their limiting distributions in an AR~1! model with a single structural break of ... Previous studies in the literature of structural change focus mainly on changes.
Econometric Theory, 17, 2001, 87–155+ Printed in the United States of America+
STRUCTURAL CHANGE IN AR(1) MODELS TE R E N C E TA I-LE U N G CH O N G The Chinese University of Hong Kong
This paper investigates the consistency of the least squares estimators and derives their limiting distributions in an AR~1! model with a single structural break of unknown timing+ Let b1 and b2 be the preshift and postshift AR parameter, respectively+ Three cases are considered: ~i! 6 b1 6 , 1 and 6 b2 6 , 1; ~ii! 6 b1 6 , 1 and b2 5 1; and ~iii! b1 5 1 and 6 b2 6 , 1+ Cases ~ii! and ~iii! are of particular interest but are rarely discussed in the literature+ Surprising results are that, in both cases, regardless of the location of the change-point estimate, the unit root can always be consistently estimated and the residual sum of squares divided by the sample size converges to a discontinuous function of the change point+ In case ~iii!, bZ 2 does not converge to b2 whenever the change-point estimate is lower than the true change point+ Further, the limiting distribution of the break-point estimator for shrinking break is asymmetric for case ~ii!, whereas those for cases ~i! and ~iii! are symmetric+ The appropriate shrinking rate is found to be different in all cases+
1. INTRODUCTION Previous studies in the literature of structural change focus mainly on changes that take place in stationary processes, for example, a process that changes from one stationary process to another+ Recently, there also have been studies on changes in nonstationary time series ~Hansen, 1992!+ An interesting case of structural change is found in Mankiw and Miron ~1986! and Mankiw, Miron, and Weil ~1987!; these authors conclude that the short term interest rate has changed from a stationary process to a near random walk since the Federal Reserve System was founded at the end of 1914+ However, the asymptotic theory on this kind of structural change still remains unexplored+ In this paper, we develop a comprehensive asymptotic theory for an AR~1! model with a single structural break of unknown timing+ Specifically, we examine the case where an AR~1! process changes from a stationary one to a nonstationary one ~or the other way around!+ In each case, the asymptotic criterion function is derived, consistency of estimators is established, and limiting This paper is based on chapter 4 of my Ph+D+ dissertation submitted to the University of Rochester+ I am grateful to my thesis adviser, Bruce Hansen, for his advice and encouragement+ My thanks also go to the editor and three anonymous referees for their helpful comments, which led to a substantial revision+ All errors are mine+ Address correspondence to: Terence Tai-Leung Chong, Department of Economics, the Chinese University of Hong Kong, Shatin, N+T+, Hong Kong; email: b792703@mailserv+cuhk+edu+hk+
© 2001 Cambridge University Press
0266-4666001 $9+50
87
88
TERENCE TAI-LEUNG CHONG
distributions of estimators are also derived+ As far as economic applications are concerned, we will restrict our discussion to the case where both the preshift and postshift parameters are in the interval ~21,1# + The plan of the paper is as follows+ Section 2 presents our basic model+ Section 3 deals with the case where the preshift and postshift parameters are both less than one in absolute value+ Section 4 studies the case where the preshift parameter is less than one in absolute value and the postshift parameter equals one+ Section 5 examines the case where the preshift parameter equals one and the postshift parameter is less than one in absolute value+ Section 6 discusses Monte Carlo experiments, and Section 7 concludes the paper+ Mathematical details are collected in Appendices A–K+ Interesting findings include the following results: ~i! in Sections 4 and 5 the asymptotic criterion function is random and has a sudden jump at the true break point; ~ii! in Section 4, the asymptotic distribution of the break-point estimator for shrinking break is asymmetric; and ~iii! in Section 5, the preshift estimator is always consistent, whereas the postshift estimator bZ 2 does not converge to b2 once the change-point estimate is lower than the true change point+
2. THE MODEL Our basic model is an AR~1! model without drift, with a structural break in the AR parameter b at an unknown time k 0 + We consider the following model: yt 5 b1 yt21 1$t # k 0 % 1 b2 yt21 1$t . k 0 % 1 «t
~t 5 1,2, + + + , T !,
(1)
where 1${% is an indicator function that equals one when the statement in the braces is true and equals zero otherwise+ We let t0 5 k 0 0T be the true break fraction and make the following assumptions+ ~A1! y0 is drawn from an independent and identical distribution with zero mean and a finite variance+ ~A2! «t ; i+i+d+~0, s 2 ! ∀t, 0 , s 2 , `, and E~«t4 ! , `+ S , ~0,1!+ ~A3! t0 [ @ st, t# We let k 5 @tT # where @{# is the greatest integer function and define sk 5 @ stT # and kN 5 @ tT S #+ Our interests are to estimate the structural parameters b1 and b2 and the time of change t0 + We estimate the following model: y[ t 5 bZ 1 yt21 1$t # @ t[ T T #% 1 bZ 2 yt21 1$t . @ t[ T T #%+
(2)
STRUCTURAL CHANGE IN AR(1) MODELS
89
The estimation method employed here is the least squares method proposed in Bai ~1994a, 1994b! and Chong ~1995!+ For any given t, the ordinary least squares ~OLS! estimators are given by @tT #
bZ 1 ~t! 5
(1 yt yt21 (1
(3)
,
@tT # 2 yt21
T
(
bZ 2 ~t! 5
@tT #11
yt yt21
(
(4)
,
T
@tT #11
2 yt21
and the change-point estimator satisfies t[ T 5 Arg min RSST ~t!,
(5)
t[~0,1!
where @tT #
RSST ~t! 5
(1
T
~ yt 2 bZ 1 ~t!yt21 ! 2 1
(
@tT #11
~ yt 2 bZ 2 ~t!yt21 ! 2+
We write kZ 5 @ t[ T T # , bZ 1 5 bZ 1~ t[ T !, and bZ 2 5 bZ 2 ~ t[ T !+ Throughout this paper, we let B1~{! and B2 ~{! be two independent Brownian motions defined on the non-negative half real line R 1 and B~{! and B~{! O be two independent standard Brownian motions on @0,1# + The symbol n denotes the weak convergence of a p d stochastic process, & represents convergence in probability, & denotes cond vergence in distribution and 5 stands for identical in distribution+ To achieve notational economy, the integral of a Brownian motion with respect to Lebesgue measure *B~r! dr is written as *B + The stochastic integral *B~r! dB~r! is written as *BdB+ 3. CASE WHERE _ b1 _ , 1 AND _ b2 _ , 1 3.1. The Asymptotic Criterion Function Consider the model in ~1! with 6 b1 6 , 1 and 6 b2 6 , 1+ A similar model has been studied by Salazar ~1982!, who provides a Bayesian analysis of structural changes in stationary AR~1! and AR~2! models with a known change point+ In our model, the change point t0 is unknown and has to be estimated by t[ T + To show the consistency of t[ T , the usual practice is to show that ~10T !RSST ~t! converges uniformly to a nonstochastic function that has a unique minimum at t 5 t0 + Thus, we will focus on the asymptotic behavior of ~10T !RSST ~t!+ The
90
TERENCE TAI-LEUNG CHONG
following lemma is useful in deriving the limiting behavior of the criterion function ~10T !RSST ~t! and in proving Theorem 1, which follows+ LEMMA 1+ Let $ yt % Tt51 be generated according to model ~1! with 6 b1 6 , 1 and 6 b2 6 , 1+ Under Assumptions ~A1!–~A3!, we have ~3a!
1
sup
T
0#tA#tB#1
1
~3b!
@t0 T #
(1
T 1
~3c!
T
~3d!
2 yt21
*(
*
@tB T # @tA T #
p
t0 s 2
&
T
2 ( yt21 @t T #11
yt21 «t 5 op ~1!;
1 2 b12 p
&
~1 2 t0 !s 2
st#t#t0
*
1 T
;
1 2 b22
0
sup
;
@t0 T #
2 2 ( yt21 @tT #11
~t0 2 t!s 2 1 2 b12
*
5 op ~1!;
sup 6 bZ 1 ~t! 2 b1 6 5 op ~1!;
~3e!
st#t#t0
5 o ~1!; * ~t 2 t!~1 2 b ! 1 ~1 2 t !~1 2 b ! * ~t 2 t !s 1 sup * T ( y 2 1 2 b * 5 o ~1!; t ~1 2 b !b 1 ~t 2 t !~1 2 b !b sup bZ ~t! 2 5 o ~1!; * t ~1 2 b ! 1 ~t 2 t !~1 2 b ! *
~3f !
sup
st#t#t0
bZ 2 ~t! 2
~t0 2 t!~1 2 b22 !b1 1 ~1 2 t0 !~1 2 b12 !b2 2 2
0
@tT #
~3g!
t0,t#tS
2
0
2 t21
@t0 T #11
2 2
1
t0,t#tS
p
2 2
0
~3h!
p
2 1
0
1
0
2 2
0
0
2 1 2 1
2
p
sup 6 bZ 2 ~t! 2 b2 6 5 op ~1!+
~3i!
t0,t#tS
Proof+ See Appendix A+ Note from ~3e! that when st # t # t0 , bZ 1~t! converges uniformly to b1 + This should be obvious because bZ 1~t! only utilizes the data generated by the process yt 5 b1 yt21 1 «t + However, from ~3f !, bZ 2 ~t! converges uniformly to a weighted average of b1 and b2 + The weight depends on the true change point, the true preshift and postshift parameters, and the location of t+ When t0 , t # t,S ~3h! and ~3i! display similar results+ From Appendix B, the asymptotic behavior of the residual sum of squares is as follows: sup st#t#t0
*
*
~t0 2 t!~1 2 t0 !~ b2 2 b1 ! 2 s 2 1 RSST ~t! 2 s 2 2 5 op ~1!, T ~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 ! (6)
sup t0,t#tS
* T RSS ~t! 2 s 2 t ~1 2 b ! 1 ~t 2 t !~1 2 b ! * 5 o ~1!+ t0 ~t 2 t0 !~ b2 2 b1 ! s 2
1
2
T
0
2 2
0
2
2 1
p
(7)
STRUCTURAL CHANGE IN AR(1) MODELS
91
Figure 1. Graph of ~10T !RSST ~t! for b1 5 0+5, b2 5 20+5, T 5 4000+
One can easily verify that ~10T !RSST ~t! converges uniformly to a piecewise concave function of t that attains a unique global minimum at t 5 t0 + A simulation of ~10T !RSST ~t! for T 5 4,000 is plotted in Figure 1+ 3.2. The Consistency and the Limiting Distributions of bZ 1 and bZ 2 If both b1 and b2 are less than one in absolute value then the process is said to be stationary throughout+ It is not difficult to show that all the OLS estimators are consistent in this case+ Bai ~1994a, 1994b! shows that, in the conventional stationary case, the change-point estimator is T-consistent+ This convergence rate is fast enough to make the limiting distributions of bZ 1 and bZ 2 behave as if the true change point t0 is known+ Theorem 1 establishes the asymptotic normality of bZ 1 and bZ 2 + THEOREM 1+ Under assumptions ~A1!–~A3!, if 6 b1 6 , 1 and 6 b2 6 , 1, the OLS estimators t[ T , bZ 1~ t[ T !, and bZ 2 ~ t[ T ! are consistent and
92
TERENCE TAI-LEUNG CHONG
6 t[ T 2 t0 6 5 Op
!T ~ bZ 1~ t[ T ! 2 b1 !
d
!T ~ bZ 2 ~ t[ T ! 2 b2 !
d
SD
S S
1 , T
D D
&
N 0,
1 2 b12 , t0
&
N 0,
1 2 b22 + 1 2 t0
(8)
Proof+ See Appendix C+ Thus, bZ 1~ t[ T ! and bZ 2 ~ t[ T ! are asymptotically normally distributed, with variances depending on b1 , b2 , and t0 + 3.3. The Limiting Distribution of t[ T as b2 Collapses to b1 In this section, we consider the limiting distribution of t[ T + For a shift with fixed magnitude, Hinkley ~1970! showed that the limiting distribution of the changepoint estimator depends on the underlying distribution of the innovation «t in a complicated manner+ Thus, statistical inference on the change point under a break of fixed magnitude is difficult to perform+ To obtain a limiting distribution of t[ T invariant to «t , we have to let the magnitude of change go to zero at an appropriate rate+ Further, to ensure the consistency of t[ T , we should let 6 b2 2 b1 6 shrink at a rate slower than the rate of convergence of bZ 1~t0 ! and bZ 2 ~t0 ! so that there still exists a relative shift in parameters+ To achieve this, we fix b1 and let b2T be a sequence of b2 such that 6 b2T 2 b1 6 r 0 and !T 6 b2T 2 b16 r ` as T r `+ Recall from equation ~5! that the change-point estimator is defined as def
t[ T 5 Arg min RSST ~t! 5 Arg min $RSST ~t! 2 RSST ~t0 !%+ t
t
Let y [ R be a finite constant+ We first examine the asymptotic behavior of RSST ~t! 2 RSST ~t0 ! at the region t 5 t0 1 y0T ~ b2T 2 b1 ! 2 + From Appendix D, we have the following expressions+ For y # 0, RSST ~t! 2 RSST ~t0 ! n 22s 2 B1
6y6 1 2 b12
S D
1
s 2 6y6 + 1 2 b12
(9)
S D
1
s 2y 1 2 b12
(10)
For y . 0, RSST ~t! 2 RSST ~t0 ! n 22s 2 B2
y 1 2 b12
where B1~{! and B2 ~{! are defined in Section 2+
STRUCTURAL CHANGE IN AR(1) MODELS
93
Let r 5 y0~1 2 b12 ! and applying the continuous mapping theorem for argmax functionals ~Kim and Pollard, 1990!, we have the following theorem+ THEOREM 2+ Under assumptions ~A1!–~A3!, suppose we fix 6 b1 6 , 1 and let b2T be a sequence of b2 such that 6 b2T 2 b1 6 r 0 and !T 6 b2T 2 b1 6 r ` as T r `+ Then the limiting distribution of t[ T is given by T ~ b2T 2 b1 ! 2 ~ t[ T 2 t0 ! 1 2 b12
d
&
S
Arg max B * ~r! 2 r[R
D
1 6r6 , 2
(11)
where B * ~r! is a two-sided Brownian motion on R defined to be B * ~r! 5 B1~2r! for r , 0 and B * ~r! 5 B2 ~r! for r $ 0+ The terms B1~{! and B2 ~{! are defined in Section 2+ Proof+ See Appendix D+ The exact distribution of the right hand side in ~11! was first derived by Picard ~1985!+ Yao ~1987! tabulates the numerical approximation of this distribution+ To understand the implication of Theorem 2, consider a Brownian motion moving along an inverted-V shape linear function with a kink at t0 + Because t[ T is the location where this motion achieves its maximum, its limiting distribution will be symmetric+ One should be careful that ~11! is not the exact distribution of the changepoint estimator for fixed shifts as we are letting the shift shrink to zero when deriving the theorem+ Rather, the theorem serves to provide a conservative confidence interval for t0 when the shift is small+ 4. CASE WHERE _ b1 _ , 1 and b2 = 1 We now consider the case where the AR~1! process shifts from a stationary process to an I ~1! process+ It is well known that the distribution theory for the least squares estimator in an AR process with a unit root or near unit root is nonstandard+ See, for example, White ~1958!; Dickey and Fuller ~1979!; Lai and Siegmund ~1983!; Ahtola and Tiao ~1984!; Chan and Wei ~1987, 1988!; Phillips ~1987, 1988!, and Perron ~1996!+ 4.1. The Asymptotic Criterion Function In our case, when b2 5 1, the criterion function ~10T !RSST ~t! behaves very differently+ The following lemma is useful in deriving its limiting behavior and in proving Theorem 3, which follows+ LEMMA 2+ Let $ yt % Tt51 be generated according to model ~1! with 6 b1 6 , 1 and b2 5 1+ Define
H
S ù t 5 t0 1 JT1 5 ~t0 , t#
j : j 5 1,2, + + + , @~1 2 t0 !T # T
and
lim Tr`
j
!T
J
5` +
94
TERENCE TAI-LEUNG CHONG
Then under Assumptions ~A1!–~A3!, we have ~4a!
*
(1
T
(
T
@t0 T #11
1
~4c!
T
~4d!
2
2
0
1 T
st#t#t0
sup
*
*(
T
E
1
B 2 5 Op ~1!;
0
*
@t0 T #
yt21 «t 5 op ~1!;
@tT #11
1
~B 2 ~1! 2 1! 5 Op ~1!;
@t0 T #
2 2 ( yt21 @tT #11
~t0 2 t!s 2 1 2 b12
*
5 op ~1!;
sup 6 bZ 1 ~t! 2 b1 6 5 op ~1!;
~4f !
st#t#t0
sup 6 bZ 2 ~t! 2 16 5 Op
~4g!
st#t#t0
SD 1
T
;
sup 6 bZ 2 ~t! 2 b1 6 5 Op ~1!;
~4h!
~4j!
~1 2 t0 !s 2
T
0#t#t0
~4i!
yt21 «t n
2 n ~1 2 t0 ! 2 s 2 ( yt21 @t T #11
sup
~4e!
yt21 «t 5 op ~1!;
T
1
~4b!
*
@t0 T #
1
st#t#t0
1
*
*
@tT #
(
yt21 «t 5 Op ~1!
for t [ JT1 ;
yt21 «t * 5 Op ~1! T * @tT( #11
for t [ JT1 ;
T
@t0 T #11
1
T
S! D SD 1
~4k! 6 bZ 1 ~t! 2 16 5 Op ~4l! 6 bZ 2 ~t! 2 16 5 Op
T
1
T
for t [ JT1 ;
for t [ JT1 +
Proof+ See Appendix E+ Note that the set JT1 is different from the set ~t0 , t# S as JT1 excludes sequences of t that converge too fast to t0 + For example, JT1 contains all given constants t such that t0 , t # t+S The sequence tT 5 t0 1 ~10log T ! also belongs to JT1+ However, the sequence tT 5 t0 1 ~10T ! is not in JT1+ From Appendix F, we have sup st#t#t0
*
*
~1 2 b1 !~t0 2 t!s 2 1 RSST ~t! 2 s 2 2 5 op ~1! T 1 1 b1
(12)
and 1 RSST ~t! T
p
&
s2 1
~1 2 b1 !t0 s 2 1 1 b1
for t [ JT1 +
(13)
STRUCTURAL CHANGE IN AR(1) MODELS
95
Note from ~12! and ~13! that ~10T !RSST ~t! converges uniformly to a downward sloping linear function for st # t # t0 and it converges to a flat line located above s 2 for t [ JT1+ Further, there is an asymptotic gap between ~10T !RSST ~t0 ! and ~10T !RSST ~t! for t [ JT1+ The transitional process begins at t0 and completes at the left boundary of JT1+ Thus the set JT1 establishes the rate at which plim~10T !RSST ~t! jumps from s 2 to a flat line located above s 2 once t . t0 + This rate serves as a conservative rate of consistency of t[ T + To examine the consistency of bZ 1 , we have to investigate the transitional behavior of ~10T !RSST ~t!+ It should be noted that, for any constant c . 0,
bZ 1 ~t0 1 cT
a21
1
k0
! 5 uT ~a, c! b1 1
(1 «t yt21 k0
2 (1 yt21
1
2
k 01@cT a #
(
1 ~1 2 uT ~a, c!! 1 1
k 011
«t yt21
k 01@cT a #
(
2 yt21
k 011
2
,
1@cT # 2 2 2 !0~ ( k10 yt21 1 ( kk0011 yt21 !+ where uT ~a, c! 5 ~ ( k10 yt21 1 _ For a , 2 , we have a
uT ~a, c!
p
&
1,
bZ 1 ~t0 1 cT a21 !
p
&
b1 ,
&
1,
and 1 RSST ~t0 1 cT a21 ! T For
1 _ 2
uT ~a, c!
p
&
s 2+
, a , 1, we have p
&
0,
bZ 1 ~t0 1 cT a21 !
p
and 1 RSST ~t0 1 cT a21 ! T
p
&
s2 1
~1 2 b1 !t0 s 2 + 1 1 b1
This implies that if the convergence rate of t[ T is faster than T 102 then bZ 1 will be consistent; otherwise it will be inconsistent+
96
TERENCE TAI-LEUNG CHONG
Note that the transition is defined only when a 5 _12 , which means the speed of the transition is 10!T + For a 5 _12 , Appendix F shows that uT
S D
1 ,c n 2
t0 t0 1 ~1 2 b12 !
E
,
c
B32
0
S
bZ 1 t0 1
c
!T
D
E E
c
t0 b1 1 ~1 2 b12 !
B32
0 c
n t0 1 ~1 2 b12 !
,
B32
0
S
c 1 RSST t0 1 T !T
D
t0 s 2 ~1 2 b1 ! 2
n s2 1
SE D
21
c
t0
B32
def
5 s 2 1 w~c!,
(14)
1 1 2 b12
0
where B3 ~{! is a Brownian motion defined on R 1 + Note that when c 5 0, p p ~10T !RSST ~t0 ! & s 2 , and as c r `, ~10T !RSST ~t0 1 ~c0!T !! & s 2 1 2 @~1 2 b1 !t0 s #0~1 1 b1 !+ Because ]w~c!0]c . 0, the transition is monotonically increasing+ Figure 2 simulates the behavior of ~10T !RSST ~t! using T 5 4,000+ For t # t0 , it is linear and decreasing+ For t . t0 , it is flat and has a sudden jump near t0 + 4.2. The Consistency and the Limiting Distributions of t[ T , bZ 1 , and bZ 2 For a fixed magnitude of the break, Theorem 3, which follows, shows that bZ 1 is asymptotically normally distributed with a variance depending on b1 and t0 , whereas bZ 2 has a scaled Dickey–Fuller distribution+ Similar to Section 3+3, we should let the magnitude of the break go to zero at a certain rate to obtain the asymptotic distribution of t[ T + However, if we make 0 b1 converge to b2 ~5 1!, the process $ yt % kt51 will be a near unit-root process+ Thus, we are dealing with a structural change from a near unit-root process to an exact unit-root process+ We fix b2 at 1 and let b1T be a sequence of b1 such that b1T r 1 and T ~1 2 b1T ! r ` as T r `+ Consider the region t 5 t0 1 y0T ~1 2 b1T ! where y [ R+ From Appendix G, we have the following expressions+ For y # 0, RSST ~t! 2 RSST ~t0 ! n 22s 2 B1 ~6y6!Ba ~ 212 ! 1 6y6s 2 Ba2 ~ 212 !+
STRUCTURAL CHANGE IN AR(1) MODELS
97
Figure 2. Graph of ~10T !RSST ~t! for b1 5 0+5, b2 5 1, T 5 4000+
For y . 0, RSST ~t! 2 RSST ~t0 !
ES y
n 2s 2
B2 ~s! 1 Ba
0
S DD 1 2
ES y
dB2 ~s! 1 s 2
B2 ~s! 1 Ba
0
S DD 1 2
2
ds,
where B1~{! and B2 ~{! are defined in Section 2+ The expression Ba ~ _12 ! is gener` ated by *0 exp~2s! dB1 ~s!+ Theorem 3, which follows, states that as b1 approaches one, the limiting distribution of t[ T is the argmax of a random process moving along an inverted-V shape linear function on the real line+ THEOREM 3+ Under Assumptions ~A1!–~A3!, if 6 b1 6 , 1 and b2 5 1, the OLS estimators t[ T , bZ 1~ t[ T !, and bZ 2 ~ t[ T ! are all consistent and
98
TERENCE TAI-LEUNG CHONG
6 t[ T 2 t0 6 5 Op
!T ~ bZ 1~ t[ T ! 2 b1 !
d
SD
S
1 , T
N 0,
&
D
1 2 b12 , t0
(15)
B 2 ~1! 2 1
T ~ bZ 2 ~ t[ T ! 2 1! n
2~1 2 t0 !
E
+
1
B
2
0
Suppose we fix b2 at one and let b1T be a sequence of b1 such that 61 2 b1T 6 r 0 and T ~1 2 b1T ! r ` as T r `+ Then the limiting distribution of t[ T is given by d
~1 2 b1T !T ~ t[ T 2 t0 !
&
S SD D ES SD SD D
Arg max y[R
1 C * ~y! 2 6y6 , 1 2 Ba 2
where C * ~y! is defined to be C * ~y! 5 B1~2y! for y # 0 and
E
C * ~y! 5 2B2 ~y! 2
0
y
B2 ~s! dB2 ~s! 2 1 Ba 2
y
0
B2 ~s! 1 1 B2 ~s! ds 1 2Ba 2
for y . 0+ The terms B1~{! and B2 ~{! are defined in Section 2+ Here Ba ~ _12 ! is generated by ` *0 exp~2s! dB1 ~s!+ Proof+ See Appendix G+ An interesting and important feature of Theorem 3 is that the distribution of t[ T is asymmetric about t0 + An intuitive explanation for this result is that ~10T !RSST ~t! for breaks with fixed magnitude is asymmetric in the neighborhood of t0 as shown in Figure 2+ As we compress the size of break, this asymmetry perseveres+ Thus, t[ T 5 arg min RSST ~t! should also be asymmetric about t0 + 5. CASE WHERE b1 = 1 and _ b2 _ , 1 Consider the case where b1 5 1 and 6 b2 6 , 1+ Intuitively, it seems that the results of this case are the mirror images of the case where 6 b1 6 , 1 and b2 5 1+ However, we show that a time reversing argument cannot be applied to model ~1!+ The findings in this section turn out to be very different from those in Section 4+
STRUCTURAL CHANGE IN AR(1) MODELS
99
5.1. The Asymptotic Criterion Function The following lemma is useful in deriving the limiting behavior of ~10T !RSST ~t! and in proving Theorem 4, which follows+ LEMMA 3+ Let $ yt % Tt51 be generated according to model ~1! with b1 5 1 and 6 b2 6 , 1+ Define
H
JT2 5 @ st, t0 ! ù t 5 t0 2
j : j 5 1,2, + + + , @t0 T # T
and
lim Tr`
j
!T
5`
J
and let JT1 be defined as in Lemma 2+ Under assumptions ~A1!–~A3!, we have ~5a! ~5b!
@t0 T #
1
(1
T
yt21 «t n
s2 2
~B 2 ~t0 ! 2 t0 ! 5 Op ~1!;
S
T
1
yt21 «t !T @t ( T #11
d
N 0,
&
0
~5c! ~5d!
~5e!
~5f ! ~5g! ~5h! ~5i!
~5j! ~5k! ~5l!
@t0 T #
1 T 1 T
(1
2
2 yt21 n s2
E
t0
2 n ( yt21 @t T #11
1 2 b22
* T ( y « * 5 O ~1!; 1
sup t[JT2
@tT #
t21
t
p
1
*
1 T
@t0 T #
(
@tT #11
*
yt21 «t 5 Op ~1!;
sup 6 bZ 1 ~t! 2 16 5 Op
t[JT2
SD 1
T
;
sup 6 bZ 2 ~t! 2 16 5 op ~1!;
t[JT2
sup t[JT1
sup t[JT1
*
1
@tT #
T
@t0 T #11
*T
1
(
*
yt21 «t 5 op ~1!;
( yt21 «t * 5 op ~1!; @tT #11 T
sup 6 bZ 1 ~t! 2 16 5 Op
t[JT1
SD 1
T
sup 6 bZ 2 ~t! 2 b2 6 5 op ~1!+
t[JT1
D
;
B 2 5 Op ~1!;
t 2 t0 1 B 2 ~t0 !
0
t[JT2
1 2 b22
0
@tT #
sup
s4
;
s2;
100
TERENCE TAI-LEUNG CHONG
Proof+ See Appendix H+ The set JT2 serves a similar purpose to that of JT1 in Lemma 2+ The term JT2 excludes sequences of t that converge to t0 at a speed faster than 10!T + For example, JT2 contains all given constants t such that st # t , t0 + The sequence tT 5 t0 2 ~10log T ! also belongs to JT2 , but the sequence tT 5 t0 2 ~10T ! does not+ We now derive the asymptotic behavior of ~10T !RSST ~t!+ From Appendix I, we have the following expressions+ For t [ JT2 , ~1 2 b2 !~1 2 t0 1 B 2 ~t0 !! 2 1 RSST ~t! n s 2 1 s, T 1 1 b2 sup ta , tb[JT2
1 6RSST ~tb ! 2 RSST ~ta !6 5 op ~1!+ T
(16)
For t 5 t0 , 1 RSST ~t0 ! T
p
&
s 2+
(17)
For t [ JT1 , ~1 2 b2 !~t 2 t0 1 B 2 ~t0 !! 2 1 RSST ~t! n s 2 1 s, T 1 1 b2 sup t[JT1
*
S D*
1 1 RSST ~t! 2 RSST t 1 T T
5 op ~1!+
(18)
Note from ~16! that, for t [ JT2 , ~10T !RSST ~t! converges to a flat line located randomly above s 2 1 @~1 2 b2 !~1 2 t0 !s 2 #0~1 1 b2 ! . s 2 + For t [ JT1 , ~10T !RSST ~t! converges to a random upward-sloping linear function of t+ The lower support for this line is s 2 1 @~1 2 b2 !~t 2 t0 !s 2 #0 ~1 1 b2 ! . s 2+ In both situations, ~10T !RSST ~t! converges to a random line whose position depends on the true change point t0 , the postshift AR parameter b2 , and the realization of the standard Brownian motion B~t0 !+ Note that the larger the magnitude value of break ~1 2 b2 !, the more easily the change will be detected+ Note that we work with JT2 and JT1 in this section to reflect the fact that plim~10T !RSST ~t! is discontinuous at both sides of t0 , whereas we work with the interval @ st, t0 # and JT1 in Section 4 because plim~10T !RSST ~t! is only discontinuous at the right-hand-side neighborhood of t0 +
STRUCTURAL CHANGE IN AR(1) MODELS
101
Figure 3 simulates the behavior of ~10T !RSST ~t! with T 5 4,000+ For t # t0 , it is a random flat line and has a sudden plunge near t0 + For t . t0 , it is linear, random, and increasing, which agrees with our theory+
5.2. The Monotonic Transition Note that for any positive constant c bounded away from 0, there is an asymptotic gap between ~10T !RSST ~t0 ! and ~10T !RSST ~t0 6 c!+ The speed of the transitional process is 10!T , and the transition is completed in JT2 and JT1 , respectively+ To investigate the consistency of the change-point estimator, it is sufficient to show that the asymptotic transition of ~10T !RSST ~t! from t 5 t0 to t 5 t0 6 c is monotonically increasing+
Figure 3. Graph of ~10T !RSST ~t! for b1 5 1, b2 5 0+5, T 5 4000+
102
TERENCE TAI-LEUNG CHONG
From Appendix J, for m 5 0,1,2, + + + and ~m0T ! r 0 as T r `, we have
S
*
sup bZ 1 t0 2 m
D
*
m 2 1 5 op ~1! T
and
S
bZ 2 t0 2 1 T
D
m ~1 2 b22 !mB 2 ~t0 !~1 2 b2 ! , 2 b2 n T ~1 2 b22 !mB 2 ~t0 ! 1 1 2 t0 1 B 2 ~t0 !
S
m T
RSST t0 2
D
n s2 1
~1 2 t0 1 B 2 ~t0 !!~1 2 b2 ! 2 mB 2 ~t0 ! 2 s ~1 2 b22 !mB 2 ~t0 ! 1 1 2 t0 1 B 2 ~t0 !
def
5 s 2 1 h 1 ~m!+ If m 5 0, we have ~10T !RSST ~t0 ! m r `,
S
m 1 RSST t0 2 T T
D
(19)
n s2 1
(20) p
&
s 2 + As
~1 2 b2 !~1 2 t0 1 B 2 ~t0 !!s 2 + 1 1 b2
Thus, we bridge the gap between ~10T !RSST ~t0 ! and ~10T !RSST ~t0 2 c!+ For t 5 t0 1 ~m0T !, m 5 0,1,2, + + + , and ~m0T ! r 0 as T r `, we have
*
S D S D
sup bZ 1 t0 1 m
*
sup bZ 2 t0 1 m
* 2 b * 5 o ~1!+
m 2 1 5 op ~1!, T
m T
2
p
From Appendix J, we also have
S
m 1 RSST t0 1 T T
D
n s2 1
~1 2 b2 !B 2 ~t0 !~1 2 b22m ! 2 s 1 1 b2
def
5 s 2 1 h 2 ~m!+
(21) p
If m 5 0, we have ~10T !RSST ~t0 ! & s 2 + As m r `, ~10T !RSST ~t0 1 ~m0T !! n s 2 1 $@~1 2 b2 !B 2 ~t0 !#0~1 1 b2 !%s 2+ Thus, we bridge the gap between ~10T !RSST ~t0 ! and ~10T !RSST ~t0 1 c!+ Note from equations ~20! and ~21! that, for all m . 0, we have h 1~m 1 1! . h 1~m! . h 1~0! 5 0 and h 2 ~m 1 1! . h 2 ~m! . h 2 ~0! 5 0+ Thus the transition is monotonic+ Note also from equation ~19! that, unless m 5 0, bZ 2 ~t0 2 ~m0T !! does not converge to b2 ! Remarks+ However, because of the special behavior of ~10T !RSST ~t! described in equations ~16!–~18! and ~20!–~21!, we will have lim Tr` Pr~ kZ Þ k 0 ! 5 0+ Thus, bZ 2 should be a consistent estimator of b2 in practice+ See Theorem 4, which follows+
STRUCTURAL CHANGE IN AR(1) MODELS
103
5.3. The Consistency and the Limiting Distributions of t[ T , bZ 1 , and bZ 2 Despite the fact that bZ 2 is extremely sensitive to the local behavior of k,Z Appendix K shows that for a fixed magnitude of break, lim Tr` Pr~ kZ Þ k 0 ! 5 0+ p This means T ~ t[ T 2 t0 ! & 0+ Hence the distribution of t[ T degenerates very fast for any fixed magnitude of break+ However, if we allow b2 to converge to T b1~5 1!, the process $ yt % t5k will be a near unit-root process with an initial 011 value drawn from an I ~1! process+ Thus, we are dealing with structural changes from an exact unit-root process to a near unit-root process with a nonstationary initial value+ We fix b1 at one and let b2T be a sequence of b2 such that !T ~1 2 b2T ! r 0 and T 304 ~1 2 b2T ! r ` as T r `+ Consider the region t 5 t0 1 y0~1 2 b2T ! 2 T 2, where y [ R+ From Appendix K, we have the following expressions+ For y # 0, RSST ~t! 2 RSST ~t0 ! n 22B~t0 !s 2 B1 ~6y6! 1 6y6s 2 B 2 ~t0 !+ For y . 0, RSST ~t! 2 RSST ~t0 ! n 22B~t0 !s 2 B2 ~y! 1 ys 2 B 2 ~t0 !, where B1~{!, B2 ~{!, and B~{! are defined in Section 2+ THEOREM 4+ Under Assumptions ~A1!–~A3!, if b1 5 1 and 6 b2 6 , 1 then t[ T , bZ 1~ t[ T !, and bZ 2 ~ t[ T ! are consistent, and Pr~ kZ Þ k 0 ! r 0, T ~ bZ 1 ~ t[ T ! 2 1! n
B 2 ~t0 ! 2 t0 2
E
(22)
,
t0
B2
0
!T ~ bZ 2 ~ t[ T ! 2 b2 ! n
O !1 2 b B~1! 2 2
1 2 t0 1 B 2 ~t0 !
+
Suppose we fix b1 at one and let b2T be a sequence of b2 such that !T ~1 2 b2T ! r 0 and T 304 ~1 2 b2T ! r ` as T r `+ Then the limiting distribution of t[ T is given by ~1 2 b2T ! 2 T 2 ~ t[ T 2 t0 !
d
&
Arg max y[R
S
D
B * ~y! 1 2 6y6 , B~t0 ! 2
where B * ~y! is a two-sided Brownian motion on R defined to be B * ~y! 5 B1~2y! O for y # 0 and B * ~y! 5 B2 ~y! for y . 0+ The terms B1~{!, B2 ~{!, B~{!, and B~{! are defined in Section 2+
104
TERENCE TAI-LEUNG CHONG
Proof+ See Appendix K+ Thus bZ 1 has a scaled Dickey–Fuller distribution+ The term bZ 2 can be described as being asymptotically normally distributed with a random variance+ This is because T
(
!T ~ bZ 2 2 b2 ! 5
@t0 T #11
«t yt21 0!T 1 op ~1!,
T
(
@t0 T #11
2 yt21 0T
the numerator follows a central limit theorem, and the denominator converges to a random variable by ~5b! and ~5d! in Lemma 3+ The rapid rate of convergence of kZ in Theorem 4 is very surprising+ The rate in this case is much faster than those in the previous two cases+ In previous cases, we only have 6 kZ 2 k 0 6 5 Op ~1!; that is, the estimation of k 0 is always subject to an error of order Op ~1!+ Now, with b1 5 1 and 6 b2 6 , 1, k 0 can be precisely estimated asymptotically, despite the fact that k 0 r ` as T r `+ Further, as b2 approaches one, the limiting distribution of t[ T is the argmax of a random process moving along an inverted-V shape linear function+ 6. MONTE CARLO EXPERIMENTS For empirical applications, we perform the following experiments to see how well our asymptotic results match the small-sample properties of the estimators+ In all experiments, the sample size is set at T 5 200 and the number of replications is set at N 5 20,000; $ yt % Tt51 is generated from model ~1!; $«t % Tt51 ; nid~0,1! and y0 ; nid~0,1! independent of $«t % Tt51+ The true change point is set at t0 5 0+5+ Experiment 1a+ This experiment verifies equation ~19!, which predicts that when b1 5 1 and 6 b2 6 , 1, the postshift structural estimator bZ 2 ~t! does not converge to b2 whenever t , t0 + We let b1 5 1; b2 5 0+5, 0, 20+5+ Let bNZ 2 ~t! be the mean of bZ 2 ~t! in 20,000 replications+ We consider those t in the neighborhood of t0 ~5 0+5!+ Note that when b1 5 1 and 6 b2 6 , 1, the estimate of b2 is inaccurate for t , t0 + The strange behavior of bNZ 2 ~t! is highlighted by italic figures in Table 1+ Note that bNZ 2 ~t! is not close to b2 even for t 5 t0 2 ~10T ! 5 0+495+ Experiment 1b+ With 6 b1 6 , 1, b2 5 1, and t close to t0 , we will show that both bZ 1~t! and bZ 2 ~t! are very close to the true parameters b1 and b2 , respectively+ Let b1 5 0+5, 0, 20+5; b2 5 1+ The results in Table 2 do not image those in Table 1+ Here bNZ 1~t! is very close to b1 + Note that in experiments 1a and 1b, the estimates of the unit root are close to one in all cases+
STRUCTURAL CHANGE IN AR(1) MODELS
105
Table 1. Results for experiment 1a bNZ 1~t!
bNZ 2 ~t!
bNZ 1~t!
bNZ 2 ~t!
bNZ 1~t!
bNZ 2 ~t!
bNZ
b1 5 1
b2 5 0+5
b1 5 1
b2 5 0
b1 5 1
b2 5 20+5
0+480 0+485 0+490 0+495 0+500 0+505 0+510 0+515 0+520
0+9819 0+9820 0+9822 0+9824 0+9826 0+9728 0+9702 0+9693 0+9689
0.7041 0.6773 0.6401 0.5869 0+4930 0+4915 0+4903 0+4899 0+4898
0+9817 0+9819 0+9821 0+9822 0+9824 0+9628 0+9622 0+9617 0+9612
0.4703 0.4163 0.3413 0.2249 0+0009 0+0012 0+0013 0+0013 0+0013
0+9819 0+9821 0+9823 0+9825 0+9826 0+9529 0+9450 0+9423 0+9409
0.1350 0.0541 20.0550 20.2160 20+4939 20+4923 20+4913 20+4909 20+4908
t
\
Experiment 2+ This experiment simulates the distributions of bZ 1 and bZ 2 + Case 1: b1 5 0+5, b2 5 20+5; Case 2: b1 5 0+5, b2 5 1; Case 3: b1 5 1, b2 5 0+5+
Figures 4 to 6 display the results for Cases 1 to 3, respectively+ For Case 1, Theorem 1 states that both bZ 1 and bZ 2 are asymptotically normally distributed+ The theorem is well supported by Figure 4 even for a sample size of 200+ For Case 2, Figure 5 shows that the small-sample distribution of bZ 1 is approximately normal, whereas bZ 2 appears to have a Dickey–Fuller distribution+ Both results agree with the theorem+
Table 2. Results for experiment 1b
t
\
bNZ
0+480 0+485 0+490 0+495 0+500 0+505 0+510 0+515 0+520
bNZ 1~t!
bNZ 2 ~t!
bNZ 1~t!
bNZ 2 ~t!
bNZ 1~t!
bNZ 2 ~t!
b1 5 0+5
b2 5 1
b1 5 0
b2 5 1
b1 5 20+5
b2 5 1
0+4899 0+4900 0+4900 0+4901 0+4902 0+4956 0+5044 0+5158 0+5290
0+9817 0+9820 0+9823 0+9827 0+9830 0+9830 0+9830 0+9830 0+9829
20+0001 20+0000 20+0000 20+0000 20+0000 0+0109 0+0301 0+0559 0+0862
0+9810 0+9815 0+9820 0+9825 0+9830 0+9830 0+9830 0+9830 0+9830
20+4902 20+4903 20+4904 20+4905 20+4905 20+4756 20+4512 20+4193 20+3819
0+9788 0+9799 0+9809 0+9819 0+9830 0+9830 0+9829 0+9829 0+9829
106
TERENCE TAI-LEUNG CHONG
Figure 4. ~a! Distribution of t0 T0~1 2 b12 ! ~ bZ 1 2 b1 !; —— ~T 5 200!, {{{{{{ ~T 5 `!+ ~b! Distribution of ~1 2 t0 !T0~1 2 b22 ! ~ bZ 2 2 b2 !; —— ~T 5 200!, {{{{{{ ~T 5 `!+
!
!
STRUCTURAL CHANGE IN AR(1) MODELS
107
Figure 5. ~a! Distribution of t0 T0~1 2 b12 ! ~ bZ 1 2 b1 !; —— ~T 5 200!, {{{{{{ ~T 5 `!+ ~b! Distribution of ~1 2 t0 !T0~ bZ 2 2 b2 !; —— ~T 5 200!, {{{{{{ ~T 5 `!+
!
108
TERENCE TAI-LEUNG CHONG
Figure 6. ~a! Distribution of t0 T~ bZ 1 2 b1 !; —— ~T 5 200!, {{{{{{ ~T 5 `!+ ~b! Distribution of T0~1 2 b22 ! ~ bZ 2 2 b2 !; —— ~T 5 200!, {{{{{{ ~T 5 `!+
!
STRUCTURAL CHANGE IN AR(1) MODELS
109
For Case 3, Theorem 4 predicts that bZ 1 should have a Dickey–Fuller distribution and bZ 2 should have a normal distribution+ Figure 6 agrees with Theorem 4+ Note that in a small sample, the variation of t[ T is nontrivial+ This variation is the main source of the small-sample bias of bZ 1 and bZ 2 + In Case 1 where both the pre- and postshift processes are stationary, the small-sample bias is very little+ However, in Cases 2 and 3, bZ 1 and bZ 2 consist of both stationary and nonstationary observations; thus their small-sample distributions should look like a mixture of normal and Dickey–Fuller distributions+ From Figures 5a and 6b, it is clear that the small-sample distribution has a larger variation and is more skew than a normal distribution+ From Figures 5b and 6a, the small-sample distribution has a smaller variation than and is not as skew as a Dickey–Fuller distribution+ Experiment 3+ In this experiment, we simulate the distribution of t[ T for shrinking break+ Case 1: b1 5 0+5, b2T 5 0+5 2 T 2108 ; Case 2: b1T 5 1 2 T 2104, b2 5 1; Case 3: b1 5 1, b2T 5 1 2 T 2508+
Thus, in each case, we simulate the behavior of T 304 ~ t[ T 2 t0 !+ Figures 7~a!–7~c! display the results of Cases 1 to 3, respectively+ For Cases 1 and 3, Figures 7a and 7c show that t[ T is symmetrically distributed, as predicted by Theorems 1 and 4+ For Case 2, Theorem 3 predicts that t[ T should be asymmetrically distributed+ Figure 7b agrees with our predictions+ Theoretically, as T r `, the domain for T 304 ~ t[ T 2 t0 ! should be unbounded+ However, because we fix T at 200 in this experiment, and t[ T is searched within the ~0,1! interval, hence the simulated distribution of T 304 ~ t[ T 2 t0 ! is bounded in the interval ~2200 304 3 0+5, 200 304 3 0+5! 5 ~226+6,26+6!+ Note that the distribution of t[ T has a “boundary effect,” especially for Case 3, in that there is a lot of mass near the boundary+ The clustering of mass at the boundary is due to the large value of Var~ bZ 2 ~t! 2 bZ 1~t!! when t nears zero or one+ Although 6 E ~ bZ 2 ~t! 2 bZ 1~t!!6 # 6 E ~ bZ 2 ~t0 ! 2 bZ 1~t0 !!6 for all t, actual realization of 6 bZ 2 ~t! 2 bZ 1~t!6 may be larger than 6 bZ 2 ~t0 ! 2 bZ 1~t0 !6 in finite sample, especially for the case where t is near the boundary+ For example, when t is near 0, bZ 1~t! will be calculated based only on very few observations, and therefore it has a large variance+ Note that t[ T is generated by the minimization of RSST ~t!, or in other words, by the maximization of 6 bZ 2 ~t! 2 bZ 1~t!6+ Thus, the probability that t[ T falls into the boundary is nontrivial+ This phenomenon will disappear as the sample size goes to infinity+ 7. CONCLUSION AR models have been used extensively in economics+ Many economic variables such as interest rates, real consumption, and real gross domestic product
110
TERENCE TAI-LEUNG CHONG
Figure 7. Distribution of T 304 ~ t[ T 2 0+5!: ~a! b1 5 0+5, b2 5 0+5 2 T 2108 ; ~b! b1 5 1 2 T 2104, b2 5 1; ~c! b1 5 1, b2 5 1 2 T 2508+
~GDP! can be well predicted by their own lags+ Some variables exhibit stationarity, whereas some variables display nonstationarity+ A shock to the production technology or a sudden change in the government policy may cause structural changes in these autoregressive models+ In particular, the changes may cause a stationary process to shift to a nonstationary one or vice versa+ In this paper, we present a structural change AR~1! model with independent and identically distributed ~i+i+d+! innovations+ We discuss three cases of struc-
STRUCTURAL CHANGE IN AR(1) MODELS
111
tural changes+ Case 1 deals with the change from a stationary process to another stationary process+ Case 2 examines the situation where a stationary process shifts to an I ~1! process+ Case 3 discusses the situation where an I ~1! process shifts to a stationary process with a nonstationary initial value+ In each case, consistency of the least squares estimators is established, and their limiting distributions are derived+ Having the limiting distributions of these estimates allows us to carry out statistical inference on the parameters of interest+ Cases 2 and 3 are intriguing+ Results of these two cases do not mirrorimage each other, which is counterintuitive+ A possible explanation for this asymmetry is that for Case 2, the initial value of the postshift process is stationary, whereas Case 3 is nonstationary+ Another explanation is that we are excluding an intercept in our model+ As pointed out by Banerjee, Lumsdaine, and Stock ~1992, p+ 278!, a more proper model for Case 3 should be, in our notations, yt 5 yt21 1$t # k 0 % 1 ~ yk0 ~1 2 b2 ! 1 b2 yt21 !1$t . k 0 % 1 «t + This process avoids a spurious sharp jump to zero at the break point+ It should also be mentioned that the disturbance terms $«t % Tt51 in this paper are assumed to be i+i+d+ In the case where both 6 b1 6 and 6 b2 6 are less than one, the assumption of i+i+d+ avoids the inconsistency of b’s Z due to serial correlation of «t + This assumption is also very helpful in calculating the long run variances of processes such as («t yt21 + However, extension to the cases of heterogeneous and0or dependent «t should be possible in the nonstationary cases at the expense of a more complicated mathematical treatment+ Thus, a generalization of our model to ARMA~ p, q! models with an intercept deserves future investigation+ REFERENCES Ahtola J+ & G+C+ Tiao ~1984! Parameter inference for a nearly nonstationary first order autoregressive model+ Biometrika 71, 263–272+ Andrews, D+W+K+ ~1987! Consistency in nonlinear econometric models: A generic uniform law of large numbers+ Econometrica 55, 1465–1471+ Andrews, D+W+K+ ~1988! Laws of large numbers for dependent non-identically distributed random variables+ Econometric Theory 4, 458– 467+ Bai, J+ ~1994a! Least squares estimation of a shift in linear processes+ Journal of Time Series Analysis 15, 453– 472+ Bai, J+ ~1994b! Estimation of Structural Change Based on Wald Type Statistics+ Working paper 94-6, M+I+T+, Department of Economics+ Banerjee, A+, R+L+ Lumsdaine, & J+H+ Stock ~1992! Recursive and sequential tests of the unit root and trend break hypothesis: Theory and international evidence+ Journal of Business and Economic Statistics 10, 271–287+ Chan, N+H+ & C+Z+ Wei ~1987! Asymptotic inference for nearly nonstationary AR~1! processes+ Annals of Statistics 15, 1050–1063+ Chan, N+H+ & C+Z+ Wei ~1988! Limiting distributions of least squares estimate of unstable autoregressive processes+ Annals of Statistics 16, 367– 401+ Chong, T+T+L+ ~1995! Partial parameter consistency in a misspecified structural change model+ Economics Letters 49, 351–357+ Dickey, D+A+ & W+A+ Fuller ~1979! Distribution of the estimators for autoregressive time series with a unit root+ Journal of the American Statistical Association 74, 427– 431+
112
TERENCE TAI-LEUNG CHONG
Hansen, B+E+ ~1992! Tests for parameter instability in regressions with I~1! processes+ Journal of Business and Economic Statistics 10, 321–335+ Hinkley, D+V+ ~1970! Inference about the change-point in a sequence of random variables+ Biometrika 57, 1–17+ Kim, J+ & D+ Pollard ~1990! Cube root asymptotics+ Annals of Statistics 18, 191–219+ Lai, T+L+ & D+ Siegmund ~1983! Fixed accuracy estimation of an autoregressive parameter+ Annals of Statistics 11, 478– 485+ Mankiw, N+G+ & J+A+ Miron ~1986! The changing behavior of the term structure of interest rates+ Quarterly Journal of Economics 101, 221–228+ Mankiw, N+G+, J+A+ Miron & D+N+ Weil ~1987! The adjustment of expectations to a change in regime: A study of the founding of the Federal Reserve+ American Economic Review 77, 358–374+ Perron, P+ ~1996! The adequacy of asymptotic approximations in the near-integrated autoregressive model with dependent errors+ Journal of Econometrics 70, 317–350+ Phillips, P+C+B+ ~1987! Time series regression with a unit root+ Econometrica 55, 277–301+ Phillips, P+C+B+ ~1988! Regression theory for near-integrated time series+ Econometrica 56, 1021– 1043+ Picard, D+ ~1985! Testing and estimating change-points in time series+ Advances in Applied Probability 17, 841–867+ Pötscher, B+M+ & I+R+ Prucha ~1989! A uniform law of large numbers for dependent and heterogeneous data process+ Econometrica 57, 675– 683+ Salazar, D+ ~1982! Structural changes in time series models+ Journal of Econometrics 19, 147–163+ White, H+ ~1984! Asymptotic Theory for Econometricians+ New York: Academic Press+ White, J+S+ ~1958! The limiting distribution of the serial correlation coefficient in the explosive case+ Annals of Mathematical Statistics 29, 1188–1197+ Yao, Y+C+ ~1987! Approximating the distribution of the maximum likelihood estimate of the changepoint in a sequence of independent random variables+ Annals of Statistics 15, 1321–1328+
APPENDIX A: PROOF OF LEMMA 1 If 6 b1 6 , 1, 6 b2 6 , 1, «t are i+i+d+, and y0 is drawn from a stationary process, then ~1! ~3a!, ~3d!, and ~3g! are consequences of the uniform law of large numbers in Andrews ~1987, Theorem 1!; ~2! ~3b! and ~3c! are consequences of the weak law of large numbers in Andrews ~1988, Theorem 2!+
* * @tT #
~3e!
sup 6 bZ 1 ~t! 2 b1 6 5 sup
st#t#t0
st#t#t0
(1 yt21 «t @tT #
2 (1 yt21
*( * @tT #
#
T @ stT #
(1
2 yt21
yt21 «t
sup st#t#t0
1
T
5 Op ~1!op ~1! 5 op ~1!+
STRUCTURAL CHANGE IN AR(1) MODELS
113
To show ~3f !, utilizing the fact that T
T
(
bZ 2 ~t! 2 b1 5
@t0 T #11
~ b2 2 b1 ! 1
T
(
@tT #11
(
2 yt21
@tT #11
«t yt21 ,
T
(
2 yt21
@tT #11
2 yt21
adding and subtracting ~1 2 t0 !s 2
T
1 2 b22
~ b2 2 b1 !
,
T
(
@tT #11
2 yt21
by the triangle inequality, and last by ~3a!, ~3c!, and ~3d!, we have
sup st#t#t0
*
bZ 2 ~t! 2
5 sup
st#t#t0
*
~t0 2 t!~1 2 b22 !b1 1 ~1 2 t0 !~1 2 b12 !b2 ~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 ! T
# sup
st#t#t0
*
T
(
@t0 T #11
( @tT #11
(
2 yt21
~ b2 2 b1 ! 1
T
2
*
@tT #11
«t yt21
T
2 ( yt21 @tT #11
2 yt21
~1 2 t0 !~1 2 b12 !~ b2 2 b1 ! ~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 !
T
( «t yt21 @tT #11 T
2 ( yt21 @tT #11
1 6 b2 2 b1 6 sup
st#t#t0
*
*
T
1 6 b2 2 b1 6 sup
st#t#t0
*
*
T
2 ( yt21 @t T #11 0
T
2 ( yt21 @tT #11
T 2
~1 2 t0 !s 2 1 2 b22 T
2 ( yt21 @tT #11
*
~1 2 t0 !s 2 1 2 b22 T
(
@tT #11
2 yt21
2
~1 2 t0 !~1 2 b12 ! ~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 !
*
114
TERENCE TAI-LEUNG CHONG
T
#
sup
T
(
2 yt21
@t0 T #11
st#t#t0
*
3 sup
st#t#t0
(
2 yt21
*
0
T
*
~1 2 b12 !~1 2 b22 !
sup 2 yt21
(
@t0 T #11
st#t#t0
*
T
( «t yt21 @tT #11 T
*
6 b2 2 b1 6T
1
T
(
@t0 T #11
1*
2 yt21
T
2 ( yt21 @t T #11 0
T
1 sup
st#t#t0
*
s 2 ~1 2 t0 !
2
1 2 b22
2 yt21
*
(
2 yt21
2
T
2
@tT #
(
@t0 T #11
1 2 b12
~ b2 2 b1 ! 1
@tT #
(1
adding and subtracting ~t 2 t0 !s 2 1 2 b22 ,
(1
2 yt21
;
@tT #
(1
2 yt21
@tT #
(1 «t yt21 2 yt21
2 yt21
T
*
*
T
0
T
2
*
s 2 ~t0 2 t!
@tT # 2 yt21
@tT #11
2 ( yt21 @t T #11
@t0 T # @tT #11
(
*2
The proof for ~3h! is analogous to ~3f !, using the fact that
~ b2 2 b1 !
1 2 b22
2 ( yt21 @tT #11
@~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 !#s 2
T
T
~1 2 t0 !s 2
T
6 b2 2 b1 6T
bZ 1 ~t! 2 b1 5
2
T
T
1
T
@~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 !#
T
(
6 b2 2 b1 6T
1
T
2 ( yt21 @t T #11
6 b2 2 b1 6~1 2 t0 !~1 2 b12 !T
st#t#t0
@t0 T #11
T
*
@t0 T #11
1 sup
#
T
( «t yt21 @tT #11
5 op ~1!+
~1 2 t0 !s 2 1 2 b22
*
STRUCTURAL CHANGE IN AR(1) MODELS
115
by the triangle inequality, and last by ~3a!, ~3b!, and ~3g! we have
sup t0,t#tS
*
t0 ~1 2 b22 !b1 1 ~t 2 t0 !~1 2 b12 !b2 t0 ~1 2 b22 ! 1 ~t 2 t0 !~1 2 b12 !
bZ 1 ~t! 2
* * @tT #
(1 «t yt21
# sup
t0,t#tS
@tT #
(1
t0,t#tS
#
sup
@t0 T #
(1
2 yt21
t0,t#tS
t0,t#tS
3 sup
t0,t#tS
T
1
2 yt21
@tT #
(1 «t yt21 T
*
1
@tT #
(1
2
2 (1 yt21
~t 2 t0 !~1 2 b12 ! t0 ~1 2 b22 ! 1 ~t 2 t0 !~1 2 b12 !
6 b2 2 b1 6T @t0 T #
(1
sup t0,t#tS
2 yt21
*
@tT #
2 yt21
*
*
2
~t 2 t0 !s 2 1 2 b22
*
2
~t 2 t0 !s 2 1 2 b22
*
@tT #
(
@t0 T #11
2 yt21
T
*
2 (1 yt21
@tT #
@t0 ~1 2
t0,t#tS
@t0 T #
5 op ~1!+
2 (1 yt21
0
~t 2 t0 !s 2 1 2 b22
T
@tT #
6 b2 2 b1 6T
(1
@tT #
2
2 ( yt21 @t T #11
@t0 ~1 2 b22 ! 1 ~t 2 t0 !~1 2 b12 !#
sup
@t0 T #
(1
*
*
~t 2 t0 !s 2 1 2 b22
T
*
@tT #
6 b2 2 b1 6~t 2 t0 !~1 2 b12 !T
1 sup
#
t0,t#tS
2 yt21
1 6 b2 2 b1 6 sup
T
1 6 b2 2 b1 6 sup
*
2 yt21
*
1*
1 ~t 2 t0 !~1 2 ~1 2 b12 !~1 2 b22 !
b22 !
@tT #
(1
«t yt21 T
*
1
6 b2 2 b1 6T @t0 T #
(1
@t0 T #
t0 s 2 2 1 2 b12
(1
b12 !#s 2
2 yt21
T
*
2 yt21
2
2 (1 yt21
T
sup t0,t#tS
1 sup
t0,t#tS
*
*
*
@tT #
(
@t0 T #11
2 yt21
T
@tT #
~t 2 t0 !s 2 2 1 2 b22
(
@t0 T #11
T
2 yt21
*2
116
~3i!
TERENCE TAI-LEUNG CHONG
sup 6 bZ 2 ~t! 2 b2 6 5 sup
t0,t#tS
t0,t#tS
*
T
( «t yt21 @tT #11 T
2 ( yt21 @tT #11
*
*( T
T
#
sup
T
(
@ tT S #11
@tT #11
2 yt21
t0,t#tS
«t yt21
T
*
5 Op ~1!op ~1! 5 op ~1!+
n
These prove Lemma 1+
APPENDIX B: DERIVATION OF EQUATIONS ~6! AND ~7! For st # t # t0 , we have @t0 T #
@tT #
RSST ~t! 5
(1 ~«t 2 ~ bZ 1~t! 2 b1 !yt21 ! 2 1 @tT(#11 ~«t 2 ~ bZ 2 ~t! 2 b1 !yt21 ! 2 T
1
( ~«t 2 ~ bZ 2 ~t! 2 b2 !yt21 ! 2 @t T #11 0
S( D @tT #
2
yt21 «t
T
5 ( «t2 2 1
1
@tT #
(1
@t0 T #
2 2~ bZ 2 ~t! 2 b1 !
2 yt21
(
@tT #11
yt21 «t
@t0 T #
1 ~ bZ 2 ~t! 2 b1 ! 2
T
2 2 2~ bZ 2 ~t! 2 b2 ! ( yt21 «t ( yt21 @tT #11 @t T #11 0
T
1 ~ bZ 2 ~t! 2 b2 ! 2
(
@t0 T #11
2 yt21 +
(B.1)
By ~B+1! and the fact that @t0 T #
bZ 2 ~t! 5 b1
2 ( yt21 @tT #11 T
(
@tT #11
2 yt21
T
1 b2
2 ( yt21 @t T #11 0
T
(
@tT #11
2 yt21
T
1
( «t yt21 @tT #11 ,
T
(
@tT #11
2 yt21
STRUCTURAL CHANGE IN AR(1) MODELS
117
we have RSST ~t! 2 RSST ~t0 !
S ( D S( D @t0 T #
2
@tT #
yt21 «t
5
1
1 1 1 31
22
1
@tT #
2 (1 yt21
2 yt21 T
~ b2 2 b1 !
2 ( yt21 @t T #11 0
T
(
@tT #11
~ b2 2 b1 !
(
@tT #11
(
T
(
2 yt21
@tT #11
@t0 T #
1
T
T
1
T
T
0
T
@t0 T #
(
@tT #11
T
2
0
T
21 2
2
2 yt21
(
(
@tT #11
(1
T
2 ( yt21 @t T #11 0
@t0 T #
24 2
T
2 ( yt21 @t T #11 0
2 ( yt21 ( yt21 «t @t T #11 @tT #11 0
T
2 yt21
1 L T ~t!,
(B.2)
S ( D S( D S ( 2
2
@tT # 1
1
@tT #
(1
2 yt21
@t0 T #11
«t yt21
@t0 T #11
2 yt21
@tT #11
«t yt21
T
(
D
2
T
2
T
(
D S( 2
T
yt21 «t
2
2 yt21
0
2
2 yt21
yt21 «t
@t0 T #
T
2 ( yt21 @tT #11
T
@tT #11
1
0
( «t yt21 @t T #11
T
2
2
T
( yt21 «t @t T #11
@t0 T #
@t0 T #11
@t0 T #
2 yt21
( «t yt21 @t T #11
2 ( yt21 @tT #11
(
1 ~ b2 2 b1 ! 2
2 ( yt21 @tT #11 T
yt21 «t
0
( «t yt21 @tT #11
2 ( yt21 ( yt21 «t @tT #11 @t T #11
(
@tT #11
2 ( yt21 @t T #11
T
2 ( yt21 @tT #11
@t0 T #
2
2 ( yt21 @tT #11
2 ( yt21 @tT #11
1
2 yt21
( «t yt21 @tT #11
@t0 T #
@t0 T #
2 2
T
2 ( yt21 @tT #11 2 ( yt21 @tT #11
~ b1 2 b2 !
2 yt21
( «t yt21 @tT #11
1
T
@tT #11
T
T
0
T
L T ~t! 5
1
T
5 2~ b2 2 b1 !
where
T
( «t yt21 @tT #11
2 yt21
2 ( yt21 @t T #11
2 2 ~ b1 2 b2 !
1
1
2
@t0 T #
(1
2
yt21 «t
@tT #11
2 yt21
+
118
TERENCE TAI-LEUNG CHONG
Let
l1 5
l2 5
S S
~1 2 t0 !~1 2 b12 !~ b2 2 b1 ! ~t0 2 t!~1 2
b22 !
1 ~1 2 t0 !~1 2
D D
2
b12 !
~t0 2 t!~1 2 b22 !~ b2 2 b1 !
and
2
~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 !
+
By ~B+1!, the fact that ~ b2 2 b1 ! 2 s 2 ~t0 2 t!~1 2 t0 ! ~t0 2 t!~1 2
b22 !
1 ~1 2 t0 !~1 2
b12 !
5 l1
~t0 2 t!s 2 12
b12
1 l2
~1 2 t0 !s 2 1 2 b22
,
and the triangle inequality, we have
sup st#t#t0
*T
1
RSST ~t! 2 s 2 2
~ b2 2 b1 ! 2 s 2 ~t0 2 t!~1 2 t0 ! ~t0 2 t!~1 2 b22 ! 1 ~1 2 t0 !~1 2 b12 !
S( D @tT #
#
*T ( « 2s * 1 1
2
1
1 sup
*
1 sup
*
1 sup
*
st#t#t0
st#t#t0
st#t#t0
1 sup
st#t#t0
1
sup
@tT #
st#t#t0
2~ bZ 2 ~t! 2 b1 !
~ bZ 2 ~t! 2 b1 ! 2
2~ bZ 2 ~t! 2 b2 !
*
2
yt21 «t
T
2 t
*
~ bZ 2 ~t! 2 b2 ! 2
2 (1 yt21
T
1 T 1 T 1 T 1 T
@t0 T #
(
@tT #11
yt21 «t
*
@t0 T #
2 2 l1 ( yt21 @tT #11
T
( yt21 «t @t T #11 0
T
~t0 2 t!s 2
*
*
2 2 l2 ( yt21 @t T #11 0
1 2 b12
~1 2 t0 !s 2 1 2 b22
*
+
We add and subtract ~ bZ 2 ~t! 2 b1 ! 2 @~t0 2 t!s 2 #0~1 2 b12 ! in the fourth term and ~ bZ 2 ~t! 2 b2 ! 2 @~1 2 t0 !s 20~1 2 b22 !# in the sixth term+ By the triangle inequality, the fact that sup6a 2 2 b 2 6 # sup6a 2 b6sup6a 1 b6, and Lemma 1, the preceding expression is bounded by
STRUCTURAL CHANGE IN AR(1) MODELS #
*T
sup 6 bZ 1 ~t! 2 b1 6 sup * ( yt21 «t * (1 «t2 2 s 2 * 1 st#t#t T 1 st#t#t T
1
1
0
*
1 sup ~ bZ 2 ~t! 2 b1 ! 2 sup
*
st#t#t0
st#t#t0
st#t#t0
st#t#t0
~t0 2 t!s 2
1 sup
1 2 b12
st#t#t0
T 1 T
@t0 T #
(
@tT #11
yt21 «t
@t0 T #
*
2 2 ( yt21 @tT #11
~t0 2 t!s 2 1 2 b12
*
sup 6~ bZ 2 ~t! 2 b1 ! 2 2 l 1 6
st#t#t0
1 sup ~ bZ 2 ~t! 2 b2 ! 2 st#t#t0
1 2 b22
1
st#t#t0
1 2 sup 6 bZ 2 ~t! 2 b2 6
~1 2 t0 !s 2
@tT #
0
1 2 sup 6 bZ 2 ~t! 2 b1 6 sup
1
119
*
*
T
( yt21 «t @t T #11 0
T
*
T
2 ( yt21 @t T #11 0
T
2
~1 2 t0 !s 2 1 2 b22
*
sup 6~ bZ 2 ~t! 2 b2 ! 2 2 l 2 6
st#t#t0
5 op ~1!+
n
This derives equation ~6!+ For t0 , t # t,S we can write the residual sum of squares as follows: @t0 T #
RSST ~t! 5
(1
@tT #
~«t 2 ~ bZ 1 ~t! 2 b1 !yt21 ! 2 1
(
@t0 T #11
~«t 2 ~ bZ 1 ~t! 2 b2 !yt21 ! 2
T
1
( ~«t 2 ~ bZ 2 ~t! 2 b2 !yt21 ! 2 @tT #11
T
5 ( «t2 2 2~ bZ 1 ~t! 2 b1 ! 1
2 2~ bZ 1 ~t! 2 b2 !
S(
@tT #11
«t yt21
T
(
(1
@t0 T #
yt21 «t 1 ~ bZ 1 ~t! 2 b1 ! 2
(1
@tT #
@tT #
0
0
2 yt21
2 yt21 ( yt21 «t 1 ~ bZ 1~t! 2 b2 ! 2 @t ( @t T #11 T #11
D
2
T
2
@t0 T #
@tT #11
2 yt21
+
(B.3)
120
TERENCE TAI-LEUNG CHONG
By ~B+3! and the fact that @t0 T #
@tT #
(1
bZ 1 ~t! 5
2 yt21
b1 1
@tT #
(1
@tT #
2 ( yt21 @t T #11 @tT #
(1
2 yt21
(1 «t yt21
b2 1
0
,
@tT #
(1
2 yt21
2 yt21
we have
RSST ~t! 2 RSST ~t0 ! 5 2~ b2 2 b1 !
1
@t0 T #
@tT #
( yt21 «t (1 @t T #11 @tT #
(1
(
~t 2 t0 !~1 2 b12 !~ b2 2 b1 !
D
and
2
2 yt21
1 L T ~t!+
(B.4)
2 yt21
2
t0 ~1 2 b22 ! 1 ~t 2 t0 !~1 2 b12 !
2 yt21
2 (1 yt21
(1
2 yt21
(1
S
(
@t0 T #11
@tT #
2 yt21
@tT #
Let
@tT #
yt21 «t
@t0 T #
@t0 T #11
1 ~ b2 2 b1 ! 2
(1
2
0
@tT #
l3 5
@t0 T # 2 yt21
l4 5
S
t0 ~1 2 b22 !~ b2 2 b1 !
D
2
t0 ~1 2 b22 ! 1 ~t 2 t0 !~1 2 b12 !
+
By ~B+3!, the fact that t0 ~ b2 2 b1 ! 2 s 2 ~t 2 t0 ! t0 ~1 2
b22 !
1 ~t 2 t0 !~1 2
5 l3
b12 !
t0 s 2 12
b12
1 l4
~t 2 t0 !s 2 ,
1 2 b22
and the triangle inequality, we have
t0,t#tS
#
* T RSS ~t! 2 s 2 t ~1 2 b ! 1 ~t 2 t !~1 2 b ! * 1
sup
*
t0 ~ b2 2 b1 ! 2 s 2 ~t 2 t0 !
2
T
2 2
0
1 T
T
(1 «t2 2 s 2
1 sup
t0,t#tS
1 sup
t0,t#tS
1 sup
t0,t#tS
* *
*
*
*
1 sup
t0,t#tS
~ bZ 1 ~t! 2 b1 ! 2 ~ bZ 1 ~t! 2 b2 !
1 T
~ bZ 1 ~t! 2 b1 !
2 T
@t0 T #
(1
2 yt21 2 l3
2
@tT #
T
@t0 T #11
(
2 1
0
yt21 «t
@t0 T #
(1
yt21 «t
t0 s 2 1 2 b12
*
*
@tT #
(
~ bZ 1 ~t! 2 b2 ! 2
@t0 T #11
T
*
2 yt21
2 l4
~t 2 t0 !s 2 1 2 b22
*
S(
1 sup
t0,t#tS
D
2
T
@tT #11
«t yt21
T
T
(
@tT #11
2 yt21
+
STRUCTURAL CHANGE IN AR(1) MODELS
121
We add and subtract ~ bZ 1~t! 2 b1 ! 2 @t0 s 2 0~1 2 b12 !# in the third term and ~ bZ 1~t! 2 b2 ! 2 @~t 2 t0 !s 2 #0~1 2 b22 ! in the sixth term+ Using sup6a 2 2 b 2 6 # sup6a 2 b6sup6a 1 b6, sup~a 2 b! 2 # ~sup6a 2 b6! 2 , and Lemma 1, the preceding expression is bounded by #
*
( «t2 2 s 2 1
T
1 1
* *
T
1
1
*
1
@t0 T #
T
(1
2 yt21 2
t0 s 2 1 2 b12
2 T
@t0 T #
(1
t0 s 2 1 2 b12
*
yt21 «t
sup ~ bZ 1 ~t! 2 b1 ! 2
sup 6~ bZ 1 ~t! 2 b1 ! 2 2 l 3 6
t0,t#tS
*
1 sup ~ bZ 1 ~t! 2 b2 ! 2 sup
*
t0,t#tS
t0,t#tS
t0,t#tS
t0,t#tS
sup 6 bZ 1 ~t! 2 b1 6
t0,t#tS
t0,t#tS
1 2 sup 6 bZ 1 ~t! 2 b2 6 sup
1 sup
*
t0,t#tS
~t 2 t0 !s 2 1 2 b22
1
@tT #
T
@t0 T #11
1
@tT #
T
(
yt21 «t
*
2 2 ( yt21 @t T #11
~t 2 t0 !s 2
0
1 2 b22
*
sup 6~ bZ 1 ~t! 2 b2 ! 2 2 l 4 6
t0,t#tS
1 sup 6 bZ 2 ~t! 2 b2 6 sup t0,t#tS
t0,t#tS
*T
1
( yt21 «t * @tT #11 T
5 op ~1!+ This shows equation ~7!+
n
APPENDIX C: PROOF OF THEOREM 1 If 6 t[ T 2 t0 6 5 Op ~10T !, or equivalently 6 kZ 2 k 0 6 5 Op ~1!, then for any h . 0, there exists an M , ` such that Pr~6 kZ 2 k 0 6 . M ! , h+ We shall prove this by using the contradiction argument+ Suppose t[ T is not T-consistent; then there exists a sequence MT . 0 such that MT r `, ~MT 0T ! r 0 as T r `, and lim Pr~6 kZ 2 k 0 6 . MT ! 5 a,
Tr`
where a is a positive constant in ~0,1# + Now, let l 5 b2 2 b1 , Z T 5 $1,2, + + + + , T %, D1T 5 $m : m [ Z T , m , k 0 2 MT %, D2T 5 $m : m [ Z T , m . k 0 1 MT %, D3T 5 $m : m [ Z T , k 0 2 MT # m # k 0 1 MT %+
122
TERENCE TAI-LEUNG CHONG
Note that
S
Pr~6 kZ 2 k 0 6 . MT ! 5 Pr infm[D1T øD2T RSST
SD m
T
, infm[D3T RSST
S DD m
T
+
Because k 0 [ D3T , we have infm[D3T RSST ~m0T ! # RSST ~t0 !, and the preceding probability is bounded by
S
SD F SD m
# Pr infm[D1T øD2T RSST 2
#
( Pr i51
S
T
m
infm[DiT RSST
T
D G D
, RSST ~t0 !
2 RSST ~t0 ! , 0 +
By using ~B+2! and ~B+4! in Appendix B, the right-hand side equals
1 51
5 Pr infm[D1T
2l
k0
T
( yt21 «t k 11 0
T
( k 11
2
( yt21 «t m11 k0
2 ( yt21 m11
2 yt21
0
2
T
2 ( yt21 m11
1 l2 1
(
2 yt21
k 011
1 51
1 Pr infm[D2T
2l
LT
k0
T
(
2 yt21
m
T
m11
k0
m
( yt21 «t (1 yt21 «t k 11 2
0
m
( k 11
k0
2 (1 yt21
2 yt21
0
(
k 011
LT
k0
m
2 yt21
(1
2 yt21
,0
2
m
2 ( yt21
1 l2 1
62
SD
62
SD m
T
,0 +
Using 2inf x 5 sup~2x!, sup~x 1 y! # sup x 1 sup y, sup xy # sup x sup y, and sup x # sup6 x6, the preceding item is bounded by # Pr~22lA 1 1 2lA 2 1 A 3 . l2 ! 1 Pr~2lA 4 2 2lA 5 1 A 6 . l2 !,
STRUCTURAL CHANGE IN AR(1) MODELS where T
A1 5
( yt21 «t k 11 0
T
(
5 op ~1!
2 yt21
k 011
by ~3a! and ~3c! of Lemma 1; k0
( yt21 «t m11
A 2 5 sup
# sup
k0
m[D1T
(
m[D1T
2 yt21
m11
*
k0
1
yt21 «t
( k 0 2 m m11
k0
(
2 yt21 0~k 0 2 m!
m11
*
5 op ~1!
by the uniform law of large numbers in Andrews ~1987, Theorem 1!;
A 3 5 sup m[D1T
3
*
T
2 ( yt21 m11
LT
k0
T
2 2 ( yt21 ( yt21 k 11 m11 0
1 * sup
k0
2
(1
*
k 011
2 ( yt21 k 11 0
1
MT
m
( yt21 «t k 11 0
m
(
2 yt21
5 op ~1!
k0
(
k 02MT
2 yt21
*
2
«t yt21
m11
T
2 ( yt21 m11
~op ~1! 1 Op ~1!! 5 Op
5 op ~1!
k 011
2 yt21
T
2
T
by ~3a! and ~3b! of Lemma 1;
m[D2T
(
k 011
1
1
S( D S( D 2
«t yt21
2 yt21
A 5 5 sup
m
2 (1 yt21
T
1 Op
k0
1
2 (1 yt21
k0
A4 5
1 T
2
m
2
k0
S S D S DD (1 yt21 «t
*1 #
yt21 «t
1
m[D1T
1
T
yt21 «t
1 sup
T
m
S( D S( D
m[D1T
5 Op
SD
*2
S D 1
MT
5 op ~1!;
2
123
124
TERENCE TAI-LEUNG CHONG
by a similar argument as in A 2 ;
A 6 5 sup m[D2T
3
*
m
2 (1 yt21 m
(
k 011
1 * sup
1
m[D2T
2
k0
2 (1 yt21 2
yt21 «t
1
2 (1 yt21
*
S
1 Op
1
2
k0
m
2 (1 yt21
T
( «t yt21 k 11 0
D S 2
2
T
T
(
D
2
T
~Op ~1! 1 op ~1!! 5 Op
1 k 01MT k 011
2 ( yt21 m11
0
1
*
( «t yt21 m11
2 ( yt21 k 11
MT
1
k0
m
yt21 «t
S S D S DD 1
*1
1
#
S( D S( D
1 sup
T
m
T
2 2 yt21 ( yt21
m[D2T
5 Op
SD
LT
k0
2 yt21
2
*2
S D 1
MT
5 op ~1!+
Because A 1 –A 6 are op ~1! and l2 is a positive constant, we have Pr~6 kZ 2 k 0 6 . MT ! # Pr~op ~1! . l2 ! 1 Pr~op ~1! . l2 ! r 2 Pr~l2 , 0! 5 0+ This means a 5 0, which contradicts the original argument that a . 0+ Thus t[ T is T-consistent+ To find the limiting distributions of bZ 1~ t[ T ! and bZ 2 ~ t[ T !, note that t[ T 2 t0 5 Op ~10T ! and
1
!T ~ bZ 1~ t[ T ! 2 bZ 1~t0 !! 5 !T
5 1$ t[ T # t0 %!T
1
@ t[ T T #
(1
@t0 T #
yt yt21 2
@ t[ T T #
(1
1 1$ t[ T . t0 %!T
(1
1
(1
2
@ t[ T T #
(1
@t0 T #
@t0 T #
«t yt21
«t yt21
@t0 T #
(1
2 yt21
yt yt21
(1
2 yt21
@ t[ T T #
(1
2 yt21
2 yt21
2 @ t[ T T #
@t0 T #
(1
(
~ b1 yt21 1 «t !yt21 1
@ t[ T T #
(1
(1
@t0 T #11
~ b1 yt21 1 «t !yt21 @t0 T #
(1
2 yt21
~ b2 yt21 1 «t !yt21 @ t[ T T #
(1
2 yt21
@t0 T #
2
2
2
2 yt21
STRUCTURAL CHANGE IN AR(1) MODELS
5 1$ t[ T # t0 %!T
1
@t0 T #
@t0 T #
@t0 T #
2 ( yt21 (1 @ t[ T #11
(1
1
@t0 T #
(1
2 yt21 @ t[ T T #
1 1$ t[ T . t0 %!T 2
@ t[ T T #11
2
T
@ t[ T T #
(
«t yt21
@ t[ T T #
(1
2 yt21
1
@t0 T #
(1
2 yt21
(
2 yt21
@t0 T #11 @ t[ T T #
(1
2
2 yt21
«t yt21
@ t[ T T #
(1
2 yt21
2
S S D S ! D S DD ! S S D S D S D S DD !
5 1$ t[ T # t0 %!T Op 1 1$ t[ T . t0 %
@t0 T #11
2 yt21
@ t[ T T #
1 ~ b2 2 b1 !
(
«t yt21
0
(1
2 yt21
@ t[ T T #
@t0 T #
2 ( yt21 (1 @t T #11 @ t[ T T #
«t yt21
125
1
T
T Op
1
Op
1
T
2 Op
T
Op
1
1
T
1
1 Op
T
T
1
1 Op
T
5 op ~1!+
Thus, bZ 1~ t[ T ! and bZ 1~t0 ! have the same asymptotic distribution+ Similarly, bZ 2 ~ t[ T ! and bZ 2 ~t0 ! have the same asymptotic distribution+ Define It 5 s~«i : i # t ! as the sigma field generated by the past history of $«t %+ @t0 T # T Because $«t yt21 , It % t51 and $«t yt21 , It % t5@t are martingale difference sequences 0 T #11 with E~«t yt21 6It21 ! 5 0
~t 5 1,2, + + + , T !,
@t0 T #
1 @t0 T # 1
( t51
E~~«t yt21 ! 2 6It21 !
p
&
s4
, `,
1 2 b12
T
E~~«t yt21 ! 2 6It21 ! T 2 @t0 T # t5@t(T #11
p
&
0
s4 1 2 b22
, `+
Applying the central limit theorem for martingale difference sequences ~see, e+g+, White, 1984, ch+ V! and by ~3b! and ~3c! in Lemma 1, we have @t0 T #
!T ~ bZ 1~ t[ T ! 2 b1 ! 5d !T ~ bZ 1~t0 ! 2 b1 ! 5
(1
S
«t yt21 0!T d
@t0 T #
(1
&
N 0,
2 yt21 0T
T
!T ~ bZ 2 ~ t[ T ! 2 b2 ! 5d !T ~ bZ 2 ~t0 ! 2 b1 ! 5
( «t yt21 0!T @t T #11 0
T
(
@t0 T #11
This proves Theorem 1+
2 yt21 0T
d
&
1 2 b12
S
N 0,
t0
D
,
1 2 b22 1 2 t0
D
+
n
126
TERENCE TAI-LEUNG CHONG
APPENDIX D: PROOF OF THEOREM 2 To derive the limiting distribution of t[ T for shrinking shift, let b2T 5 b1 1 ~10!g~T !!, where g~T ! . 0, with g~T ! r ` and @ g~T !0T # r 0 as T r ` + For t 5 t0 1 y@ g~T !0T # and y # 0, by ~B+2! in Appendix B and the facts that
S( D @t0 T #
2
yt21 «t
~i! L T ~t! 5
1
@t0 T #
(1
S( T
12
2 (1 yt21
T
( yt21 «t @tT #11
12
2
T
2 ( yt21 @t T #11 0
T
2 ( yt21 @t T #11 0
2 yt21
@tT #
yt21 «t
2
T
(1
@t0 T #
«t yt21
@t0 T #11
1
11 2 2 D 11 2 2 (1 yt21 «t (1
2 yt21
2 @t0 T #
@tT #
T
( yt21 «t @t T #11
2 ( yt21 @tT #11
0
5 Op ~1!~1 2 ~1 2 op ~1!! ~1 1 op ~1!!! 2
1 Op ~1!~1 2 ~1 1 op ~1!! 2 ~1 2 op ~1!!! 5 op ~1!; @t0 T #
T
2 ( yt21 ( yt21 «t @tT #11 @t T #11
~ii! ~ b2T 2 b1 !
0
T
5
Op ~g~T !!Op ~!T !
!g~T !Op ~T !
2 ( yt21 @tT #11
5 Op
S! D g~T ! T
5 op ~1!; T
~iii!
2 ( yt21 @t T #11 0
p
&
T
(
@tT #11
1,
2 yt21
we have RSST ~t! 2 RSST ~t0 ! @t0 T #
5 22~ b2T 2 b1 ! 52
(
@t0 T #
@tT #11
yt21 «t 1 ~ b2T 2 b1 ! 2
@6 y6 g~T !#21
2
!g~T !
(0
yk02t21 «k02t 1
SS D
n 22s 2 B1
6y6
1 2 b12
2
1
1
6y6
D
2 yt21 1 op ~1!
@6 y6 g~T !#21
g~T !
2 ~1 2 b12 !
(
@tT #11
,
where B1~{! is a Brownian motion defined on R 1 +
(0
yk202t21 1 op ~1!
STRUCTURAL CHANGE IN AR(1) MODELS
127
Similarly, for t 5 t0 1 y@ g~T !0T # and y . 0, from ~B+4! in Appendix B and the fact 2 that @y0~1 2 b2T !# r @y0~1 2 b12 !# as T r `, we have RSST ~t! 2 RSST ~t0 ! @tT #
5 2~ b2T 2 b1 !
@tT #
2 yt21 1 op ~1! ( yt21 «t 1 ~ b2T 2 b1 ! 2 @t ( @t T #11 T #11 0
52
0
@yg~T !#21
2
!g~T !
(0
~2yk01t «k01t11 ! 1
SS D
n 22s 2 B2
y
1 2 b12
2
y 2~1 2 b12 !
D
@yg~T !#21
1
(0
g~T !
yk201t 1 op ~1!
,
where B2 ~{! is another Brownian motion defined on R 1 independent of B1~{!+ Define r 5 y0~1 2 b12 ! and apply the continuous mapping theorem for argmax functionals ~see Kim and Pollard, 1990!+ We have T ~ b2T 2 b1 ! 2 1 2 b12
~ t[ T 2 t0 ! 5 r[ 5 Arg min $RSST ~t! 2 RSST ~t0 !% r
d
&
H S S D
Arg min 22s 2 B * ~r! 2 r
6r6
DJ
1
5 Arg max B * ~r! 2 r
1 2
2
6r6 ,
where B * ~r! is a two-sided Brownian motion on R defined to be B * ~r! 5 B1~2r! for r # 0 and B * ~r! 5 B2 ~r! for r . 0+ This proves Theorem 2+ n
APPENDIX E: PROOF OF LEMMA 2 Conditions ~4a! and ~4d! are special cases of ~3a!+ Conditions ~4e! and ~4f ! are identical to ~3d! and ~3e!, respectively+ ~4b!
1 T
T
T
1
2 ~ yt2 2 yt21 2 «t2 ! ( yt21 «t 5 2T @t ( @t T #11 T #11 0
0
5 5
T 2 k0 2T 1 2 t0 2
1
T
yT2 T 2 k0
SS ! 2
yk20
2
T 2 k0
yk 0 T 2 k0 yk20
T 2 k0
1
2
2
( «t2 @t T #11 0
T 2 k0
2
T
1
!T 2 k
0 T
1
( «j j5k 11 0
«t2 T 2 k 0 @t ( T #11 0
D
+
D
2
128
TERENCE TAI-LEUNG CHONG
Because yk 0
p
!T 2 k
&
0
T
1
and
!T 2 k
0
( «j n sB~1!, j5k 11
0
0
by the continuous mapping theorem, we have
S!
yk 0 T 2 k0
1
T
1
!T 2 k
( «j j5k 11
0
0
D
2
n s 2 B 2 ~1!+
Therefore 1 T
T
( yt21 «t n @t T #11
~1 2 t0 !s 2 2
0
~4c!
1 T
2
T
2 5 ( yt21 @t T #11
~B 2 ~1! 2 1! 5 Op ~1!+
~T 2 k 0 ! 2 T
0
2
S
T
0
T
5 ~1 2 t0 ! 2
yk 0
(
@t0 T #11
n ~1 2 t0 ! 2 s 2
E
«j
011
S
op ~1! 1
t2k 022
1
!
D
2
t21
1
( !T 2 k 1 !T 2 k j5k( @t T #11 0 0 T 2 k0
(
D
1 T 2 k0
2
«k01j11
j50
1 T 2 k0
1
B 2 ~s! ds 5 Op ~1!,
0
where s 5 ~t 2 k 0 2 2!0~T 2 k 0 !+ To prove ~4g!, we use the triangle inequality, ~4b!–~4e!, and the facts that @t0 T #
sup st#t#t0
@t0 T #
2 # ( ( yt21 @tT #11 1
2 yt21 +
Then we have
sup 6 bZ 2 ~t! 2 16 5 sup
st#t#t0
st#t#t0
#
*
@t0 T #
~ b1 2 1!
T
(
(
@tT #11
(
@t0 T #11
«t yt21
2 yt21
S
2 ( yt21 @tT #11 @t0 T #
6 b1 2 16
1
5
@tT #11
T
«t yt21 1
T
1
@t0 T #11
(
@t0 T # 2 yt21 1
*
(1
T
( «t yt21 @t T #11 0
Op ~T ! 1 op ~T ! 1 Op ~T ! Op ~T 2 !
*(
@t0 T #
2 yt21 1 sup
5 Op
st#t#t0
*D
SD 1
T
+
@tT #11
«t yt21
*
*
STRUCTURAL CHANGE IN AR(1) MODELS
129
~4h! By the triangle inequality and ~4g!, sup 6 bZ 2 ~t! 2 b1 6 # sup 6 bZ 2 ~t! 2 16 1 61 2 b1 6 5 Op ~1!+
st#t#t0
st#t#t0
Conditions ~4i! and ~4j! can be proved by using an argument similar to that in ~4b!+ ~4k! For t [ JT1 ,
6 bZ 1 ~t! 2 16 5
*
@t0 T #
(1
~ b1 2 1!
@t0 T #
(1
2 yt21 1
@tT #
«t yt21 1
@t0 T # 2 yt21 1
@t0 T #
# 6 b1 2 16
(
@t0 T #11
2 yt21
* ** @t0 T #
2 yt21
1
@tT #
( @t T #11
«t yt21
@tT #
(1
(1
(
@t0 T #11
(1
0
1
@t0 T #
(1
2 yt21
«t yt21 2 yt21
*
@tT #
(
@t0 T #11
«t yt21
@tT #
*
(
2 yt21
T
*( *
@t0 T #11
@t0 T #
# 6 b1 2 16
@t0 T #
1
(1
T
2 yt21
T
1
@~t 2 t0 !T #
2
AT
*
T @~t 2 t0 !T # 2
@tT #
( @t T #11 0
yt21 «t T
AT 1
@t0 T #
(1
*
2 yt21
«t yt21
1
T
,
where AT 5
@~t2t0 !T #
(0
S!
1
D
2
yk01t @~t 2 t0 !T #
1
+
@~t 2 t0 !T #
The proof is completed by utilizing ~4a!, ~4e!, ~4i! and the following facts+ ~i! By the definition of JT1 , T0@~t 2 t0 !T # 2 5 o~1! for t [ JT1 + ~ii! By the definition of JT1 and by using the continuous mapping theorem, we have 1
AT n s2
E
5 Op ~1!+
1
B2
0
~4l! By the fact that sup t[JT1
* ( * *( * T2
T
@tT #11
2 yt21
#
T2
T
k11 N
2 yt21
5 Op ~1!
130
TERENCE TAI-LEUNG CHONG
and by ~4j!,
sup 6 bZ 2 ~t! 2 16 5 sup
t[JT1
t[JT1
*
T
( yt21 «t @tT #11 T
2 ( yt21 @tT #11
*
*( T
#
1 T
*( * T2
T
k11 N
2 yt21
@tT #11
sup t[JT1
yt21 «t
T
*
5 Op
SD 1
T
+
n
These prove Lemma 2+
APPENDIX F: DERIVATION OF EQUATIONS ~12!–~14! For 6 b1 6 , 1, b2 5 1, by ~B+1! in Appendix B, the triangle inequality, and the fact that sup6ab6 # sup6a6sup6b6, we have
*T 1 #* (« 2s * T
sup
1
st#t#t0
RSST ~t! 2 s 2 2
~1 2 b1 !~t0 2 t!s 2 1 1 b1
*
T
2 t
2
1
st#t#t0
*T ( y « * 1
1 sup 6 bZ 1 ~t! 2 b1 6 sup
st#t#t0
t21
1 sup
st#t#t0
*
st#t#t0
~ bZ 2 ~t! 2 b1 ! 2
1 2 sup 6 bZ 2 ~t! 2 16 st#t#t0
*
1 T
*
@t0 T #
1
(
T
@tT #11
1 T
S
2 2 ( yt21 @tT #11
T
( yt21 «t @t T #11 0
5 op ~1!+
sup 6T ~ bZ 2 ~t! 2 1!6
st#t#t0
yt21 «t
@t0 T #
T
*
T
1
t
1
1 2 sup 6 bZ 2 ~t! 2 b1 6 sup st#t#t0
@tT #
2 ( yt21 2 @t T #11
D
0
T2
*
~1 2 b1 !~t0 2 t!s 2 1 1 b1
*
STRUCTURAL CHANGE IN AR(1) MODELS
131
All six preceding terms above are op ~1! by Lemma 2+ To show that the fourth term is also op ~1!, we use the fact that ~1 2 b1 !0~1 1 b1 ! 5 ~1 2 b1 ! 20~1 2 b12 ! and the triangle inequality that sup6a 2 b6 5 sup6a 2 c 1 c 2 b6 # sup6a 2 c6 1 sup6c 2 b6, with @t0 T #
(
a 5 ~ bZ 2 ~t! 2 b1 ! 2 c 5 ~ bZ 2 ~t! 2 b1 ! 2
2 yt21
@tT #11
b5
,
T
~1 2 b1 !~t0 2 t!s 2
and
,
1 1 b1
~t0 2 t!s 2 +
1 2 b12
Conditions ~4e! and ~4h! in Lemma 2 imply that sup6a 2 c6 and sup6c 2 b6 are both op ~1!+ This derives ~12!+ For t [ JT1 , by ~B+3! in Appendix B and the triangle inequality, we have
*T
1
~1 2 b1 !t0 s 2
RSST ~t! 2 s 2 2 #
*
1 T
1 1 b1
T
(1 «t2 2 s 2
*
* *
1 2~ bZ 1 ~t! 2 b1 !
*
1
@t0 T #
T
(1
yt21 «t
@t0 T #
(1
1 ~ bZ 1 ~t! 2 b1 ! 2 1 ~ bZ 1 ~t! 2 1! 2
2 yt21
2
T
~1 2 b1 !t0 s 2 1 1 b1
*
**
@tT #
(
1 2~ bZ 1 ~t! 2 1!
@t0 T #11
yt21 «t
T
*
2 1 * ~ bZ 2 ~t! 2 1! yt21 «t *+ ( yt21 T @tT( @t T #11 #11 @tT #
1 T
T
1
0
In the third term, we use the facts that ~1 2 b1 !0~1 1 b1 ! 5 @~1 2 b1 ! 2 #0~1 2 b12 ! and the triangle inequality that 6a 2 b6 5 6a 2 c 1 c 2 b6 # 6a 2 c6 1 6c 2 b6, with @t0 T #
a 5 ~ bZ 1 ~t! 2 b1 !
(1
2
2 yt21
b5
,
T
~1 2 b1 !t0 s 2
and
1 1 b1
c 5 ~ bZ 1 ~t! 2 b1 ! 2
t0 s 2 1 2 b12
Thus, the preceding equation is bounded by #
*
1 T
* * ( * (
T 1
1
1 ~ bZ 1 ~t! 2 b1 ! 2 1 1
1
( «t2 2 s 2 1 2
t0 s 2 12
b12
*
@t0 T #
T
yt21 «t 6 bZ 1 ~t! 2 b1 6
1
@t0 T #
2 yt21 2
T
1
t0 s 2 1 2 b12
*
6 bZ 1 ~t! 1 1 2 2b1 66 bZ 1 ~t! 2 16 1 26 bZ 1 ~t! 2 16
@~t 2 t0 !T # 2 T
1 6 bZ 2 ~t! 2 16
~ bZ 1 ~t! 2 1! 2
*T
1
@~t 2 t0 !T #
( yt21 «t * + @tT #11 T
@tT #
1 2
(
@t0 T #11
2 yt21
*
1
@tT #
T
@t0 T #11
(
yt21 «t
*
+
132
TERENCE TAI-LEUNG CHONG
All seven preceding terms above are op ~1! by Lemma 2 and by the definition of JT1+ This derives equation ~13!+ To derive ~14!, first note that for any positive constant c,
S
bZ 1 t0 1
1
D S D
c
1
5 uT
!T
2
b1 1
,c
(1 «t yt21 k0
2 (1 yt21
S S DD
1 1 2 uT
2
k0
1
2
,c
1
@k 01c!T #
(
«t yt21
k 011
11
k 01c! T
(
k 011
2 yt21
2
,
where k0
uT
2 (1 yt21
S D 1
2
,c 5
(1
+
@k 01c!T #
k0
2 yt21
(
1
2 yt21
k 011
By ~3b! and the fact that yk20 0T 1 T
@k 01c! T #
( k 11
2 yt21 5
0
1
@c! T #
T
(1
p
&
0, plus the invariance principle that
2 yt1k n s2 021
E
c
B32 ,
0
where B3 ~{! is a Brownian motion defined on R 1 , we have ~i! uT
S D 1
2
t0
,c n
t0 1 ~1 2 b12 !
E
;
c
B32
0
S
~ii! bZ 1 t0 1
c
!T
D
E E
n
c
t0 1 ~1 2 b12 !
B32
Additionally, we have sup c[R 11
~iv!
sup c[R 11
*bZ St 1 !T D 2 1* 5 o ~1!; c
2
1 T
*
0
@k 01c! T #
(
k 011
B32
0
0
~iii!
c
t0 b1 1 ~1 2 b12 !
p
*
yt21 «t 5 op ~1!;
+
STRUCTURAL CHANGE IN AR(1) MODELS ~v!
sup c[R 11
1 T
*
T
( yt21 «t @k 1c! T #11 0
*
133
5 Op ~1!+
From ~B+3! in Appendix B, ~4e! of Lemma 2, and ~i!–~v! given immediately preceding, we have 1 T
S
RSST t0 1
c
!T
D
5
1 T
SS !D D SS !D D ( SS !D D ( T
c
c
2 2 bZ 1 t0 1
SS SS
1 bZ 1 t0 1
2 bZ 2 t0 1
1 T
T
!T c
T
2 yt21
@k 01c! T #
1
( T
1
( yt21 «t @k 1c! T #11
T
SS SS !D D c
0
D ! D
n s2 1
S
1
2
T
(
T
E E
B32
B32
0
t0 1 ~1 2
b12 !
E
D
2
0
B32
s2
SE D
21
t0
B32
0
+ 11 2
E
b12
t0 s 2 1 2 b12
c
0
t0 s 2 ~1 2 b1 ! 2 c
2
2
c
0
c
2
1 T
k0
2 1 op ~1! (1 yt21
2 yt21 1 op ~1!
k 011
c
t0 1 ~1 2 b12 !
2 b1
@k 01c! T #
1
~1 2 b1 !~1 2 b12 !
t0 ~ b1 2 1!
5 s2 1
This derives equation ~14!+
T
c
bZ 1 t0 1
21
yt21
k 011
T
(1 «t2 1 op ~1! 1
yt21 «t
k 011
T
21
(1 yt21 «t
1
T
2
T
k0
@k 01c! T #
1
D D ! D D 21
1
k0
1
T
21
c
1 bZ 1 t0 1
1
2
2 b1
T
2 b1
T
1
1 bZ 1 t0 1
5
c
( «t2 2 2 bZ 1 t0 1
B32
134
TERENCE TAI-LEUNG CHONG
APPENDIX G: PROOF OF THEOREM 3 To show 6 t[ T 2 t0 6 5 Op ~10T !, it is sufficient to prove that A1–A6 defined in Appendix C are all op ~1! in the case where 6 b1 6 , 1 and b2 5 1+ Now T
A1 5
( yt21 «t k 11
5 op ~1!
0
T
(
2 yt21
k 011
by ~4b! and ~4c! of Lemma 2; k0
(
yt21 «t
m11
A 2 5 sup
5 op ~1!
k0
m[D1T
(
2 yt21
m11
by the uniform law of large numbers in Andrews ~1987, Theorem 1!;
A 3 5 sup m[D1T
3
*
T
2 ( yt21 m11
LT
k0
T
2 2 ( yt21 ( yt21 k 11 m11
m
T
0
1 * sup
S
1 sup m[D1T
k0
D S 2
( yt21 «t 1
2
k0
m[D1T
*
1
T2
*1 #
1
(
k 011
( yt21 «t 1
D
m
2 (1 yt21
2
k 011
1 Op
2
(
k 011
1
MT
«t yt21
2
T
2 yt21
m11
T
(
k 02MT
2 yt21
*
T
«t yt21
k0
(
2 yt21
S( D S( D T
1
1
T
2
m
2 (1 yt21
S S D S DD
5 Op
SD
2 yt21
m11
~op ~1! 1 Op ~1!! 5 Op
*2
S D 1
MT
5 op ~1!;
2
STRUCTURAL CHANGE IN AR(1) MODELS
135
k0
A4 5
(1 yt21 «t
5 op ~1!
k0
(1
2 yt21
by ~4a! and ~4e! of Lemma 2; m
A 5 5 sup
ym2 2 yk20 2 5 sup
0
m
m[D2T
(
k 011
A 6 5 sup m[D2T
3
m
( yt21 «t k 11
*
2 ( yt21 k 11
2
m
m
(
k 011
1 *
LT
k0
2 2 yt21 ( yt21
S
1
k0
1 sup m[D2T
1
T
m
T
D S 2
k0
*
*1
1
#
k0
D
2
( yt21 «t 1
m
1
2 (1 yt21
m
2 (1 yt21
2 (1 yt21
(
k 011
1 Op
2
«t yt21
2
T
m11
T
2 ( yt21 k 11
2 ( yt21 m11
0
1
MT2
2 yt21
2
*
T
«t yt21
1 k 01MT k 011
S( D S( D 2
T
S S D S DD 1
SD
2
( yt21 «t
m[D2T
5 op ~1!+
0
2 (1 yt21
sup
5 Op
0
m
m[D2T
2 yt21
( «t2 k 11
*2
~Op ~1! 1 Op ~1!! 5 op ~1!+
~To prove that A 5 5 op ~1!, explicit formulae would be needed+ I am not able to provide a general proof because of the unknown asymptotic properties of 2 supm[D2T ~ ( km011 yt21 «t !0~ ( km011 yt21 ! under nonstationarity of yt + No study on the uniform convergence under this kind of nonstationarity has been done in the literature+ However, the referees and I believe that the result is true and intuitive+! Because A 1 –A 6 are op ~1!, thus t[ T is T-consistent+ To find the limiting distribution of bZ 1~ t[ T !, note that t[ T 2 t0 5 Op ~10T ! and
!T ~ bZ 1~ t[ T ! 2 bZ 1~t0 !! 5 !T
1
@ t[ T T #
(1
@t0 T #
yt yt21 2
@ t[ T T #
(1
2 yt21
(1
yt yt21
@t0 T #
(1
2 yt21
2
136
TERENCE TAI-LEUNG CHONG
5 1$ t[ T # t0 %!T
1
@t0 T #
@t0 T #
(
@ t[ T T #11
1 1$ t[ T . t0 %!T
2 yt21
(1
@ t[ T T #
(1
1
(
@ t[ T T #11
2
@t0 T #
(1
2 yt21
@ t[ T T #
2
@t0 T #
«t yt21
@ t[ T T #
(1
2 yt21
1
(1
@t0 T #
(1
2 yt21
(
@t0 T #11
2
«t yt21
@ t[ T T #
(1
2 yt21
2
2 yt21
@ t[ T T #
(1
@t0 T #11
2 yt21
@ t[ T T #
1 ~ b2 2 b1 !
(
«t yt21
0
@ t[ T T #
2 yt21
@ t[ T T #
@t0 T #
2 ( yt21 (1 @t T #11
«t yt21
2 yt21
S S D S ! D S DD ! S S D S D S D S DD !
5 1$ t[ T # t0 %!T Op
1
T
1 1$ t[ T . t0 % T Op
1
Op
1
T
2 Op
T
Op
1
1
T
1 Op
T
1
T
1 Op
1
T
5 op ~1!+ Thus, bZ 1~ t[ T ! and bZ 1~t0 ! have the same asymptotic distribution+ @t0 T # Because $«t yt21 ,It % t51 is a martingale difference sequence, with E~«t yt21 6It21 ! 5 0, @t0 T #
(1
p
E~~«t yt21 ! 2 6It21 !
&
s4 1 2 b12
, `+
Applying the central limit theorem for martingale difference sequences and the fact that 1 T
@t0 T #
(1
2 yt21
p
&
t0 s 2 1 2 b12
,
we have @t0 T #
d
!T ~ bZ 1~ t[ T ! 2 b1 ! 5 !T ~ bZ 1~t0 ! 2 b1 ! 5
(1
«t yt21 0!T d
@t0 T #
(1
2 yt21 0T
&
S
N 0,
1 2 b12 t0
D
+
STRUCTURAL CHANGE IN AR(1) MODELS
137
To find the limiting distribution of bZ 2 ~ t[ T !, note that t[ T 2 t0 5 Op ~10T ! and T ~ bZ 2 ~ t[ T ! 2 bZ 2 ~t0 !!
5T
1
T
( yt yt21 @ t[ T #11 T
T
(
@ t[ T T #11
T
( yt yt21 @t T #11
2
0
T
(
2 yt21
1
2 yt21
@t0 T #11 @t0 T #
5 1$ t[ T # t0 %T 2
2
T
0
T
T
(
@ t[ T T #11
(
2 yt21
@t0 T #11
(
1 ~ b1 2 b2 !
@ t[ T T #11
(
1
2 yt21
T
@ t[ T T #11 @ t[ T T #
(
1
@ t[ T T #11
2 yt21
0
T
(
@ t[ T T #11
(
2 yt21
@t0 T #11
2 yt21
2 @ t[ T T #
T
T
T
@ t[ T T #11
2 ( yt21 ( «t yt21 @t T #11 @t T #11 0
«t yt21
(
2 yt21
@t0 T #
1 1$ t[ T . t0 %T
@t0 T #
T
2 ( yt21 ( «t yt21 @ t[ T #11 @t T #11
(
2
@t0 T #11 T
(
2 yt21
«t yt21
@ t[ T T #11
2 yt21
S S D S D S D S DD S S D S D S DD
5 1$ t[ T # t0 %T Op
1
T
1 1$ t[ T . t0 %T Op
Op
2
1
T
2
1
1 Op
T
1
Op
1
T
2 Op
T
1 Op
2
+
1
T2
1
T2
5 op ~1!+ Thus, bZ 2 ~ t[ T ! and bZ 2 ~t0 ! have the same asymptotic distribution+ Applying ~4b! and ~4c! in Lemma 2, we get 1 d
T ~ bZ 2 ~ t[ T ! 2 1! 5 T ~ bZ 2 ~t0 ! 2 b2 ! 5
T
T
( «t yt21 @t T #11
1 T2
0
T
(
@t0 T #11
2 yt21
n
B 2 ~1! 2 1 2~1 2 t0 !
E
+
1
B
2
0
These derive the limiting distribution of bZ 1~ t[ T ! and bZ 2 ~ t[ T ! for fixed magnitude of break+ To derive the limiting distribution of t[ T for shrinking shift, we fix b2 at one and let b1T 5 1 2 @10g~T !# , where g~T ! . 0, with g~T ! r ` and @ g~T !0T # r 0 as T r `+ Let y be a finite constant, and B1~{! and B2 ~{! be defined as in Section 2+ For t 5 t0 1 y@ g~T !0T # and y # 0, we have L T ~t! 5 op ~1! where L T ~t! is defined in Appendix B, and
138
TERENCE TAI-LEUNG CHONG
yk02t21
!g~T !
5
FS
1
!g~T ! k 02t22
( i50
5
ns
SS
12
E
`
D
k 02t22
1
12
g~T !
k 02t22
D D
g~T !
S
12
i50
g~T ! i0g~T !
1
(
y0 1
«k02t212i
!g~T !
d
exp~2s! dB1 ~s! 5 sBa
0
SD
1 g~T !
D
i
«k02t212i
G
1 op ~1!
1
2
,
where Ba ~ _12 ! is a Brownian motion defined on R 1 + Because yk02t21
!g~T !
n sBa
SD 1
2
and
@6 y6 g~T !#21
1
!g~T !
(0
«k02t n sB1 ~6y6!,
we have @6 y6 g~T !#21
1
(0
!g~T !
yk02t21 «k02t
!g~T !
n s 2 Ba
@6 y6 g~T !#21
1
(0
g ~T ! 2
SD SD 1
2
B1 ~6y6!
and
1
yk202t21 n 6y6s 2 Ba2
2
+
Further, because @t0 T #
~1 2 b1T !
T
2 ( yt21 ( yt21 «t @tT #11 @t T #11 0
T
(
@tT #11
5
Op ~g 2 ~T !!Op ~T !
2 yt21
g~T !Op ~T ! 2
5 Op
S D g~T ! T
5 op ~1!
and T
2 ( yt21 @t T #11
p
0
&
T
(
@tT #11
1,
2 yt21
equation ~B+2! in Appendix B becomes RSST ~t! 2 RSST ~t0 ! @t0 T #
5 22~1 2 b1T ! 52
2
(
@tT #11
@t0 T #
yt21 «t 1 ~1 2 b1T ! 2
@6 y6 g~T !#21
g~T !
n 22s 2 Ba
(0
SD 1
2
yk02t21 «k02t 1 B1 ~6y6! 1 6y6s 2 Ba2
g ~T ! 1
2
2 yt21 1 op ~1!
@6 y6 g~T !#21
1 2
SD
(
@tT #11
+
(0
yk202t21 1 op ~1!
STRUCTURAL CHANGE IN AR(1) MODELS
139
Similarly, for t 5 t0 1 y@ g~T !0T # and y . 0, from ~B+4! in Appendix B, we have L T ~t! 5 op ~1!, @t0 T #
@tT #
(1
~1 2 b1T !
(
yt21 «t
@t0 T #11
2 yt21
Op ~T !Op ~g 2 ~T !!
5
@tT #
g~T !Op ~T 2 !
2 (1 yt21
yk 0
!g~T !
S
ns
E
g~T !
`
D
k0
1
5 12
5 Op
S D g~T ! T
k 021
y0
!g~T !
( i50
1
d
@yg~T !#21
1
(0
g~T !
S! D
2
yk01t
g~T !
5
@yg~T !#21
1
(0
g~T !
S!
@yg~T !#21
yk01t
(0 !g~T ! «k 1t11 n s 2
!g~T !
0
B2 ~s! 1 Ba
y
B2 ~s! 1 Ba
0
i
«k02i
g~T !
S DD S DD 1
g~T !
1
,
2
yk 0
0
ES
ES
1
( «k 1t2i 1 !g~T ! g~T ! i50
0
1
12
t21
1
y
n s2
S D! SD
exp~2s! dB1 ~s! 5 sBa
0
5 op ~1!,
D
2
2
2
1
2
ds,
dB2 ~s!+
Hence, equation ~B+4! in Appendix B becomes RSST ~t! 2 RSST ~t0 !
5
@tT #
@tT #
0
0
2 yt21 1 op ~1! ( yt21 «t 1 ~1 2 b1T ! 2 @t ( @t T #11 T #11
5 2~1 2 b1T !
@yg~T !#21
2
(0
g~T !
yk01t «k01t11 1
ES y
n 2s 2
B2 ~s! 1 Ba
0
5 22s 2 Ba2
F
3 2
S DD 1
2
1
@yg~T !#21
(0
g ~T ! 2
yk201t 1 op ~1!
ES y
dB2 ~s! 1 s 2
B2 ~s! 1 B
0
SD
S DD 1
2
2
ds
1
2
E
B2 ~y! 2 1 Ba 2
SD
y
0
E
B2 ~s! dB2 ~s! 2 1 Ba2 2
SD
y
0
S
D
G
1 B2 ~s! B2 ~s! 11 ds 2 y + 1 1 2 2Ba Ba 2 2
SD
SD
140
TERENCE TAI-LEUNG CHONG
By the continuous mapping theorem for argmax functionals, we have ~1 2 b1T !T ~ t[ T 2 t0 ! 5 y[ 5 Arg max $RSST ~t! 2 RSST ~t0 !% y[R
d
&
H
Arg max 22s 2 Ba2 y[R
5 Arg max y[R
SD 1
2
S S D DJ 1 C * ~y! 2 6y6 1 2 Ba 2
S SD D
1 C * ~y! 2 6y6 + 1 2 Ba 2
n
This proves Theorem 3+
APPENDIX H: PROOF OF LEMMA 3 ~5a! By the fact that ~ yk0 0!T ! n sB~t0 ! and ~ y02 0T ! 5 op ~1!, we have 1 T
@t0 T #
(1
yt21 «t 5
n
~5b!
1 2T s2 2
@t0 T #
(1
1
2 ~ yt2 2 yt21 2 «t2 ! 5
2T
S
@t0 T #
yk20 2 y02 2
(1
«t2
D
~B 2 ~t0 ! 2 t0 ! 5 Op ~1!+
T
1
yt21 «t !T @t ( T #11 0
5
T
1
!T t5k(11 0
5
5
1
!T 1
SS
S
k0
y0 1 ( «j
k0
( !T j51
j51
S
1 op ~1!+
«j
D
t21
b2t2k021 yk0 1
D(
(
b2t2i21 «i «t
T
k 011
b2t2k021 «t 1
T
k 011
(
i5k 011
b2t2k021 «t
D
1
1
T
t21
0
0
( ( b2t2i21 «i «t t5k 11 i5k 11 T
!T t5k(11 0
S
«t
t21
(
i5k 011
b2t2i21 «i
D
D
STRUCTURAL CHANGE IN AR(1) MODELS
141
T k0 By Assumption ~A2! that «’s are i+i+d+, $«j ( t5k b2t2k021 «t , Ij % j51 and 011 t21 t2i21 T $«t ( i5k011 b2 «i , It %t5k011 will both be martingale difference sequences with
S
T
( b2t2k 21 «t 6Ij21 t5k 11
E «j 1 T
0
0
k0
(E j51
S
(
T
T
( E t5k 11 0
~ j 5 1,2, + + + , k 0 !,
D D
b2t2i21 «i 6It21
50
0
0
i5k 011
50
( b2t2k 21 «t t5k 11
T
«j
t21
E «t 1
SS
D
SS
«t
D
2
p
6Ij21
t21
( b2t2i21 «i i5k 11 0
t0 s 4
&
, `,
1 2 b22
~t 5 k 0 1 1, k 0 1 2, + + + , T !,
D D 2
p
6It21
~1 2 t0 !s 4
&
+
1 2 b22
By the central limit theorem for martingale difference sequences and by the independence of the two martingale difference sequences given previously, we have
S
T
1
yt21 «t !T @t ( T #11
d
N 0,
&
0
~5c! ~5d!
@t0 T #
1 T 1 T
(1
2
@t0 T #
1 2 b22
1
y0
@tT #
1
2 5 ( yt21 T t5k(11 @t T #11 0
S! D 0
5
1
1 T
yk 0
( «i i51
1
!T
S
T
T
;
2
2
1
n s2
T
E
t0
B 2 5 Op ~1!;
0
t21
( b2t2i21 «i i5k 11
b2t2k021 yk0 1
0
@tT #
2
yk 0
12
D
t21
(1 !T
2 yt21 5
@tT #
s4
(
t5k 011 @tT #
( t5k 11
S
0
@tT #
( t5k 11 0
D
2
b22~t2k021!
S
t21
b2t2k021
(
i5k 011
t21
( b2t2i21 «i i5k 11 0
b2t2i21 «i
D
D
2
+
The first term converges weakly to @s 2 B 2 ~t0 !#0~1 2 b22 !+ Because 6 yk0 0!T 6 5 Op ~1! and ~10!T !sup t.k0 6«t 6 5 op ~1!, the second term is bounded by #2 #2 #2
*! * * ( * ! *S ( * * yk 0
@tT #
sup
T
t[JT1
yk 0
`
T
k 011
k 011
b2t2k021
b2t2k021 1 2 b2
* !T *S ~1 2 6 b 6! yk 0
1
2
2
1
1 2 b2t2k021 1 2 b2 `
1
( k 11 0
*
*!
1
b22~t2k021! 1 2 b2
T
sup 6«t 6 t.k 0
*D !
1 ~1 2 6 b2 6!~1 2 b22 !
1
D!
T
1 T
sup 6«t 6 t.k 0
sup 6«t 6 5 op ~1!+ t.k 0
142
TERENCE TAI-LEUNG CHONG
Last, because 6 b2 6 , 1 and «i are i+i+d+ with finite fourth moment, we have lim Tr`
1
@tT #
T
t5k 011
(
S(
Var
SS
D DD S (
~t 2 t0 !s 2
2
t21
E
i5k 011
5
b2t2i21 «i
t21
( b2t2i21 «i i5k 11 0
,
1 2 b22
2
D
4
t21
,E
i5k 011
b2t2i21 «i
, `+
By the uniform weak law of large numbers for dependent and heterogeneous processes, sup t[JT1
*
1 T
@tT #
( t5k 11 0
S
t21
( b2t2i21 «i i5k 11 0
D
~t 2 t0 !s 2
2
2
1 2 b22
*
5 op ~1!+
Thus, 1 T
@tT #
2 n ( yt21 @t T #11
t 2 t0 1 B 2 ~t0 !
~5e! ~5f !
sup t[JT2
sup t[JT2
*T
1
*
1 T
s 2+
1 2 b22
0
sup * ( (1 yt21 «t * 5 t[J * 1 !T !T
@tT #
@tT #
d
( @tT #11
*
t[JT2
d
0
yt21 «t T
@tT #11
t0
sup s 2
&
t[JT2
*( ! !* * E * @t0 T #
yt21 «t 5 sup
t[JT2
t
* *
T
BdB 5 Op ~1!;
@tT #
~5g!
sup 6 bZ 1 ~t! 2 16 5 sup
t[JT2
5
~5h!
1 T
@tT #
(1
#
2 yt21
t[JT2
*
T
yt21 «t
T2
1
(1
SD
t[JT2
2 yt21
1
T
; @t0 T #
T
2 1 ( «t yt21 1 ( «t yt21 ( yt21 @t T #11 @tT #11 @t T #11 0
@t0 T #
(
@tT #11
T
2 yt21 1
# sup
2 ( yt21 @t T #11 0
@t0 T #
t[JT2
( @tT #11 T
2 yt21
( «t yt21 @t T #11 0
T
(
@t0 T #11
2 yt21
*
+
1 sup
(
@t0 T #11
T
*
T
0
6 b2 2 16
1
1
sup
@ stT #
T
~ b2 2 1!
*( * @tT #
Op ~1!Op ~1! 5 Op
sup 6 bZ 2 ~t! 2 16 5 sup
t[JT2
(1 yt21 «t
t[JT2
BdB 5 Op ~1!;
sup s 2
&
T 2
@t0 T #
* E * t
yt21 «t
t[JT2
*
2 yt21
@t0 T #
(
@tT #11
«t yt21
@t0 T #
(
@tT #11
2 yt21
*
*
STRUCTURAL CHANGE IN AR(1) MODELS
143
Using the fact that @t0 T #
S D
2
yt21
( @tT #11 ! T
1 @~t0 2 t!T #
n s2
E
t0
t
B 2 5 Op ~1!,
the definition of JT2 , ~5d!, and ~5f !, the first term is bounded by T
6 b2 2 16
sup
k0 2 k
t[JT2
2 ( yt21 @t T #11
T ~k 0 2 k!
sup t[JT2
0
@t0 T #
(
@tT #11
T
2 yt21
5 Op ~T 2102 !Op ~1!Op ~1! 5 op ~1!+
The second term is bounded by
*( * @t0 T #
1
sup t[JT2
k0 2 k
*( * T ~k 0 2 k!
sup t[JT2
@t0 T #
@tT #11
2 yt21
sup t[JT2
@tT #11
«t yt21
T
5 Op ~T 2102 !Op ~1!Op ~1! 5 op ~1!+ The last term is op ~1! by ~5b! and ~5d! of this lemma+ Thus, sup t[JT2 6 bZ 2 ~t! 2 16 5 op ~1!+
~5i!
sup t[JT1
*
1 T
@tT #
( yt21 «t @t T #11 0
*
5 sup
t[JT1
# sup
t[JT1
* *
1 sup
1 T
@tT #
( @t T #11 0
1
@tT #
T
@t0 T #11
t[JT1
*
S
(
1 T
b2t2k021 yk0 «t
@tT #
(
t21
b2t2k021 yk0 1
t21
(
@t0 T #11 i5k 011
(
i5k 011
D*
b2t2i21 «i «t
* *
b2t2i21 «i «t +
Because 6 yk0 0!T 6 5 Op ~1! and ~10!T ! sup t.k0 6«t 6 5 op ~1!, the first term is bounded by @10~1 2 b2 !#6 yk0 0 k 0 6~10!T !sup t.k0 6«t 6 5 op ~1!+ t21 t21 Further, because E~ ( i5k b2t2i21 «i «t ! 5 0 and Var~ ( i5k b2t2i21 «i «t ! # 011 011 2 4 s 0~1 2 b2 ! , `, by the uniform weak law of large numbers for dependent and heterogeneous processes ~Pötscher and Prucha, 1989; Andrews, 1987!, the second term converges to zero+ This shows ~5i!+
!
144
TERENCE TAI-LEUNG CHONG
To show ~5j!, by ~5b!, ~5i!, and the triangle inequality, we have sup t[JT1
*
~5k!
1 T
T
( yt21 «t @tT #11
* * #
1 T
T
( yt21 «t @t T #11 0
*
1 sup
t[JT1
*
1
@tT #
T
@t0 T #11
*
(
yt21 «t 5 op ~1!+
sup 6 bZ 1 ~t! 2 16
t[JT1
5 sup
t[JT1
*
(
@t0 T #11
2 yt21 1
(
@t0 T #11
«t yt21
2 (1 yt21
*
0
*
@t0 T #
2 1 ( ( yt21 @t T #11 1
«t yt21 1 sup
t[JT1
*
*
@tT #
(
@t0 T #11
«t yt21
*
@t0 T #
(1
5
(1
@tT #
«t yt21 1
@tT #
T
6 b2 2 16 #
@t0 T #
@tT #
~ b2 2 1!
Op ~T ! 1 Op ~T ! 1 op ~T ! Op ~T ! 2
5 Op
2 yt21
SD 1
T
by ~5a!, ~5c!, ~5d!, and ~5i!+
~5l!
sup 6 bZ 2 ~t! 2 b2 6 5 sup
t[JT1
t[JT1
*
T
( «t yt21 @tT #11 T
(
@tT #11
2 yt21
*
#
T
sup
T
(
@ tT S #11
2 yt21
t[JT1
*
T
( «t yt21 @tT #11 T
*
5 Op ~1!op ~1! 5 op ~1! by ~5d! and ~5j!+ These prove Lemma 3+
n
APPENDIX I: DERIVATION OF EQUATIONS ~16!–~18! To show that ~10T !RSST ~t! converges uniformly to a random flat line above s 2 for t [ JT2 , it is sufficient to show that ~i! ~10T !RSST ~t! converges pointwise to a random value above s 2 for all t [ JT2 and p ~ii! sup tb , ta[JT2 6~10T !RSST ~tb ! 2 ~10T !RSST ~ta !6 & 0 for any tb , ta [ JT2 +
STRUCTURAL CHANGE IN AR(1) MODELS
145
To show ~i!, use Lemma 3 and ~B+1! in Appendix B+ For t [ JT2 , b1 5 1, and 6 b2 6 , 1, we have
1 T
RSST ~t! 5
T
@t0 T #
@tT #
T
(1 «t2
(1 yt21 «t
2 ~ bZ 1 ~t! 2 1!
T
(
yt21 «t
@tT #11
2 2~ bZ 2 ~t! 2 1!
T
@t0 T #
1
1
T
T
2 ( yt21 @tT #11
~T ~ bZ 2 ~t! 2 1!! 2
2
T2
2
!T
~ bZ 2 ~t! 2 b2 !
T
1 ~ bZ 2 ~t! 2 1! 2
2 ( yt21 @t T #11 0
T
( yt21 «t @t T #11 0
!T T
1 2~ bZ 2 ~t! 2 1!~1 2 b2 !
2 ( yt21 @t T #11 0
T
T
1 ~1 2 b2 ! 2
2 ( yt21 @t T #11 0
T
T
5
(1 «t2 T
n s2 1
T
2 op ~1! 2 op ~1! 1 op ~1! 2 op ~1! 1 op ~1! 1 op ~1! 1 ~1 2 b2 ! 2 ~1 2 b2 !~1 2 t0 1 B 2 ~t0 !! 1 1 b2
2 ( yt21 @t T #11 0
T
s 2 . s 2+
To show ~ii!, use ~B+1! in Appendix B, the triangle inequality, and Lemma 3+ We have
5
* T RSS ~t ! 2 T RSS ~t !* 1
sup tb , ta[JT2
1
T
sup tb , ta[JT2
*
b
T
a
@tb T #
2 ~ bZ 1 ~tb ! 2 1!
(1
@t0 T #
(
yt21 «t 2 2~ bZ 2 ~tb ! 2 1!
T
@tb T #11
T
@t0 T #
1 ~ bZ 2 ~tb ! 2 1! 2
2 ( yt21 @t T #11 b
T
yt21 «t
T
1 2~ bZ 2 ~ta ! 2 bZ 2 ~tb !!
( yt21 «t @t T #11 0
T @ta T #
T
1 ~~ bZ 2 ~tb ! 2 b2 ! 2 2 ~ bZ 2 ~ta ! 2 b2 ! 2 ! 2
2 ( yt21 @t T #11 0
T
@t0 T #
1 2~ bZ 2 ~ta ! 2 1!
( yt21 «t @t T #11 a
T
1 ~ bZ 1 ~ta ! 2 1!
@t0 T #
(
2 ~ bZ 2 ~ta ! 2 1!
2
@ta T #11
T
2 yt21
*
(1
yt21 «t T
146
TERENCE TAI-LEUNG CHONG
*( * @tb T #
yt21 «t
# sup 6 bZ 1 ~tb ! 2 16 sup tb[JT2
1
tb[JT2
T
*(
@t0 T #
@tb T #11
1 2 sup 6 bZ 2 ~tb ! 2 16 sup tb[JT2
tb[JT2
yt21 «t
T
*
@t0 T #
2 ( yt21 @t T #11
1 sup ~ bZ 2 ~tb ! 2 1! 2 sup tb[JT2
12
b
T
tb[JT2
*
T
( yt21 «t @t T #11 0
T
*
sup 6 bZ 2 ~ta ! 2 bZ 2 ~tb !6
tb , ta[JT2
T
2 ( yt21 @t T #11
1
sup 6~ bZ 2 ~tb ! 2 b2 ! 2 2 ~ bZ 2 ~ta ! 2 b2 ! 2 6
0
T
tb , ta[JT2
1 sup 6 bZ 1 ~ta ! 2 16 sup ta[JT2
ta[JT2
*
@ta T #
(1
yt21 «t T
*
*(
@t0 T #
1 2 sup 6 bZ 2 ~ta ! 2 16 sup ta[JT2
ta[JT2
@ta T #11
T
yt21 «t
*
@t0 T #
(
1 sup ~ bZ 2 ~ta ! 2 1! ta[JT2
2
@ta T #11
sup ta[JT2
2 yt21 p
T
&
0+
This derives equation ~16!+ For t 5 t0 , 1 T
RSST ~t0 !
5
T
S( D @t0 T #
T
(1 «t2
2
T
yt21 «t
2
1
@t0 T #
T
(1
2 2~ bZ 2 ~t0 ! 2 b2 !
2 yt21
( yt21 «t @t T #11 0
T
T
1 ~ bZ 2 ~t0 ! 2 b2 ! 2
5
1 T
T
2 ( yt21 @t T #11 0
T
(1 «t2 1 op ~1! 1 op ~1! 1 op ~1!
p
&
s 2+
This derives equation ~17!+ To show that ~10T !RSST ~t! converges uniformly to a random linear function above s 2 for t [ JT1 , it is sufficient to show that
STRUCTURAL CHANGE IN AR(1) MODELS
147
~i! ~10T !RSST ~t! converges pointwise to a random linear function above s 2 for all t [ JT1 and p ~ii! sup t[JT1 6~10T !RSST ~t! 2 ~10T !RSST ~t 1 ~10T !!6 & 0 for all t [ JT1 + To show ~i!, by Lemma 3 and ~B+3! in Appendix B, we have @t0 T #
T
1 T
(1 «t2
RSST ~t! 5
T
(1
2 2~ bZ 1 ~t! 2 1!
@t0 T #
(1
yt21 «t 1 ~ bZ 1 ~t! 2 1!
T
2
T
@tT #
@tT #
(
2 2~ bZ 1 ~t! 2 b2 !
2 yt21
@t0 T #11
(
yt21 «t
@t0 T #11
1 ~ bZ 1 ~t! 2 b2 ! 2
T
2 yt21
T
T
(
2 ~ bZ 2 ~t! 2 b2 !
yt21 «t
@tT #11
T @tT #
T
(1 «t2
5
T
n s2 1
(
2 op ~1! 1 op ~1! 2 op ~1! 1 ~ bZ 1 ~t! 2 b2 ! ~1 2 b2 ! 2 ~t 2 t0 1 B 2 ~t0 !!
2
2 yt21
@t0 T #11
2 op ~1!
T
s 2 $ s 2+
1 2 b22
To show ~ii!, use ~B+3! in Appendix B, the triangle inequality, and Lemma 3+ We have
* T RSS ~t! 2 T RSS St 1 T D* 1
sup t[JT1
1
1
T
T
*S S
5 sup 2 bZ 1 t 1 t[JT1
@t0 T #
1 T
D
2 bZ 1 ~t!
(1
D
yt21 «t T @t0 T #
S
S S D DD
1 ~ bZ 1 ~t! 2 1! 2 2 bZ 1 t 1
1
21
T
2
(1
2 yt21
T
@tT #
SS D
1 2 bZ 1 t 1
1
T
2 bZ 1 ~t!
D
(
yt21 «t
@t0 T #11
SS D D
1 2 bZ 1 t 1
T
1
T
2 b2
y@tT # «@tT #11
@tT #
S
S S D DD
1 ~ bZ 1 ~t! 2 b2 ! 2 bZ 1 t 1 2
1
T
2 b2
2
(
@t0 T #11
2 yt21
T T
SS D D
2 bZ 1 t 1
1
T
2 b2
2 y2 @tT #
SS D D
2 bZ 2 t 1
1
T
2 b2
T
SS D
1 bZ 2 t 1
y@tT # «@tT #11 T
*
1
T
2 bZ 2 ~t!
D
( yt21 «t @tT #11 T
T
148
TERENCE TAI-LEUNG CHONG
*( * @t0 T #
yt21 «t
#2
1
T
*bZ St 1 T D 2 bZ ~t!* 1
sup t[JT1
1
1
* ( *SS @t0 T #
1
1
T
2 yt21
1
T
1 2 sup
t[JT1
1 2 sup
t[JT1
T sup 6 bZ 1 ~t! 2 16
2
t[JT1
D S 2
* S D * * * S D * * bZ 1 t 1
bZ 1 t 1
1
2 bZ 1 ~t! sup
T
1 T sup
1
(
yt21 «t
@t0 T #11
T
y@tT # «@tT #11
2 b2 sup
*
@tT #
t[JT1
T
t[JT1
T
t[JT1
S D *D D
bZ 1 t 1
1
T
21
2
*
* @tT #
1 sup
T t[J1
S S D D* SS D D * ~ bZ ~t! 2 b ! 1
1 sup bZ 1 t 1 t[JT1
1
T
2
2 bZ 1 t 1
2 b2
2
sup
1 sup
*bZ St 1 T D 2 b * sup *
t[JT1
&
1
2
t[JT1
1
2
2
sup T t[J1
2 yt21
T
T
t[JT1
*bZ St 1 T D 2 bZ ~t!* sup 2
2 b2
T
(
@t0 T #11
2 y@tT #
1 sup
t[JT1
p
2
1
2
*
T
( yt21 «t @tT #11 T
y@tT # «@tT #11
t[JT1
T
*
*
0+
This derives equation ~18!+
APPENDIX J: DERIVATION OF EQUATIONS ~19!–~21! For b1 5 1, 6 b2 6 , 1 and for t 5 t0 2 ~m0T !, m 5 1,2,3, ++ +, and ~m0T ! r 0 as T r `, we have ~Ja!
1 T
yk202m n s 2 B 2 ~t0 ! 5 Op ~1!;
STRUCTURAL CHANGE IN AR(1) MODELS
~Jb!
T
m[$1,2, + +%, ~m0T !r0
~Jc!
m[$1,2, + +%, ~m0T !r0
S
*
k0
(
k 02m11
yt21 «t 5 op ~1!;
*bZ St 2 T D 2 1* 5 o ~1!; m
sup
m
bZ 2 t0 2
*
1
sup
T
1
0
D S
5 a t0 2
p
m T
D
1
k0
11
S S
1 1 2 a t0 2
( «t yt21 k 2m11 0
k0
(
k 02m11
m T
DD
1
2 yt21
2
T
b2 1
( «t yt21 k 11 0
T
(
k 011
2 yt21
2
,
where
S
a t0 2
m T
D
k0
1 5
(
T k0
1 T
k 02m11
(
k 02m11
2 yt21
1
2 yt21
1 T
T
(
k 011
mB 2 ~t0 !
n 2 yt21
mB 2 ~t0 ! 1
1 2 t0 1 B 2 ~t0 ! 1 2 b22
Further, because k0
( «t yt21 k 2m11 0
k0
(
k 02m11
T
5 op ~1!,
( «t yt21 k 11 0
T
(
2 yt21
k 011
5 op ~1!,
2 yt21
we have
S
~Jd! bZ 2 t0 2
m T
D
n
~1 2 b22 !mB 2 ~t0 ! 1 ~1 2 t0 1 B 2 ~t0 !!b2 ~1 2 b22 !mB 2 ~t0 ! 1 1 2 t0 1 B 2 ~t0 !
Therefore
S
bZ 2 t0 2
m T
D
2 b2 n
~1 2 b22 !mB 2 ~t0 !~1 2 b2 ! ~1 2 b22 !mB 2 ~t0 ! 1 1 2 t0 1 B 2 ~t0 !
This derives equation ~19!+
+
+
+
149
150
TERENCE TAI-LEUNG CHONG
Using ~B+1! in Appendix B, ~5a!, ~5b!, and ~5d! in Lemma 3, and ~Ja!–~Jd!, we have 1 T
S
RSST t0 2
m T
( «t2 5
2
k0
yt21 «t
2
T
S( D k 02m
T 1
D
SS
1
2 2 bZ 2 t0 2
k 02m
(1
T
2 yt21
m T
D D 21
(
k 02m11
yt21 «t
T
k0
SS
1 bZ 2 t0 2
m T
(
D D
2 k 2m11 0
21
2 yt21
T T
SS
2 2 bZ 2 t0 2
m T
D D 2 b2
( yt21 «t k 11 0
T
T
SS
1 bZ 2 t0 2
5
T
SS
1 op ~1! 1 op ~1! 1 bZ 2 t0 2
SS
1 bZ 2 t0 2
m T
D D 21
0
T
m T
2 ( yt21 2 k 11
D D 2 b2
0
T
~1 2 t0 1 B 2 ~t0 !!~1 2 b2 ! 2 ms 2 B 2 ~t0 ! ~1 2 b22 !mB 2 ~t0 ! 1 1 2 t0 1 B 2 ~t0 !
This derives equation ~20!+ For t 5 t0 1 ~m0T !, we have ~Je!
1
sup m[$1,2, + +%, ~m0T !r0
~Jf !
1 T
k 01m
2 n ( yt21 k 11
T
*(
k 01m k 011
*
yt21 «t 5 op ~1!;
s 2 B 2 ~t0 !~1 2 b22m ! 1 2 b22
0
by ~5d! in Lemma 3; ~Jg!
sup m[$1,2, + +%, ~m0T !r0
~Jh!
1 T
T
1 T 1
*
T
( yt21 «t k 1m11 0
T
*
5 op ~1!;
2 2 # yt21 5 Op ~1! ( yt21 T k( k 1m11 11 0
2 b2
2 ( yt21 2 k 2m11
T
n s2 1
D D
0
T
k0
T
(1 «t2
m T
2 ( yt21 2 k 11
0
+
1 op ~1!
STRUCTURAL CHANGE IN AR(1) MODELS by ~5d! in Lemma 3; ~Ji! ~Jj!
*bZ St 1 T D 2 1* 5 O S T D; m *bZ St 1 T D 2 b * 5 o ~1!+ m
sup m[$1,2, + +%, ~m0T !r0
sup m[$1,2, + +%, ~m0T !r0
151
1
0
2
0
1
p
2
p
Using ~B+3! in Appendix B, by ~5a!–~5c! of Lemma 3 and ~Je! to ~Jj! given previously, we can write the residual sum of squares as follows: k0
T
1 T
S
RSST t0 1
m T
D
( «t2 5
1
SS
m
2 2 bZ 1 t0 1
T
T
D D 21
(1 yt21 «t T
k0
SS
1 bZ 1 t0 1
m T
D D 21
2
2 (1 yt21
T k 01m
SS
2 2 bZ 1 t0 1 k 01m
(
3
k 011
m T
D D 2 b2
S(
(
yt21 «t
k 011
T
T
2
D
m T
D D
2
2 b2
2
T
2 yt21
SS
1 bZ 1 t0 1
k 01m11
yt21 «t
T
T
2 ( yt21 k 1m11 0
k 01m
T
5
(1 «t2
SS
m
1 op ~1! 1 op ~1! 1 op ~1! 1 bZ 1 t0 1 T T 1 op ~1!
n s2 1
D D 2 b2
(
2 k 11 0
2 yt21
T
~1 2 b2 !s 2 B 2 ~t0 !~1 2 b22m ! 1 1 b2
+
This derives equation ~21!+
APPENDIX K: PROOF OF THEOREM 4 It is not difficult to show that t[ T is T-consistent by using a similar proof as in Appendixes C and G+ We want to prove a stronger result that lim Tr` Pr~ kZ Þ k 0 ! 5 0+
152
TERENCE TAI-LEUNG CHONG
Because kZ # k 0 1 Op ~1!, for any h . 0, there exists an M , ` such that Pr~6 kZ 2 k 0 6 . M ! , h+ Therefore, Pr~ kZ Þ k 0 ! 5 Pr~6 kZ 2 k 0 6 . M ! 1 Pr~6 kZ 2 k 0 6 # M # h 1 Pr~6 kZ 2 k 0 6 # M M
#h1
S S D S S D
( Pr m51
M
1
( Pr m51
and
RSST t0 2
RSST t0 1
m
T
m
T
kZ Þ k 0 !
and
kZ Þ k 0 !
D
2 RSST ~t0 ! , 0
D
2 RSST ~t0 ! , 0 +
Let h 1~m! and h 2 ~m! be defined as in equations ~20! and ~21!, respectively+ The preceding probability is equal to M
5h1
M
( Pr~h1~m! 1 D1T , 0! 1 m51 ( Pr~h2 ~m! 1 D2T , 0!, m51
where
S S
D 1T 5 RSST t0 2 D 2T 5 RSST t0 1
m T m T
D D
2 RSST ~t0 ! 2 h 1 ~m! 5 op ~1!, 2 RSST ~t0 ! 2 h 2 ~m! 5 op ~1!+
Because h 1~m! and h 2 ~m! are increasing with m, this implies every individual probability given earlier is declining with m+ Thus, the preceding term is bounded by h 1 M Pr~h 1~1! 1 D 1T , 0! 1 M Pr~h 2 ~1! 1 D 2T , 0!+ Further, because h 1~1! and h 2 ~1! are positive and of order Op ~1!, there exists a T large enough such that Pr~h 1~1! 1 D 1T , 0! # h and Pr~h 2 ~1! 1 D 2T , 0! # h for any h . 0+ Therefore for any h . 0, Pr~ kZ Þ k 0 ! # ~2M 1 1!h for all large T+ Because M is finite, we have lim Tr` Pr~ kZ 5 k 0 ! 5 1 and plim bZ 1~ t[ T ! 5 plim bZ 1~t0 ! 5 b1 5 1, plim bZ 2 ~ t[ T ! 5 plim bZ 2 ~t0 ! 5 b2 + These prove the consistency of bZ 1 and bZ 2 + To find the limiting distributions of bZ 1~ t[ T ! and bZ 2 ~ t[ T !, note that for any given real value x, lim Pr~T ~ bZ 1 ~ t[ T ! 2 b1 ! # x!
Tr`
5 lim Pr~T ~ bZ 1 ~t0 ! 2 b1 ! # x! lim Pr~ kZ 5 k 0 ! Tr`
Tr`
1 lim Pr~T ~ bZ 1 ~ t[ T ! 2 b1 ! # x6 kZ Þ k 0 ! lim Pr~ kZ Þ k 0 ! Tr`
Tr`
5 lim Pr~T ~ bZ 1 ~t0 ! 2 b1 ! # x!+ Tr`
Thus, bZ 1~ t[ T ! and bZ 1~t0 ! have the same asymptotic distribution+ Similarly, bZ 2 ~ t[ T ! and bZ 2 ~t0 ! have the same asymptotic distribution+ By ~5a!–~5d! in Lemma 3, we have
STRUCTURAL CHANGE IN AR(1) MODELS
153
@t0 T #
(1
d
T ~ bZ 1 ~ t[ T ! 2 1! 5 T ~ bZ 1 ~t0 ! 2 1! 5
«t yt21 0T
@t0 T #
(1
B 2 ~t0 ! 2 t0
n
2
2 yt21 0T 2
E
,
t0
B2
0
T
!T ~ bZ 2 ~ t[ T ! 2 b2 ! 5d !T ~ bZ 2 ~t0 ! 2 b2 ! 5
( «t yt21 0!T @t T #11 0
T
(
@t0 T #11
n
2 yt21 0T
O !1 2 b B~1! 2 2
1 2 t0 1 B 2 ~t0 !
,
where B~{! and B~{! O are defined in Section 2+ These show the limiting distributions of bZ 1~ t[ T ! and bZ 2 ~ t[ T ! for a fixed magnitude of break+ These derive the limiting distribution of bZ 1 and bZ 2 for fixed magnitude of break+ To derive the limiting distribution of t[ T for shrinking break, we fix b1 at one and let b2T 5 1 2 @10!Tg~T !# , where g~T ! . 0, with g~T ! r ` and g~T !0!T r 0 as T r `+ Let y be a finite constant and B~{!, B1~{!, and B2 ~{! be defined as in Section 2+ For t 5 t0 1 y@ g~T !0T # and y # 0, we have ~i! L T ~t! 5 op ~1!; @t0 T #
~ii! ~ b2T 2 1!
T
2 ( yt21 ( yt21 «t @tT #11 @t T #11 0
T
52
2 ( yt21 @tT #11
Op ~Tg~T !!Op ~T !
!Tg~T !Op ~T
2
!
5 Op
S! D g~T ! T
5 op ~1!; T
~iii!
2 ( yt21 @t T #11
1;
&
T
(
@tT #11
~iv!
p
0
2 yt21 @6 y6 g~T !#21
1
(0
!Tg~T ! 5 5
yk02t21 «k02t @6 y6 g~T !#21
1
(0
!Tg~T ! yk 0
@6 y6 g~T !#21
(0
52
yk 0
S!
1
!T !g~T !
D
t
yk0 2 ( «k02i «k02t i50
@6 y6 g~T !#21
1
!T !g~T ! 3
S
(0
«k02t 2
g~T !
!T
t
1
( «k 2i «k 2t g~T ! i50 0
0
D
1 g~T !
@6 y6 g~T !#21
(0
~2«k02t ! 2 Op
S! D g~T ! T
n 2s 2 B~t0 !B1 ~2y!;
154
TERENCE TAI-LEUNG CHONG
~v!
@6 y6 g~T !#21
1
(0
Tg~T ! 5
yk202t21 @6 y6 g~T !#21
1
(0
Tg~T !
5 6y6 1 5 6y6
yk20
22
!
t
yk0 2 ( «k02i i50
g~T ! yk0
T g~T !
1
T
g~T !
yk20
S
T
!T
1 Op
2
@6 y6 g~T !#21
@6 y6 g~T !#21
(0
D S!
(0
S!
( «k 2i g~T ! i50
t
1
( «k02i
g~T !
S! D S D g~T !
t
1
i50
g~T !
1 Op
T T T Thus, equation ~B+2! in Appendix B becomes
D
0
2
D
1 g~T !
n 6y6s 2 B 2 ~t0 !+
RSST ~t! 2 RSST ~t0 ! @t0 T #
52
2
@t0 T #
(
5 22~ b2T 2 1!
@tT #11
yt21 «t 1 ~ b2T 2 1! 2
(
@tT #11
@6 y6 g~T !#21
(0
!Tg~T !
~2yk02t21 «k02t ! 1
2 yt21 1 op ~1! @6 y6 g~T !#21
1 g~T !T
(0
yk202t21 1 op ~1!
n 22s 2 B~t0 !B1 ~6y6! 1 6y6s 2 B 2 ~t0 !+ Similarly, for t 5 t0 1 @yg~T !0T # and y . 0, by ~B+4! in Appendix B and the facts that ~vi! L T ~t! 5 op ~1!, @t0 T #
2 ~vii! ~ b2T 2 1!
(1
@tT #
yt21 «t
(
@t0 T #11
2 yt21
52
@tT #
2 (1 yt21
Op ~T !Op ~Tg~T !!
!Tg~T !Op ~T 2 !
5 Op
5 op ~1!, @t0 T #
~viii!
(1
2 yt21
p
@tT #
(1
&
1,
2 yt21
equation ~B+4! in Appendix B becomes RSST ~t! 2 RSST ~t0 ! @tT #
5 2~ b2T 2 1! 52
2
(
@t0 T #11
@yg~T !#21
(0
!Tg~T !
5 22
yk 0
@tT #
yt21 «t 1 ~ b2T 2 1! 2
1
!T !g~T !
yk01t «k01t11 1
Tg~T !
«k01t11 1 y
n 22s 2 B~t0 !B2 ~y! 1 ys 2 B 2 ~t0 !+
yk20 T
2 yt21 1 op ~1!
@yg~T !#21
1
@yg~T !#21
(0
(
@t0 T #11
(0
1 op ~1!
yk201t 1 op ~1!
S! D g~T ! T
STRUCTURAL CHANGE IN AR(1) MODELS
155
Applying the continuous mapping theorem for argmax functionals, we have ~1 2 b2T ! 2 T 2 ~ t[ T 2 t0 ! 5 d
&
T g~T !
~ t[ T 2 t0 ! 5 y[ 5 Arg min $RSST ~t! 2 RSST ~t0 !%
H S
y[R
Arg min 22s 2 B 2 ~t0 ! y[R
5 Arg max y[R
B * ~y! B~t0 !
2
1 2
S D
B * ~y! B~t0 !
2
1 2
6y6
DJ
6y6 ,
where B * ~y! is a two-sided Brownian motion on R defined to be B * ~y! 5 B1~2y! for y # 0 and B * ~y! 5 B2 ~y! for y . 0+ n