**3. European Union**

4 Renewable Energy – Trends and Applications

and economic growth has achieved mixed results. For twenty OECD countries, Apergis & Payne (2010) estimated a panel vector error correction model, and found bidirectional causality between RE and economic growth, both in the short and long run. A bidirectional causality between RE and economic growth was also detected by Apergis *et al*. (2010), for a group of 19 developed and developing countries. For the US, Menyah & Wolde-Rufael (2010) found only a unidirectional causality running from GDP to RE. Conversely, when it comes to analysing the relationship between RE and economic growth, the empirical literature is thin. Menegaki (2011) is one of the exceptions, studying the situation in Europe. Indeed, focused on 27 EU Members, using panel error correction, the author did not confirm the presence of Granger causality running from RE to economic growth, either in the short or long run. These results lead the author to conclude that the consumption of RE makes a minor contribution to GDP. In fact, it seems that the nature of the relationship between RE

The literature on the empirical link between restraining emissions of carbon dioxide (CO2) and economic growth has shown some unexpected results. Menyah & Wolde-Rufael (2010) only found unidirectional causality running from CO2 to RE. In the same way, Apergis *et al*. (2010) conclude that the consumption of RE does not contribute to reducing CO2 emissions. Their explanation is grounded in the well-known problem of storing energy, as well as the intermittency characteristic of renewables. The failure to store energy, for example from wind or solar sources, requires the simultaneous use of established sources of energy, such as natural gas or even the highly polluting coal. This scenario leads to two effects on the installed capacity and on energy dependency. On the one hand, it implies the maintenance and even the enlargement of productive capacity that becomes idle for long periods, which generates economic inefficiencies. On the other hand, the intermittency may not even contribute to the reduction of a country's energy dependence goal, such as documented by

The root of the lack of consensus in literature, with regard to the relationship between RE and economic growth, could come from different theoretical and practical perspectives. On the one hand, it is admissible that the effect of RE on economic growth could vary largely according to both the geographical area and the time span analysed. On the other hand, there could be a variable omission bias problem. In fact, the research may be disregarding the importance of other variables, such as the simultaneous consumption of oil, coal, nuclear or natural gas. These variable omissions could lead to wrong conclusions on causality between each energy source and economic growth, when analysed separately. The cost of

Under the well-known premise that energy plays a crucial role in the economic growth process, the question that arises is what will the particular role of renewable sources of energy be on economic growth? To find the answer to this question, as stated before, possible bias resulting from the omission of variables must be avoided, and it is necessary to assess the simultaneous explanatory power of the main sources of energy driving economic growth.

The problem of GDP growth analysis has a long path in economic literature. The mainstream does not sustain any relationship between economic growth and its volatility. Nevertheless, the relationship between economic growth volatility and the trend in growth has been the object of increasing attention in literature. Indeed, macroeconomists have long

this is that inconsistent and erroneous results may be achieved.

**2.2 Volatility and economic growth** 

and economic growth still has a long way to go before consensus is achieved.

Frondel *et al*. (2010).

For a long time now, the EU has taken the lead in the fight against climate change. As stated before, one of the tools that the EU has used is to set targets for the use of RE, in each of the EU Member States. Some of the important milestones along the way have been the White Paper for a Community Strategy and Action Plan, Energy for the Future: Renewable Sources of Energy", in December, 1997 and the EU Directives 2001/77/EC and 2009/28/EC. In parallel with concerns about climate change, concerns are also emerging in Europe about economic growth, which has been generally modest.

#### **3.1 Current picture**

The commitment to renewables made by the EU has been translated into real achievements with regard to the contribution of these sources to the energy supply. For the period 1990- 2007, and for the EU of 27, we looked at the picture of the evolution of GDP growth rates, as well as the evolution of the contribution of renewable sources to total energy supply (CRES), as a percentage.

Fig. 1. Economic growth and renewable energy use in EU27

Figure 1 suggests that the rising trend in the use of RE is contemporaneous with different behaviours of economic growth. The periods of greatest growth in the use of renewable sources were simultaneous with contractions in economic growth. This gap is clearly visible in the mid-1990s and the 2000s, and the CRES variable clearly accelerates during 2000s.

When we analyse this reality in detail, we find that data is missing for some of the 27 countries, in particular with regard to variables related to the use of other energy sources. Thus, the EU Members for which the data is available, for all the variables considered and for the time span under review, are: Austria, Belgium, the Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Luxembourg, the Netherlands, Poland, Portugal, the Slovak Republic, Slovenia, Spain, Sweden, and the United Kingdom. For the time span 1990-2007, this chapter is focused on this panel of 21 EU Members. For this panel, we made a first inspection by country, into the relationship between the growth in renewables' use and economic growth. To do so, we calculated both the average rate of the growth rates of contribution of renewables to total energy supply, and the average rate of economic growth in the period 1991-2007. This information is summarised in Figure 2.

In general, it was observed that the highest rates of growth in the use of RE are associated with countries with lower economic growth. The highest average rate of economic growth during this period was found for Ireland. For this country, the rate of growth in the use of renewables (4.22%) was markedly lower than that rate of economic growth (6.48%). The highest rates of growth in the use of renewable sources in this period are usually associated with countries that have shown lower economic growth rates. Estonia, the Slovak Republic and the Czech Republic have the highest average growth rates of the use of renewables (nearly 12.1%, 11.3% and 9.5%, respectively), but they have low rates of economic growth (2.82%, 2.88% and 1.99%, respectively). Note that the average economic growth rate is 2.76% and the average growth of use of renewables is 5.17%. Germany has one of the highest growth rates of renewables (8.7%) during this period, but its average economic growth was only 1.58%.

Fig. 2. Economic growth and renewable energy use

#### **3.2 Variables**

6 Renewable Energy – Trends and Applications

parallel with concerns about climate change, concerns are also emerging in Europe about

The commitment to renewables made by the EU has been translated into real achievements with regard to the contribution of these sources to the energy supply. For the period 1990- 2007, and for the EU of 27, we looked at the picture of the evolution of GDP growth rates, as well as the evolution of the contribution of renewable sources to total energy supply (CRES),

economic growth, which has been generally modest.

1990

Fig. 1. Economic growth and renewable energy use in EU27

Year

1991

1992

1993

1994

1995

1996

1997

1998

Figure 1 suggests that the rising trend in the use of RE is contemporaneous with different behaviours of economic growth. The periods of greatest growth in the use of renewable sources were simultaneous with contractions in economic growth. This gap is clearly visible in the mid-1990s and the 2000s, and the CRES variable clearly accelerates during 2000s. When we analyse this reality in detail, we find that data is missing for some of the 27 countries, in particular with regard to variables related to the use of other energy sources. Thus, the EU Members for which the data is available, for all the variables considered and for the time span under review, are: Austria, Belgium, the Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Luxembourg, the Netherlands, Poland, Portugal, the Slovak Republic, Slovenia, Spain, Sweden, and the United Kingdom. For the time span 1990-2007, this chapter is focused on this panel of 21 EU Members. For this panel, we made a first inspection by country, into the relationship between the growth in renewables' use and economic growth. To do so, we calculated both the average rate of the growth rates of contribution of renewables to total energy supply, and the average rate of economic growth in the period 1991-2007. This information is

1999

2000

2001

2002

2003

2004

2005

2006

2007

%

**3.1 Current picture** 

as a percentage.

CRES Growth

summarised in Figure 2.

This chapter is focused on analysing the relationship between economic growth and the use of different sources of energy. We define as dependent variable the Logarithm of real Gross Domestic Product for country *c*, at period *t*, (*LGDPct*). The explanatory variables arise from the literature and are in accordance with those previously discussed. Therefore, in addition to the energy sources, we also control for energy consumption per capita, dependency on foreign energy and economic growth volatility. We then present and discuss the variables, their measurement and the expected contributions to economic growth. Given that volatility is a built variable instead of an observable one, we will explain in detail the process followed for its computation.


GDP growth has often been modelled as an autoregressive time series with random disturbances having conditional heteroskedastic variances. GDP growth, in particular, has been modelled as a GARCH type processes. The GARCH model is, in effect, sufficient to allow different macroeconomic regimes by letting the volatility of the economic growth evolve over time. It also assumes that a large change in GDP growth, either positive or negative, is probably followed by other large changes in subsequent years. Other methods of computing volatility, such as variance (or standard deviations), imply loss of observations and have several handicaps. Alternatively, they treat positive or negative changes in some way (the squares of economic growth rates) and were therefore excluded from our analysis.

We fit an autoregressive (AR) process with GARCH errors to the natural logarithm of the GDP per capita growth rates, assuming that the distribution of the error process is the normal distribution. This option results from the well-known characteristics of persistence of GDP growth. Indeed, we begin with the simplest model, namely the AR(1)-GARCH (1,1). Given the well-known propensity of the GARCH model to generate high estimated values at the beginning and end of time span, the AR(1)-GARCH(1,1) was estimated using raw data for the time span of 1971 to 2010.

The model estimated has a mean equation and an equation for conditional variance, which are, respectively (1) and (2):

$$\text{DLGDPPC}\_t = \gamma\_0 + \gamma\_1 \text{DLGDPPC}\_{t:1} + \varepsilon\_{t:t} \tag{1}$$

and

8 Renewable Energy – Trends and Applications




GDP growth has often been modelled as an autoregressive time series with random disturbances having conditional heteroskedastic variances. GDP growth, in particular, has been modelled as a GARCH type processes. The GARCH model is, in effect, sufficient to allow different macroeconomic regimes by letting the volatility of the economic growth evolve over time. It also assumes that a large change in GDP growth, either positive or negative, is probably followed by other large changes in subsequent years. Other methods of computing volatility, such as variance (or standard deviations), imply loss of observations and have several handicaps. Alternatively, they treat positive or negative changes in some way (the squares of economic growth rates) and were therefore excluded

We fit an autoregressive (AR) process with GARCH errors to the natural logarithm of the GDP per capita growth rates, assuming that the distribution of the error process is the normal distribution. This option results from the well-known characteristics of persistence of GDP growth. Indeed, we begin with the simplest model, namely the AR(1)-GARCH (1,1). Given the well-known propensity of the GARCH model to generate high estimated values at the beginning and end of time span, the AR(1)-GARCH(1,1) was estimated using raw

The model estimated has a mean equation and an equation for conditional variance, which

0 1 *t t* -1

, (1)

*DLGDPPC DLGDPPC <sup>t</sup>* 

if the former effect prevails.

literature to compute risk.

data for the time span of 1971 to 2010.

are, respectively (1) and (2):

growth.

from our analysis.

$$
\sigma\_t \sigma\_t^2 = \alpha + \alpha \sigma\_{\mathcal{E}t-1}^2 + \beta \sigma\_{t-1}^2 \tag{2}
$$

where *<sup>t</sup>* is the error term. In the above model, equation (1) is the conditional mean equation and equation (2) is the conditional variance equation. The conditional standard deviation term, *<sup>t</sup>* , represents the measure of GDP per capita growth volatility. One can also view *<sup>t</sup>* as a measure of economy wide risk.

Since we are more interested in the level of volatility than in the volatility itself ( *<sup>t</sup>* ), we proceed to establish the trend of volatility (*VOLGDPPCct*) applying the well-known Hodrick & Prescott (1997) – HP filter to the volatility obtained from the AR(1)-GARCH(1,1). Following a standard procedure of the related literature on HP filter, we use the value of λ =100 as the smoothing parameter.

Figure 3 shows the computed trend volatility. In general, there is no uniform behaviour pattern for the countries. For the time span analysed we observe the three possible kinds of trend: increase, decrease, and stability. For example, Austria and Spain reveal a period of stability until the end of the 1990s and a marked decline thereafter. In their turn, countries like Ireland, Luxembourg, and Poland show a trajectory of declining volatility. On the contrary, countries like France and Hungary reveal an increasing path with regard to volatility.

Fig. 3. Volatility trend


#### **3.3 Method**

This chapter makes use of panel data techniques to assess the nature of the effects of the several energy sources, and other drivers, on economic growth. Complex compositions of errors could be present in panel data analysis. The general model to estimate is:

$$LGDP\_{ct} = \alpha + \delta LCRES\_{ct-1} + \sum\_{k=1}^{k} \beta\_k X\_{kct} + d\_c + d\_t + \mu\_{ct} \tag{3}$$

where *LCRESct−1* is the share of renewables of country c in period t−1. The dummy variables *<sup>c</sup> d* and *<sup>t</sup> d* refer to country and time, respectively. In the error term *ct c c t ct* , 1 , *ct* is serially uncorrelated, but correlated over countries.

To deal with the complexity of the errors, good econometric practices suggest performing the analysis by first making a visual inspection of the nature of the data, followed by a battery of tests to detect the possible presence of heteroskedasticity, panel autocorrelation, and contemporaneous correlation. We use the Modified Wald test (Baum, 2001) in the residuals of a fixed effect regression, to appraise the existence of groupwise heteroskedasticity. The Modified Wald test has <sup>2</sup> distribution and tests the null of: 2 2 *<sup>c</sup>* , for *c N* 1,..., . The Wooldridge test assesses the presence of serial correlation. It is normally distributed *N*(0,1) and it tests the null of no serial correlation. We use the parametric testing procedure proposed by Pesaran (2004), the non-parametric test from Friedman (1937) and the semi-parametric test proposed by Frees (1995 and 2004), either for fixed effects or random effects, to test the countries' independence. Pesaran's test is a parametric testing procedure and follows a standard normal distribution; Frees' test uses Frees' Q-distribution; Friedman's test is a nonparametric test based on Spearman's rank correlation coefficient. All these tests - Pesaran, Frees and Friedman - test the null of cross-section independence.

Within a panel data analysis, the presence of such phenomena discourages the use of the common Fixed Effects (FE) and Random Effects (RE) estimators, due to the inefficiency in coefficient estimation and to biasedness in the estimation of standard errors they could cause. In this case, the appropriate estimators to be used are the Feasible Generalised Least Squares (FGLS) and the Panel Corrected Standard Errors (PCSE). In our sample, the number of cross sections (21) is larger than the number of time periods (18) and, therefore, the best suited estimator to deal with the presence of panel-level heteroskedasticity and contemporaneous correlation is the PCSE (Reed & YE, 2009).

The PCSE estimator allows the use of first-order autoregressive models for *ct* over time in (3), it allows *ct* to be correlated over the countries, and allows *ct* to be heteroskedastic (Cameron and Triverdi, 2009). We begin by estimating a pooled OLS model (model *I*) and then we work on a panel data structure by applying the PCSE estimator. We will estimate the model presupposing the various assumptions about variances across panels and serial correlations, with the aim of checking the robustness of the results. The assumptions made throughout the models are as follows: model *II* - correlation over countries and no autocorrelation; model *III* – country-level heteroskedastic errors and common first-order autoregressive error (AR1); model *IV* - correlation over countries and autocorrelation AR(1); and model *V* - correlation over countries and autocorrelation country-specific AR(1).

#### **3.4 Data**

10 Renewable Energy – Trends and Applications


enlarged. If the second effect overcomes, then a negative signal is achieved.


and the latter is relatively recent, the expected effect may not be obvious *a priori*.

This chapter makes use of panel data techniques to assess the nature of the effects of the several energy sources, and other drivers, on economic growth. Complex compositions of

1

where *LCRESct−1* is the share of renewables of country c in period t−1. The dummy variables

To deal with the complexity of the errors, good econometric practices suggest performing the analysis by first making a visual inspection of the nature of the data, followed by a battery of tests to detect the possible presence of heteroskedasticity, panel autocorrelation, and contemporaneous correlation. We use the Modified Wald test (Baum, 2001) in the residuals of a fixed effect regression, to appraise the existence of groupwise heteroskedasticity. The

distribution and tests the null of: 2 2

Wooldridge test assesses the presence of serial correlation. It is normally distributed *N*(0,1) and it tests the null of no serial correlation. We use the parametric testing procedure proposed by

*LGDP LCRES X d d*

1

(3)

*k ct ct k kct c t ct k*

,

 *<sup>c</sup>* 

*ct c c t ct* 

 , 1 ,

, for *c N* 1,..., . The

*ct* is

errors could be present in panel data analysis. The general model to estimate is:

serially uncorrelated, but correlated over countries.

Modified Wald test has <sup>2</sup>

*<sup>c</sup> d* and *<sup>t</sup> d* refer to country and time, respectively. In the error term

**3.3 Method** 

The data used in this chapter come from several sources. Table 1 summarises the variables, their sources and their descriptive statistics. The time span is 1990-2007, and we collect data for 21 EU Members, those for which there are available data for all the variables.



378 2.5407 1.2422 1.0622 8.7522

378 52.2925 29.6911 -50.83 99.8

378 0.3614 0.2753 0 0.97

378 0.0698 0.0983 0 0.51

Factbook 2010 376 1.5965 1.0126 -1.6094 3.4404

**Variable Definition Source Obs Mean SD Min Max** 

Own calculation. Raw data from World Bank World Development Indicators, and International Financial Statistics of the

IMF

OECD

Ratio electricity generation to coal (TWh) / total elect. generation (TWh). EU Energy in Figures 2010 DG TREN

Ratio electricity generation to oil / total elect. Generation. EU Energy in Figures 2010 DG TREN

EU Energy in Figures 2010 DG TREN

*VOLGDPPC ct*

*LCRESct-1*

*IMPTDPct*

*SCOALEGct*

*SOILEGct*

*Per capita* GDP volatility

Logarithm of the factor of contribution of renewables to total primary energy supply, lagged one period

Import

coal to electricity generation

dependency of energy (%)

Contribution of

Contribution of oil to electricity generation


Table 1. Data: definition, sources and descriptive statistics

First following a visual inspection of the data, we analyse the correlation coefficients, which are disclosed in the correlation matrix (table 2). In general, the correlation coefficients did not arouse any particular concern about the existence of collinearity among explanatory variables, although the correlation of *VOLGDPPC* with *LGDP* may be a possible exception.


Table 2. Correlation matrix

In order to dispel any doubt we proceed as follows: i) we estimate the models excluding the variable volatility, concluding that there is no change in the coefficients' signals; ii) we compute the Variance Inflation Factor (VIF) test for multicollinearity (see table 3). The mean VIF is only 2.35 and the largest individual VIF is 4.21. From all this we conclude that collinearity is not a concern.


Table 3. Variance Inflation Factor

Once the first inspection of the data had been made, we proceeded by testing the intrinsic characteristics of the data, namely by assessing the presence of the phenomena previously reported, i.e., heteroskedasticity, panel autocorrelation, and contemporaneous correlation. Table 4 reveals the specification tests we computed.


Table 4. Specification tests

From table 2, the null hypothesis of no first-order autocorrelation is rejected, as suggested by the Wooldridge test. From the Modified Wald statistic, we observe that the errors exhibit groupwise heteroskedasticity. As far as the contemporaneous correlation is concerned, all the tests are unanimous in their conclusions. They support the rejection of the null of crosssectional independence, and thus the residuals do not appear to be spatially independent. The use of the PCSE is therefore sustained.
