Академический Документы
Профессиональный Документы
Культура Документы
22
y2 . Since the projection matrix MX is 1 = y1 For the model (7.97), u equal to I y1 (y1 y1 )1 y1 in this case, it annihilates y1 . Therefore, y2 = MX y2 , 1 = MX y1 MX MX u which is the rst result that was to be proved. The numerator of both statistics is u MX y2 , MX u 1 = n1/2 n1/2 u (S7.17) (S7.16)
= MX u. If we where we have used (S7.16) and the standard result that u substitute I y1 (y1 y1 )1 y1 for MX in (S7.17), the right-hand side becomes u y2 u y1 (y1 y1 )1 y1 y2 . n 1 /2 and (y1 y1 )1 y1 y2 both tend to as n , the probability limit Because of this expression is the random variable
n
plim n1/2 u ( y1 y2 ).
(S7.18)
The denominator of tGNR is 1 MX u 1 )1/2. s(n1 u 1 has The rst factor evidently tends to u under the null hypothesis, since u , asymptotically. The second factor is the square no explanatory power for u root of 2 y2 MX y2 , 1 MX u 1 = n1 n1 u Copyright c 2003, Russell Davidson and James G. MacKinnon
23
where the equality follows from (S7.16). By essentially the same argument as the one that led to (S7.18), this is asymptotically equal to n1 2 y2 y2 n1 4 y1 y1 . (S7.19)
2 Using the fact that yt has variance u /(1 2 ), it is easy to see that the 2 limit in probability of both n1 y1 y1 and n1 y2 y2 is u /(1 2 ). Therefore, expression (S7.19) is asymptotically equal to 2 2 2 u u (1 2 ) 2 4 2 ( ) = = 2 u . 1 2 1 2
(S7.20)
Thus the denominator of tGNR is asymptotically equal to u times the square 2 root of (S7.20), or u . Transposing (S7.18), which is of course a scalar, and dividing it by this denominator yields the desired result that
n 2 plim tGNR = u plim n1/2 ( y1 y2 ) u.
(S7.21)
The analysis for tSR is similar to, but simpler than, the one for tGNR . Both statistics have the same numerator, which is asymptotically equal to (S7.18). The rst factor in the denominator of tSR is s , which evidently tends to u 1 u 1 )1/2. Since n1 under the null hypothesis, and the second factor is (n1 u times the sum of squares of the residuals lagged once must be asymptotically the same as n1 times the sum of squared residuals, this second factor clearly has a probability limit of
n
under the null hypothesis. Therefore, the denominator of tSR is asymptotically equal to 1/ times the denominator of tGNR , and we nd that
n 2 plim tSR = u plim n1/2 ( y1 y2 ) u,
which was the result to be proved. In order to show that tGNR is asymptotically distributed as N(0, 1) under the null hypothesis, we must show that has mean 0 and variance 1 and that it is asymptotically normally distributed. That it has mean 0 follows immediately from the fact that, under the null hypothesis, ut is uncorrelated with yt1 and yt2 . That it is asymptotically normally distributed follows from the fact that we can apply a central limit theorem to both of the quantities
n n
n1/2
t=2
ut yt1
and
n1/2
t=3
ut yt2 ,
(S7.22)
24
2 since the dierence between these quantities, when divided by u , is asymptotically equal to .
It remains to show that the variance of is 1. Since the factors of n1/2 oset the fact that both terms in (S7.22) are summations, it is clear that the 4 variance of is 1/u times
2 2 2 2 Var( ut yt1 ut yt2 ) = E( 2 u2 t yt1 2 ut yt1 yt2 + ut yt2 ).
(S7.23)
Therefore, the variance of itself is just 1, and we conclude that N(0, 1). Because tSR is asymptotically equal to times tGNR , it must be asymptotically distributed as N(0, 2 ). Since | | < 1, the asymptotic variance of tSR is always less than that of tGNR .