Annals of Data Science (2019) 6(4):787–806 797

where the first row corresponds to t = 3 up to t = T and Z S is instrumental variable for

the system GMM estimator. The added moment restrictions can be written as:
( ) ⎛ εi3 ⎞
E Zli εi = 0 with εi = ⎜ ⋮ ⎟
⎜ ⎟
⎝ εiT ⎠
The system model consists of (T − 2) stacked equations in first differences and
(T − 2) stacked equations in levels for t = 3, … , T. The system GMM estimator is:
( ) (
𝛿̂ [ ]� [ ])−1 ([ ]� )
𝛾̂GMMsystem = = ΔY−l , ΔX Zs WZs � ΔY−l , ΔX ΔY−l , ΔX Zs WZs � Δy
𝛽̂
(3.7)
Blundell and Bond [5] developed the moment conditions for the equations in lev-
els and in the lagged differences as instruments that the additional moment’s restric-
tions remain valid also in case of weak instruments and explicitly define the assump-
tions on the initial conditions that need to be satisfied for the system estimator to be
valid.
System GMM estimator uses only the most recent lag of the first differences as
instruments for the equations in levels, as the studies are already using the lagged
levels as instruments for the equations in first differences. The use of additional
lagged first-differences would result in redundant moment condition [5, 17]. Works
of literature have paid attention to the choice of the weighting matrix for the one-
step GMM procedure and for the first step in a two-step procedure. Windmeijer [23]
found that two-step GMM performs better than one-step GMM in estimating coef-
ficients, with lower bias and standard errors. And the reported two-step standard
errors, with his correction, are quite accurate.
The specified model, considering the model in this empirical application, was
replicated by following the Stata commands of Roodman [17]. According to Rood-
man [17], as a developer of the xtabond2 code, discussed pointers for the proper
estimation of dynamic panel estimators. First, short panel data is appropriate
for the estimators. If T is large, dynamic panel bias becomes insignificant, and a
more straightforward (standard or traditional) fixed effects estimator works. Sec-
ond, include time dummies in the estimation helps to make assumption more likely
to hold such as the autocorrelation test and the robust estimates of the coefficient
standard errors assume no correlation across individuals and in the idiosyncratic dis-
turbances. Last, but not least, report all specification choices’ with these estimators
involves many choices, and researchers should report the ones they make: difference
or system GMM; first differences or orthogonal deviations; one or two-step estima-
tion; and non-robust or Windmeijer-corrected.

3.4 Specification Tests

3.4.1 Test for Validity of Instrument Sets

Sargan [18] and Hansen [9] shown test for the assumption about the absence of any
(asymptotic) correlation between the instrumental variables and the disturbances.

13
798 Annals of Data Science (2019) 6(4):787–806

In addition to the former assumption, the choice of “good” instruments lies in the
potency of the correlation between the endogenous regressors and the instruments.
Indeed, Blundell and Bond [5], show that a small correlation results in erratic
parameter estimates. Moreover, Baltagi [4] showed that there is a significant amount
of the deleterious effects of weak instruments. Sargan (Hansen) over identification
restriction (OIR) test, which is given by:
[N ]−1
∑ ( )( )�
m = Δu�� Z Z�i Δ̂ui Δ̂ui Zi Z� (Δ̂u) ∼ χ2
L−K
i=1

where L refers to the number of columns of Z and Δ̂u denotes the residuals from a
two-step estimation.
Under the assumption of no serial correlation in εit , Δεit follow an MA(1) process
and, the series y­ i,t-2, ­yi,t-3, …, ­yi,t-T are valid instruments for estimating this model.
However, if ε�it s are serially correlated, this series no longer constitutes a valid instru-
ment set. This implies that one can test ­H0 or 𝜀it is serially uncorrelated against ­H1
by comparing the difference between Sargan and Hansen statistics corresponding to
two instrument sets: Z ­ 0 contains the instruments defined by the series y­ i,t-2, ­yi,t-3, …,
­yi,t-T and ­Z1 is an instrument set not dependent on the assumption of 𝜀it not being
serially correlated. Indeed, to increase the test’s power, one might be more specific
for ­H1 and test H ­ 0 against ­H1 with the latter hypothesizing, denote the difference
between the two Sargan and Hansen statistics by D ­ Qsh. Under the null this is distrib-
uted as χ2p −p where ­ P0 and ­ P1 are the number of instruments in ­ Z0 and ­ Z1,
0 1
respectively.

3.4.2 Test for the Absence of Serial Correlation in ε

The existence of serial correlation in 𝜀it will typically overthrow the use of lagged
values and first differences of the endogenous variable as instruments. So, it is cru-
cial to test for such serial correlation. Arellano and Bond [2] proposed, that is made
based on result of the model estimated in first differences. Let Δ̂ε be the vector of
residuals from the model in first differences, Δ̂ε−2, its second lag value, and Δ̂ε∗
the reduction of the vector Δ̂ε allowing computation of the product Δ̂ε−2 Δ̂ε∗. This
test provides a measure of the importance of serial correlation of order 2 once the
model is written in first differences. If they ε�it s are serially uncorrelated given by
Δεit = εit − εi,t−1 follow an MA (1) process and thus, are not correlated at order 2.
On the contrary, if Δ𝜀it appears to be correlated of order two, one can infer that the
disturbances 𝜀it exhibit some serial correlation. The test statistics is given by:
�
m2 = Δ̂ε−2 Δ̂ε∗ ∕ξ1∕2

The test is one-sided as it will exhibit positive serial correlation. As test statistic,
m2 show that, under the null of no serial correlation in Δεit at order 2. One rejects H ­ 0
of no serial correlation.

13
Annals of Data Science (2019) 6(4):787–806 799

3.5 Panel Unit Root Test: Harris–Tsavalis (HT) Test

Panel unit root tests of Harris and Tzavalis derived a unit-root test for a short panel
that assumes the time dimension, T, is fixed or short. Their simulation results sug-
gest that the test has favorable size and power properties for N greater than 25, and
they report that power improves faster as T increases for a given N than when N
increases for a given T in balanced panel data. HT assumed that is independent and
identically distributed with constant variance across panels. Because of the bias
induced by the inclusion of the panel means in this model, the expected value of
the estimator is not equal to unity under the null hypothesis. Notice that, the HT test
assumes that all panels share the same autoregressive parameter [4].

4 Results and Discussion

4.1 Panel Unit Root Test (PURT)

Harris–Tzavalis panel unit test described in methodology Sect. 3.5 requires cross-
sectional independence, which is a strong assumption to make in the macroeconomic
environment. However, there is a way to manage this issue, such as by assuming that
the cross-sectional dependence is following a common trend across cross-sections.
Thus, the issue mitigated by subtracting period means across cross-sections from
each individual observation to eliminate a possible trend common to all cross-sec-
tions. To remove the cross-sectional dependence from the individual panel the Stata
command demean was used. Under the null hypothesis, the panel variables contain
a unit root versus alternative stationary. The test statistic and p-values for each vari-
able in level and the first difference are reported in Table 1.

Table 1  Panel unit root test Harris–Tzavalis test statistic

results of the variables at level
and first difference. Source: Level First difference
Own computation
Ethiopia’s coffee exports − 12.719*** –
Ethiopia’s GDP − 8.579*** –
Ethiopia’s openness to trade − 4.279*** –
Ethiopia’s institutional quality − 19.698*** –
Weighted distance − 32.099*** –
Real exchange rate 8.309 − 12.957***
Real ­GDPa 3.197 − 19.095***
Populationa 2.296 − 13.767***
Openness to ­tradea 2.108 − 22.204***
Ethiopia’s population 3.383 − 12.102***
a
Of importing countries
***, **, and * represent 1, 5 and 10% significance levels, respec-
tively

13
800 Annals of Data Science (2019) 6(4):787–806

The tests suggested that five variables which are Ethiopia’s coffee exports, Ethio-
pia’s real gross domestic product, Ethiopia’s measure of institutional quality, Ethio-
pia’s openness to trade, and weighted distance, were stationary and no further adjust-
ment is needed to make them stationary. The rest of the variables were unit root and
the necessary adjustments were made to avoid the problem that arises from spuri-
ous regression. All variables which were a real gross domestic product of import-
ing countries, Ethiopia’s population, openness to trade of importing countries, real
exchange rate, and the population of importing countries which were found to con-
tain unit root when tested at the level were found to be stationary when tested after
differencing.

4.2 Dynamic Panel Data Model Estimation

The robust regression result of the linear dynamic panel gravity model of Ethi-
opia’s coffee export performance was estimated by two-step GMM estima-
tion approach and presented in Table 2. It shows Ethiopia’s coffee exports per-
formance was found to be positively and significantly influenced by real gross
domestic product and openness to trade of the importing countries, population,
real gross domestic product, and institutional quality of the supply side or Ethio-
pia’s, and lagged Ethiopia’s coffee exports. But the weighted distance was found

Table 2  Ethiopia’s coffee ECE performance by Two-step GMM estimate

exports (ECE) performance
model. Source: own Variables System model
computation
Coefficientsa WCSEb

One lag of ECE 0.466 (23.200)*** 0.137

GDPc 1.559 (6.830)*** 2.236
Real exchange rate − 0.107 (− 0.06) 1.303
Openness to ­tradec 1.045 (3.080)*** 1.617
Weighted distance − 0.640 (− 2.600)*** 1.019
Populationc − 0.159 (− 0.08) 9.796
Ethiopia’s GDP 0.258 (6.070)*** 0.184
Ethiopia’s openness to trade 0.026 (1.62) 0.085
Ethiopia’s population 1.308 (9.810)*** 0.999
Ethiopia’s institutional quality 3.094 (7.510)*** 1.812
No. of observations 710
No. of groups 71
No. of instruments 73
a
Test statistics in parenthesis
b
Windmeijer’s corrected standard errors
c
Of importing countries
***, **, and * represent 1%, 5% and 10% significance levels, respec-
tively

13
Annals of Data Science (2019) 6(4):787–806 801

to have a significant negative effect on coffee export. Finally, the remain three
explanatory variables: the population of the importing countries, Ethiopia’s meas-
ure of openness to trade index, and real exchange rate were found to be insignifi-
cant even at 10% level of significance.
The coefficient on real GDP of importing countries was found to be statisti-
cally significant at 1% level. The positive value is consistent with the theoretical
prediction of gravity trade flow that anticipates that trade volumes increase with an
increase in a partner’s economic size. The estimated coefficient of 1.56 of the real
GDP of importing countries suggests that a 1% increase in real GDP of importing
countries will result in roughly a 1.56% increase in the flows of Ethiopia’s coffee
exports, ceteris paribus. Importing countries with larger GDP indicate for higher
demand for Ethiopian coffee. This is consistent with the Orindi [15]. Using the grav-
ity model, the study discovered that the real GDP of importing countries had a posi-
tive effect on the value of bilateral trade between Kenya and the 25 trading partners
considered in the study for the years 1964–2008.
On the other hand, the estimated coefficient of Ethiopia’s real GDP was 0.26
which implies that keeping other variables constant, a percent increase in Ethiopia’s
real GDP will result in roughly 0.26% increase in Ethiopia’s coffee exports. Supply-
side GDP improvement is an indication for better production power. The result is in
conformity with the study on African countries used a dynamic panel data set for
48 African countries over the period 1987–2006 to identify the key determinants
of export performance and found that supply capacity had a positive effect on the
export performance [12].
Besides, Ethiopia’s population was found to significantly and positively deter-
mine Ethiopia’s coffee exports flow. The estimated coefficient was 1.31 which
implies that keeping other variables constant, a 1% increases in Ethiopia’s popula-
tion results in a 1.31% increase in Ethiopia’s coffee exports. It also indicates that the
rate of adsorption in response to a change in population size is weaker. The result is
consistent with Dlamini et al. [6], they analyzed the factors determining sugar export
from Swaziland to trading partners using a gravity model and used annual panel
dataset for the period 2001 to 2013, and found that exporter population had a posi-
tive effect on the performance of the export sector.
This study also found that the degree of openness to trade or trade liberaliza-
tion implies a substantial reduction in tariff and non-tariff barriers has an undeni-
able impact in the bilateral trade process. The estimated coefficient for the degree of
openness to trade of importing countries was 1.05 which means that, ceteris paribus,
a percentage improvement in the degree of openness to trade increase Ethiopia’s cof-
fee exports flow to these countries by around 1.05%. Thus, there the rise in Ethio-
pia’s coffee over those periods can be attributed to the improvement in the open-
ness of importing countries to trade. Hence, openness is an important determinant
of Ethiopia’s coffee export earnings and the improvements in the performance of
Ethiopia’s coffee exports can be partly attributed to the improvement in the open-
ness of importing countries. The result is in conformity with the Babatunde [3] ana-
lyzed Sub-Saharan African export performance and used panel data set from 1980 to
2005 and found that the export performance of exporter’s had influence by openness
to trade of importers’.

13