Ardl Pak
Ardl Pak
Ardl Pak
+ + + =
2 6 1 5
(5)
The second step is related to search the minimum t-ratio by a test to test the hypothesis that 1 = :
= =
=
+ A + + + =
k
i
k
i
t t i t t i
k
i
t i t
x c x TB TB x
1 1
1 1 1 2 2
1
1 1 1
(6)
We have included the dummy variable
it
DTB in the estimated equation so as to make sure that
) , ( min
2 1
t
IO
t congregates i.e. converges to distribution:
2
1
2
1
1 2 1
2 1
)] ( [
inf ) , ( min
K
H
t
t
IO
. =
(7)
21
Once the order of integration is determined, the second stage involves testing for the existence of
cointegration between the series in a multivariate framework in the presence of structural breaks.
For this purpose, we adopt the autoregressive distributed lag (ARDL) bounds test (Pesaran et al.
2001) to test the existence of long-run relationship between income inequality, international
remittances and economic growth. The Bounds test has several advantages over the widely used
cointegration test (e.g. Johansen cointegration test). First, the ARDL bounds test is more robust
for small sample studies and availability of critical values for sample size 30 (Narayan, 2005)
contributes to the popularity of the method. Second, the method does not require the order of
integration to be similar like other cointegration approaches such as Johansen-Juselius or Engle-
Granger approach. Third, Pesaran et al. (2001) argued that, based on Monte Carlo results, this
procedure is robust even with the presence of endogenous regressors in the model, irrespective of
whether the regressors are I(1) or I(0). The bounds test involves the testing of an unrestricted
error-correction model (UECM) using
t
Y ,
t
I and
t
R which are given by:
t t F t F t F
i t
n
i
iF
n
i
i t iF
n
i
i t iF F t
InR e InI e InY e
InR d InI c InY b DUM t a InY
+ + + +
A + A + A + + + = A
= =
=
1 3 1 2 1 1
0 0 1
0
(8)
t t G t G t G
i t
n
i
iG
n
i
i t iG
n
i
i t iG G t
InY e InR e InI e
InY d InR c InI b DUM t a InI
+ + + +
A + A + A + + + + = A
= =
=
1 3 1 2 1 1
0 0 1
0
(9)
t t X t X t X
i t
n
i
iX
n
i
i t iX
n
i
i t iX X t
InGI e InY e InR e
InI d InY c InR b DUM t a InR
+ + + +
A + A + A + + + = A
= =
=
1 3 1 2 1 1
1 0 1
0
(10)
22
where Ais the first difference operator. In the model, b, c and d captures the short-run dynamics
while the es captures the long-run effects and DUM is dummy variable to capture the structural
break stemming in the series
10
. In order to test the absence of a long run relationship in equation
(8), we restrict the coefficient (using F-test or Wald test) of e
1G
, e
2G
and e
3G
to be zero (H
o
: e
1F
=
e
2F
= e
3F
=
0) against the alternative hypothesis that at least one is not equal to zero. This is
denoted as F
Y
(Y|I, R).Similarly, for equation (3) and (4) we test the null hypothesis for no
cointegration as (H
o
: e
1G
= e
2G
= e
3G
=0) and (H
o
:e
1X
=e
2X
=e
3X
=0), respectively. This is denoted as
F
I
(I|Y, R) and F
R
(R|Y, I). The asymptotic distributions of the test statistics are non-standard
regardless of whether the variables are I(0) or I(1). For this purpose, we used Narayans (2005)
computed sets of asymptotic critical values. The first set of asymptotic critical values assume
variables to be I(0) and the other as I(1) which is known as lower bounds (LCB) and upper
bounds critical values (UCB), respectively
11
. If the computed F-statistic is more than UCB, we
can than reject the null hypothesis of no cointegration and vice versa. The results are inconclusive
if calculated F-statistic is between upper and lower critical bounds. Since the selection of lags is
important, we relied on the Schwarz Bayesian Criterion (SBC) to select the optimal lag length.
Additionally, to ensure that the model satisfy all assumption of regression, a series of diagnostic
tests namely Lagrange multiplier (LM) test for serial autocorrelation in the presence of lagged
variables, Ramsey/RESET test for functional form, Bera-Jarque for residuals normality and
Heteroscedasticity test based on the regression of squared residuals on squared fitted values are
performed. The CUSUM and CUSUMSQ test is applied to examine the model stability.
23
IV.I Sensitivity Analysis
Theoretical findings suggest that income inequality affects economic growth and, to an equal
extent, economic growth may affect inequality. Hence, both income inequality and economic
growth are endogenous and placing either variable on the right hand side violates the exogeneity
assumptions. We tackle this issue by carefully specifying an ARDL model with an appropriate lag
structure. Pesaran et al. (2000) proved that it is sufficient to simultaneously correct for the
residual serial correlation and the endogenous regressors problem using appropriate orders of the
ARDL model. The single equation approach of the ARDL also allows us to check the robustness
of the estimates. When we use the Vector Autoregressive (VAR) model on a system of variables,
we were also able to mitigate the problem because in VAR no such conditional factorisation is
made a priori. Instead, variables can be tested for exogeneity later, and restricted to be exogenous
then. These considerations motivate our choice of the ARDL and VAR model for studying the
interdependencies between income inequality, international remittances and economic growth.
We conduct several sensitivity analyses to tackle the problem of endogeneity. First, we set up
three simultaneous equations by treating each variable as endogenous variable. This allows us to
identify whether desired changes in their values take place. In doing so, we also vary the lag
length of our regression. We also rerun the equations by omitting the income inequality and
economic growth, separately, to check the robustness of the regression. This is equivalent to
performing reduced form of the equation by expressing each endogenous variable as a function of
only the predetermined variables. In all cases, we can only detect significant relationship when
economic growth serves as the dependent variable. In other words, in long run, international
remittance and income inequality tend to influence economic growth.
24
Second, the Granger-causality testing methodology seems to be one of an ideal tool to examine
the influence of each variable empirically. For the context of this paper this means that if after
lagged economic growth and contemporaneous income inequality are controlled for Granger-
causality running from lagged inequality to GDP growth is found to be significantly positive, then
this is evidence in favour of income inequality acting as an endogenous variable. If, however,
negative Granger-causality in the medium run and no Granger causality in the long run are found,
then this speaks in favour of income inequality being exogenous. Since our Granger causality is
performed in a multivariate setting, spurious causality can also arise, when the third variable is
introduced in the model. For this purpose, we conclude that no causality found in multivariate
setting only when there is also no causality in a bivariate setting. This again allows us to check the
robustness of our results.
IV.II Innovative Accounting Approach
Although cointegration test is able to identify the long-run forcing variables of economic growth,
the direction of causality will be less clear at this stage. In other words, cointegration does not
provide indication about the causality of series interdependencies, which however is an essential
enquiry in our study. The evidence of cointegration is only a necessary but not sufficient
condition for rejecting Granger non-causality. Therefore, the presence of cointegrating among the
variables leads us to perform the Granger causality test. If the series are cointegrated, the causality
testing should be based on a Vector Error-Correction Model (ECM) rather than on an unrestricted
VAR model (Johansen, 1988; Johansen and Juselius, 1990). Nonetheless, the Granger causality
tests do not determine the relative strength of causality effects beyond the selected time span
25
(Shahbaz et al. 2012). Due to the limitation of the VECM Granger causality test, we include
innovative accounting approach (IAA) to investigate the dynamic causal relationships among
income inequality, international remittances and economic growth. The uniqueness of the IAA is
that it avoids the problem of endogeneity and integration of the series. This approach has an
advantage compared to the VECM Granger causality test because the latter only shows a causal
relationship between the variables within the sample period while the former illustrates the extent
of causal relationship ahead the selected sample period. The IAA includes forecast error variance
decomposition and impulse response function. This procedure decomposes forecast error variance
for each series following a standard deviation shock to a specific variable and enables us to test
which series is strongly impacted and vice versa.
For instance, if a shock in income inequality has significant effects of economic growth but a
shock occurring in economic growth only affect very minimum the variations of income
inequality. Then, this is inferred as a unidirectional causality runs from income inequality to
economic growth. If economic growth explains more of the forecast error variance of income
inequality; then we deduce that economic growth causes income inequality. The bidirectional
causality exists when shocks in income inequality and economic growth have a strong impact on
the variability of income inequality and economic growth respectively. If shocks occur in both
series do not have any impact on the economic growth and income inequality then there is no
causality between the variables. Impulse response function helps us to trace out the time path of
the impacts of shocks of variables in the VAR. One can determine how much income inequality
responses due to its own shock and shock in economic growth. We support the hypothesis that
economic growth causes income inequality of the impulse response function indicates significant
26
response of income inequality to shocks in economic growth than other variables. A strong and
significant reaction of income inequality to shocks in economic growth implies that income
inequality causes economic growth. This study incorporates income inequality, international
remittances and economic growth to examine the relationship between economic growth and its
determinants in the VAR model. A VAR system takes the following form (Shan, 2005):
t i t
k
i
i t
V V + =
=
1
(11)
where, ) , , (
t t t t
R I Y V = and ) , , (
R I Y t
=
i
are the estimated coefficients and is a vector of error terms.
V. Empirical Results
Although bounds test does not require the knowledge of order of integration, yet, the test is
crucial to avoid having series with higher order (e.g. I(2)). Table-2 reports the unit root properties
of the data series with and without trend term. It is evident that all unit root tests yield similar
results. The series are non-stationary in their levels but become stationary after taking the first
differences. Although, it can be concluded that all series are I(1) at the 1% and 5% significant
level but at 10% level some of the series are found to be I(0).
Table-2: Unit Root Analysis
Variables ADF PP KPSS
Intercept Intercept Intercept Intercept Intercept Intercept
27
and Trend and Trend and Trend
t
Y ln
-2.226 -2.118 -2.110 -1.508 0.728** 0.179**
t
Y ln A
-3.888* -4.463* -4.008* -4.492* 0.352 0.131
t
I ln
-1.751 -3.283*** -1.501 -3.222*** 0.577** 0.181**
t
I ln A
-7.992* -7.991* 10.331* 15.181* 0.259 0.144
t
R ln
-2.629*** -2.507 -2.757*** -2.718 0.244** 0.169
t
R ln A
-4.559* -4.384* -4.559* -4.378* 0.198 0.102
Note: *, ** and *** denotes significant at 1%, 5% and 10%, respectively. SIC is used to
select the lag length for ADF. The bandwidth for PP and KPSS test is selected using
Newey-West method using Barlett-Kernel. Null hypothesis for ADF and PP is that series
are non-stationary while for KPSS series are stationary, respectively.
The results of ADF, PP and KPSS may be biased and unreliable because these unit root tests do
not seem to have information about structural break arising in the series. This is issue is solved by
applying the Clemente et al. (1998) accommodating single and two structural breaks. The unit
root test by Clemente et al. (1998) uses innovative outlier (IO) and additive outlier (AO) models.
The IO model captures the steady changes in mean of the variables. The sudden changes in the
mean of the series are plugged out by AO model. The AO model is more reliable and suitable
than IO model because it provides information about sudden structural changes. Our results are
reported in Table-3 show that economic growth (Y
t
), foreign remittances (R
t
) and income
inequality (I
t
) have unit root problem in the presence of structural breaks at level
12
. This implies
that series are found to be stationery at first difference i.e. the variables are I(1).
28
Table-3: Structural Break Unit Root Test
Model: Trend Break Model
Variable Innovative outliers Additive outliers
T
B1
T
B2
Test statistics K T
B1
T
B2
Test statistics K
t
Y ln
1998 --- -2.464 3 1998 ---- -4.581** 2
1998 2003 -5.286 4 1990 1998 -6.583* 5
t
I ln
1989 --- -3.324 3 2000 ---- -4.937** 3
1989 2000 -5.054 3 1982 2000 -6.659* 2
t
R ln
1981 --- -0.691 2 1982 ---- -5.736* 3
1978 2002 -3.885 4 1991 2003 -6.510* 3
Note: T
B1
and T
B2
are the dates of the structural breaks; k is the lag length. * and ** show significant at 1%
and 5% levels respectively. T
B1
and T
B2
indicate first and second structural breaks.
This justifies the use of ARDL cointegration. The results of the ARDL bounds test are reported in
Table-3. In equation (8) with
t
Y ln as dependent variable, we note that the computed F-statistic
(4.906) is above the upper bound critical value (4.428). It indicates that there is a strong evidence
to reject the null hypothesis of no cointegration at 5% significant level once we used income
inequality and international remittances as forcing variables. Likewise, for the other two
equations, (9) and (10), we fail to reject the null hypothesis of no cointegration.
It should be noted here that linking income inequality, international remittances and economic
growth may also lead to a biased estimated coefficient if per capita income is regarded as an
29
endogenous variable. Yamamura and Shin, (2009) and J ackman et al. (2009) suggested the
possible influence of income inequality and international remittances on economic growth
respectively. Nevertheless, when income inequality and international remittances served as the
dependent variables in equations 9-10, examining the long-run relationship then we failed to
establish any cointegration. In other words, in case of Pakistan, we fail to track any long-run
convergence in income inequality and international remittances equations (9-10). This also
confirms the problem of endogeneity is less obvious in our case. In addition, we also examine
whether there exists an endogenous relationship between international remittance, income
inequality and per capita income by applying the Durbin-Wu-Hausman test (the augmented
regression test) suggested by Davidson and MacKinnon, (1993). The results suggested that
endogeneity is not significant. The regression also passes a series of diagnostic tests and the
stability test-CUSUM and CUSUMQ test.
Table-4: Cointegration Analysis: Bounds Tests
Equation F-statistic Lag 95% critical value bounds
a
) , \ (
t t t Y
R I Y F
4.906** 3 LCB: 3.538 UCB: 4.428
) , \ (
t t t I
R Y I F
2.261 3
) , \ (
t t t R
I Y R F
2.105 3
Diagnostic Tests
Equation SERIAL
2
REMSAY
2
NORMAL
2
ARCH
2
CUSUM CUSUMsq
8 1.079
[0.390]
3.528
[0.448]
1.798
[0.406]
0.387
[0.683]
Stable Unstable
13
30
9 0.706
[0.525]
2.074
[0.187]
1.796
[0.407]
0.034
[0.966]
Stable Stable
10 2.233
[0.177]
1.928
[0.215]
0.112
[0.945]
0.012
[0.987]
Stable Stable
Note:
a
Critical values are obtained from Narayan, (2005). The lag selection is based
on SBC. [ ] and ** denotes the probability and the significant level at 0.05,
respectively. NORMAL
2
is for normality test, SERIAL
2
for LM serial correlation test,
ARCH
2
for autoregressive conditional heteroskedasticity and REMSAY
2
for Remsay
Reset test.
Having found cointegration when
t
Y ln serves as dependent variable, we proceed to estimate the
long and short-run coefficient. It should also be noted that long-run estimates are reliable due to
the fact that we fail to detect any significant endogeneity in the model. Table-4 reports the results
of the estimation. In long-run, income inequality is found to be significant at 1% with a positive
impact. In other words, in long-run, income inequality increases economic growth in Pakistan.
This does not support the claim of Pritchett, (1997) and Stiglitz, (2002) that globalization has
contributed to income inequality in the poorest developing countries at least in case of Pakistan.
Despite being a middle income economy
14
, Pakistans inequality gap is still widening. In
Pakistan, a mild increase in inequality from 0.357 to 0.369 is recorded during 1976-1985 and
slight decrease from 1985-1988 (from 0.363-0.348) while from 1989 to 2006 it has continuously
increased (from 0.365-0.421). This suggests that income inequality is still on the rise and
consequently contributes to growth in per capita income, as a whole.
31
International remittance is found to be positively significant at 5%. The positive impact of
remittances is consistent with the findings of Adams, (1991) in case of Egypt; Rodriguez, (1998)
for Philippines and, Iqbal and Sattar, (2005) for Pakistan. Again, in case of remittances, both
short and long-run estimates show a positive sign indicating no evidence of U-curve relationship
for international remittance. Despite the fact that international remittances are significant, it
suggests that availability of remittances is limited to certain groups. Stark et al. (1986) argued that
impact of foreign remittances depend on migration history where migrants might not have equal
opportunity to migrate. Therefore, migrant who are well informed on foreign labor market are
usually those who are in a better income bracket might have more opportunity. This might have
contributed to the widening gap in income inequality in case of Pakistan. However, relatively,
international remittance has a smaller impact than that of per capita income. This may be due to
two reasons. First, the small effect of international remittance may be due to informal transfer of
remittances that remain the main limitation in this paper. International remittances are channeled
from two possible ways in Pakistan, the formal channel via banking systems and informal way
that is known as hawala or hundi. In 2001, it is predicted that 20% of remittances to Pakistan
entered through formal channels while vast majority uses the informal system. Second, the small
magnitude
15
(size of coefficient) of international remittances also seems to suggest that other
factors may have more profound effect on income inequality than international remittance.
Perhaps, it is the domestic deregulation and external liberalization that impacted income
inequality in Pakistan more than the international remittances themselves.
Table-5: Long and Short Run Error Correction Model Estimates
Pane-A: Long Run Estimates-Dependent Variable:
t
Y ln
32
Variable Coefficient Std. Error T-Statistic
Constant -3.563** 0.653 -5.458
t
I ln
0.221* 0.061 3.615
t
R ln
0.057** 0.024 1.923
Pane-B: Short Run(Error Correction Model) -Dependent Variable:
t
Y ln A
Constant -0.007 0.094 -0.070
t
I ln A
0.149** 0.049 2.997
t
R ln
0.038** 0.017 2.243
1 t
ECM
-0.675* 0.182 -3.707
Adj-R
2
0.353
F-statistic 5.263*
Diagnostic Test F-statistic P-value
NORMAL
2
2.614 0.271
SERIAL
2
0.018 0.893
ARCH
2
0.3111 0.778
WHITE
2
0.544 0.456
REMSAY
2
0.334 0.563
Note: * and ** denote the significant at 1% and 5% levels respectively.
NORMAL
2
for
white heteroskedasticity and
REMSAY
2