Science of the Total Environment 651 (2019) 2886–2898
Contents lists available at ScienceDirect
Science of the Total Environment
j ourna l homepage: www.e lsev ie r .com/ locate /sc i totenvDoes energy consumption follow asymmetric behavior? An assessment
of Ghana's energy sector dynamicsSamuel Asumadu Sarkodie a,⁎, Aba Obrumah Crentsil b, Phebe Asantewaa Owusu c
a Department of Environmental Sciences, Faculty of Science and Engineering, Macquarie University, NSW 2109 Australia
b Institute of Statistical, Social & Economic Research (ISSER), University of Ghana, P.O. Box LG 74, Legon, Ghana
c Sustainable Environment and Energy Systems, Middle East Technical University, Northern Cyprus Campus, Kalkanli, Guzelyurt, TRNC 99738/Mersin 10, TurkeyH I G H L I G H T S G R A P H I C A L A B S T R A C T• Does energy evolves in different states
by transitioning over a finite set of
states?
• We ascertain if energy consumption fol-
low an asymmetric behavior.
• Weexamine the unobserved factors un-
derpinning energy crisis in Ghana.
• Markov-switching, NIPALS regression,
and neural network analysis are used.⁎ Corresponding author.
E-mail address: asumadusarkodiesamuel@yahoo.com
https://doi.org/10.1016/j.scitotenv.2018.10.147
0048-9697/© 2018 Elsevier B.V. All rights reserved.a b s t r a c ta r t i c l e i n f oArticle history:
Received 10 September 2018
Received in revised form 10 October 2018
Accepted 10 October 2018
Available online 13 October 2018
Editor: Damia Barcelo
JEL classification:
Q43
Q47
Q41The study answered the following questions: First, does energy evolves in different regimes by transitioning over
a finite set of unobserved states? Second, does energy consumption follow an asymmetric behavior over “energy
boom” and energy scarcity? and, Third, are there unobserved factors underpinning energy crisis? We employed
Markov-switching dynamic regression to examine the asymmetric effect, NIPALS regression to examine energy
determinants and neural network analysis for prediction. The neural network model suggests a 99% prediction
of energy consumption by the predictor variables. It was evident that energy consumption evolves in two states
by transitioning over a finite set of unobserved states. The 11.6% growth in energy consumption is expected to
occur in 4.1 years while energy crisis is expected to last for 3.7 years. Technological advancement and the devel-
opment of green energy through foreign direct investment are essential to improve energy sector portfolio.
© 2018 Elsevier B.V. All rights reserved.Keywords:
Markov-switching
Energy efficiency
Neural network
NIPALS
Ghana(S.A. Sarkodie).
S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898 28871. Introduction
Energy production and consumption remain the largest contributor
of anthropogenic carbon dioxide emissions, as such, mitigation option
requires measures that promote energy efficiency and the substitution
of conventional energy sources with renewable energy technologies
(Owusu and Asumadu, 2016; Sarkodie and Strezov, 2018). While the
Sustainable Development Goal Seven seeks to ensure the availability
and accessibility of clean and modern energy technologies (United
Nations, 2015), the underlying factors that propel energy consumption
levels should not be underestimated.
The determination of factors affecting energy consumption and eco-
nomic growth has been a topical subject of many studies, since the
1970s. However, the existing evidence on the nexus between economic
growth and energy consumption has been inconclusive. Yet, an under-
standing of the determinants of energy consumption and its modeling
in emerging economies is important. Studies on the relationship be-
tween energy consumption and economic growth have been found to
be very complex due to the four possible impact scenarios, namely
(Inglesi-Lotz and Pouris, 2016; Sarkodie and Adom, 2018): growth, con-
servative, feedback and neutrality hypotheses.
The growth hypothesis postulates a unidirectional causality fromen-
ergy consumption to economic growth. This implies that energy saving
policies may hinder growth because such an economy is dependent on
energy to grow. Second, the conservative hypothesis refers to a unidi-
rectional causality from economic growth to energy consumption.
Here, energy-saving policies may have little or no negative effects on
growth (Destek and Sarkodie, 2019). However, if the causal relationship
from energy consumption to growth is positive, the adoption of energy-
saving policies can lead to a decline in growth and employment. Con-
versely, if the causal relationship from economic growth to energy con-
sumption is negative, the use of energy saving policies can lead to an
increase in output. The feedback hypothesis refers to a bidirectional cau-
sality between energy consumption and economic growth. Thus, energy
conservation policies can restrict the economy and vice versa. The neu-
trality hypothesis argues that there is no causal relationship between
energy consumption and economic growth. Accordingly, reducing en-
ergy consumption is ineffective for economic growth.
The lack of consensus evidence in energy-growth correlation is pri-
marily due to the omission of other potential determinants in the
modeling of energy demand function (Chang, 2015). As such, we pro-
pose using econometric methods that allow for the consideration of
an asymmetric effect and viable data series in studying the causal fac-
tors of energy consumption.
The aim of the study is to answer the following questions: First, does
energy evolves in a different state (regime) by transitioning over a finite
set of unobserved states? Second, does energy consumption follow an
asymmetric behavior over “energy boom” and energy scarcity? Third,
are there unobserved factors underpinning energy crisis? and Fourth,
what are the probabilities that a country will experience “energy
boom” or “energy crisis”?
To the best of our knowledge, no study in existing literature con-
siders trio dynamic models that minimize the complexities of available
models in the literature. First, this study overcomes multicollinearity, a
problem with time series variables by using both NIPALS and neural
network models. Second, the study controls for structural breaks and
discontinuous shifts in regression regimes at an unknown point using
the Markov-switching dynamic regression. The Markov-switching dy-
namic regression has been widely used in financial economics
(Hamilton, 1989), political (Jones et al., 2010) and health sciences
(Martínez-Beneito et al., 2008) but has not been applied in energy eco-
nomics. Importantly, the Markov-switching dynamic regression can
easily switch the states according to the Markov process; the speed of
adjustment/correction is quick after a change of state and has the ability
to deal with a high number of variable observations. Third, the study
draws attention to unobserved variables that play a critical role inenergy demand-side management and energy conservation while ex-
amining the predictive power of trio models, thus, providing new evi-
dence with policy implications.
The remainder of the study consists of section two “Literature
Review”, section three “Methodology”, section four “Results and
Discussion”, and section five “Conclusion”.
Nomenclature
3G Third Generation of Broadband Cellular Network Technology
4G Fourth Generation of Broadband Cellular Network Technology
ARDL Autoregressive Distributed Lag
GSM Global System for Mobile communication
ICT Information and Communication Technology
LMDI Logarithmic Mean Divisia Index
MAD Mean Absolute Deviation
MAPE Mean Absolute Percentage Error
NEPAD New Partnership for Africa's Development
NIPALS Nonlinear Iterative Partial Least Squares
OECD Organisation for Economic Co-Operation and Development
RSME Root Mean Square Error
SSE Error Sum of Squares
VECM Vector Error Correction Model
VIP Variable Importance of Projection
2. Literature review
An investigation of the causal nexus between economic activities
and energy consumption and Greenhouse gas emissions has been stud-
ied in several empirical works. According to Sarkodie and Owusu
(2016), the majority of the existing literature can be categorized into
three; the first category of research, examines the causal effect of envi-
ronmental pollution, energy consumption, and macroeconomic vari-
ables by testing the validity of the environmental Kuznets curve
hypothesis.
Usingfixed effectsmodel and themethod of least square generalized
linear regression in China between 1995 and 2010, Zhang and Lin
(2012), illustrated that the demographic intensities, GDP, industrial
production, and energy consumption have an impact on CO2 emissions.
Ahmed et al. (2016) examined the causal effect of carbon dioxide emis-
sions, GDP, and energy consumption in Brazil, South Africa, China and
India using a panel data spanning from 1970 to 2013 using the fully
modified least squares method. Their results confirmed the validity of
the environmental Kuznets curve hypothesis and found evidence of bi-
directional causality between carbon dioxide emissions and energy
consumption.
According to (Sarkodie and Owusu, 2016) the second category of re-
search examines the causal effect of environmental pollution, energy
consumption, andmacroeconomic variableswithout testing the validity
of the environmental Kuznets curve hypothesis. Employing a multivar-
iate co-integration analysis, ARDL and vector error correction modeling
techniques to investigate in Ghana for the period 1971–2013, Asumadu
and Owusu (2016b) examined the relationship between carbon dioxide
emissions, GDP, energy consumption and population. Their results sug-
gested the existence of mutual causality between Ghana's energy con-
sumption and GDP.
Another study by Owusu and Samuel (2016) investigated the rela-
tionship between carbon dioxide emissions, energy consumption, pop-
ulation and GDP in Ghana using VECM technique for the period
1980–2012. Their findings suggested that the continuous increase in
population growth within the study period has resulted in a substantial
increase in energy demand and CO2 emissions in Ghana. A bidirectional
causality was observed between carbon dioxide emissions and energy
consumption.
In the sameway,Mohiuddin et al. (2016) examined the relationship
between carbon dioxide emissions, energy consumption (EC), GDP, and
electricity production from oil, coal and natural gas, in Pakistan from
2888 S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898
Table 1
Data variable description.
Variable name Variable code
Total greenhouse gas emissions (kt of CO2 equivalent) GHG
Energy consumption (kg of oil equivalent per capita) ENUS
GDP (current LCU) GDP
Mobile cellular subscriptions MCS
Population POP
Total fisheries production (metric tons) TFP
Industry, value added (current US$) INA
Household final consumption expenditure (current LCU) HFICE
Foreign direct investment, net inflows (% of GDP) FDIN
Energy imports, net (% of energy use) EGIM
Agricultural machinery AGM
Electric power transmission and distribution losses (% of output) EPTDL1971 to 2013 and found evidence of long-run equilibrium relationship
running from EC, electricity production from coal, electricity production
from natural gas, electricity production from oil and GDP to carbon di-
oxide emissions.
The third category of research according to (Sarkodie and Owusu,
2016) examines the causal effect of environmental pollution and agri-
cultural variables. Using Chinese official statistical data, Zou et al.
(2015), investigated the emissions of greenhouse gases from agricul-
tural irrigation to inform strategies for reasonable use of water re-
sources and emission reduction. The study found out that the total
carbon dioxide equivalent (CO2-e) emission from agricultural irrigation
is 36.72–54.16 Mt. Emissions from energy activities in irrigation (in-
cluding water pumping and conveyance) account for 50%–70% of total
emissions from energy activities in the agriculture sector. Groundwater
pumpingwas the biggest emission source, accounting for 60.97% of total
irrigation emissions.
Similarly using the LMDI technique, Li et al. (2014) investigated ag-
ricultural CO2 emissions in China from 1994 to 2011. Their findings sug-
gested that economic development acts to increase CO2 emissions
significantly whiles Agricultural subsidy acts to reduce CO2 emissions
effectively.
The fourth category of research (Faucheux and Nicolaï, 2011; Hamdi
et al., 2014; Moyer and Hughes, 2012) examines the causal effect of en-
ergy consumption, CO2 emission, and information communication tech-
nology usage. Employing panel unit root test accounting for the
presence of cross-sectional dependence, a panel cointegration test, the
Pooled Mean Group regression technique and Dumitrescu-Hurlin cau-
sality test, Salahuddin and Alam (2016) examined the short- and
long-run effects of ICT use and economic growth on electricity con-
sumption using OECD panel data for the period of 1985 to 2012. Their
findings suggested electricity consumption in both the short- and the
long-run had a direct relationship to the use of ICT and economic
growth. In a similar study using data for a period of 1985–2012 in
Australia, Salahuddin and Alam (2015) examines the short- and long-
run effects of Internet usage and economic growth on electricity. Their
results indicated the presence of a unidirectional between Internet
usage to economic growth and electricity consumption.
Examining the trend of worldwide electricity consumption Van
Heddeghem et al. (2014) and showed that three key ICT categories,
namely, communication networks, personal computers, and data cen-
ters, has increased in 2012 from its level in 2007 due to increase in elec-
tricity consumption.
The majority of the aforementioned literature assured the existence
of a closed-form relationship between energy consumption andmacro-
economic factorsmostly, focusing on a casualtywith less than three var-
iables. As a contribution to literature, an attempt is made to investigate
the causal relationship between energy consumption, agricultural ma-
chinery, foreign direct investment net inflows, economic growth, total
greenhouse gas emissions, industrialization, total fisheries production,
net energy imports, electric power transmission and distribution losses,
household final consumption expenditure, mobile cellular subscrip-
tions, and population using time series data from 1971 to 2014 in
Ghana.
3. Methodology
3.1. Data
To conduct an assessment of energy consumption in Ghana, the
study employs data from 1971 to 2014 from theWorld BankWorld De-
velopment Indicator database (World Bank, 2016). Twelve data vari-
ables are used in the study namely; Total greenhouse gas emissions
(kt of CO2 equivalent), Energy consumption (kg of oil equivalent per
capita), GDP (current LCU), Mobile cellular subscriptions, Population,
Total fisheries production (metric tons), Industry, value added (current
US$) as a proxy for industrialization, Household final consumptionexpenditure (current LCU), Foreign direct investment, net inflows (%
of GDP), Energy imports, net (% of energy use), Agricultural machinery
and Electric power transmission and distribution losses (% of output)
presented in Table 1. Energy use employed as a proxy for energy con-
sumption is defined by theWorld Bank as “the use of primary energy be-
fore transformation to other end-use fuels, which is equal to indigenous
production plus imports and stocks change, minus exports and fuels sup-
plied to ships and aircraft engaged in international transport” (World
Bank, 2016). Fig. 1 presents the trend of the study variables. A visual as-
sessment of Fig. 1 shows that the trend of the variables exhibits some
unexplainable behaviors and complexities that can only be ascertained
through a dynamic regression model elaborated in the subsequent
section.
3.2. Model estimation
3.2.1. Markov-switching dynamic regression
The Markov-switching dynamic regression model was initiated by
Goldfeld and Quandt (1973); Quandt (1972) and was first employed
by Hamilton (1989) in economics to examine the observed asymmetric
behavior in the growth rate of GDP. TheMarkov-switching dynamic re-
gression model is “rich enough” to capture the dynamic and switching
behavior of macroeconomic and energy-related variables thus, allows
quick adjustment/correction of the state classification measures
(Hamilton, 1989; Zhu et al., 2017).
The general specification of the Markov-switching dynamic regres-
sion model is expressed as:
Yt ¼ φs þ Xtα þ Ztβs þ εs ð1Þ
where Yt is the dependent variable, φs is the “state-dependent” inter-
cept, Xt and Zt represent exogenous variables in a vector form, α repre-
sents the “state-invariant” coefficients, βs is the “state-dependent”
coefficients and εs represents the independent and identically distrib-
uted normal error with a “state-dependent” variance σ2s and a zero
mean.
The two-state Markov state-switching model considers that the re-
sponse of the independent variables' return to structural energy con-
sumption shocks is dependent on the state (St) at time (t). St denotes
an observable two-state and 1st order Markov process. The transition
probability of the two-state is expressed in a matrix as:

 
¼ p11 p21P p12 p22 ð2Þ
where pij= P(St = j|St−1 = i), by means of∑ 2j=1pij=1, i (i=1,2) de-
notes the number of states involved in the Markov process.
3.2.2. Neural network
The neural network is a function of some hidden nodes derived
from the non-linear functions of the original inputted variables.
S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898 2889
440 2,080 50
2,040
400 40
2,000
30
360 1,960
20
320 1,920
10
1,880
280
1,840 0
240 1,800 -10
75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15
Year Year Year
ENUS (kg of oil equivalent per capita) AGM EGIM (% of energy use)
30 10 1.6E+11
25 8
1.2E+11
20 6
15 4 8.0E+10
10 2
4.0E+10
5 0
0 -2 0.0E+00
75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15
Year Year Year
EPTDL(% ofoutput) FDIN (% of GDP) GDP (currentLCU)
160,000 1E+11 1.6E+10
8E+10
120,000 1.2E+10
6E+10
80,000 8.0E+09
4E+10
40,000 4.0E+09
2E+10
0 0E+00 0.0E+00
75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15
Year Year Year
GHG(kt ofCO2 equivalent) HFICE (currentLCU) INA (current US$)
40,000,000 28,000,000 500,000
24,000,000 450,000
30,000,000
400,000
20,000,000
20,000,000 350,000
16,000,000
300,000
10,000,000
12,000,000 250,000
0 8,000,000 200,000
75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15 75 80 85 90 95 00 05 10 15
Year Year Year
MCS POP TFP (metric tons)
Fig. 1. Trend of variables.A one-layer feed-forward neural network employed in the study is
expressed as:
0 X  X ! 1k n
PredENUS ¼ f@εo þ h φ þ X w ε Aj i ij j ð3Þ
j¼1 i¼1
where PredENUS is the predicted values of energy consumption
representing the neural network output, f(.) is the non-linear trans-
fer function of the independent variables (inputs) Xi, h(.) is the hid-
den layer activation function applied to the nodes with
corresponding biases of the hidden layer φj, wij denotes the weight
from the input layer to the hidden layer, εo and εj represent the out-
put biases and the weight values from the hidden layer to the output
layer.
3.2.3. NIPALS regression
The Nonlinear Iterative Partial Least Squares (NIPALS) regression
analysis by Sarkodie and Adom (2018); Wold et al. (2001) beginswith the centering and scaling of the dependent variable (Y) and inde-
pendent variables (X). The next step involves the initialization of u =
Y with a subsequent repetition of Eqs. (4)–(9) to reach convergence
thus,
w ¼ X0u=ðu0=uÞ ð4Þ
w≔w=kwk ð5Þ
t ¼ Xw ð6Þ
c ¼ Y 0t=ðt0=tÞ ð7Þ
c≔c=kck ð8Þ
u ¼ Yc ð9Þ
wherew and c are the X and Y loadingswith a unit norm thus, c is a 1 × 1
unit vector converging in a single NIPALS iteration while t and u are the
X and Y scores.
2890 S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898Subsequently, X and Y are regressed on t and u as:
p ¼ X0t=ðt0=tÞ ð10Þ
q ¼ Y 0u=ðu0=uÞ ð11Þ
The next step is a deflation of the matrices of X and Y expressed as:
X≔X−tp0 ð12Þ
Y≔Y−tc0 ð13Þ
The deflation of the matrices is repeated d times gathering the
vectors (t, p, u, q) into matrices to produce a preferred factorization
into X and Y scores T and U, X and Y loadings P and Q, weights of X
and Y W and C derived by gathering w and c vectors into n × d and m
× d matrices and output errors/residuals E and F expressed as:
X ¼ TP0 þ E ð14Þ
Y ¼ UQ 0 þ F ð15Þ
To predict Y from X, the matrix of the regression coefficient (B) is
expressed as:
B ¼ W  C0 ð16Þ
Therefore, the final NIPALS regression is given as:
Ŷ ¼ XB þ ðμY−μ BX Þ ð17Þ
where
B ¼ Σ −1X BΣY 0E ¼ EΣX0 ð18Þ
4. Results and discussion
4.1. Descriptive analysis
This section beginswith a descriptive statistical analysis of the study
variables presented in Table 2. The average energy consumption within
the last four decades was almost 347 kg of oil equivalent per capita, aTable 2
Descriptive statistical analysis.
Statistic ENUS AGM EGIM EPTDL FDIN GDP
Mean 347 1957 23 10 2 12,300,000,000
Median 360 1948 21 5 1 334,000,000
Maximum 418 2076 46 29 10 113,000,000,000
Minimum 272 1807 −7 2 −1 250,000
Std. Dev. 38 66 12 9 3 26,300,000,000
Skewness −0.4048 0.0252 0.2317 0.7448 1.1955 2.5042
Kurtosis 2.2932 2.3680 3.2326 1.7977 3.0377 8.5307
Jarque-Bera 2.0693 0.5863 0.4816 6.5649 9.5298 102.0677
Probability 0.3553 0.7459 0.7860 0.0375a 0.0085a 0.0000a
Correlation 347 1957 23 10 2 12,300,000,000
ENUS 1
AGM 0.1420 1
EGIM −0.4579 −0.2795 1
EPTDL −0.8121 −0.2866 0.8336 1
FDIN 0.1392 −0.3927 0.3901 0.1809 1
GDP −0.6687 −0.5217 0.8882 0.883 0.3257 1
GHG −0.3163 −0.4708 0.8281 0.6683 0.5188 0.7622
HFICE −0.6686 −0.5175 0.8917 0.8856 0.3272 0.9998
INA −0.2143 −0.8340 0.6515 0.5335 0.5676 0.7826
MCS −0.6724 −0.4713 0.6926 0.725 0.142 0.8995
POP −0.3531 −0.8179 0.6831 0.6561 0.5621 0.7973
TFP 0.0389 −0.7153 0.3993 0.2929 0.5241 0.4365
a Rejection of the Null hypothesis of Normal distribution at 5% significance level.minimum and maximum consumption of 272 and 418 kg of oil equiva-
lent per capita. Total greenhouse gas emissions experienced a rise from
19,454 kt of CO2 equivalent to 130,473 kt of CO2 equivalent with an av-
erage of 45,194 kt of CO2 equivalent. Economic growth grew from al-
most GH₵ 250,000 in 1971 to GH₵ 113 billion in 2014 with a mean of
GH₵ 12.3 billion. Mobile cellular subscriptions (i.e. MTN, Vodafone,
Tigo, Airtel, and Expresso) has experienced a significant growth in the
communication industry from zero in 1971 to 30,360,771 in 2014 at
an average of 3,826,884 subscriptions. It is important to note that mo-
bile cellular subscription is employed as a proxy for assessing the
trend of mobile phone usage in Ghana. There has been an exponential
growth in population, from 8,827,273 to 26,786,598 in 2014 at an aver-
age growth of 16,307,944 people. Since Ghana is an agrarian country,
the use of agriculturalmachinery grew from1807 tractors to 2, 076 trac-
tors, with a mean of 1957 tractors and the total fisheries production ap-
preciated from at least 219,327 metric tons to a maximum of 495,683
metric tons at an average of 341,821 metric tons. The economic value
of industrialization has increased from US$ 252,000,000 to US$
12,900,000,000 with a mean of US$ 2,380,000,000. There has been a
huge increase in household final consumption expenditure from GH₵
193,500 to GH₵ 75.5 billion with a mean of GH₵ 8.64 billion. Foreign di-
rect investment net inflows grew from −1% of GDP to 10% of GDP at a
mean of 2%. Ghana's net energy imports grew from −7% of energy use
to 40% of energy use, with an average of 23% of energy use. Electric
power transmission and distribution losses grew from 2% of output to
29% of output, with an average of 10% of output energy production.
Table 2 reveals that except for energy consumption, the remaining var-
iables are positively skewed. Apart from EGIM, FDIN, GDP, HFICE, INA,
and MCS, the remaining variables exhibit a platykurtic distribution. It
is further revealed that only ENUS, AGM, EGIM and TFP are normally
distributed hence, the application of a logarithmic transformation to
the study variables. The correlation analysis reveals that except AGM,
FDIN, and TFP, the remaining variables have a negative relationship
with energy consumption. However, due to the limitation of correlation
as a descriptive analysis and its inability to determine the causal factors,
the study proceeds with inferential statistical analysis.
4.2. Unit root test
According to Perron (1989), unit roots and structural breaks have a
close relationship, thus, traditional unit root tests are “biased towards
false unit root null when the data are trend stationary with a structuralGHG HFICE INA MCS POP TFP
53,508 8,640,000,000 2,380,000,000 3,826,884 16,307,944 341,821
27,815 275,000,000 1,190,000,000 1071 15,689,386 355,305
130,473 75,500,000,000 12,900,000,000 30,360,771 26,786,598 495,683
19,454 193,500 252,000,000 0 8,827,273 219,327
40,160 17,400,000,000 3,080,000,000 8,255,112 5,399,093 76,964
0.6903 2.3418 2.0955 2.1368 0.3432 0.0209
1.6243 7.8792 6.5000 6.2290 1.9109 1.9782
6.6476 83.8612 54.6592 52.5978 3.0385 1.9173
0.0360a 0.0000a 0.0000a 0.0000a 0.2189 0.3834
53,508 8,640,000,000 2,380,000,000 3,826,884 16,307,944 341,821
1
0.7678 1
0.7129 0.7812 1
0.5265 0.8954 0.6507 1
0.7856 0.7992 0.9255 0.5744 1
0.6024 0.4403 0.7746 0.2030 0.8279 1
S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898 2891
Table 3
Zivot-Andrews and Vogelsang and Perron Breakpoint Unit Root Test.
Variable Zivot-Andrews Unit Root Test Vogelsang and Perron Breakpoint Unit Root Test
Level 1st difference Level 1st difference
BD 5% (−4.8/−4.42/−5.08) BD 5% (−4.8/−4.42/−5.08) Intercept Both Intercept Both
I/T/B Intercept Trend Both I/T/B Intercept Trend Both T-Statistic Prob T-Statistic Prob T-Statistic Prob T-Statistic Prob
AGM 2006/2001/2006 −476.149 −3.425 −402.779 2006/2007/2006 −6.974 −6.756 −10.506 −0.6269 ≥0.50 −2.3194 N0.99 −583.7652 b0.01 −576.6685 b0.01
EGIM 2007/2007/2007 −1.249 −3.268 −3.773 2008/2005/2010 −6.462 −6.332 −7.076 −1.5665 N0.99 −1.6612 N0.99 −6.4037 b0.01 −6.3733 b0.01
ENUS 2000/2007/2000 −4.39 −1.994 −3.817 2006/2004/2000 −6.446 −6.412 −8.276 −1.7003 N0.99 −2.1678 N0.99 −7.1421 b0.01 −7.2557 b0.01
EPTDL 2000/1994/2000 −7.62 −3.075 −6.848 2000/2002/2000 −8.253 −8.024 −9.369 −3.092 b0.10 −3.374 b0.10 −10.8199 b0.01 −10.6852 b0.01
FDIN 1978/1984/1977 −3.103 −3.582 −3.887 1987/1994/1987 −6.319 −5.744 −6.283 −1.0768 ≥0.50 −2.0194 ≥0.50 −5.692 b0.01 −6.6877 b0.01
GDP 1978/1987/1983 −3.157 −4.87 −5.133 1978/1979/1977 −6.896 −7.43 −7.899 −1.4878 ≥0.50 −0.7256 N0.90 −5.8905 b0.01 −6.3077 b0.01
GHG 1998/1986/1983 −16.443 −3.14 −4.359 2000/1999/1998 −9.799 −9.494 −10.216 −2.2368 N0.99 −4.2133 b0.025 −6.2801 b0.01 −5.6891 b0.01
HFICE 1978/1987/1998 −3.183 −4.378 −16.374 1978/1979/1977 −7.022 −7.452 −7.722 −1.7283 N0.41 −0.5719 N0.90 −5.9161 b0.01 −6.4708 b0.01
INA 2006/1982/1981 −3.808 −4.194 −4.183 1984/1987/1983 −5.838 −5.361 −6.16 −0.2445 ≥0.50 −3.4588 b0.05 −5.453 b0.01 −5.4624 b0.01
MCS 1992/1979/1992 −6.699 −2.07 −13.516 1992/1993/1992 −7.022 −5.897 −7.99 −5.7968 b0.01 −3.5356 b0.10 −32.4316 b0.01 −36.5975 b0.01
POP 1986/1991/1986 −7.423 −8.172 −7.505 1991/1980/1978 −2.454 −4.244 −4.1 −4.8497 N0.10 −3.1525 N0.80 −7.734 b0.01 −8.2349 b0.01
TFP 1986/1999/1996 −4.184 −4.184 −4.558 2000/1987/1988 −8.907 −8.91 −9.046 −4.3873 N0.05 −4.213 N0.25 −9.5847 b0.01 −9.3623 b0.01
NB: I = Intercept, T = Trend, B=Both (Intercept & Trend), BD = Break Date.break”. Hence, the study employs Vogelsang and Perron (1998) and
Zivot and Andrews (2002) unit root tests for structural breaks. Both
tests presented in Table 3 reveals that all the data series are stationary
atfirst difference, therefore, integrated of order one [I(1)]. Zivot and An-
drews unit root test in Fig. 2 reveals two prominent regimes after/be-
fore/between the break dates.
The structural break test provides a series of information that needs
to be examined. Agricultural machinery observed a minimum
breakpoint in 2006 due to a decline in the investment in tractors by
the Government, however, the trend changed during a regime change.
Industrialization observed aminimumbreakpoint in 2006 due to the
rise of the agricultural and services sectors' contribution to the GDP
therefore much attention was focused on the two. However, industrial-
ization took a rising turn at the commencement of mining crude oil in
2010, quadrupling the share of the industrial sector in Ghana's eco-
nomic growth (Ackah et al., 2014).
Total fisheries production experienced a minimum breakpoint in
1986due to a decline of outboardmotors that serve as an important fea-
ture in canoe fishing. The decline of outboardmotors was due to the ex-
ponential increase in the price of purchasing, running (price of premix
fuel) and maintenance (Wayo Seini, 1995).
The mobile industry began in 1992 in Ghana with over 100%mobile
penetration rate, thus, the reason whymobile cellular subscriptions ex-
perienced a minimum breakpoint in 1992 (AMGOO, 2014).
Household final consumption expenditure, GDP, and foreign direct
investment net inflows observed a minimum breakpoint in 1978 due
to the 1978 coup d'état that made Ghana ungovernable. This action
led to civil unrest, the dismissal of 1000 public employees, about eighty
strike actions taken by civil servants and a declaration of a state of emer-
gency, thus, affecting the country's economic policies including house-
hold income (Photius, 2011). In addition, the inflation rate increased
to over 92% per year in 1978, affecting both economic growth and for-
eign direct investment net inflows (Aryeetey et al., 2000).
The total greenhouse gas emissions had a minimum breakpoint in
1998, which was due to the introduction of the forestry development
master plan in 1996 that promotes sustainable forest harvesting. The
policymadeGhana a net sink due to the high echelons of carbon capture
sequestration in the land use, land-use change, and forestry sector. In
addition, there was a reduction in energy consumption in 1998 com-
pared to other years, hence, increasing the net greenhouse gas removals
(MLNR, 2012).
The net energy import observed a minimum breakpoint in 2007
due to an increase in Ghana's current account deficit from 6.9% in
2006 to 8.6% in 2007, resulting in a trade deficit caused by a higher
than expected oil import bill (OECD, 2008). As such, the import
cover declined in 2007. Ghana discovered its first indigenous oil
and gas in the Jubilee fields in 2007 which saved Ghana US
$1000,000 per day on the importation of fuel for thermal power gen-
eration (USAID, 2016).
Energy consumption observed a minimum breakpoint in 2000 due
to the poor energy sector policies and poor economic state, which
turned the country into a highly indebted poor country in 2001. This
propelled households to spend relatively small of their expenditure on
energy due to high electricity tariffs and prices of petroleum products
while salaries and wages remained stagnant (ESMAP, 2006). Electric
power transmission and distribution losses experienced a minimum
breakpoint in 2000 due to the introduction of a customer base of
about 120,000 by the Northern Electricity Department. The distribution
system consisted of 8000 km of sub-transmission lines, 22 bulk supply
points, 30,000 km distribution networks and 1800MVA installed trans-
former capacitywhichwas very efficient leading to a reduction in trans-
mission and distribution losses but however dwindled in the
subsequent years (ESMAP, 2006).
Theminimumbreakpoint in population occurred in 1986whichwas
due to a decline in fertility rate as a result of the earlier introduction of
contraceptives, an increase in under-5 infant mortality rate of 155 per
2892 S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898
Fig. 2. Structural break using Zivot-Andrews Unit Root Test.
Table 4
Linear regression analysis.
Source SS df MS Number of obs 44
F(11, 32) 15.82
Model 0.4586 11 0.0417 Prob N F 0.0000
Residual 0.0843 32 0.0026 R-squared 0.8447
Adj R-squared 0.7913
Total 0.5429 43 0.0126 Root MSE 0.0513
ENUS Coef. Std. Err. t P N t [95% Conf. Interval]
AGM 0.0130 0.0069 1.89 0.0670 −0.0010 0.0270
EGIM −0.0520 0.0247 −2.11 0.0430 −0.1023 −0.0017
EPTDL −0.0808 0.0180 −4.49 0.0000 −0.1174 −0.0442
FDIN 0.0353 0.0175 2.02 0.0510 −0.0002 0.0709
GDP 0.1391 0.2666 0.52 0.6060 −0.4040 0.6822
GHG 0.0499 0.0337 1.48 0.1480 −0.0186 0.1185
HFICE −0.0732 0.2549 −0.29 0.7760 −0.5924 0.4460
INA 0.0503 0.0451 1.11 0.2730 −0.0416 0.1422
MCS −0.0078 0.0063 −1.23 0.2260 −0.0207 0.0051
POP −0.9794 0.7216 −1.36 0.1840 −2.4492 0.4905
TFP 0.1014 0.0785 1.29 0.2050 −0.0584 0.2612
_cons 18.1719 10.6136 1.71 0.0970 −3.4474 39.79111000 live births and a decline in the proportion of married women
(UN, 2001).
The structural breaks have policy implications but requires a meth-
odology (revealed in the subsequent sections) capable of explaining
the two regimes in order to prevent speculative assumptions.
4.3. Asymmetric effect of predictors on energy consumption
Table 4 shows the linear regression analysis of the study. The linear
regression is employed as a baseline result to examine the impact of the
predictor variables on energy consumption without the switching ef-
fect. Table 4 reveals that only net energy imports and electric power
transmission and distribution losses are significant at 5% level. To test
the asymmetric effect of the predictor variables on energy consumption,
the study estimates the Markov-switching dynamic regression which
considers the switching effect. The results of the Markov-switching
dynamic regression in Table 5 shows statistical significance for all
the predictor variables. Essentially, the linear regression and the
Markov-switching dynamic regression have the same sign (i.e.
negative/positive) and significant at 5% level. Table 5 shows that
S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898 2893
Table 5
Markov-switching dynamic regression.
Variable Coef. Std. Err. z P N z
ENUS AGM 0.0097 0.0024 4.1000 0.0000
EGIM −0.0231 0.0086 −2.6900 0.0070
EPTDL −0.0815 0.0061 −13.3500 0.0000
FDIN 0.0347 0.0059 5.8800 0.0000
GDP 0.2973 0.0917 3.2400 0.0010
GHG 0.0336 0.0116 2.9000 0.0040
HFICE −0.2636 0.0875 −3.0100 0.0030
INA 0.0310 0.0153 2.0200 0.0430
MCS −0.0098 0.0022 −4.5100 0.0000
POP −0.4742 0.2412 −1.9700 0.0490
TFP 0.0541 0.0272 1.9900 0.0470
State 1 _cons 11.4730 3.5443 3.2400 0.0010
State 2 _cons 11.5625 3.5436 3.2600 0.0010
sigma 0.0175 0.0019agriculturalmachinery, foreign direct investment net inflows, economic
growth, total greenhouse gas emissions, industrialization and total fish-
eries production have a positive effect on energy consumptionwhile net
energy imports, electric power transmission, and distribution losses,
household final consumption expenditure, mobile cellular subscrip-
tions, and population have a negative effect on energy consumption.
In this study, state 1 is classified as “energy crisis” with high volatility
while state 2 is classified as “energy boom”with low volatility. The out-
put in Table 4 shows that energy consumption will grow by 11.6% dur-
ing the energy boom periods while growth will decline by 0.1% during
the period of energy crisis (11.5%). Table 6 presents the post-
estimation of Markov-switching dynamic regression. The expected du-
ration and the transition probability of entry into the two states are es-
timated. The output in Table 6 reveals that 11.6% growth in energy
consumption is expected to occur in 4.1 years while energy crisis is ex-
pected to last for 3.7 years. Further evidence shows a 73% probability
(p11) of remaining in 3.7 years of energy crisis while there is a 75%
chance of staying in 4.1 years of energy boom (p22). The probability
(p12) of switching from energy crisis to energy boom is 27%, and the
probability (p21) of changing from an energy boom to energy crisis is
25%. The predictive power of state 1 and state 2 on energy consumption
is examined as a sensitivity analysis presented in Fig. 3. We employ
MAPE and R-square as error and predictive power metrics. Fig. 3
(a) reveals that state 1 has a MAPE of 0.91% while state 2 has a MAPE
of 0.86% (Appendix A).We select state 2withminimumMAPE to exam-
ine its predictive power presented in Fig. 3(b). It is evident that the pre-
dictor variables in state 2 explain 81% (i.e. R2 = 0.81) of observed
dynamics in Energy consumption.
4.4. Neural network estimation
This section examines the predictive power of the neural network
model via a causal relationship between energy consumption, agricul-
tural machinery, foreign direct investment net inflows, economic
growth, total greenhouse gas emissions, industrialization, total fisheries
production, net energy imports, electric power transmission andTable 6
Post estimation of Markov-switching dynamic regression.
Post estimation Estimate Std. Err. [95% Conf. Interval]
Expected duration
State 1 3.7299 1.3439 2.0402 8.1644
State 2 4.0710 1.4814 2.1931 8.9044
Transition probabilities
p11 0.7319 0.0966 0.5099 0.8775
p12 0.2681 0.0966 0.1225 0.4901
p21 0.2456 0.0894 0.1123 0.4560
p22 0.7544 0.0894 0.5440 0.8877distribution losses, household final consumption expenditure, mobile
cellular subscriptions and population. Fig. 4 shows the exact input-
output diagramof themodel that employs one layer. The study employs
a TanHactivation function [i.e. TanH {(TanH=e2x− 1/e2x+1)where x
is a linear combination of the predictors} which is a sigmoid function
that converts a value between −1 and 1 (see Appendix A)] at the
nodes of the hidden layer. The resultant Predicted ENUS equation
using a one-layered feedforward network with three hidden nodes is
presented as:
PredENUS ¼ 5:5510þ−0:1194 : H11 þ−0:1140
: H12 þ 0:3825 : H13 ð19Þ
where, the output equations based on the hidden layer from the model
are expressed as:
H11 : TanHð0:5  ð112:3534þ−0:3939 : AGMþ−1:8115 : EGIM
þ−0:2193 : EPTDLþ−1:2334 : FDINþ 0:2788 : GDP
þ 1:2323 : GHGþ 0:9370 : HFICEþ 0:5512 : INA ð20Þ
þ−0:2972 : MCSþ−3:3136 : POPþ−7:7411 : TFPÞÞ
H12 : TanHð0:5  ð145:9786þ−0:0661 : AGMþ−2:2040 : EGIM
þ−0:7693 : EPTDL þ 0:1199 : FDINþ−0:2646 : GDP
þ−2:9475 : GHGþ−0:2758 : HFICEþ−1:9081 : INA
þ−0:4148 : MCSþ−4:8500 : POPþ 2:2950 : TFPÞÞ ð21Þ
H13 : TanHð0:5  ðð−160:7658Þ þ−0:0542 : AGMþ−0:4877 : EGIM
þ−0:4435 : EPTDL þ−0:0613 : FDINþ 0:7052 : GDP
þ−0:5345 : GHGþ−0:3905 : HFICEþ 0:2477 : INA
þ−0:9188 : MCSþ 10:2696 : POPþ−0:2103 : TFPÞÞ ð22Þ
Thus, H11, H12 and H13 are the hidden nodes from the one-layered
feedforward neural network model.
Fig. 5 presents the training and validation plots of the neural net-
work model while the corresponding goodness of fit estimates are
displayed in Table 7. Due to the flexibility of the neural network
model, there may arise problems related to overfitting data, which
will, in turn, predict future observations poorly. This temptation of
overfitting is prevented via the application of a “penalty” on the param-
eters of the model and the assessment of the predictive power of the
model using an independent dataset for validation. Fig. 5 reveals evi-
dence of 35 observations for the training set and 9 observations for
the validation set. This study employs the K-Fold validation method
advantaged in the efficient use of a small sample dataset and produces
accurate predictions. The output results in Table 7 shows that the train-
ing and validation set predict energy consumption at almost 99% R-
Squared value, 0.01 RSME value, 0.01 MAD value, 0.01 SSE value and
0.16% MAPE value (Appendix A). The predictive power and error met-
rics confirm the relationship between energy consumption and the pre-
dictor variables.
4.5. Determinants of energy consumption
To examine the determinants of energy consumption, the study em-
ploys the NIPALS regression analysis. NIPALS regression is a powerful
multivariate analysis that is superior to other conventional econometric
methods in terms of accuracy, predictability and variable projections.
The model allows for the imputation of missing data using their
means. The NIPALS regression analysis begins with the selection of a vi-
ablemodel using the number of factors available. Appendix B presents a
model comparison summary. The output in Appendix B shows that the
2894 S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898
Fig. 3. (a) Prediction of ENUS by states (b) Markov-switching dynamic regression model.model with the asterisk (*) is superior compared to the others, accord-
ingly, selected for further analysis. The selected model has a corre-
sponding number of factors in Appendix C required for coefficient
estimation. Appendix C shows that almost all the variables have VIP
greater than or equal to 0.8, a requirement for selecting important var-
iables (Samuel and Owusu, 2017). The variables can be classified into
two categories based on their VIP value; EPTDL, EGIM, and TFP are
highly important variables (VIP N 1) while the remaining predictor var-
iables aremoderately important variable thus, 0.8 ≤VIP ≤ 1. Now,we es-
timate the coefficient of the regression analysis using the 11 factors
presented in Appendix D and its corresponding output in Appendix B.
The NIPALS estimated coefficients in Table 8 have the same sign as
the linear regression (Table 4) and the Markov-switching dynamic re-
gression (Table 5) analysis, hence, confirming the validity of the coeffi-
cients in explaining the relationship between the response and the
predictor variables. The output in Table 8 reveals that a 1% increase in
agriculturalmachinery, foreign direct investment net inflows, economic
growth, total greenhouse gas emissions, industrialization, and totalfish-
eries production increases energy consumption by 0.35%, 0.47%, 5.00%,
0.32%, 0.47%, and 0.21%. In contrast, net energy imports, electric power
transmission and distribution losses, household final consumptionexpenditure, mobile cellular subscriptions, and population decreases
energy consumption by 0.30%, 0.65%, 2.62%, 0.49%, and 2.93%. The
next step is to examine the independence of the residuals and the pre-
dictive power of the NIPALS regression. Appendix E shows a diagnostic
plot of the residuals while Fig. 6 depicts the NIPALS Prediction Plot. The
residual normal quantile plot shows that the residuals of the model are
normally distributed.We examine the prediction error usingMAPE. The
errormetric reveals aMAPE of 0.64% (Appendix A). Fig. 6 shows that the
predictor variables can explain about 85% of the dynamics in energy
consumption.
TheMarkov-switching dynamic regression shows that linear regres-
sion is incapable of examining the relationship between energy con-
sumption and the predictors in the presence of structural breaks.
Using the linear regression as a baseline and the NIPALS regression as
a validation method for the study, the Markov-switching dynamic re-
gression showed the asymmetric effect of the predictors on energy con-
sumption. It was evident in Table 5 and Table 8 that agricultural
machinery, foreign direct investment net inflows, economic growth,
total greenhouse gas emissions, industrialization, and total fisheries
production have a positive effect on energy consumption. It is positive
in the sense of increasing the demand for energy consumption.
S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898 2895
Table 7
The goodness of fit estimation of the model.
ENUS Training Validation
Measures Value Value
RSquare 0.9881 0.9931
RMSE 0.0121 0.0093
MAD 0.0093 0.0081
-LogLikelihood −104.8593 −29.3045
SSE 0.0051 0.0008
Sum Freq 35 9
Fig. 4. One-layered feedforward neural network model for energy consumption.
Fig. 5.Model: (a) training and (b) validation.Agriculturalmachineries are dependent on fuel for its operations inme-
dium and large-scale agricultural production. Therefore, as large-scale
agricultural production enhances, agricultural machineries evolves,
hence, increasing energy demand (fossil fuels). Our study is supported
by Kuang et al. (2017), they found a relationship between agricultural
mechanization and energy consumption in China. Their study con-
cluded that energy consumption and efficient agricultural mechaniza-
tion are strongly linked with agricultural economic growth.
As of 2010, fisheries production accounted for 7.3% of the agricul-
tural GDP in Ghana (Asumadu and Owusu, 2016a). Thus, Government
invests more to boost the sustainable production through the provision
of subsidies on premix fuel. Fisheries production impact energy con-
sumption in the sense of premix fuels used to operate outboard motors
for canoe fishing. Tyedmers (2004) argues that energy in the form of
fuel is used directly in fishing vessels for vessel propulsion. In large-
scale fishing, high-intensity lamp batteries, onboard processing, auto-
mated jigging machines, refrigeration, freezing, vessel construction,
andmaintenance are all powered bydiesel-fueled generators, therefore,
increasing energy consumption (Tyedmers, 2004).
Contrary to the work of Adom (2015) that suggests a negative effect
of Foreign direct investment net inflows on energy intensity in South
Africa, Sadorsky (2010) confirms our results, thus, a positive effect of
Foreign direct investment net inflows on energy consumption
(Sadorsky, 2010). Foreign direct investment net inflows play a critical
role in Ghana's energy consumption. Legislation and structural frame-
works that provide an enabling environment to attract foreign investors
give Ghana a comparative advantage over other Sub-Saharan African
countries in terms of foreign investments (Sarkodie and Strezov,
2019). The multi-party democracy system, the rule of law, good gover-
nance (i.e. “first country to be reviewed under the Africa Peer Review
Mechanism of NEPAD”.), peaceful socio-political environment, out-
standing hospitality and personal security, the availability of export
free zone territory, access to tariff-free exports to the USA via the
Africa Growth and Opportunity Act, the vast ongoing oil exploration,
endowed natural resources, and among others entice foreign direct in-
vestment to the energy sector (CGA, 2017). For example, a compact
was signed in 2014 between Ghana and the Millennium ChallengeTable 8
Model coefficients for centred and scaled data.
Coefficient ENUS Plot
Intercept 0.0000
AGM 0.3579
EGIM −0.3045
EPTDL −0.6534
FDIN 0.4725
GDP 5.0147
GHG 0.3252
HFICE −2.6233
INA 0.4684
MCS −0.4919
POP −2.9376
TFP 0.2107
2896 S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898
Fig. 6. NIPALS Prediction Plot.Corporation for an investment of US$ 498.2 million to enable the reno-
vation of Ghana's energy sector (i.e. electricity production, distribution,
and accessibility to clean energy) and to motivate foreign and private
investment (USAID, 2017). Foreign direct investment net inflowsmedi-
ate in either the expansion of Ghana's energy sector orwidening the en-
ergy demand, which increases energy consumption. Sadorsky (2010)
argues that financial development is a sign of prosperity and economic
growth that strengthens customer and business confidence thereby in-
creasing the economic demand for energy-intensive goods.
Economic growth has a positive effect on energy consumption with
corresponding positive foreign direct investment net inflows. Devel-
oped countries with sustained and high economic growth rates can in-
stitute policies and measures that enable access to constant electricity
while developing and least developing countries stand a risk of energy
supply interruptions due to inconsistent economic growth, which
Ghana is not an exception (DiSano, 2002). The positive effect of the
total greenhouse gas emissions on energy consumption means that en-
vironmental policies and renewable energy policies are put into effect to
repeal and replace fossil fuel energy sources with green energy while
meeting the energy demand.
The positive effect of industrialization means it stimulates energy
consumption. Ghana's peak demand has been increasing due to loads
emanating from the industrial and commercial sectors. The upstream
activities of Ghana's petroleum sector involve the production and refin-
ing of crude oil and other petroleum products to meet the needs of the
industrial sector. For example, there has been a development of the
West Africa gas pipeline to feed industries with fuel (CGA, 2017;
Owusu and Sarkodie, 2016).
In contrast, Table 5 and Table 8 revealed a negative effect of energy
imports, electric power transmission, and distribution losses, household
final consumption expenditure, mobile cellular subscriptions, and pop-
ulation on energy consumption. The study expected a positive effect of
energy imports, household final consumption expenditure, mobile cel-
lular subscriptions, and population on energy consumption. However,
the results proved otherwise, but surprisingly, all the three regressions
have the same sign which means the models have no misspecification
problem. Theoretically, as energy imports, household final consumption
expenditure, population, and mobile cellular subscriptions increases, it
is expected that energy demand which translates into consumption
will increase. As household consumption expenditure increases,members of the household are able to afford energy-intensive goods
and services thus, consuming more energy. However, the negative ef-
fect of household consumption expenditure maybe due to energy effi-
ciency and conservation practices such as using green energy
technologies and services that reduce energy consumption. Increasing
population becomes burdensome on the peak energy demand, thus, in-
creasing energy demand and consumption. Social media like Facebook,
WhatsApp, Snap Chat, Instagram and among others propel the use of
mobile phones in Ghana. The use of 4G, 3G, and wireless networks in
mobile phones drain more battery compared to GSM, as such, the fre-
quency of charging mobile phones increases, thus, affecting electricity
demand. It is reported that base stations in cellular networks account
for over 50% of energy consumption in mobile phones (Han and
Ansari, 2013). It is in this regard that mobile cellular subscription is an
important factor in assessing the determinants of energy consumption.
Increasing levels of electric power transmission and distribution losses
reduce energy consumption, which affects energy efficiency and
conservation.
5. Conclusion
This study examined the asymmetric behavior of energy consump-
tion by assessing Ghana's energy sector dynamics. Using a time series
data from 1971 to 2014, the study employed Markov-switching dy-
namic regression to examine the asymmetric effect, NIPALS regression
to examine the determinants of energy consumption and neural net-
work analysis for prediction.
Using Vogelsang and Perron, and Zivot and Andrews unit root tests
for structural breaks, the study revealed some historical information
that has policy implications but required a methodology that could ex-
plain the two regimes exhibited in the structural break plots. While the
linear regression used as a baseline of the study had limitations, the
Markov-switching dynamic regression revealed the presence of asym-
metric effect in the presence of structural breaks.
It was evident in bothMarkov-switching andNIPALS regression that
agriculturalmachinery, foreign direct investment net inflows, economic
growth, total greenhouse gas emissions, industrialization, and totalfish-
eries production have a significant positive effect on energy consump-
tion. However, both models showed a significant negative effect of
energy imports, electric power transmission, and distribution losses,
S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898 2897household final consumption expenditure, mobile cellular subscrip-
tions, and population on energy consumption.
The study revealed that energy consumption evolves in two differ-
ent states (energy boom and energy scarcity) by transitioning over a fi-
nite set of unobserved states. Energy consumption is expected to grow
by 11.6% during the energy boom periods while growth will decline
by 0.1% during energy crisis (11.5%). The expected duration and the
transition probability of entry into the two states show that 11.6%
growth in energy consumption is expected to occur in 4.1 years while
energy crisis is expected to last for 3.7 years. There is 73% probability
of staying in energy crisis for 3.7 years while the chance of staying in
4.1 years of energy boom is 75%. The probability of switching from en-
ergy crisis to energy boom is 27%, while the probability of changing
from an energy boom to energy crisis is 25%.
The NIPALS regression revealed unobserved or unreported factors
like electric power transmission, and distribution losses, andmobile cel-
lular subscriptions underpinning energy crisis. There was evidence of
continuous shift in the regression from agricultural machinery, foreign
direct investment net inflows, economic growth, total greenhouse gas
emissions, industrialization, total fisheries production, net energy im-
ports, electric power transmission and distribution losses, household
final consumption expenditure, mobile cellular subscriptions, and pop-
ulation to energy consumption at a known point.
Even though the neural network is disadvantaged in terms of inter-
pretability yet revealed an accurate prediction of energy consumption
compared to theMarkov-switching dynamic and the NIPALS regression
models. The neural networkmodel suggests a 99% relationship between
energy consumption and the predictors.
As a policy implication, there is the need for improved and sustain-
able agricultural mechanization systems in Ghana to enhance the effi-
cient use of energy while increasing agricultural productivity. As
foreign direct investment net inflows propel technological advance-
ment and the development of green energy and energy efficiency, is es-
sential for the Government of Ghana to improve the renewable energy
policy that will provide more incentives and bolster foreign and private
investment into the energy sector. There is the need for sustained and
high economic growth rates in Ghana through the institution of finan-
cial policies and measures that increases financial development, which
will help the development and access to constant electricity thereby in-
creasing productivity.
Future research should aim at expanding the scope of the study to
include the role of research development and income inequality on en-
ergy consumption.
Acknowledgement
The usual Disclaimer applies to the final version of this paper. SA
Sarkodie is grateful to Macquarie University, Australia for the Interna-
tional Macquarie University Research Training Program (iMQRTP)
Scholarship.
Declaration
There is no conflict of interest.
Appendices A–E. Supplementary data
Supplementary data to this article can be found online at https://doi.
org/10.1016/j.scitotenv.2018.10.147.References
Ackah, C., Adjasi, C., Turkson, F., 2014. Scoping Study on the Evolution of Industry in
Ghana. WIDER Working Paper.
Adom, P.K., 2015. Determinants of energy intensity in South Africa: testing for structural ef-
fects in parameters. Energy 89, 334–346. https://doi.org/10.1016/j.energy.2015.05.125.Ahmed, K., Shahbaz, M., Kyophilavong, P., 2016. Revisiting the emissions-energy-trade
nexus: evidence from the newly industrializing countries. Environ. Sci. Pollut. Res.
23, 7676–7691.
AMGOO, 2014. Mobile industry in Africa: the changing face of mobile in Ghana. Retrieved
from. http://www.amgoo.com/blog/mobile-industry-in-africa-the-changing-face-of-
mobile-in-ghana.
Aryeetey, E., Harrigan, J., Nissanke, M., 2000. Economic Reforms in Ghana: The Miracle
and the Mirage. Africa World Press.
Asumadu, S.S., Owusu, P.A., 2016a. The relationship between carbon dioxide and agricul-
ture in Ghana: a comparison of VECM and ARDL model. Environ. Sci. Pollut. Res. Int.
23, 10968–10982. https://doi.org/10.1007/s11356-016-6252-x.
Asumadu, S.S., Owusu, P.A., 2016b. Carbon dioxide emissions, GDP, energy use, and pop-
ulation growth: a multivariate and causality analysis for Ghana, 1971-2013. Environ.
Sci. Pollut. Res. 23, 13508–13520. https://doi.org/10.1007/s11356-016-6511-x.
CGA, 2017. Investing in Ghana's energy sector. Retrieved from. http://www.
chinagoabroad.com/en/guide/investing-in-ghana-s-energy-sector.
Chang, S.-C., 2015. Effects of financial developments and income on energy consumption.
Int. Rev. Econ. Financ. 35, 28–44.
Destek, M.A., Sarkodie, S.A., 2019. Investigation of environmental Kuznets curve for eco-
logical footprint: the role of energy and financial development. Sci. Total Environ.
650, 2483–2489. https://doi.org/10.1016/j.scitotenv.2018.10.017.
DiSano, J., 2002. Indicators of Sustainable Development: Guidelines and Methodologies.
United Nations Department of Economic and Social Affairs, New York.
ESMAP, 2006. Ghana: sector reform and the pattern of the poor, energy use and supply.
Retrieved from. http://documents.worldbank.org/curated/en/
421071468252576846/pdf/435860WP0Box0327368B0PUBLIC1.pdf.
Faucheux, S., Nicolaï, I., 2011. IT for green and green IT: a proposed typology of eco-
innovation. Ecol. Econ. 70, 2020–2027.
Goldfeld, S.M., Quandt, R.E., 1973. A Markov model for switching regressions. J. Econ. 1,
3–15.
Hamdi, H., Sbia, R., Shahbaz, M., 2014. The nexus between electricity consumption and
economic growth in Bahrain. Econ. Model. 38, 227–237.
Hamilton, J.D., 1989. A new approach to the economic analysis of nonstationary time se-
ries and the business cycle. Econometrica: Journal of the Econometric Society
357–384.
Han, T., Ansari, N., 2013. On optimizing green energy utilization for cellular networks with
hybrid energy supplies. IEEE Trans. Wirel. Commun. 12, 3872–3882. https://doi.org/
10.1109/TCOMM.2013.051313.121249.
Inglesi-Lotz, R., Pouris, A., 2016. On the causality and determinants of energy and electric-
ity demand in South Africa: a review. Energ Source Part B 11, 626–636.
Jones, B.D., Kim, C.-J., Startz, R., 2010. Does congress realign or smoothly adjust? A discrete
switching model of congressional partisan regimes. Statistical Methodology 7,
254–276.
Kuang, Z., Yanbin, Q., Xin, D., 2017. Linking between agricultural mechanization, energy
consumption and economic growth. Journal of Agricultural Mechanization Research
3, 002.
Li,W., Ou, Q., Chen, Y., 2014. Decomposition of China's CO2 emissions from agriculture uti-
lizing an improved Kaya identity. Environ. Sci. Pollut. Res. Int. 21, 13000–13006.
https://doi.org/10.1007/s11356-014-3250-8.
Martínez-Beneito, M.A., Conesa, D., López-Quílez, A., López-Maside, A., 2008. Bayesian
Markov switching models for the early detection of influenza epidemics. Stat. Med.
27, 4455–4468.
MLNR, G, 2012. Climate Investment Funds Forest Investment Program.
Mohiuddin, O., Sarkodie, S.A., Obaidullah, M., 2016. The relationship between carbon di-
oxide emissions, energy consumption, and GDP: a recent evidence from Pakistan. Co-
gent Eng. 3, 1210491. https://doi.org/10.1080/23311916.2016.1210491.
Moyer, J.D., Hughes, B.B., 2012. ICTs: do they contribute to increased carbon emissions?
Technol. Forecast. Soc. Chang. 79, 919–931.
OECD, 2008. Ghana. Retrieved from. https://www.oecd.org/dev/emea/40577770.pdf.
Owusu, P., Asumadu, S.S., 2016. A review of renewable energy sources, sustainability is-
sues and climate change mitigation. Cogent Eng. 3, 1167990. https://doi.org/
10.1080/23311916.2016.1167990.
Owusu, P.A., Samuel, S.A., 2016. Multivariate co-integration analysis of the Kaya factors in
Ghana. Environ. Sci. Pollut. Res. Int. 23, 9934–9943. https://doi.org/10.1007/s11356-
016-6245-9.
Owusu, P., Sarkodie, S., 2016. A review of Ghana's energy sector national energy statistics
and policy framework. Cogent Eng. 3, 1155274. https://doi.org/10.1080/
23311916.2016.1155274.
Perron, P., 1989. The great crash, the oil price shock, and the unit root hypothesis.
Econometrica: Journal of the Econometric Society 1361–1401.
Photius, 2011. Ghana the Akuffo Coup, 1978. Retrieved from. http://www.photius.com/
countries/ghana/national_security/ghana_national_security_the_akuffo_coup_
197~150.html.
Quandt, R.E., 1972. A new approach to estimating switching regressions. J. Am. Stat. Assoc.
67, 306–310.
Sadorsky, P., 2010. The impact of financial development on energy consumption in
emerging economies. Energy Policy 38, 2528–2535. https://doi.org/10.1016/j.
enpol.2009.12.048.
Salahuddin, M., Alam, K., 2015. Internet usage, electricity consumption and economic
growth in Australia: a time series evidence. Telematics Inform. 32, 862–878.
Salahuddin, M., Alam, K., 2016. Information and communication technology, electricity
consumption and economic growth in OECD countries: a panel data analysis. Int.
J. Electr. Power 76, 185–193.
Samuel, A.S., Owusu, P.A., 2017. The impact of energy, agriculture, macroeconomic and
human-induced indicators on environmental pollution: evidence from Ghana. Envi-
ron. Sci. Pollut. Res. Int. 24, 6622–6633. https://doi.org/10.1007/s11356-016-8321-6.
2898 S.A. Sarkodie et al. / Science of the Total Environment 651 (2019) 2886–2898Sarkodie, S.A., Adom, P.K., 2018. Determinants of energy consumption in Kenya: a NIPALS
approach. Energy 159, 696–705. https://doi.org/10.1016/j.energy.2018.06.195.
Sarkodie, S.A., Owusu, P.A., 2016. Carbon dioxide emissions, GDP per capita, industrializa-
tion and population: an evidence from Rwanda. Environ. Eng. Res. 22, 116–124.
https://doi.org/10.4491/eer.2016.097.
Sarkodie, S.A., Strezov, V., 2018. Assessment of contribution of Australia's energy production
to CO2 emissions and environmental degradation using statistical dynamic approach.
Sci. Total Environ. 639, 888–899. https://doi.org/10.1016/j.scitotenv.2018.05.204.
Sarkodie, S.A., Strezov, V., 2019. Effect of foreign direct investments, economic develop-
ment and energy consumption on greenhouse gas emissions in developing countries.
Sci. Total Environ. 646, 862–871. https://doi.org/10.1016/j.scitotenv.2018.07.365.
Tyedmers, P., 2004. Fisheries and energy use. Encyclopedia of Energy 2, 683–693.
UN, 2001. Africa, in sub-Saharan, workshop on prospects for fertility decline in high fer-
tility countries. Retrieved from. http://www.un.org/esa/population/publications/
prospectsdecline/makinwa.pdf.
United Nations, 2015. Sustainable development goals. Retrieved from. https://
sustainabledevelopment.un.org/?menu=1300.
USAID, 2016. Ushering in Ghana's gas revolution. Retrieved from. https://www.usaid.gov/
power-africa/newsletter/june2016/ushering-in-ghana.
USAID, 2017. Power Africa: investment brief for the electricity sector in Ghana. Retrieved
from. https://www.usaid.gov/sites/default/files/documents/1860/Ghana%20_IG_
2015_05_03.pdf.Van Heddeghem, W., Lambert, S., Lannoo, B., Colle, D., Pickavet, M., Demeester, P., 2014.
Trends in worldwide ICT electricity consumption from 2007 to 2012. Comput.
Commun. 50, 64–76.
Vogelsang, T.J., Perron, P., 1998. Additional tests for a unit root allowing for a break in the
trend function at an unknown time. Int. Econ. Rev. 1073–1100.
Wayo Seini, A., 1995. Economics of Marine Canoe Fisheries in Ghana.
Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics.
Chemom. Intell. Lab. Syst. 58, 109–130.
World Bank, 2016.World development indicators. Retrieved from. http://data.worldbank.
org/country.
Zhang, C., Lin, Y., 2012. Panel estimation for urbanization, energy consumption and CO2
emissions: a regional analysis in China. Energy Policy 49, 488–498.
Zhu, H., Su, X., You, W., Ren, Y., 2017. Asymmetric effects of oil price shocks on stock
returns: evidence from a two-stage Markov regime-switching approach. Appl. Econ.
49, 2491–2507.
Zivot, E., Andrews, D.W.K., 2002. Further evidence on the great crash, the oil-price shock,
and the unit-root hypothesis. J. Bus. Econ. Stat. 20, 25–44.
Zou, X.X., Li, Y., Li, K., Cremades, R., Gao, Q.Z., Wan, Y.F., Qin, X.B., 2015. Greenhouse gas
emissions from agricultural irrigation in China. Mitig. Adapt. Strateg. Glob. Chang.
20, 295–315. https://doi.org/10.1007/s11027-013-9492-9.