2. Methodology
According to the previous studies, the residential electricity use, measured as Kwh per capita,
is considered to be a function of the GDP per capita (constant 2010 US$), the real electricity prices
for Kwh and the imports of goods and services as a share of GDP [
49
–
51
]. Additionally, in this paper,
the squared and cubed values of GDP per capita have also been considered as explanatory variables of
the electricity consumption, as in Yin et al. [
47
] and Pablo-Romero et al. [
48
]. Introducing these variables
gives more flexibility to the demand function, allowing the study of how the electricity consumption
varies as the income variable grows. In that regard, the electricity consumption elasticity with respect
to GDP, may not be constant through time, varying with GDP per capita. Thus, if the estimated
coefficient for GDP per capita is positive, and those related to the GDP per capita squared and the
GDP per capita cubed, are both negatives, from which an inverted N-shape is obtained. The second
threshold level is reached when the electricity elasticity, with respect to income, becomes zero on
a decreasing elasticity trend [
10
]. Nevertheless, if the estimated coefficients related to GDP variables
have other signs, alternative types of relationships will be defined between electricity consumption
and GDP [
52
].
The ARDL cointegration testing approach by Pesaran et al. [
53
], and extended to introduce
break points, has been used to analyze the long-run and short-run relationships between the
residential electricity consumption and the defined explanatory variables, through the 1970–2013
period. According to Pesaran et al. [
53
], this model can be used even if variables are not integrated
to the same order, but being I(1), I(0) or fractionally integrated. It is worth noting that the ARDL
cointegration testing approach by Pesaran et al. [
53
] holds no feedbacks running from dependent
variable to independent variable at the levels. Nevertheless, McNown et al. [
54
] have developed the
Energies 2018, 11, 1656
4 of 18
ARDL bounds test with bootstrap techniques which can solve the endogeneity problems. Therefore,
if there are endogeneity problems, it is convenient to use the ARDL bounds test with bootstrap
techniques. In this paper, no endogeneity problems are detected (see Section
4.5
). Therefore the ARDL
cointegration testing approach by Pesaran et al. [
53
] has been used.
The breakpoint unit root test has been used to test a possible presence of structural breaks in
the studied variables, as Perron [
55
] considered that most macroeconomic series present transitory
fluctuations. Therefore, if the unit root test does not take into account the structural break point,
selected exogenously, the decision toward exists against rejecting will be biased. Following Perron [
55
]
and Zivot and Andrews [
56
], there are three types of break point: those related to changes in the level
of the time series (change in the intercept), those related to the change in the rate of growth (change in
the trend) and the result of both (change in intercept and trend). In order to examine the stationarity
of each variable, the two-break LM unit root test by Lee-Strazicich is used [
57
]. The null hypothesis
implies that there is a unit root, while the alternative hypothesis implies the series is breakpoint
stationary. The breaks are determined endogenously. The Crash or A model, that captures a change in
the level of the series, has been adopted in this study.
According to previous literature (see for example, Narayan and Smith [
58
], Belloumi [
59
],
Charfeddine et al. [
60
], among others), the Pesaran et al. [
53
] model may be implemented in three
steps by sequentially estimating three functions. The first function is defined in order to estimate the
conditional error correction of the ARDL model. The ordinary least squares method (OLS) is used.
Equation (1) may be expressed as follows:
DlogElec
t
= c + α
1
logElec
t
−
1
+ α
2
logGDPC
t
−
1
+ α
3
(logGDPC
t
−
1
)
2
+ α
4
(logGDPC
t
−
1
)
3
+α
5
logP
t
−
1
+ α
6
logImp
t
−
1
+
p
−
1
∑
i
=
1
β
1i
DlogElec
t
−
i
+
p
−
1
∑
i
=
0
β
2i
DlogGDPC
t
−
i
+
p
−
1
∑
i
=
0
β
3i
D(logGDPC
t
−
i
)
2
+
p
−
1
∑
i
=
0
Dostları ilə paylaş: |