Академический Документы
Профессиональный Документы
Культура Документы
p
jZ1
r
ij
Dy
i;tj
C3
it
iZ1; 2; 3; .N; tZ1; 2; 3; .T 3
8
According to Pedroni (2000, 2001), for heterogeneous cross-section data, the group-mean estimator is
much more useful than the within-group estimator in the sense that test statistics constructed using
group-mean estimators are designed to test the null hypothesis, so that the values under alternative
hypotheses are not constrained to be the same.
9
To reduce size distortion, the initial value in FMOLS should be a value under the null hypothesis or
a theory-based derived value.
10
Maddala and Wu (1999) and Choi (2001) also suggest the same kinds of panel unit root tests using
a Fisher statistic.
75 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
where p is selected to make the residuals uncorrelated over time. The null hypothesis
is that b
i
Z0 for all i, whereas the alternative hypothesis is that b
i
!0 for some i.
The ADF type t-statistics of IPS can be written as follows:
^
t
NT
Z
1
N
N
i
t
iT
p
i
4
where t
iT
( p
i
) is the ADF t-statistic for country i based on the country-specic ADF
regression, as in Eq. (3). IPS modied the standardized t-bar statistic as follows:
tZ
N
p
_
^
t
1
N
N
iZ1
Et
iT
p
i
; 0 : b
i
Z0
_
1
N
N
iZ1
Vart
iT
p
i
; 0 : b
i
Z0
_ 5
The t-bar statistic has a standard normal distribution as N and T0N and
N/T 0k, where k is a nite positive constant.
Recently, attention has been paid to cointegration tests and estimation in panel
models, including studies by Pedroni (1995, 1999, 2000), Kao (1999), Kao and
Chiang (2000), and Phillips and Moon (1999).
11
One of the most troublesome
questions in those papers is how to tackle heterogeneous short-run dynamics and
heterogeneity across members, for the heterogeneity problem seems to be prevalent
in panel data. We use the cointegration test and estimation method of Pedroni (1995,
1999, 2000) to handle the heterogeneity problem. As a general form, the following
type of regression will be considered.
IR
it
Za
i
Cd
t
Cb
i
SR
it
C3
it
iZ1; 2; 3.N; tZ1; 2; 3.T 6
This formulation allows for considerable heterogeneity: xed eect (a
i
), and
individual deterministic trend (d
i
) and a heterogeneous slope coecient (b
i
). The
regression form in Eq. (6) is very powerful in the sense that it incorporates all kinds
of country and time heterogeneities that are usual in the savingsinvestment relation.
IR
i
and SR
i
are assumed to be integrated of order one for each member i in panels.
Under the null hypothesis of no cointegration, the residual 3
it
is also integrated of
order one.
Pedroni (1995, 1999) proposes seven residual-based tests for the null of no
cointegration, allowing for heterogeneous xed eects and deterministic trends as
well as heterogeneous short-run dynamics. Four are based on pooling the residuals
of the regression for the within-group and the other three are based on pooling the
residuals for the between-group.
12
In both tests, rejection of the null hypothesis
11
For a recent literature survey, see Baltagi and Kao (2000).
12
Pedronis (1995) four within-group panel cointegration tests are the panel r-statistic, panel
parametric t-statistic (panel ADF), panel t-statistic, and panel variance ratio statistic. He proposes the
group r-statistic, non-parametric group t-statistic, and parametric group t-statistic (group ADF) as the
between-group test statistics for panel cointegration.
76 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
means that the variables under consideration are cointegrated. Pedronis (1995)
Monte Carlo simulation shows that the power of the between-group statistic is
higher than that of the within-group statistic in the case of small samples. Our work
will do cointegration tests using both the between-group and the within-group
methods.
3.2. The between-group panel FMOLS and DOLS estimators
Now, how should we estimate the cointegration vector? We will examine two
panel cointegration estimators: the between-group fully modied OLS (FMOLS)
and dynamic OLS (DOLS). The FMOLS is popular in conventional time series
econometrics, for it is believed to eliminate endogeneity in the regressors and serial
correlation in the errors. Pedroni (2000, 2001) proposes two methods to apply this
fully modied method to panel cointegration regression: the pooled (or within-
group) panel FMOLS estimator and the group-mean (between-group) FMOLS
estimator.
13
We will use the between-group FMOLS estimator, because it allows for
a more exible alternative hypothesis and suers much less from small sample size
distortion than the within-group estimator.
14
The group-mean panel FMOLS estimator for Eq. (6) can be written as:
^
b
GFM
Z
1
N
i
_
_
T
tZ1
_
SR
i;t
SR
i
_
IR
i;t
T
^
g
i
T
tZ1
_
SR
i;t
SR
i
_
2
_
_
7
where
IR
i;t
Z
_
IR
i;t
IR
i
_
^
U
21;i
^
U
22;i
DSR
i;t
and
^
g
i
Z
^
G
21i
C
^
U
0
21;i
^
U
21;i
^
U
22;i
^
G
22;i
C
^
U
0
22;i
Here
^
U
i
Z
^
U
0
i
C
^
G
i
C
^
G
i
#
is the estimated long-run covariance matrix of the stationary
vector consisting of the estimated residuals from the cointegration regression and the
dierences in savings rate.
^
U
0
21i
is the long-run covariance between the stationary
error terms (3
it
in Eq. (6)) and the unit root autoregressive disturbances.
^
U
2
22i
is the
long-run covariance among the dierence in savings rates.
^
G
i
is a weighted sum of the
autocovariances and a bar over these letters denotes the mean for i members.
15
The associated t-statistic for the between-group FMOLS estimator takes the
following form:
t
^
b
GFM
Z
1
N
p
N
iZ1
^
b
FM;i
b
_
^
U
1
11;i
t
_
SR
i;t
SR
i
_
2
_
1=2
8
13
Phillips and Moon (1999) also propose that the FM method yields an optimal estimator for the
cointegration coecient in Gaussian cointegration regression models.
14
Pedroni (2000) shows that the between-group FMOLS estimator has a much smaller size distortion
than the within-group estimator by Monte Carlo simulation, whereas both are unbiased.
15
Interested readers can refer to Pedroni (2000, 2001) for more details.
77 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
where b is a value under the null hypothesis. The above t-statistic is standard normal
as T and N approach innity.
Next, the DOLS panel cointegration estimator of the FH coecient is considered.
Kao and Chiang (2000) apply dynamic OLS (DOLS) to panel cointegration
estimation. Here, the DOLS estimator is slightly dierent from the original
formulation of Kao and Chiang (2000), because we are interested in the between-
group estimator. DOLS uses the past and future values of DSR
it
as additional
regressors. The between-group panel DOLS regression can be written as follows:
IR
i;t
Za
i
Cd
t
Cb
i
SR
i;t
C
Ki
kZKi
g
ik
DSR
i;tk
Cu
i;t
9
^
b
DOLS
Z
_
1
N
N
iZ1
_
T
tZ1
Z
i;t
Z
#
i;t
_1_
T
tZ1
Z
i;t
~
IR
i;t
__
1
10
where Z
i,t
is the 2(KC1)1 vector of regressors Z
i;t
ZSR
it
SR
i
; DSR
i;tk
; .;
DSR
i;tCk
,
~
IR
i;t
ZIR
i;t
IR
i
. A bar over a letter denotes a mean and the subscript 1
outside the brackets indicates the rst elements of the vector used to obtain the
pooled slope coecient. The associated t-statistic for the group-mean estimator can
be constructed as:
t
^
b
DOLS
Z
1
N
p
N
iZ1
_
^
b
D;i
b
_
_
1
^
s
2
i
t
_
SR
i;t
SR
i
_
2
_
1=2
11
where
^
s
2
i
is the long-run variance of the residuals from the DOLS regression and
^
b
D;i
is the conventional DOLS estimator. This t-statistic is standard normal as T and N
approach innity, as in the FMOLS estimator.
4. Empirical results
The data are taken from the International Financial Statistics of the IMF. The
data are annual and cover the period from 1960 to 1998 for 11 Asian countries.
16
We
check whether the individual time series data are nonstationary using ADF and PP
(PhillipsPerron test) and examine whether the panel data series have unit roots
using the LL and IPS methods.
Table 1 shows that the individual savings and investment rates are nonstationary,
except in Sri Lanka. Only the savings rate in Sri Lanka is stationary at the 5%
signicance level. Table 2 shows that the null of the unit roots for the panel data for
16
India, Indonesia, Japan, Korea, Malaysia, Myanmar, Pakistan, Philippines, Singapore, Sri Lanka,
and Thailand.
78 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
the savings and investment rates cannot be rejected by both the LL and IPS tests.
Therefore, we can implement a test for panel cointegration between savings rate and
investment.
Table 3 shows the outcomes of cointegration tests between the investment and
savings rates. We use three within-group tests and three between-group tests to check
whether the panel data are cointegrated. The results of the within-group test show
that the null hypothesis of no cointegration can be rejected at the 1% signicance
level. The between-group tests reject the null of no cointegration at the 5%
signicance level. Therefore, the savings and investment rates appear to be
cointegrated at a reasonable signicance level.
Finally, we estimate the cointegrating vector using two methods: the group-mean
FMOLS and DOLS estimators. We consider two cases: with time dummy and
without time dummy. A time dummy is usually included to check for cross-sectional
dependency across countries. In general, savings and investment rates can be aected
by the international business cycle. Therefore, a time dummy is included to control
for international business cycle eects.
Table 4 shows the estimate of the saving retention ratio by period, using the
between-group panel cointegration technique. First, we look at the case of a time
Table 1
Unit root tests for saving rate and investment rate by individual country, 19601998
Country Savings rate Investment rate
ADF test PP test ADF test PP test
India 1.44 1.74 1.71 1.27
Indonesia 1.14 0.96 0.97 1.06
Japan 0.87 1.42 2.06 2.09
Korea 1.43 1.41 1.99 1.76
Malaysia 0.34 0.22 2.07 1.58
Myanmar 1.74 1.77 2.77 2.73
Pakistan 1.67 1.53 3.16 3.14
Philippines 1.02 1.05 2.43 2.02
Singapore 0.41 0.47 1.68 1.56
Sri Lanka 3.39** 3.46** 1.88 1.90
Thailand 0.51 0.51 1.95 1.65
Notes: the number of lags is one in Myanmar and Pakistan and zero in others. ** denotes the 5%
signicance.
Table 2
Panel unit root tests for saving and investment rates, 19601998
Method Savings rate Investment rate
LL IPS LL IPS
t Value 1.03 1.82 1.23 1.16
Panel critical value
5% 5.43 1.88 5.43 1.88
10% 5.13 1.98 5.13 1.98
Notes: the null hypothesis in the above panel unit root is that all individual series are nonstationary.
79 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
dummy for each period. The group-mean FMOLS estimate of the FH coecient
is 0.54, while the DOLS estimate is 0.62, for the entire period. These estimates
are slightly smaller than those in other papers. However, the coecient is still
signicantly nonzero with either method. It is recognized that these values are similar
to the two-thirds typically found in empirical FH studies.
Now, we consider two sub-periods (19601979 and 19801998).
17
Table 4 shows
that the FH coecient using FMOLS is 0.39 for the period 19801998, whereas it
is 0.58 for 19601979. The DOLS estimate also decreases from 0.76 in the rst sub-
period to 0.42 in the second sub-period. The estimated coecients (0.39, 0.42) for the
period 19801998 are much lower than previous results,
18
although they are still
signicantly dierent from zero. Without the time dummy, the coecients are 0.37
(FMOLS) or 0.44 (DOLS) for the second sub-period, and are also much smaller than
those for the rst sub-period. These coecients seem very small, when we consider
the conclusion of FH, i.e., two-thirds of domestic savings stay in the country. It is
interesting that the estimates by FMOLS are smaller than or similar to those by
DOLS. These results contrast with the results of Ho (2002) that coecients estimated
by DOLS are much smaller than those by FMOLS. These estimated coecients
remind us of Corbins (2001) results. Her FH coecient was around 0.4 for the
period using the gold standard (18851913), when capital across countries was
almost perfectly mobile.
What do the signicant reduction of the FH coecients between the two sub-
periods and the relatively small numbers mean? It is well acknowledged that the
capital account liberalization policies and development of telecommunication and
information technology has increased capital movement across countries in Asia
since the 1980s. A signicant decrease in the FH coecient during the period 1980
1998 is consistent with the recognized facts. Therefore, a reduction in the HF
coecients for the period 19801998 seems to tell us that international capital
mobility has increased after the 1980s in Asian countries from the FH perspective.
However, the nonzero coecients imply that capital is not perfectly mobile across
these countries. In addition, these estimated results imply that the FH coecient is
still a good measure for international capital mobility, when it is estimated by panel
cointegration estimators.
Table 3
Panel cointegration tests, 19601998
Within-group Between-group
r-Stat Panel t-stat Panel ADF-stat r-Stat Group t-stat Group ADF-stat
3.48 3.51 3.72 1.98 2.80 2.97
Notes: the test statistics are normalized so that the asymptotic distribution is standard normal.
17
The period is divided into before and after 1980, because many Asian countries started to remove
capital controls in the 1980s.
18
Coakley et al. (1998) summarize the cross-section and time series coecients in previous researches.
The former is above 0.62 except in one case and the latter above 0.6 with several exceptional cases.
80 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
5. Conclusion
This paper has studied the international capital mobility of Asian countries in
terms of the FH coecient, applying recently developed panel cointegration
methods. We apply two classes of panel cointegration test and used between-group
FMOLS and DOLS estimators in order to control for heterogeneous short-run
dynamics and heterogeneous error terms.
It is found that the savings and investment rates in panel data are nonstationary
but that they are cointegrated. The FH coecients using FMOLS and DOLS are
0.39 and 0.42, respectively, for the period 19801998, which is much smaller than
those in other previous researches, whereas they are 0.58 and 0.76 for 19601979.
The estimated FH coecients after 1980 are almost the same as those for the period
of the gold standard system, when capital is considered to have been highly mobile
between countries. The decrease in the estimated FH coecient from 0.58 to 0.39
(group-mean FMOLS) or from 0.76 to 0.42 (group-mean DOLS) might be
interpreted as evidence that capital has been more mobile after the 1980s in Asian
countries. This increasing capital mobility might reect capital account liberaliza-
tion. Finally, the FH coecient might be a good candidate for measuring the degree
of international capital mobility, when it is estimated in the between-group panel
cointegration frameworks.
Acknowledgements
We gratefully acknowledge the helpful comments of anonymous referees. We are
grateful to Peter Pedroni for his sincere and helpful comments.
This work was supported by the Korea Research Foundation Grant (KRF-2000-
013-CA0120).
References
Baltagi, H.B., Kao, C., 2000. Nonstationary panels, cointegration in panels and dynamic panels: a survey.
Advances in Econometrics 15, 751.
Table 4
Panel cointegration estimation
Period 19601998 19601979 19801998
Panel with
time dummy
GM FMOLS 0.54(18.39) 0.58(14.58) 0.39(6.26)
GM DOLS 0.62(21.11) 0.76(16.11) 0.42(7.22)
Panel without
time dummy
GM FMOLS 0.69(21.70) 0.61(17.84) 0.37(6.34)
GM DOLS 0.84(26.25) 0.76(18.82) 0.44(1.60)
Notes: the value in parenthesis denotes the t-value for zero coecient. As explained in Section 3,
asymptotic distribution of t-statistic is standard normal as T and N go to innity. **GM denotes the
group-mean (or between-group).
81 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182
Bayoumi, T., 1990. Savinginvestment correlations: immobile capital, government policy or endogenous
behavior? IMF Sta Papers 37, 360387.
Choi, I., 2001. Unit root test for panel data. Journal of International Money and Finance 20, 249272.
Coakley, J., Kulasi, F., Smith, R., 1998. The Feldstein and Horioka puzzle and capital mobility: a review.
International Journal of Finance and Economics 3, 169188.
Coakley, J., Fuertes, A.M., Spagnolo, F., 2001. The FeldsteinHorioka puzzle is not as bad as you think.
Available from: http://papers.ssrn.com/sol3/papers.cfm.
Coiteux, M., Olivier, S., 2000. The saving retention coecient in the long run and in the short run:
evidence from panel data. Journal of International Money and Finance 19, 535548.
Corbin, A., 2001. Country specic eect in the FeldsteinHorioka paradox: a panel data analysis.
Economics Letters 72, 297302.
Feldstein, M., Horioka, C., 1980. Domestic saving and international capital ow. The Economic Journal
90, 314329.
Ho, Tsung-wo, 2002. The FeldsteinHorioka puzzle revisited. Journal of International Money and
Finance 21, 555564.
Im, K.S., Pesaran, M.H., Shin, Y.C., 1996. Testing for Unit Roots in Heterogeneous Panels. Discussion
Paper, University of Cambridge.
Jansen, W.J., 2000. International capital mobility: evidence from panel data. Journal of International
Money and Finance 19, 507511.
Kao, C., 1999. Spurious regression and residual-based tests for cointegration in panel data. Journal of
Econometrics 90, 144.
Kao, C., Chiang, M.H., 2000. On the estimation and inference of a cointegrated regression in panel data.
Advances in Econometrics 15, 179222.
Krol, R., 1996. International capital mobility: evidence from panel data. Journal of International Money
and Finance 15, 467474.
Levin, A., Lin, C.F., 1992. Unit Root Tests in Panel Data: Asymptotic and Finite Sample Properties.
UCSD Discussion Paper 9223.
Maddala, G.S., Wu, S., 1999. A comparative study of unit root tests with panel data and a new simple test.
Oxford Bulletin of Economics and Statistics, special issues 631652.
Moon, H.R., Phillips, P.C.B, 1998. A Reinterpretation of the FeldsteinHorioka Regression from
a Nonstationary Panel Viewpoint. Mimeo, Yale University.
Obstfeld, M., 1993. International Capital Mobility in the 1990s. NBER Working Paper, Working Paper
no. 4534.
Pedroni, P., 1995. Panel Cointegration: Asymptotic and Finite Sample Properties of Pooled Time
Series Tests, With an Application to the PPP Hypothesis. Working Paper no. 95-013, Indiana
University.
Pedroni, P., 1999. Critical values for cointegration tests in heterogeneous panels with multiple regressors.
Oxford Bulletin of Economics and Statistics, special issues 653670.
Pedroni, P., 2000. Fully modied OLS for heterogeneous cointegrated panels. Advances in Econometrics
15, 93130.
Pedroni, P., 2001. Purchasing power parity tests in cointegrated panels. Review of Economics and
Statistics 83, 723741.
Phillips, P.C.B., Moon, H., 1999. Linear regression limit theory for nonstationary panel data.
Econometrica 67, 10571111.
Taylor, A.M., 1996. International Capital Mobility in History: The Saving and Investment Relationship.
NBER Working Paper, Working Paper no. 5743.
Tesar, L.L., 1991. Saving, investment and international capital ows. Journal of International Economics
31, 5578.
82 H. Kim et al. / Journal of International Money and Finance 24 (2005) 7182