MerQuR+Statistics Glossary

Statistics Glossary → Hausman test

Hausman testconcept

A test of a model assumption that compares an estimator that is consistent and efficient under the null hypothesis with one that is consistent in any case but less efficient.

The logic is that if the assumption holds, both estimators converge to the same value and their difference is only chance; if it fails, only the robust estimator stays correct and the difference becomes systematic. The statistic is H = (b₁ − b₀)' (V₁ − V₀)⁻¹ (b₁ − b₀), which follows a chi-square distribution with as many degrees of freedom as coefficients compared.

Its best-known use is choosing between fixed and random effects in panel data: a significant result suggests that the unit effects are correlated with the predictors and that the random-effects estimates may be biased. The same idea is used for endogeneity in instrumental-variables regression (Wu-Hausman) and for the IIA test in multinomial logit. In small samples the variance-difference matrix may not be positive definite, and with heteroscedasticity or clustering robust versions are needed. A non-significant result does not prove that the random-effects model is right.

An everyday example

Ten years of data on 50 firms are used to estimate the effect of R&D spending on profitability. The fixed-effects model finds an effect of 0.8, the random-effects model 1.4. The Hausman test is significant: unmeasured firm traits such as management quality probably drive both R&D and profitability, inflating the random-effects estimate.

In MerQur

As the fixed- versus random-effects comparison in Panel Data Regression; as the endogeneity diagnostic (Wu-Hausman) in Instrumental Variables Regression; as the IIA test (Hausman-McFadden) in Multinomial Logistic Regression; and as 'Extra output: Spatial Hausman test' in the Spatial Error Model.

Analyses where this concept appears

Multinomial Logistic RegressionPanel Data RegressionInstrumental Variables Regression (2SLS)Spatial Error Model (SEM)

Related concepts

Fixed and random effectsIndependence of irrelevant alternatives (IIA)