ASYMPTOTICALLY EXACT CONFIDENCE

INTERVALS OF CUSUM AND CUSUMSQ TESTS:

A Numerical Derivation Using Simulation Technique
Hisashi TANIZAKI
Faculty of Economics, Kobe University, Nada-ku, Kobe 657, JAPAN

ABSTRACT: In testing a structural change, the approximated confidence intervals

are conventionally used for CUSUM and CUSUMSQ tests. This paper numerically
derives the asymptotically exact confidence intervals of CUSUM and CUSUMSQ tests.
It can be easily extended to nonnormal and/or nonlinear models.
KEY WORDS: CUSUM test, CUSUMSQ test, Monte-Carlo simulation, Asymptotically exact confidence interval

INTRODUCTION

There is a great amount of literature on use of recursive residuals, e.g., Brown, Durbin
and Evans (1975), Galpin and Hawkins (1984), Harvey (1989), Johnston (1984), Ploberger
(1989), Ploberger, Kramer and Alt (1989), and Westlund and Tornkvist (1989). Especially,
Brown, Durbin and Evans (1975) described an important application of recursive residuals
in testing for structural change over time. The technique is appropriate for time series data,
and might be used if one is uncertain about when a structural change might have taken
place. We have two tests; cumulative sum (CUSUM) test and cumulative sum of squares
(CUSUMSQ) test The null hypothesis is that the coefficient vector is the same in every
period; the alternative is simply that it (or the disturbance variance) is not. The test is
quite general in that it does not require a prior specification of when the structural change
takes place. However, it is known that the power of the test is rather limited compared to
that of the Chow test. The test is frequently criticized on this basis. However, the Chow
test is based on a rather definite piece of information, namely, when the structural change
takes place. If this is not known or must be estimated, the advantage of the Chow test
diminishes considerably (see Greene (1990) and Kramer (1989)). See Galpin and Hawkins
(1984) for an application.
One of the reasons why the CUSUM test is less powerful is that the confidence interval
of the test is approximated. For the CUSUM test statistic, we cannot derive the explicit
distribution, and therefore the approximated confidence interval is conventionally used in
testing the stability of the coefficient. Although it is known that the CUSUMSQ test statistic
is distributed as a beta random variable, the confidence interval of the CUSUMSQ test
is also approximated. Therefore, in this paper, an attempt is made to obtain the exact
confidence intervals of the CUSUM and CUSUMSQ tests asymptotically using the MonteCarlo simulation technique. Moreover, the power is compared for both the approximated
confidence intervals (the confidence intervals conventionally used) and the simulated ones
(the confidence intervals proposed in this paper).
1

OVERVIEW OF CUSUM AND CUSUMSQ TESTS

Consider the following regression model:

ut N (0, 2 ),

yt = xt + ut ,

t = 1, , T,

where is a k 1 unknown parameter vector. yt is a dependent variable while xt is a 1 k

vector of independent variables. ut is assumed to be normally distributed with mean zero
and variance 2 . Define Xt and Yt as follows:


y1
x1
y2
x2

Yt =
Xt =
.. ,
... .
.
xt
yt
The null hypothesis of no structural change for the model is specified as:
H0 : 1 = 2 = = T = and 12 = 22 = = T2 = 2 ,
where t denotes the vector of coefficients in period t and t2 the disturbance variance in
that period. The null hypothesis would be violated if remained constant but 2 varies.
Let bt be the ordinary least squares estimate of t using the data Yt and Xt , i.e., bt =
(Xt0 Xt )1 Xt0 Yt . It is well known that bt is obtained recursively (see Brown, Durbin and
Evans (1975), Johnston (1984) and Riddell (1975) for the recursive algorithm). The recursive
residual is defined as:
yt xt bt1
,
(1)
t = p
0
Xt1 )1 x0t
1 + xt (Xt1
which has mean zero and variance 2 .
The CUSUM test statistic for testing structural change is given by:
Wt =

t
X

.
i
b,

t = k + 1, , T,

(2)

i=k+1
T
X
1
i2 , which implies the unbiased estimate of 2 . The expected value
T k
i=k+1
of Wt is zero and the distribution of Wt is symmetric about zero if the disturbance in the
regression model is symmetric. Since we cannot obtain the explicit distribution of Wt , we
conventionally test as follows. The null hypothesis is accepted if Wt , t =
k + 1, , T ,
stay within
a
pair
of
straight
lines
which
pass
through
the
points
(k,
c
w T k) and

(T, 3cw T k), where cw is a parameter depending on the significance level chosen for
the test. It is known that we have cw =1.143 when =0.01, cw =0.948 when =0.05 and
cw =0.850 when =0.10 (see Brown, Durbin and Evans (1975) and Johnston (1984)).
The CUSUM test is usually used with the CUSUMSQ test, which plays a role of complement to the CUSUM test. Using the CUSUM test, we can see when the structural
change occurs. However, the CUSUM test is not very powerful. Even though the structural
change clearly takes place in a period, the null hypothesis is often accepted. Conversely, the
CUSUMSQ test is too powerful but we cannot know the period when the structure changes.
Accordingly, both tests are complementarily used.
The CUSUMSQ test statistic which is the alternative test to the CUSUM test is represented as:
t
T
. X
X
St =
i2
i2 ,
(3)

where
b2 =

i=k+1

i=k+1

Table 1. =0.10, 0.05, 0.01 and T k =10, 20, 30, 40, 50

T k

0.10
0.05
0.01

10

20

30

40

50

.39075
.44641
.54210

.31325
.35277
.43071

.26767
.30081
.36649

.23781
.26685
.32459

.21630
.24245
.29456

tk T t
which is distributed as a Beta random variable with parameter (
,
). The expected
2
2
tk
value of St is
. Also, the confidence interval is conventionally approximated, which is
T k
tk
given by a pair of straight lines cs
, where cs depends on both the sample size T k
T k
and the significance level . For =0.10, 0.05, 0.01 and T k =10, 20, 30, 40, 50, cs is given
by Table 1, which represents a both-sided test. See Durbin (1969) and Johnston (1984) for
cs . Durbin (1969) pointed out that the test statistic gives us a good approximation when
T k is large (see Harvey (1981)).
An example of T k = 50 and = 0.01, 0.05, 0.10 is taken in Figures 1a and 1b.
90%, 95% and 99% denote the confidence intervals for each significance level, i.e., =
0.10, 0.05, 0.01.
The confidence intervals displayed in Figures 1a and 1b are conventionally used for
CUSUM and CUSUMSQ tests.
There is some evidence that the CUSUM test is less powerful than the CUSUMSQ test in
small sample. Some Monte-Carlo experiments by Garbade (1977) also suggest that the latter
may not be very powerful in comparison with tests based on variable parameter models.
It is known that the confidence intervals of these two tests are the approximated ones.
The exact confidence interval cannot be obtained explicitly for the CUSUM test, because
its distribution is not known. Also, for the CUSUMSQ test, the exact confidence interval
is not utilized, even though its distribution is known. Thus, because the exact distribution
is not utilized for these two tests, the CUSUM test is not powerful and the CUSUMSQ
test is too powerful in small sample. In the next section, we obtain the asymptotically
exact confidence intervals using the simulation technique such as the method of simulated
moments (see McFadden (1989)).

NUMERICAL DERIVATION OF CONFIDENCE INTERVALS

In this section, we construct both exact confidence intervals of CUSUM and CUSUMSQ
tests. The intervals are obtained by generating random numbers.
For the CUSUM test, the confidence interval is simulated as follows. First, generate
T k normal random numbers, which are denoted by t , t = k + 1, , T , and compute
the CUSUM test statistic (2). Repeat this procedure m times. Denote the CUSUM test
(j)
statistic in the j-th smallest value of the m simulation runs by Wt , which implies that
(j)
(1)
Wt , j = 1, , m, are sorted by size for all t = k + 1, , T . That is, Wt is the smallest
(m)
value and Wt
the largest one. We take m =20,000 in this paper.
3

Figure 1a. CUSUM test

.
. . .
. . .
. . .
.
.
.
.
. . .
....... ......
. . .
. .......
. . .
....... .....
.........
. . .
.... .......
....................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
....
. .
....... ...... ........................................
. . .
. .......
..
. . .
....... ...... .......................................
. . .
. .......
...
....... .............................................
. . .
.
.
.
.
.
.
.
.
.
.
. .......
................
. . .
...... .............................................
. . .
.... .......
.....
. . . .. ....... .................................................
..............................
. . .
.
.
.
.
.
.
.
.
.
.
.
.
. . .... ....... .....................................
............................................

20

10

Wt

.............................................................................................................................................................................................................................................................................................................................................................................

.............................................
. . . .... ....... ...............................................
...... ....... .....................
. . .
....... ...... ...........................
. . .
..
. .......
. . .
....... ..............................................
. . .
. .......
..
. . .
....... .............................................
. . .
..
. .......
....... ..............................................
. . .
..
. .......
. . .
....... ...... ........................................
. . .
. .......
..
. . .
....... ...... .......................................
. . .
....................
. .......
....................
....... ......
. . .
. .......
. . .
....... ......
. . .
. .......
. . .
. . .
. . .
. . .
. . .

10

20

10

20

30

40

50

tk
............................................

90%

....... ....... ....... .......

95%

. . . . . . . .

99%

Figure 1b. CUSUMSQ test

1.0

.. ...
.....
.
. ......
......
.
...... ......
......
. ......
......
.................
......
.
.
.
.
.
.
.
.
.
.
..... .....
......
.
. ......
......
.
...... ......
.
......
. ......
......
.
.................
......
.
.
.
.
.
.
.
.
.
.
.
.
..... .....
......
.
. ......
......
.
...... ......
.
......
. ......
.....
......
.
.................
......
.
......
.
.
.
.
.
.
.
.
.
.
.
...... ..
..... .......
......
...... ....
.
......... ..
... ...........
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. ....
...... ...
......
.
.
...... ......
...... .....
.
......
.
. ......
.
...... ..
......
.
...... .........
...... ....
.
......
.
... ...........
......
.
......... ..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. ....
...... ...
......
.
.
...... ......
...... .....
.
......
.
. ......
.
...... ..
......
.
...... ........
...... .....
.
......
.
... ............
.
......... ..
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
....
......
.
...... ..
.
...... ......
...... .....
.
......
.
. ......
.
...... ..
......
.
...... ........
...... .....
.
......
.
... ............
.
.......... ..
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
....
......
.
...... ..
.
...... ......
...... .....
.
......
.
. ......
.
...... ..
......
.
...... ........
.
...... .....
......
.
... ............
.
.......... ..
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
....
.
...... ..
......
.
..................
.
......
...... .....
.
.
...... ..
......
.
...... ......
.
...... .....
......
.
. ......
.......... ..
......
.
.
.
.
.
.
.
..................
.
.
.
.
. .
.
.
...... ..
......
.
..................
...... ......
......
.
......
...... .
.
.....
......
...... .....
.
.......... ..
......
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
...... ..
......
.
...... ......
......
.
...... .
......
.
...... .....
......
.
.......... ..
......
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
......
...... ..
.
...... ......
......
.
...... .
......
.
......
...... ......
.
.......... .
......
.
.
.
.
.
.
.
.
.
.
.
.
..
.. ..
.

0.8

0.6
St
0.4

0.2

0.0

10

20

30

40

50

tk
............................................

90%

....... ....... ....... .......

95%

. . . . . . . .

99%

Denote the significance level through all periods by and that in each period by w . Let
(U )
(L)
Wt be the upper bound and Wt be the lower one, which implies (U L)/m =1 w
in the case of 100(1 w )% confidence interval at time t, where w is the significance level
at time t for the CUSUM test. If we assume that the distribution of the disturbance ut is
symmetric about zero, that of the CUSUM test statistic is also symmetric, and therefore we
(U )
(L)
have Wt = Wt for all t. In this procedure, we need to obtain the confidence interval
(L)
(U )
[Wt , Wt ], t = k + 1, , T , and the significance level at each time (i.e., w ), given .
Note that we have the relationship 1 = (1 w )T k if Ws is independent of Wt for
t 6= s. Since Wt is clearly correlated with Ws for t 6= s, we have 1 6= (1 w )T k .
(U )
(L)
Summarizing above, we obtain Wt , Wt and w , given , satisfying the following
three conditions:
(U )

(L)

(U )

Prob(Wk+1 > Wk+1 > Wk+1 , , Wt

(L)

> Wt > Wt
(U )
WT

(U )

Prob(Wt
(U )

Wt

(L)

> Wt > Wt
(L)

= Wt

,,
(L)

> WT > WT ) = 1 ,

) = 1 w ,

(4)
(5)

for all t = k + 1, , T.

(6)

Equation (4) corresponds to the joint density function of Wk+1 , , WT , while Equation (5)
is based on the marginal density function of Wt . Here, we apply the simulation technique
(U )
(L)
used in Diebold and Rudebusch (1991) to obtain Wt , Wt and w , given .
It might be appropriate to assume that there is a certain relationship between w and
, which is represented as:
= fw (w ),
where fw () is unknown but derived by the simulation technique. fw (w ) denotes the prob(L)
(U )
ability that {Wt }Tt=k+1 lies outside the interval {(Wt , Wt )}Tt=k+1 , given the significance
1
level at time t (i.e., w ). In practice, we use the conventional Newton-Raphson nonlinear
optimization procedure to obtain w , i.e.,

(i+1)
(i)
(i)
(i)
w
= w
+ dw
fw (w
) ,
for i = 1, 2, ,
(7)
(i)

(i)

(i)
). The superscript (i) denotes the i-th
where dw can be interpreted as dw = 1/fw0 (w
(1)
iteration. We take w = , which is the initial value for w . However, since the function
fw () is unknown and accordingly we cannot obtain the first derivative of fw (), we choose
(i)
(i1)
(0)
dw = w dw , dw = 1 and w = .5. Moreover, taking into account the convergence speed,
at each i we consider choosing either the above Newton-Raphson procedure or the following:

(i+1)
(i)
w
=

,
if

1
(8)

0.1,
w
(i)
(i)
fw (w )
fw (w )
(1)

(i+1)

(i)

where w = . The convergence criterion is: |w

w | < 0.0001. For the significance
(i+1)
level at each time (i.e., w ), we choose w
satisfying the convergence criterion.
Thus, summarizing above, the procedure constructing the confidence interval for the
CUSUM test is as follows:
1 Given , we can compute the probability f ( ) by the simulation. Accordingly, we obtain such
w
w
w
w
1
that fw (w ) is equal to . That is, w = fw
() is derived. See Figure 2 for relationship between w and
fw (w ).

(i) Generate m sets of T k standard normal random numbers (i.e., m(T k) standard
random draws),2 which corresponds to m sets of the normalized recursive residuals
{t /}Tt=k+1 . For each set of the T k random draws, we compute {Wt }Tt=k+1 and
{St }Tt=k+1 .
The CUSUM test statistic (i.e., {Wt }Tt=k+1 ) and the CUSUMSQ test statistic (i.e.,
{St }Tt=k+1 ) obtained from the jth set of the T k random draws are denoted by
(j)

(j)

{Wt }Tt=k+1 and {St }Tt=k+1 .

(i)

(U )

(L )

(ii) Given the random draws, and w , obtain Wt i and Wt i for t = k + 1, , T

(U )
(L )
(U )
(L )
(i)
such that (Ui Li )/m = 1 w
and Wt i = Wt i . Wt i and Wt i denote the
U -th and L-th largest values at time t in the i-th iteration.
(ji ) T
}t=k+1 ,

(iii) Count the number of the series {Wt

(L )
{(Wt i ,

(U )
Wt i )}Tt=k+1 .

j = 1, , m, outside the interval

Note that
(j )

[ the number of {Wt i }Tt=k+1

(L )
(U )
falling in the interval {(Wt i , Wt i )}Tt=k+1 ]
(i)
,
fw (w
)=1
m
which corresponds to the significance level through all periods, i.e., . Also, note that
(U )
(L )
we have the restriction Wt i = Wt i because the probability density function of
the CUSUM test statistic is symmetric.
(i+1)

(i)

(iv) Using the following optimization procedure, update from w to w

(i+1)
w

(1)

if

(i)

f ((i) ) w ,
w
w
=

(i)
(i)
(i)

w + dw fw (w ) ,
(i1)

(i)

where w = , dw = w dw

(i+1)

(v) Repeat (ii)-(iv) until w

(i)

fw (w )

1 0.1,

otherwise,

(0)

, dw =1 and w is a constant value.

(i+1)

is stable, i.e., |w

(i)

w | < 0.0001.

For the CUSUMSQ test, we can take the almost same procedure as above. Similarly,
after m simulation runs, denote the j-th largest value of the CUSUMSQ test statistics in
(j)
(j)
m simulations by St , where St , j = 1, , m, are already sorted by size. That is, for
(1)
(m)
all t, St is the smallest value and St the largest one. Also, m =20,000 is taken for the
CUSUMSQ test.
For the CUSUMSQ test, let s be the significance level in each period. Given , we
(L)
(U )
compute St , St and s satisfying:
(U )

(L)

(U )

Prob(Sk+1 > Sk+1 > Sk+1 , , St

(L)

> St > St , ,
(U )

ST

(L)

> ST > ST ) = 1 ,

(9)

2 In the case where the error u is nonnormal, we generate the random numbers corresponding to the
t
underling assumption of ut .

Figure 2. w vs f (w ) and s vs f (s ): T k = 50 and m = 20, 000

1.0

. ....... . ...................................................................................................................
..... ...... ..............
. . .............
.....
....
.....
.
.
.
.
....
....
..
...
...
...
..
.
.
.
.
.
.
..
...
...
..
... ....
.
.
.. .
... ....
.
.. ...
.. ...
.
.. ...
.. ...
..
.. ..
... ....
.. ....
.. ...
..
.. ..
.. ....
.... ....
. ...
..
.. ..
...........................................
......
w
w
........
....
.
.
....... ....... ....... ......
...
s
s
.......
.
.....
......
.
.
..
...
...
..
..
.
....

0.8

f (w ),
f (s )

0.6
0.4

vs f ( )

0.2

vs f ( )

0.0
0.0

(U )

Prob(St

0.2

0.4
0.6
w , s

0.8

1.0

(L)

> St > St ) = 1 s ,

(10)

(L)
St

(U )
= 1 ST +kt ,
(U )
(L)
St = ST +kt ,

(11)
for t = k + 1, , T 1.
(U )

(12)

(L)

Given the significance level , we consider computing St , St and s for the CUSUMSQ
test. We have to obtain s which is constant over time. If t , t = k + 1, , T , are mutually
independently and identically distributed, then the density function of St is equivalent to
that of 1 ST +kt for all t = k + 1, , T 1. Accordingly, we have the relationships:
(L)
(U )
(U )
(L)
St = 1 ST +kt and 1 St = ST +kt . We obtain s which satisfies = fs (s ). fs (s )
(L)

(U )

denotes the probability which {St }Tt=k+1 lies outside the interval {(St , St )}Tt=k+1 , given
the significance level at time t (i.e., s ). The procedure constructing the confidence interval
for the CUSUMSQ test is as follows.

(i)

if |
1| 0.1,

(i)
f ((i) ) s ,
fs (s )
s
s
(i+1)
s
=

s(i) + ds(i) fs (s(i) ) ,

otherwise,
(1)

(i)

(i1)

(0)

where s = , ds = s ds
, ds = 1 and s is a constant value. Note that we need to have
(U )
(L)
(U )
(L)
(L)
(U )
(U )
(L)
St and St such that St St is minimized, St = 1 ST +kt and 1 St = ST +kt
for t = k + 1, , T 1. In this paper, we take s = .5. Also, the convergence criterion is:
(i+1)
(i)
|s
s | < 0.0001.

Taking
an example of the case: T k = 50 and m = 20, 000, w , fw (w ) and s ,
fw (s ) are displayed in Figure 2.
7

fw () and fs () are monotone and accordingly their inverse functions exist. The proposed
approach of constructing the confidence intervals is guaranteed to work in all circumstances,
even for nonnormal cases.3
In Table 2, we check the convergence speed in simulating the confidence intervals, where
= 0.10, 0.05, 0.01, T k =50 and m =20,000 are taken. We can see that the convergence
is very quick.
We concretely show the computation procedure, taking an example of the case: CUSUM
test and = 0.10. Constructing the confidence interval based on w = 0.1000 and com(L )
(U )
puting the probability which lies outside the interval (Wt 1 , Wt 1 ) for some t, we have
(1)
(1)
(2)
fw (w ) = 0.5374. Using equation (8), w is updated to w , which is 0.0186. Based on
(2)
w = 0.0186, again, we compute the confidence interval and the corresponding significance
(2)
(3)
(4)
level, i.e., fw (w ) = 0.1527. For the computation from fw (w ) to w , the convergence
criterion is switched to equation (7). The procedure is completed by the sixth iteration
(5)
(6)
(5)
(6)
since w is the same value as w , i.e., w =w =0.0112. As it is seen in Table 2, the
convergence speed is quite fast. By the results obtained from Table 2, we have the confidence intervals in Figures 3a and 3b, where the case of T k = 50, m = 20, 000 and
= 0.10, 0.05, 0.01 is displayed.
Thus, we have the confidence intervals of CUSUM and CUSUMSQ tests as shown in
Figures 3a and 3b. However, the intervals in Figures 3a and 3b are not smooth, which
implies the obtained confidence intervals are not reliable. In order to improve this problem,
repeating the procedure n times, we take the arithmetic average of n sets of upper bounds
and lower bounds for the confidence intervals under each significance level . As n goes to
infinity, by law of large number, the arithmetic average of the intervals becomes the true
intervals.4 n = 2, 000 is taken in Figures 4a and 4b, where each simulated 90%, 95% and
99% confidence intervals are drawn.
According to the procedure shown in Figures 4a and 4b, we can construct the intervals
for any T k and any , given m and n.

3 and denote the significance levels at a single time, while f ( ) and f ( ) denote the significance
w
s
w
w
s
s
levels over all time periods. When w increases, fw (w ) also clearly increases. And similarly, when s
increases, fs (s ) also clearly increases. Therefore, fw (w ) and fs (s ) are monotone. Accordingly, the
1
inverse functions, i.e., fw
() and fs1 (), exist for all nonnormal errors.
4 As m increases, the simulated confidence intervals become smooth. However, an increase in m yields an
increase in data storage. We cannot take an extremely large m in practice because of a computer and/or
compiler limitation. Therefore, we consider repeating the procedure (i)(v) n times, which implies obtaining
n sets of simulated confidence intervals.

Table 2. Convergence of Simulated Confidence Intervals (T k = 50)

i
1
2
3
4
5
6
7
1
2
3
4
5
6
7
1
2
3
4
5
6

CUSUM Test
CUSUMSQ Test
(i)
(i)
(i)
(i)
w
fw (w )
s
fw (s )
= 0.10
.1000
.5374
.1000
.6933
.0186
.1527
.0144
.1737
.0122
.1064
.0083
.1101
.0106
.0954
.0058
.0819
.0112
.1001
.0071
.0956
.0112

.0073
.0988

.0074

= 0.05
.0500
.3285
.0500
.4507
.0076
.0713
.0055
.0783
.0053
.0532
.0035
.0546
.0045
.0459
.0024
.0393
.0051
.0518
.0031
.0474
.0049
.0498
.0032
.0503
.0050

.0032

= 0.01
.0100
.0914
.0100
.1285
.0011
.0138
.0008
.0143
.0008
.0107
.0005
.0104
.0006
.0083
.0004
.0090
.0007
.0096
.0005
.0090
.0008

.0005

Figure 3a. CUSUM test

. .
. . . . . . . .
. . .
.......
. . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.... ...
. . . . .
....... ....... ... ...............................
. . . .
... .......
.....
. . . . .
....... ...... .... ..........................................
.
.
.
.
.
.
.
.
.
.
...... ..........................................................
. . .
. ...... .
..
. .
....... ...... ..............................
. . .
....... ....... .....................
. . .
.. ........................................
.
.
.
.
.
. .
.
.
.
.
.
.. ............................
.
.
.
.
.
.
.
.
.
.. ...... ...............
. . ....... ............................
. . . ...........................
. ..............................
. . ..
. ..............
..............
..........
....
.
.
.
.
.................................................................................................................................................................................................................................................................................................................................................................................
....
...
............
..............
. ................
. .................
. . .................
. ... ...........................
. . ....... ............................
.. ...... ...............
. .
. ...... ................
. .
. ....... ....................
. .
....... ...... .....................
. .......
..
. . .
....... .............................................
. . .
. ....... .
.
. .
...... ......................................................
. . .
....
....... ......
. . .
. ....... .... ...................................................
. . . .
... ....... ....
......................
. . . .
... ....... .
......................
. . . . .
...... .......
....... .......
. . . .
.......
. . .
. . . .
. . . . . .
.

20

10

Wt

10

20

10

20

30

40

50

tk
............................................

90%

....... ....... ....... .......

95%

. . . . . . . .

99%

Figure 3b. CUSUMSQ test

1.0

0.8

0.6
St
0.4

0.2

0.0

. ..............................................................................
.. .
. . ....................
...... ....
. .........................
.
......
. . ............................
......
....
. ................
......
.
.
.
.
.
.....
.
.
. ... .....
...
......
. ...........................
......
.......
.
.
.
.
.
.
.
.
.
.
.
.
.
..
......
. . ....... ........
.... .
......
. . .... ..........
... ...
......
. . ..... ...............
.... . .
......
.
..........
......
.......... .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
. .
.... .. .
.
......
. . .......
.... .
......
.
..... .......
..... ..... . .
......
. .....
. .
......
.....
.................
. .
......
............ .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
.
....
......
.
...... .........
..... ..... .
......
. . ....... ..........
..... .. .
......
..
.
......
..... ...
.
.. ..........
......
..... ... . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..... ...
. .........
......
.
..... ... . .
..............
......
.
..... ...
......
.
..... ...
.
...................
......
.
........... . .
.
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.. ......
...... .
......
. . ...............
.
......
...... .....
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
...... .....
......
.
.
..............
.... ..
......
.
.
..... ...
......
.. ......
........ ...
......
.
.
..... .......
......
.
...... .......
.
.
.
.
.
. .... .........
.
.
.
.
.
.
..... ..
.. ...
.
......
...... ..... . . .
......
. ..... ........
....... ..
......
.. .....
.
....... .....
.
......
.............
. ...... ........
......
.
.
.
.
.
.
.
.
.
.
.
.
. ....
.
.
.
...... ......
......
. ................
. .
....... .
......
.
.
........ .....
......
.
.....
....... .. .
......
.
. .............
......
........ ..... . .
.
.
.
.
.
.
.
.
.
.
.
. .
. . ......
......
....... ... . . .
......
...............
.........
........... . . .
......
...... ... .
......
.......
.............. . .
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
....
......
...... ..... .
......
.............. . .
.......
......
........... .
......
......... ....... . .
......
......
...................... .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.. .....
............ . . .
................................................................... .

10

20

30

40

50

tk
............................................

90%

....... ....... ....... .......

10

95%

. . . . . . . .

99%

Table 3. Maximum Standard Errors of Simulated Confidence Intervals

(m = 20, 000 and n = 2, 000)
T k

10

20

30

40

50

0.10

CUSUM
CUSUMSQ

.0391
.0038

.0755
.0041

.1027
.0040

.1259
.0039

.1510
.0038

0.05

CUSUM
CUSUMSQ

.0474
.0049

.1010
.0044

.1351
.0046

.1739
.0046

.2037
.0044

0.01

CUSUM
CUSUMSQ

.0758
.0084

.1847
.0097

.2745
.0090

.3465
.0083

.4337
.0079

Each value in Table 3 denotes the maximum value of the T k arithmetic standard errors
obtained from n sets of confidence intervals.
For the intervals of CUSUM test, the standard error is large when T k increases and/or
is small. For the CUSUMSQ test, the standard error increases as is small. In any case,
it is seen from Table 3 that all the standard errors are enough small compared with length
of the confidence intervals. It is clear that the standard errors in Table 3 approach zero as
n increases.
When comparing the conventional confidence intervals and the simulated confidence
intervals, the findings are as follows. For the CUSUM test (see Figures 1a and 4a), the
simulated confidence interval is larger than the conventional one at the middle of the period,
while the former is smaller than the latter otherwise. As for the CUSUMSQ test (see Figures
1b and 4b), similarly, the simulated one is larger than the conventional one in the middle,
while the former is smaller than the latter in the tails.
Finally, note as follows. For any underlying distribution of the disturbance ut in the
regression model, the computational procedure can be applied in the exactly same fashion.
In the following section, we compare the power for the numerical confidence intervals
and the conventional ones under the normal disturbance.

POWER OF CUSUM AND CUSUMSQ TESTS

Using the simulated confidence intervals obtained in the last section, we perform MonteCarlo experiments to examine the power of CUSUM and CUSUMSQ tests. Since the recursive residuals are identically, independently and normally distributed, we generate normal
random draws for t , t = k + 1, , T . Suppose that at the middle point of T k the structural change takes place. Thus, we can assume that t , t = k + 1, , T , are distributed as
follows:
t N (0, 1),

for t = k + 1, , (T k)/2,

t N (, 2 ),

for t = (T k)/2 + 1, , T,

11

Figure 4a. CUSUM test

. . .
. . . .
. . . .
. . . .
. .......
.
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
....... .......
. . .
....... .......
.........................
. . .
...... ....... .................................................
. . .
....... .......
..........
. . .
...... ...... .............................................
.
.
.
.
.
.
.
.
.
.
. .
. .......
................
....... ............................................
. . .
.. .......
. .
......
....... .... ......................
. .
. . . ....... ......................................
.
. . ..... .........................
. . ....... ......................
. .... ............................
. .............................
. .
. .............
. ...........
.........
....
.
.
.
.
.................................................................................................................................................................................................................................................................................................................................................................................
....
...
.........
..............
. ...............
. .................
. . ...............
. .... ............................
. . ....... ......................
. . ..... ............................
. .
....... ........................
..... ..... .................
. .
.....
.. .......
. .
. .
....... ...........................................
. . .
.. .......
...
....... .................................................
. . .
............
....... .......
. . .
....... ....... ...............................................
. . .
..........................
....... .......
. . .
...........................
....... .......
. . .
...
....... .......
. . .
....... .......
. . .
.......
. . . .
. . . .
. . . .
. . . .
.

20

10

Wt

10

20

10

20

30

40

50

tk
............................................

90%

....... ....... ....... .......

95%

. . . . . . . .

99%

Figure 4b. CUSUMSQ test

1.0

0.8

0.6
St
0.4

0.2

0.0

....................................................
. ...........................
...... ..
. . ...............
...... ....
. . ...................
......
.
. . ..........................
......
.
.....
. .................
......
.
.
.
.
.
.
.
....
.
. .... ....
......
.......
. . .........................
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.... ....
.. .
......
. . .... ...............
... ..
......
......
... ...
. . .... ...............
.... .
......
. . ..... ...............
.......... .
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. ....
.
...
......
. . ..................
..... .. .
......
..... .... .
......
. . ..................
.....
......
.
..... ...... . .
......
..............
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. ....
.... .
......
.
.................
..... ..... . .
......
.
.
.....
......
.
.................
.
..... ....
......
.
.
....... ..
......
.
.
.
...............
.
.
.
.
.
.
.
.
.
.
. .
.
. ...
......
..... .... . .
.
................
......
..........
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.... ..
...... ......
.
......
.
..... ...
......
. ......
.
...... .... . .
......
................
.
...... ..
......
.
..... .... . .
......
.
.
.
................
.
.
.
.
.
.
.
.
.
.
.
.
. ...
..... ......
......
.
..... .
.
.............
......
.
...... .....
.
......
.. ....
.
...... .
......
.
..............
.
...... ......
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.... .....
...... .....
......
.
. ..... .........
...... ..
.
......
......
...... ....
.
.. ....
.
......
...... ..
.
. ..... .........
......
...... ...... . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
...... ...... . .
......
. ..... .......
...... .
......
....... ...... . .
......
. .........
....... .
......
. .............
........ ..... . .
......
.
.
.
.
.
.
.
.
.
.
.
.
. . ......
....... .... . . .
......
....... .....
......
.........
........ .. . .
......
........ ......
......
.......
............... . . .
......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.. ..
...
.....
........ ..... . . .
......
..............
......
......
.......... . . .
......
......... .... . .
......
......
....................... .
.
.
.
.
. ...........
.
.
.
.
.
.
.
.
.
.
. ...
................ .
................................................................... .

10

20

30

40

50

tk
............................................

90%

....... ....... ....... .......

12

95%

. . . . . . . .

99%

where we examine testing not only a shift in location but also that in scale. Then we compute
each of the test statistics, and record whether or not rejection occurs for tests with each
size. The experiment is repeated M times, and then compute the sample powers (for each
statistic and the test size) as the relative q
rejection frequency. We can obtain the estimated

standard errors for the sample powers as pb(1 pb) M , where pb denotes the sample power.
Since we perform the experimentp10,000 times, i.e., M =10,000, the estimated standard
errors are enough small (at most .5(1 .5)/10, 000 0.005).
, T k, and are chosen as follows:
= 0.10, 0.05, 0.01,
T k = 10, 20, 30, 40, 50,
= 0.0, 0.5, 1.0, 1.5,
= 1.0, 1.5, 2.0, 2.5.
In order to obtain the simulated confidence intervals, m sets of random draws are generated,
i.e., m =20,000 as in Section 3. In this section, different M sets of normal random draws
are generated to examine the power for both CUSUM and CUSUMSQ tests. The results
are in Tables 4a 4c and 5a 5c. In the case of =0.0 and =1.0, theoretically the
sample power should be equal to the significance level . However, for the conventional
90% confidence interval of the CUSUM test, the sample power is .0236 when T k =10,
.0474 when T k =20, .0572 when T k =30, .0633 when T k =40, and .0690 when
T k =50 (see Table 4a, Conventional 90% Confidence Interval, Case: =0.0 and =1.0).
This implies that the conventional confidence interval of the CUSUM test gives us a poor
approximation even if T k is large. For the conventional 90% confidence interval of the
CUSUMSQ test, the power is .1445 when T k =10, .1253 when T k =20, .1205 when
T k =30, .1133 when T k =40, and .1133 when T k =50 (see Table 5a, Conventional
90% Confidence Interval, Case: =0.0 and =1.0), which implies that the power of the
confidence interval of the CUSUMSQ test is too large.
The other findings are as follows. The CUSUM test is more powerful in the location
parameter while the CUSUMSQ test is more powerful in the scale parameter . Both
tests are powerful as T k is large for almost all the cases. Given , the CUSUM test is less
powerful as is large. Therefore, the CUSUM test is sensitive to the structural change of
the shifting parameter and the CUSUMSQ test should be used in testing heteroscedasticity.

SUMMARY

In this paper, the asymptotically exact confidence intervals for the CUSUM and CUSUMSQ
tests were derived using the simulation technique. As the number of simulation (i.e., m
and/or n) increases, the approach proposed in this paper clearly gives us the asymptotically
exact confidence interval. The main results are as follows: (i) for the conventional confidence
intervals, the power of the CUSUM test is too low. The confidence interval of the test gives
us a poor approximation even when T k goes to large (see T k = 50), and the power of the
CUSUMSQ test is too large, (ii) the CUSUM test is more sensitive to the location parameter
than the CUSUMSQ test while the latter is more sensitive to the scale parameter than
the former, and (iii) given , the CUSUM test is poor as increases.
Finally, even in the case where the recursive residuals are not normal, the simulation
technique to obtain the confidence intervals of the CUSUM and CUSUMSQ tests can be
easily applied. That is, the asymptotically exact confidence intervals proposed in this paper
can be derived for any regression model with the nonnormal disturbance. However, note
that the proposed confidence intervals have larger simulation errors for large sample size
T k and small significance level than for small T k and large .
13

Table 4a. CUSUM Test ( = 0.10)

Conventional Interval
1.0
1.5
2.0
2.5

0.0
0.5
1.0
1.5

.0236
.0244
.0298
.0411

.0109
.0119
.0141
.0185

.0057
.0066
.0074
.0092

0.0
0.5
1.0
1.5

.0474
.0713
.1840
.3722

.0229
.0405
.1034
.2266

.0151
.0246
.0612
.1374

0.0
0.5
1.0
1.5

.0572
.1220
.3762
.7118

.0297
.0686
.2301
.5153

.0207
.0462
.1435
.3409

0.0
0.5
1.0
1.5

.0633
.1697
.5670
.8931

.0325
.1013
.3764
.7301

.0248
.0651
.2436
.5398

0.0
0.5
1.0
1.5

.0690
.2176
.7096
.9699

.0337
.1289
.4931
.8700

.0250
.0839
.3237
.6907

T k
.0032
.0038
.0038
.0049
T k
.0109
.0176
.0405
.0870
T k
.0168
.0349
.0966
.2281
T k
.0208
.0482
.1635
.3850
T k
.0227
.0593
.2205
.5123

14

1.0
= 10
.0991
.1047
.1367
.1866
= 20
.0948
.1330
.2779
.5124
= 30
.0991
.1710
.4542
.7812
= 40
.0955
.2061
.6118
.9138
= 50
.1008
.2496
.7338
.9747

Exact Interval
1.5
2.0

2.5

.0509
.0578
.0806
.1177

.0340
.0382
.0517
.0722

.0241
.0267
.0357
.0505

.0440
.0717
.1779
.3523

.0285
.0486
.1120
.2338

.0217
.0357
.0777
.1606

.0414
.0913
.2847
.5958

.0267
.0618
.1833
.4162

.0226
.0473
.1264
.2878

.0406
.1172
.4119
.7661

.0282
.0740
.2733
.5811

.0240
.0548
.1863
.4202

.0385
.1386
.5126
.8815

.0254
.0896
.3388
.7105

.0226
.0619
.2332
.5311

Table 4b. CUSUM Test ( = 0.05)

Conventional Interval
1.0
1.5
2.0
2.5

0.0
0.5
1.0
1.5

.0037
.0044
.0039
.0034

.0017
.0010
.0012
.0012

.0008
.0003
.0005
.0003

0.0
0.5
1.0
1.5

.0190
.0296
.0877
.1914

.0074
.0131
.0406
.0930

.0034
.0068
.0198
.0464

0.0
0.5
1.0
1.5

.0250
.0602
.2346
.5337

.0103
.0310
.1235
.3244

.0066
.0176
.0686
.1842

0.0
0.5
1.0
1.5

.0294
.0934
.4121
.7901

.0131
.0483
.2367
.5752

.0087
.0312
.1331
.3735

0.0
0.5
1.0
1.5

.0336
.1260
.5661
.9182

.0142
.0677
.3434
.7561

.0105
.0395
.2035
.5370

T k
.0003
.0002
.0002
.0001
T k
.0028
.0040
.0102
.0247
T k
.0045
.0115
.0429
.1102
T k
.0076
.0212
.0787
.2341
T k
.0086
.0282
.1275
.3532

15

1.0
= 10
.0494
.0501
.0686
.0952
= 20
.0482
.0689
.1750
.3571
= 30
.0477
.0919
.3165
.6437
= 40
.0463
.1217
.4736
.8386
= 50
.0510
.1498
.6074
.9356

Exact Interval
1.5
2.0

2.5

.0225
.0262
.0344
.0502

.0140
.0169
.0207
.0292

.0092
.0103
.0145
.0178

.0198
.0343
.0924
.2125

.0117
.0200
.0539
.1250

.0083
.0140
.0346
.0772

.0172
.0470
.1781
.4352

.0120
.0305
.1046
.2714

.0099
.0215
.0699
.1667

.0177
.0606
.2874
.6428

.0124
.0390
.1688
.4359

.0107
.0273
.1025
.2893

.0168
.0792
.3785
.7926

.0114
.0467
.2267
.5765

.0098
.0311
.1445
.3899

Table 4c. CUSUM Test ( = 0.01)

Conventional Interval
1.0
1.5
2.0
2.5

0.0
0.5
1.0
1.5

.0000
.0000
.0000
.0000

.0000
.0000
.0000
.0000

.0000
.0000
.0000
.0000

0.0
0.5
1.0
1.5

.0009
.0021
.0049
.0136

.0001
.0005
.0016
.0031

.0000
.0002
.0004
.0013

0.0
0.5
1.0
1.5

.0024
.0085
.0448
.1491

.0003
.0032
.0148
.0530

.0002
.0012
.0059
.0190

0.0
0.5
1.0
1.5

.0046
.0160
.1299
.4152

.0013
.0059
.0461
.2012

.0008
.0027
.0194
.0820

0.0
0.5
1.0
1.5

.0042
.0278
.2438
.6672

.0014
.0117
.1032
.3892

.0007
.0064
.0416
.1919

T k
.0000
.0000
.0000
.0000
T k
.0000
.0000
.0002
.0003
T k
.0001
.0004
.0027
.0083
T k
.0006
.0014
.0087
.0370
T k
.0007
.0035
.0203
.0944

16

1.0
= 10
.0093
.0117
.0142
.0177
= 20
.0097
.0146
.0511
.1218
= 30
.0090
.0231
.1235
.3361
= 40
.0104
.0343
.2282
.5874
= 50
.0097
.0471
.3370
.7717

Exact Interval
1.5
2.0

2.5

.0044
.0046
.0064
.0079

.0021
.0027
.0027
.0031

.0013
.0014
.0013
.0016

.0022
.0053
.0190
.0505

.0012
.0027
.0075
.0215

.0010
.0016
.0037
.0094

.0022
.0103
.0504
.1650

.0011
.0057
.0240
.0784

.0007
.0036
.0116
.0408

.0033
.0147
.0997
.3428

.0019
.0075
.0431
.1738

.0015
.0046
.0255
.0869

.0025
.0198
.1592
.5175

.0017
.0105
.0765
.2825

.0012
.0075
.0378
.1563

Table 5a. CUSUMSQ Test ( = 0.10)

Conventional Interval
1.0
1.5
2.0
2.5

0.0
0.5
1.0
1.5

.1445
.1522
.1909
.2858

.2436
.2651
.3389
.4363

.4153
.4348
.4873
.5692

0.0
0.5
1.0
1.5

.1253
.1343
.2398
.4715

.3250
.3775
.5133
.6905

.6239
.6578
.7414
.8365

0.0
0.5
1.0
1.5

.1205
.1443
.3163
.6366

.4190
.4904
.6584
.8505

.7825
.8113
.8773
.9421

0.0
0.5
1.0
1.5

.1133
.1397
.3729
.7673

.4993
.5825
.7669
.9301

.8723
.8984
.9466
.9835

0.0
0.5
1.0
1.5

.1133
.1474
.4417
.8602

.5712
.6600
.8483
.9675

.9343
.9478
.9777
.9949

T k
.5808
.5886
.6238
.6832
T k
.8289
.8417
.8754
.9200
T k
.9374
.9452
.9635
.9836
T k
.9800
.9831
.9898
.9960
T k
.9945
.9961
.9973
.9991

17

1.0
= 10
.0977
.1003
.1256
.1891
= 20
.0989
.1037
.1697
.3304
= 30
.0993
.1114
.2249
.4884
= 40
.1007
.1171
.2629
.6130
= 50
.1039
.1188
.3089
.7291

Exact Interval
1.5
2.0

2.5

.1700
.1891
.2418
.3220

.3067
.3215
.3726
.4439

.4549
.4642
.5013
.5573

.2430
.2856
.3998
.5617

.5120
.5467
.6303
.7421

.7324
.7505
.8018
.8580

.3234
.3868
.5386
.7384

.6770
.7103
.7967
.8921

.8878
.8955
.9258
.9594

.3904
.4597
.6485
.8553

.7897
.8166
.8924
.9586

.9542
.9574
.9764
.9900

.4464
.5311
.7423
.9235

.8661
.8967
.9481
.9847

.9852
.9871
.9942
.9973

Table 5b. CUSUMSQ Test ( = 0.05)

Conventional Interval
1.0
1.5
2.0
2.5

0.0
0.5
1.0
1.5

.0722
.0728
.0940
.1339

.1312
.1480
.1898
.2501

.2445
.2581
.2990
.3591

0.0
0.5
1.0
1.5

.0621
.0717
.1398
.2979

.2058
.2459
.3639
.5288

.4746
.5118
.5977
.7163

0.0
0.5
1.0
1.5

.0610
.0742
.1988
.4768

.2933
.3584
.5218
.7359

.6655
.7001
.7927
.8885

0.0
0.5
1.0
1.5

.0609
.0797
.2463
.6243

.3730
.4490
.6511
.8632

.7900
.8222
.8986
.9630

0.0
0.5
1.0
1.5

.0578
.0766
.3072
.7513

.4405
.5331
.7572
.9335

.8786
.9040
.9528
.9874

T k
.3723
.3770
.4099
.4563
T k
.7086
.7291
.7783
.8410
T k
.8862
.8959
.9254
.9597
T k
.9580
.9634
.9784
.9912
T k
.9862
.9898
.9952
.9978

18

1.0
= 10
.0490
.0483
.0646
.0989
= 20
.0487
.0563
.0974
.2060
= 30
.0493
.0571
.1329
.3442
= 40
.0504
.0615
.1684
.4735
= 50
.0551
.0632
.2040
.6043

Exact Interval
1.5
2.0

2.5

.0943
.1109
.1466
.2047

.1965
.2077
.2488
.3094

.3161
.3281
.3620
.4133

.1526
.1823
.2723
.4214

.3785
.4181
.5006
.6181

.6188
.6387
.6981
.7749

.2198
.2705
.4142
.6220

.5589
.5969
.6984
.8226

.8170
.8324
.8762
.9269

.2743
.3407
.5285
.7684

.6897
.7308
.8236
.9248

.9167
.9268
.9541
.9787

.3287
.4077
.6351
.8705

.7910
.8306
.9071
.9683

.9684
.9739
.9860
.9952

Table 5c. CUSUMSQ Test ( = 0.01)

Conventional Interval
1.0
1.5
2.0
2.5

0.0
0.5
1.0
1.5

.0175
.0175
.0169
.0205

.0328
.0374
.0478
.0592

.0695
.0743
.0857
.1003

0.0
0.5
1.0
1.5

.0126
.0154
.0295
.0627

.0589
.0745
.1195
.1812

.1782
.2005
.2524
.3331

0.0
0.5
1.0
1.5

.0137
.0150
.0516
.1649

.1021
.1387
.2385
.4102

.3666
.4051
.5039
.6446

0.0
0.5
1.0
1.5

.0146
.0193
.0772
.2897

.1487
.1984
.3632
.6127

.5369
.5829
.7035
.8393

0.0
0.5
1.0
1.5

.0115
.0185
.1043
.4392

.2049
.2713
.4901
.7732

.6757
.7250
.8360
.9372

T k
.1121
.1154
.1296
.1426
T k
.3451
.3672
.4164
.4944
T k
.6543
.6769
.7372
.8201
T k
.8343
.8538
.8984
.9488
T k
.9321
.9426
.9667
.9874

19

1.0
= 10
.0112
.0102
.0125
.0231
= 20
.0086
.0087
.0233
.0650
= 30
.0106
.0107
.0347
.1344
= 40
.0103
.0138
.0539
.2208
= 50
.0101
.0154
.0675
.3309

Exact Interval
1.5
2.0

2.5

.0224
.0258
.0386
.0592

.0607
.0657
.0806
.1084

.1151
.1214
.1388
.1678

.0472
.0616
.1012
.1784

.1691
.1877
.2482
.3442

.3523
.3773
.4373
.5249

.0737
.1014
.1900
.3539

.3123
.3496
.4500
.5984

.6057
.6294
.6969
.7917

.1032
.1436
.2815
.5212

.4498
.4960
.6233
.7826

.7747
.7991
.8575
.9216

.1403
.1957
.3803
.6798

.5790
.6294
.7654
.8957

.8923
.9091
.9408
.9742

REFERENCES
Brown, R.L., Durbin, J. and Evans, J.M. (1975). Techniques for Testing the Constancy of
Regression Relationships over Time, Journal of the Royal Statistical Society, Ser.B,
vol.37, pp.149 192.
Durbin, J. (1969). Tests for Serial Correlation in Regression Analysis based on the Periodogram of Least-Squares Residuals, Biometrica, vol.56, pp.1 15.
Diebold, F.X. and Rudebusch, G. (1991). Unbiased Estimation of Autoregressive Models,
unpublished manuscript.
Galpin, J. and Hawkins, D. (1984). The Use of Recursive Residual in Checking Model Fit
in Linear Regression, The American Statistician, vol.38, pp.94 105
Garbade, K. (1977). Two Methods for Examining the Stability of Regression Coefficients,
Journal of the American Statistical Association, vol.72, pp.54 63.
Greene, W.H. (1990). Econometric Analysis, Macmillan Publishing Company, Chap.7,
pp.224 227.
Harvey, A.C. (1981) Time Series Models, Philip Allen Publishers Limited, Oxford.
Harvey, A.C. (1989) Forecasting, Structural Time Series Models and the Kalman Filter,
Cambridge University Press.
Johnston, J. (1984) Econometric Methods, McGraw-Hill Book Company, pp.384 392.
Kramer, W. (1989) The Robustness of the Chow Test to Autocorrelation among Disturbances, Statistical Analysis and Forecasting of Economic Structural Change (P.
Hackl, Eds.). Springer-Verlag.
McFadden, D. (1989) A Method of Simulated Moments for Estimation of Discrete Response
Models without Numerical Integration, Econometrica, Vol.57, No.5, pp.995 1026.
Ploberger, W. (1989) The Local Power of the CUSUM-SQ Test against Heteroscedasticity, Statistical Analysis and Forecasting of Economic Structural Change (P. Hackl,
Eds.). Springer-Verlag.
Ploberger, W., Kramer, W. and Alt, R. (1989). A Modification of the CUSUM Test in
the Linear Regression Model with Lagged Dependent Variables, Econometrics of
Structural Change (W. Kramer, Eds.). Physica-Verlag Heidelberg.
Riddell, W.C. (1975). Recursive Estimation Algorithms for Economic Research, Annals
of Economic and Social Measurement, vol.4, pp.397 406.
Westlund, A.H. and Tornkvist, B. (1989). On the Identification of Time for Structural
Changes by MOSUM-SQ and CUSUM-SQ Procedures, Statistical Analysis and Forecasting of Economic Structural Change (P. Hackl, Eds.). Springer-Verlag.

20