Econometric Theory and Methods

Answers to Starred Exercises

74

Solution to Exercise 10.16

10.16 Show that the articial OPG regression (10.73) possesses all the properties needed for hypothesis testing in the context of a model estimated by maximum likelihood. Specically, show that the regressand is orthogonal to the regressors G( ) when the latter are ; evaluated at the MLE , when the estimated OLS covariance matrix from (10.73) evaluated at multiplied by n, consistently estimates the inverse of the asymptotic information matrix; the OPG regression (10.73) allows one-step estimation: If the OLS para, where is any from (10.73) are evaluated at = meter estimates c ` +c root-n consistent estimator of , then the one-step estimator 1 /2 ` is asymptotically equivalent to , in the sense that n ( 0 ) and 0 ) tend to the same random variable as n . n1/2 (

The rst result is very easy to demonstrate. The inner product of the regres) is G( sand with the matrix of regressors G ) = 0, = g ( G by the rst-order conditions (10.14). Thus we see that the regressand is indeed . orthogonal to the matrix of regressors when the latter is evaluated at The second result is also easily shown. The OLS covariance matrix from the is articial regression (10.73) evaluated at G )1, where s2 = s2 (G n . nk (S10.37)

Clearly, s2 tends to 1 as n . Moreover, as we saw in Section 10.4, G consistently estimates I( ). It follows that n times expression n1 G (S10.37) must consistently estimate I1 ( ). For the nal result, we start with the usual expression for the OLS estimates , multiplying each factor by the approfrom regression (10.73) evaluated at priate powers of n for asymptotic analysis: G )1 n1/2 G . = (n1 G n 1 /2 c (S10.38)

A Taylor expansion of the right-hand side of (S10.38) around 0 yields the result that = (n1 G0 G0 )1 n1/2 G0 n 1 /2 c 0 ) , + (n1 G0 G0 )1 n1 H0 n1/2 (
a

(S10.39)

Copyright c 2003, Russell Davidson and James G. MacKinnon

Econometric Theory and Methods

Answers to Starred Exercises

75

where G0 G(0 ) and H0 H (0 ). There should be an additional term in (S10.39), but it is asymptotically negligible and is therefore ignored. Taking probability limits of both sides, and using the information matrix equality, we see that
n

0 ), = I1 (0 ) plim n1/2 g0 plim n1/2 ( plim n1/2 c

n n

which can be rearranged to yield

n

+c 0 ) = I1 (0 ) plim n1/2 g0 . plim n1/2 (

n

(S10.40)

The right-hand side of (S10.40) is the probability limit of the right-hand side 0 ). of the asymptotic equality (10.38), the left-hand side of which is n1/2 ( `= +c , we conclude that Therefore, since
n

0 ) = plim n1/2 ( ` 0 ), plim n1/2 (

n

which establishes the one-step property for the articial regression (10.73).

Copyright c 2003, Russell Davidson and James G. MacKinnon