Escolar Documentos
Profissional Documentos
Cultura Documentos
1
(L) y
t
x
1, t
;
2
(L) y
t
x
2, t
;
(1 L
2
)Ly
t
x
3, t
;
(1 L
2
)L
2
y
t
x
4, t
; (1)
where L is a lag operator and
1
(L) (1 L L
2
L
3
)L;
2
(L) (1 L L
2
L
3
)L: (2)
The unit root hypotheses may be tested by estimating the regression
3
4
y
t
0
X
4
i1
i
x
i, t
X
p
i1
4
y
ti
e
t
, (3)
where
4
1 L
4
:
If there is a root at the zero frequency then
1
0. If there is a seasonal
root at 1,
2
0. If there are seasonal roots at the complex values,
3
Note that in the model presented here,
3
and
4
are dened in reverse to the denitions of
Hylleberg et al. (1990).
570 BULLETIN
#Blackwell Publishers 1999
3
0, and
4
0. Hylleberg et al. recommend that the rst two hypo-
theses are tested using a one sided t-test. The alternatives are
1
,0,
2
,0. When testing for complex roots these authors recommend either a
joint F-test of
3
4
0, against the alternative of
3
6 0 and/or
4
6 0. Alternatively, they propose a two tailed test for
3
0, and on the
failure to reject this hypothesis then a one tailed test for
4
0 against
4
,0. Hylleberg et al. generate the two tailed critical values statistics for
the t-statistics for
3
0. However, it is easier to generate an F-statistic for
this hypothesis, thereby restricting the test to one which is automatically
one tailed. In what follows, t
i
or F
i
will be used to denote the t-statistic or
F-statistics for the hypothesis that
i
0 and F
ij
will be used to denote the
F-statistic for the hypothesis that
i
j
0.
As with the Dickey-Fuller (D-F) test the distribution of t and F-statistics
are non-standard. Hylleberg et al. have tabulated the relevant critical values,
where seasonal dummies and trends are also included. Beaulieau and Miron
(1993) outline the analogous test for monthly series and give the appropriate
critical values.
III. ALTERNATIVE MODELS
Zivot and Andrews (1992), and Perron (1994) consider two types of
alternative specications: the `additive outlier model' and, the `innovational
outlier model'. In this section, the analogous models are presented for the
case where the data may contain seasonal roots. Two alternative forms
which include seasonality but which are analogous to the `additive outlier'
specication are considered below.
y
t
0
0
D
t
X
3
i1
i
s
i, t
(4)
D
t
X
3
i1
i
s
i, t
u
t
;
and,
y
t
0
0
D
t
X
3
i1
i
s
i, t
(5)
D
t
X
3
i1
i
s
i, t
t D
t
t u
t
:
The additional variables are dened as:
T total sample size;
s
i, t
ith seasonal dummy at time t;
SEASONAL UNIT ROOT TESTS WITH STRUCTURAL BREAKS 571
#Blackwell Publishers 1999
D
t
0 if t ,hTi; and,
D
t
1 if t > hTi where hTi is the integer part of T, and 0:15 ,
,0:85.
Furthermore,
4
u
t
X
4
i1
i
z
i, t
X
p
i1
4
u
ti
e
t
(6)
where (using the denitions in Equation Set 2):
1
(L)u
t
z
1, t
;
2
(L)u
t
z
2, t
;
(1 L
2
)Lu
t
z
3, t
; (7)
(1 L
2
)L
2
u
t
z
4, t
;
4
1 L
4
:
The rst of these models allows for a break in the deterministic seasonal
behaviour of the data, along with a break in the mean, or mean innovation,
of the process. The second model allows for a changing linear trend. Once
these components are extracted the residual u
t
has a polynomial which
has potential roots at 1, 1, i and i. The standard HEGY procedure
tests the null hypotheses that
i
0, i 1, 2, 3, 4, under the auxiliary
assumptions
o
0,
0,
j
0, j 1, 2, 3.
If it is known where the structural break occurs, then this could easily be
accommodated by running the regressions on Equations 4 and 5, extracting
the residuals and estimating Equation 6. The tests for seasonal roots could
be conducted in the manner described above, except that the distribution of
the test statistics will be altered. This is analogous to the tests suggested by
Perron. Perron (1989, 1990) demonstrated that the distribution of the t-tests
are dependent upon the point at which the structural break is introduced.
Zivot and Andrews (1992) developed a test where the point of the structural
break is not known. They proposed a test which was based on the minimum t-
statistic, from a series of regressions placing the break point at every point in
the data. This paper suggests that the same approach can be taken in testing
for seasonal roots. The following procedure can be conducted:
Run the regressions 4 or 5 for the integer parts of T for all values of
0:15 < < 0:85; and,
Take the minimum t and maximum F-statistics:
a. inf
(t
i
()), i 1, 2, 4; and,
b. sup
(F
3,4
()), and sup
F
3
()
572 BULLETIN
#Blackwell Publishers 1999
from the regression of residuals in Equation 6. From hereon, these
statistics will be denoted as min(t
i
) max(F
3
) and max(F
34
).
The `innovational outlier' versions of models in Equations 4 and 5 have
the deterministic variables entering directly into the autoregressive repre-
sentation of the original variable. Sequential regressions are performed (for
each integer value hrTi) of the form
4
y
t
^
0
^
0
D
t
X
3
i1
^
i
s
i, t
X
4
i1
^
i
x
i, t
D
t
X
3
i1
^
i
s
i, t
^
t D
t
^
t ^u
t
, (8)
for all 2 (:15, :85):
where, as before D
t
0 if t ,hTi; and, D
t
1 if t > hTi where hTi is
the integer part of T.
The regression in Equation 8 corresponds to the model in Equation 5. In
order to obtain the corresponding model to Equation 4, the trends would be
removed.
Zivot and Andrews (1992) proved that the minimum statistics have
limiting distributions in the case of zero frequency unit root tests. Here, it is
suggested that the minimum (or maximum) statistics have limiting distribu-
tions in the tests for seasonal roots. Under the assumption that these limiting
distributions exist, the critical values can be generated using Monte-Carlo
trials. Section V presents the critical values generated from Monte-Carlo
trials for the sequential statistics (min(t
i
), max(F
3
) and max(F
34
)) using the
innovational outlier models in Equation 8 under the null hypothesis that
4
y
t
e
t
, e
t
I N( ,
2
).
IV. THE IMPACT OF STRUCTURAL CHANGE ON HEGY TESTS
This section examines the power of seasonal root tests where the data is
generated by a process:
y
t
P
4
i1
i
s
i, t
:D
t
P
4
i1
i
s
i, t
e
t
;
e
t
I N(0, 1);
(
1
,
2
,
3
,
4
) (0, 1, 1, 2); (9)
:5, T 120; and,
0, 1, 2:
The data generating process in Equation 9 has a unit root at the zero
frequency but no unit roots at seasonal frequencies. At 0 there is no
alteration in the seasonality. At 2 there is a full switching of the
SEASONAL UNIT ROOT TESTS WITH STRUCTURAL BREAKS 573
#Blackwell Publishers 1999
deterministic seasonality (a full reversal of seasons) in the middle of the
sample period.
The HEGY tests were run, including seasonal dummies and a trend in the
auxiliary regression. Using the critical values in Hylleberg et al., the
proportion of observations which gave test statistics for t
2
0
, and F
3
\
4
0
rejecting the null hypothesis of a seasonal root, at the 5 percent levels are in
Table 1 (i.e. it is the proportion of time an incorrect hypothesis is rejected at
a 5% of signicance).
The results demonstrate that with a large sample (30 years of quarterly
data) seasonal root tests hasve high power to reject the hypothesis of
seasonal roots. However, the power of these tests falls as the size of the
structural break increases. Once there is an actual switch in the `direction'
of the seasonality, the power of seasonal root tests quickly becomes low.
The full `switching case' renders the test almost completely without power
at sample sizes of 40 or less, and the power is poor at sample sizes of 80
and greater.
V. CRITICALVALUES FORTHE SEQUENTIALTEST
The asymptotic convergence of the HEGY type statistics is proved in Chan
and Wei (1988) along with Beaulieau and Miron (1993), and the conver-
gence of the sequential unit root test is proved in Zivot and Andrews (1992).
There has not, as yet, been any proof of the convergence of the test statistics
from a sequential HEGY test. This section demonstrates, using Monte Carlo
evidence, that the sequential unit root tests do have asymptotically con-
vergent distributions. The null model used to generate the data is
4
y
t
e
t
, where e
t
I N( ,
2
): (10)
The size of the variance of the process (
2
6 0) will not effect the
asymptotic distribution of the test statistics, as it simply plays the role of a
TABLE 1
Power of HEGY Tests Under Structural Breaks
alpha Stat T 40 T 80 T 120
0 t
2
.44 .91 .99
F
34
.70 .97 .99
1 t
2
.12 .48 .76
F
34
.14 .48 .69
2 t
2
.03 .20 .56
F
34
.01 .06 .40
Notes: The null hypotheses
2
0, and
3
4
0 (in Equation 3) are being tested against the
alternatives that j
2
j 6 0; and
3
6 0 or
4
6 0 respectively. The DGP is as in Equation 9,
with no seasonal unit roots, but changes in the values of the deterministic seasonality.
574 BULLETIN
#Blackwell Publishers 1999
scaling parameter. However, as in the case of the standard D-F (Dickey and
Fuller, 1979) test, the value of (henceforth referred to as the `drift') may
affect the large sample distributions of the test statistics if a trend term is
not added to the regression. Therefore, the trials are conducted under three
scenarios:
Case 1) constant in the regression no drift or trend in DGP;
Case 2) constant in the regression drift but no trend in DGP;
Case 3) constant trend in the regression no drift or trend in DGP;
Case 4) constant trend in the regression drift but no trend in DGP.
These four data generating processes represent only a small subset of the
possible DGPs which could be explored (i.e. trends and unit roots in the
DGP and roots at some frequencies and not others). It was felt that these
DGPs encompassed most of the practically interesting cases, and conformed
with the values presented in Hylleberg et al. (1990). The cases with and
without drift were included so to demonstrate that the trend would effect
only the zero frequency statistics as in the case of the standard Dickey-
Fuller case, unless a trend term was introduced although Hylleberg et al.
(1990) did not consider this in their paper.
4
The asymptotic critical values for the sequential tests are presented in
Table 2. These values are calculated from 10,000 trials for each sample size
of T 40, 80, 120 and 160. Using the method advocated in Gregory and
Hansen (1996), the critical values presented in Table 2 are calculated as
Crit(T, P)
0
1
T
1
error, (11)
where T is the sample size and P is the level of condence (99, 97.5, 95
and 90). The standard errors were also computed, and while they are not all
presented here, the maximum standard error for any value in each column is
listed in the last row of Table 2.
Examining Cases 3 and 4 in Table 2, the critical values test for
1
0,
using the sequential test, are dependent on whether 0 if there is no
trend in the regression. The tests are invariant (asymptotically) to the value
of if a trend term is added to the regression. While there are considerable
differences between the critical values in Cases 1 and 2 in Table 2, the
inclusion of a trend negates the effect of a drift in the generating process.
Inspecting all the third to last columns of Table 2, the tests for seasonal
roots
2
0,
3
0,
3
0 \
4
0, appear to be invariant to the value
of .
4
Although the results are not presented here, additional trials were conducted using alternative
values of . The results were insignicantly different from those presented here (where 1)
indicating that asymptotically at least, it is only the existence of a drift term that matters, not its
value. Readers should be mindful of the fact that if DGP has a trend as well as a unit root and
drift, the critical values will change.
SEASONAL UNIT ROOT TESTS WITH STRUCTURAL BREAKS 575
#Blackwell Publishers 1999
(i) Power of Sequential Tests
The power values for the sequential tests generated under the same DGPs as
in Table 1 are presented in Table 3. It is evident that the relative power of
the sequential tests are moderately reduced in small sample sizes (40 or
less) where there are no structural breaks. However, in samples of 80 or
TABLE 2
Asymptotic Critical Values
min(t
1
) min(t
2
) max(F
3
) min(t
4
) max(F
34
)
1
0
2
0
3
0
4
0
3
4
0
99 4.78 4.78 13.60 5.44 16.10
97.5 4.46 4.47 11.32 5.13 14.53
95 4.24 4.23 9.13 4.83 13.13
90 3.97 3.95 7.29 4.61 11.70
case 1
99 3.53 4.69 13.98 5.41 15.86
97.5 3.23 4.40 11.40 5.16 14.34
95 2.98 4.19 9.51 4.85 12.97
90 2.69 3.92 7.49 4.57 11.60
case 2
99 5.52 4.78 13.38 5.45 15.91
97.5 5.24 4.44 10.84 5.14 14.32
95 4.99 4.21 9.05 4.87 13.03
90 4.71 3.94 7.17 4.59 11.61
case 3
99 5.54 4.77 13.39 5.44 15.93
97.5 5.25 4.46 10.94 5.13 14.37
95 4.97 4.21 9.02 4.88 13.12
90 4.71 3.95 7.12 4.60 11.69
case 4
Max SE .03 .03 .21 .03 .15
TABLE 3
Power of Sequential Tests Under Structural Breaks
alpha Stat T 40 T 80 T 120
0 min(t
2
) .32 .84 .98
max(F
34
) .65 .97 .99
1 min(t
2
) .32 .85 .99
max(F
34
) .67 .99 1.00
2 min(t
2
) 1.00 1.00 1.00
max(F
34
) .73 .99 1.00
Notes: The null hypothesis
2
0, and
3
4
0 (in Equation 8) are being tested against the
alternatives that j
2
j 6 0; and,
3
6 0 or
4
6 0 respectively. The DGP is as in Equation 9,
with no seasonal unit roots, but changes in the values of the deterministic seasonality.
576 BULLETIN
#Blackwell Publishers 1999
more the loss in power is small and almost non-existent in samples of 120
or more. Conversely, where the data contain structural breaks there are quite
dramatic improvements in the power of the tests. For instance, taking the
case where T 80 and 1, the probability of correctly rejecting a
seasonal unit root is .48 when using either the t and F-statistics generated
by the standard HEGY procedure, whereas for the sequential tests, this has
increased to over .85 and .99 for the t and F-statistics respectively.
(ii) Note on Lag Length Selection
In empirical work, the selection of the lag length in the HEGY regressions
can be of critical importance. Here, a sequential procedure is used to select
the lag length in the HEGY regression. An upper value of 12 lags was
selected. In both the standard and sequential tests, the procedure iteratively
eliminated the last lag in the regression if it was insignicant (if jtj ,1:64).
Perron (1994) recommends the use of this approach over other methods,
such as the Akaike Information Criteria.
VI. APPLICATION TO FOOD PRICES IN THE U.S.
One rather obvious potential source of seasonality is the weather, which
inuences the price of agricultural commodities in particular, due to their
characteristically limited potential for storage, and dependence on exogen-
ous factors such as temperature, rainfall, number of sunlight hours and so
forth. As Hylleberg (1994) notes, changes in production techniques storage
possibilities, biological innovations etc. have the potential to change sea-
sonal patterns. The seasonal behaviour of a number of agricultural commod-
ities are examined in this section.
(i) Data
The data used in this section are ten price series published by the United
States Department of Agriculture (USDA) in Agricultural Price Summaries.
The data are quarterly prices for U.S. agricultural outputs. The quarters used
are January, April, July and October. The data are from the rst quarter of
1965 to the last quarter in 1991. The prices have been deated using the
U.S. retail price index, and all the real prices have been logged.
(ii) Results
Standard HEGY Tests
The results of the standard seasonal root tests are presented in Table 4. The
regressions include a constant and time trend, therefore the alternative is
trend stationarity in the case where
1
0 is being tested. Examining the
results in Table 4, it is apparent that only two of the series (Vegetables and
SEASONAL UNIT ROOT TESTS WITH STRUCTURAL BREAKS 577
#Blackwell Publishers 1999
Fresh Fruit) reject a unit root at the zero frequency at the 5 percent level of
signicance. These two series also reject seasonal roots at the 10 percent
level and below.
Four of the series fail to reject seasonal roots for all frequencies at the 5
percent level of signicance (All Crops, Fresh Fruit, Tobacco and Meat
Animals). The All Crops series rejects
3
4
0. Fresh Fruit rejects
2
0 but not
3
4
0. The other two series (Tobacco and Meat
Animals) fail to reject seasonal unit roots at any of the seasonal frequencies.
Sequential HEGY Tests
The results of the seasonal root tests are presented in Table 5. All the
variables reject unit roots at all the seasonal frequencies at the 10 percent
level of signicance or below, except for the All Crops variable which fails
to reject
2
0. The All Crops variable also failed to reject unit roots using
the standard HEGY test. In addition several of the variables now reject unit
roots at the zero frequency (Vegetables, Poultry and Eggs, Cotton and Fresh
Fruit). Only Fresh Fruit rejected a unit root at the zero frequency using the
standard tests.
Overall the conclusions which are reached using the sequential tests are
substantially different from those derived from the standard HEGY proce-
dure. As was expected, the sequential tests also rejected seasonal roots for
all those series which have the seasonal root hypothesis rejected using the
standard test. However, in general the results of the sequential tests give less
credence to the hypothesis that the data contain seasonal unit roots.
TABLE 4
Standard Seasonal Root Tests
Null Hypothesis
Prices
1
0
2
0
3
4
0
Lag
length
Farm Produce 1.89 3.82
15.30
0
Feed Grain 2.77 7.02
28.31
0
All Crops 2.22 2.52
20.08
7
Vegetable 3.74
5.11
28.31
0
Poultry Egg 2.64 4.02
24.65
2
Cotton 1.43 3.64
42.8
2
Fresh Fruit 3.89
2.84
5.83 12
Tobbacco 2.33 2.08 5.21 12
Dairy 1.14 3.12
17.70
12
Meat Animals 2.47 2.24 5.41 2
Crit Vals 5% (10%) 3.53 (3.22) 2.94 (2.63) 6.60 (5.52)
Notes:
48.40
4
Feed Grain 3.96 7.00
37.45
0
All Crops 4.56 3.04
28.19
7
Vegetable 5.55
6.78
44.00
12
Poultry Egg 5.09
7.45
38.16
12
Cotton 5.87
5.59
73.14
2
Fresh Fruit 5.60
4.02
30.97 7
Tobbacco 4.49 4.10
22.68 12
Dairy 4.35 5.95
34.94
1
Meat Animals 3.91 5.39 33.06
7
Crit Vals 5% (10%) 4.97 (4.71) 4.21 (3.95) 13.12 (11.69)
Notes:
19.66
9
Sequential 3.40 4.56
22.71
9
RPW Standard .89 2.15 15.50
7
Sequential 4.34 6.03
56.36
7
EP Standard 1.85 3.88
17.89
10
Sequential 3.39 4.71
22.17
10
FP Standard 1.59 7.37
36.90
5
Sequential 4.81
7.72
47.97
10
IP Standard 1.46 6.08
53.16
0
Sequential 3.79 6.08
58.05
0
EGAP Standard 2.62 3.83
18.5
9
Sequential 3.57 4.64
21.60
9
INTEREST Standard 2.89 6.92
43.70
0
Sequential 7.20
7.38
43.12
6
VM2 Standard 2.60 2.65
21.36
11
Sequential 3.71 5.37
25.50
11
CRITVALS Standard 3.53 2.94 6.60
5% Sequential 4.97 4.21 13.12
Notes: