Econometric Theory and Methods

Answers to Starred Exercises

34

Solution to Exercise 8.24

8.24 Using the same methods as those in Sections 6.5 and 6.6, show that the nonlinear version (8.89) of the IVGNR satises the three conditions, analogous to those set out in Exercise 8.20, which are necessary for the use of the IVGNR in hypothesis testing. What is the nonlinear version of the IV variant of the HRGNR? Show that it, too, satises the three conditions under the assumption of possibly heteroskedastic error terms.

The nonlinear version of the IVGNR is y x( ) = PW X( )b + residuals. (8.89)

IV , the inner product of the regressand with the When it is evaluated at matrix of regressors is , ) PW X (y x IV ) and X IV ). This inner product is equal to 0 because X ( x( where x IV must satisfy the moment conditions (8.84). Thus the IVGNR satises the rst condition necessary for its use in hypothesis testing. IV have no explanatory Since the regressors of the IVGNR evaluated at power for the regressand, the OLS covariance matrix estimator is n PW X )1. 2 (X nk This is n/(n k ) times the covariance matrix estimator (8.88). The factor of n/(n k ) arises because the OLS estimator of 2 divides by n k , and the IV estimator divided by n. Thus the IVGNR also satises the second condition. For the third condition, we need to show that, if the IVGNR is evaluated at , then a root-n consistent estimator
a IV 0 ) = 0 ) + n1/2 b . n1/2 ( n1/2 (

(S8.34)

By the standard formula for the OLS estimator, = (X PW X )1X PW (y x ) b PW X )1X PW (x0 x + u). = (X (S8.35)

is consistent, a rst-order Taylor series approximation shows that Since

a 0 ). x0 ) = n1/2 X0 ( n1/2 (x

Copyright c 2003, Russell Davidson and James G. MacKinnon

Econometric Theory and Methods

Answers to Starred Exercises

35

Substituting this into (S8.35) and adding appropriate powers of n yields the result that = (n1X ) + n 1 /2 u . PW X )1 n1X PW n1/2 X0 (0 n 1 /2 b
a

(S8.36)

is assumed to be root-n consistent, we can replace X here by X0 Because without changing the asymptotic equality. Therefore, (S8.36) implies that
a = 0 ). n 1 /2 b (n1X0 PW X0 )1 n1/2 X0 PW u n1/2 (

(S8.37)

Here we have used the fact that (n1X0 PW X0 )1 (n1X0 PW X0 ) = I. As was noted in (8.86), the rst term on the right-hand side of (S8.37) is asympIV 0 ). Therefore, (S8.37) implies that totically equal to n1/2 (
a = IV 0 ) n1/2 ( 0 ), n1/2 b n1/2 (

which can be rearranged to yield (S8.34). The nonlinear version of the IV variant of the HRGNR is = PU ()PW X() U 1 ( )PW X( )b + residuals, (S8.38)

where U ( ) is an n n diagonal matrix, the t th diagonal element of which is yt xt ( ). IV , the inner product of the regressand with When (S8.38) is evaluated at the matrix of regressors is 1PW X PUP U WX PW X (X PW U U PW X )1X PW U U 1PW X = U (X PW U U PW X )1X PW X . PW X =u = 0, the last line of PW X Since the moment conditions (8.84) imply that u (S8.39) is just a zero vector. Therefore, as required, we have shown that the = IV . regressand in (S8.38) is orthogonal to the regressors when = IV , is The estimated OLS covariance matrix from (8.90), evaluated at n PW U 1P 1PW X X U UPW X nk
1

(S8.39)

The rst factor here is /(n k ), which is the OLS estimate of 2 from regression (S8.38). Because the regressand is orthogonal to the regressors, the SSR is precisely . The second factor can be rewritten as PW U 1 U PW X (X PW U U PW X )1X PW U U 1PW X X PW X (X PW U U PW X )1X PW X = X
1 1

PW P WX (X PW X )1, PW X )1X = (X Copyright c 2003, Russell Davidson and James G. MacKinnon

Econometric Theory and Methods

Answers to Starred Exercises

36

IV ) which is the analog, for nonlinear IV estimation, of the HCCME Var h ( given by (8.65). Finally, we need to show that (S8.38) satises the one-step property. By exactly the same reasoning as we used in Exercise 8.20 to prove the analogous result for the HRGNR in the linear IV case, it can easily be shown that = (X PW X )1X PW (x0 x + u). b This is precisely equation (S8.35). The rest of the argument for the nonlinear IVGNR (8.89) goes through unchanged in the heteroskedasticity-robust case, and we conclude that (S8.34) holds for the heteroskedasticity-robust IVGNR (S8.38) just as it does for the ordinary IVGNR (8.89).

Copyright c 2003, Russell Davidson and James G. MacKinnon