Examining the stability of the long-run relationship between energy and GDP in the presence of an industrialization process

ABSTRACT In this paper, we examine the extent to which the relationship between energy consumption and real gross domestic product can be affected by an effective economic development policy based on industrialization. The Spanish Stabilization and Liberalization Plan of 1959 is regarded as a “natural experiment” that is especially interesting for our purposes. To explore long-run causality and possible changes in its direction, we have exploited statistical information for a long period (1900–2008). We applied a moving-window regression procedure based on the augmented vector autoregressive methodology of Toda and Yamamoto (1995). Causality results are highly sensitive to the industrialization process. For the sub-period preceding the development policy, we consistently found greater support for the “conservation hypothesis”. This contrasts sharply with the set of results obtained from the data belonging to the subsequent sub-period, where evidence is in favor of the “growth hypothesis”.


Introduction
Nowadays, there is growing interest in studying the relationship between energy consumption and economic growth, since knowledge about this issue may be particularly useful as a guide in formulating recommendations for energy and environmental policies. The aim of this paper is to contribute to this burning debate by empirically investigating whether (and, if so, in which direction) the causal relationship between primary energy use and economic growth can be affected by an accelerated industrialization. The successful policy reforms carried out following the Spanish Stabilization and Liberalization Plan (SLP) of 1959 constitute a "natural experiment" that is particularly interesting for our purposes. This development policy marked the end of an era that was characterized by a lack of productive resources and energy restrictions derived from the situation of autarky, and led the way toward a new period where economic openness and liberalization sharply favored an intense process of industrialization and technological modernization.
Since the seminal paper of Kraft and Kraft (1978), a large number of studies have been interested in testing long-run causality between energy consumption and economic growth. As can be seen from a review of the literature on the topic, 1 the empirical evidence is rather mixed, even when the same country is analyzed. 2 The heterogeneity of the results could reasonably be attributed, at least in part, to some difference in the periods analyzed. In particular, it can be hypothesized that the level of industrial development of a country is, to a greater or lesser extent, progressively changing over time, and empirical results are sensitive to this sort of changes. Furthermore, the lack of robustness of some of the methodologies that are usually applied may also play an important role. On the one hand, older causality techniques such as Granger (1969) or Sims (1972) can lead to spurious results when time series are non-stationary (Park and Phillips, 1989;Stock and Watson 1989). On the other hand, the application of the Engle and Granger (1987) and Johansen and Juselius (1990) methodologies by using small samples, can be criticized on the grounds of their low power to test for cointegration between variables (Zapata and Rambaldi 1997).
We exploit time series over a long period of time , which will allow us to explore the extent to which economic development based on an industrialization process may affect the possible causality between energy use and GDP. To do so, we will apply the methodology developed by Toda and Yamamoto (1995) (hereinafter: TY), which takes into account the possibility of unit roots and avoids the problems related to the power and size properties of cointegration tests. This methodology has been widely used in the last few years in the literature on this issue. The recent set of papers by Bowden and Payne (2010), Payne and Taylor (2010), Payne (2011a), Payne (2011b), and Payne (2012) are some interesting examples.
Next, we start by focusing on the policy conditions in the first half of the twentieth century, and then go on to look at the main market-oriented reforms related with the SLP. Section 3 describes both the data and the econometric methodology. Section 4 presents the results. Finally, the main conclusions will be discussed.

An overview of the period analyzed
The first part of the twentieth century was characterized by a low level of industrialization in which the local production of coal was almost the only energy source used. The utilization of this energy source was promoted at this initial stage from 1914 onwards as a result of an increase in external demand deriving from the First World War. This fact gave a certain impulse to the process of industrialization, which was later interrupted by the Spanish Civil War (1936)(1937)(1938)(1939). From then on, under the first two decades of the dictatorial mandate of General Francisco Franco, a high state of interventionism and an autarkic regime were established. One of the cornerstones of the interventionist policy was a plan for the generation of domestic energy resources based on the construction of dams and hydroelectric power plants. However, during this period of interventionism and international isolation, economic development was still relatively low compared to the rest of the neighboring European countries, which led the country into a situation of considerable backwardness. The limitation on the amount of foreign currency as well as the administrative obstacles hindering the importation of productive inputs were seen as some of the major causes of bottlenecks curbing the possibilities to grow.
In 1959, given the domestic economic situation, the dictatorial government carried out a set of complementary policy measures for development called the SLP. 3 The short-run objectives were the stabilization of domestic prices, the liberalization of domestic economic activities, and to boost foreign trade. Interestingly, the reduction of administrative obstacles hindering imports as well as achieving international convertibility of the local currency (peseta) were some of the essential measures applied to promote external trade. An exceptional increase in GDP and energy consumption took place a few years after the implementation of the development plan. Although the incorporation of new local energy sources played a relevant role (i.e., hydro and thermal power plants), the expansion of energy use was mainly sustained by the importation of petroleum products. Hence, according to data from the Spanish Ministry of Industry, Energy and Tourism, fuel oil imports increased six-fold in just fifteen years after the introduction of the SLP.
Following the implementation of the SLP, the real GDP grew significantly until 2008 (i.e., when the international economic crisis took place). With few exceptions, such as the short stage of stagnation of energy consumption derived from the world's second oil crisis in 1979, this economic growth was accompanied by a sharp rise in energy consumption.

Data and methodology
We consider annual data on primary energy consumption and GDP in Spain covering the period from 1900 to 2008. On the one hand, the energy consumption is expressed in tons of oil equivalent. The primary energy consumption series consisted in the period up to 1972 and was obtained from Sudrià et al. (1995), while the data from 1973 to 2008 was taken from the Spanish Ministry of Industry, Energy and Tourism. On the other hand, GDP is expressed in Euros and in real terms (on the basis of 1995), in accordance with the variation in the domestic consumer price index. The real GDP variable up to 2000 was drawn from Prados de la Escosura (2003), while more recent data were collected from the Bank of Spain's Statistical Bulletin (December 2011). The evolution of both time series throughout the entire period is shown in Figure 1.
To analyze the long-run causality between variables we employ the TY approach. The procedure requires the estimation of an "augmented" VAR model in levels with a (p þ d max ) order, where pis the optimal lag length and d max represents the maximal order of integration. Granger non-causality is then tested on the first p coefficient matrices by the standard Wald statistic (ignoring the last d max lagged vectors of the variables). The following specification is considered: 4 (1) where y t ¼ ec t ; gdp t À Á ; is an nx1 vector of endogenous variables composed of the log of primary energy consumption and the log of GDP, respectively, and y tÀi is the corresponding lag term of order i. Additionally, γ i is the (nxn) matrix of autoregressive coefficients of vector y tÀi for i ¼ 1; 2; . . . ; p þ d max ; α ¼ α ec ; α gdp À Á is the nx1 ð Þ intercept vector; and β ¼ β ec ; β gdp refers to Figure 1. Evolution of real GDP and energy consumption for Spain . 4 As in other papers on this issue (e.g., Fuinhas and Marques 2012), we acknowledge that some bias from omitted variables could result from the use of simple bivariate models. We are obviously limited to the availability of time series for the analyzed period. However, rather than improving the accuracy of the causality test for a given sub-period, we are more interested in the sensitivity of the results when faced with changes in the degree of industrialization.
the nx1 ð Þ trend vector. Lastly, u t ¼ u ec ; u gdp À Á represents the nx1 ð Þ vector of error terms, which is assumed to be a white-noise process.

Causality analysis
With the aim of identifying possible changes in causality from 1900 to 2008, we implemented moving-window OLS regressions based on Eq. (1). We have introduced an impulse dummy for the Civil War period (1936)(1937)(1938)(1939), and a trend dummy for the years following 1936. For each window, the optimal lag length (p) has been selected according to the Akaike Information Criterion (AIC), while the maximal order of integration (d max Þ was obtained from the unit root tests proposed by Phillips and Perron (1988) (hereinafter: PP), and Elliot et al. (1996) (hereinafter: Dickey-Fuller GLS, DF-GLS). In Table 1 we provide the results for both unit root tests and long-run causality statistics by considering a window size of 40 observations.
Results from the PP and DF-GLS unit root tests indicate that the gdpvariable is non-stationary in levels but it becomes stationary after taking the first differences (at the 0.01 level of significance). However, both unit root tests produce different results for the stationarity of ecin some cases. Note that TY approach is particularly suitable for application when the variables are integrated of different orders. Then, we are able to safely carry out the procedure after identifying that the maximum order of integration d max is 1 for all the windows.
With regard to the long-run causality test, the set of empirical results from the exploratory analysis can be straightforwardly divided into three well-differentiated groups. First, there are those derived from moving-window regressions based on a sample from the sub-period 1900-1955. From this set of results, we can reject the null hypothesis that ec does not cause gdp at least at the 0.10 significance level. Second, results derived from data coming from both this first sub-period and the later one are rather mixed (i.e., non-causality, causality from energy to GDP, and causality from GDP to energy consumption). Third, the results become broadly consistent again when statistical information from the whole sub-period 1961-2008 is considered. However, in contrast to the first sub-period, we can now reject the null hypothesis that gdp does not cause ec from all movingwindow results at the 0.01 significance level.
Additionally, we now perform the causality analysis by separating the total sample directly into the two sub-periods detected above: one before the implementation of the SLP  and the other after . This empirical strategy will allow us to check the robustness of the general outcome obtained previously. The results for both sub-periods are reported in Table 2. According to the AIC and unit root tests, we can safely assume a VAR with 2 lags (p = 1 and d max = 1) from 1900 to 1955 and a VAR with 5 lags (p = 3 and d max = 2) from 1961 to 2008. A battery of diagnostic tests indicates that such specifications present the desired econometric properties in both sub-periods. 5 Consistent with the results from the moving-window approach, the causality tests for separated time periods suggest that before the SLP of 1959 the direction of causality runs from GDP to energy use. That is, in the first part of the last century, empirical evidence is broadly consistent with what is well known in the literature as the "conservation hypothesis". From the beginning of the sub-period characterized by an accelerated industrialization process to the present, the empirical evidence indicates that causality runs from energy use to GDP; that is, evidence of the "conservation hypothesis" clearly disappears in favor of what is widely known as the "growth hypothesis".

Concluding remarks
In this paper, we have empirically investigated the extent to which the causal relationship between primary energy use and economic growth can be affected by a substantial change in the growth of 5 Moreover, the stability conditions for VAR from Eq. (1) are satisfied for both sub-periods. The inverse roots of the AR characteristic polynomial are available upon request from the corresponding author. economic development based on a process of industrialization. The Granger causality analysis used in our work is based on the augmented VAR methodology developed by TY in a moving-window framework with the aim of exploring possible changes in causality over time. We found broad consistency of evidence in favor of the "conservation hypothesis" (i.e., unidirectional causality running from GDP to energy consumption) only when data information until 1955 was used. When the sample window corresponds to a mixture of data information from both sub-periods (i.e., before and after the economic reforms), results on causality are rather mixed. Finally, from data belonging to the sub-period after the economic development policy, we found consistent evidence in  Perron (1988), andElliott et al. (1996), respectively. In this regard, I(0) represents the non-rejection of the null hypothesis of stationarity of the series, whereas I(1) means that variable has one unit root. Lags refer to the optimal lag length according to AIC (p) plus the maximum integration order of the series (d max ). The p-values are in brackets and we use ***, ** and * to indicate the rejection of the null hypothesis of non-causality at the 1%, 5% and 10% significance levels (α), respectively. We account for the Spanish Civil War in estimations by incorporating an impulse dummy for the time period 1936-39 and a trend dummy for the years after 1936. favor of the fulfillment of the "growth hypothesis" (i.e., causality runs from energy consumption to GDP). Thus, our results suggest that, although the conservation energy policies are probably adequate in contexts with low economic growth and industrial stagnation, they may not be entirely appropriate in economies with an accelerated industry-based development, despite the possible beneficial effect on the environment. In these latter cases, in order to preserve the quality of the environment, we should alternatively consider the possibility of promoting the use of clean energy sources rather than trying to implement policies to reduce energy consumption. However, the existing evidence on the question under study is still too limited and, therefore, further studies need to be conducted to draw final conclusions. We expect that the availability of time data series for other regions which have also undergone an economic development "miracle" some time in their history will help us to improve the knowledge available on this question.

Funding
Financial support from the European Union FEDER funds and the Spanish Ministry of Economy and Competitiveness (ECO2014-58975-P) are gratefully acknowledged.