EXPORT-LED GROWTH HYPOTHESIS IN THE MENA REGION: A MULTIVARIATE COINTEGRATION, CAUSALITY AND STABILITY ANALYSIS HUSEIN, Jamal* Abstract
Empirical evidence supporting the export-led growth (ELG) hypothesis has been mixed and inconclusive. Some studies may have been misspecified since they tested the ELG hypothesis using bivariate models. Others used multivariate cointegration framework but presumed (either explicitly or implicitly) stability of the cointegrating relations. This study examines the ELG hypothesis for eight Middle East and North Africa (MENA) countries in a multivariate framework by including terms of trade as a third variable. We utilize Johansen and Juselius cointegration procedure and error correction modeling to test the ELG hypothesis. The empirical evidence supports the existence of a “stable”
long-run equilibrium relationship among real output, real exports, terms of trade, and finds strong support for the ELG hypothesis in all but one of the MENA countries analyzed.
JEL Classification: C32, F43, O53
Keywords: Economic growth, Exports, Cointegration, Causality, Stability 1. Introduction
Many empirical studies have sought to test the validity of the export-led growth hypothesis (ELG), which tests whether export growth driven by export promotion policies enhances overall economic growth. Central to this debate is whether strong economic performance can be led by exports, as proponents of the ELG hypothesis contend, or can be growth driven. As such, the determination of the causality pattern between exports and economic growth has important policy implications concerning the appropriate growth and development strategies and policies to pursue. In early work that studied the relationship between exports and economic growth, a positive correlation between exports and output growth or a significant positive coefficient of the export variable in the growth equation (in simple or multiple OLS regression) was considered a confirmation of the ELG hypothesis (Balassa 1978, 1982, 1984; Feder 1983; Dollar 1992;
and many others). However, these studies did not confirm the direction of such a causal relationship between exports and economic growth. A positive and significant coefficient in the growth equation can be compatible with causality from exports to economic growth, from economic growth to exports (known as the growth-led exports (GLE) hypothesis), or a two-way (bidirectional) causality between the two variables.
Since early work in this area adopted a cross-sectional framework and did not examine the issue of the direction of causality between export growth and output growth, a recent number of country specific time-series studies, adopting both bivariate and multivariate models, have tested the validity of the ELG hypothesis. However, the empirical evidence based on those studies is mixed and often contradictory. The lack of consistent causal pattern between exports and economic growth in ELG studies in general
*Jamal Husein, Angelo State University, Accounting, Economics and Finance Department, ASU Station # 10908, San Angelo, TX 76909-0908, USA, email: [email protected]
162
and for the MENA region in particular may be explained, in part, by differences in the measures of exports used, the sampling period, and methodologies adopted. 1
This study investigates the causal relationship between exports and economic growth for eight MENA countries: Algeria, Egypt, Iran, Israel, Morocco, Sudan, Tunisia, and Turkey. We examine those MENA countries, where data is available, to assess whether promoting exports may contribute to economic growth in light of their shift to relatively more liberal trade regimes in the 1980s and 1990s.
Many developing countries, including those in the MENA region, have pursued trade liberalization and export promotion policies as a desirable strategy for development and growth, especially after the strong export performance and enviable economic growth of some East Asian economies. This change in the development strategy is based on the expectations that trade liberalization in general and export promotion in particular would result in the following: i) exports expand the market and help exploit economies of scale;
ii) exports relax the binding foreign exchange constraint to allow importation of needed capital and intermediate goods, for domestic production and exports, thus expanding the economy’s production possibilities; iii) exports enhance domestic efficiency via increased competition; and iv) exports promote the diffusion of technical knowledge through foreign buyers’ suggestions and learning by doing.
This study contributes to the existing relevant literature in four ways. First, it applies recent techniques in time series analysis by using Johansen and Juselius (1990) multivariate cointegration frameworks and error correction models to test for causality between exports and output growth. In testing the causality, we tackle the critical issues of unit root testing and lag length that some previous studies have failed to address.
Second, none of the studies for MENA countries included the terms of trade as an additional variable. The inclusion of a third relevant variable, terms of trade, corrects for misspecification bias associated with many bivariate models that tested the validity of the ELG hypothesis. Terms of trade inclusion reflects the possible linkage of the real exchange rate (and the possible effects of “restrictive” trade policies) and real output.
Moreover, terms of trade plays a significant role in influencing the size of imports, exports, and the trade balance. Third, this study uses the most up to date annual data and a relatively larger sample size to examine the causal relationship between exports and economic growth for the MENA countries. Fourth, unlike many ELG studies for developed and developing countries that assumed (either explicitly or implicitly) stability of the cointegrating vector, this study tests the constancy of the estimates by applying Hansen-Johansen (1999) formal stability tests.
2. The literature
Most ELG studies on the MENA countries did not apply cointegration tests and error correction models when testing for causality between exports and economic growth.
In the presence of cointegration, causality inferences based on standard Granger causality (SGC) tests are inappropriate (Granger, 1988). In addition, the few studies that adopted cointegration tests used Engle and Granger (1987) cointegration methodology rather than the more reliable Johansen (1988) and Johansen and Juselius (1990) method.
1MENA region includes 21 Arabic speaking countries (Arab League) in addition to Turkey, Iran, and Israel.
The following brief survey of previous studies focuses on time-series analyses that tested the ELG hypothesis for the MENA region. Using a bivariate SGC, Jung and Marshall (1985) found support for ELG in Egypt (1965-1979), no causality in the cases of Morocco, Tunisia, and Turkey, and negative bidirectional causality for Israel (1950- 1978). Hutchinson and Singh (1992) used GDP without exports in a bivariate model to test the ELG hypothesis for the period 1950-1985. They found no evidence of causality in Egypt, Morocco, and Tunisia. Kugler and Dridi (1993) used Johansen’s methodology in a bivariate model and found no evidence of cointegration in Egypt. In another bivariate SGC study, Dodaro (1993) found no evidence of causality between GDP growth and growth of real exports in Algeria, Jordan, Morocco, Sudan, and Tunisia. However, they found evidence of causality in the cases of Egypt and Israel.
Reizman et al. (1996) employed a bivariate SGC test and found evidence of causality in Algeria, Egypt, and Tunisia. They found no evidence of causality in Israel, Jordan, Morocco, Sudan, and Turkey. However, when they included imports as a third variable, the ELG hypothesis was supported in Jordan and Sudan. Pompino (1996) used cointegration and error-correction model (ECM) and found support for unidirectional causality from output growth to exports, i.e., growth-led exports (GLE), in Algeria and Tunisia, but found no causality for Morocco, Sudan, and Turkey. When Pompino introduced investment as a third variable, bidirectional causality was found for Turkey and Tunisia, but no change was found in the results for Algeria, Morocco, and Sudan.2
Xu (1996) used cointegration and ECM approaches and found no evidence of long-run relationship between exports and economic growth in Israel, Morocco, Tunisia, and Turkey. He found support for GLE in Israel and Tunisia.
In a bivariate study, Abual-Foul (2004) tested the validity of the ELG hypothesis for Jordan over the period 1976-1997. Citing a lack of a long sample for Jordan, the study carried no unit root tests and hence did not test for cointegration. As an alternative, the study employed three bivariate models, namely, a VAR in levels (assumed data is stationary), a VAR in first differences (the assumption is data is integrated of order one), and an ECM (assumed same level of integration, and the presence of cointegration).
Based on the three bivariate models, Abu-Foul found evidence of unidirectional causality from exports to output.
Abu-Qarn and Abu-Bader (2005) included imports as a third variable when testing the ELG hypothesis for nine MENA countries. They used cointegration and ECM approaches, and when aggregate exports were considered, no evidence of cointegration was found for Egypt (1963-1999), Jordan (1976-1998), Morocco (1963-1999), and Tunisia (1963-1998). However, they found evidence of cointegration in Iran, Israel, Sudan, and Turkey. The ECM showed marginal evidence of unidirectional causality from exports to economic growth only in the case of Iran; on the other hand, the GLE evidence was found in the cases of Israel, Sudan, and Turkey. They also applied SGC tests for Egypt, Jordan, Morocco, Tunisia, and Algeria and found no evidence of causality. When using manufactured exports, Abu-Qarn and Abu-Badr found unidirectional causality from manufactured exports to economic growth in Israel and Turkey, the two countries that have relatively high shares of manufactured exports. They found no causality between
2This study used nominal instead of real data, hence any causality or lack of between exports and economic growth cannot be inferred.
164
manufactured exports and economic growth in Algeria and Jordan, and bidirectional causality for Morocco and Tunisia.
Husein (2009) included terms of trade as a third variable and tested the validity of the ELG hypothesis for Jordan (1969-2005). He used Johansen-Juselius and Saikkonen and Lütkepohl (S&L) cointegration and error correction methods. He found that real output, real exports, and terms of trade are cointegrated. Granger causality tests based on the ECM suggested that the causal link between real exports and real GDP growth is bidirectional.
3. Data, methodology and empirical results Data and definitions of variables
In testing the ELG hypothesis for the selected MENA countries, this study employs the latest econometric techniques, i.e., Johansen and Juselius (1990) cointegration tests, and uses recent available annual data. We consider the following three variables: real gross domestic product (Y), real total exports (X), and net barter terms of trade (TOT).
The sample used includes the following countries for the specified periods:
Algeria, Egypt, Israel, Morocco, (1960-2005), Tunisia (1961-2005), Iran (1974-2005), Sudan and Turkey (1973-2005). The main source of our data is the World Bank’s World Development Indicators (WDI).3 Real GDP, real exports, and terms of trade were all transformed into log form. Log transformation can reduce the problem of heteroscedasticity since it compresses the scale in which the variables are measured, thereby reducing a tenfold difference between two values to a twofold difference (Gujarati, 1995). In addition, first differences of log transformation can be interpreted as growth rates.
Unit root testing
The first step in this study is to investigate the integration properties of the data.
If the variables are integrated (non-stationary), then the issue is to what degree they are integrated.4 If all variables in the data are integrated of order one, I(1), we proceed to test whether they are cointegrated using Johansen and Juselius methodology (details of this approach can be found in Johansen (1988) and Johansen and Juselius (1990)).
The Phillips-Perron (1988) unit root test is used in this study, in conjunction with the Augmented Dickey-Fuller (1979) test to address the issue of integration of the time- series data.5 The Phillips and Perron (PP) is considered here because it accounts for possible correlation in the first differences of the time-series using a nonparametric correction and allows for the presence of a non-zero mean and a deterministic time trend.
In addition, Perron (1989) has suggested that Augmented Dickey-Fuller (ADF) tests may falsely conclude the presence of a unit root in a time-series subject to a structural break.
In general, the PP test is based on the estimate of the following regression:
0 1 1
t t t
Y a a t Y
(1)
3 See Appendix for Data sources and further definitions.
4If a time series requires differencing d times before it becomes stationary, it is integrated of order d, i.e., I(d).
5 The joint use of both tests attempts to overcome the common criticism that unit root tests have limited power in finite samples to reject the null hypothesis of nonstationarity.
where a0 is a drift and t represents a time trend. Once the regression is estimated, the null hypothesis of unit root may be tested, i.e., H0: 0 against HA: 0. While the ADF test corrects for higher order serial correlation by adding lagged differenced terms of Yt on the right hand side of equation 1, the PP test makes a correction to the t-statistic of the α coefficient by using heteroscedasticity autocorrelation consistent estimates:
j
1
1 T
t t j
T t j
2 01
2 1
1
p
j j
w j
p
where
t’s are the estimated residuals from equation 1, T is number of observations, and P is a proper truncation lag that assures white noise residuals. Finally, the above calculated estimates are used to compute the PP t-statistic, tpp:
2
1 2 0
2
b b
pp
w Ts
t t
w w s
where s is the standard error of the test regression in equation 1, Sb and tb are the estimated standard error of α and the standard t-statistic for testing the null that γ1 = 0, respectively. The asymptotic distribution of the PP t-statistic is the same as the ADF t- statistic; hence, the same critical values of both tests are used to determine if the null hypothesis is rejected or not. If the calculated PP or ADF test statistic for Yt is less than its critical value, then the series Yt is said to be stationary or integrated of order zero, I(0).
If that is not the case, then the tests are performed on the Yt’s first differences. If the first differences are found to be stationary, then the order of integration is said to be one, I(1).
The results of the PP and ADF unit root tests are presented in Table 1. If the calculated PP or ADF test statistic for the natural logarithm of Y, X, or TOT (denoted as LY, LX, LTOT) is less than its critical value, then the series is said to be stationary or integrated of order zero, I(0). If that is not the case, then the tests are performed on the series’ first differences (∆LY, ∆LX, and ∆TOT). If the first differences are found to be stationary, then the order of integration is said to be one, I(1).
Table 1 provides the unit root test results of the Phillips-Perron (PP) and the Augmented Dickey-Fuller (ADF) tests for LY, LX, and LTOT for all countries. It can be seen that in none of the level variables are the computed PP and ADF statistics less than its 95% critical value. Therefore, all variables are nonstationary at the 95% level of significance. In first difference, the calculated ADF and PP for the same variables are less than the 95% critical values. We conclude that all first differenced variables are stationary or I(0). Thus all level variables are integrated of the same order or I(1).
166
Table 1. Results of the ADF and PP unit root tests
ADF Test Country Variable Level 1st- Difference
Phillips-Perron Test Level 1st-Difference Algeria LYb -3.66** -7.92* -1.57 -8.59*
LXa -0.37 -4.28* -1.16 -10.02* LTOTa -1.71 -5.87* -1.81 -5.87*
______________________________________________________________________________
Egypt LYb -2.17* -3.88** -1.33 -3.67**
LXb -2.24 -4.78* -2.41 -4.47* LTOTa -1.10 -4.42* -1.81 -4.29*
______________________________________________________________________________
Iran LYb -3.46 -6.03* -0.70 -3.08
LXb -1.87 -4.36* -1.44 -4.19**
LTOTa -0.80 -6.27* -1.03 -6.29*
________________________________________________________________________________________________________________________
Israel LYb -1.88 -4.61* -2.20 -4.50*
LXb -2.55 -6.46* -2.55 -6.48* LTOTa -1.90 -5.11* -2.32 -8.90*
_______________________________________________________________________________________________________________________
Morocco LYb -1.37 -10.4* -1.49 -11.01*
LXb -3.11 -6.16* -3.09 -8.00* LTOTc -0.38 -6.24* -0.87 -10.55*
_______________________________________________________________________________________________________________________
Sudan LYb -1.79 -4.40* -2.04 -4.63*
LXb -2.27 -6.36* -2.34 -11.01* LTOTa -2.09 -6.52* -2.03 -6.75*
_______________________________________________________________________________________________________________________
Tunisia LYb -1.72 -7.44* -1.66 -7.44*
LXb -1.49 -7.32* -1.46 -8.01* LTOTa -2.77 -5.09* -2.24 -4.92*
_______________________________________________________________________________________________________________________
Turkey LYb -2.70 -6.02* -2.82 -6.01*
LXb -3.37 -6.39* -3.30 -6.40* LTOTa -2.87 -6.18* -2.98** -7.15*
Notes: (1) a = intercept (1% and 5% critical values are -3.59 and -2.93, respectively); b = trend and intercept (1% and 5% critical values are -4.18 and -3.51, respectively); (2) * = significant at 1%
level; ** = significant at 5% level; (4) optimal lag lengths chosen by Akaike information criterion (AIC) for the ADF test and by Newey-West automatic truncation lag for the PP test.
Cointegration and error-correction models
Since LY, LX, and LTOT are integrated of the same order, this paper proceeds to test for cointegration properties and examines causality between LX and LY in a cointegration and error correction framework. In general, a set of variables are cointegrated if a linear combination of the integrated series is stationary, i.e., if Yt ~ I(d) and Xt ~ I(d), the following regression is run:
t t t
Y X (2)
If the residuals, µt, are I(0), then Xt and Yt are said to be cointegrated. We implement Johansen and Juselius approach to test for cointegration relationships, r, since
it allows for the presence of multiple cointegration relationships. Consider the following basic vector auto regressive (VAR) model of order p:
1 1 ...
t t p t p t
Y A Y A Y (3)
where Yt is a vector of nonstationary variables, in this study a 3 ×1 vector containing the natural logarithm of real GDP, real exports, and the terms of trade, Ai is a 3 ×3 matrix of parameters, and µt is i.i.d. (independently and identically distributed) k dimensional Gaussian error term with µt ~ (0, Σµ). Since Yt is I(1), the VAR can be written in the first- differenced error-correction (EC) form by subtracting Yt-1 from both sides of equation 3 and rearranging terms such as:
1 1 ... 1 1 1
t t p t p t t
Y Y Y Y
(4)
where Гi = - (Ai+1 +…+ Ap) and П =- (Ik – A1 -…- Ap) for i = 1, …, p -1
Because ∆Yt does not contain stochastic trends by the assumption that all variables are at most I(1), it follows that ПYt-1 is the only term that includes I(1) variables.
Hence, ПYt-1 must be I(0) and as a result it contains the cointegration relations. The Γj’s ( j =1, …, p – 1) are often referred to as short-run parameters, and ΠYt-1 is referred to as containing the long-run parameters. The model in equation 4 will be abbreviated as VECM(p – 1). The focus of the Johansen and Juselius technique is on the parameter matrix П, which contains information about the long-run relationship among the variables in the data vector.
The rank r of this matrix П, rk(П), determines the number of cointegrating vectors in the VAR system. If matrix П has a full rank, i.e., rk(П)= k, the vector Yt is stationary. Instead, if matrix П has a rank that equals zero, rk(П)=0, then П is a null matrix and equation 4 corresponds to a traditional VAR model in first difference. Finally, if matrix П has a reduced rank (0 < r < 3 in this case), then there exists k × r matrices α and β, each with rank r such that П = α β` and β`Yt are I(0) even though Yt itself is I(1). r is the number of cointegrating relations and each column of β is the cointegrating vector.
The matrix α contains the weights attached to the cointegrating relations in the individual equations of the model and sometimes is referred to as the loading matrix. In this case, equation 4 is a vector error-correction model of order p-1, VECM (p-1).
Several extensions of the VECM in equation 4 are usually necessary to represent the main characteristics of a data set of interest. Including deterministic terms such as a constant, a linear trend term, and seasonal and other dummy variables, may be required for a proper representation of the process. A general VECM that includes all such terms is:
'
1 1 ... 1 1 1
t t p t p t t t
Y Y Y Y D
(5)
Where Dt contains all regressors associated with deterministic terms and Φ is a matrix of parameters. In the presence of cointegration, Granger Causality concerns the influence of the Γ and α parameters on the levels of the endogenous variables, i.e., LY, LX, and LTOT.
Granger causality can be investigated in the framework of the VECM in equation 5. For example, the VECM in equation 5 with P level lags and one cointegration relation, has the following short-run parameter matrices, Γj, and long-run parameter matrix, αβ`:
,11 ,12 ,13 11 11 1, 1
,21 ,22 ,23 21 11 12 13 21 1, 1
31 31 1, 1
,31 ,32 ,33
j j j t
j j j j t
j j j t
ec
and ec
ec
(6)
168
where sub-index j is the lag length of the VECM (j = 1, …, p - 1), and ec1,t-1 is the error correction term lagged one period . Granger causality from LX to LY (ELG hypothesis) in the presence of cointegration is determined by testing the following null hypothesis:
1,12 2,12 ... j,12 11ec1,t1 0 LX does not Ganger cause LY LX( LY)
By rejecting the null hypothesis, one can conclude that LX Granger-causes LY. The above test differs from SGC tests since it includes an error correction term, ec1,t-1, that accounts for cointegration among the variables. It is worth mentioning that, if there is a cointegration relation among the variables, there must also be Granger causality in at least one direction, i.e., one of the coefficients of the error correction terms must be significantly different from zero. Similarly, Granger causality from LY to LX (GLE hypothesis) can be determined by testing the following null hypothesis:
1,21 2,21 ... j,21 21ec1,t1 0 LYdoes not Granger cause LX LY( LX)
The Johansen and Johansen and Juselius (1990) cointegration technique allows estimation of the cointegrating relationships among the I(1) variables (LY, LX, LTOT) using a maximum likelihood (ML) procedure that tests for the rank of П and estimates the parameters of β. The cointegrating rank, r, can be tested using a likelihood ratio (LR) tests that is known as the trace test. The LR test, (λTrace), for the null hypothesis that there are at most r cointegrating vectors is computed as follows:
1
( ) ln(1 )
k Trace i
j r
r T
(9)where r1,...,k are the k – r estimated values of the characteristic roots (eigenvalues) obtained from the estimated П matrix.
Lag order and cointegration rank
The above λTrace test is sensitive to the choice of the appropriate lag length, p, of the VAR model. An appropriate lag length ensures that the error terms in the VECM are Gaussian. In this study, we performed tests for different lag orders and checked the robustness of the results in conjunction to following an information criterion such as Akaike information criterion (AIC) as suggested by Enders (2004).
Since the full sample contains 46 annual observations in the cases of Algeria, Egypt, Israel, Morocco, Tunisia, and 33 annual observations in the cases of Iran, Sudan, Turkey, and inference on lag order determination is based on classical asymptotic theory, a main criterion that we utilize is parsimony. In addition to using model selection criteria such as Akaike information criterion (AIC)6, we also performed a number of specification tests, covering serial correlation and the normality of residuals for models based on different p lags and one cointegration relation. For Algeria, Egypt, Iran, Israel Sudan and Turkey, the choice of the lag order for each country is based on AIC. In the cases of Morocco and Tunisia, AIC suggested one level lag, but LM tests performed
6 The general approach to using the criterion is to fit VAR(m) models with orders m = 0, …, pmax
(pmax = 4 in this study) and to choose an estimator of the order p that minimizes the “preferred”
criterion.
rejected 1 lag in favor of 3 lags for these two countries. Therefore, we apply cointegration tests using three level lags for Morocco and Tunisia.7
Table 2. Johansen tests for cointegrating rank, r
Null Critical Values
Country Hypotheses p λtrace 5% p-value
Algeria r = 0 4 30.97 29.68 0.02
r ≤ 1 3.87 15.41 0.90
r ≤ 2 0.44 3.76 0.50
Egypt r = 0 3 36.96 34.91 0.03
r ≤ 1 18.77 19.96 0.08
r ≤ 2 7.75 9.24 0.09
Iran r = 0 4 51.33 42.44 0.00
r ≤ 1 20.38 25.32 0.21
r ≤ 2 8.32 12.25 0.23
Israel r = 0 3 40.64 35.07 0.01
r ≤ 1 15.03 20.26 0.22
r ≤ 2 5.57 9.16 0.21
Morocco r = 0 3 43.13 34.91 0.00
r ≤ 1 17.98 19.96 0.11
r ≤ 2 6.23 9.24 0.17
Sudan r = 0 3 44.89 34.91 0.00
r ≤ 1 14.44 19.96 0.26
r ≤ 2 3.22 9.24 0.54
Tunisia r = 0 3 49.86 34.91 0.00
r ≤ 1 18.26 19.96 0.09
r ≤ 2 6.87 9.24 0.13
Turkey r = 0 1 46.07 42.91 0.02
r ≤ 1 18.25 25.87 0.33
r ≤ 2 7.26 12.52 0.33
Notes: r = cointegration rank. Critical values for the cointegration tests taken from Osterwald-Lenum(1992)
Using Johansen’s method, the number of cointegrating relationships, r, among LY, LX, and LTOT is determined by the λTrace test statistic. From Table 2, the hypothesis of no cointegration can be rejected at 95% level of significance for all countries in favor of one cointegration relationship. Hence, a long run relationship exists between GDP, exports, and terms of trade for the eight MENA countries. Now that a unique cointegrating relationship between LY, LX, and LTOT has been established, point estimates, standard errors, and t-statistics of the long-run relationships are presented in Table 3. It can be seen that LX has a positive and significant long-run impact on LY in all countries.
Testing For Granger Causality
The results of Granger causality between LY and LX, and whether the export-led growth (ELG) or growth driven exports (GLE), or both hold true for the selected MENA countries are reported in Table 4.
7 LM and Specification test results are not reported but can be supplied upon request.
170
Table 3. ML estimates of the cointegrating vector (normalized on LY)
Country Variable Cointegrating Standard Error [t-statistics] Vector (β)
Algeria LY - - -
LX 1.01 0.05 [-20.2]
LTOT 0.45 0.04 [-11.7]
Egypt LY - - -
LX 1.47 0.22 [-6.58]
LTOT 1.61 0.60 [-2.65]
Iran LY - - -
LX 0.14 0.05 [-2.55]
LTOT 0.22 0.04 [-3.55]
Israel LY - - -
LX 1.03 0.06 [-15.0]
LTOT -1.74 0.76 [2.28]
Morocco LY - - -
LX 0.83 0..20 [-4.23]
LTOT 4.00 2.32 [-1.72]
Sudan LY - - -
LX 0.51 0.04 [-10.4]
LTOT -0.21 0.47 [4.580]
Tunisia LY - - -
LX 0.66 0.06 [-10.2]
LTOT -0.17 0.16 [-1.05]
Turkey LY - - -
LX 0.37 0.03 [-10.7]
LTOT 0.16 0.30 [-0.52]
Table 4: Granger non-causality tests (between LY and LX) for the cointegrated VAR models with one cointegration relation
Asymptotic Bootstrapped Country Null Hypothesis Wald Test p-value p-value α (t-stat.) Algeria LX → LY 19.83 0.00 0.013 -0.25 (-3.13)
LY → LX 1.79 0.77 0.858 -0.12 (-0.88) Egypt LX → LY 20.46 0.00 0.033 -0.04 (-4.03) LY → LX 23.93 0.00 0.011 -0.07 (-2.34) Iran LX → LY 20.60 0.00 0.063 -0.41 (-2.82) LY → LX 14.94 0.00 0.117 -1.60 (-1.81) Israel LX → LY 22.06 0.00 0.023 0.09 (4.06) LY → LX 39.04 0.00 0.000 0.25 (4.49) Morocco LX → LY 30.68 0.00 0.003 0.04 (4.88) LY → LX 26.46 0.00 0.004 0.04 (3.12) Sudan LX → LY 3.59 0.30 0.447 -0.16 (-1.76)
LY → LX 58.8 0.00 0.000 1.00 (4.35) Tunisia LX → LY 47.92 0.00 0.000 -0.21 (-6.69)
LY → LX 19.55 0.00 0.012 -0.25 (-3.34) Turkey LX → LY 39.96 0.00 0.000 -0.32 (-4.45)
LY → LX 7.37 0.00 0.044 -0.84 (-3.41)
Note: α is the error correction term coefficient lagged one period. We report both asymptotic and bootstrapped p-values of the Wald test for Granger causality between LX and LY are reported.
The estimates of the VECM in equation 5 are applied to detect the direction of Granger causality through the influence of Γ and α parameters on the levels of endogenous variables.8 For five countries, namely, Egypt, Israel, Morocco, Tunisia, and Turkey, Granger causality between real output and real exports is bidirectional. An indication that export growth and economic growth are simultaneously reinforcing in these countries, i.e., it is possible that they experienced periods when economic growth was driven by exports as advocated by ELG proponents and also periods when exports were growth driven (GLE).
In Algeria and Iran, we find evidence of unidirectional causality running from exports to output growth, while in the case of Sudan the GLE hypothesis seems to hold.
As shown, for seven out of the eight countries examined in this study, the empirical evidence strongly supports the ELG hypothesis.
4. Stability analysis
To test whether the estimated long-run parameters are stable, we apply Hansen- Johansen (1999) formal tests. Long-run stability means that those parameters of the cointegration relationship are invariant overtime. As has been emphasized by Brüggemann et al. (2003), it is of some importance to formally investigate the stability of the cointegrating vectors further, once a long-run relationship has been identified.
For cointegrated VAR models, Hansen and Johansen (1999) suggested applying a fluctuation test to the nonzero eigenvalues of the reduced rank matrix. The fluctuation test rejects stability when the recursively estimated eigenvalues fluctuate excessively. The test may be applied to the eigenvalues themselves, λi, giving rise to the test statistic Sup λi, or to the transformation ξi = log (λi/(1- λi)), giving rise to the test statistic Sup ξi.
To examine the constancy of the cointegration space, we consider two types of Nyblom tests. The first (supermum, SupQs) test statistic, is based on the maximum value of a weighted LM-type test statistic over the experimentation period, and the second (mean, MeanQs) test, on the average of this statistic.9
Constancy of the non-zero eigenvalues
Table 5 report the Hansen-Johansen fluctuation tests, Sup λi and ξi = log (λi/(1- λi)). As can be seen, the null hypothesis of constant eigenvalue(s) cannot be rejected for the eight MENA countries using both asymptotic and bootstrapped p-values (showing exact level of significance) over the test period.10
8 Nb, the number of bootstrap samples used to estimate all bootstraped critical values in this study,
= 1,000.
9 We implement the score function suggested by Brüggemann et al. (2003) for the two Nyblom type tests instead of the first order approximation used by Hansen and Johansen (H-J).
Brüggemann et al. (2003) and Warne (2005) suggested that SupQs and MeanQs tests are superior to H-J’s SupQ and MeanQ since the latter suffer from numerical problems in simulation exercises, leading to small sample distributions that are far away from the limit distributions. All the stability tests and the asymptotic and bootstrap critical values are computed using the program structural VAR.
10 It is worth mentioning that formal tests do not require trimming of the sample; however, we use 10% of the sample as a base period and examine constancy over the remainder.
172
Table 5. Hansen-Johansen fluctuation tests of the stability of the non-zero eigenvalue for the cointegrated VAR with p lags and one cointegration relation
Country Test asym. boots. Test asym. boots
Sup λi p-val p-val Sup ξi p-val p-val P Period Algeria 0.98 0.28 0.17 0.99 0.27 0.27 4 1977-2005 Egypt 1.25 0.08 0.36 1.16 0.13 0.50 3 1973-2005 Iran 0.11 0.99 0.56 0.12 1.00 0.56 4 1992-2005 Israel 0.76 0.60 0.34 0.94 0.33 0.31 3 1973-2005 Morocco 1.24 0.08 0.14 1.38 0.04 0.19 3 1973-2005 Sudan 0.47 0.97 0.58 0.57 0.89 0.64 3 1986-2005 Tunisia 0.72 0.66 0.41 0.92 0.35 0.43 3 1974-2005 Turkey 0.21 1.00 0.57 2.75 0.00 0.43 1 1978-2005 Notes: For H-J fluctuation tests and Nyblom type tests, both asymptotic p-values and bootstrapped p-values are reported. All the stability tests and their critical values are computed using the program structural VAR by Andres Warne.
Constancy of the cointegration space
Table 6 reports the two Nyblom type tests, SupQs and MeanQs, for testing the stability of the cointegrating vector, β, for the eight MENA countries. As can be seen, the hypothesis of inconstant β is strongly rejected for all countries using both the asymptotic and bootstrapped p-values. To sum up, based on the two Nyblom type tests, we conclude that the cointegration space is constant for the eight MENA countries.
Table 6. Nyblom supermum and mean test for the constancy of β in the cointegrated VAR model with one cointegration relation
asym. boot. asymp boot.
Country SupQs p-value p-val MeanQs p-val. p-val
Algeria 0.97 0.51 0.25 0.45 0.20 0.11
Egypt 1.86 0.14 0.06 0.17 0.74 0.84
Iran 0.40 0.99 0.85 0.17 0.85 0.77
Israel 1.15 0.49 0.35 0.49 0.13 0.19
Morocco 1.20 0.46 0.28 0.38 0.24 0.32
Sudan 1.11 0.53 0.31 0.56 0.09 0.14
Tunisia 1.25 0.42 0.27 0.68 0.05 0.08
Turkey 0.72 0.84 0.62 0.20 0.60 0.60
5. Conclusion
Using the latest econometric advances in time series analysis, this paper investigates the ELG hypothesis for selected MENA countries using Johansen multivariate approach to cointegration. Explicitly, whether promoting exports, as has been suggested by many international institutions and prominent economists is a key strategy to enhancing economic growth. Our empirical results suggest that real GDP, real exports, and terms of trade are cointegrated in the MENA countries we analyze, implying a long-run relationship between real GDP, real exports, and terms of trade. The empirical evidence from Granger causality tests suggests that the causal link between real exports and real GDP growth is bidirectional in Egypt, Israel, Morocco, Tunisia, and Turkey.
This is an indication that export growth and economic growth are simultaneously reinforcing in the above six countries, i.e., it is possible that they experienced periods when economic growth was driven by their exports as advocated by ELG proponents and
also periods when their exports were growth driven (GLE). In addition, the empirical evidence suggests a unidirectional causality from exports to output growth (ELG) in Algeria and Iran, and a unidirectional causality from output growth to exports (GLE) in Sudan.
From the study’s findings, public policies that put forward emphasis on export growth and on domestic production are recommended for the five countries that experienced bidirectional causality between exports and economic growth, and to further enhance their economic growth, export promotion policies are recommended for Algeria and Iran. Moreover, there is strong support for the GLE hypothesis in Sudan; as such, emphasis should be on domestic production if the goal is to promote economic growth.
Finally, unlike most if not all ELG previous studies, this study formally tested the stability of the cointegrating vector and the empirical results reject instability of the cointegrating vector in all countries analyzed.
References
Abu-Bader, S. and Abu-Qarn, A. (2005), “The Validity of the ELG Hypothesis in the MENA Region: Cointegration and Error Correction Model Analysis,” Applied Economics, 36, 1685- 95.
Abual-Foul, Bassam, (2004), “Testing the Export-Led Growth Hypothesis: Evidence from Jordan”, Applied Economics Letters, 11, 393-96.
Brüggemann, A., Donati, P., and Warne, A. (2003) Is the Demand for Euro Area M3 Stable?
European Central Bank Working Paper No. 255.
Davidson, R. and MacKinnon, J.G., (1996), “The Size Distortion of Bootstrap Tests,”
Working paper, Department of Economics, University of Queens’s, Ontario.
Dhawan, U. and Biwal, B. (1999), “Re-examining the Export-Led Growth Hypothesis: A Multivariate Cointegration Analysis for India,” Applied Economics, 31, 525-30.
Dickey, D.A. and Fuller, W.A. (1979), “Distribution of the Estimators of Autoregressive Time Series with a Unit Root,” Journal of the American Statistical Association, 74, 427-31.
Dodaro, S. (1993), “Exports and Growth: A Reconsideration of Causality,” The Journal of Developing Areas, 27, 227-44.
Enders, W. (2004), Applied Econometric Time Series, John Wiley & Sons.
Engle, R. F., and Granger, C. W. J. (1987), “Co-Integration and Error Correction:
Representation, Estimation and Testing,” Econometrica, 55, 252-76.
Granger, C. W. J. (1988), “Some Recent Developments in A Concept of Causality,” Journal of Econometrics, 39, 199-211.
Gujarati, D. N. (1995), Basic Econometrics, McGraw-Hill.
Hansen, H. and Johansen, S. (1999), “Some Tests for Parameter Constancy in Cointegrated VAR Models,” Econometrics Journal, 2, 306-333.
Husein, J. (2008). “Traditional Export Demand Relation: A Co-integration and Parameter Constancy Analysis of Jordan”, International Journal of Applied Econometrics and Quantitative Studies, Vol. 5-2
Husein, J. (2009), “Export-Led Growth Hypothesis: A Multivariate Cointegration and Causality Evidence for Jordan,” The Journal of Developing Areas, 42 (2).
Johansen, S. (1988), “Statistical Analysis of Cointegrating Vectors,” Journal of Economic Dynamics and Control, 12, 231-54.
Johansen, S. (1992), “Determination of Cointegration Rank in the Presence of a Linear Trend,” Oxford Bulletin of Economics and Statistics, 54, 383-97.
174
Johansen, S. and Juselius, K. (1990), “Maximum Likelihood Estimation and Inference on Cointegration: With an Application to Demand for Money,” Oxford Bulletin of Economics and Statistics, 52, 169-210.
Jung, S. W. and Marshall, P. J. (1985), “Macroeconomic Determinants of Economic Growth:
Cross-Country Evidence,” Journal of Monetary Economics, 16, 141-63.
Kugler, P. and Dridi, J. (1993), “Growth and Exports in LDCs: A Multivariate time series study,” International Review of Economics and Business, 40, 759-67.
Perron, P. (1988), “Trend and Random Walks in Macroeconomic Time Series,” Journal of Economic Dynamics and Control, 12, 297-32.
Philips, P. C. B., and Perron, P. (1988), “Testing for A Unit Root in Time Series Regression,”
Biometrica, 75, 335-46.
Pomponio, X. Z. (1996), “A Causality Analysis of Growth and Export Performance,” Atlantic Economic Journal, 24, 168-76.
Riezman, R. G., Summers, P. M., and Whiteman, C. H., (1996), “The Engine of Growth or its Handmaiden? A Time Series Assessment of Export Led Growth,” Empirical Economics, 21, 77-113.
Trenkler, C. (2004), “Determining p-values for systems cointegration tests with a prior adjustment for deterministic terms,” mimeo, Humboldt-Universität zu Berlin.
Xu, Z. (1996), “On the causality Between Export Growth and GDP Growth: An Empirical Investigation,” Review OF International Economics, 4, 172-84.
APPENDIX. Data Definition and Sources
Variables:
LY Natural logarithm of Gross Domestic product (GDP) in constant units (1980 =100 in cases of Algeria, Morocco, 1982 = 100 for Iran, 1990 = 100 for Sudan, Turkey and Tunisia, 1992 = 100 for Egypt, 2000 = 100 for Israel).
LX Natural logarithm of exports of goods and services in constant units (1980 =100 in cases of Algeria, Morocco, 1982 =100 for Iran, 1990 =100 for Sudan, Turkey and Tunisia, 1992 =100 for Egypt, 2000 =100 for Israel)
LTOT Natural logarithm of terms of trade:
( )
1 0 0
( )
u n it v a lu e in d e x o f e x p o rts e x p o r t p r ic e T O T
u n it v a lu e in d e x o f im p o rts i m p o rt p ric e
.
For export and import price indexes: 1980 =100 in cases of Algeria, Morocco, 1982 =100 for Iran, 1990 =100 for Sudan, Turkey and Tunisia, 1992 =100 for Egypt, 2000 =100 for Israel.
Annual Data:
1960-2005: Algeria, Egypt, Israel, Morocco(a) (b) 1961-2005: Tunisia (a) (b)
1974-2005: Iran (a)(c)
1973-2005: Sudan and Turkey (c)
Note: Real GDP and Real exports data for all countries are in thousands of local currency units (LCU) except for Sudan, Turkey, and Israel (data is in millions of U.S. Dollars) Sources: a World Bank, World Development Indicators (WDI online). b International Financial Statistics of the International Monetary Fund (IMF, various issues). c United Nations National Accounts Statistics.
Journal published by the EAAEDS: http://www.usc.es/economet/eaa.htm