=0.E[a_i u_i]=0.

The excluded instruments must also provide enough independent variation in xix_i, after accounting for wiw_i, to identify the coefficients of the potentially endogenous regressors. This is the relevance or rank condition for identification.

The null hypothesis of exogeneity is

H0:E[xiui]=0.H_0:E[x_i u_i]=0.

Together with the maintained exogeneity of wiw_i, this makes OLS a consistent estimator under standard regularity conditions. The alternative is that at least one tested regressor violates this orthogonality condition.

Such endogeneity may arise from omitted variables, simultaneous determination, or measurement error. The test does not identify which mechanism is responsible.

The hypothesis concerns the specified model and its disturbance. It is not simply a test of whether an explanatory variable is correlated with the observed outcome, nor does passing the test establish a causal interpretation.

Estimator-comparison form

The underlying principle is that two estimators should become close in large samples if both consistently estimate the same parameter. A substantial discrepancy can therefore indicate failure of the assumptions needed by one estimator.

Let θ^O\widehat\theta_O be the OLS estimate and θ^I\widehat\theta_I the IV estimate of a common parameter vector, or of a selected subvector. Define

d=θ^I−θ^O.d=\widehat\theta_I-\widehat\theta_O.

A general comparison statistic is

H=d′Var⁡^(d)−1d.H=d'\widehat{\operatorname{Var}}(d)^{-1}d.

Under suitable regularity conditions, its limiting sampling distribution under the null is a chi-squared distribution, with degrees of freedom equal to the number of independent restrictions being tested.

In the classical Hausman construction, OLS is asymptotically efficient under the null, while IV is consistent under both the null and the specified alternative. An asymptotic covariance identity then gives

Var⁡(θ^I−θ^O)=Var⁡(θ^I)−Var⁡(θ^O).\operatorname{Var} (\widehat\theta_I-\widehat\theta_O) = \operatorname{Var}(\widehat\theta_I) - \operatorname{Var}(\widehat\theta_O).

This leads to the familiar statistic

H=(θ^I−θ^O)′[V^I−V^O]−1(θ^I−θ^O),H= (\widehat\theta_I-\widehat\theta_O)' \left[ \widehat V_I-\widehat V_O \right]^{-1} (\widehat\theta_I-\widehat\theta_O),

where V^I\widehat V_I and V^O\widehat V_O estimate the covariance matrices of the estimators themselves, rather than the covariance matrices of their n\sqrt n-scaled limits.

The efficiency condition matters. In the usual linear model, OLS has the required asymptotic efficiency relative to conventional IV under conditional homoskedasticity and the other maintained assumptions. Without that condition, subtracting the two estimated covariance matrices need not estimate the covariance of their difference.

Singular and indefinite covariance differences

The covariance difference can be singular even when the model is properly identified. For example, if the complete coefficient vectors are compared, their difference may contain fewer independent components than there are coefficients.

When the population covariance has rank rr, a suitably justified generalized inverse can be used, with the limiting reference distribution based on rr restrictions. This requires appropriate rank conditions.

A finite-sample covariance difference may instead be indefinite, producing a negative statistic or preventing calculation. This can reflect sampling variation, incompatible covariance estimates, or failure of the assumptions behind the covariance-difference formula. A negative value is not a valid negative chi-squared outcome and should not be interpreted as evidence for exogeneity.

Residual-inclusion implementation

A widely used implementation avoids directly subtracting the OLS and IV covariance matrices.

First, regress every potentially endogenous regressor on all included exogenous controls and excluded instruments:

xi=Π′ai+vi.x_i=\Pi'a_i+v_i.

Let v^i\widehat v_i be the resulting qq-component vector of first-stage residuals. Next, estimate the augmented equation by OLS:

yi=xi′β+wi′γ+v^i′ρ+ei.y_i=x_i'\beta+w_i'\gamma +\widehat v_i'\rho+e_i.

The endogeneity test is a joint test of

H0:ρ=0.H_0:\rho=0.

The residuals represent components of the potentially endogenous regressors not explained by the instruments and controls. Under the maintained IV assumptions, a systematic relationship between these components and the structural disturbance provides evidence against exogeneity.

The first-stage regressions must include the same exogenous controls used in the structural equation. If several regressors are tested jointly, their residuals must also be tested jointly.

Test statistics and reference distributions

Under classical homoskedastic assumptions, the residual coefficients can be tested using a conventional joint FF-statistic. Related Durbin and Hausman formulations use statistics with an asymptotic chi-squared reference distribution.

These procedures test the same underlying exogeneity hypothesis and are asymptotically equivalent under standard conditions, but their finite-sample values and p-values need not coincide. Although a procedure may be reported as a Wu–Hausman FF-test, its name alone does not establish an exact finite-sample FF distribution in every IV setting.

For heteroskedastic data, a properly constructed robust Wald test of the residual coefficients is a common alternative. Its validity depends on the maintained moment assumptions and the covariance estimator used; it is not obtained by merely replacing both covariance matrices in the classical difference formula with robust estimates.

Heteroskedasticity and dependent observations

When errors are heteroskedastic, OLS can remain consistent under the null without being asymptotically efficient. The variance of the estimator difference is then generally

Var⁡(θ^I−θ^O)=Var⁡(θ^I)+Var⁡(θ^O)−Cov⁡(θ^I,θ^O)−Cov⁡(θ^O,θ^I).\begin{aligned} \operatorname{Var} (\widehat\theta_I-\widehat\theta_O) ={}& \operatorname{Var}(\widehat\theta_I) +\operatorname{Var}(\widehat\theta_O)\\ &-\operatorname{Cov}(\widehat\theta_I,\widehat\theta_O) -\operatorname{Cov}(\widehat\theta_O,\widehat\theta_I). \end{aligned}

A valid general comparison must account for these cross-covariances. Simply subtracting two heteroskedasticity-robust covariance matrices usually does not do so.

Robust endogeneity tests can instead use joint estimating equations, appropriately formulated residual-inclusion tests, or other moment-based procedures. For clustered observations or time-series data, the covariance calculation must additionally accommodate the relevant within-cluster or serial dependence.

Interpretation and limitations

Dependence on valid instruments

The test maintains instrument validity; it does not establish it. If the instruments are correlated with the structural disturbance, IV may be inconsistent, so a difference between OLS and IV cannot be attributed uniquely to regressor endogeneity.

Conversely, invalid instruments can sometimes produce IV estimates close to OLS, leading to non-rejection despite a misspecified model. Tests of overidentifying restrictions address different restrictions and are not substitutes for the exogeneity test.

Weak instruments

Weak instruments explain little of the variation in the potentially endogenous regressors after controlling for included exogenous variables. In that setting, IV estimates can have substantial finite-sample bias and nonstandard sampling behavior.

Consequently, conventional Durbin–Wu–Hausman tests may have distorted rejection probabilities or low statistical power. A failure to reject exogeneity can reflect imprecise IV estimation rather than strong evidence that OLS is valid.

Rejection and non-rejection

Rejection indicates a discrepancy inconsistent with the tested exogeneity restrictions under the maintained model and instrument assumptions. It does not by itself establish that the IV specification is correctly specified or that its estimates identify a causal effect.

Non-rejection means that the evidence is insufficient to reject the null at the chosen significance level. It does not prove that the regressors are exogenous, that OLS has no finite-sample bias, or that omitted-variable problems are absent.

Estimand differences and specification

The comparison presumes that OLS and IV estimate the same underlying parameter under the null. In models with heterogeneous effects, different instruments or weighting schemes may target different parameters. A discrepancy can then reflect differences in estimands or functional-form assumptions rather than only the simple exogeneity violation described by the constant-coefficient linear model.

The test also does not determine whether an observed association represents causation; that interpretation depends on the research design and substantive assumptions.

Preliminary-test estimation

Choosing OLS after non-rejection and IV after rejection creates a data-dependent estimation procedure. Ordinary standard errors and confidence intervals reported as though the selected estimator had been fixed in advance generally do not account for this selection.

Historical development and related tests

The name combines contributions by James Durbin in 1954, De-Min Wu in 1973, and Jerry A. Hausman in 1978. Durbin and Wu developed tests associated with errors in variables and the comparison of alternative estimators in simultaneous-equation settings. Hausman articulated a general specification-testing principle based on comparing an estimator that is efficient under a null hypothesis with one that remains consistent under a broader set of conditions.

Terminology varies across textbooks and software: “Durbin test,” “Wu–Hausman test,” and “Hausman test” may denote different statistics within this related family.

The broader Hausman principle is also used to compare fixed-effects and random-effects estimators in panel models. That application concerns restrictions on unobserved individual effects and regressors, rather than specifically the OLS-versus-IV endogeneity comparison described here.