Energy, Efficiency Gains and Economic Development: When Will Global Energy Demand Saturate?

Not anytime soon. Using a novel dataset covering 127 countries and spanning two centuries, we find evidence for an energy Kuznets curve, with an initial decline of energy demand at low levels of per capita income followed by stages of acceleration and then saturation at high-income levels. Historical trends in energy efficiency have reduced energy demand, globally, by about 1.2 percent per year and have, thus, helped bring forward a plateau in energy demand for high income countries. At middle incomes energy and income move in lockstep. The decline in the manufacturing share of value added, globally, accounted for about 0.2 percentage points of the energy efficiency gains. At the country level, the decline (rise) of the manufacturing sector has reduced (increased) US (China) energy demand by 4.1 (10.7) percent between 1990 and 2017.

Abstract

Not anytime soon. Using a novel dataset covering 127 countries and spanning two centuries, we find evidence for an energy Kuznets curve, with an initial decline of energy demand at low levels of per capita income followed by stages of acceleration and then saturation at high-income levels. Historical trends in energy efficiency have reduced energy demand, globally, by about 1.2 percent per year and have, thus, helped bring forward a plateau in energy demand for high income countries. At middle incomes energy and income move in lockstep. The decline in the manufacturing share of value added, globally, accounted for about 0.2 percentage points of the energy efficiency gains. At the country level, the decline (rise) of the manufacturing sector has reduced (increased) US (China) energy demand by 4.1 (10.7) percent between 1990 and 2017.

I. Introduction

Will global economic growth endanger our planet’s capacity to sustain life? Or will higher incomes and human ingenuity be able to preserve it? These questions are particularly relevant today as the debate on climate change mitigation policies intensifies while rising income standards, especially in developing countries, are accompanied by an ever-increasing demand for energy services, which in the early stages of development is frequently satisfied by burning carbon-intensive fossil fuels (i.e., coal, oil, and natural gas). If income and energy are tightly connected then to reduce primary energy’s consumption we may have to forfeit income growth. 2

An analog question—whether environmental concerns could pose limits to growth—was already posed in the context of local pollution more than two decades ago (Grossmann and Krueger 1995 and Stokey 1998). In that context, an inverted U-shaped relation between income and pollution was found which became known as the environmental Kuznets curve (EKC), see Selden and Song (1994).3 The empirical evidence for the EKC, however, was later disputed (Harbaugh et al. 2002). Thus, the facts have not unequivocally supported an EKC for local pollutants so far.

Today, however, the most concerning energy externalities are perhaps global (not local) and their effects are less visible and to be felt more prominently by future generations. Stokey (1998)’s rationalization may, thus, not apply to energy emissions. Moreover, even though a transition from high-carbon to low-carbon energy sources will be necessary to reduce carbon emissions, a substantial change in the energy mix at the global level will probably take decades. It is, thus, worth abstracting from the energy mix and ask: what is the historical relationship between economic development and primary energy demand? 4 Since some high-income countries have seen their energy demand plateau in recent years, it begs the question of whether this phenomenon can be generalized and whether it can help predict a slowdown in energy demand as emerging markets grow. Finally, a better understanding of the relationship between levels of income and energy can help assess the feasibility of climate change targets (such as carbon emissions over GDP) and develop and calibrate integrated assessment models (IAMs) that are used to examine climate policies. 5 Following Grossman and Krueger (1995), we postulate and test the existence of an energy Kuznets curve that allows for a peak in primary energy sources as income grows. This relationship is proxied by a cubic polynomial in GDP per capita which is the lowest-order polynomial that enables us to capture both a peak in energy demand (i.e. a static saturation point) and an inflexion point where the income elasticity of energy demand is the highest. This reduced-form approach has been favored to structural models (see Bhattacharyya 1996 and 2019 for a survey) as being more flexible (i.e., imposing less assumptions), but it requires a sufficiently long history and wide cross-sectional variation. Our dataset starts in 1850 and includes 127 countries in total across various stages of development.

To identify savings in energy consumption beyond and above those associated with income growth, following Griffin and Schulman (2005) and van Benthem and Romani (2009), we add year fixed effects to our empirical specification, and also control for the share of the manufacturing sector in value added. Year fixed effects proxy for omitted variables that move similarly across all countries, especially technology when focusing on low frequencies. Energy saving technologies, by reducing energy demand regardless of the country-specific income level, induce a downward shift, over time, of the energy Kuznets curve. Disentangling the effect of income and technology on energy demand is useful and important, especially in relation to policies in support of R&D. Utilizing the trend of the year fixed effects to calculate the rate of energy efficiency gains we introduce the concept of dynamic saturation: the level of income at which income growth is perfectly offset by energy efficiency growth, leaving energy demand unchanged, ceteris paribus.

Results strongly support the presence of a Kuznets relationship between log (per capita) energy and log (per capita) income. This contrasts with results in the literature that find or assume a constant elasticity of energy use with respect to income (Liddle and Huntington 2020, Huntington et al. 2019, Burke et al. 2016, and Csereklyei et al. 2016, among others). Medlock and Soligo (2001) and van Benthem and Romani (2009) found a quadratic relation between energy and income, but the concavity of the relation is unstable because a quadratic relation cannot capture the Kuznets curve’s inflexion point where concavity flips. So, the estimates of the concavity depend on whether the sample consists mostly of low-income countries or high-income countries. 6

The elasticity we find, instead, is low and increasing at low-income levels, peaks at middle-income levels, and declines afterwards. The inflexion point in the energy-income relationship (i.e., the maximum income elasticity) is at about $10,000 (2011 USD) below the global per capita income in 2015 of $15,000. At that income level the energy’s income elasticity is about 1. The peak of the Kuznets curve (i.e., the income level that, statically, determines energy saturation) is high, above $107,000. The Kuznets’ peak, however, is not precisely estimated because there are not enough observations at very high-income levels. This problem has induced some studies to prefer a linear or quadratic specification to a cubic one (Liddle and Huntington 2020). Instead, this is where the usefulness of the dynamic saturation concept becomes apparent. Since energy efficiency gains have been substantial over the sample, estimated at about 1.2 percent per year, the income necessary to reach dynamic saturation is about $43,000, which is already reached by 15 countries in North America, Europe, and the Asia-Pacific region.7 Indeed, Monte Carlo simulations reveal sharp differences in variance and skewness between the two distributions of estimates for the static and dynamic saturation point, with the 90 percent confidence interval of the former ($81,170–162,250) being much wider than that of the latter ($32,970–59,730).

Thus, dynamic saturation, a novelty in the literature, can offer a more satisfactory explanation of the energy consumption plateaus observed for advanced economies than earlier work and helps reconcile conflicting results from previous studies. Csereklyei et al. (2016), for example, argues that there is no energy-income decoupling yet and that energy demand (in some sectors of the economy) will continue to increase into the foreseeable future. Van Benthem (2015), instead, estimates that saturation occurs around an income level of $35,000, a level already surpassed by many economies. Absolute decoupling of energy and income growth in the near future, however, does not seem plausible on theoretical grounds (for example, service-sector-led economic growth still requires energy), as argued by e.g., Parrique et al. (2019) and Smil (2017, 2019). It also cannot explain the tendency for energy consumption, despite its negative trend, to bounce back in years with above-average income growth. Our estimates show that at $107,000 per capita absolute energy-income decoupling is still far away. So rather than relying on a negative (or zero) income elasticity, dynamic saturation depends on a combination of factors that allow for a relative decoupling—i.e., a declining but still positive income elasticity and continuous energy efficiency gains that bring about energy saturation at high levels of income.

Historically, efficiency gains have, thus, helped to slow down energy demand globally, thereby reducing the magnitude of the climate change mitigation challenge. An important implication of dynamic saturation is that further declines in energy consumption are by no means guaranteed and will depend on future energy efficiency gains to continue to outpace the effect of income growth on energy consumption. Future efficiency gains seem, thus, critical in the fight against climate change, as global energy demand is unlikely to saturate anytime soon because the income elasticity of energy demand is still close to one for emerging markets.

This paper also contributes to the literature by highlighting the important role played by manufacturing in reducing energy consumption. It finds that while manufacturing offshoring and the global decline of manufacturing are quantitatively relevant forces, neither can wholly explain the decline in energy consumption in advanced economies. Jakob et al. (2012) argue that the decoupling of income growth and energy consumption for advanced economies is in part explained by the weaker relationship between economic growth and manufacturing at higher income levels. Cohen et al. (2017) show that once the impact of trade (and thus manufacturing offshoring, among other factors) is considered, the evidence for a decoupling of growth and emissions in advanced economies becomes weaker. This amounts to comparing the income elasticity of emissions consumption to the income elasticity of emissions production. In contrast to these papers, our work quantifies how the global manufacturing decline lowers the level of income at which countries dynamically saturate, and to what extent saturation is brought forward in AEs and delayed in EMs due to the offshoring of manufacturing.

Quantitatively, our results indicate that the global decline in the manufacturing share has reduced global energy demand by about 8.7 percent between 1971 and 2017. This contribution has been stronger for AEs, as manufacturing activities have relocated to EMs. In absence of manufacturing relocation, energy demand in the US would have been 4.1 percent higher, while energy demand in China would have been 10.7 percent lower. Looking forward, however, the scope of both the relocation and reduction of the manufacturing share seems more limited

Our analysis complements existing papers by including novel (historical) data from the following periods: the post global financial crisis era (2009–2017), the golden era of capitalism (1948–1970), the first period of globalization (1850–1913) and the interwar period (1919–1938). We find strong evidence for the Kuznets relationship between income and energy in all of our three historical panels. The point estimates of dynamic saturation from the historical panels are typically larger, at around $68,000 (67,000) for the panel 1850–2017 (1900–2017), though it should be noted they contain fewer countries. Results point to larger efficiency gains in the order of 1.6 (1.3) percent per year for the panel 1850–2017 (1900–2017) compared to 1.2 percent for the period 1971–2017, perhaps indicating that (some of) the low-hanging energy efficiency fruits have already been harvested by past innovators (Gordon 2012, Bloom et al. 2020).

We also perform a battery of robustness tests. Results are largely unchanged when we exclude one-by-one countries from our main sample but for two exceptions: efficiency gains are smaller (larger) when US (China) is excluded. Adding energy exporters and transition economies to our main sample one at a time does not overthrow the Kuznets relationship between energy and income. It does point, however, to different energy-income dynamics for energy exporters while transition economies introduce a break in efficiency gains in early 1990s. We also tested for a partial adjustment mechanism (see Medlock and Soligo (2001), van Benthem and Romani (2009), van Benthem (2015)), but it does not alter the main results.

Finally, the baseline econometric analysis purposely omits energy prices. In fact, while energy prices (proxied by real oil prices) did fluctuate globally during the sample period, such fluctuations, and their effects on energy demand, are captured by the year fixed effects—this is corroborated by the positive correlation between oil prices and year-fixed effects’ deviations from their trend. The omission of energy prices, however, should not affect the estimated trend in efficiency gains because, during the sample period 1971–2017, energy prices do not show any trend. Level differences in energy prices across countries are captured by isolating energy exporters or introducing country fixed effects.

The paper is structured as follows. In the next section we present our conceptual framework and motivating stylized facts about global energy demand and manufacturing. Section III presents our empirical approach to study the relation between energy consumption and income per capita. Our specification allows us to test for static and dynamic saturation points in the relation between energy consumption and income. Section IV describes the construction of our energy consumption variable and provides data sources and descriptive statistics for our main panel as well as three novel historical data panels. Section V presents empirical results for our main panel. Section VI shows that qualitatively results hold in a battery of robustness tests; in addition, it repeats the empirical analysis for our three historical panels. Section VII concludes. Additional robustness checks and detailed information on the construction of the historical dataset are reported in several appendices.

II. Conceptual Framework and Stylized Facts

Energy and Durable Goods

Energy is an essential input in economic systems as most durable goods (i.e., their services) require energy to function.8 Energy and durable goods service’s consumption are, thus, complements for consumers and producers alike. Hence, for given methods of production, increases in energy demand are driven either by an increase in the number of households and producers that have access to durable goods (i.e., extensive margin) or by a more intensive use of them (i.e., intensive margin). As a country reaches higher incomes, however, both the intensive and extensive margins of durable goods services’ consumption can reach a “satiation point” (or offset each other); the associated income level, in turn, is the income that determines a saturation in energy demand. The implication is that, beyond business cycle frequencies, energy demand decouples from income growth and movements in energy prices.

This idea has already been explored in the literature. Farrell (1954) and Bonus (1973) derived an S-shaped logit curve for appliance ownership and, more recently, models in Gertler et al. (2011) and (2016) predict an S-shaped relation between income and electricity demand and household durable good purchases (such as domestic appliances and automobiles). Similarly, Dargay, Gately, and Sommer (2007) and Dargay and Gately (1999) find it for car ownership.9 At the aggregate level, it seems plausible that these relations combined would show a maximum elasticity of energy at middle-income levels and a saturation point beyond which energy and income, seemingly, decouple—i.e., they would imply an energy Kuznets curve.

Energy and Structural Change

The steady increase in global energy demand from 1971 to 2017 by 150 percent (more than 3 percent per year) masks stark dissimilarity between AEs and EMs (Figure 1). The latter, especially China and more recently India, have driven most of the energy growth of the last 15 years while the former’s contribution has been minimal leading to a decline in AEs world energy consumption shares (Wolfram et al 2012). Some high-income countries seem to have already peaked in terms of per capita and even total energy demand (chart). However, this is not necessarily a proof of saturation, as previously defined, since we must control for per capita income growth and structural change such as changes in the methods of production (i.e., energy efficiency gains) and, for advanced economies, the decline of manufacturing in favor of services.

Figure 1
Figure 1

Energy Consumption and GDP per capita (1850–2017)

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: ktoe = kilotonnes of oil equivalent. For specific sample years by country please see list of country groups in Historical Data Appendix.

Indeed, industrial energy demand, which is a complement to services from capital goods, increases rapidly during the early industrialization stage when energy-intensive manufacturing industries expand. As economies develop, the service sector becomes more relevant while the manufacturing sector declines as share of value added and, especially, as share of total employment—given also its relatively higher productivity growth. In addition, since at least the 90s the world has seen a substantial reallocation of manufacturing to EMs.

In AEs, the manufacturing share of value added (measured at current prices) has declined steadily since the 1970s by about 10 percentage points. In part this is due to the decline in the relative price of manufacturing goods (i.e., a relative increase in productivity). Globally, however, the manufacturing share has declined less and only till the beginning of the century as energy-intensive industries moved or re-located from AEs to EMs, especially China (Acemoglu et al 2016, Baily and Bosworth 2014, Fort, Pierce, Schott 2018) (Figure 2). Since the manufacturing sector is relatively energy intensive, these trends probably explain part of the decline in energy per capita observed in advanced economies and the fast energy demand growth of emerging markets.

Figure 2
Figure 2

Global Manufacturing Share

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: Global manufacturing share of value added (measured at current prices) derived using energy consumption weights by income group. Income group based on World Bank categorization.

We, thus, expect a Kuznets curve (i.e., a S-shaped curve) governing the relation between the stage of development and energy demand, like the one found by Grossman and Krueger (1994) for the environment, that is shifted over time by structural changes. Energy saving technologies, in fact, can anticipate actual saturation by shifting down the energy-income curve because the same economic activities (such as heating, cooling, transports, …) require less energy while economies with a higher manufacturing share require relatively more energy.10 Energy efficiency gains can come about through various mechanisms, including increased thermal efficiency of fossil fuel power plants and improved fuel economy of internal combustion engines. The next sections will offer a quantification of the Kuznets curve.

III. Empirical Analysis

Using an unbalanced panel dataset of low-, medium-, and high-income countries, we study the relation between energy demand and income, isolating the role of the manufacturing sector and controlling for population density (i.e., population and land area). Country fixed effects are introduced for robustness to control for country-specific factors that could explain different energy needs (such as latitude). Furthermore, we use time dummies to capture energy efficiency gains, fluctuations in global energy demand that are not fully captured by income, and movements in energy prices. However, since the relative price of energy is mean reverting in our sample, we interpret the trend component of the time dummies as energy efficiency.11 12

Our empirical approach generalizes earlier contributions to the literature on energy demand, including Medlock and Soligo (2001) and van Benthem (2015) and it follows the seminal work of Grossman and Krueger (1995). 13 In particular, we estimate the following specification relating (log of) primary energy demand E to (log of) population, pop, a 3rd order polynomial in (log) income per capita, y, the manufacturing share and a set of control variables:

Eit=β0+βppopit+β1yit+β2(yit)2+β3(yit)3+λt+βAAi+βmmit+εit(1)

where λt are year fixed effects, At is a set of control variables including country fixed effects or (log) land area; mit is a measure of the manufacturing share in value added (measured at current prices), and the indices i and t refer to countries and years, respectively. In the baseline regression, the manufacturing share is in deviation from the global share and orthogonalized with respect to the rest of the regressors. To gain insights into its role, the manufacturing share is also introduced just in deviation from the global share and as is.

The IEA baseline sample is annual and spans from 1971 to 2017 covering 97 countries (excluding energy exporters and transition economies) and two major energy price cycles. The historical dataset starts in 1850, but it features fewer countries (its results are presented in Section VI). In the Robustness section we introduce energy exporters and transition economies one by one to our main sample of 97 countries.

Since small countries can have very specific energy issues (especially in terms of initial energy endowment or energy needs) we rely on weighted least squares (WLS) as it is more unlikely that these idiosyncrasies play a major role in large countries. Also, large countries will drive future global emissions and energy demand, by giving large countries more weight we can speak more directly to climate-related questions. WLS are preferred to country fixed effects (FE) because the latter do not let us exploit persistent cross-sectional income differences. The combined used of WLS and FE is, instead, problematic (see Section IV).

The above specification allows us to test for a peak (saturation), a trough (graduation), and an inflexion point (max elasticity) in the relationship between energy and income. The (log) per capita income levels that define static saturation and graduation from biomass, yk with k = s, g, where s (g) stands for saturation (graduation), solves the following quadratic equation

δ1+2δ2yk+3yk2=0,

where δi = βi3 for i = 1, 2 and β3 * 0. The inflexion point (i.e., maximum elasticity) is given by -δ2/3.14 The special case of β2 = β3 = 0 and β1 = βp corresponds to the case where log energy demand scales up proportionally to log GDP. The (log) saturation and graduation points are defined as

yk=(δ2+/δ223δ1)/3(2)

The (log) dynamic saturation point, z, is calculated by total differencing equation (1) with respect to time and, thus, solves the following equation

(β1+2β2z+2β3z2)y˙+λ˙+βpp˙+βmm˙=0

Where y˙ is income growth, λ˙ is the annual energy efficiency loss, m˙ is the change in the manufacturing share, and p˙ is population growth. Assuming a constant manufacturing share, m˙=0, the solution takes the same form of (2) but with a modified δ1

z=(δ2+δ223δ˜1)/3(3)

where

δ˜1def¯¯δ1+(λ˙+βpp˙)/(β3y˙).

If δ22>3δ1 (i.e., given a static saturation point), as far as λ˙<βpp˙andβ3y˙>0, dynamic saturation exists and is lower than its static counterpart, z < ys. The intuition is simple, energy saving (i.e., λ˙<0) shifts down the Kuznets curve while income growth represents the speed at which a point moves over the curve. When the two opposite forces offset each other, we have dynamic saturation in energy demand.15 In absence of energy savings and with constant population, static and dynamic saturation coincide. Income growth, however, is a “free parameter”. We discipline its choice by estimating a beta-convergence relation between growth and the level of income, y˙=a^+b^y, over the sample (Barro and Sala-i-Martin (2004)).16

The static saturation point may be hard to estimate as it probably implies an income level that few or no countries have reached yet. Since z < ys, however, it is more likely that some countries are at or close to the dynamic saturation point. We will exploit this consideration later and analyze the estimated covariances among coefficients that enter the saturation point’s calculation.

Data

For our main analysis we constructed an annual panel from 1971 to 2017. This panel is slightly unbalanced, with the number of countries increasing from 88 in 1971 to 97 in 2008 and thereafter.17 In addition, we constructed three historical balanced panels: the first and the longest, Group 1, covers 1850 to 2017 (11 countries), Group 2 covers 1900 to 2017 (31 countries), and Group 3 covers 1950 to 2017 (65 countries). By digitalizing and bringing together various historical data sources, our main data contribution is to expand the number of observations in the period 1850–1949 and 1950–1970, which respectively contain 2126 and 2086 observations. To the best of our knowledge, our historical energy panel dataset contains significantly more observations than previous work in the literature.18

Throughout the paper, energy demand reflects the total primary energy supply (TPES), which measures the total amount of energy that is made available to domestic end-users in the economy (in million tonnes of oil equivalent, MTOE). In contrast to total final consumption (TFC, used, for example, in van Benthem 2015 and Csereklyei et al. 2016), TPES captures energy efficiency gains along the entire supply chain due to its upstream character.19 TPES is also more pertinent than TFC when studying environmental and climate related questions.

The global coverage of energy consumption, GDP per capita, population, and the manufacturing share increases slowly and steadily between 1850 and 1949, and then expands greatly from 1950 onwards. Manufacturing data is an exception and doesn’t cover much more countries until the 1970s, a jump that coincides with the publication of the UN national accounts data.

Table 1 presents descriptive statistics for the four panels. The raw data shows, perhaps not surprisingly, that the range of values as well as the standard deviation for energy consumption per capita, GDP per capita, and the manufacturing share are larger or at least as large in our main panel of 97 countries than in our historical panels. The standard deviation of GDP and the range of values for energy consumption growth (log-diff) are smaller in Group 1 and increase with the extension of the cross-section, highlighting its importance. The relative importance of the cross-section compared to the time dimension, can in part be attributed to the fact that the United States and many other advanced economies had already reached (upper-)middle-income status between 1850–1900. Also, members of Group 1 show a high degree of co-movement.

Table 1:

Descriptive Statistics

article image
Sources: Maddison; Mitchell; World Bank; United Nations; and IMF staff calculationsNote: See country table for further details on countries included in each group. Obs = observations, Std. Dev. = standard deviation. All variables in logarithmic scale except the manufacturing variables. Manufacturing Share (dev.) is in deviation from its global average, while Manufacturing Share (ort.) has been orthogonalized to a third order income p.c. polynomial.

IV. Empirical Results

Results strongly support the presence of a Kuznets curve where the energy demand elasticity to income increases at low-income levels, peaks at middle-income levels at around one, and declines toward zero at high-income levels. A constant income elasticity of energy demand is rejected by the data (Table 2 and Figure 3).20 This result is in contrast with other results in the literature that suggest or impose a constant income elasticity (such as Burke et al 2016 and Csereklyei et al 2016, among others).

Figure 3
Figure 3

Kuznets Curves

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: The figure shows demeaned Kuznets curves estimated in Table 2. Dashed blue and black lines are the linear and quadratic specifications, respectably. Dotted yellow and red lines are country fixed effects specifications using WLS and OLD, respectively. The solid grey line is our baseline specification while the solid black line is specification 6 where manufacturing is introduced as is. Black line is unitary elasticity.
Table 2:

Primary Energy Demand Determinants by Specification

article image
Sources: Maddison; Mitchell; World Bank; United Nations; and IMF staff calculationsNote: Based on years 1971–2017. Robust standard errors clustered at the country level in parentheses. All variables in logarithmic scale except “manufacturing share” and “year”. Average elasticity is calculated at 15,000 2011 international USD. All regressions have time fixed effects; fe = country fixed effects; ols = ordinary least squares; wls = weighted least squares. Manufacturing Share (dev.) is in deviation from its global average, while Manufacturing Share (ort.) has been orthogonalized to a third order income p.c. *p < 0.05, **p < 0.01, ***p < 0.001

In the baseline specification, (4), the manufacturing share has been orthogonalized to the rest of other explanatory variables. In this way, time dummies and income still subsume the role of manufacturing across time and stages of development while its coefficient correctly represents the semi-elasticity of energy demand to the manufacturing share (Table 2, compare specifications 4 and 6).

A trend line fits very well the evolution of time dummies supporting the view that their low frequency movement is mostly capturing a global shift in technologies and methods of production rather than common changes in energy prices, taxes or subsidies, regulation, or other global factors (Figure 4). Indeed, cyclical movements in energy efficiency—i.e. deviations from the trend line—are negatively correlated to (real) oil price levels and positively correlated to GDP growth (see Appendix C). Moreover, we reject the ADF unit root test for real oil prices at the 5 percent level in the 1971–2017 sample (real oil prices in 2017 are about as they were in 1975 and about equal to the 1971–1981 average). Hence, trend efficiency gains must represent the average adoption, across countries and over the medium term, of energy efficient technologies—controlling for the stage of economic development. These gains substantially reduced energy demand over the sample period, by about 1.2 percent per year in the baseline specification (Table 2, specification 4), and are strongly significant across specifications (except when country fixed effects are used, see discussion below).21 Comparing specification 5 and 6 we can identify the contribution of manufacturing to efficiency gains. The global decline in the manufacturing sector share has contributed about 0.2 percentage points to energy efficiency gains, mostly between the 1980s and 1990s (Table 2, Figure 2, and Figure 4).

Figure 4
Figure 4

Energy Efficiency

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: The blue line represents the time fixed effects estimated in regression (4), which is the baseline, with 95 percent confidence intervals (blue dashed lines). The black line is a linear trend estimated for the period 1971–1989 (1992–2017). Other time fixed effects shown map the specifications from Table 2; fe = country fixed effects; spec = specification.
Figure 5
Figure 5

The Energy Kuznets Curve

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: The figure shows 1000 Monte Carlo draws of demeaned Kuznets curves for specification 4 of Table 2. The dash red line is the point estimate Kuznet curve. Dot dashed line is the unitary elasticity. The dynamic saturation vertical bar is calculated using the average efficiency gains estimated for specification 4 and income growth as described in Section III.

A parsimonious linear specification can be appealing, however, the estimated income elasticity, 0.9, is lower than in the baseline specification because it must strike a balance between high- and medium-income countries’ elasticities (Table 2). In fact, for high-income countries, higher estimated elasticities would require unreasonably high efficiency gains to square the data (see Figure 7 upper panel). Adding a quadratic income term does not help, as it tries to capture both the increase in income elasticities at middle-income levels and the flattening of the relation at higher income levels. This struggle makes quadratic estimates not significant. The cubic specification, instead, solves the above problem increasing the income elasticity for middle-income countries while reducing it for the high income. In addition, it captures the initial decline of energy demand in response to income growth in most LICs which is, probably, the result of graduation from biomass—an inefficient source of energy.22

Figure 6
Figure 6

Distribution of Saturation Points Estimates; spec 4

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: The top (bottom) panel shows the distribution of 10,000 Monte Carlo draws of the static (dynamic) saturation point, from specification 4 (Table 2). The dash red line is the point estimate Kuznet curve. Dot dashed line is the unitary elasticity.
Figure 7
Figure 7

Estimation Uncertainty

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: The top (bottom) panel shows a scatter plot of 10,000 Monte Carlo draws of the average energy efficiency loss and the income elasticity (static saturation) point, from specification 1 (4), see Table 2.

The maximum income elasticity is precisely estimated and statistically slightly higher than 1 (Table 3). Under the baseline specification it is reached at about $10,000 (2011 USD) income per capita— below the 2015 global per capita income of $15000 and already reached by various emerging markets. Passed this income level the elasticity starts to decline, but slowly. The economy reaches energy saturation at about $107,000 (2011 USD). The point is far in the future for any country and imprecisely estimated. 23 The distribution of the static saturation point is also strongly skewed to the right, as shown by Monte Carlo simulations which suggest a 95 percent confidence interval between $81,000 to $162,000 (Figure 6 and Table 3).24 This result is not completely surprising since we are controlling for energy saving, but it poses an estimation problem because estimates mostly rely on out-of-sample projections of the Kuznets curve as no country has reached such income level.

Table 3:

Median Estimates and Confidence Intervals

article image
Note: Based on Monte Carlo simulation of 100,000 samples. The income elasticities for the quadratic and cubic specifications are calculated at $15000 per capita GDP. [ ] represent 10th and 90th percentile confidence intervals, repsectively. For the description of the specifications see Table 2.

The dynamic saturation point is less subject to this problem as z < ys (see Section III). Assuming income growth as laid out in the previous section, the median estimate of the dynamic saturation point is about $42,850, lower than the static saturation’s one, already passed by 15 economies including the US, Germany, Japan and Singapore, and more precisely estimated (see Table 3 and Figure 6). 25

The contribution of manufacturing to energy demand is substantial. A 1 percentage point (standard deviation) increase in the manufacturing share increases a country’s energy demand by 1.6 (6) percent. As noticed, the trend decline in global manufacturing covaries with time dummies. This suggests that some of the energy efficiency gains stem directly from productivity improvements in the manufacturing sector, globally. This average contribution is about 0.2 percentage points per year which is the difference between the estimates of energy efficiency in specification (6) and (5). The role for offshoring and relocating manufacturing production outside advanced economies is substantial. For example, it accounts for as much as 3.1 (4.1) percent reduction in energy demand in the US between 1990 and 2010 (1990–2017) while the relocation of manufacturing to China between 1990–2010 (1990–2017) accounts for a 14 (10.7) percent increase in energy consumption.

The role of manufacturing can also be seen comparing the Kuznets curve under specification (4) and (6). Since higher income levels are associated with lower manufacturing shares, when we control for the manufacturing share not only energy efficiency’s estimates are reduced but also the Kuznets curve is shifted to the right increasing both saturation points substantially.

Finally, not surprisingly, energy demand moves in lockstep with population as its coefficient is very close to 1 in most specifications. When it is substantially above one, it casts a doubt on the econometric specification used. The estimated coefficient on land area is significant and implies that when population density halves (i.e., the land area doubles for a given population) energy demand increases by about 10 percent (Table 2). Doubling both population and land area more than double energy demand, suggesting that “big countries” consume relatively more. Other variables (not shown), such as latitude or winter/summer temperature, are often not significant, possibly due to collinearity with income.26

The presence of population “inflates” the overall R-squared; however, even when estimated in per capita terms (having replaced land with population density) all results are retained and the R-square is about 0.94, suggesting that the baseline specification captures a substantial variation of per capita energy demand across time and countries.27

Country fixed effects flatten dramatically the Kuznets curve reducing the income elasticity and the static saturation point to the extent that energy efficiency gains are no longer needed to explain the observed slowing pace of energy demand in various advanced economies (Figure 3 and Table 2). The covariation of the estimates of the efficiency gain and the saturation point is suspiciously high (the higher the gains the higher the saturation, Figure 7 bottom panel) and the coefficient on population is questionable pointing to an identification problem introduced by the country fixed effects that is absent in the baseline specification. For example, a coefficient on population higher than 1 implies that the relatively higher per capita energy growth of emerging markets could be explained, erroneously, by their faster population growth—a 1 percentage point higher population growth would imply a 10 percent faster energy growth! The problem introduced by country fixed effects is in part due to the use of weighted least squares which further reduce the role of the cross-section as weights put the emphasis on a few big energy consumers. Indeed, in Specification 8, using OLS, results are more in line with the baseline specification. In the next section, we will analyze robustness and repeat the regressions on our novel historical dataset.

V. Historical Regressions and Robustnesst

A. Historical Regressions

Group 1 consists of 10 advanced economies (e.g., Belgium, Germany, US, …) that experienced a relatively similar development path and similar business cycles. The sample starts in 1850. The use of time dummies is, thus, problematic because it explains most of the energy variations. We have, thus, replaced time dummies with a time trend. Signs point to a Kuznets curve, but coefficients are significant only when country fixed effects are introduced. Group 2 has already enough cross-sectional variation to re-introduce time dummies, while the sample starts in 1900 and includes Japan, Italy, the USSR, and Brazil. Similarly, the sample of Group 3 starts in 1950 and includes China, Poland, Turkey and South Africa. In contrast to Group 1 and like our main panel, the coefficients on the polynomial are significant for Group 2 and 3 even without fixed effects.

Overall, the Kuznets curve is well-preserved in the Group 2 and Group 3 historical panels, and results point to strong efficiency gains over the sample starting after WWI (Figure 8). The post-WWII boom period of 1950–1970, which saw the rapid expansion of electricity generation and automobile transportation and the concomitant rise of oil consumption in advanced economies, appears to be an exception, as energy efficiency looks rather constant during this era of rapid growth (Figure 8).28

Figure 8
Figure 8

Historical Energy Efficiency Gains

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Energy Demand dataset, see Historical Data Appendix.Note: The blue line represents the time dummies estimated in regression (4), Table 2. The dotted line is the time trend of specification 3, group 1; the light gray and dashed lines are time dummies of specification 8 and 13, see Table Historical Regressions. Series have been rebased to year 1971.
Table 4:

Historical Regressions

article image
Sources: Buera and Kabobski; Maddison; Mitchell; World Bank; United Nations; and IMF staff calculationsNote: The sample is a balanced panel for 1850–2017 (Group 1), 1900–2017 (Group 2) and 1950–2017 (Group 3). Robust standard errors clustered at the country level in parentheses. All variables in logarithmic scale expect “manufacturing share” and “year”. Average elasticity is calculated at 15,000 2011 international USD. All regressions incorporate year fixed effects, except Group 1; fe = country fixed effects; ols = ordinary least squares; wls = weighted least squares. Manufacturing Share (ort.) has been orthogonalized to a third order income p.c. Please see appendix for country list. *p < 0.05, **p < 0.01, ***p < 0.001

For each historical panel, the dynamic saturation point is calculated based on the trend in the efficiency gains over the entire historical sample. The point estimates of dynamic saturation from the historical panels are typically larger, at around $68,000 (67,000) for the panel 1850–2017 (1900–2017), but since these panels contain fewer countries than our main panel, it is not straightforward to determine whether the Kuznets curve and its characteristics such as the dynamic saturation point have changed over time, and in what direction. Results point to larger efficiency gains in the order of 1.6 (1.3) percent per year for the panel 1850–2017 (1900–2017) compared to 1.2 percent for the period 1971–2017, perhaps indicating that (some of) the low-hanging energy efficiency fruits have already been harvested by past innovators (Gordon 2012, Bloom et al. 2020).

B. Robustness

In this section we perform a series of robustness tests.

Testing for outliers. We drop one country at the time to see identify potential outliers (Figure 9). Dropping the US (China) from the regression reduces (increases) the energy efficiency gains by about 0.4 (0.2) percentage point while the peak of the Kuznets curves decreases (increases) by $25,000 ($60,000). Since efficiency gains and static saturation move in opposite direction dynamic saturation is less affected. It is, thus, important to include both countries—especially when using weighted least squares. This is in part due to the opposing evolution of the manufacturing sector in the two countries. Indeed, the rise of China as manufacturing powerhouse increased her energy demand by [14] percent (see Section Results). Interestingly, omitting Ethiopia from the sample increases the saturation point, probably, because Ethiopia helps capture graduation from biomass.

Figure 9
Figure 9

Robustness to Outliers

Citation: IMF Working Papers 2020, 253; 10.5089/9781513561240.001.A001

Source: Historical energy datasetNote: Scatter represents static saturation points and average energy efficiency losses for robustness exercise output. Red diamond denotes baseline specification. Labels indicate the country that is dropped from the baseline specification.

Energy Exporters and Transition Economies. We now introduce energy exporters and transition economies one at a time to see their impact on the estimation.29 We would expect energy exporters to increase the saturation point as energy sources are locally abundant (i.e. cheap) and, usually have better energy infrastructure. Also, oil and gas discoveries may positively correlate with income per capita moving both energy consumption and GDP together even at high income levels. These effects are exemplified by adding Qatar. Transition economies pose a different problem, mostly affecting time dummies in the early 90s as energy-inefficient countries entered the sample. Also, most of them experienced a collapse of (inefficient) domestic manufacturing sectors during the transition which then led them to a fast recovery of energy efficiency. The most notable example is Russia, whose introduction reduces both the average rate of energy efficiency gains and the static saturation point.

Additional variables. Results are qualitatively robust to different specifications, including “distance from the equator” and “average winter and summer temperature” (not shown), which serve as proxies for intrinsic demand for cooling and heating energy services, reducing the significance of income. This is not surprising, however, as Hall and Jones (1999) and other scholars have used distance from the equator to instrument output per worker.

A dynamic panel. Finally, a dynamic panel (where energy is introduced as lagged dependent variable) and a specification where energy is in per capita terms deliver qualitatively similar results. The use of the dynamic panel, however, has not been chosen as baseline because the interpretation of the coefficients is more problematic, introducing a short- and long-term Kuznets curve.

Per capita specification and cross-sectional dependence. A specification where energy demand is already expressed in per capita terms is useful to have a better idea of the goodness of fit of the regression. When we specify the baseline specification of Section Results using per capita energy (dropping population and replacing land area with density) we find qualitatively similar results, but R-squared is somewhat lower than in Table 2’s regressions but still as high as 0.94. When we specify specification 8 (OLS FE) of Section Results using per capita energy we find an overall R-squared of 0.79 while the within and between R-squared equal respectively 0.56 and 0.84.

Finally, even though we cannot rule out cross-sectional dependence, its effects on estimates are probably modest given the use of year dummies. To inspect robustness to cross-sectional dependence, we ran 47 cross-section regressions, one for each year of our main panel. Results show point estimates that corroborate the Kuznets curve previously found and are statically significant from 2006 onwards (and the signs of the three polynomial coefficients often point in the right direction in the years before that), highlighting the importance of using recent data. Extending this cross-sectional angle, we also repeated regression Table 2 by employing the between estimator, which discards the time variation in the data. The Kuznets curve remains highly significant. Focusing alternatively on the time-dimension, we used our longest historical panel to regress, by country, energy demand per capita on population, the income polynomial, four lags of real oil prices and a year trend. For 10 out of 11 countries of Group 1, we find significant evidence of a Kuznets curve, and for 7 out of 11 countries the year trend is negative and statistically significant, indicating energy efficiency gains.30

VI. Conclusions

A Kuznets curve seems to characterize the relationship between energy demand and economic development (proxied by per capita income). Moreover, once controlling for manufacturing and the effect of common factors, our preferred specification can explain a large part of energy demand variation over time and across countries. The main benefit of estimating a Kuznets curve is allowing a non-constant income elasticity of energy demand that can capture what seems a typical pattern in the data: an initial decline of energy demand at low levels of income that is followed by stages of acceleration and then saturation at high-income levels. This cannot be captured by neither linear nor quadratic specifications.

The highest value for the elasticity of energy to income is slightly above 1 and is reached at middle-income levels. Based on this elasticity alone, which is very precisely estimated, it seems emerging markets will have to forego growth to slow down their energy consumption. Fortunately, energy efficiency gains have historically played a crucial role in reducing energy demand. These efficiency gains have reduced energy demand by about 1.2 percent per year on average in the past fifty years. In the longer sample, however, energy efficiency gains started to become significant only after WWI when the second industrial revolution was near its end and coal started to give way to more efficient fossil fuels.

In the post WWII period, estimated energy efficiency gains are in part due to fast productivity growth in the manufacturing sector. In fact, the decline of the global manufacturing share (mostly induced by strong productivity growth in manufacturing output till the early 2000s) explains roughly 0.2 percentage points of the annual rate of global energy efficiency gains. At the country level, the role of offshoring and the relocation of manufacturing production has been substantial in some cases. For example, changes in the manufacturing share of GDP between 1990 and 2017 can account for as much as a 4.1 percent decline in US energy consumption and for a 10.7 percent increase of China’s energy consumption. Looking forward, it seems that questionable that we will see a re-shuffling of the manufacturing sector across countries to the extent we have seen in the past, and thus this impact on energy demand will diminish.

The peak of the Kuznets curve (i.e., static saturation) is far into the future, even for AEs, so absolute energy-income decoupling is still far away, being above $107,000 per capita income. Energy efficiency gains, however, allows the introduction of a dynamic concept of peak energy demand which is better estimated and more robust to misspecification—because estimates of efficiency gains and static saturation usually move in opposite directions. So rather than relying on a zero-income elasticity, dynamic saturation depends on a combination of factors that allow for a relative decoupling—i.e., a declining but still positive income elasticity and continuous energy efficiency gains that bring about energy saturation at high levels of income. The income necessary to reach dynamic saturation is about $43,000, which is already reached by 15 countries in North America, Europe, and the Asia-Pacific region. An important implication of dynamic saturation is that further declines in energy consumption are by no means guaranteed and will depend on future energy efficiency gains to continue to outpace the effect of income growth on energy consumption.

Finally, we find that high oil prices (a proxy for energy prices) have probably affected innovation in energy efficiency but slowly. It is, thus, likely that the decade of high oil and energy prices which ended in 2014 helped stimulate energy efficiency gains in the recent years. Prospects of cheap energy, however, will probably reduce the incentive to innovation in the energy and end-use energy services sectors, in absence of policy interventions.

Energy, Efficiency Gains and Economic Development: When Will Global Energy Demand Saturate?
Author: Mr. Christian Bogmans, Lama Kiyasseh, Mr. Akito Matsumoto, and Mr. Andrea Pescatori