Professional Documents
Culture Documents
187
u 2 2t
2 2
2 1 + (MZ PX y )2 t 2+ 2 1 a
(15.91)
1 ) may be As we saw in the previous exercise, the terms that involve 1t ( omitted when calculating Cox tests for linear regression models. Therefore, from (15.54), a typical element of the test regressor must be
2t (2 )
E 1
2t (2 )
(S15.44)
The t th contribution to
2 (2 )
is
2 2
2t (2 )
1 1 1 2 = log 2 log 2
u 2 2t . 2 2 2
(S15.45)
We saw in the solution to the previous exercise that, under a DGP in H1 2 2 is equal, with an error of , the expectation of log 2 with parameters and 1 2 2 1 2 order n only, to log(1 + a ), where a is dened as plim n1 MZ X 2 . An entirely similar argument, based on a Taylor expansion, shows that the 2 2 is 1/a with error of order n1 only. expectation of 1/ 2 The residual u 2t is the t th element of the vector MZ y = MZ u + MZ X . 2 2 The expectation of u 2 2t is therefore 1 (1 ht ) + (MZ X )t , where ht is the th t diagonal element of PZ . A natural estimate of this expectation is
2 1 + (MZ PX y )2 t,
2t (2 ))
is (S15.46)
1 1 1 2 2 log 2 log( 1 + a )
2 1 + (MZ PX y )2 t . 2+ 2 2 1 a
Subtracting expression (S15.46) from expression (S15.45) and multiplying the result by 2 (which we do for simplicity because the scale of the test regressor has no eect on the value of the test statistic) yields
2 log 2 2 u 2 1 + (MZ PX y )2 t 2t 2 2 + log( + ) + 1 a 2 2 2 2 1 + a 2 2 1 + a 2 2
= log
2 u 2 1 + (MZ PX y )2 t 2t + , 2 2+ 2 2 1 a
which is expression (15.91). Copyright c 2003, Russell Davidson and James G. MacKinnon