Technical Note: Speci…cation Sensitivities in Right-Tailed Unit Root Testing for Financial Bubbles Shu-Ping Shi The Australian National University Peter C. B. Phillips Yale University, University of Auckland, University of Southampton & Singapore Management University Jun Yu Singapore Management University February 8, 2011
Before proving Proposition 4.1, we …rst introduce a few lemmas. 1 j=0
(L) "t =
j "t j ;
where
1 j j=0
< 1 and f"t g is an i:i:d P[T r] sequence with mean zero,P variance and …nite fourth moment. De…ne MT (r) = 1=T s=1 us t with r 2 0 ; 1] and t = P s=1 us . We have: P1 P1 P[r (1) Pts=1 us = (1) ts=1 "s + t 0 , where t = j=0 j " j and j=0 j "t j , 0 = 1 , which is absolutely summable. = j i=1 j+i 1 P[T r] 2 p (2) T t=1 "t ! 2 r: P[T r] L (3) T 1=2 t=1 "t ! W (r) : h i P[T r] P L (4) T 1 t=1 ts=11 "s "t ! 12 2 W (r)2 r : Rr P[T r] L (5) T 3=2 t=1 "t t ! rW (r) 0 W (s) ds : P p [T r] (6) T 1 t=1 t 1 0 "t ! 0: Lemma 0.1 Let ut =
2
p
(7) T 1=2 [T r] 0 ! 0: p L (8) T MT (r) ! (1) W (r) : h i P L [T r] (9) T 1 t=1 t 1 "t ! 12 (1) 2 W (r)2 r : Rr P[T r] L (10) T 3=2 t=1 t 1 ! (1) 0 W (s) ds: Rr P[T r] p (11) T 5=2 t=1 t 1 t ! (1) 0 W (s) sds: 1
j
(12) T (13) T
R L 2 2 (1)2 r t=1 t 1 ! 0 P[T r] p 3=2 u ! 0; 8j t=1 t 1 t j 2
P[T r]
W (s)2 ds: 0:
Pt
1 Lemma 0.2 De…ne yt = ~ t + s=1 us and ut = (L) "t = 1 j <1 j=0 j "t j ; where j=0 j 2 and f"t g is an i:i:d sequence with mean zero, variance and …nite fourth moment. Then, we have Z r [T r] X L 3=2 (a) T yt 1 "t ! ~ rW (r) W (s) ds : 0
t=1
(b) T
2
[T r] X
~ 2 r : 2
p
yt
1
!
yt2
1
!
yt
1 ut j
t=1
[T r]
(c) T
3
X t=1
(d) T
2
[T r] X
~2 3 r : 3
p
t=1
p
! 0; 8j
0:
The proof of Lemma 0.1 and Lemma 0.2 can be found in Phillips (1987), Phillips and Perron (1988), Phillips and Solo (1991) and Hamilton (1994). P Lemma 0.3 De…ne yt = ~ T t + ts=1 us ; ~ T = (1) T and ut = (L) "t = 1 j=0 j "t j ; 1 where j=0 j j < 1 and f"t g is an i:i:d sequence with mean zero, variance 2 and …nite fourth moment. Then, if > 1=2, we have (a1) T
1
[T r] X
yt
1 "t
t=1
(b1) T
3=2
[T r] X
2
[T r] X
yt
yt2
1
L
1
t=1
(d1) T
3=2
[T r] X
yt
1 2
L
t=1
(c1) T
L
! !
!
= 1=2, we have (a2) T
1
[T r] X t=1
yt
1 "t
L
!
(1)
Z
(1)
i r :
r
W (s) ds:
0
2
(1)2
Z
r
W (s)2 ds:
0
1 ut j
t=1
and if
h (1) W (r)2
2
rW (r)
p
! 0; ; j = 0; 1; Z
0
2
:
r
W (s) ds +
1 h W (r)2 2
r
i
:
[T r] X
3=2
(b2) T
L
yt
1
t=1
!
[T r]
2
(c2) T
X
L
yt2
1
t=1
[T r] X
3=2
(d2) T
(1)2
!
yt
1 r+ 2
Z
1 3 r + 3
2
(1)
1 ut j
t=1
r
W (s) ds :
0
Z
r
Z
W (s)2 ds + 2
0
r
W (s) sds :
0
p
! 0; ; j = 0; 1;
:
Proof: (a) From (1) of Lemma 0.1, T
1
[T r] X
yt
1 "t
1
=T
t=1
[T r] X
(1)
t T
t=1
=
1
+
t 1 X
(1)
X
1
t"t
1
(1) T
t=1
+ L
!
1
(1) T
[T r] t 1 X X
:
i r Rr 0
(b) From Lemma 0.1(10), T
3=2
[T r] X
yt =
1
"s "t + T
h (1) W (r)2 n (1) rW (r)
1 2 2
(1) T
(1) (1)
L
!
[T r] X
[T r] X
(1)2 T
yt2 =
t=1
L
!
(
2(1+ )
[T r] X
T
[T r] X t=1
yt
1 ut j
t 1
t2 + T
2
= T
[T r] X
[T r] X
(~ T t +
3
2
W (r)
[T r] X
r
io
if
> 1=2
if
= 1=2
t
(s) R r ds 0 W (s) ds 2 t
if if
+ 2 (1) T
2
> 1=2 = 1=2 [T r] X
t
t
t=1
+2
1=2;
t=1
1 2
h
t=1
t=1 t=1 R 2 (1)2 r W (s)2 ds h 0 Rr 2 1 3 (1) 3 r + 2 0 W (s)2 ds 3=2
"t
0
3=2
t+T
t=1 Rr 0 W 1 r 2 +
(d)From (5) and (13) of Lemma 0.1, when 3=2
"t
W (s) ds +
(c) From (11) and (12) of Lemma 0.1, 2
[T r] X
"t
t=1
[T r] X
3=2
t=1
T
0
t=1
t=1 s=1
8 <
t 1
s=1
[T r]
(1) T
"s +
!
t ) ut j
Rr 0
i if W (s) sds if
> 1=2 = 1=2
=
[T r] X
3=2
(1) T
tut
3=2
+T
j
t=1
=
(1)
3=2
(L) T
[T r] X
t ut j
t=1
[T r]
X
t"t
j
+T
t=1
3=2
[T r] X
t ut j
t=1
p
! 0:
We now derive the asymptotic distributions of SADF. Proof of Proposition 4.1. First, consider Case 1. The regression model is yt =
p 1 X
k r
yt
k
+
r yt 1
+ "t :
k=1
Let r be the window size (fractional) with r 2 [r0 ; 1], r0 the smallest fraction considered and t = [T r]. Assume the initial value y0 = 0. Under the null hypothesis that r = 0; we have P 1 1 2 2 p 1 p 1 yt = ts=1 us ; where ut = r (L) "t with r (L) = 1 : rL rL r L The deviation of the OLS estimate ^r from the true value r is given by 2 3 12 3 [T r] [T r] X X ^r Xt Xt0 5 4 X t "t 5 (1) r =4 t=1
t=1
where Xt = [ut 1 ut 2 : : : ut p+1 yt 1 ]0 ; r = [ 1r 2r : : : pr 1 r ]0 . P[T r] From (13) of Lemma 0.1, we know that the probability limit of t=1 Xt Xt0 is block diagonal. P[T r] P[T r] Therefore, we only need to obtain the last row of t=1 Xt Xt0 and t=1 Xt "t to calculate the ADF statistics. Then, the deviation of the OLS estimate ^ r from the true value is characterized by T 1 yt 1 " t T ^r = ; (2) T 2 yt2 1 where denotes summation over t = 1; 2; ; [T r] : Since Equation (2) is a continuous function of T 1 yt 1 "t and T 2 yt2 1 , based on (9) and (12) of Lemma 0.1, we have L
T ^r !
W (r)2 r : Rr 2 r (1) 0 W (s)2 ds
To calculate the t-statistic of ^ r ; we need to …nd the standard error of ^ r : Since the variance ^ of r is 82 3 12 39 2 3 1 [T r] [T r] [T r] = < X X X V ar ^r = V ar 4 Xt Xt0 5 4 Xt "t 5 = 2r 4 Xt Xt0 5 : : ; t=1
t=1
4
t=1
Therefore, variance of ^ r is
2 r
multiplied by the last element of 2 r ; 2 yt 1
V ar ^ r =
hP
[T r] 0 t=1 Xt Xt
i
1
; which is
so the variances of T ^ r can be calculated such that V ar T ^ r =
2 r
T
2
1 : W (s)2 ds
L
yt2 1
2Rr r (1) 0
!
Hence, the t-statistic of ^ r is ^ ADFr =
r
=
se ^ r
T ^r se T ^
L
!
r
2
hR
W (r)2 r 0
r
W (s)2 ds
i1=2 :
By CML, the asymptotic distribution of the SADF statistic is 8 9 > > < = W (r)2 r L SADF (r0 ) ! sup : h i 1=2 > r2[r0 ;1] > : 2 R r W (s)2 ds ; 0 Second, consider Case 3 where the regression model is yt =
r
+
r yt 1
+
p 1 X
k r
yt
k
+ "t :
k=1
P Under the null hypothesis that r = T and r = 0; we have yt = ~ T t + ts=1 us ; where 1 2 2 p 1 p 1 1 ~ T = r (1) T and ut = r (L) "t with r (L) = 1 . rL rL r L The deviation of the OLS estimate ^r from the true value r is given by ^r
r
2 3 [T r] X =4 Xt Xt0 5 t=1
12
3 [T r] X 4 X t "t 5 ;
(3)
t=1
where Xt = [~ T + ut 1 ~ T + ut 2 : : : ~ T + ut p+1 1 yt 1 ]0 ; = [ 1r 2r : : : pr 1 r r ]0 . As we can observe, the limiting distributions of (a1) (d1) of Lemma 0.3 are exactly the same as (9); (10); (12); (13) of Lemma 0.1 respectively. Therefore, if > 1=2, the limiting distribution of the SADF statistic under Case 2 is identical with that of the case when the regression includes a constant and the true process is a random walk without drift. From (d1) 5
P[T r] of Lemma 0.3, we know that the probability limit of t=1 Xt Xt0 is block diagonal. Therefore, P[T r] we only need to obtain the last 2 2 components of t=1 Xt Xt0 and the last 2 1 component P[T r] of t=1 Xt "t , which are 1 yt 1
yt yt2
1
"t
and
yt
1
respectively, where denotes summation over t = 1; 2; and (a1) of Lemma 0.3, the scaling matrix should be equation (3) by T , results in br b
T
r
r
=
r
and the matrix
T
(1)
hP
T
T 1=2 0 0 T
Rrr
W (s) ds
[T r] t=1 Xt "t
1
yt
r)
L
!
1 ( 2) ( 2)
T
;
> ;
=
1 "t
r
; [T r] : Based on (3) of Lemma 0.3 = diag T 1=2 ; T : Pre-multiplying 18
1
( 2) 1
"t
Under the null hypothesis that T 1=2 (^ r T ^r
i
i
9 > =
T 1=2 0 = 0 T 1 Rr r (1) r R 0 W (s) ds 2 2 r 2 r r (1) 0 W (s) ds
yt yt2
1
hP
1 T
( 2) ( 2)
[T r] 0 t=1 Xt Xt
1
r 0 1
2 3 [T r] X 14 Xt Xt0 5 t=1
yt
L
r
T 1
T 1=2 0 0 T
> :
1
Consider the matrix
!
8 > <
T
;
1 "t
T
> < > :
T
T
1
2 3 [T r] X 14 X t "t 5 t=1
1 yt
3=2
T T
1
3=2 2
A0;r = C0;r =
r
yt 1 yt2 1
;
T 1=2 "t T 1 yt 1 " t
L
!
"
1 2 2 r
rW h
1
C0;r D0;r
=
(r)
(1) W (r)2
r
1 B0;r r + A20;r
r
2 r
6
2
Z
r
r
i
#
:
B0;r C0;r + A0;r D0;r A0;r C0;r rD0;r
W (s) ds; B0;r = W (s)2 ds; r (1) 0 0 h i 1 2 2 (1) W (r) r : r W (r) ; D0;r = 2 r r r (1)
:
(4)
with Z
> ;
;
= 0;
r A0;r A0;r B0;r
( 2) 1
9 > =
Therefore, we have A0;r C0;r rD0;r : B0;r r + A20;r
L
T ^r !
To calculate the t-statistic of ^ r ; we need to …nd the standard error of ^ r . We know that ^r ^
var
1
2 r
=
r
yt
1
so the variances of T ^ r can be calculated as follows: ( 1 T 1=2 (^ r T 1=2 0 r) 2 var = r 0 T T ^r 2 r
L
!
1
yt yt2
1 1
1 yt
yt yt2
1
B0;r A0;r A0;r r
B0;r r + A20;r
1 1
T 1=2 0 0 T
1
)
1
:
Hence, the t-statistic of ^ r is T ^r ADFr = se T ^
1 2r
L
r
!
r1=2
By CMT, we have L
SADF (r0 ) ! sup
8 > <
1 2r
r2[r0 ;1] > : r1=2
h
W (r)2
nR r 0
h
0
i
W (s)2 dsr
W (r)2
nR r
r
r
i
W (s)2 dsr
Rr
W (s) dsW (r) o : Rr 2 1=2 0 W (s) ds
0
9 > = W (s) dsW (r) 0 : o Rr 2 1=2 > ; W (s) ds
Rr
0
P[T r] If = 1=2, we know that the probability limit of t=1 Xt Xt0 is block diagonal from (d2) P[T r] of Lemma 0.3. Similarly, we only need to obtain the last 2 2 components of t=1 Xt Xt0 and P[T r] the last 2 1 component of t=1 Xt "t to calculate the ADF statistics. Based on (3) of Lemma p T ; T . Consider the 0.1 and (a2) of Lemma 0.3, the scaling matrix should be T = diag hP i [T r] 1 0 matrix T 1 t=1 Xt Xt T ; ( 2) ( 2)
p
L
!
T 0 "
0 T
1
1 yt
1
yt yt2
1 1
p
T 0
r (1)
1 2r
+
Rr r 0
W (s) ds
(1)2
0 T h
1
Rr 1 2 r + r 0 W (s) ds R Rr 2 r W (s)2 ds + 2 r 0 r 0
(1) 1 3 3r
+ 7
i W (s) sds
#
=
1 0
0 (1)
1 2r
p
!
T 0 "
r
( 2) 1
(1)
r
0
0
r
r
n
T (^ r T ^r
+
;
(s) ds Rr 2 r 0 W (s) sds
1 "t
Rr
rW (r) "
(1)
r)
0
rW
r
=T
and
r
L
!
0
(r)
W (s) ds +
rW (r)
Under the null hypothesis that p
Rr 1 r + r 2 0 W R 2 r W (s)2 ds + r 0
1 0
0 (1)
"t yt
r
=
i
1
0 T
1 3 3r
W (s) ds
r 0
[T r] t=1 Xt "t
T
L
+
hP
1
and the matrix
Rrr
Rr
W (r)2
W (r)
0
W (s) ds +
r
= 0;
0 (1)
r
1 2 r
h
r 1
Ar;
r
h
1 2 r
r
io
W (r)2
1
Ar; Br;
#
r r
r
i
Cr; Dr;
#
:
r r
where Ar; Br; Cr;
r
r
r
1 = r+ 2 1 = r3 + 3
r
Z
r
W (s) ds;
0
2 r
Z
r
2
W (s) ds + 2
0
= W (r) ; Dr;
r
r
Z
= rW (r)
Z
r
W (s) sds
0
r
W (s) ds +
0
1 2
r
h
W (r)2
i r :
We can see that ^ r converges at rate T to the following distribution L
T ^r !
0
r
(1)
r
1
Ar;
Ar; Br;
r
1
Cr; Dr;
r r
r
r
=
(1)
r
rDr; r Ar; r Cr; Br; r r A2r; r
r
:
We know that var
p
T (^ r T ^r
r)
= L
!
2 r
( p
2 r
T 0
1 0
0 T 0 (1)
8
1
1 yt
1
1
r Ar;
r
yt yt2 Ar; Br;
p
T 0
1 1 1
r r
1 0
0 T 0 (1)
1
) 1
:
1
;
Hence, the t-statistic of ^ r is: ADFr =
T ^r se T ^
rDr;
L
!
r
r1=2
Ar; r Cr;
r
Br; r r
r
1=2
A2r;
:
r
By CMT, we have L
SADF (r0 ) ! sup
r2[r0 ;1]
(
rDr;
r
Ar; r Cr;
r1=2 Br; r r
A2r;
r
1=2 r
)
:
Third, consider Case 4. The regression model is the same as Case 2. P Under the null hypothesis that r = 0 and r = r , we have yt = ~ r + ut = ~ r t + ts=1 us ; where 1 1 2 2 p 1 p 1 . ~ r = (1) r and ut = (L) "t with (L) = 1 r L rL rL The deviation of the OLS estimate ^r from the true value r is given by 2 3 12 3 [T r] [T r] X X ^r Xt Xt0 5 4 X t "t 5 (5) r =4 t=1
t=1
where Xt = [~ r + ut 1 ~ r + ut 2 : : : ~ r + ut k 1 yt 1 ]0 , r = [ 1r 2r : : : pr 1 r r ]0 . P[T r] From (d) of Lemma 0.2, we know that the probability limit of t=1 Xt Xt0 is block diagonal. P[T r] Therefore, we only need to obtain the last 2 2 components of t=1 Xt Xt0 and the last 2 1 P[T r] component of of Lemma 0.2, t=1 Xt "t . Based on (3) of Lemma 0.1 and (a) hP i the scaling matrix 3 [T r] 1 1 1=2 0 should be T = diag T ; T 2 . Consider the matrix T t=1 Xt Xt T ; ( 2) ( 2)
1
T 1=2 0 1 yt 1 0 T 3=2 hP i [T r] and the matrix T 1 X " t t t=1
yt yt2
( 2) 1
T 1=2 0 0 T 3=2
1
Under the null hypothesis that T 1=2 (^ r T 3=2 ^ r 2
=
2r r 6r 3 ~ r 1
1
r
=
rW
L
1 "t r
and
!
~r
r
= 0,
1 2 r 2 ~rr ! 1 1 2 3 2 2 ~rr 3 ~rr Rr rW (r) + 3 R0 W (s) ds r 2 0 W (s) ds r rW (r)
r)
r
L
!
1 2 2 ~rr
1 2 2 ~rr 1 2 3 3 ~rr
;
"t yt
T 1=2 0 0 T 3=2
1
L
9
r
rW (r)
1
~r :
r
(r) Rr 0
W (s) ds
:
r W (r) Rr rW (r) 0 W (s) ds
;
(6)
We can see that L T ^ r ! 6r
3
~r
1
r
rW (r)
2
p
r 1 2 ~ 2 rr
2 r
r
W (s) ds :
0
The variances of T ^ r can be calculated as follows: ( 1 T 1=2 (^ r T 1=2 0 r) 2 var = r 0 T 3=2 T 3=2 ^ r !
Z
1 yt
1
1
1 2 2 ~rr 1 2 3 3 ~rr
1 1
4r 1 6r 2 ~ r 1
2 r
=
yt yt2
T 1=2 0 0 T 3=2
1
)
1
6r 2 ~ r 1 12r 3 ~ r 2
Hence, the t-statistic of ^ r is T 3=2 ^ r ADFr = se T 3=2 ^
L
r
!
Rr 0
By CMT, we have L
SADF (r0 ) ! sup
r2[r0 ;1]
(R r 0
Rr sdW (s) 0 W (s) ds : Rr 1=2 2 0 s ds
) Rr sdW (s) W (s) ds 0 : Rr 2 ds 1=2 s 0
Finally, consider Case 5 where the regression model is yt =
r+
r yt
1+
rt +
p 1 X
k r
p 1 X
k r ut k
yt
k
+ "t
k=1
This equation can equivalently be written as yt =
r
+
r
t+
rt +
+ "t
k=1
where ut k = yt k (1) r r , and r r (1) ; r = r (1) r (1 r) ; r = r; r = r + 1 1 2 2 p 1 p 1 = y (1) (t 1) with (L) = 1 L L L : Under the null t 1 r t r r r r hypotheses Pthat r = 0 and r = 0 and with the assumption that y0 = 0; we have ut = r (L) "t and t = ts=1 us : Consider a hypothetical regression of yt on ut 1 ; ; ut p+1 ; a constant, t and a time trend with the sample from 1 to [T r] ; producing the OLS estimates 2 3 2 3 12 3 ^r 1 t yt t 1 2 4 ^ 5=4 5 4 5 t 1 yt t 1t t 1 t 1 r 2 t yt t t ^r t 1t 10
The maintained hypothesis is that = r (1) r ; = 0 and = 0; we have = (L) 0 ; = 0 and = 0: The deviation of the OLS estimation from these true values are given by 2 3 2 3 12 3 ^r 1 t "t r (1) r t 1 2 4 5=4 5 4 5 ^ t 1 "t t 1 t 1t t 1 r 2 t"t t t ^r t 1t p T ; T; T 3=2 and based on Based on Lemma 0.1, the scaling matrix should be T = diag Lemma 0.2, we have 2 p 3 12 32 p 3 1 1 t T 0 0 T 0 0 t 1 2 4 0 T 54 0 T 0 5 4 0 5 t 1 t 1t t 1 2 3=2 t t 0 0 T 0 0 T 3=2 t 1t 2 3 T 1 1 T 3=2 t 1 T 2 t = 4 T 3=2 t 1 T 2 2t 1 T 5=2 t t 1 5 T 3 t2 Rr 3 1 2 r (1) r R 0 W (s) ds r 2 R R L r 2 (1)2 r W (s)2 ds ! 4 (1) r 0r W (s) ds (1) r 0 W (s) sds 5 r 0 R r 1 2 1 3 (1) r 0 W (s) sds 2r 3r Rr 2 32 32 1 2 r W (s) ds r 1 0 0 1 2 0 R R R (1) r 0 5 4 0r W (s) ds R0r W (s)2 ds 0r W (s) sds 5 4 0 =4 0 r 1 2 1 3 0 0 1 0 2r 3r 0 W (s) sds 2
T
2
t
T
5=2
t
t 1
and
2 p
T 4 0 0 2
0 T 0
0 0 T 3=2 0
r
=4 0 0
2 r
r
0
3 5
(1)
12
"t
3
2
(r)
i
3 0 0 5: 1
r
3
7 (1) W (r)2 r 5 Rr t"t rW (r) W (s) ds 3 0 32 0 h W (r) i 7 6 2 1 0 54 r 5: 2 W (r) Rr r rW (r) 0 W (s) ds 4
t 1 "t
L 6 5! 4
rW h
0 (1) 0
1 2 2 r
r
Under the null hypothesis that r = 0 and r = 0; 2 p 3 T (^ r r (1) r ) 4 5 T ^r 3=2 T ^r Rr 2 3 12 3 1 2 r W (s) ds r 1 0 0 2 0 R Rr Rr L 2 5 5 4 r W (s) ds ! r4 0 r (1) r 0 0 0 W (s) ds 0 W (s) sds R r 1 1 2 3 0 0 1 2r 3r 0 W (s) sds 11
1
2
1 4 0 0 2 L
where
0 (1) r 0 r
!4 0 0
r
0 r (1) 0
Ar =
Z
3 0 0 5 1 1
3 32 0 h W (r) i 6 7 2 1 r 5 r r (1) 0 5 4 2 W (r) Rr 0 1 rW (r) 0 W (s) ds 3 2 3 1 1 0 r Ar 2 r2 Dr 5 4 5 4 Ar Br Cr Er 5 0 1 2 1 3 Cr 3 r Fr r 2r 12
1 4 0 0 32
0
r
W (s) ds; Br =
0
Z
r
W (s) ds; Cr =
0
Dr = W (r) ; Er =
2
1h W (r)2 2
Z
r
W (s) sds; Z r i r ; Fr = rW (r) W (s) ds: 0
0
Therefore, we can see that ^ converges at rate T to the following distribution T ^r
L
^ r (1) = T r !
Er r 3
6Fr (2Cr Ar r) + 2Dr r (3Cr 2Ar r) : 12Ar Cr r 12Cr2 + Br r3 4A2r r2
The variances of T ^ can be calculated as follows: 02 p 31 T ^r V ar @4 T ^ r 5A T 3=2 ^ r 82 p 3 12 32 p 3 > 1 t T 0 0 T 0 0 < t 1 2 54 0 T = 2r 4 0 T 0 5 4 0 5 t 1 t 1t t 1 > 2 : t t 0 0 T 3=2 0 0 T 3=2 t 1t 2 3 1 T 1 1 T 3=2 t 1 T 2 t = 2r 4 T 3=2 t 1 T 2 2t 1 T 5=2 t 1 t 5 2 5=2 T t T t t 1 T 3 t2 3 12 3 2 32 0 0 0 0 r Ar 12 r2 r r L 1 1 !4 0 0 5 4 Ar Br Cr 5 4 0 0 5: r (1) r (1) 1 3 1 2 Cr 3 r 0 0 0 0 r r 2r
Therefore,
V ar T ^ r
L
r (1) !
12Ar Cr r
r3 12Cr2 + Br r3
4A2r r2
:
Hence, the t-statistic of ^ r is ^ ADFr =
r
se ^ r
L
!
Er r 3
6Fr (2Cr
r3=2 (12Ar Cr r 12
Ar r) + 2Dr r (3Cr 12Cr2
+ Br
r3
2Ar r) : 2 2 4Ar r )1=2
1
9 > = > ;
1
By CMT, the asymptotic SADF distribution is
L
SADF (r0 ) !
sup r22[r0 ;1]
(
Er r 3
6Fr (2Cr
r3=2 (12Ar Cr r
Ar r) + 2Dr r (3Cr 12Cr2 + Br r3
2Ar r)
4A2r r2 )1=2
)
:
REFERENCES Hamilton, J., 1994, 1994, Time Series Analysis. Princeton University Press. Phillips, P.C.B., 1987, Time series regression with a unit root, Econometrica 55, 277-301. Phillips, P.C.B., and Perron, P., 1988, Testing for a unit root in time series regression. Biometrika, 75(2):335–346. Phillips, P.C.B., and Solo, V., 1992, Asymptotics for linear processes. The Annals of Statistics, 20:971–1001.
13