Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a notforprofit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.
Wiley is collaborating with JSTOR to digitize, preserve and extend access to Journal of Applied Econometrics.
http://www.jstor.org
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
Econ.
22: 383400
(2007)
(www.interscience.wiley.com)
?WILEY
?':
*4&#
iTltetSCieTICe?
BTr^rTTTTSTrnTTTrrnTif
^?
10.1002/jae.931
Departement
Department
de sciences
Los Angeles,
CA, USA
University
of Economics,
of Southern California,
de Montreal,
Montreal,
CIREQ and CIRANO, Universite
Quebec,
economiques,
Canada
SUMMARY
This
studies
paper
risk. We
focus
nonstationarities
in a panel
of Canadian
model
the crosssectional
and US
rates
interest
of different
maturities
and
as a linear dynamic
the panel
our data into common
and idiosyncratic
in turn.
and decompose
factor model,
that we analyze
components
some of the idiosyncratic
in our panel.
Our results
the presence
of a single nonstationary
factor
Since
suggest
we conclude
are cointegrated.
can
are stationary,
that these series
the dominant
factors
components
Finally,
on methods
which
dependence
within
be interpreted as level and slope factors as in the term structure literature. Copyright ? 2007 JohnWiley &
Sons,
Ltd.
1. INTRODUCTION
In earlier empirical studies, using standard methods, nominal interest rates of different maturities
are typically found to be nonstationary and cointegrated
(see, for example, Campbell and Shiller,
1987; Evans and Lewis, 1994). On the other hand, early results in the nonstationary panel literature
that suppose that the series in the panel are independent are more favorable to stationarity (see,
for example, Wu and Chen, 2001).
In this paper, we analyze a panel of 25 monthly Canadian and US interest rates of different
maturities and risk. First, we document significant crosssectional
correlation among the series in
the panel. To model the correlation, we employ a dynamic factor model. Our analysis then focuses
on a decomposition
of the panel into common and idiosyncratic components
in order to answer
the following three questions: (i) Is the observed data stationary or not and, if it is nonstationary,
common components
or nonstationary
is it because of nonstationary
idiosyncratic components?
correlations and how
(ii) How many common factors are necessary to capture the crosssectional
Do
the
estimated
factors
many of these common factors are nonstationary?
(iii)
represent some
observable variables of interest?
The first question will be answered by carrying out a series of panel unit root tests and stationarity
tests that have been developed
on the specification of the test, this will test
recently. Depending
are nonstationary or not. The first part of the
whether the idiosyncratic or common components
second question will be answered by using the panel information criteria proposed by Bai and Ng
(2002), while the second part of the question will be answered using a combination of tests and
*
Correspondence
Montreal,
Quebec,
Copyright
to: Benoit
Canada
2007
de sciences economiques,
Perron, Departement
H3C 3J7. Email: benoit.perron@umontreal.ca
John Wiley
CIREQ
and CIRANO,
Universite"
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
de Montreal,
384
a new
information criterion developed by Bai (2004). Finally, the third question will be answered
the
results in Bai (2004).
using
We find that interest rates are characterized by a single nonstationary factor and some stationary
in other words, they are cointegrated as in Campbell and Shiller (1987)
idiosyncratic components;
and Evans and Lewis (1994). Secondly, we find that the first two factors in our interest rate panel
have interpretations as level and slope factors as in the term structure literature.
The paper is organized as follows. Section 2 introduces our dataset and documents strong cross
for the analysis of
of the methodology
sectional correlation. Sections 3 provides an overview
factors and our related empirical results. In Section 4, we discuss our analysis of the idiosyncratic
components by focusing specifically on the application of panel unit root tests with factors. Finally,
Section 5 concludes.
2. DATA
section introduces our panel of interest rates briefly. Throughout, we suppose that we have
panel data zu of individual i that is observed at time t. Let n and T denote the size of the cross
section and time series dimensions,
respectively. We model our panel using the decomposition
common
and idiosyncratic components as in Bai and Ng (2004):
among deterministic,
This
(1)
where
du
idiosyncratic components.
is nonstationary while the idiosyncratic one
Except for the case where the common component
it is sometimes more convenient
for characterizing
the
is stationary (hence zu is cointegrated),
as
an
and
Perron
form
in
Moon
to
in
of
model
zu
(1)
express
autoregressive
stationarity properties
(2004):
zu =
zl
where
are
yit
unobservable
error
terms
<*i+ zl
Pizl_i
a factor
with
yit
(2)
+ yit
structure:
Pi ft + eit (3)
where
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22:
383400
(2007)
DOI:
10.1002/jae
Our panel consists of 14 monthly Canadian interest rates and 11 US rates. These vary by both
maturity and risk. The included Canadian rates are 1, 3, and 6month Tbills, federal government
bonds with a maturity of 1, 2, 3, 4, 7, and 10 years, commercial paper with a maturity of 1month
and Scotia indices of yields on corporate bonds with midterm
and longterm
and 3 months,
maturities. The US rates are Treasury securities with 3 months, 6 months, and 1, 2, 3, 5, 7, and
10 years to maturity,
1month commercial
indices of yields on corporate
paper, and Moody's
bonds with AAA and BAA ratings. The panel spans the January 1985April
2004 period for a
total of 232 observations
for each rate.
Table I presents estimates of the shortrun correlation matrix for these data.1 We have divided
the data into Canadian and US rates. There are high correlations among yields with similar maturity
in a given country. In particular, the correlation among long rates is very high, as should be the
case if the expectations hypothesis were true. There is much lower correlation within a country
between short rates and long rates and across countries for the same maturity. For example, the
3month Canadian Tbill has a correlation of 0.933 with the 6month Canadian Tbill but of only
0.405 with the 3month US Treasury. Thus, the data are supportive of a model with high degree
of correlation among crosssectional
units such as our factor model.
3. ANALYSIS OF FACTORS
The first step in our analysis of the nonstationary properties of our panel is to analyze the behavior
of the common factors. This entails estimating the total number of common factors and determining
how many of these are nonstationary. We will also attempt to relate the estimated factors to
observable variables of interest. The first subsection discusses our methodology,
while the second
one
reports
our
results.
empirical
3.1. Methodology
Estimation of Number of Factors
One of our main goals is to determine the number of factors in the possibly nonstationary panel
model (l)(3). To this end, we will employ information criteria as suggested by Bai and Ng (2002).
These information criteria will be applied to factors estimated by principal components either on
residuals (following
the Moon and Perron, 2004, approach) or on first differences
(following Bai
and Ng, 2004).
To estimate the true number of factors, K, Bai and Ng (2002) propose tominimize
the following
criterion functions:
PC(r)
IC(r)
o2e(r) + rGnT,
r is the number
where
longrun
Copyright
correlation
2007
matrix
John Wiley
is similar
& Sons,
J.
Appl
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
n_o
1361
10 1m3m
Mid
1m
Long
10 1m AAA
fcP
mth
mths
mths
yr
yrs
yrs
yrs
yrs
yrs
CP
BA
corp.
CP
corp.
mths
mths
yr
yrs
yrs
yrs
yrs
yrs
CP
?BAA
0.108
0.148
0.193
0.241
0.346
0.380
0.444
0.455
0.486
0.134
0.152
0.136
0.479
0.365
0.456
0.525
0.615
0.731
0.775
0.822
0.855
0.870
0.299
0.942
??
AAA
0.100
0.153
0.214
0.274
0.379
0.417
0.484
0.489
0.121
0.524
0.155
0.122
0.499
0.481
0.412
0.570
0.663
0.769
0.815
0.860
0.893
0.906
0.309
matrix.
First
of
monthly
Canadian
and
US
interest
rates
19852004
^
10.970
m
com.
paper
0.165
0.309
0.354
0.360
0.304
0.308
0.258
0.229
0.215
0.311
0.254
0.264
0.215
0.244
0.661
0.698
0.660
0.528
0.465
0.392
0.361
0.319
^
10
years
0.122
0.205
0.276
0.347
0.453
0.485
0.556
0.559
0.206
0.144
0.584
0.501
0.146
0.516
0.500
0.650
0.766
0.886
0.930
0.989
Ju
7years
0.137
0.228
0.304
0.377
0.484
0.511
0.574
0.590
0.226
0.158
0.160
0.499
0.527
0.543
0.694
0.811
0.926
0.963
0.989
differences
S5years
0.147
0.242
0.316
0.388
0.492
0.506
0.559
0.550
0.556
0.239
0.169
0.171
0.456
0.502
0.598
0.744
0.857
0.958
0.984
0.267
0.162
0.494
0.436
0.913
0.807
0.512
0.411
0.342
0.560
0.543
0.191
0.260
0.663
0.988
0.520
0.546
0.186
3years
2years
0.193
0.304
0.373
0.436
0.523
0.524
0.548
0.526
0.222
0.515
0.295
0.228
0.408
0.476
0.732
0.863
0.952
1year
0.248
0.359
0.414
0.459
0.500
0.501
0.461
0.473
0.504
0.279
0.347
0.289
0.365
0.856
0.429
0.953
Canadian
US
<l
rates
6months
0.278
0.372
0.404
0.426
0.431
0.434
0.420
0.365
0.380
0.306
0.362
0.315
0.294
0.361
0.937 *
O
0.323
0.405
0.401
0.426
0.418
0.340
0.394
0.310
0.326
0.380
0.300
0.237
0.365
0.325
3months
g 0.333
0.759
0.670
0.513
0.322
0.954
0.948
0.900
0.953
0.514
0.860
0.292
Mid
corporate
Long
corporate
0.233
0.440
0.591
0.667
0.885
0.757
0.815
0.911
0.940
0.266
0.440
0.277
1m
bank.
ace.
0.952
0.909
0.761
0.651
0.404
0.469
0.517
0.376
0.313
0.993
0.922
ffl
3m
com.
paper
0.863
3
2
0.979
years
years
0.924
0.642
0.698
20.976
0.830
0.621
0.474
0.955
0.883
0.276
0.710
10
years
0.509
0.299
0.363
0.910
0.948
0.644
0.755
0.392
0.456
1m
>com.
paper
0.848
7years
0.334
0.545
0.690
0.889
0.778
0.928
0.977
O
5years
0.359
0.575
0.716
0.802
0.924
0.961
g
0.567
0.414
0.862
0.783
0.910
0.834
1year
0.583
0.851
0.966
E
0.704
0.649
0.466
0.535
0.466
0.928
0.962
0.933
0.697
6months
a.
0.872
3
j*>Canadian
rates
months
?I
D_
o
??
rates
US
fc

Ooo
8c3
s? So?8
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
that acts as penalty for more complex models. The penalty function has to satisfy the conditions
(i) Gnj > 0 and (ii) min{w, T}Gnj > oo, as both n and T go to infinity. As we increase the
number of factors, the fit must improve (i.e., a% goes down), but the penalty term increases. The
the appropriate criterion.
estimated number of factors is the integer that minimizes
Bai and Ng (2002) and Moon and Perron (2004) demonstrate that estimation of the total number
either the PC or IC criterion is consistent in the sense that the probability
of factors by minimizing
that the estimated number of factors equals the true one approaches one as both n and T become
large.
Bai and Ng
suggest
three specific
GpcXnT
GPcxnT
Gpc,3,nT
^2
<?e (Kmzx)?
n + T (? nT \
m
nT
\n + TJ
=
^(Xmax)^^nT
=
(5)
T})\
?.??
(
\ min{n,T}
estimated
where a^(Kmax) is the variance of the idiosyncratic
component
and
number of factors leading to criteria PCi, PC2, and PC3 respectively,
GicxnT
Gicxm
triC,3,nT
= ??
n+ T\ (?? nT \
In
nT
minjn,
T)
the maximum
(6)
T}),
fln(rmn{n,T})\
I
r";7f\?
with
\n + T J
n+ T
= ?zrln(min{/i,
nT
=
criterion:
ln(min{/i,r}),
/ln(min{?,
^2
^C^max)
for the PC
function
leading to the ICi, IC2, and IC3 criteria, respectively. Note that, just as in BIC, the IC has the
advantage of not requiring the estimation of a scaling factor in the penalty function. We also
consider the modified BIC criterion (called BIC3):
r
ln(nT) (1)
because simulation evidence suggests that it performs better in selecting the number of factors
when min (n, T) is small (<20), as is often the case in empirical applications. Bai and Ng rejected
this criterion because it does not satisfy the required conditions for consistency when either n or
T dominates the other one exponentially,
but this appears to be a rather unusual case. For small
n and T of roughly the same magnitude,
this criterion performed best in their simulation among
those
they considered.
For
large panels,
function
are essentially
equivalent.
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
388
the above information criteria on either first differences or residuals, our next task is to determine
how many of these factors are nonstationary. For this purpose, we will draw from two approaches.
The first one is an extension of the information criteria above. The second is a testing procedure
proposed in Bai and Ng (2004) that tests the rank of the longrun covariance matrix of the factors.
Bai (2004) proposed new information criteria to select the number of nonstationary
factors in a
panel. These are closely related to the information criteria in Bai and Ng (2002) discussed above,
but they are applied to the levels of the series rather than the first differences or the residuals. The
three
are
criteria
n+ T
^~o
( nT \
,
IPCdr) = a2u(r)+ raTa2u(Kmm)?? ln I
?^
j
IPC2(r)
IPCi(r)
uit =
=
=
a2u(r) +
a2u(r) +
ln(min{?,
T}),
(8)
raTa2u(K^)?^
+ T~
raTa2u(Kmax)n
T
ln(nD
ni
where
ccT=
zu
dit
Thus,
in Bai and Ng
(2005).
on Estimated
Factors
has
obtained inferential results for
Bai
(2004)
Recently,
allow one to test whether some observable variable can
the estimated
factors
The inherent problem with
a
some
are
to
basis
normalization
space
only). The
(they
Inference
factors
the linear
regression:
Rt=a
Copyright
2007
John Wiley
& Sons,
+ #Ft +
Ltd.
rit.
J.
(9)
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
where Rt is some observed series and Ft are the estimated factors. In our results below, we will
report the R2 of this regression for each individual series in the panels and some spreads (in levels
and in first differences to prevent against a spurious regression) for one and two factors. Of course,
we could also search for observables
that are proxied by these factors outside of our panels, but
the series in the panels are natural first candidates.
Bai also proposed a procedure for testing whether one of the factors represents the observed
series Rt and the fitted values from (9), Rt.
series, Rt. The idea is to compare the observed
intervals for Rt requires the factors to be nonstationary with
The construction of these confidence
stationary idiosyncratic errors, otherwise the above regression is a spurious regression (see Phillips,
(i.e., for each t), which means that if a factor truly
1986). These distributional results are pointwise
we
see
should
observed
series
the
Rt,
(under independence) a% of the observations fall
represents
?
interval constructed in this fashion. In consequence, we will also
outside of a 1 a% confidence
that fall outside the appropriate 95%
report in our tables below the percentage of observations
confidence
interval. If the asymptotic analysis is accurate and one of the included factors in the
above regression represents the observed series, we should expect to see entries close to 5% in
these
3.2.
columns.
Empirical
Results
apply our procedures to the analysis of our panel of nominal yields. We first
the total number of factors. Typically,
the term structure literature employs
the number of factors does not
three factors (see Litterman and Scheinkman,
1988). However,
seem to be well estimated by the Bai and Ng information criteria. The ICi suggests the presence
number we allowed), while BIC suggests seven factors.
of eight factors (the maximum
our
the MQC
factors. Using
Table II reports
evidence regarding the number of nonstationary
number of factors.
factor, regardless of the maximum
statistic, we find a single nonstationary
of eight nonstationary
However, results with information criteria are more fragile. With amaximum
In this section, we
start by estimating
factors, the IPC criterion finds four nonstationary factors while the BIC finds three. After setting
the maximum number of factors to four, the IPC finds three factors and the BIC finds two factors.
If we allow only two factors, all criteria find a single one as with theMQC statistic. Finally, the
KPSS statistic rejects stationarity for only one factor. Thus, our results point to the presence of a
single
nonstationary
factor.
To determine whether our estimated factors proxy some variables of interest, we regressed each
interest rate on a constant and either the first one or two estimated factors. The results from this
regression are presented in Table III for each interest rate in the panel and for some interest rate
spreads. The first two columns of the table report the R2 from regression (9) when a single factor
Table
Max.
II. Estimated
number
of nonstationary
factors
12345678
11111111
112
1112
MQC
IPC
BIC
4
3
3
3
Note:
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
4
3
4
3
factors with
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
390
K=\
Levels
Canadian
1month
Tbill
Tbill
91.4
93.5
95.5
1 year
97.0
2 years
98.1
3 years
97.8
96.2
5 years
7 years
93.9
10
91.3
years
Longterm
Midterm
First differences
Levels
First
K=\
K=2
differences
rates
3month
Tbill
6month
1month
3month
1month
K=2
commercial
paper
commercial
paper
bankers'
acceptances
(Scotia)
(Scotia)
corporate
corporate
US rates
3month Treasury
6month Treasury
1year
Treasury
2year
Treasury
3year
Treasury
5year
Treasury
Treasury
7year
10year Treasury
1month commercial
paper
Spreads
Canadian
10y3m
3m US3m
Canadian
91.8
93.5
91.9
85.9
91.8
78.6
79.8
81.4
84.6
86.6
88.5
88.4
87.6
77.5
85.2AAA
Moody's
81.3BAA
Moody's
42.7
42.8
2.8
10y3mUS
42.3
67.2
76.0
79.3
79.8
78.3
76.2
72.5
66.0
48.5
66.7
49.5
56.2
67.1
98.6 82.4
63.8
48.3
99.4 96.9
57.8
22.0
99.4 89.1
44.8
37.9
98.9
83.6
32.8
98.2
27.2
79.9
54.7
97.9
33.6
79.8
53.9
97.1
51.7
82.3
67.2
95.8
80.3
58.2
94.5
63.4
78.7
69.8
99.1 89.5 60.3
32.3
99.6 96.9
56.0
14.7
99.289.7 60.8
32.8
88.2 67.9
74.1
78.5
92.675.8 65.5
82.8
33.7
40.2
47.1
47.9
46.4
44.6
44.2
41.0
20.2
32.6
29.6
78.7 34.4
82.3
91.8
79.8 45.0
80.6
90.5
82.1 57.9
84.1
88.8
87.5 69.0
81.5
75.9
73.7
71.6
91.3 73.6
72.8
96.2 76.0
43.5
65.5
25.0
97.7 79.3
98.3 77.6
67.2
18.1
77.8 20.4
85.8
93.5
61.2
69.2
96.1
37.1
65.5
63.1
61.6
91.3
8.0
42.7
3.0
93.9 84.1
57.6 77.1
32.3
53.9
70.3
37.5
92.2
94.0
87.9
100.0
36.9
85.3
is included on the righthand side. The table reveals that all rates are highly correlated with the first
factor, with the highest correlation for the 2year Canadian rate. The next two columns report the
R2 from regression (9) when a second factor is added to the regression. We see a large increase in
the R2 for the spreads, in particular the Canadian term spread. Finally, the last two columns provide
intervals for the rotated factors using the Bai (2004)
the rejection rates of the 95% confidence
observable variable, we would
the corresponding
the
estimated
factor
If
proxied
methodology.
corroborates the information
to
time.
in
these
columns
The
information
of
5%
the
expect
reject
from the R2. The lowest rejection rate with one factor is with the 2year Canadian rate, and the
of a second factor, has most impact on the Canadian term spread. There is much less
we see
on
the US term spread or on the spread between the two countries. Nevertheless,
impact
that all rejection rates in the table are much above the nominal 5% level. This suggests that the
first two factors do not represent one of the rates or spread in our panel.
Figure 1 plots the time series of the 2year Canadian bond rate against the limits of the 95%
confidence interval for the rotated first factor. For illustration we use the case with a single factor,
addition
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
14 I1111
0I1111
V'
1990
1985
1995
2005
2000
Figure 1.Time plot of 2year Canadian bond (solid line) vs. 95% CI for rotated factor (dashed lines)?single
factor
but the picture with two or even three factors is almost identical. We see that the fit is generally
pretty good, but that large parts of the series lie outside the bands. If the factor proxied the 2year
Canadian rate, we would see (under independence over time) 5% of the observations outside the
limits of the confidence band. Since about 27% of them do, we must conclude that no single
interest rate in the panel can proxy for our first estimated factor.
In the empirical literature on the term structure, many models with unobservable
factors are
to account for the
used. In these studies, it is common to find that three factors are necessary
or
term structure (see Litterman and Scheinkman,
et
Andersen
Diebold
and Li,
1988;
al., 2004;
recent
these
for
with
three
associated
with
factors
2006;
'level', 'slope' and
usually
examples),
'curvature'. It is clear that our first factor can be labeled a level factor that shifts the entire set
of yields. Usually,
this level factor is proxied by a short rate, but our results suggest that a 1 or
rate
that this first factor affects all rates similarly
2year
provides a better fit. Further evidence
comes from the fact that the loadings are similar for all rates in the panel (results available from
the authors upon request).
The second factor, the 'slope' or steepness of the yield curve, is usually proxied by a spread
between a long rate and a short rate. It is therefore reasonable that the inclusion of a second
factor in the regression has a large impact on term spreads, mostly Canadian but also American
as evidenced by the much higher R2 (multiplied by 10 in first differences for both term spreads).
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
392
I1111
V'"^'
3
_4 I_i111
19901985
2005
2000
1995
Figure 2. Time plot of Canadian spread (solid line) vs. 95% CI for rotated factors (dashed lines)?two
factors
of these two factors seems to have much to do with the yield differentials between the two
countries; however. Figure 2 plots the Canadian term spread and the limits of the 95% confidence
bands for the case with two factors. Once again, we see that the overall fit is pretty good, but that
the confidence
interval does not contain the observed series for many periods, mostly associated
with large swings (either up or down) in the spread. Note that there are only two episodes when
Neither
the Canadian yield curve was inverted (negative spread) for more than one month,
in early 1986
and late 1989early
interval appears rather narrow, and this could be due
1990. The confidence
to the fact that we have treated this second factor as nonstationary
in the construction. Our results
above
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
closely linked to levels of economic activity. They find a high correlation between their second
factor and a measure of capacity utilization. This correlation is also the source of the use of the
term spread to forecast business cycle fluctuations as in Estrella and Mishkin
(1998). The yield
curve tends to be downwardsloping
(short rates are higher than longer rates) prior to or at the
beginning of recessions. Our second factor is therefore likely to be related to the business cycle.
4. ANALYSIS OF IDIOSYNCRATICCOMPONENTS
step in our analysis of the nonstationary
properties of our panel is to look at the
our
in
uit
(1). The analysis of the nonstationarity of the
decomposition
idiosyncratic components,
more
is
much
than
the analysis of the stationarity properties of
developed
idiosyncratic component
is
factors and proceeds in two steps: in the first step some estimate of the idiosyncratic components
a
a
root
test
is
of
in
second
The
the
while
unit
estimation
obtained,
step panel
idiosyncratic
applied.
is equivalent to obtaining estimates of the deterministic
and common components.
component
to obtain estimates of
In the previous section, we used principal components on first differences
the factors. The idiosyncratic components could then be obtained as the residuals after removing
Our
second
factor loadings).
The second step in our analysis consists in
unit root. We will use a total of seven unit root
correlation is captured via a factor model. This
T to diverge since a large n is necessary for
whereas
the alternative
hypothesis
: pt =
1 for
all
i=
1,
...,
:pt <
1 for some
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
22: 383400
(2007)
DOI:
10.1002/jae
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
394
and Wu,
1999;
Early panel unit root tests (e.g., Levin et al, 2002; Im et al, 2003; Maddala
is
Choi, 2001) assumed independence of observed data across individual units. This assumption
clearly unrealistic in empirical settings as suggested by the large correlations reported in Table I.
once factors have been extracted from the data, similar methods
can be applied to the
However,
In other words, once we have estimated idiosyncratic components, we
idiosyncratic components.
could apply the pooling approach of Levin et al (2002), the averaging approach of Im et al.
and Wu (1999) or Choi (2001) to test
(2003), or the pvalue combination
approach of Maddala
the joint null hypothesis of nonstationary
idiosyncratic components.
The seven tests we consider are those of Moon and Perron (2004), Bai and Ng (2004), Phillips
and Sul (2003), Choi (2006a), Pesaran (2007, this issue), a panel version of the Sargan and
in Moon
and Perron (2005), and a common point optimal test
(1983) test developed
Bhargava
= 1. These tests
c
et
Moon
in
with
al.
(2006)
vary according to the way they eliminate
proposed
the factors and the way they aggregate the individual information. We will first use the defactoring
of Moon and Perron (2004) before applying these last two tests.
The CIPS test developed by Pesaran (2007, this issue) is slightly different in nature. Pesaran
showed that by augmenting
the usual ADF regression with the first difference and the first lag
of the crosssectional
mean, one can account for the crosssectional
dependence
arising through
a single stationary factor. Thus, no direct estimation of the idiosyncratic component
is needed in
method
this approach. This can be beneficial for small panels where estimation of factors is difficult.
As with the factors, we will perform stationarity tests on the estimated idiosyncratic components.
In order to do so, it is first necessary to project the estimated idiosyncractic
components onto the
a
to
constant
the
The
this approach has
and
factors.
of
space orthogonal
validity
nonstationary
been shown in Bai and Ng (2005). Note, however,
that we can only perform individual KPSS
tests on the estimated idiosyncratic components. Pooling is not possible owing to the presence of
factors as the nonstationarity will be transmitted to the residuals in a nonvanishing
nonstationary
way under the null hypothesis of stationarity.
that do not rely on a factor
Other tests for panel unit roots with crosssectional
dependence
structure are available. Recent tests of this kind that have been proposed by Chang (2002), Breitung
and Das (2005), Shin and Kang (2006), and Choi and Chue (2007, this issue), while an earlier one
was proposed by Taylor and Sarno (1998). These tests allow for general crosssectional
dependence
of the error terms and, typically, do not need to let the number of crosssections,
n, diverge since
no estimation of the factors is necessary.
4.1.
Empirical
Results
Our unit root test results are presented in Table IV. For the tests that require the estimation of the
number of factors (Moon and Perron, Bai and Ng, and the two optimal tests combined with the
Moon and Perron defactoring procedure), we report results as the number of factors varies between
one and eight since the estimated idiosyncratic components depends on the assumed number of
does not appear to be
factors. We report results in this way because the number of crosssections
sufficient to get a good estimate of the number of factors, as we saw above. For the tests that are
based on individual ADF tests and the CIPS test, we choose the number of lagged first differences
to be included by AIC in all cases with a maximum
of 12 lags. When needed, longrun variances
are computed with a quadratic spectral kernel estimate with bandwidth selected
and covariances
the rule of Andrews
(1991) and with prewhitening.
following
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22:
383400
(2007)
DOI:
10.1002/jae
Table IV. Results of unit root tests on idiosyncratic components. Sample of 25 monthly Canadian and US
interest rates, January 1985April 2004
12345678 K=
MoonPerron
6.496*
2.217*
BaiNg
5.611*
(Z)
PhillipsSul
Choi
(Z)
Pesaran
3.059*
(CIPS)
1.263
SarganBhargava
MoonPerron0.792
Phillips (c= 1)
Number of stationary
25
5.923*
0.912
3.807*
2.790*
4.622*
1.214
KICl
KBic3
4.261*
1.124
4.678*
0.965
4.703*
5.190*
4.589*
7.196*
8
8
7
7
1.942*
2.125*
2.604*
2.585*
2.406*
2.451*
3.671*
4.115*
8
8
7
7
7.914*
1.489
0.574
25
2.685*
2.637*
2.352*
1.934*
25
24
2423
25
23
components
*
at the 5% level.
Significant
the results from the application of panel unit root tests are clear and do not depend
Fortunately,
much on the estimated number of factors. With the exception of the MoonPerronPhillips
and
one
or
two
tests
the
Bai
test
with
factors
and
and
with
and
two,
four,
five,
SarganBhargava
Ng
six factors, we reject the null hypothesis of a unit root in all idiosyncratic components.
The results from the KPSS tests confirm these results. The last row of Table IV includes the
number of idiosyncratic series for which we cannot reject stationarity at the 5% level. Regardless
of the number of factors we allow, we cannot reject stationarity for at least 23 series. Moreover,
we
reject stationarity for a total of only six out of the 200 tests that we perform. As discussed
above, it is not possible to aggregate these individual tests in order to control the overall rejection
these results are indicative and further support our claim that most, if not
probability. However,
all, idiosyncratic components are stationary.
The presence of nonstationary
that
factor(s) and stationary idiosyncratic
components means
the nominal yields in our panel are cointegrated. Similar conclusions
factors
and
(nonstationary
stationary idiosyncratic components) are obtained when applying our approach to the yields of each
in the term
country separately. This corroborates many empirical results supporting cointegration
structure dating back to Campbell and Shiller (1987) and Evans and Lewis (1994) and provides
in the term structure such as that of Carstensen
support for models of cointegration
(2003).
5. CONCLUSION
This paper has analyzed a panel of interest rates with different maturities and risk characteristics.
Our application
of recently developed methods
for nonstationary
the units are
panels where
correlated through a factor structure suggests that a single nonstationary
factor is sufficient to
model these data. However,
the number of stationary factors needed to model the crosssectional
correlation
is not well
identified. Since we can reject the nonstationarity
of all idiosyncratic
these results suggest that interest rates are cointegrated. The dominant factor in
components,
the interest rate panel is a level factor that is highly correlated with all rates and could be the
result of inflationary expectations. The second factor has an interpretation as a slope factor, that is
the differential between a long rate and a short rate, since it affects short and long rates differently
and might be a measure of the business cycle.
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
396
APPENDIX: SIMULATIONEVIDENCE
to document
In this appendix, we present limited simulation
the behavior of the
evidence
tests for the nonstationarity
of the idiosyncratic
in this paper. Other
considered
components
simulation evidence can be found in Gutierrez
(more comprehensive)
(2006) and
comparative
et al. (2005).
Gengenbach
The datagenerating process is a slightly modified version of the datagenerating process of Moon
and Perron (2004). The modification
greatly increases the level of crosssectional
dependence by
replacing the _V (0, 1) factor loadings with U [0, 1] factor loadings.
The datagenerating
process
is given by equations
Zit
with
Oti +
(2):
z?it
A41 + yit
All
(3):
tPift + eit.
(ftj,eit)~i.i.d.N(0,I2)
while
neous,
Af
(0,
are Pij ~
i.i.d.U
components
are heteroge
1).
is considered for pt =
~ U
0.99, Pi
[0.98, 1]. Finally,
common versus
of
importance
= 10, 20, and 100) and two
(n
even though it is not necessary
first thing to notice is that all tests that require the estimation of the number of factors (MPP,
for the
Moon
and Perron, Sargan and Bhargava, and Bai and Ng) have severe size distortions
of the number of factors as reported inMoon
smallest choice of n. This is due to overestimation
number of factors
and Perron (2004). For n = 10, the ICi criterion tends to choose the maximum
allowed (8). For n > 20, this problem disappears since the estimated number of factors is almost
always equal to the true one. The CIPS and Phillips and Sul tests are not affected by this problem
Copyright
2007
John Wiley
& Sons,
Ltd.
/.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22:
383400
(2007)
DOI:
10.1002/jae
MPP
c=
x=
r=
MoonPerron
c=
10
20
100
10
20
100
100
100
100
300
300
300
30.8
35.4
4.3
4.9
30.3
4.4
4.7
6.0
5.9
34.5
6.2
5.3
10
20
100
10
20
100
100
100
100
300
300
300
40.1
4.5
5.0
35.5
4.0
4.4
45.0
6.4
CIPS
Sargan
Choi
BaiNg
PhillipsSul
2
5.6
23.0
17.5
16.9
5.3
2.9
3.9
24.5
3.3
3.7
7.0
7.1
24.5
69.7
6.2
6.0
5.6
6.1
4.9
5.2
5.7
5.9
6.4
5.1
5.6
6.5
40.6
8.2
9.2
35.0
7.2
6.4
6.0
5.4
5.8
5.4
5.1
5.7
33.1
2.4
29.5
7.4
7.0
29.8
7.4
5.6
40.7
5.9
5.0
3.8
29.3
2.8
4.0
16.4
6.7
5.7
5.9
7.5
17.7
23.5
35.0
6.7
7.0
17.3
6.3
6.6
64.3
9.4
13.4
33.5
5.3
6.1
6.8
5.3
5.0
6.9
Note:
VI.
Table
MPP
c=
t=1
t =
Note:
10
20
100
10
20
100
100
100
100
300
300
300
10
20
100
10
20
100
100
100
100
300
300
300
51.1
56.9
99.6
74.3
of tests against
linear
DGP: zu = ?/o +
z?,
=
+ *? '
4
P?4i+*#/'
1)
ai0,ft,eit~iA.dJN(0,
i.i.d.U
[0, 1]
ft
~
U [0.98, 1]
Pi
Power
MoonPerron
c=
alternative
CIPS
Sargan
39.8
42.2
Choi
BaiNg
PhillipsSul
95.8
100.0
57.3
64.9
99.7
80.0
97.1
100.0
51.7
69.9
99.8
76.6
97.6
100.0
8.0
8.9
11.0
18.0
25.8
45.3
57.9
45.3
89.7
72.9
78.8
93.1
63.8
53.2
90.6
78.7
81.3
93.4
59.8
60.3
92.1
75.5
84.1
93.7
7.0
7.5
7.2
16.8
20.9
36.0
99.1
67.8
93.7
100.0
25.5
44.1
94.9
48.9
97.1
100.0
59.6
87.4
99.8
88.0
98.7
99.9
11.6
17.9
50.9
28.9
52.3
97.8
48.6
33.0
88.6
66.5
75.9
93.8
24.1
36.4
82.4
42.2
83.5
97.9
56.7
74.8
90.6
75.0
87.8
96.3
8.8
13.9
36.3
20.0
31.2
69.2
See Table V.
since the first one does not estimate factors directly, while the second one imposes (correctly in
this case) the presence of a single factor. As expected,
the Choi test has large size distortions.
These are reduced when T is large and r is small. The Moon and Perron and Bai and Ng tests
tend to overreject for n > 20. The MPP tends to slightly underreject for c = 1 and overreject for
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
398
Table
VII.
Sizeadjusted
of
tests
linear
alternative
CIPS
Sargan
BaiNg
7.4
8.1
9.3
18.8
25.0
42.2
10.0
53.2
99.3
23.9
95.8
100.0
51.1
89.8
22.4
81.6
6.1
7.0
6.3
15.9
20.5
8.9
43.2
89.8
20.0
80.1
93.4
32.3
94.1
power
against
DGP:zit=ctio+z?t
=
+ eit
zl
Pizl_l+Tpift
~i.i.d._V(0,
1)
otio,ft,eit
~ i.i.d.
U [0, 1]
Pi
~ U
Pi
[0.98, 1]
nT
MPP
c=
c=\
10
r=l
20
100
10
20
100
t =
3
20
100
10
20
100
Note:
Each
MoonPerron
10.3
60.6
99.6
22.3
96.2
100.0
10.4
62.0
99.7
23.7
96.6
100.0
100
100
100
300
300
10 8.6
47.7
89.7
8.4
49.0
89.9
20.8
80.3
93.4
300
11.3
62.2
99.7
27.3
96.5
100.0
10.6
the percentage
of replications
in which
entry represents
5% test using the empirical
critical values. The number
appropriate
PhillipsSul
100
100
100
300
300
300
19.1
80.0
93.4
Choi
8.4
36.2
93.0
20.4
96.0
100.0
10.2
30.3
53.5
16.3
96.6
45.9
85.8
28.6
98.4
99.9
96.9
7.0
31.0 11.7
77.0
16.0
80.7
97.4
49.8
8.2
24.3
33.0
51.2
30.9
66.0
19.0
31.2
79.0
89.7
65.8
is rejected
for the
c=
2. Nonetheless,
it is clear that the combination
of the Moon and Perron (2004) defactoring
procedure with this common point optimal tests leads to reasonable size control.
The sizeadjusted power results suggest that the MPP, Moon and Perron, and SarganBhargava
tests dominate and are pretty much equivalent. The Bai and Ng test is next, followed by Phillips
et al. (2006) that the
and Sul, and finally CIPS. Table VII also reproduces the result in Moon
?
10). Finally, the table
power of their test is not sensitive to the choice of c (except maybe for n
also shows that, as expected, power goes down with the degree of crosssectional
correlation; i.e.,
= 3 than for r = 1.
power is lower for r
In conclusion,
in this study, the CIPS test of Pesaran (2007, this
among the tests considered
issue) is best at controlling size, in particular for smaller panels. It is also the easiest to compute.
its power for alternatives
that are close to the unit root null hypothesis
is quite low.
However,
in situations with small panels and alternatives of interest that
It should therefore be emphasized
are not too close to unity. In other cases, tests that estimate the factor model are called for. In
is suspected, the Bai and Ng procedure must be given precedence
since
particular, if cointegration
it is the only valid procedure in such cases.
ACKNOWLEDGEMENTS
on Unit Roots
in
of this paper circulated under the titles 'Panel Evidence
'An
and
and
Interest
Rates
with
Crosssectional
Analysis
Dependence'
Exchange
Empirical
thank
in Panels of Exchange Rates and Interest Rates with Factors'. We
of Nonstationarity
at Texas
three anonymous
Werner Ploberger, Hashem Pesaran,
referees, seminar participants
on Panel
and Toronto and the 11th International Conference
A&M, Rice, Rochester,
Queen's
versions
Previous
Rates
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
REFERENCES
mean
in the shortterm
J. 2004.
Stochastic
TG, Benzoni
L, Lund
drift, and jumps
volatility,
rate. Mimeo,
and University
of Minnesota.
Northwestern
University
and autocorrelation
Andrews
D.
1991. Heteroskedasticity
covariance
matrix
consistent
estimation.
metrica
59: 817858.
interest
Andersen
Bai
J. 2004.
Bai
S.
J, Ng
191221.
2002.
common
crosssection
Estimating
122:
Econometrics
trends
stochastic
in nonstationary
data.
panel
Econo
Journal
of
137183.
the
Determining
number
of
in approximate
factors
factor
models.
70:
Econometrica
Bai J, Ng S. 2004. A PANIC attack on unit roots and cointegration. Econometrica 72: 11271177.
Bai J, Ng S. 2005. A new look at panel testing of stationarity and the PPP hypothesis. In Indentification and
in Econometric
Inference
Models:
in Honor
Essays
of Thomas
J. Rothenberg,
Andrews
D,
Stock
J (eds).
Breitung
120.
S. 2005.
J, Pesaran
Breitung
The Econometrics
Panel
MH.
2007.
under
roots
Unit
Data
of Panel
tests
root
unit
(3rd
and
edn),
cross
sectional
Statistica
dependence.
as L.
in panels.
In Maty
cointegration
P (eds). Kluwer
L, Sevestre
and P.
Academic:
Matyas
59:
Neerlandica
Sevestre,
Dordrecht
(Forthcoming).
JY, Shiller
Campbell
RJ.
1987.
tests of present
and
Cointegration
value
models.
Journal
of Political
Economy
95: 10621087.
K.
Carstensen
2003.
term
Nonstationary
and
premia
of
cointegration
term
the
structure.
EconomicsLetters
80: 409413.
Y.
2002.
Chang
Econometrics
Choi
I. 2001.
Choi
I. 2006a.
Nonlinear
root
Unit
IV
root
unit
tests
in
panels
with
crosssectional
tests
for panel
root
unit
Combination
data.
tests
Journal
of International
Money
correlated
crosssectionally
C. B. Phillips,
Corbae
of Peter
311333.
UK;
and Finance
for
in Honor
Essays
Press: Cambridge,
In The Palgrave
Handbook
Nonstationary
panels.
of Econometrics,
Macmillan:
511539.
UK;
(eds). Palgrave
Basingstoke,
Choi
T. 2007.
tests for nonstationary
I, Chue
Subsampling
hypothesis
panels
rates and stock prices.
Journal
Econometrics
22: 233264.
of Applied
and
Research:
Applied
Cambridge
Choi
I. 2006b.
Diebold
FX,
University
Li C.
Journal
dependency.
of
261292.
110:
2006.
Forecasting
the
term
B.
2006.
structure
of government
bond
20:
249272.
In Frontiers
panels.
of Analysis
B (eds).
D, Durlauf
S, Hansen
Vol.
with
yields.
1,Mills
T, Patterson
to exchange
applications
Journal
of Econometrics
130: 337364.
Diebold
FX,
Rudebusch
factor
approach.
Estrella
A, Mishkin
Economics
and
Evans
Lewis
MD,
Journal
Gengenbach
dencies:
Gutierrez
Aruoba
GD,
Journal
of Econometrics
FS.
1998. Predicting
Statistics
80: 4561.
KK.
of Monetary
C, Palm
1994.
Do
Economics
The
131:
US
recessions:
risk
stationary
33: 285318.
and
macroeconomy
309338.
financial
premia
explain
variables
it all?
the yield
as
Evidence
curve:
a dynamic
indicators.
leading
from
the
latent
Review
term
of
structure.
JP. 2005.
unit root tests in the presence
Panel
of crosssectional
FC, Urbain
depen
and implications
for modelling.
of Maastricht.
Mimeo,
comparison
University
L. 2006.
Panel
unit roots tests for crosssectionally
a Monte
correlated
Carlo
panels:
comparison.
C, Mignon
V.
2004.
Second
generation
panel
unit
root
tests.
THEMACNRS,
Universite
de Paris
X.
Mimeo.
Copyright
2007
John Wiley
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
400
Im KS,
Pesaran
Shin
MH,
Y.
2003.
Testing
Unit
root
roots
for unit
in heterogeneous
Journal
panels.
of Econometrics
115: 5374.
Levin
A, Lin F, Chu C.
108:
of Econometrics
Litterman
R, Scheinkman
Maddala
GS,
Wu
S.
2002.
tests
in panel
data:
and finitesample
asymptotic
Journal
properties.
124.
J. 1988. Common
A
1999.
factors
comparative
of
study
bond returns.
affecting
unit root tests with
Journal
panel
of Fixed
data and
Income
a new
1: 5461.
test.
simple
Perron
HR,
B.
2004.
Testing
for
B.
2005.
Asymptotic
de Montreal.
a unit
root
local
power
in panels
with
factors.
dynamic
Journal
of Econometrics
122: 81126.
Moon
HR, Perron
effects. Mimeo
Moon
Universite
Perron
HR,
B,
of Econometrics.
Pesaran MH.
2007.
1986.
Sul D.
PCB,
Phillips
dence.
Econometrics
JD,
Sargan
Gaussian
Bhargava
random
Incidental
2006.
panel
simple
Applied Econometrics
PCB.
Phillips
311340.
PCB.
Phillips
root
unit
. ratio
of pooled
test
trends
and
Understanding
Copyright
S.
and
2007
2001.
of
crosssection
2003.
Dynamic
panel
Journal
6: 217259.
A.
walk.
1983.
Testing
Econometrica
estimation
residuals
51:
in econometrics.
regressions
spurious
and
from
homogeneity
in panels
root
unit
with
tests.
fixed
Journal
Journal
dependence.
of
least
squares
Mean
Statistics
John Wiley
reversion
63:
of
of Econometrics
Journal
under
testing
for
regression
cross
being
section
generated
33:
depen
by
the
153174.
S. 2006. An
instrumental
variable
DW,
Kang
approach
134: 215234.
Journal
of Econometrics
dependence.
of real exchange
Sarno L.
1998. The
behavior
Taylor MP,
Economics
46: 281312.
Journal
of International
J, Chen
Economics
of panel
the power
in the presence
roots
for unit
22: 265312.
for panel
Shin
Wu
tests
interest
rates
rates
in
the
unit
during
eurocurrency
root
the
tests
under
postBretton
market.
cross
sectional
Woods
period.
Oxford
Bulletin
of
459473.
& Sons,
Ltd.
J.
Appl.
Econ.
This content downloaded from 202.185.32.3 on Sat, 09 May 2015 02:28:55 UTC
All use subject to JSTOR Terms and Conditions
22: 383400
(2007)
DOI:
10.1002/jae
Mult mai mult decât documente.
Descoperiți tot ce are Scribd de oferit, inclusiv cărți și cărți audio de la editori majori.
Anulați oricând.