Energy Intensity: A Decomposition and Counterfactual Exercise for Latin American Countries Energy Intensity: A Decomposition and Counterfactual Exercise for Latin American Countries

Energy Intensity: A Decomposition and Counterfactual Exercise for Latin American Countries* Raul Jimenez a** a b

Jorge Mercado b

Inter-American Development Bank and University of Rome Tor Vergata

Energy Division, Infrastructure Department, Inter-American Development Bank.

Abstract This paper investigates trends in energy intensity over the last 40 years. Based on a sample of 75 countries, it applies the Fisher Ideal Index to decompose the energy intensity into the relative contributions of energy efficiency and economic structure. Then, the determinants of these energy indexes are examined through panel data regression techniques. Special attention is lent to Latin American countries (LAC) by comparing its performance to that of a similar set of countries chosen through the synthetic control method. When analyzed by income level, energy intensity has decreased in a range between 40 and 54 percent in low and medium income countries respectively. Efficiency improvements drive these changes, while the structural effect does not represent a clear source of change. The regression analysis shows that per capita income, petroleum prices, fuel-energy mix, and GDP growth are main determinants of energy intensity and efficiency, while there are no clear correlations with the activity component. In the case of LAC the energy intensity decreased around 20 percent which could be interpreted as an under-performance. However, the counterfactual exercise suggests that LAC has closed the gap with respect to its synthetic control.

Keywords: energy intensity; decomposition; panel data; synthetic control method. JEL Code: O5; O13; Q40; Q43

*

The opinions expressed in this article are strictly those of the authors and do not necessarily reflect those of the Inter-American Development Bank (IDB), its Board of Executive Directors or the countries they represent. A previous version was published as a working paper by the IDB. The authors are grateful for the support of Ramon Espinasa and the Research Department at the IDB, as well as for the helpful comments and suggestions of Lenin Balza, Diego Margot, Juan Jose Miranda, Tomas Serebrisky, Rodolfo Stucchi and four anonymous peer reviewers. All remaining errors are our own responsibility. **

Corresponding author: [email protected], address: 1300 New York Avenue, N.W. Washington, DC 20577; phone: 1-202-623-2170.

1

Energy Intensity: A Decomposition and Counterfactual Exercise for Latin American Countries

1. INTRODUCTION As both energy prices and concerns about global warming continue to increase, measures to improve the energy use have become important components of public policy agenda. In particular, there is a focus on identifying factors that influence change in energy intensity and distinguishing the contribution of energy efficiency from other relevant factors. This information is useful as it provides a basis for policy decisions and evaluation. Further, energy efficiency represents a cost-effective strategy to address crosscutting issues such as energy security, climate change and competitiveness. In this context, this paper aims to investigate trends in energy intensity based on a sample of 75 countries with annual data during the period 1971-2010. To this end, three specific objectives are addressed. First, analyze the evolution of the aggregate energy intensity and its main components. Then, identify the main determinants of these energy indexes. Finally, the article lends special interest to the Latin American region by evaluating its relative performance in terms of energy intensity and efficiency. Following the World Energy Outlook (IEA, 2012) energy intensity would have decreased about 20% in the World and 35% in OECD countries between 1980 and 2010. Accordantly, empirical literature suggests a downward trend in energy intensity, with the efficiency effect as its most important source of variation. However, the magnitudes of those improvements tend to be heterogeneous depending on the case and period analyzed. Previous studies could be divided in two groups. One with a rich and large body decomposing and examining trends in energy intensity within a specific sector, where the manufacture has received great attention. The other group has been less explored and bases its analysis on more aggregate data mainly at multi-sector level within a country. With respect to previous research in the industrial sector, some relevant figures emerge of the well-studied cases as China, India and United States. China represents a notable case of improvement decreasing its level of energy intensity more than 70 percent between 1980–2010. Sinton and Levine (1994); Zhang (2003), and Ma and Stern (2008) show that this was mostly a 2

sustained decrease in industrial energy intensity, with efficiency explaining most of this variation. As reported by Ke et al. (2012), during the period 1996–2010, the efficiency effect explains 30 percent of the energy savings in industrial energy consumption. Another remarkable fact occurring in the industrial sector is that intensity actually increase in the period 2003 to 2005, related to a notably increase in the levels of energy consumption. Still, the industrial energy intensity continued to decrease. In contrast, studies of the Indian industrial sector found mixed results from 1981 to 2005, showing only slight improvement in energy intensity (see Reddy and Ray, 2011). Interesting cases where both efficiency and activity have played a role in reducing the overall energy intensity index are found in studies of the United States. Hasanbeigi, Rue du Can, and Sathaye (2012) show that in California, from 1997 to 2008, the energy intensity ratio decreased 43 percent mainly explained by two events: (i) a shift in value added participation from the oil and gas extraction sector to the electric and electronic manufacturing sector, which uses less energy per value added; and (ii) an escalation in energy prices that led the industries to improve efficiency in order to reduce energy costs. Over a similar period, Huntington (2010) analyzes 65 U.S. industries in the commercial, industrial, and transportation sectors, showing that an estimated 40 percent of reduction in aggregate energy intensity was due to structural change. In one of the first studies available on energy intensity at the state/country level, Metcalf (2008) performs a decomposition exercise at state level in the United States for the period between 1970 and 2001. He finds a reduction in energy intensity of approximately 75 percent as a result of efficiency improvements. Further, through a panel data analysis, he shows that rising per capita income and higher energy prices play an important part in lowering energy intensity. Bernstein et al. (2003) analyze a similar period using a sample of 48 states in the U.S., finding that population, energy prices, climate temperatures, and indicators of sector activities, are strongly correlated with energy intensity. In a recent study, Voigt et al (2014) perform a decomposition analysis finding that intensity decreased by 18 percent on a sample of 40 major economies over the period 1995 and 2007. The results also suggest that this improvement is largely attributable to technological change. Using a similar approach, Bhattacharya and Shyamal (2001) decompose the aggregate energy intensity of India into pure intensity or efficiency and the economic activity composition 3

effect. They take broad sectors including agriculture, industry, and transport for the period between 1980 and 1986, finding that the efficiency effect contributed significantly to energy conservation. This paper focuses on energy intensity indicators at broad end-use sectors at the country level. This implies the observation of (aggregate) energy indexes (i.e., the indicators of energy intensity and its decomposition into efficiency and the activity mix) at the country level. For this purpose, we adopt the monetary-based definition, where energy efficiency improvement generally means using less energy to produce the same amount (value added) of services or output (Nanduri, 2002 and Ang, 2004). In this context, the paper has three main contributions. Strengthen the literature by analyzing a greater sample over a longer period than previous studies. A further step is provided by the analysis of the determinants of energy intensity and its components. Second, the paper shows results by income level set of countries with a focus in Latin American region, where appears to be lacking of evidence. Finally, a methodological contribution to this specific literature is the comparison analysis using the synthetic control method in order to overcome heterogeneity issues in a benchmark exercise. The paper is structured as follows. Next section provides methodological strategies for (i) the decomposition of aggregate energy intensity into activity and pure intensity, which is interpreted as efficiency, (ii) the specification of the panel data analysis in order to evaluate the determinant of those three indexes (intensity, efficiency and activity), and (iii) the synthetic control method used to construct a comparison set of countries to evaluate the relative performance of Latin America. Section 3 presents the empirical results of these methodologies, and Section 4 concludes.

2. EMPIRICAL STRATEGIES 2.1. Decomposition through the Fisher Ideal Index A key limitation in empirical analysis is related with availability of data. Based on different levels of data disaggregation, methodological contributions have been made in order to estimate energy efficiency measures. Those methods are mostly based on decomposing energy intensity into different factors, including energy efficiency, economic structure, production levels, and/or 4

fuel sources. The more disaggregate the data, the more accurate the efficiency contribution estimations would be. The election of the specific method to be used depends on the objectives and data availability. Some extensive methodological studies and surveys on decomposition methods can be found in Boyd, Hanson, and Sterner (1988); Ang and Lee (1994); Ang and Liu (2003); Ang (2004); Boyd and Roop (2004); and Ang, Huang, and Mu (2009). They suggest a certain degree of academic consensus that using price index numbers is preferred when dealing with aggregate data at the country level. Following those recommendations, the method applied herein to perform the decomposition is the Fisher Ideal Index. Its main advantage is that it does not have residual term, referring to a portion of the change in intensity which is not assigned to a particular source; that is, a portion of energy intensity that remains unexplained (Boyd and Roop, 2004). The presence of residual term makes it difficult to interpret the relative importance of factors being evaluated. Specifically, Ang, Mu, and Zhou (2010) emphasize that the perfect decomposition methods should be adopted in the case of cross-country/region studies. In addition, as mentioned by Ang (2004; 2006), Boyd and Roop (2004), and Ang and Liu (2003), these methods are also preferred in the case of two-factor decomposition due to their theoretical foundation and their adaptability, as well as the ease in interpreting their results. In our case, energy intensity is decomposed into its efficiency and activity components. Besides the references above Ang and Lee (1994), Greening et al (1997), Ang, Mu and Zhou (2010) provide a compressive review and applications of alternative decomposition methods. In this context, the problem is set in terms of total energy consumption (E) and total production (Y), as well as sub-indexes for economic sector (i) and years (t). In our application 𝑖 refers to the agricultural, industrial, services, and residential sectors. Thus, the aggregate energy intensity (e) can be written as: 𝑒𝑡 =

𝐸𝑡 𝑌𝑡

= ∑𝑛𝑖

𝐸𝑖𝑡 𝑌𝑖𝑡 𝑌𝑖𝑡 𝑌𝑡

= ∑𝑛𝑖 𝑒𝑖𝑡 𝑠𝑖𝑡 (1)

Expression 1 indicates that a change in 𝑒𝑡 may be due to changes in the sector energy intensity (𝑒𝑖𝑡 ) and/or the product mix or compositional effect (𝑠𝑖𝑡 ). By construction, the energy uses in the different sectors need to form a partition (i.e., they must not overlap), but the measures of economic activities do not need to satisfy this condition. The last represents one of the main operative/practical advantages of this approach. What is more, they do not even need 5

to be in the same units, facilitating the identification of good indicators to account for the activity mix (𝑠𝑖𝑡 ). Following the index number theory, we proceed to derive the two components of the Fischer index. Dividing equation (1) by the aggregate energy intensity for a base year (𝑒0 = ∑ni ei0 si0 ),

∑n ei0 si0

and factorizing by ∑in i

ei0 si0

∑n eit sit

and ∑in i

eit sit

, it is obtained the Laspeyres and Paasche indexes

respectively. Laspeyres indexes

:

∑𝑛 𝑒𝑖0 𝑠𝑖𝑡

𝐿𝑎𝑐𝑡 = ∑𝑛𝑖 𝑡 𝑖

Paasche indexes

:

𝑒𝑓𝑓

𝐿𝑡

𝑒𝑖0 𝑠𝑖0

∑𝑛 𝑒𝑖𝑡 𝑠𝑖𝑡

𝑃𝑡𝑎𝑐𝑡 = ∑𝑛𝑖 𝑖

𝑒𝑓𝑓

𝑃𝑡

𝑒𝑖𝑡 𝑠𝑖0

∑𝑛 𝑒𝑖𝑡 𝑠𝑖0

= ∑𝑛𝑖 𝑖

𝑒𝑖0 𝑠𝑖0

∑𝑛 𝑒𝑖𝑡 𝑠𝑖𝑡

= ∑𝑛𝑖 𝑖

𝑒𝑖0 𝑠𝑖𝑡

Equations (2.1) and (2.2) reflect the components that could be attributed to changes in the activity mix or to pure intensity changes, which will be interpreted as efficiency effect. Then, the activity and efficiency index are constructed as follows: 𝑒𝑓𝑓

𝑎𝑐𝑡 𝐹𝑡𝑎𝑐𝑡 = √𝐿𝑎𝑐𝑡 (2.1) 𝑡 𝑃𝑡

𝐹𝑡

𝑒𝑓𝑓 𝑒𝑓𝑓

= √𝐿𝑡 𝑃𝑡

(2.2)

They are the Fisher Ideal Indexes, which is a geometric mean of the Laspeyres and Paasche indicators. By multiplying both, it can be recovered the aggregate energy intensity index: 𝑒𝑡 𝑒0

𝑒𝑓𝑓

≡ 𝐼𝑡 = 𝐹𝑡𝑎𝑐𝑡 𝐹𝑡

… (3)

Then, the method allows a perfect decomposition of the aggregate energy intensity index into 𝐹 𝑒𝑓𝑓 and 𝐹 𝑎𝑐𝑡 indexes with no residual. By taking the logarithm of (3), it is possible to observe the additive contribution of the activity-mix effect and the energy efficiency effect to the total variation in energy intensity. For a detailed review and derivation of this method see Ang, Liu and Chung (2004); Boyd and Roop (2004), and de Boer (2008). It is important to highlight that at working with aggregate end-use data; it will not be possible to detect shifts between subsectors in each broad activity. Thus, the current study does not capture structural changes between sub-industries with high-energy intensity versus lowenergy intensity within the industrial sector. To identify specific trends in each subsector, or in products and services, it would be necessary to use more detailed information.

6

A potential drawback to this strategy is that the estimations herein could be sensitive to the degree of data disaggregation. For example, within a broad activity, changes from more energy-intensive sub-activities to less energy-intensive sub-activities could lead to overestimate the gains in energy efficiency (and vice versa). Then, it is possible to interpret a result as an energy efficiency effect when it is really a compositional effect within a broad activity. In general, it is preferable to have more disaggregated good quality data to obtain better estimates. In the case of California industry, an interesting finding by Metcalf (2008) is that a higher level of disaggregation did not significantly affect his estimations. However, Huntington (2010) found contrasting results using a more detailed dataset.1 2.2. Panel Data Determinants Analysis In line with the approach taken by Galli (1998) and Metcalf (2008) the current paper relies on a dynamic panel data specification to analyze the determinants of the energy indexes. By adding the lagged dependent variable, it allows modeling the state dependence of the energy indexes and estimates its partial adjustment process. That is, energy indexes could react slowly to changes in the explanatory variables. Besides, having the lagged dependent variable makes it possible to estimate the elasticity of the short and long run, where 𝛾 is interpreted as the speed of the adjustment to the long-run equilibrium relationship. In equation 4, the dependent variable (𝑦) refers to intensity, efficiency, or the activity index. That is, it will be performed three regressions for each energy index calculated through the Fisher Ideal Method as explained in section 1.1. The matrix (X) represents the set of variables of interest suggested by the literature –e.g. Bernstein, et al., 2003; Metcalf, 2008– and includes per capita income, energy prices, population growth, fossil fuel energy consumption, and the investment capital ratio. Besides, we also include, as part of X, the economic growth rate and rent from natural resources. In order to account for invariable characteristics specific to each country, we include the country fixed effect (𝜇). In addition, to account for effects that change over time, the specification contains a trend by country (𝑡). The inclusion of this cross-

1

It is important to mention that both authors use different datasets and analyze different periods. In their study of the energy intensity trend in China, Ma and Stern (2008) provide another example where the data disaggregation could affect the decomposition results. They found that the contribution of the industry mix goes from positive to negative, after performing the decomposition with more detailed data.

7

section specific effects, as well as previous covariates ( 𝑋 ), are expected to capture the heterogeneity across countries over time. The proposed specification is as follows: 𝑦𝑖𝑡 = 𝛽𝑋𝑖𝑡 + 𝛾𝑦𝑖𝑡−1 + ∑𝑖 𝛼𝑖 𝜇𝑖 + ∑𝑖 𝜃𝑖 𝑡𝑖 + 𝜀𝑖𝑡 … (4) With respect to the expected relationship between the explanatory variables and the energy indexes, there is a certain degree of consensus about the effect of energy prices on intensity and efficiency. However, there is no conclusive evidence about the effects of the other variables. In the case of prices, higher prices would lead to reduced intensity through improving efficiency and/or turning to less intense activities.2 Sue Wing (2008) emphasizes three channels through which prices would influence energy intensity: (i) production input substitution due to changes in relative energy prices, given constant technology; (ii) innovation, capturing both secular scientific progress and inducement effects of high energy prices; and (iii) changes in the composition of the stock of quasi-fixed inputs to production.3 The effects of per capita income on the energy indexes remains an issue of empirical discussion, as the level of energy intensity could change according to the level of economic development; see for example Galli (1998) and Metcalf (2008). On one hand, it is expected that income would put pressure on the demand for energy, increasing intensity. On the other, as income broadly reflects the stage of development, it is expected that it would correlate positively with the degree of efficiency, reducing energy intensity. This justifies considering the square of per capita income to allow nonlinearities that capture both effects. The effects of new investments (measured through the investment capital ratio) on energy indexes are also not certain. While they would improve energy intensity and efficiency by making the stock of available capital more productive, they could also be targeted primarily at enhancing production capacities without significant effects on energy savings. Further, investments oriented toward improving energy efficiency are usually very specific, and not necessarily aligned with other types of investments.

2

It would be preferable to account for energy prices, however since there is no uniform data on energy prices for all countries, we use international petroleum prices in real terms from 2005 as a proxy. Oil prices play a significant role several oil-imported economies and even in those oil producers countries with market oriented industries. 3 In a study of 35 industries in the United States during the period 1958–2000, Sue Wing shows that the energy prices influenced a decline in energy intensity, mainly due to the quasi-fixed variable costs, particularly vehicle stocks and disembodied autonomous technological progress.

8

With respect to the population dimension, fast-growing population rates may be associated with agglomeration economies that tend to make energy use more efficient. However, these economies of scale depend on infrastructure growing fast enough to cover the needs of the growing population. For example, a direct consequence of population and infrastructure growing at different rates is traffic congestion, which leads to greater use of fossil fuels per the same unit of distance traveled. The fossil fuel mix, measured as the ratio of fossil fuel energy consumption to total energy use, does not have a clear influence on the energy indexes. In recent studies, Ma and Stern (2008), and Shahiduzzaman and Khorshed (2013) suggest an inverse relationship. It can also be argued that high fuel consumption makes a country sensitive to price variation, providing an incentive for increased efficiency. However, it is important to note that there is little evidence of the mechanism by which this relationship operates. For example, the level of fossil fuel consumption could be endogenous, resulting from abundance in resources, which could provide a perverse incentive to maintain a high use of fossil fuels without improving efficiency. For this reason, we include as a regressor the rent from natural resources, which is expected to capture the effects of being a country with relative abundance in extractive resources over the energy indexes. Based on the literature on natural resources and economic growth (e.g., Sachs and Warner, 1995), one could argue that a country rich in fossil fuels, with subsidized energy prices, would not have an incentive to change its fuel mix or invest in more energy efficient technologies, leading it to maintain a high level of energy intensity. Moreover, to take into account the performance of the economy, we include the Gross Domestic Product (GDP) growth rate as another co-variable. It is expected that a country’s economic growth will encourage energy efficient investments and/or boost other sectors in the economy that have differing energy intensities. The method of estimation for eq. (4) is Least Square Dummy Variable (LSDV). It is expected that this estimator would perform well in samples with large T, which is the case herein, since we restrict our exercise to the countries with the longest sets of information. Still in the appendix 3, it is provided two robustness exercises under Bias Corrected LSDV or Kiviet estimator which is suggested by the Monte Carlo experiments (Judson and Owen, 1999, and 9

Galiani and Gonzalez-Rozada, 2002), and the most commonly used Arellano and Bond estimator. The presence of Heteroskedasticity was tested (Modified Wald Test) and corrected through the estimation of robust standard errors. The presence of unit root was rejected (tests of Im-Pesaran-Shin and Fisher) in the residuals of eq. 4 for each of the energy indexes as dependent, suggesting that our variables are co-integrated. Since the power of previous test could be low due to the presence of structural breaks, the test of Zivot-Andrews which allows for multiple structural breaks was also applied to the residual of eq. (4) as well as to each variable by country. In the case of the residuals, in levels, the test rejects the presence of unit root in most countries supporting previous results. The dependent and independent variables have in most countries unit root in level but are stationary in first differences. No systematic autocorrelation in the residuals of eq. (4) were found until lag 10 (Arellano and Bond test). Only in lag 5 for the intensity and activity indexes, and in lag 4 for efficiency; serial correlation cannot be rejected at 5 percent of statistical significance. We interpret these results in favor of specification (4) in the sense that there is not serial correlation which could lead to bias estimations.

2.3. Synthetic Control Method for the Average Latin American Country As will be shown in next section, over the last decades LA region seems to have a particular pattern of energy intensity trends. Then, in order to perform a credible comparison of the energy indexes of Latin America it is necessary to construct a similar set of countries. A suitable methodology for this task is the synthetic control approach (Synth) proposed by Abadie and Gardeazabal (2003) and detailed by Abadie, Diamond, and Hainmueller (2010). This method would allow us to build a unit comparable to the Latin American region in terms of energy indexes. The authors emphasize that this approach goes a step further than the panel data analysis by avoiding the shortcoming of pooling countries side by side, regardless of whether they have similar or radically different characteristics. Even after controlling for such differences, the regression approach is not clear about the relative contribution of each comparison unit.

10

Synth is a data-driven procedure that allows us to construct a comparison unit as a weighted average from the available comparison countries. That is, since it is often difficult to find a single country that approximates the most relevant characteristics of Latin American countries, this procedure allows for combining countries in order to provide a better comparison unit. The advantages of this method are: (i) as a data-driven procedure, this method reduces discretion in the choice of peers, forcing researchers to demonstrate affinities between the comparison units; (ii) it makes explicit the weights used to build the comparison unit; and (iii) because the weights can be restricted to be positive and sum to one, this method provides a safeguard against extrapolation. Further, Abadie, Diamond, and Hainmueller (2010) demonstrate that the conditions of Synth are more general than the conditions under which linear panel data or difference-in-differences estimators are valid. That is, it generalizes the traditional fixed effects model by allowing the effects of unobserved, confounding characteristics to vary over time. As described in Abadie, Diamond, and Hainmueller (2010; 2011), Synth could be applied when multiple units are exposed to an intervention; as for example, the evaluation of policies in states or countries. In particular, our strategy considers the characteristics of the average Latin American country to build a convex combination of non-Latin American countries with similar characteristics, and provides equal weights to each country to avoid overrepresenting a given country. The average is taken because three countries (Brazil, Mexico, and Argentina) represent more than 60 percent of the GDP and the total energy consumption in the LAC region (see figure 4). Thus, searching for a synthetic of the aggregate LAC –instead of the average– region would over-represent the biggest economies. The selection of the characteristics (or predictors) by which the comparison unit is chosen is usually based on literature standards. The validity of the predictors is a key factor for the validity of this method. In our context, this requires to identify the determinants of the energy indexes. This exercise was performed in the previous section when selecting the set of variables in 𝑋. The panel data estimation also provides some insights into the relevance of each variable and the final variables to be considered as predictors (see equation 4). Following Abadie and Gardeazabal (2003), 𝑋𝜏 represents the matrix of predictors by countries which is partitioned into 𝑋1 and 𝑋0 . The problem is minimize (𝑋1 − 𝑋0 𝑊)′𝑉(𝑋1 − 11

𝑋0 𝑊) subject to 𝑤𝑗 ≥ 0 and ∑𝑗 𝑤𝑗 =1; in order to find the optimal vector of weight (𝑊 ∗ ). 𝜏 refers to the condition (𝜏 = 1; 0) to be evaluated and 𝑗 indicates each country. Solving this problem allows finding a comparison unit only if 𝑋1 lies on the predictors’ support, avoiding extrapolations. The comparison unit takes the form of 𝑋0 𝑊 ∗ (≈ 𝑋1 ) and is called the synthetic control. 𝑉 represents a diagonal matrix, whose elements reflect the importance of each predictor. Following Abadie, Diamond, and Hainmueller (2011), an optimal choice of 𝑉 assigns weights that minimize the mean square error of the synthetic control estimator—that is, the expectation of (𝑋1 − 𝑋0 𝑊)′(𝑋1 − 𝑋0 𝑊). Note that 𝜏 has a time dimension connotation, for example the occurrence of an event in a sub-set of countries in a given year. Then, Synth would choose a control group based on preevent characteristics, and to attribute post-outcome differences only to the occurrence of that event. Our strategy does not have such a source of temporal variability, but only the distinction between Latin American and non-Latin American countries. This means that we must choose a year in which we assume an event occurs. This arbitrary decision makes the results potentially sensitive to the year chosen. The results could also be sensitive to the time window in which we restrict the algorithm to match the predictors—that is, changing the time window in which we match the predictors could change the gap between Latin America and its synthetic control. This would occur because each possible window would return a different set of comparison countries and/or weights. To address this problem, we apply Synth recursively, which allows us to capture the average gap-trend of Latin American energy indexes, taking into account different time windows or periods. We use the three following strategies to choose the time windows: a)

Enlarging matching periods, where the windows are chosen from (1972) to (1972 +

p), with 𝑝𝜖[3 (3)27]. The cut-off is given by (1972+p). The weights of the sets of countries that resemble Latin America are estimated for each period. That is, the synthetic is constructed in the period before each cut-off and the energy indexes are evaluated after each cut-off. This strategy allows for the introduction of more memory each time, starting from early 1970s, the window in which we look for a synthetic LAC gradually increases until a maximum of 27 years (1972-1999).

12

b)

Reducing matching periods, with windows from year (1972 + q) to (1999 ), with

𝑞𝜖[0 (3) 24]. Here, we reduce the window by starting the matching exercise from a later year each time. This strategy gradually reduces the period of years with which the average synthetic LAC is constructed. Each period allows to find a counterfactual for more recent LAC characteristics (from the first window 1972-1999 to the last window 1997-1999). c)

Moving matching periods with windows chosen from year (1972 + r) to (1981 + r),

with 𝑟𝜖[0(6)18]. Each window has nine years to construct a synthetic average Latin American country. Under this strategy the cut-off is given by (1981 + 𝑟). The windows move every 6 years ( 𝑟 ) until the matching period of (1990 − 1999) . This strategy captures a set of counterfactuals representative of the characteristics of the average Latin American country in a given period. We arbitrarily choose values for 𝑝, 𝑞, and 𝑟. In all cases, the time windows extend until the year 1999, which gives us 11 years to perform the comparison exercise. To summarize the estimations we average the results of the recursions by strategy. Section 3.3 presents these results. Note that under this approach, the pool of countries and weights used to construct the synthetic counterfactual could change depending on the period analyzed. An advantage of this is that allows to construct a synthetic LAC not only in terms of its characteristic in a given period, but also in terms of its characteristic along different periods, capturing the changes that the region has experienced.

3. EMPIRICAL RESULTS This section presents the main results of the strategies previously described. Annex 1 provides details about the data used for the exercise. 3.1. Energy Intensity Trends Figure 1 presents the trends in energy indexes, contrasting the Latin American region with others income level set of countries. In accordance with previous literature it shows that energy intensity has decreased in all regions mainly led by the efficiency effect. In general, the activity effect has less impact for all income levels; however, it is notoriously more relevant in high-income countries, especially those belonging to the Organization for Economic Co13

operation and Development (OECD). This structural component contributes to a 10 percent decrease in energy intensity in the case of high-income countries. In contrast, the structural effects of medium-income countries contribute to an 8 percent increase in energy use. We observed that all income classifications, with the exception of those in Latin America, consistently reduced energy intensity (and efficiency) by between 40 and 54 percent during this period. The literature of convergence in energy intensity has already identified this peculiar behavior, whereby the differences in intensity levels within a region have tended to decrease over the last four decades, except in Latin American countries (Liddle, 2010; Duro and Padilla, 2011; IEA, 2012; Mulder and de Groot, 2012). Figure 1: Energy Intensity Decomposition

1

High Income Countries

LAC Countries

Middle Income Countries

.8

Activity

.6

Efficiency

.4

Intensity

1970

1980

1990

2000

2010 1970

1980

1990

2000

2010 1970

1980

1990

2000

2010

Source: Authors’ elaboration. Note: LAC = Latin American countries.

In the case of Latin American countries, we have observed a 17 percent decrease in energy intensity over the last 40 years. During the 1970s, the intensity decreased by about 8.5 percent; between 1980 and 2000, it slightly increased, showing great volatility; and between 2000 and 2010, energy intensity decreased by another 10.6 percent (with respect to the 1970 level). In general, the efficiency effect explains all the changes, while the activity effect remained almost invariable. Table 1 shows the additive contribution of each energy index to the change in energy intensity. 14

Table 1: Variation Explained by Each Energy index (Base year 1971=100) Year

Intensity

1980 1990 2000 2010 1980 1990 2000 2010

Activity

Efficiency

Intensity

Activity

Efficiency

Latin American countries -8.53 2.23 -2.09 1.31 -6.57 -2.66 -17.18 0.78

-10.76 -3.40 -3.91 -17.96

Medium-income countries -9.88 4.69 -20.62 4.96 -42.46 5.08 -54.22 6.89

-14.57 -25.58 -47.54 -61.11

Low-income countries -12.88 -0.44 -24.11 -0.31 -29.19 -0.30 -42.83 1.28

-12.45 -23.80 -28.89 -44.12

High-income countries -15.70 -0.60 -35.31 -1.97 -42.15 -4.73 -50.45 -5.96

-15.10 -33.34 -37.42 -44.49

Source: Authors’ elaboration.

We should be careful in interpreting these results, since they hide great heterogeneity at the country level. For example, figure 2 presents decomposition trends by Latin American countries, showing the different paths and variances inside the region. In particular, we observe that during the period 1980–2000, LAC region in aggregate was mainly influenced by similar trends in Argentina, Brazil, Mexico, and Venezuela. That is, those aggregate measures tend to over-represent the biggest economies. This is a result of the relative weight of those economies in terms of GDP and energy consumption. Note that the four countries mentioned account for 76 percent of those variables (see figure 3). Besides, Latin America is one of the least energy-intensive regions in the world. By income classification, it is mainly composed of middle-income countries that, on average, use 165 kg of oil equivalent per US$1,000 GDP (at constant 2005 PPP), just above the high-income group and far below middle- and low-income countries (see Figure 4). This fact could suggest that, even for its stage of development, Latin America can be characterized as a low-intensity region. However, despite its absolute ratio of energy intensity, the economic dynamics of the region raise questions about its performance having into account its specific characteristics. These facts illustrates the importance of finding a set of similar countries to perform appropriate benchmark, as well as to perform an analysis of the average Latin American country from the final indexes and predictors. Previously to address this task, it is necessary to identify those specific characteristics which are relevant to define the trends presented.

15

CHL

80

100

COL

60 150 100 50 120

PRY

100

60

60

40

60

80 80

JAM

80

80

100

100

200

40

80 100 80 100

PER

80

120

HND

60

PAN 100

NIC

120

70 100 120

DOM

40 60 80

0

60

50

80

100

150

CUB

MEX

120

90

90 80

150 100

90 100

CRI

140

BRA 100 110

250

BOL

200

100 110 120

ARG

100 110

Figure 2: Latin America Energy Indexes

150

80

Intensity Efficiency

60

200

100

Activity mix

40

100

60 40

1970 1980 1990 2000 2010

VEN 200

URY 100

400

TTO

300

SLV

80

100 120

1970 1980 1990 2000 2010

1970 1980 1990 2000 2010

1970 1980 1990 2000 2010

Source: Authors’ elaboration.

16

1970 1980 1990 2000 2010

Figure 3: Relative Weight of Latin American and Caribbean Countries in GDP and Energy Use (Average GDP and TFC, 2000–2010)

Brazil

Mexico

Argentina

Colombia

Venezuela

Other LAC

TFC Kg oil equivalent, %

GDP PPP 2005, % 18%

18% 35%

5%

36% 8%

6%

5%

9%

11% 22%

27%

Source: Authors’ elaboration. Notes: Other LAC includes Chile, Peru, Cuba, Dominican Republic, Costa Rica, El Salvador, Bolivia, Uruguay, Trinidad and Tobago, Panama, Jamaica, Honduras, Paraguay, and Nicaragua. Gross domestic product (GDP); Total Final Energy Consumption (TFC)

Figure 4: Energy Intensity and Income, 2010 Energy Intensity and Income

300 200

NIC VEN PRY HND BOL

100

Energy Intensity*

400

500

Average Intensity by Income Classification

JAM SLV

ARG BRAMEX CHL URY CRI DOM

COL PER PAN

LIC

MIC

LAC

HIC

0

Income Classification

10000

20000

30000

40000

50000

Income percapita, constant 2005 PPP

*Kg of oil equivalent per US$1,000 GDP at constant 2005 PPP

Source: Authors’ elaboration. Notes: LIC = low-income countries; MIC = medium-income countries; LAC = Latin American Countries; HIC = high-income countries.

3.2. Determinants of Energy Intensity This section applies the regression analysis described in eq. (4) to identify the main drivers of the energy indexes. The exercise takes advantage of the entire available sample – 75 countries between 1971 and 2010– in order to increase the power of the results. Table 2 shows the results 17

of the three regressions for each index previously calculated (intensity, efficiency and activity). We start by noting that the lag of the energy indexes as an explanatory variable is always statistically significant, through a dynamic panel specification. This intuitively supports the argument that energy indexes do not respond immediately to changes in economic variables, although these have effects that materialize over time. Income is also statistically significant, both at level and its square, suggesting some degree of concavity, as expected from Figure 4. Intuitively, energy intensity declines as income increases, but at a decreasing rate. On the other hand, real petroleum prices have a significant influence on increasing efficiency and reducing intensity. This suggests that increasing petroleum prices over the last two decades have been a strong incentive for improving energy use, as shown in the previous section. The energy mix is closely related to energy intensity and efficiency, but not to the activity component. This suggests that countries that consume a higher proportion of fossil fuels in terms of total energy consumption tend to be more energy-intensive and less efficient. Specifically, keeping everything else constant, a 1 percent increase in fossil fuel consumption is often related to an increase in both the intensity and the efficiency indexes by an estimated 0.14 percent. The rents from natural resources and the GDP growth rate are both relevant in explaining the variability of the energy indexes. The former tend to increase intensity and reduce efficiency, without a strong relationship with activity. The economic growth rate tends to reduce energy intensity and increase efficiency, probably by increasing the use of fixed assets oriented toward production, converging to an optimal point of energy use. Table 2 also presents the income and price elasticities. An increase of 1 percent in per capita income tends to reduce the intensity and efficiency indexes by about 1.9 and 1.7 percent respectively. Equivalently, a 1 percent increase in real petroleum prices tends to reduce intensity and efficiency by 0.05 and 0.04 percent, respectively. The low impact of energy prices is notable, probably because we use international petroleum prices instead of energy prices, which could lead to some bias. In general, energy tariffs have some subsidies, so variations in international energy prices do not correspond exactly across countries. In this sense, we could be underestimating the conditional correlations between prices and energy indexes. 18

Table 2: Energy Indexes Regressions ln(intensity) Adjustment parameter

ln(efficiency)

ln(activity)

0.764**

0.734**

0.670**

(0.0183)

(0.0267)

(0.0398)

-0.549**

-0.515**

-0.0846*

(0.109)

(0.118)

(0.0436)

0.0266**

0.0224**

0.00658**

(0.00654)

(0.00729)

(0.00308)

-0.0123**

-0.0109**

-0.00112

(0.00307)

(0.00365)

(0.00117)

0.226

0.363

-0.150

(0.240)

(0.309)

(0.115)

0.179**

0.172**

0.00833

(0.0416)

(0.0683)

(0.0435)

0.139**

0.142**

0.00170

(0.0477)

(0.0523)

(0.0123)

0.00456

0.000626

0.00367

(0.0228)

(0.0184)

(0.00602)

-0.466**

-0.447**

-0.00838

(0.0411)

(0.0495)

(0.0226)

Country fixed effect

Yes

Yes

Yes

Trend effect

Yes

Yes

Yes

3.76**

3.89**

1.77**

(0.503)

(0.558)

(0.286)

Income elasticity

-1.898

-1.714

0.036

Price elasticity

-0.052

-0.041

-0.003

Observations

2845

2845

2845

Adjusted R-squared

0.940

0.923

0.781

ln(GDP per capita, constant 2000 PPP) ln(GDP per capita sq., constant 2000 PPP) ln(petroleum prices) Population growth (%) Natural resources rents (%) Fossil fuel energy consumption (%) Investment/capital ratio (%) GDP growth rate

Constant

Source: Authors’ elaboration. Notes: Standard errors in parentheses * p