Bachelor Thesis Economics and Business Specialization: Economics and Finance
Faculty of Economics and Business Academic year: 2016 – 2017
Electricity consumption as a predictor of economic growth:
instrumental regression on dynamic panel data of 82 countries
Name:
Stefan van Schaik
Student number: 6118984
Supervisor:
Ron van Maurik
2 Statement of originality
This document is written by Stefan van Schaik, who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
3 Abstract
This thesis investigates the effect of electricity consumption growth on economic growth by making use of two stage least squares (TSLS) regression. A panel data consisting of 82 countries for the period 1972-2013 is constructed. Instrumental variables are used to counter simultaneous causality between electricity use growth and economic activity growth. The instrumental variables used are the ratio of tax revenue and GDP, a factor which controls for extreme average temperatures, the urbanization grade, net energy import, and the first lag of electricity consumption growth. Control variables used are population growth, the unemployment rate, and capital formation growth. Endogenous regressors are the first lag of GDP growth and electricity consumption growth. Also, Granger causality is tested for on simple ALD models. The results indicate a significant effect of electricity consumption growth on GDP growth, while Granger causality runs from GDP growth to electricity consumption growth, but not vice versa. No prove for a simultaneous causal relationship is found.
4 Table of contents Abstract 3 1. Introduction 5 2. Literature review 6 3. Data 9 4. Methodology 12 5. Results 15 6. Conclusion 19 7. References 21 8. Appendix 23
5
1. Introduction
Energy consumption, especially electricity consumption, differs from most other consumption goods in that it can be conceived as both a necessity good as a luxury good. People need a minimum of electricity for running their household and the level of electricity consumption increases when income rises. The more income one receives, the more luxury goods one is able to purchase. More importantly, with higher wages, people are able to afford bigger houses to live in, which in turn almost always lead to more use of energy. This disparate purpose of consuming makes electricity consumption an interesting subject for economic research.
In addition to changes in demand for electricity of middle- and high-income countries, it is also worth paying attention to low-income country tendencies in electricity consumption. While the question for middle- and high-income countries remains at what level electricity consumption develops, for low-income countries it is more relevant looking to the size of the population with actual access to electricity. For instance, according to data of the World Bank (2016) only 5.06% of the population in South Sudan had access to electricity in 2012. However, when looking to the world population, the percentage of people having access to electricity increased from 75.65% in 1990 to 84.58% in 2012. Since in this period the world population grew by 34.2%, this is an even more impressive result, which demonstrates the increasing dependency of electricity for low-income countries, too.
Another trend in the energy sector is the expanding attention to energy production from
renewable sources. Although coordinating a global policy of reducing carbon emissions seems difficult to arrange, considering extensive negotiations during climate conferences, it is of increasing interest
developing and innovating less polluting ways of energy consumption, hence reducing the reliance on fossil fuels. The most common sources of renewable energy are wind power, hydropower, solar energy, geothermal energy, bio energy, and energy storage. The first and foremost market trying to make these ambitions reality is the automobile sector. Making advanced hybrid, plug-in hybrid, and electric vehicles available for the mass should result in fewer dependency on exhaustible resources like oil and gas in the near future. Data of the U.S. Energy Information Administration (2016, p. 10) show that renewable energy is the world’s fastest-growing source of energy, at an average rate of 2.6% per year. It is also forecasted that electricity production in the period 2012 to 2040 will increase by 1.9% per year on average. This means electricity generation will increase by a total of 54.2% in the coming 22 years.
Ferguson, Wilkinson and Hill (2000, p. 934) where among the firsts to study the relationship between electricity consumption and economic development. Their findings show a strong correlation between electricity use and wealth creation. Moreover, the correlation between electricity use and wealth creation is stronger than the correlation between total energy use and wealth. However, the presence of a strong correlation does not necessarily imply a causal relationship. Not only could electricity production growth result in higher economic output, it is also reasonable to think of economic growth being an important factor of electricity demand growth. The problem measuring the relation between these
6
variables is the possibility of simultaneous causality. Indeed, understanding the causal relationship between electricity consumption and economic growth is helpful in defining and implementing environmental and energy policies. For example, the results should be beneficial in determining appropriate rates of energy taxes.
This thesis therefore examines the relationship between electricity consumption growth and economic growth. An unbalanced panel data is constructed covering 42 years for 82 countries. While the majority of other studies about this subject makes use of cointegration techniques and vector
autoregression (VAR) models, this article tries demonstrating the relationship by applying instrumental regression methods. The model being used is also tested with country and time fixed effects. By this way variables that are constant across countries but evolve over time and vice versa are controlled for.
The remainder of this thesis is structured as follows. The next section contains a literature study and evaluates multiple results of other relevant articles. In Sector 3 the data collection process is discussed and descriptive statistics are included. Section 4 introduces the empirical methodology of this thesis and explains the used econometric models. In section 5 the findings of this research are presented. Finally, section 6 concludes and summarizes the results of this thesis and is followed by attachments containing this article’s references, tables, figures, and used regression output.
2. Literature review
Numerous studies have been undertaken to assess the ambiguous relationship between electricity consumption growth and economic development. However, it was only until the last decade of the previous century the scientific research about electricity consumption as a determinant of economic growth did expand. This section explains relevant developments in economical energy research in the last decades. First, primary scientific research is described. Next, because oil price fluctuations have been regarded as an important determinant of economic growth, this relationship will be briefly outlined. Oil price fluctuations may be seen as adjacent to electricity consumption, because electricity prices are influenced by the price of crude oil and both are inputs of energy. In the last part of this section the attention is moved to electricity consumption as a distinct source of economic growth. Other articles with different econometrical methods are discussed.
Prior to the study of Ferguson, Wilkinson and Hill (2000), which is mentioned in the previous section, Ferguson et al. (1997, p. 258) already define a strong correlation between electricity use and wealth creation, while not finding a significant relationship between total energy use and wealth. This result is surprising, since the World Bank introduced the ratio gross domestic product (GDP) per tonne of oil equivalent (toe) as a meaningful indicator of a nation’s level of development. A tonne of oil
equivalent is a unit of energy defined as the amount of energy released by burning one tonne of crude oil. According to this article it appears that electricity, rather than energy, represents a strong impetus for economic growth.
7
However initially, scientists paid more attention finding out the effects on economic activity of energy consumption in general than the use of electricity in particular. For example, Beenstock and Willcocks (1981, p. 226) examine income- and price-elasticities of demand for energy. They demonstrate that the income-elasticity of demand for energy is significantly greater than unity, while the price-elasticity is close to zero, being highly inelastic. These findings show that energy demand is hardly affected by the price paid for it, though the level of income is of considerable influence on energy demand.
The research of causality between energy consumption and economic growth did commence with the work of Kraft and Kraft (1978). In this study the authors find a causal relationship running from gross national product (GNP) to energy input in the United States. No causality is detected running from energy consumption to GNP. The relationship is later examined by others using different samples and with mixed results. For instance, Yu and Hwang (1984, p. 188) do not find any causal relationship between the two variables, while Stern (1993, p. 148) argues that energy consumption is at least a modest determinant of economic growth.
Before scientists shifted part of their focus to electricity consumption as an indicator of
economic activity, oil prices were being considered as an important indicator of a nation’s future welfare. For instance, Hamilton (1983, p. 246) discovered a systematic relationship between oil prices and
economic output in the postwar period 1948 to 1972. He generalized the argument by stating that energy prices play a vital role in macroeconomic performance after 1960. In the context of these findings, the research of Hooker (1996, p. 206) is of high interest. In this article, he finds Granger causality from oil price to real GDP and from oil price to the rate of unemployment for the period 1948 to 1973, but not for the period 1973 to 1994. The data is split into two groups, while the OPEC embargo of 1973 is used as the breaking point. Possible explanations of the difference in result of the two subsamples are that oil prices have become endogenous to the United States economy and that the used linear vector
autoregression equations misspecify the nature of the interaction between the oil price and the economy. In a more recent article it is stated that the consequences of oil price changes depend on whether a country is a net importer or exporter of oil (Jiménez-Rodríguez & Sánchez, 2005, p. 224). By making use of vector autoregression techniques it follows that real GDP growth of oil importing economies is negatively affected by increases in oil prices, while for oil exporting economies the effect of an increase of the oil price is inconclusive. For example, the GDP of Norway is positively affected by an oil price increase, while for the United Kingdom a higher oil price has a significant negative impact on economic development.
Also, lot of attention is placed on the symmetry of the relationship of oil prices and economic activity. Mork (1989, p. 740) criticizes Hamilton’s research for using a sample with solely rising oil prices. In addition to this article, Mork finds an insignificant effect of oil price decreases on real GNP growth. Mory (1993, p. 159) gives a similar statement. He concludes that an increase of the oil price has a powerful adverse outcome on GNP, while oil price declines do not influence GNP positively, nor negatively.
8
The main difference between oil and electricity as a source of commercial energy is that oil can be stored and transported without much technological difficulties, while electricity storage is much more expensive and harder to accomplish (Kousksou et al., 2014, p. 67). Due to the high cost of electricity storage and little capacity on the electrical grid, only as much electricity is generated as is required. When consumption need for oil is below expectation, the excess of oil can be stored for future use, while once generated, electricity is transmitted directly through the distribution system to end users. Furthermore, renewable energy sources depend heavily on the kind of weather and if it is day- or nighttime. This has implications for the supply and demand of electricity. Ibrahim, Ilinca and Perron (2008, p. 60) state that because of the absence of a regular supply of electricity generated by renewable sources, the supply cannot be adjusted to changes in consumer demand. Hence, supply of electricity roughly meets demand and electricity consumption growth reflects the growth of production at the equilibrium level.
Electricity is an essential element in an economy as a whole, both from the perspective of producers as well as from a consumer’s angle, and its importance is increasing thanks to advances in renewable energy technology. Therefore, the subject has been of great interest for research the last two decades. Various studies have been done exploring the exact nature of the relationship between electricity and economic activity. Two main econometric methods that have been used thus far are cointegration analysis using vector error correction models and Granger causality tests. Two variables are cointegrated when they share a common stochastic trend (Granger, 1981, p. 128). Cointegration entails the existence of a causal link, but no conclusions can be made about the direction of this relationship. Granger causality verifies if past values of one variable contains information that significantly helps forecasting a second variable, compared to predicting the second variable by its past values alone (Granger, 1969, p. 431). Although the definition suggests testing for undeniable causality, it merely demonstrates one variable’s ability in predicting the second variable. Causality still has to be deducted with the use of economic theory.
Both approaches have been applied in multiple studies recent years, while different samples and time periods were being used. Some research focused on a single nation. Ghosh (2002, p. 128) for example, studies the relationship of electricity consumption and economic growth in India. He observes Granger causality running from GDP to electricity consumption, while not finding evidence for Granger causality the opposite way. This implies energy conservation policies by national governments may continue without harming economic activity. This is contrary to the findings of Altinay and Karagol. While not finding a causal relationship between aggregated energy use and GDP growth in a preliminary study (Altinay & Karagol, 2004, p. 993), they do find Granger causality running from electricity
consumption to real GDP, making use of Turkish macroeconomic data (Altinay & Karagol, 2005, p. 855). Shiu and Lam (2004, p. 52) investigate the case for China. They detect electricity consumption and real GDP being cointegrated, while Granger causality running from electricity consumption to real GDP but not vice versa. Polemis and Dagoumas (2013, p. 807) however, find bidirectional Granger causality for Greece. A similar result is found for South-Korea (Yoo, 2005, p. 1631).
9
Because different conclusions are derived from causality tests based on individual countries, as well as from different time period within the same country, some research is done trying to get more generalized inferences by making use of longitudinal data. Lee and Chang expand the research by
including sixteen Asian nations. They identify energy consumption Granger causing GDP in the long run, while not finding evidence for the reverse to be true, nor do they find short term effects (Lee & Chang, 2009, p. 63). Ciarreta and Zarraga concentrate on a more interconnected energy network by looking at the European continent. They detect a negative short-run relationship from energy consumption to GDP (Ciarreta & Zarraga, 2010, p. 3795). An explanation for this result is the existence of capacity constraints of electricity generation. Perhaps the most comprehensive research to date is done by Karanfil and Li (2015, p. 276). Out of a panel data of 160 countries, they construct several subsamples, based on the nations’ geographic region and income classification. They define different results for the subsamples, hence expressing their concern that the conclusions made are highly sensitive to regional characteristics. Therefore, one has to be careful when describing the extent to which outcomes can be generalized.
Apparently, contrasting results remain the issue for this research subject, both when single country time series are used and when information of multiple nations is combined. That is why this research is trying to get more insight of the matter by adopting a whole different approach, namely instrumental regression.
3. Data
In this section the data collection process is described and the variables used in the models are defined and further explained. Furthermore, summary statistics are described.
A panel data set is constructed out of 82 countries for the period 1972 to 2013, which comprises 42 years. The countries used are listed in table 3 in the appendix. The complete panel data set is
unbalanced, since for several variables data are missing. Unfortunately, complete data for some variables are not found. Especially a large number of information is missing of the unemployment rate and capital formation growth. Still, the data set used for the regression models consists of 1,631 useful cross-sectional observations.
The benefits of using panel data instead of cross-sections and time series are listed by Baltagi (2008, p. 6). One of these advantages is that panel data control for individual heterogeneity. For example, the ratio of people having religious beliefs may vary among countries, while being invariant across time. For example, when studying the effect of some independent variables on alcohol use, it is likely that very religious people limit their alcohol use or fully refrain from it. If no variables explaining the degree of religion are included in the model, then bias of the estimation results arises. Panel data are able to control for these country- and time-invariant variables, whereas cross-sectional or time series studies cannot. Another advantage using a panel data approach is that the possibility of multicollinearity decreases,
10
because panel data have more informative data, variability, and degrees of freedom. One of the main drawbacks of using panel data is the difficult process of data collection.
Except the information used to compute the temperature factor, data is collected from the online database of the World Bank (2016). The World Bank makes use of data of the International Energy Agency (IEA), the International Monetary Fund (IMF), the Organization for Economic Co-operation and Development (OECD), the International Labour Organization (ILO), and the United Nations Statistics Division (UNSD). Temperature data is retrieved from a data set constructed by Harris et al. (2014), on behalf of the Climactic Research Unit (CRU) at the University of East Anglia.
Two variables are adjusted, since they otherwise could not be tested for in the model. Electricity consumption growth is computed by dividing the difference of two consecutive years in electricity consumption by the electricity consumption in the first year. The variable electricity consumption measures the production of power plants and combined heat and power plants less transmission, distribution, and transformation losses and own use by heat and power plants. Furthermore, the temperature factor is computed by taking the square of the difference between the average annual temperature and fifteen, divided by ten. The average annual temperature is expressed in degrees Celsius. The reasoning behind the temperature factor is that departures from 15°C lead to higher electricity use, either by using more electric heating for low annual temperature countries, or by using more air
conditioning and ventilation systems for high annual temperature countries. The scale of the temperature factor roughly goes from 0 to 5, while the majority of the countries have a value between 0 and 1. This is shown in table 4 in the appendix. Hence, by using a temperature factor as a regressor on electricity consumption, extreme temperature countries having an extra positive impact on electricity consumption is assumed.
The control variables included in the model are population growth, unemployment rate, and capital formation growth. Population growth is put into the model, because higher growth rates are associated with lower economic growth levels (Solow, 1956, p. 76). The population growth rate is
expressed as a percentage. Also, the unemployment rate is used. A high unemployment rate may stimulate productivity of employees, while it affects economic activity by earning less income. A low
unemployment rate may cause workers to shirk on their job. Therefore, a socially optimal unemployment rate exists (Shapiro & Stiglitz, 1984, p. 440). The unemployment rate refers to the share of the labor force that is without work, but available for and seeking employment and is expressed as a percentage. It is an indicator of the overall development in a country. Capital formation growth indicates net investment or equivalently the saving rate multiplied by economic output. Increases of the saving rate lead to higher equilibrium growth of economic activity (Solow, 1956, p. 76). Capital formation growth is expressed as a percentage.
The instrumental variables included in the model are the total of tax collected compared to GDP, the aforementioned temperature factor, the urbanization rate, the ratio energy imports against total energy use, and the first lag of electricity consumption growth. Apart from the temperature factor, these variables
11
are expressed as percentages. The instrumental variables correlate with the endogenous variable electricity consumption growth, while they do not correlate with the dependent variable GDP growth. The tax-to-GDP ratio is used as an instrumental variable, because it is assumed that a higher tax-to-tax-to-GDP ratio indicate higher energy taxes, which may lead to less electricity consumption. The urbanization rate is a degree of concentration of the population. It is hypothesized that if more people live in urban areas, then the easier it is developing and maintaining a high-quality energy infrastructure, which leads to higher accessibility to electricity. The higher net energy imports, the more dependent a nation is on other countries for its energy. This increases the vulnerability of energy supply, which has a negative impact on electricity use on a macroeconomic level. A negative value indicates the country being a net exporter of energy. Also, the first lag of electricity consumption growth is added as an instrument.
Summary statistics of GDP and electricity consumption variables are provided in table 4, which is included in the appendix. The minimum value of GDP per capita of $ 76 belongs to Nepal in 1973, while the maximum value is $ 113,727, which belongs to Luxembourg in 2013. Mean GDP per capita is $ 9,094. Also, Nepal has the lowest value of average annual electricity consumption. Therefore, it is not surprising Nepal has the lowest urbanization rates in the data set, too. In addition to minimum values of GDP per capita, the low electricity consumption is explained by Nepal’s challenging landscape. Due to this, both energy and transport infrastructure are underdeveloped and poorly maintained. In contrast, Iceland has the highest value of electricity consumption in 2013. This number is explained by the country’s aluminum industry, which accounts for approximately 70% of Iceland’s total energy use (Orkustofnun, 2017). The mean of electricity consumption is 3,287 kWh, while the average value for the Netherlands is 6,821 kWh in 2013. Both GDP and electricity consumption per capita are displayed in figure 1 for the period 1971 to 2015. The figure makes clear that electricity consumption increases at a fairly constant rate, while GDP per capita grows faster, but with more fluctuation. This is in accordance with the results in table 4, in which is shown that GDP per capita has a higher standard deviation than electricity consumption per capita.
Both the minimum and maximum values of GDP growth are achieved by Iraq during the Gulf War, which lasted from August 1990 to February 1991. Figure 1 in the appendix shows a line graph of GDP growth and electricity consumption growth of the world population. This figure confirms the higher variance of GDP compared to electricity consumption. Moreover, it appears from the graph that both growth rates move in the same direction. This strengthens the hypothesis that electricity use and economic development are cointegrated. Figure 4 in the appendix shows the scatter plot of GDP growth and electricity consumption growth. The scatter plot shows the significant correlation of GDP growth and electricity consumption growth, r(3430) = .2798, p < .001.
Moreover, table 4 shows descriptive statistics of the control variables and instrumental variables of the model being used. The world’s population growth and unemployment rate are graphed in figure 5. Population growth is gradually declining from 2% to 1%, while the unemployment rate fluctuates around 6%. The highest unemployment rate belongs to Iraq in 2003, which suffered from the invasion and
12
occupation initiated by the United States. Further, in figure 6 the line graphs of the tax-to-GDP ratio and urbanization rate are plotted. The graph makes clear that the tax-to-GDP ratio is rather constant over time, whereas the urbanization rate is increasing at a linear rate during the period 1970 to 2015.
4. Methodology
This section contains the model setup. In addition, the thread of simultaneous causality is explained. Furthermore, several tests are done validating the models being used.
To analyze the effect of electricity consumption growth on GDP growth, the following model is constructed and tested for:
GDPgrowth𝑖,𝑡= 𝛽0+ 𝛽1GDPgrowth𝑖,𝑡−1+ 𝛽2elecgrowth𝑖,𝑡+ 𝛽3popgrowth𝑖,𝑡+ 𝛽4unempl𝑖,𝑡+
GDPgrowth𝑖,𝑡=𝛽5capformgrowth𝑖,𝑡+ 𝑢𝑖,𝑡 (1)
The subscripts i and t respectively denote the specific country and year of the different variables. The
intercept 𝛽0 represents a constant intercept. The first lag of GDP growth is added as an endogenous
variable, since last year’s GDP is correlated with this year’s GDP. Hence, the model contains of two endogenous variables, three exogenous variables, and five instrumental variables. Robust standard errors are used.
First the model is tested with ordinary least squares (OLS) regression and robust standard errors. However, when testing for an effect of electricity use on GDP, OLS regression cannot be applied, because it leads to inconsistent estimates. This is due the fact that there is simultaneous causality. A causal effect runs both from electricity consumption to GDP and from GDP to electricity consumption, as the results in section 2 demonstrate. Figures 2, 3 and 4 in the appendix support the assumption this is indeed the case. When regressing GDP growth on electricity consumption growth a significant relationship is shown, β = .1721, t(3430) = 17.07, p < .001. Simultaneous causality has implications for the internal
validity of the model and indicates electricity consumption is correlated with the error term. The error term represents omitted factors that determine GDP growth. Hence, electricity consumption growth is an endogenous variable. If endogeneity exists, then the error term does not have conditional mean zero.
To counter the simultaneous causality issue, instrumental variables (IV) regression is applied, again with robust standard errors. Multiple instrumental variables are used to predict electricity consumption growth. The reasons and logic behind the selection of the instrumental variables are mentioned in the previous section. The instrumental variables used to predict electricity consumption growth, are the tax-to-GDP ratio, the temperature factor, the urbanization rate, the energy imports ratio and the first lag of electricity consumption growth. A valid instrumental variable must satisfy the
instrument relevance condition and the instrument exogeneity condition. An instrument is relevant when the variance of the instrument correlates with the variance of the endogenous variable. An instrument is
13
exogenous if the variance of the instrument is uncorrelated with the error term of the regression model. If these two conditions are satisfied, then the instrument used, captures that part of the variance of the endogenous variable that is exogenous to the model. Therefore, the coefficient of the endogenous variable can consistently be estimated by IV regression.
Various tests are done verifying consistency of the model being used. First, electricity
consumption growth and the GDP growth lag are tested for endogeneity with two Hausman-Durbin-Wu tests. First, one of the endogenous variables is regressed on all exogenous and instrumental variables of the model. Next, the residuals of this regression are computed. Finally, these residuals are added to equation (1). Under the null hypothesis these values have no effect on GDP growth and OLS regression is consistent, while under the alternative hypothesis the tested coefficient is significantly different from zero, which justifies the use of IV regression. In the appendix, the output of the Hausman-Durbin-Wu tests is shown. The coefficient of the residuals while testing GDP growth lag significantly differs from zero, β = -.1499, t(1630) = -2.50, p = .013. When testing for electricity consumption growth, the same
conclusion is made, β = -.2880, t(1630) = -4.23, p < .001. This proves that both the first lag of GDP and
electricity consumption growth are endogenous variables.
Testing instruments for instrumental relevance is complex when having multiple endogenous regressors. To test whether the used instruments are relevant with one endogenous variable, an F-test is
used. First, the endogenous variable is regressed on all exogenous and instrumental variables. Then, the
F-statistic is tested against the null hypothesis that all instruments are zero. As a rule of thumb, the
instruments used in the model are of sufficient strength if the F-statistic exceeds ten (Stock and Watson,
2014, p. 490). The computed F-statistic indicates that the instrumental relevance condition is satisfied
with only the first lag of GDP growth as an endogenous regressor, F(5,1622) = 62.48, p < .001 and with
only GDP lag as an endogenous variable, F(5,1622) = 12.49, p < .001. The output of the F-tests for
instrumental relevance can be found in the appendix.
If the instruments are not exogenous, then the IV estimators are inconsistent. This would imply the estimators converge to values other than the true population parameters. That is why the instrumental variables must be exogenous. To test for exogeneity, a J-test is performed. First, IV regression is applied
on equation (1). Next, the residuals are calculated. Subsequently, these values are regressed on all exogenous and instrumental variables using OLS. Now, the F-statistic of the null hypothesis that all
instruments have a coefficient of zero is calculated. The J-statistic can be determined by multiplying the
number of instruments to the derived F-statistic. The J-statistic of the used model is equal to 1.15. Under
the null hypothesis that all the instruments are exogenous, the J-statistic is chi-squared distributed with
the number of instruments minus the number of endogenous regressors as the number of degrees of freedom. This implies that the p-value of the J-statistic is equal to .76, which is greater than the 5%
significance level. Thus, the null hypothesis is not rejected and instrumental exogeneity is verified for. The statistical output used for the J-test is provided in the appendix.
14
The regression equation makes use of panel data which variables are observed in two dimensions. These dimensions are country and year. Hence, each value in the data set has a unique combination of three characteristics: the country the value relates to, the year the value is effectuated and the model’s variable that the value presents. However, it is possible there are variables affecting GDP growth, which are not included in the model. Because of the possible existence of omitted variable bias, the model is also tested with entity and time fixed effects. By this way, it is possible to eliminate the effect of omitted variables that differ across entities but are constant over time and omitted variables that differ over time but are constant across countries. For example, culture differs greatly among countries in the panel data, while having great influence on a number of macroeconomic factors. Cultural change is a long-lasting process which happens incrementally. Hence, it is a country fixed effect and time-invariant. Since culture’s quantifiability and measurability is complicated it is excluded from the model. An example of an entity-invariant time fixed effect is scientific knowledge. The amount or quality of all scientific
information evolves over time and can be used anytime by every country, in contrast to information in patents. One can argue that scientific knowledge is an important determinant of industrial development (Narin, Hamilton & Olivastro, 1997, p. 330). Nonetheless, scientific knowledge is disregarded in the model. The combined use of country fixed effects and time fixed effects eliminates any omitted variables bias arising from both kind of unobserved variables. The following equations represent the fixed effects regression models:
GDPgrowth𝑖,𝑡= 𝛼𝑖+ 𝛽1GDPgrowth𝑖,𝑡−1+ 𝛽2elecgrowth𝑖,𝑡+ 𝛽3popgrowth𝑖,𝑡+ 𝛽4unempl𝑖,𝑡+
GDPgrowth𝑖,𝑡=𝛽5capformgrowth𝑖,𝑡+ 𝑢𝑖,𝑡 (2)
GDPgrowth𝑖,𝑡= 𝜆𝑡+ 𝛽1GDPgrowth𝑖,𝑡−1+ 𝛽2elecgrowth𝑖,𝑡+ 𝛽3popgrowth𝑖,𝑡+ 𝛽4unempl𝑖,𝑡+
GDPgrowth𝑖,𝑡=𝛽5capformgrowth𝑖,𝑡+ 𝑢𝑖,𝑡 (3)
GDPgrowth𝑖,𝑡= 𝛼𝑖+ 𝜆𝑡+ 𝛽1GDPgrowth𝑖,𝑡−1+ 𝛽2elecgrowth𝑖,𝑡+ 𝛽3popgrowth𝑖,𝑡+ 𝛽4unempl𝑖,𝑡+
GDPgrowth𝑖,𝑡=𝛽5capformgrowth𝑖,𝑡+ 𝑢𝑖,𝑡 (4)
The intercept 𝛼𝑖 is the country fixed effect, while the intercept 𝜆𝑡 is the time fixed effect. The intercepts are multiplied by a dummy variable 𝐷𝑖 or 𝐷𝑡 respectively representing a country or year. Then, the constants 𝛼1+ ⋯ + 𝛼𝑛 are all country fixed effects, while the constants 𝜆1+ ⋯ 𝜆𝑇 are all time fixed effects. By including entity and time fixed effects in the regression model, respectively for every particular country and year a different intercept is estimated. Still, the estimated slope coefficients of the regression model are the same for each country and year. Robust standard errors are used.
Finally, just like multiple other studies mentioned in section 2, the data is tested for Granger causality running from GDP growth to electricity consumption growth and running from electricity consumption growth to GDP growth. First, both variables are tested for stationarity by using the
Im-15
Pesaran-Shin test. Stationarity is crucial, since if future values happen to be different from historical values, then past values cannot be used to forecast forthcoming values. The Im-Pesaran-Shin is used, because the test is also applicable to unbalanced and dynamic panels (Im, Pesaran & Shin, 2003, p. 65). The test is a standardized t-bar test statistic based on the augmented Dickey–Fuller statistics averaged
across the groups. The null hypothesis that all panels are nonstationary is rejected, since the p-value of the
𝑍𝑡̃-bar-statistic of both GDP growth and electricity consumption growth is essentially zero. The output of the Im-Pesaran-Shin unit root tests can be found in the appendix.
Since both variables are at least partially stationary, OLS estimations are done on simple autoregressive distributed lag (ADL) models:
GDPgrowth𝑖,𝑡= 𝛽0+ 𝛽1GDPgrowth𝑖,𝑡−1+ 𝛽2elecgrowth𝑖,𝑡−1+ 𝑢𝑖,𝑡 (5)
elecgrowth𝑖,𝑡= 𝛽0+ 𝛽1GDPgrowth𝑖,𝑡−1+ 𝛽2elecgrowth𝑖,𝑡−1+ 𝑢𝑖,𝑡 (6)
If Granger causality is running from electricity consumption growth to GDP growth, then the estimated coefficient of the lagged value of electricity consumption growth is significantly different from zero. Likewise, if Granger causality is running from GDP growth to electricity consumption growth, then the estimated coefficient of the first lag of GDP growth is significantly different from zero. Again, the existence of Granger causality does not imply strict economic causality, merely it shows the ability of one variable predicting another.
5. Results
In this section the results of the performed regressions are shown, analyzed, and linked with the existing literature as mentioned in section 2.
In table 1 the regression results are listed. Column (1) shows the outcome of the model estimated by OLS. However, since both the first lag of GDP growth and electricity consumption growth are endogenous variables, these estimations are biased. Both variables are correlated with the error term in the model, which implies the crucial assumption that the conditional mean of the error term is equal to zero is not satisfied. By performing IV regression, that part of the endogenous variables that correlates with the error term is isolated and that part that does not correlate is captured and used for estimating the dependent variable. The next four columns of table 2 present IV estimations. In panel data, it is possible to control for omitted variables that differ from one country to the next but remain constant over time and for variables that are the same across countries but vary over time. To do so, one has to perform regression with entity and time fixed effects. Column (3) shows the estimation results of regression with country fixed effects, column (4) presents the estimated coefficients of the time fixed effects regression, while the results in column (5) are controlled for both entity and time fixed effects. The estimations of the
16 Table 1 Regression analysis
(1) (2) (3) (4) (5)
GDP growth lag 0.3229*** (11.45) 0.3559*** (3.43) 0.3303*** (3.41) 0.3667*** (2.92) 0.2007* (1.67) Electricity consumption growth 0.1853*** (7.80) 0.3007*** (3.67) 0.1661*** (3.69) 0.3080*** (3.50) 0.0676 (1.16) Population growth -0.3930*** (-4.98) -0.4159*** (-4.08) -0.7978*** (-3.95) -0.4228*** (-3.69) -0.8499*** (-4.18) Unemployment rate -0.0296** (-2.12) -0.0180 (-1.00) -0.0773* (-1.68) -0.0202 (-1.10) -0.1288** (-2.42) Capital formation growth 0.0271* (1.83) 0.0236 (1.49) 0.0255 (1.46) 0.0197 (1.47) 0.0209 (1.32)
R-squared 0.3286 -0.2748 0.3575 -0.3456 0.4513
Type of regression OLS IV IV IV IV
Country fixed effects No No Yes No Yes
Time fixed effects No No No Yes Yes
Note: t-values are given in parentheses under the coefficients; * = significant at the 10% level; ** = significant at the 5% level, *** = significant at
the 1% level.
OLS regression are highly significant, but these results are biased. However, the sign of all coefficients is in line with the hypotheses in section 3. GDP growth of the current year is positively related to GDP growth of the previous year. Furthermore, it is in accordance with Solow’s findings (1956, p. 76) that the relationship between both population growth and the unemployment rate are negatively related to GDP growth. Capital formation growth has a small significant effect on economic activity.
When comparing the IV regression results to the estimated coefficients of the OLS regression, it is worth noting that the signs of the estimated parameters do not change. Although t-values of the IV
regressions are generally less than their OLS counterparts, significance does only slightly decrease. The significance of GDP growth lag, electricity consumption growth, and population growth are still highly significant. However, the unemployment rate and capital formation growth parameters are significant under OLS, while this is not the case under IV estimation. A possible cause is that the OLS regression had 2,040 useful cross-sectional observations, while the IV models only had 1,631 observations. This is due the fact that the IV regression used the tax-to-GDP ratio as an instrument predicting the values of electricity consumption growth, while information about the ratio is limited in the dataset. Only data are used of a country for a particular year, if all of the explanatory and instrumental variables are known. If one of these variables is missing, then the whole cross-section is left out of the regression. Because the tax-to-GDP ratio is more often unknown in poorer and non-European countries, the data used for the IV regression is somewhat biased compared to the data used by the OLS regression.
17
Although sample bias may exist, the estimated coefficients of GDP growth lag, electricity
consumption growth, and population growth are quite alike when comparing the OLS regression with the IV regression without fixed effects. The estimated coefficients of the endogenous variables, especially electricity consumption growth, are higher in the IV model. This is not in accordance with the expected direction of the simultaneity bias. When electricity consumption growth causes GDP growth and vice versa, then the following equations represent this relationship:
GDPgrowth = 𝛽1elecgrowth + 𝑢1 (7)
elecgrowth = 𝛽2GDPgrowth + 𝑢2 (8)
For simplicity reasons, no further exogenous variables are put in the equations. Consequently, the simultaneity bias is equal to the covariance of electricity consumption growth and error term u1
(Wooldridge, 2008, p. 552): cov(elecgrowth, 𝑢1) =1−𝛽𝛽2 1𝛽2𝐸(𝑢1 2) = 𝛽2 1−𝛽1𝛽2𝜎1 2 (9)
Since GDP growth and electricity consumption growth are positively correlated, both β1 and β2 are
between 0 and 1. This indicates positive covariance, which implies the direction of the bias must be upward. When using IV regression instead of OLS regression, this technique ought to be removing this bias. However, since the results in table 1 suggest downward bias, the rising coefficients of the IV regression do not point out simultaneity bias. Instead, the rising coefficients indicate attenuation bias. This bias arises when some measurement error is included in the model. This could be caused by using an inaccurate set of instruments. If some measurement error is included in an independent variable, then this would lead to a biased OLS estimate towards zero (Wooldridge, 2008, p. 320). Therefore, despite a significant effect of electricity consumption growth on GDP growth in the IV model, no causal effect without error is detected.
The explained variance of the IV model is slightly less compared to the OLS model. This is no real issue though. When an explanatory variable is correlated with the error term, the variance of the independent variable cannot be divided into explained sum of squares and residual sum of squares. However, the R-squared is still showed, because it shows how much the fit decreases when applying IV regression.
When analyzing differences in result between the IV regression without fixed effects and the IV models with country fixed effects and time fixed effects, there are a few dissimilarities. The original IV model and the IV regression with time fixed effects are quite alike. This suggests no or few country-invariant variables are omitted in the model. Yet, when country fixed effects are added, the effect of electricity consumption growth on economic activity diminishes, while retaining significance. In addition,
18
the negative effect of population growth on economic growth gets stronger. Apparently, time-invariant variables not included in the model and having distinctive effects on GDP per country exist. When these variables are controlled for, some influence of electricity consumption growth on GDP growth is captured by population growth. Perhaps this has to do with the selection of instrumental variables. The instruments used to predict electricity use growth may have not properly included country-specific elements that influence GDP growth. For instance, possibly the weather effect on electricity consumption is better estimated by both a country’s average number of sun hours per day and freezing days per year than by its annual temperature. Still, the effect of electricity consumption growth on GDP growth is significant when country fixed effects are included. Also, the unemployment rate coefficient increases and gets significant.
The significance of electricity consumption growth vanishes completely when including both country and time fixed effects in the model. A possible explanation for this is that too little variation in electricity consumption growth is left to accurately explain GDP growth effects. This model completely ignores between-country variation and focuses only on within-country variation. Also, time trends are disregarded. However, the estimated coefficient of the unemployment rate gets significant at the 5% level, while being negative. This implies that higher values of the unemployment rate reduce GDP growth. The sign is in accordance with Okun’s Law, which states that a fall of the unemployment rate by 1% leads GNP to rise by 2 to 3% (Prachowny, 1993, p. 331). Shapiro and Stiglitz (1984, p. 440) propose the existence of a socially optimal rate of employment. In the estimated IV models, only linear relations were investigated. To check for a possible non-linear relationship of unemployment rate on GDP and to check for robustness of the model, the square unemployment rate is added to equation (1), while performing IV regression on this. In the appendix, the output of this regression is shown. None of the coefficients in column (2) of table 1 change much, except for the non-significant coefficient of the unemployment rate. Also, the significance of the estimated parameters is not influenced by adding an extra variable to the model.
Apart from the biased OLS regression, the results in table 1 show no significant effect of capital formation growth on economic activity. This is surprising, since capital formation growth is an indicator of changes in the saving rate. While Solow (1956, p. 76) states that increases in the saving rate have a positive influence on GDP growth, no such relation is found here.
To further check robustness of the model, another IV regression is performed. This time, countries beginning with the letters A or B are eliminated from the data. That means data of Algeria, Argentina, Australia, Austria, Belgium, Bangladesh, Bolivia, and Brazil are excluded. The output of this regression can be found in the appendix. The conclusions regarding the model’s robustness are mixed. The effect of previous GDP growth decreases, while estimated coefficients of electricity consumption growth, population growth, and the unemployment rate increase. Moreover, the unemployment rate gets significant. However, in addition to the increase in significance of the individual parameters, the explained variance of the model diminishes considerably. Although the overall difference in estimated coefficients is
19 Table 2 Granger causality analysis
Direction Lag Coefficient*** t-statistic p-value
Electricity consumption growth → GDP growth 1 0.0093*** 0.89 0.375
GDP growth → Electricity consumption growth 1 0.1133*** 3.98 <0.001
Note: * = significant at the 10% level; ** = significant at the 5% level, *** = significant at the 1% level.
moderate, it is concluded that reducing the number of observations does have some implications for the estimations.
In table 2 the results of the OLS regressions for determining Granger causality are shown. According to these results, Granger causality is running from GDP growth to electricity consumption growth. Contrary to the research of several authors (Shiu & Lam, 2004; Altinay & Karagol, 2005; Yoo, 2005; Polemis & Dagoumas, 2013), no Granger causality is running from electricity consumption growth to GDP growth. These findings suggest no existence of a simultaneous causal relationship, which is in line with the findings of the instrumental regressions. Thus, higher economic growth leads to a higher rate of electricity use, while electricity consumption has no significant effect on a nation’s economic activity.
6. Conclusion
In this section the overall process of this thesis is discussed, while the main results are listed and briefly explained.
Multiple research has been conducted aiming to identify the relationship between economic activity and energy use. Since electricity’s role in daily life is getting increasingly integral both for
producers and consumers, the past few years more focus is shifted to electricity as an energy determinant of economic growth. However, scientific studies to date have not found a conclusive answer on the causal direction of the relationship between electricity consumption and economic output.
This thesis investigates the relation of GDP growth and electricity consumption growth by making use of instrumental regression. This is an unusual approach, for most of the existing literature primarily makes use of cointegration and vector autoregression techniques. Also, while the majority of the research papers focus on time series of a single country, for this thesis an extensive panel dataset is constructed, consisting of 82 countries and 42 years. By including a lag of GDP growth into the model, the panel data becomes dynamic, which makes estimation of the parameters more complex. In addition to IV regression, the existence of Granger causality is studied by performing OLS regression on simple ALD models with one lag.
The results of the IV regressions indicate a significant effect of electricity consumption growth on economic activity. Compared to the OLS estimation results, the significant coefficients of the IV regression were less in magnitude. However, when simultaneity bias exists, the bias would lead to a
20
downward direction of the IV estimations. Since the contrary is found, the results suggest no simultaneity bias exist. The upward bias of the explanatory variables is possibly caused by sample selection bias, attenuation bias, or measurement errors. Further research is needed for identifying the exact nature of this issue.
However, when studying the relationship between electricity consumption growth and GDP growth, one has to take into account the remarks by Karanfil and Li (2015, p. 276). When constructing a panel data, conclusions have to be drawn with care. Unobserved geographical and cultural differences may influence the analysis. That is why country fixed effects are added to the model. While doing this, the magnitude of the effect of electricity consumption growth on GDP growth decreases. When time fixed effects are added to the model, the estimated coefficients are similar to the original model. This indicates the existence of omitted time-invariant variables.
Moreover, Granger causality is analyzed by performing OLS regression on simple ADL models with one lag. The results indicate Granger causality running from economic growth to electricity use, but no Granger causality running from electricity consumption to economic activity. In this case, GDP growth precedes electricity consumption growth. Hence, one should not be deterred by government policies focusing on conservative energy. Raising taxes on energy or adopting other policies that
discourage energy use does not necessarily harm economic growth. On the contrary, to sustain economic growth decision-makers should pursue energy policies designed for stimulating more efficient energy consumption.
Future research should aim at constructing a more balanced panel data, because this increases the range of different estimation techniques. Also, having an equal amount of observations for both OLS and IV regression would eliminate the sample selection bias possibly occurred during this research. Moreover, the set of instruments should be carefully reconsidered. For instance, to estimate the weather effect on electricity consumption, the temperature factor may be replaced by two distinct variables: one variable for low temperature countries and one for high temperature countries. Examples of possible variables are the number of freezing days and the number of sunny days per year, respectively. However, finding complete data of suitable instruments may be the hardest aspect in future research.
21 References
Altinay, G. & Karagol, E. (2004). Structural break, unit root, and the causality between energy consumption and GDP in Turkey. Energy Economics, 26(6), 985-994.
Altinay, G. & Karagol, E. (2005). Electricity consumption and economic growth: Evidence from Turkey.
Energy Economics, 27(6), 849-856
Baltagi, B. (2008), Econometric Analysis of Panel Data. New York: John Wiley & Sons (fourth edition).
Beenstock, M. & Willcocks, P. (1981). Energy consumption and economic activity in industrialized countries: The dynamic aggregate time series relationship. Energy Economics, 3(4), 225-232.
Ciarreta, A. & Zarraga, A. (2010). Economic growth-electricity consumption causality in 12 European countries: A dynamic panel data approach. Energy Policy, 38(7), 3790-3796.
Ferguson, R., Hill, R., Craggs, C. & Forbes, I. (1997). Benefits of electricity generation. Engineering Science and Education Journal, 6(6), 255-259.
Ferguson, R., Wilkinson, W. & Hill, R. (2000). Electricity use and economic development. Energy Policy, 28(13), 923-934.
Ghosh, S. (2002). Electricity consumption and economic growth in India. Energy Policy, 30(2), 125-129.
Granger, C. (1969). Investigating Causal Relations by Econometric Models and Cross-Spectral Methods.
Econometrica, 37(3), 424-438.
Granger, C. (1981). Some properties of time series data and their use in econometric model specification.
Journal of Econometrics, 16(1), 121-130.
Hamilton, J. (1983). Oil and the Macroeconomy since World War II. Journal of Political Economy, 91(2), 228-
248.
Harris, I., Jones, P., Osborn, T. & Lister, D. (2014). Updated high-resolution grids of monthly climatic observations—the CRU TS3.10 dataset. International Journal of Climatology, 34(3): 623-642.
Hooker, M. (1996). What happened to the oil price-macroeconomy relationship? Journal of Monetary Economics, 38(2), 195-213.
Ibrahim, H., Ilinca, A. & Perron, J. (2008). Energy storage systems—Characteristics and comparisons. Renewable and Sustainable Energy Reviews, 12(5), 1221-1250.
Im, K., Pesaran, M. & Shin, Y. (2003). Testing for unit roots in heterogeneous panels. Journal of Econometrics, 115(1), 53-74.
Jiménez-Rodríguez, R. & Sánchez, M. (2005). Oil price shock and real GDP growth: empirical evidence for some OECD countries, Applied Economics, 37(2), 201-228.
Kraft, J. & Kraft, A. (1978). On the relationship between energy and GNP. Journal of Energy and Development, 3(2), 401-403.
Karanfil, F. & Li, Y. (2015). Electricity consumption and economic growth: Exploring panel-specific differences. Energy Policy, 82, 264-277.
Kousksou, T., Bruel, P., Jamil, A., El Rhafiki, T. & Zeraouli, Y. (2014). Energy storage: Applications and challenges. Solar Energy Materials and Solar Cells, 120(A), 59-80.
Lee, C. & Chang, C. (2008). Energy consumption and economic growth in Asian economies: A more comprehensive analysis using panel data. Resource and Energy Economics, 30(1), 50-65.
Mork, K. (1989). Oil and the Macroeconomy When Prices Go Up and Down: An Extension of Hamilton’s Results. Journal of Political Economy, 97(3), 740-744.
Mory, J. (1993). Oil Prices and Economic Activity: Is the Relationship Symmetric? The Energy Journal, 14(4), 151-161.
Narin, F., Hamilton, K. & Olivastro, D. (1997). The increasing linkage between U.S. technology and public science. Research Policy, 26(3), 317-330.
Orkustofnun (2017). Electricity Generation. Available at: http://www.nea.is/geothermal/electricity- generation (Accessed on 23 January 2017).
Polemis, M. & Dagoumas, S. (2013). The electricity consumption and economic growth nexus: Evidence from Greece. Energy Policy, 62, 798-808.
Prachowny, M. (1993). Okun’s Law: Theoretical Foundations and Revised Estimates. The Review of Economics and Statistics, 75(2), 331-336.
Shapiro, C. & Stiglitz, J. (1984). Equilibrium Unemployment as a Worker Discipline Device. The American Economic Review, 74(3), 433-444.
Shiu, A. & Lam, P. (2004). Electricity consumption and economic growth in China. Energy Policy, 32(1),
22
Solow, R. (1956). A Contribution to the Theory of Economic Growth. The Quarterly Journal of Economics, 70(1), 65-94.
Stern, D. (1993). Energy and economic growth in the USA: A multivariate approach. Energy Economics, 15(2), 137-150.
Stock, J. & Watson, M. (2014). Introduction to Econometrics. New Jersey: Pearson Education (third updated
edition).
U.S. Energy Information Agency (2016). International Energy Outlook 2016. Available at: http://www.eia.gov/outlooks/ieo/pdf/0484(2016).pdf (Accessed on 12 January 2017). Wooldridge, J. (2008). Introductory Econometrics: A Modern Approach. Boston: Cengage Learning (fourth
edition).
World Bank (2016). World Development Indicators & Global Development Finance Databank. Available at: http:// databank.worldbank.org (Accessed on 22 December 2016).
Yoo, S. (2005). Electricity consumption and economic growth: evidence from Korea. Energy Policy, 33(12),
1627-1632.
Yu, E. & Hwang, B. (1984). The relationship between energy and GNP: Further results. Energy Economics, 6(3), 186-190.
23 Appendix
Table 3
List of countries per continent
Africa Algeria, Cameroon, Congo-Brazzaville, DR Congo, Cote d'Ivoire, Egypt, Gabon, Ghana, Kenya, Morocco, Nigeria, Sudan, Senegal, South Africa, Tunisia, Zambia, Zimbabwe
Asia Bangladesh, China, Hong Kong, India, Indonesia, Iran, Iraq, Israel, Japan, Kuwait, Myanmar, Malaysia, Nepal, Oman, Pakistan, Philippines, Saudi Arabia, Singapore, South Korea, Sri
Lanka, Syria, Thailand, Turkey
Australia Australia
Europe Austria, Belgium, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Italy, Luxembourg, Malta, Netherlands, Norway, Portugal, Spain, Sweden, United Kingdom
North America Canada, Costa Rica, Cuba, Dominican Republic, El Salvador, Guatemala, Honduras, Jamaica, Mexico, Nicaragua, Panama, Trinidad and Tobago, United States
South America Argentina, Bolivia, Brazil, Chile, Colombia, Ecuador, Paraguay, Peru, Uruguay, Venezuela
Table 4
Summary statistics selected variables
Observations Mean deviation Standard Minimum Maximum
GDP per capita 3,360 9,094 14,039 76.225 113,727
Electricity consumption per capita 3,441 3,287 5,019 5.803 54,799
Main model variables
GDP growth 3,432 1.993 5.026 -64.997 53.933
Electricity consumption growth 3,440 3.988 8.224 -56.030 103.162
Population growth 3,440 1.731 1.152 -3.339 9.932
Unemployment rate 2,285 7.599 4.850 0 29.9
Capital formation growth 2,856 5.551 29.751 -376.223 1,058.283
Instrumental variables
Tax revenue / GDP 2,243 17.040 8.128 0.086 95.161
Temperature factor 3,444 0.766 0.619 0 5.018
Urbanization rate 3,444 58.939 23.114 4.198 100
Energy imports / total energy use 3,444 -79.126 554.141 -16,718.71 100
Note: GDP per capita is denominated in U.S. dollars; electricity consumption per capita is denominated in kWh; GDP growth, electricity
consumption growth, and capital formation growth are denominated in percentage changes; unemployment rate, tax revenue / GDP, urbanization rate, and energy imports / total energy use are denominated as a percentage.
24 Figure 1
GDP per capita and electricity consumption per capita of the world population
Figure 2
25 Figure 3
GDP growth and electricity consumption growth of the world population
Figure 4
26 Figure 5
Population growth and the unemployment rate of the world population
Figure 6
27 Figure 7
Stata output Hausman-Durbin-Wu test on the first lag of GDP growth
Figure 8
28 Figure 9
29 Figure 10
30 Figure 11
Stata output J-test on instrumental variables
31 Figure 12
Stata output Im-Pesaran-Shin test on GDP growth
Figure 13
32 Figure 14 Stata output OLS regression
Figure 15 Stata output IV regression
33 Figure 16
35 Figure 17
36 Figure 18
39 Figure 19
Stata output Granger causality OLS regression running from electricity consumption growth to GDP growth
Figure 20
40 Figure 21
Robustness check: square of unemployment added to the model
Figure 22
