• If a strong instrumental variable is available, the IV
estimator is consistent and approximately normally distributed in large samples.
• If we use a weak instrument, or an instrument that is
invalid in the sense that it is not uncorrelated with the regression error, then IV estimation can be as bad as, or worse than, using the OLS estimator. HAUSMAN TEST The Hausman test (also called the Durbin-Wu- Hausman test) is used to check whether an estimate for an unknown parameter is consistent or not. It is also applied in the context of linear regression to decide whether to choose a fixed effect model or a random effect model. THE HAUSMAN TEST FOR ENDOGENEITY The question we address in this section is how to test for the presence of a correlation between an explanatory variable and the error term, so that we can use the appropriate estimation procedure. If the null hypothesis is true, both the OLS estimator b and the instrumental variables estimator β̂ are consistent. CONT……. Thus, in large samples the difference between them converges to zero. That is, q = ( b − β̂ ) → 0. Naturally, if the null hypothesis is true, use the more efficient estimator, which is the least squares estimator. If the null hypothesis is false, the OLS estimator is not consistent, and the instrumental variables estimator is consistent. Consequently, the difference between them does not converge to zero in large samples. THE LOGIC OF THE HAUSMAN TEST We presented the Hausman test for whether or not an explanatory variable is endogenous using an artificial regression. Let us explore how this test works. The simple regression model is y = β1 + β2x + e (10.32) If x is correlated with the error term e, then x is endogenous and the OLS estimator is biased and inconsistent. An instrumental variable z must be correlated with x but uncorrelated with e in order to be valid. A correlation between z and x implies that there is a linear association between them. This means that we can describe their relationship as a regression x = γ1 + θ1z + v CONT……… A correlation between z and x implies that there is a linear association between them. This means that we can describe their relationship as a regression x = γ1 + θ1z + v TESTING INSTRUMENT VALIDITY A valid instrument z must be contemporaneously uncorrelated with the regression error term, so that cov( zi , ei ) = 0. If this condition fails then the resulting moment condition, like (10.16), is invalid and the IV estimator will not be consistent. In order to compute the IV estimator for an equation with B possibly endogenous variables, we must have at least B instruments. The validity of this minimum number of required instruments cannot be tested. CONT…. If all the instruments are valid, then we would expect all the IV estimates to be similar. Rather than do this, there is a test of the validity of the surplus moment conditions that is easier to compute. CONT….