β
4i
D +
p
−
1
∑
i
=
0
β
3i
D (logGDPC
t
−
i
)
3
+
p
−
1
∑
i
=
0
β
5i
DlogP
t
−
i
+
p
−
1
∑
i
=
0
β
6i
DlogImp
t
−
i
+
∑ I
k
+ γ
t
+ ε
t
(1)
where, log is the natural logarithm, D indicates first difference, Elec is the residential electricity
consumption in per capita terms, GDPC is the Gross Domestic Product per capita (at constant 2010 US$),
P is the real household electricity price, imp is the Imports of goods and services,
∑ I
k
are the dummy
variables that capture the regime changes in the model, c is a constant, γ is a temporal dummy in years
for the 1970–2013 period and ε
t
is the error term. In addition, the lag is calculated by using the VAR
optimal model, minimizing the AIC, SIC and HIC information criteria.
The introduction of the GDPC variable and its squared and cubed terms in Equation (1) may
generate multicollinearity problems among the variables [
61
], which may be analyzed by using
the values of the variance inflation factors (VIFs). Nevertheless, according to Pablo-Romero and
Sánchez-Braza [
62
], these multicollinearity problems may be mitigated by converting each explanatory
variable to deviations from the geometric mean of the sample. These new variables are denoted
respectively as elec, y, y
2
, y
3
, p, and imp instead of LogElec, LogGDPC, (LogGDPC)
2
, (LogGDPC)
3
, LogP
and LogImp, respectively. Thus, the Equation (1) may be rewritten as follow:
Delec
t
=
c
+
α
1
elec
t−1
+
α
2
y
t−1
+
α
3
y
2
t−1
+
α
4
y
3
t−1
+
α
5
p
t−1
+
α
6
imp
t−1
+
p−1
∑
i=1
β
1i
Delec
t−i
+
p−1
∑
i=0
β
2i
Dy
t−i
+
p−1
∑
i=0
β
3i
Dy
2
t−i
+
p−1
∑
i=0
β
4i
Dy
3
t−i
+
p−1
∑
i=0
β
5i
Dp
t−i
+
p−1
∑
i=0
β
6i
Dimp
t−i
+
∑ I
k
+
γ
t
+
ε
t
(2)
Once the Equation (2) has been estimated, the bounds testing approach to cointegration should
be implemented to test the presence of cointegration between the studied variables. To this end, the
Fisher-test (F-stat) for the lagged level variables joint significance is used. The null hypothesis to be
Energies 2018, 11, 1656
5 of 18
tested is H
0
: α
1
= α
2
= α
3
= α
4
= α
5
= α
6
= 0, indicating no cointegration relationship between the
studied variables. This hypothesis is rejected when the calculated F-test value exceeds the upper critical
bounds value [
53
].
If the null hypothesis is rejected, then, in a second step, the conditional ARDL (q1, q2, q3, q4, q5,
q6, q7) long-run model capturing the long-run dynamic should be estimated. The function form may
be written as follow [
59
]:
elec
t
=
c
+
q1
∑
i=1
Dostları ilə paylaş: |