ORIGINAL PAPER Export-led growth in the UAE: multivariate causality between primary exports, manufactured exports and economic growth Athanasia S. Kalaitzi 1 • Emmanuel Cleeve 1 Received: 2 November 2016 / Revised: 2 August 2017 / Accepted: 10 August 2017 / Published online: 18 September 2017 Ó The Author(s) 2017, corrected publication 11/2017. This article is an open access publication Abstract The principal question that this research addresses is the validity of the Export-Led Growth hypothesis (ELG) in the United Arab Emirates (UAE) over the period 1981–2012, focusing on the causality between primary exports, manufac- tured exports and economic growth. Unit root tests are applied to examine the time- series properties of the variables, while the Johansen cointegration test is performed to confirm or not the existence of a long-run relationship between the variables. Moreover, the multivariate Granger causality test and a modified version of Wald test are applied to examine the direction of the short-run and long-run causality respectively. The cointegration analysis reveals that manufactured exports con- tribute more to economic growth than primary exports in the long-run. In addition, this research provides evidence to support a bi-directional causality between man- ufactured exports and economic growth in the short-run, while the Growth-Led Exports (GLE) hypothesis is valid in the long-run for UAE. Keywords Export-led Growth Diversification Economic growth Causality UAE JEL Classification O47 F43 C22 & Athanasia S. Kalaitzi [email protected]; [email protected]Emmanuel Cleeve [email protected]1 Department of Accounting, Finance and Economics, Manchester Metropolitan Business School, All Saints Campus, Oxford Road, Manchester M15 6BH, UK 123 Eurasian Bus Rev (2018) 8:341–365 https://doi.org/10.1007/s40821-017-0089-1
25
Embed
Export-led growth in the UAE: multivariate causality between primary exports ... · 2018-08-13 · Export-led growth in the UAE: multivariate causality between primary exports, ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
ORI GIN AL PA PER
Export-led growth in the UAE: multivariate causalitybetween primary exports, manufactured exportsand economic growth
Athanasia S. Kalaitzi1 • Emmanuel Cleeve1
Received: 2 November 2016 / Revised: 2 August 2017 / Accepted: 10 August 2017 /
Published online: 18 September 2017
� The Author(s) 2017, corrected publication 11/2017. This article is an open access publication
Abstract The principal question that this research addresses is the validity of the
Export-Led Growth hypothesis (ELG) in the United Arab Emirates (UAE) over the
period 1981–2012, focusing on the causality between primary exports, manufac-
tured exports and economic growth. Unit root tests are applied to examine the time-
series properties of the variables, while the Johansen cointegration test is performed
to confirm or not the existence of a long-run relationship between the variables.
Moreover, the multivariate Granger causality test and a modified version of Wald
test are applied to examine the direction of the short-run and long-run causality
respectively. The cointegration analysis reveals that manufactured exports con-
tribute more to economic growth than primary exports in the long-run. In addition,
this research provides evidence to support a bi-directional causality between man-
ufactured exports and economic growth in the short-run, while the Growth-Led
Exports (GLE) hypothesis is valid in the long-run for UAE.
In order to examine whether a causal relationship exists between primary exports,
manufactured exports and economic growth in the UAE, the following tests are
applied: (a) Unit root tests in order to examine the stationarity of the variables
included in the model, (b) Cointegration test to confirm or not the Export-led
Growth hypothesis and (c) the multivariate Granger causality test and the modified
Wald test (MWALD) to investigate the direction of the short-run and long-run
causality respectively.
3.3 Unit root test
Initially, the Augmented Dickey-Fuller (ADF) test is conducted in order to test for
the presence of a unit root (Enders 1995). The ADF test is based on the following
three equations:
DYt ¼ cYt�1 þXp
i¼1biDYt�i þ et ð5Þ
DYt ¼ a0 þ cYt�1 þXp
i¼1biDYt�i þ et ð6Þ
DYt ¼ a0 þ cYt�1 þ a2t þXp
i¼1biDYt�i þ et; ð7Þ
where a0 and a2 represents the deterministic elements.
Equation (5) is a random walk, Eq. (6) is a random walk with intercept only,
while the last equation is a random walk with intercept and time trend (Gujarati
2003). In addition the random errors are assumed to be uncorrelated and identically
distributed with zero mean and variance r2 {et * ii (0, r2) for t = 1, 2,���}. In
each case, the null hypothesis is that c = 0; Ho the time series is not stationary, while
the alternative hypothesis is that c\ 0; Ha the time series is stationary.
In addition, this research applies the Phillips and Perron (1988) test (PP), which is
a generalization of the Dickey-Fuller procedure that allows for serial correlation and
heteroskedasticity in the error terms (Enders 1995). This test involves the following
equations:
Yt ¼ c�0 þ c�1yt�1 þ lt ð8Þ
Yt ¼ c�0 þ c�1yt�1 þ c�2 t � T/2ð Þ þ lt; ð9Þ
where c*0 and c*1 are the deterministic elements, T is the number of observations,
while lt is the error term. In particular, the Phillips-Perron t-statistics are modifi-
cations of the ADF t-statistics that take into account the less restrictive nature of the
error process (Enders 1995). The PP test is performed by following the method
suggested by Doldado et al. (1990) regarding the inclusion of constant and trend.
Fig. 1 Plots of the time series, 1981–2012. Source: Gross Domestic Product is taken from the WDI-World Bank, Gross Fixed Capital formation and Imports are taken from IFS- IMF (years 1999–2000 aretaken from UAE National Bureau of Statistics and years 2010–2012 are taken from World Bank). Primaryand Manufactured exports are obtained from WTO- Time Series on International Trade and population isobtained from UAE National Bureau of Statistics. The graphs are produced by using the econometricsoftware Eviews 8
b
Eurasian Bus Rev (2018) 8:341–365 349
123
According to Verbeek (2012, p. 294) ‘‘Not all series for which we cannot reject
the unit root hypothesis are necessarily integrated of order one’’. For this reason this
research also applies the test proposed by Kwiatkowski et al. (1992), where the null
hypothesis is a stationary process. The Kwiatkowski-Phillips-Schmidt-Shin (KPSS)
statistic is based on the residuals from the Ordinary Least Squares (OLS) regression
of yt on the exogenous variables xt (constant and time trend):
Yt ¼ d x0
t þ ut
The KPSS statistic is defined as: KPSS =P
tSðtÞ2.
T2f0ð ÞWhere f0 is an estimator of the residual spectrum at frequency zero and S(t) is a
cumulative residual function: S tð Þ ¼Pt
r¼1 ur; based on the residuals ut from the
equation Yt ¼ d x0t þ ut.
However, when applying the ADF, PP and KPSS tests, a structural break can be
identified as evidence of non-stationarity, even if the series is stationary within each
of the periods before and after the break. Since UAE economy has been subject to
oil shocks during the period 1981–2012, the Saikkonen and Lutkepohl (2002) unit
root test (SL) with a structural break is conducted in order to assess the stationarity
of the variables. The SL test involves the following models:
Yt ¼ lo þ l1t þ dd1t þ ut ð10Þ
Yt ¼ lo þ dd1t þ ut; ð11Þ
where l0 is the constant term, l1 and d are the coefficients of the trend term and the
shift dummy variable respectively, while ut is the error term. In particular, d1t is a shift
dummy variable with break date Tbreak: d1t = 0, for t\Tbreak and d1t = 1, for t[Tbreak.
3.4 Cointegration test
After testing for stationarity of the variables, the Johansen cointegration test
(Johansen 1988, 1995) is performed in order to confirm the existence of a long-run
relationship between the variables. The Johansen cointegration test is considered to
have better properties than the other cointegration tests, such as the two-step Engle-
Granger Cointegration technique (Engle and Granger 1987). As Gonzalo (1994)
notes, Johansen’s cointegration test satisfies the three elements in a cointegration
system, ‘‘first the existence of unit roots, second the multivariate aspect, and third
the dynamics. Not taking these elements into account may create problem is
estimation’’ (Gonzalo 1994, p. 223). Johansen’s methodology estimates cointegrat-
ing vectors using a maximum likelihood procedure, taking its starting point in the
VAR of order p given by:
Xt ¼ lþXp
i¼1AiXt�i þ et; ð12Þ
350 Eurasian Bus Rev (2018) 8:341–365
123
where Xt is a (nx1) vector of variables that are I(1), l is a (nx1) vector of constants,
Ai is an (n x n) matrix of parameters, while et is a (nx1) vector of random errors.
Subtracting Xt -1 from each side of this equation and letting I be an (nxn) identity
matrix, this VAR can be re-written as:
DXt ¼ lþPXt�1 þXp�1
i¼1CiDXt�i þ et ð13Þ
whereCi ¼ �Xp
j¼iþ1Aj; P ¼
Xp
i¼1Ai� I:
D is the difference operator, Ci and P are the coefficient matrices, while the rank of
matrix P provides information about the number of cointegrating vectors. In the case
where the coefficient matrix P has rank r\ n, but not equal to zero, this means that
there is cointegration and r is the number of cointegrating vectors. It is important to
note that in a VAR model with n variables, there can be at most r = n - 1 cointegrating
relationships. In this case, P can be expressed as P = ab’ where a and b are n x r
matrices. The elements of the matrix a are known as the adjustment matrix parameters
in the vector error correction model and the matrix b is the cointegrating matrix. The
number of cointegrating vectors can be determined by using the likelihood ratio (LR)
trace test statistic suggested by Johansen (1988). The LR trace statistic is adjusted for
small sample size, as proposed by Reinsel and Ahn (1992).1
The LR trace statistic is given by the following equation:
Jtrace ¼ �TXn
i¼rþ1lnð1 � kiÞ; ð14Þ
where T is the sample size and k is the eigenvalue. The trace test, tests the null
hypothesis of at most r cointegrating vectors against the alternative hypothesis of n
cointegrating vectors.
3.5 Short-run Granger causality test
In order to conduct the Granger causality test, a VAR model is estimated, by
including the optimal lag length of each variable in each equation (Gujarati 2003).
The VAR model with six endogenous variables (LYt, LKt, LLt, LPXt, LMXt,
LIMPt) can be expressed as follows:
LYt ¼ a10 þXp
j¼1b1jLYt�j þ
Xp
j¼1c1jLKt�j þ
Xp
j¼1d1jLLt�j þ
Xp
j¼1f1jLPXt�j
þXp
j¼1h1jLMXt�j þ
Xp
j¼1l1jLIMPt�j þ e1t ð15Þ
LKt ¼ a20 þXp
j¼1b2jLYt�j þ
Xp
j¼1c2jLKt�j þ
Xp
j¼1d2jLLt�j þ
Xp
j¼1f2jLPXt�j
þXp
j¼1h2jLMXt�j þ
Xp
j¼1l2jLIMPt�j þ e2t ð16Þ
1 Trace statistics is adjusted by using the correction factor (T – n*p)/T proposed by Reinsel and Ahn
(1992), where T is the sample size, while n and p is the number of the variables and the optimal lag length
respectively.
Eurasian Bus Rev (2018) 8:341–365 351
123
LLt ¼ a30 þXp
j¼1b3jLYt�j þ
Xp
j¼1c3jLKt�j þ
Xp
j¼1d3jLLt�j þ
Xp
j¼1f3jLPXt�j
þXp
j¼1h3jLMXt�j þ
Xp
j¼1l3jLIMPt�j þ e3t ð17Þ
LPXt ¼ a40 þXp
j¼1b4jLYt�jþ
Xp
j¼1c4jLKt�jþ
Xp
j¼1d4jLLt�jþ
Xp
j¼1f4jLPXt�j
þXp
j¼1h4jLMXt�jþ
Xp
j¼1l4jLIMPt�jþ e4t ð18Þ
LMXt ¼ a50 þXp
j¼1b5jLYt�jþ
Xp
j¼1c5jLKt�jþ
Xp
j¼1d5jLLt�jþ
Xp
j¼1f5jLPXt�j
þXp
j¼1h5jLMXt�jþ
Xp
j¼1l5jLIMPt�jþ e5t ð19Þ
LIMPt ¼ a60 þXp
j¼1b6jLYt�jþ
Xp
j¼1c6jLKt�jþ
Xp
j¼1d6jLLt�jþ
Xp
j¼1f6jLPXt�j
þXp
j¼1h6jLMXt�jþ
Xp
j¼1l6jLIMPt�jþ e6t: ð20Þ
LYt represents the variable of economic growth, while LKt, LLt, LPXt, LMXt,
and LIMPt represent the right-hand side variables of the Eq. (4). In addition,
exogenous variables can be included in the VAR model, such as structural breaks,
without adding equations to the system. In the case where the variables are
cointegrated, the causality can be tested by estimating the following restricted VAR
model (VECM):
DLYt ¼Xp
j¼1b1jDLYt�jþ
Xp
j¼1c1jDLKt�jþ
Xp
j¼1d1jDLLt�jþ
Xp
j¼1f1jDLPXt�j
þXp
j¼1h1jDLMXt�jþ
Xp
j¼1l1jDLIMPt�j� kyECTt�1 þ e1t; ð21Þ
DLKt ¼Xp
j¼1b2jDLYt�jþ
Xp
j¼1c2jDLKt�jþ
Xp
j¼1d2jDLLt�jþ
Xp
j¼1f2jDLPXt�j
þXp
j¼1h2jDLMXt�jþ
Xp
j¼1l2jDLIMPt�j� kkECTt�1 þ e2t ð22Þ
DLLt ¼Xp
j¼1b3jDLYt�j þ
Xp
j¼1c3jDLKt�j þ
Xp
j¼1d3jDLLt�j þ
Xp
j¼1f3jDLPXt�j
þXp
j¼1h3jDLMXt�j þ
Xp
j¼1l3jDLIMPt�j � kLECTt�1 þ e3t ð23Þ
DLPXt ¼Xp
j¼1b4jDLYt�jþ
Xp
j¼1c4jDLKt�jþ
Xp
j¼1d4jDLLt�jþ
Xp
j¼1f4jDLPXt�j
þXp
j¼1h4jDLMXt�jþ
Xp
j¼1l4jDLIMPt�j� kpxECTt�1 þ e4t ð24Þ
DLMXt ¼Xp
j¼1b5jDLYt�jþ
Xp
j¼1c5jDLKt�jþ
Xp
j¼1d5jDLLt�jþ
Xp
j¼1f5jDLPXt�j
þXp
j¼1h5jDLMXt�jþ
Xp
j¼1l5jDLIMPt�j�kmxECTt�1 þ e5t ð25Þ
352 Eurasian Bus Rev (2018) 8:341–365
123
DLIMPt ¼Xp
j¼1b6jDLYt�j þ
Xp
j¼1c6jDLKt�j þ
Xp
j¼1d6jDLLt�j þ
Xp
j¼1f6jDLPXt�j
þXp
j¼1h6jDLMXt�j þ
Xp
j¼1l6jDLIMPt�j � kimpECTt�1 þ e6t
ð26Þ
where D is the difference operator, bij, cij, dij, fij, hij, lij and kij are the regression
coefficients and ECTt -1 is the error correction term derived from the cointegration
equation.
Once the models have been estimated, multivariate specification tests are used to
assess whether the models are well specified and stable. In particular, these tests
include the Jarque-Bera Normality test, the Portmanteau and Breusch-Godfrey LM
test for the existence of autocorrelation, the White Heteroskedasticity test, the
Multivariate ARCH test and the AR roots stability test. In addition, the cumulative
sum of recursive residuals (CUSUM) and the CUSUM of squares (CUSUMQ) tests
are applied in order to assess the parameter constancy of the ECM estimates.
Specifically, the CUSUM test detects systematic changes, while CUSUMQ test
detects haphazard changes in the parameters (Brown et al. 1975). The CUSUM test
proposed by Brown et al. (1975) is based on the statistic:
Wt ¼Xt
kþ1wt= s t ¼ k þ 1; . . .T ; ð27Þ
where s is the standard deviation of the recursive residuals (wt), which is defined as:
wt ¼ yt� x0tbt�1
� ��1 þ x0t Xt�1
0Xt�1ð Þ�1xt
� �1=2
;
where the numerator yt - x0t bt–1 is the forecast error, bt-1 is the estimated coefficient
vector up to period t-1 and xt0 is the row vector of observations on the regressors in
period t. The Xt-1 denotes the (t-1) 9 k matrix of the regressors from period 1 to
period t–1. If the b vector changes, Wt will tend to diverge from the zero mean value
line, while if b vector remains constant, E(Wt) = 0. The test shows parameter
instability if the cumulative sum of the recursive residuals lies outside the area
between the two 5% critical lines, the distance between which increases with t.
The CUSUM of Squares test uses the square recursive residuals, wt2 and is based
on the plot of the statistic:
St ¼Xt
kþ1
w2t
!=XT
kþ1
w2t
!; where t = k + 1; . . .; T: ð28Þ
The expected value of St, under the null hypothesis of bt’s constancy is E(St) =
(t - k)/(T - k), which goes from zero at t = k to unity at t = T. In this test the St are
plotted together with the 5% significance lines and, as in the CUSUM test,
movements outside the 5% critical lines of parameter stability, indicates instability
in the equation during the sample period. In the case, where CUSUM test or
CUSUMQ test shows evidence of structural instability, an exogenous variable
should be included in order to obtain more efficient estimates.
Eurasian Bus Rev (2018) 8:341–365 353
123
After estimating the restricted or unrestricted VAR model and assessing the
constancy of the estimated parameters, this research applies the multivariate
causality test (Granger 1969, 1988). The causality from primary exports to
economic growth can be examined by conducting the Chi square test and the null
hypothesis ‘‘primary exports do not Granger cause economic growth’’
(H0 :Pp
j¼1 f1j ¼ 0) is tested against the alternative hypothesis ‘‘primary exports
Granger cause economic growth’’ (HA :Pp
j¼1 f1j 6¼ 0). To examine the causality
from manufactured exports to economic growth, the null hypothesis ‘‘manufactured
exports do not Granger cause economic growth’’ (H0 :Pp
j¼1 h1j ¼ 0) is tested
against the alternative hypothesis ‘‘manufactured exports Granger cause economic
growth’’ (HA :Pp
j¼1 h1j 6¼ 0). Moreover, the null hypothesis ‘‘economic growth
does not Granger cause primary exports’’ (H0 :Pp
j¼1 b4j ¼ 0) is tested against the
alternative hypothesis ‘‘economic growth Granger causes primary exports’’
(HA :Pp
j¼1 b4j 6¼ 0). Finally, the null hypothesis ‘‘economic growth does not
Granger cause manufactured exports’’ (H0 :Pp
j¼1 b5j ¼ 0) is tested against the
alternative hypothesis ‘‘economic growth Granger causes manufactured exports’’
(HA :Pp
j¼1 b5j 6¼ 0).
3.6 Long-run Granger causality test
This paper also examines the long-run causality between disaggregated exports and
economic growth in the UAE. The majority of the recent studies examine the
direction of the long-run causality in an ECM context (Herzer et al. 2006; Awokuse
2007; Mishra 2011; Hosseini and Tang 2014). In these studies, the long-run
causality is determined by the significance of the error correction coefficient in each
equation. Nevertheless, in the case of multivariate ECMs, it is not possible to
indicate which explanatory variable causes the dependent variable. For example, if
the coefficient ky of the error correction term in equation (21) is significant (ky =0),
a long-run causality runs from the explanatory variables to the dependent variable,
but it is not possible to indicate whether exports cause economic growth. For this
reason, this paper applies the modified version of the Granger causality test
(MWALD) proposed by Toda and Yamamoto (1995). In the present paper, the Toda
and Yamamoto Granger causality test involves the following model:
LYt ¼ a10 þXpþdmax
j¼1b1jLYt�j þ
Xpþdmax
j¼1c1jLKt�j þ
Xpþdmax
j¼1d1jLLt�j
þXpþdmax
j¼1f1jLPXt�j þ
Xpþdmax
j¼1h1jLMXt�j þ
Xpþdmax
j¼1l1jLIMPt�j
þ e1t ð29Þ
LKt ¼ a20 þXpþdmax
j¼1b2jLYt�j þ
Xpþdmax
j¼1c2jLKt�j þ
Xpþdmax
j¼1d2jLLt�j
þXpþdmax
j¼1f2jLPXt�j þ
Xpþdmax
j¼1h2jLMXt�j þ
Xpþdmax
j¼1l2jLIMPt�j
þ e2t ð30Þ
354 Eurasian Bus Rev (2018) 8:341–365
123
LLt ¼ a30 þXpþdmax
j¼1b3jLYt�j þ
Xpþdmax
j¼1c3jLKt�j þ
Xpþdmax
j¼1d3jLLt�j
þXpþdmax
j¼1f3jLPXt�j þ
Xpþdmax
j¼1h3jLMXt�j þ
Xpþdmax
j¼1l3jLIMPt�j
þ e3t ð31Þ
LPXt ¼ a40 þXpþdmax
j¼1b4jLYt�j þ
Xpþdmax
j¼1c4jLKt�j þ
Xpþdmax
j¼1d4jLLt�j
þXpþdmax
j¼1f4jLPXt�j þ
Xpþdmax
j¼1h4jLMXt�j þ
Xpþdmax
j¼1l4jLIMPt�j
þ e4t ð32Þ
LMXt ¼ a50 þXpþdmax
j¼1b5jLYt�j þ
Xpþdmax
j¼1c5jLKt�j þ
Xpþdmax
j¼1d5jLLt�j
þXpþdmax
j¼1f5jLPXt�j þ
Xpþdmax
j¼1h5jLMXt�j þ
Xpþdmax
j¼1l5jLIMPt�j
þ e5t ð33Þ
LIMPt ¼ a60 þXpþdmax
j¼1b6jLYt�j þ
Xpþdmax
j¼1c6jLKt�j þ
Xpþdmax
j¼1d6jLLt�j
þXpþdmax
j¼1f6jLPXt�j þ
Xpþdmax
j¼1h6jLMXt�j þ
Xpþdmax
j¼1l6jLIMPt�j
þ e6t ð34Þ
where p is the optimal lag length, selected by minimizing the value of Schwartz
Information Criterion (SIC), while dmax is the maximum order of integration of the
variables in the model. In particular, the selected lag length (p) is augmented by the
maximum order of integration (dmax) and the Chi square test is applied to the first p
VAR coefficients.
4 Empirical results
4.1 Unit root test
Table 2 presents the results of the ADF, PP, KPSS and SL unit root tests at levels.
The ADF, PP and SL results indicate that the null hypothesis of non-stationarity
cannot be rejected for all the variables at 5% significance level. In addition, the
KPSS test results indicate that the null hypothesis of stationarity is rejected for all
the variables at conventional levels of significance.
In contrast, after taking the first difference of LY, LK, LPX, LMX and LIMP, the
null hypothesis of unit root can be rejected at 1% level of significance, while the
first-differenced series of LL is found to be stationary at 5% significance level. In
addition, the KPSS unit root test results indicate that the null hypothesis of
stationary process cannot be rejected for all the variables at 1% significance level.
Hence, all the test results indicate that the time series for the period 1981–2012 are
integrated of order one I(1).
Eurasian Bus Rev (2018) 8:341–365 355
123
Since all variables are I(1), the cointegration test is applied to investigate the
existence of a long-run relationship between the variables. The results are presented
in Table 3.
Table 2 ADF, PP, KPSS and SL test results at logarithmic level