WPS8322 Policy Research Working Paper 8322 Relationship between Energy Intensity and Economic Growth New Evidence from a Multi‐Country Multi‐Sector Data Set Uwe Deichmann Anna Reuter Sebastian Vollmer Fan Zhang South Asia Region Office of the Chief Economist January 2018 Policy Research Working Paper 8322 Abstract This paper revisits the relationship between energy inten- more than 30 percent after the level of per capita income sity and economic growth, using a flexible piecewise reaches $5,000. Based on index decomposition, the analysis linear regression model. Based on a panel data set of 137 also finds that although structural change is important for economies during 1990–2014, the analysis identifies intensity levels in all countries, the efficiency effect is more a threshold effect of income growth on energy intensity important in higher-income countries. The results suggest change: although energy intensity is negatively correlated that when countries move beyond lower-middle-income with income growth throughout the entire sample and levels, energy efficiency policies become far more critical study period, the declining rate significantly slows by for sustaining the rate of improvement in energy efficiency. This paper is a product of the Office of the Chief Economist, South Asia Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at fzhang1@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team     Relationship between Energy Intensity and Economic Growth:   New Evidence from a Multi‐Country Multi‐Sector Data Set1    Uwe Deichmanna, Anna Reuterb, Sebastian Vollmerb and Fan Zhanga  a World Bank, USA  b University of Göttingen, Germany                                JEL: Q43, Q48, O47                                                                   1  Financial support from the World Bank’s Knowledge for Change Program is gratefully acknowledged. The  findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not  necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and  its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they  represent.   I.  Introduction  The energy sector plays an important role in economic development. But energy consumption is also  associated with human‐induced climate change because of the dominant role of fossil fuels in power  production.  Energy  use  tends  to  increase  with  development  as  countries  shift  from  labor‐intensive  agriculture to capital and energy intensive industries. As the structural transformation proceeds, they  subsequently move into information‐intensive services. Energy intensity therefore initially increases  with rising incomes and then decreases—a pattern comparable to the Environmental Kuznets Curve  (EKC) that describes the relation between per capita GDP and environmental degradation.   The EKC pattern in the income and pollution relationship has been extensively documented (Dasgupta  et al., 2002; Carson, 2010). Recently, there is also a growing literature on the validity of EKC in the  context of energy use. For example, Luzzati and Orsini (2009) analyze the relationship between per  capita GDP and aggregate energy consumption for a sample of 113 countries over 1971 to 2004.  Their  findings  based  on  different  econometric  techniques  (parametric  and  semi‐parametric)  and  across  different sample groups (world, cross‐countries, and individual countries) do not support the energy‐ EKC hypothesis. On the other hand, Medlock and Soligo (2001), van Benthem and Romani (2009) and  van  Benthem  (2015)  document  an  S‐shaped  relationship  between  energy  intensity  and  economic  growth. Specifically, they find that the income elasticity of energy demand peaks at a GDP per capita  level  between  $5,000  and  $10,000,  and  trends  towards  zero  for  high‐income  levels.    Burke  and  Csereklyei  (2016)  present  disaggregated  analysis  of  the  elasticity  of  energy  demand.  Their  findings  suggest  substantial  heterogeneity  in  income  elasticity  of  energy  use  across  sectors:    industry  and  services  sectors  are  most  responsive  to  income  growth,  followed  by  residential  and  agricultural  sectors.  Another  strand  of  related  literature  tests  the  hypothesis  of  cross‐country  convergence  of  energy  intensity.  Specifically,  these  studies  examine  whether  cross‐country  variation  in  energy  intensity  is  getting smaller (so‐called sigma‐convergence) and whether less efficient countries reduce their energy  intensity  faster  than  more  efficient  ones  (beta‐convergence).  While  the  majority  of  studies  find  convergence  across  developing  and  developed  countries  (such  as  Mielnik  and  Goldemberg,  2000;  Markandya  et  al.,  2006;  Ezcurra,  2007;  Csereklyei  et  al.,  2016;  Burke  and  Csereklyei,  2016),  others  argue for a more nuanced picture of convergence in energy intensity. For example, Le Pen and Sevi  (2010) do not find evidence for global convergence for a group of 97 countries during 1971 – 2003.  When looking at subgroups, they find non‐convergence is “less strongly rejected” for the Middle East,  OECD and Europe sub‐groups.  Ezcurra (2007) show that developing counties tend to converge at a  higher  level  of  energy  intensity,  while  for  developed  countries  at  least  two  different  levels  of  convergence  are  observed.    Stern  (2012)  shows  that  divergence  in  energy  efficiency  is  mostly  associated with economies that are lacking in economic progress.   A  third  strand  of  literature  explores  the  determinants  of  the  growth  rate  of  energy  intensity.  For  example, using  data for 51 sectors in  19 OECD countries  from  1980 to 2005, Mulder and de Groot  (2011)  find  that  reductions  in  energy  intensity  have  been  driven  more  by  within‐sector  energy  efficiency  improvements  than  by  changes  in  the  composition  of  activities  at  the  economy  level  or  within the manufacturing and service sectors. In another paper, Mulder and de Groot (2012) show that  aggregate convergence in energy intensity of 18 OECD countries from 1970 to 2005 is almost entirely  driven by convergence of within‐sector energy intensity levels, and not by convergence of the sector  composition  of  economies.  By  using  output  distance  functions,  Wang  (2013)  decomposes  energy  2    intensity change across 100 countries from 1980 to 2010 into five components: technological catch‐ up,  technological  progress  and  changes  in  the  capital‐energy  ratio,  labor‐energy  ratio  and  output  structure. Technological progress, capital accumulation and changes in output structure contributed  positively to lowering energy intensity while a decrease in the labor‐energy ratio increased intensities.  As this brief survey shows, the existing EKC and convergence literature does not present conclusive  evidence  on  the  relationship  between  energy  intensity  and  economic  growth.  In  this  paper  we  therefore revisit this question using a large data set of 137 economies over the period of 1990‐2014.   In our main estimation model, we adopt a novel piece‐wise regression method which imposes less  structure on the relationship between the level of economic growth and energy intensity. We find a  negative correlation between GDP per capita and energy intensity across the entire sample and study  period. However, the downward slope flattens out at higher values of GDP. The inflection point is at a  per  capita  income  level  of  around  $5,000.  We  also  find  evidence  of  cross‐country  convergence  in  energy intensity. Finally, based on index decomposition, we investigate the extent to which shifts in  economy‐wide energy intensity are affected by structural vs. efficiency changes at the country level.  The analysis shows that while structural change is important for intensity levels in all countries, the  efficiency  effect  is  more  important  for  higher‐income  countries.  This  finding  supports  the  income  threshold effect identified earlier: the relatively automatic efficiency gains (as a result of income gains)  are at lower levels of GDP (and thus the steeper slope), while the harder problem (improving efficiency)  is more important in higher income countries. The results suggest that when countries move beyond  lower‐middle  income  levels,  energy  efficiency  policies  become  far  more  important  if  they  want  to  continue the rate of improvement in energy efficiency.     II. Data   Energy  consumption  data  are  obtained  from  the  International  Energy  Agency  (IEA)  World  Energy  Statistics  and  Balances  data  set.  We  calculate  aggregate  energy  intensity  as  total  final  energy  consumption divided by total GDP (GDP per capita times population). Sector‐level energy intensity is  defined as energy consumption of agriculture, industry and services divided by sector value added.  Data on GDP and value added in agriculture, services, and industry and manufacturing sectors are from  the World Bank World Development Indicators (WDI). 2 For comparison across countries and over time,  all values are converted into PPP‐based constant 2005 prices.  We also control for other country characteristics, including exports and imports, age structure of the  population,  and  population  density.  All  these  factors  may  lead  countries  to  converge  to  different  steady‐state levels of energy intensity. Data on these variables are also obtained from WDI.   Overall, we have an unbalanced panel of 137 economies from 1970 to 2014, and a balanced panel of  64 countries during the same period. Since the attrition rate is relatively high, it may be difficult to tell  whether a change over time is picking up a real development or just a change in the composition of  the  sample.  However, eliminating countries with missing data could cause sample selection  bias,  if  countries with missing data are systematically different from those that have complete observations  over the study period. In the following we report results from the unbalanced panel. For robustness                                                               2  Last Updated: 05/26/2017.  3    check,  we  also  estimate  the  models  using  the  balanced  panel  and  with  data  from  the  Penn  World  Tables (1980‐2014). Conclusions of the paper are robust to these alternative data sources.  Table 1 reports descriptive statistics for the 137 economies in the unbalanced sample. The variables  reported are GDP (PPP) per capita, total final energy consumption and population size. The numbers  are period averages. The sample covers the entire spectrum from poor to rich countries.     III. Relationship between energy intensity and economic development  A. Descriptive analysis  We  are  interested  in  the  relationship  between  energy  intensity  and  economic  development.  One  hypothesis that can be derived from the literature is that the relationship between energy intensity  and economic development follows a Kuznet’s curve (inverted u‐shape). Energy intensity increases in  early  stages  of  development  (industrialization)  and  it  decreases  in  later  stages  of  development  (transition to a services economy). Structural change of the economy is a long process that is unlikely  to be reflected in a panel of only 44 years. However, it is still possible to study this hypothesis. If it were  true,  we  would  expect  a  negative  association  between  energy  intensity  and  GDP  per  capita  for  developed countries and a positive association for developing countries (and possibly no association  for countries that are in the middle of the transition).   Another  hypothesis  is  that  there  is  a  negative  association  between  energy  intensity  and  economic  development with no trend shift. The theoretical justification behind this hypothesis is that economic  development  increases  opportunities  to  make  production  processes  more  efficient  and  therefore  reduces energy intensity. If this negative association was preceded by an initial positive one, as the  EKC hypothesis suggests, it would have pre‐dated the compilation of reliable cross‐country data. The  remaining possibilities are that there is no association between growth and energy intensity or that  energy intensity keeps rising with incomes (a positive association). These patterns are not supported  by any theories or evidence.  Exploratory data analysis suggests it is appropriate to take logs of both variables. Taking logs does not  change the direction of the relationship and therefore does not affect our conclusions regarding the  main  hypotheses.  But  the  logarithm makes  an  exponential  shape  linear  and  allows  correlation  and  regression analysis of the data.  As a first step we calculate the correlation coefficient between energy consumption per $PPP and GDP  per capita PPP for all 137 economies in the data set (Table A1). For about 40 percent of the countries  we  find  a  very  strong  negative  linear  association  between  the  two  variables  with  a  correlation  coefficient of ‐0.9 or smaller. For another 35 percent of the countries we find a moderate to strong  negative association with correlation coefficients between ‐0.9 and ‐0.5. Of the remaining countries,  16 show a weak to very weak negative association and 15 countries show a weak positive association.  The overall impression from the correlation analysis is that there is a fairly strong negative association  between energy consumption per $PPP and GDP per capita PPP for most countries. If the relationship  followed  a  Kuznets  curve,  we  would  expect  a  negative  association  in  developed  countries  and  a  positive  association  in  developing  countries.  Even  a  cursory  look  at  Table  A1  contradicts  this  hypothesis: We find the same strong negative linear association for countries like India, Nigeria and  4    China as for countries like the United States, the United Kingdom and Germany. On the other hand,  the  relatively  rich  EU  members  Spain  and  Portugal  show  only  a  weak  negative  or  even  a  positive  association between the two variables.   B. Econometric analysis  In the econometric analysis, we first specify the following fixed effects model to test the relationship  between energy intensity and economic development  ∙           (1)  Where  EIit is log energy intensity of country  i in year  t,  GDPit  is log GDP per capita. Previous studies  suggest  the  importance  of  controlling  for  sector  composition  of  an  economy  when  analyzing  the  relationship between income and energy intensity (Burke and Csereklyei, 2016; Medlock and Soligo,  2001). We therefore include on the right‐hand‐side of equation (1) a vector  , which is a set of control  variables  for  the  composition  of  final  energy  demand.  It  includes  the  percentage  of  total  energy  consumption  used  in  industry,  transport,  residential,  services,  agriculture  and  non‐energy  use.  The  reference category is non‐specified other energy consumption, all variables are coded between 0 and  1.   is  a  set  of control variables  including  value  added  in  industry,  agriculture,  manufacturing  and  services as well as exports and imports. All variables are measured as a percentage of total GDP and  coded between 0 and 100.   is a set of demographic variables which could affect energy consumption  patterns. It includes population density as well as the percentage of the population aged 14 or younger  and  65  and  older  (the  working  age  population  is  the  reference  category).  Population  density  is  measured as the number of people per 1,000 square kilometers of area, and the other two variables  are  coded  between  0  and  100.  Finally,   is  a  country  fixed  effect  capturing  time‐invariant  country  specific characteristics;    is a time fixed effect controlling for yearly shocks that are common to all  countries; and   is the error term.   The results of these regressions are shown in Table 2. In the first column we report the results of a  simple pooled OLS specification, in the second column we add country fixed effects, and in the third  column we add time fixed effects. The coefficient of log GDP per capita is negative and highly significant  in all three cases, but its magnitude changes when the country and time fixed effects are included.  Adding  the  control  variables  for  the  composition  of  energy  consumption  changes  the  explanatory  power of the model, but the coefficient of log GDP per capita remains relatively stable. Adding the  control variables for the structure of the economy also increases the explanatory power moderately,  but also diminishes the size of the coefficient of log GDP per capita. Adding the demographic controls  again increases the explanatory power of the model and enlarges the magnitude of the coefficient of  log  GDP  per  capita  again.  Overall,  these  regressions  are  strong  evidence  for  a  negative  association  between GDP per capita and energy intensity. An increase of GDP per capita by 1 percent is associated  with a decrease of energy intensity by 0.62 percent.    The scatter plot in Figure 1 and also the country by country correlation analysis in the previous section  indicate  that  the  relationship  between  log  GDP  per  capita  and  log  energy  intensity  is  not  perfectly  linear, because there are a few countries with no or a positive association. We thus include a squared  term of log GDP per capita to equation (1) to allow for a more flexible functional form and also to test  for an inverted U‐shaped relationship.    ∙ ∙       (2)  5    The  results  are  reported  in  Table  3.  They  are  qualitatively  similar  to  the  results  from  the  previous  regressions, but thanks to the more flexible functional form the explanatory power is always higher  than in Table 2. The coefficients of log GDP per capita are always negative and significant, and the  coefficients of log GDP per capita squared are always positive and significant. This is an indicator for a  U‐shaped relationship. However, when we calculate the minimum of the parabola ( / 2∙ ),  we find that the minimum is always near the right border of the observed values of log GDP per capita  (except for the pooled OLS regression). It is thus not a real minimum, in fact we are fitting the left‐hand  side  of  a  parabola  to  the  data,  thus  the  relationship  between  log  GDP  per  capita  and  log  energy  intensity is always negative but flattening out for higher values of log GDP per capita. In other words,  these regressions do not show evidence for a u‐shaped or inverted u‐shaped relationship between the  two variables.  To adopt a more flexible functional form without imposing structure on the relationship between GDP  and  energy  intensity,  we  apply  a  piecewise  linear  model  similar  to  the  one  Myrskylä  et  al.  (2009)  implemented. The starting point is a piecewise linear relationship that fits two different linear models  to the left and to the right of a critical value of log GDP per capita:    ∙ ∙ ∙ ∙   (3)  where   is an indicator variable that is equal to 1 to the left of the critical value of log GDP per  capita and 0 otherwise.   is an indicator variable that is equal to 1 to the right of the critical value  of log GDP per capita and 0 otherwise. If   is negative and   is positive, then the relationship is  V‐shaped (or U‐shaped). If   is positive and   is negative, then the relationship is inverted V‐ shaped. If both coefficients have the same  sign, it is interesting to compare the magnitudes of the  coefficients,  because  the  piecewise  linear  model  can  also  show  that  parts  of  the  relationship  are  steeper/flatter than others.  Before we explain how to obtain the critical value of log GDP per capita, we first note that we do not  estimate equation (3) directly but rather in its differences‐in‐differences version  ∆ ∙∆ ∙ ∙∆ ∙ ∙∆ ∆ ∆      (4)  where  ∆  is  the  difference  operator  ∆ .  The  differencing  implicitly  controls  for  the  country fixed effects and accounts for autocorrelation in the residuals. The coefficients obtained from  the  differences‐in‐differences  model  comes  closer  to  a  causal  interpretation  than  the  coefficients  obtained from a levels equation. As additional robustness check, we estimate the same model with  the first lag of the log GDP per capita variable.  ∆ ∙∆ ∙ ∙∆ ∙ ∙∆ ∆ ∆   (5)  To  find  the  optimal  value  for  the  threshold  between  the  two  linear  models,  we  use  the  following  numerical  procedure:  We  estimate  equation  (4)  for  any  possible  critical  value  within  the  range  of  observations (within 5.8 and 11.7) in steps of 0.1. We calculate the likelihood for each model and then  pick the critical value for which the model has the largest likelihood (maximum likelihood estimation).  In Figure 2 we show a plot of the likelihoods of all models with critical values between 5.8 and 11.3,  the maximum is at 8.5. Thus, we estimate equations (4) and (5) with critical values of 8.5, that is    is equal to 1 if log GDP per capita is smaller than 8.5 and 0 otherwise, and   is 1 if log GDP per  capita is greater or equal 8.5 and 0 otherwise.   6    The  results  for  the  two  differences‐in‐differences  regressions  are  reported  in  Table  4.  For  both  equations, the coefficients are negative and highly significant, both to the right and to the left of the  critical  value  of  log  GDP  per  capita.  However,  the  slope  is  much  steeper  (we  find  a  more  negative  coefficient) to the left of the critical value than to the right. These results confirm the previous finding  from  the  fixed  effects  estimations.  The  relationship  between  log  GDP  per  capita  and  log  energy  intensity is negative for the entire observation period, but the slope flattens out for higher values of  GDP per capita. Specifically, when per capita GDP reaches around $5,000, the declining rate of energy  intensity  with  respect  to  economic  growth  significantly  slows  down  by  more  than  30  percent.  This  finding suggests that energy productivity increases as income grows. However, the income elasticity  of energy demand increases with income. As a result, after exceeding a threshold of around $5,000  GDP per capita, the efficiency dividends of economic growth become much smaller. More aggressive  energy  efficiency  policies  are  needed  to  sustain  rapid  and  continued  improvement  in  energy  productivity as income levels surpass the threshold.        7    IV. Convergence of energy intensity  In an exploratory data analysis (time series plots of energy intensity) we got the impression that energy  intensity is converging to the same level across countries. We will now investigate this hypothesis more  systematically.  Typically,  convergence  analysis  is used  in the  context  of economic  growth  and goes  back to Barro and Sala‐i‐Martin (1991). In order to achieve convergence, or in other words for poor  countries to catch up to GDP per capita levels of rich countries over time, two things are necessary:  beta convergence and sigma convergence. For beta convergence we study the association between  the average growth rate of GDP per capita and the initial level of GDP per capita for a given time period.  We speak of beta convergence if the regression coefficient (beta) of the initial level of GDP per capita  is  negative  –  as  an  explanatory  variable  for  the  growth  rate  of  GDP  per  capita.  The  negative  beta  coefficient implies that initially poorer countries had higher average growth rates than initially richer  countries. Sigma convergence means that the standard deviation of GDP per capita decreases over  time. If we find both, we speak of convergence.  How  do  these  concepts  translate  to  convergence  of  energy  intensity?  For  energy  intensity,  convergence means that countries that use a lot of energy per dollar of GDP over time catch up to  levels of countries that use less energy per dollar of GDP. Thus, we speak of beta convergence in this  context if higher rates of reduction in energy consumption per dollar of GDP are associated with higher  initial levels of energy consumption per dollar of GDP. Different from convergence of GDP per capita,  in  this  context  beta  convergence  is  implied  by  a  positive  beta  coefficient  in  the  regression.    The  meaning of sigma convergence is exactly the same as in the GDP per capita context.  We calculate the total change in log energy consumption per dollar of GDP between 1990 and 2014  and  divide  it  by  the  level  in  1990.  This  is  a  simple  measure  for  the  rate  of  reduction  in  energy  consumption per dollar of GDP. We then run a regression with the rate of reduction as a dependent  variable and the initial level of log energy consumption per dollar of GDP as the explanatory variable  according to equation (6):   ∆ log       (6)  ∆Iit is the growth rate of energy intensity of country  i during the period from 1990 to 2014.  Ii0 is the  energy intensity of country i in year 1990 and  is the error term. The result is shown in the left panel  of Figure 3. Indeed, the beta coefficient is positive and highly significant (beta= .029 and p<0.01). Quite  remarkably, the variation in the initial level of log energy consumption per dollar of GDP explains about  42 percent of the variation in the rate of reduction across countries. We thus find strong evidence for  beta convergence.   In the right panel of Figure 3, we show fitted normal densities for log energy consumption per dollar  of GDP for 1990 and 2014. One can clearly see that the density became much narrower over time (and  also that the mean shifted to the left). Indeed, the standard deviation decreases from 0.649 in 1990 to  0.511  in  2014.  We  thus  find  evidence  for  sigma  convergence.  As  we  find  both  beta  and  sigma  convergence,  we  can  thus  conclude  that  countries  with  high  initial  levels  of  energy  intensity  are  converging to the levels of countries with low energy intensity.  8    In  Figure  4  we  show  box  plots  for  log  energy  intensity  for  1990,  2000,  2010  and  2014.  A  box  plot  displays the median, the first quartile, the third quartile and the largest and smallest adjacent values.3  The following observations are noteworthy: First quartile, median, third quartile and largest value all  decrease over time. The outside values get closer to the largest adjacent observation over time. The  smallest adjacent value also decreases over time, though the largest change occurred between 1990  and 2000. After 2000, the smallest adjacent value stayed fairly constant. The smallest value can be  considered as the “technology frontier” or “minimal energy consumption required per unit of output”.  According to this picture, the frontier did not move very much between 2000 and 2014, rather the rest  of the distribution moved closer to the frontier. Given that there is a generally increasing trend for per  capita income across the sample, this result appears to be consistent with our earlier conclusion that  as income rises above a certain threshold, reducing energy intensity becomes more difficult.   To examine whether there is also a difference in the speed of convergence on different sides of the  income  threshold,  we  divide  the  sample  into  two  subgroups  based  on  whether  a  country‐year  observation  lies  below  or  above  the  cutoff  point  at  log  per  capita  GDP  of  8.5.  We  calculate  the  conditional  convergence  rate  according  to  log 1  (Barro  and  Sala‐i‐Martin,  1992;  Islam, 1995), where  T is the length in the time dimension under consideration (25 years in our case)  and   is the coefficient obtained from equation (6). The results are reported in Table 5. The rate of  unconditional convergence for groups below and above the income threshold is roughly 0.10 percent  and 0.14 percent, respectively. The difference between the two groups is small and the lower‐income  sample even has a lower convergence rate. The result could reflect different economic growth rates  of the two groups.    V. Index Decomposition  In this section, we investigate how much of the improvement in energy intensity is due to structural  change and how much is due to efficiency improvements, and whether the income threshold effects  identified earlier play a role in the relative importance of structural vs. efficiency changes. To this end  we use the Fisher Ideal Index Decomposition for structural change in energy intensity as described in  Boyd and Roop (2004). We consider three sectors: industry, agriculture and services. As described in  Boyd  and  Roop  (2004)  we  first  calculate  a  Laspeyres  and  a  Paasche  index  of  energy  intensity  for  structure and efficiency using the following formulas from Boyd and Roop (2004):         S: agriculture, industry and services share of GDP  I: energy intensity  0: first year                                                               3  Values that are more than 1.5 times the interquartile range (distance between first and third quartile) away  from the median are plotted separately.   9    T: last year  i: industry, agriculture, service  The Fisher index is the geometric mean of the Laspeyres and the Paasche index. The aggregate Fisher  index is the product of the structural part and the intensity part.  Table 6 displays the values of the Fisher index (aggregate, structure and efficiency) for the year 2014  (base year 1990=1) in ascending order of the aggregate index. The structural index ranges between 0.5  and 1.3 while the efficiency intensity ranges from 0.2 to 1.6.4 To examine whether structural change  or efficiency improvement plays a more important role in influencing aggregate energy intensity, we  estimate the following country fixed‐effects model.   Fagg= FStr+ Feff+ui+ei                          (7)  Fagg is the aggregate intensity index,  Fstr is the structural index and  Feff is the efficiency index.  ui  are  time‐invariant  country  fixed  effects.  ei  is  the  error  term.  Similar  to  the  convergence  analysis,  we  conduct both full sample analysis and sub‐sample analysis based on the previously identified income  threshold.  Table 7 reports the estimation results, where column  (1) corresponds to the full sample  estimation,  and  columns  (2)  and  (3)  correspond  to  sub‐samples  below  and  above  the  income  threshold, respectively.  The  results  suggest  that  structural  change  has  made  a  higher  contribution  to  variation  in  energy  intensity for the full sample of observations. However, for country‐year observations that are above  the income threshold, efficiency change plays a more important role in driving the change in aggregate  intensity.       VI. Conclusion  Reducing the amount of energy required to produce a unit of output is a priority in efforts to slow  climate change. This paper contributes to the literature on the relationship between energy use and  economic growth by better describing the energy intensity trajectory as incomes grow. Specifically,  our  results  confirm  a  negative  correlation  between  GDP  per  capita  and  energy  intensity  as  well  as  cross‐country  convergence  in  energy  intensity.  We  also  identify  a  transition  from  a  relatively  rapid  lowering of energy intensity to slower gains. Using a flexible econometric approach, we show that as  countries reach a per capita income of about $5,000 (PPP)—slightly above the level considered by the  World  Bank  as  upper‐middle  income—the  downward  sloping  curve  of  energy  use  per  unit  of  GDP  flattens  out.  Index  decomposition  identifies  the  relative  contribution  of  structural  change  versus  energy  efficiency  to  lowering  energy  intensity.  It  shows  that  the  former  is  important  at  all  income  levels, while energy efficiency is more important at higher income levels.   The results suggest that we can expect to see relatively rapid improvements in energy intensity as the  economies in today’s poor countries grow. However, as they move beyond lower‐middle income, the  role for energy efficiency policies then becomes far more critical if a country wants to continue the  rate of improvement in energy efficiency (Deichmann and Zhang 2013). There is now a much better                                                               4  Malta is dropped as an outlier.  10    understanding of different policy instruments for improving energy efficiency—even if we do not yet  sufficiently understand why they are not adopted at a scale that would seem beneficial (Gillingham  and Palmer 2014). Energy efficiency gains likely will require numerous sector‐specific instruments such  as targeted revenue‐neutral environmental taxation (eco‐taxes) and other price instruments, stricter  regulations, and public investments.     References  van Benthem, A.A. and M. Romani (2009) Fuelling Growth: What Drives Energy Demand  in Developing Countries? Energy Journal 30(3): 147‐170  Barro, R.J., Sala‐i‐Martin, X., 1992. Convergence. J. Polit. Econ. 100, 223–251.  Boyd, G. A., & Pang, J. X. (2000). Estimating the linkage between energy efficiency and  productivity. Energy policy, 28(5), 289‐296.  Burke, P. J., & Csereklyei, Z. (2016). Understanding the energy‐GDP elasticity: A sectoral  approach. Energy Economics, 58, 199‐210.  Burnett, J. W., & Madariaga, J. (2017). The convergence of US state‐level energy intensity. Energy  Economics, 62, 357‐370.  Carson, R. T. (2010). The environmental Kuznets curve: seeking empirical regularity and theoretical  structure. Review of Environmental Economics and Policy, 4(1), 3‐23.  Csereklyei, Z., Rubio‐Varas, M. D. M., & Stern, D. I. (2016). Energy and Economic Growth: The Stylized  Facts. Energy Journal, 37(2), 223‐255.  Csereklyei, Z., & Stern, D. I. (2015). Global energy use: Decoupling or convergence?. Energy  Economics, 51, 633‐641.  Dasgupta, S., Laplante, B., Wang, H., & Wheeler, D. (2002). Confronting the environmental Kuznets  curve. The Journal of Economic Perspectives, 16(1), 147‐168.  Deichmann, U., & Zhang, F. (2013). Growing green: the economic benefits of climate action. World  Bank, Washington, D.C.  Duro, J. A., Alcántara, V., & Padilla, E. (2010). International inequality in energy intensity levels and  the role of production composition and energy efficiency: an analysis of OECD countries. Ecological  Economics, 69(12), 2468‐2474.  Elias, R. J., & Victor, D. G. (2005). Energy transitions in developing countries: a review of concepts and  literature. Program on Energy and Sustainable Development, Working Paper. Stanford: Stanford  University.  Ezcurra, R. (2007). Distribution dynamics of energy intensities: a cross‐country analysis. Energy  Policy, 35(10), 5254‐5259.  Gillingham, K., & Palmer, K. (2014). Bridging the energy efficiency gap: Policy insights from economic  theory and empirical evidence. Review of Environmental Economics and Policy, 8(1), 18‐38.  Goldemberg, J., & Prado, L. T. S. (2011). The decline of the world’s energy intensity. Energy  Policy, 39(3), 1802‐1805.  Herrerias, M. J. (2012). World energy intensity convergence revisited: A weighted distribution  dynamics approach. Energy policy, 49, 383‐399.  11    Herrerias, M. J., & Liu, G. (2013). Electricity intensity across Chinese provinces: New evidence on  convergence and threshold effects. Energy Economics, 36, 268‐276.  Islam, N., 1995. Growth empirics: a panel data approach. Q. J. Econ. 4, 1127–1170.  Le Pen, Y., & Sévi, B. (2010a). On the non‐convergence of energy intensities: evidence from a pair‐ wise econometric approach. Ecological Economics, 69(3), 641‐650.  Le Pen, Y., & Sévi, B. (2010b). What trends in energy efficiencies? Evidence from a robust test. Energy  Economics, 32(3), 702‐708.  Luzzati, T., & Orsini, M. (2009). Investigating the energy‐environmental Kuznets curve. Energy, 34(3),  291‐300.  Markandya, A., Pedroso‐Galinato, S., & Streimikiene, D. (2006). Energy intensity in transition  economies: Is there convergence towards the EU average?. Energy Economics, 28(1), 121‐145.  Maza, J. Villaverde (2008). The world per capita electricity consumption distribution: signs of  convergence? Energy Policy, 36(11), 4255‐4261.  Medlock, K.B. and R. Soligo, 2001, Economic Development and End‐Use Energy Demand.  Energy Journal 22(2): 77‐105.  Mohammadi, H., & Ram, R. (2012). Cross‐country convergence in energy and electricity consumption,  1971–2007. Energy economics, 34(6), 1882‐1887.  Mulder, P., & de Groot, H. L. (2012). Structural change and convergence of energy intensity across  OECD countries, 1970–2005. Energy Economics, 34(6), 1910‐1921.  Mulder, P., & Groot, H. L. F. (2011). Energy Intensity Across Sectors and Countries: Empirical Evidence  1980‐2005. CPB Netherlands Bureau for Economic Policy Analysis.  Nilsson, L. (1993). Energy intensity in 31 industrial and developing countries 1950–88. Energy, 18 (4),  309‐322.  Saboori, B., & Sulaiman, J. (2013). Environmental degradation, economic growth and energy  consumption: Evidence of the environmental Kuznets curve in Malaysia. Energy Policy, 60, 892‐905.  Smulders, S., & De Nooij, M. (2003). The impact of energy conservation on technology and economic  growth. Resource and Energy Economics, 25(1), 59‐79.  Stern, D. I. (2012). Modeling international trends in energy efficiency. Energy Economics, 34(6), 2200‐ 2208.  van Benthem, A. A. (2015). Energy leapfrogging. Journal of the Association of Environmental and  Resource Economists, 2(1), 93‐132.  Voigt, S., De Cian, E., Schymura, M., & Verdolini, E. (2014). Energy intensity developments in 40 major  economies: Structural change or technology improvement?. Energy Economics, 41, 47‐62.  Wang, C. (2013). Changing energy intensity of economies in the world and its decomposition. Energy  Economics, 40, 637‐644.  Worrell, E., Laitner, J. A., Ruth, M., & Finman, H. (2003). Productivity benefits of industrial energy  efficiency measures. Energy, 28(11), 1081‐1098. 12    Figure 1: Scatterplot of log energy intensity and log GDP per capita      Figure 2: Likelihood for cut‐off in piecewise linear regression        13    Figure 3: Energy Intensity Convergence: Beta and Sigma Convergence          Figure 4: Box plots of log energy consumption per dollar for selected years 5                                                                        5  This graph does not include Brunei, which had a log energy intensity of ‐18 in 1980 and thus constitutes an  extreme outlier (because of oil price effects).   14    Table 1 Summary statistics (period averages) for the 137 economies   Country  GDP p.c.  C(Energy)  Population  Country  GDP p.c.  C(Energy)  Population  Qatar   120,371    8,363   971  Hungary   19,404    17,949   10,159  United Arab  Emirates   91,285    29,787   4,593  Slovak Republic   19,142    11,544   5,372  Brunei   83,208    860   341  Russian Federation   18,346    454,862   145,585  Kuwait   80,735    9,845   2,396  Gabon   18,176    2,406   1,299  Luxembourg   78,399    3,446   454  Croatia   18,112    6,937   4,467  Norway   57,065    19,581   4,594  Lithuania   17,508    5,784   3,389  Singapore   55,826    10,518   4,207  Malaysia   16,782    32,673   24,128  Switzerland   50,961    19,507   7,363  Poland   16,114    63,573   38,279  United States   45,368    1,465,972   286,382  Chile   15,868    19,762   15,510  Bahrain   42,200    3,673   841  Latvia   15,738    4,146   2,317  Oman   41,481    6,673   2,542  Argentina   15,601    47,624   37,888  Denmark   41,260    14,363   5,378  Venezuela, RB   15,567    36,434   25,364  Netherlands   40,688    59,770   16,015  Mexico   14,536    100,630   105,562  Saudi Arabia   39,294    79,833   23,095  Kazakhstan   14,408    37,531   15,785  Austria   38,803    24,594   8,122  Turkey   14,384    61,553   65,149  Ireland   38,762    10,093   4,025  Cuba   13,768    8,324   11,106  Hong Kong SAR,  China   38,451    7,862   6,610  Uruguay   13,666    2,728   3,298  Sweden   37,422    34,295   9,025  Romania   13,640    26,475   21,652  Germany   37,219    230,486   81,591  Iran, Islamic Rep.   13,605    113,367   67,204  Canada   37,139    186,403   31,501  Lebanon   12,816    3,337   3,699  Belgium   37,075    39,658   10,457  Montenegro   12,240    805   613  Australia   35,960    69,457   19,867  Panama   12,197    2,190   3,151  Italy   35,147    127,407   57,841  Brazil   12,192    164,671   179,867  Finland   34,788    24,407   5,217  Mauritius   12,065    665   1,186  France   34,390    158,732   62,060  Suriname   11,818    515   480  Japan   34,190    312,992   126,642  Bulgaria   11,510    10,647   7,926  United Kingdom   33,148    142,358   60,066  Algeria   11,191    19,515   32,155  Cyprus   29,805    1,408   973  Botswana   11,075    1,431   1,791  Spain   29,434    82,561   42,517  Thailand   10,721    60,940   63,170  New Zealand   29,240    12,180   3,975  Costa Rica   10,718    2,608   4,004  Israel   25,725    11,519   6,516  South Africa   10,575    60,770   45,775  Slovenia   25,406    4,646   2,010  Serbia   10,451    8,941   7,464  Greece   25,367    17,934   10,801  Iraq   10,425    18,483   25,292  Portugal   24,804    17,508   10,306  Belarus   10,262    20,989   9,847  Malta   24,177    352   394  Macedonia   9,491    1,665   2,019  Libya   23,898    9,076   5,492  Colombia   9,338    22,091   41,398  Czech Republic   23,395    27,003   10,332  Azerbaijan   8,777    8,456   8,264  Korea, Rep.   22,810    127,167   47,167  Jordan   8,600    3,799   5,198  Trinidad & Tobago   21,923    8,803   1,285  Ecuador   8,458    7,890   13,066  15    Estonia   20,317    3,175   1,395  Dominican Rep.   8,446    4,358   8,824  Jamaica   8,320    2,178   2,612  Myanmar   2,034    11,930   48,184  Tunisia   8,128    5,772   9,804  Senegal   1,974    1,755   10,653  Egypt, Arab Rep.   7,956    37,571   71,593  Bangladesh   1,892    17,234   134,500  Peru   7,620    11,515   26,521  Cambodia   1,861    3,539   12,447  Turkmenistan   7,471    11,636   4,569  Tajikistan   1,855    2,292   6,581  Bosnia & Herzeg.   7,148    2,626   3,912  Tanzania   1,729    13,673   36,987  Kosovo   7,138    1,076   1,830  Benin   1,653    2,302   7,587  Namibia   7,015    1,119   1,921  Nepal   1,638    8,315   24,044  Ukraine   6,982    87,717   48,604  Haiti   1,637    2,118   8,836  Indonesia   6,807    122,971   217,661  Eritrea   1,516    566   3,755  Albania   6,622    1,613   3,064  Togo   1,251    1,391   5,257  Paraguay   6,588    3,793   5,452  Niger   813    1,820   12,570  El Salvador   6,533    2,700   5,802  Ethiopia   785    27,916   71,227  Sri Lanka   6,332    7,331   18,973  Congo, Dem. Rep.   667    15,007   52,737  Guatemala   6,125    6,256   12,390  Mozambique   659    7,055   19,728  Mongolia   6,094    2,227   2,485          China   5,526    1,120,968   1,268,418          Morocco   5,180    9,892   29,466          Congo, Rep.   5,065    904   3,332          Georgia   4,940    3,369   4,334          Angola   4,720    6,949   16,746          Philippines   4,713    23,033   80,880          Bolivia   4,688    3,566   8,668          Armenia   4,460    1,967   3,120          Yemen, Rep.   3,849    3,846   19,007          Honduras   3,809    3,264   6,480          Pakistan   3,767    56,478   144,810          Nigeria   3,747    86,100   131,933          Nicaragua   3,593    1,873   5,137          Moldova   3,250    2,673   3,624          Uzbekistan   3,199    34,771   25,394          Vietnam   3,147    31,652   79,062          India   3,059    359,441   1,087,461          South Sudan   2,915    537   7,703          Côte d'Ivoire   2,884    4,652   17,104          Sudan   2,606    8,388   29,785          Ghana   2,565    5,282   20,189          Zambia   2,533    5,605   11,412          Cameroon   2,486    5,722   17,013          Kyrgyz Republic   2,468    2,718   5,014          Kenya   2,336    10,274   33,288          16    Zimbabwe   2,053    8,415   12,729              17    Table 2: Panel regressions. Dependent variable: log energy intensity, Eq. (1)    (1)  (2) (3) (4) (5) (6)                      log GDP p.c.  ‐0.26***  ‐0.64***  ‐0.62***  ‐0.68***  ‐0.58***  ‐0.62***    (0.007)  (0.012)  (0.017)  (0.016)  (0.020)  (0.021)  E(Industry, %)    0.78***  0.40***  0.39***    (0.079)  (0.095)  (0.096)  E(Transport, %)    ‐0.35***  ‐0.34***  ‐0.29***    (0.088)  (0.097)  (0.098)  E(Residential, %)    ‐0.79***  ‐0.70***  ‐0.72***    (0.070)  (0.081)  (0.081)  E(Services, %)    ‐0.27**  ‐0.15  ‐0.11    (0.135)  (0.148)  (0.148)  E(Agriculture, %)    ‐0.05  ‐0.25  ‐0.24    (0.189)  (0.191)  (0.190)  E(Non‐energy, %)    1.53***  1.11***  1.21***    (0.144)  (0.164)  (0.164)  VA(Agriculture, %)    0.00***  0.00***    (0.001)  (0.001)  VA(Industry, %)    0.00**  0.00***    (0.001)  (0.001)  VA(Manufacturing, %)    0.01***  0.01***    (0.001)  (0.001)  VA(Services, %)  ‐0.00***  ‐0.00***  (0.001)  (0.001)  Exports, %  ‐0.00**  ‐0.00***    (0.000)  (0.000)  Imports, %    0.00*  0.00**    (0.000)  (0.000)  Pop. 0‐14, %    ‐0.01***    (0.002)  Pop. 65+, %    ‐0.00    (0.004)  Pop. Density    0.00***    (0.000)      Observations  3,238  3,238  3,238  3,238  2,717  2,698  R‐squared  0.284  0.487  0.490  0.595  0.582  0.591  Country Effects  No  Yes  Yes  Yes  Yes  Yes  Year Effects  No  No  Yes  Yes  Yes  Yes  Standard errors in parentheses    *** p<0.01, ** p<0.05, * p<0.1          18    Table 3: Panel regressions. Dependent variable: log energy intensity, Eq. (2)    (1)  (2)  (3)  (4)  (5)  (6)                log GDP p.c.  ‐2.06***  ‐1.08***  ‐1.14***  ‐1.54***  ‐1.54***  ‐1.57***    (0.094)  (0.104)  (0.105)  (0.096)  (0.120)  (0.124)  log GDP p.c. sq.  0.10***  0.02***  0.03***  0.05***  0.06***  0.06***    (0.005)  (0.006)  (0.006)  (0.006)  (0.007)  (0.007)  E(Industry, %)        0.86***  0.46***  0.44***          (0.078)  (0.094)  (0.095)  E(Transport, %)        ‐0.27***  ‐0.29***  ‐0.27***          (0.088)  (0.096)  (0.097)  E(Residential, %)        ‐0.83***  ‐0.77***  ‐0.80***          (0.069)  (0.081)  (0.081)  E(Services, %)        ‐0.32**  ‐0.21  ‐0.15          (0.134)  (0.146)  (0.147)  E(Agriculture, %)        ‐0.07  ‐0.26  ‐0.29          (0.187)  (0.188)  (0.188)  E(Non‐energy, %)        1.48***  0.99***  1.08***          (0.142)  (0.162)  (0.163)  VA(Agriculture, %)          0.00  0.00**            (0.001)  (0.001)  VA(Industry, %)          0.00**  0.00***            (0.001)  (0.001)  VA(Manufacturing, %)          0.01***  0.01***            (0.001)  (0.001)  VA(Services, %)          ‐0.00***  ‐0.00**            (0.001)  (0.001)  Exports, %          ‐0.00***  ‐0.00***            (0.000)  (0.000)  Imports, %          0.00**  0.00**            (0.000)  (0.000)  Pop. 0‐14, %            ‐0.01***              (0.002)  Pop. 65+, %            ‐0.01**              (0.004)  Pop. Density            0.00**              (0.000)                Observations  3,238  3,238  3,238  3,238  2,717  2,698  R‐squared  0.357  0.490  0.494  0.605  0.593  0.601  Country Effects  No  Yes  Yes  Yes  Yes  Yes  Year Effects  No  No  Yes  Yes  Yes  Yes  Standard errors in parentheses            *** p<0.01, ** p<0.05, * p<0.1                19    Table 4: Differences‐in‐Differences Regressions, Eqs. (4) and (5)    Eq. (4)  Eq. (5)             ‐0.64***  ‐0.14***  (0.029)  (0.033)    ‐0.43***  ‐0.09***  (0.029)  (0.031)  Observations  3,099  2,968  R‐squared  0.224  0.073  Standard errors in parentheses  *** p<0.01, ** p<0.05, * p<0.1    Table 5: Beta‐convergence Analysis     (1)  (2)  (3)  VARIABLES  Full sample  Below Above             l_base  ‐0.0190***  ‐0.0166***  ‐0.0200***    (0.002)  (0.004)  (0.003)    Observations  2,748  552  1,692  R‐squared  0.028  0.029  0.024  Standard errors in parentheses    *** p<0.01, ** p<0.05, * p<0.1      20    Table 6: Aggregate, Structural and Efficiency Index in 2014 based on Fisher Ideal Index  Decomposition (1990=1)  Country  Total  Struct.  Intens.  Country  Total  Struct.  Intens.  Azerbaijan  0.185  0.976  0.190  Korea, Rep.  0.813  0.986  0.825  Armenia  0.238  0.902  0.264  Cyprus  0.826  0.679  1.216  Romania  0.355  0.801  0.444  Nigeria  0.827  1.006  0.822  Belarus  0.392  0.944  0.415  Morocco  0.830  0.963  0.862  Bulgaria  0.407  0.706  0.576  Costa Rica  0.832  0.867  0.959  Georgia  0.412  0.934  0.441  Japan  0.841  0.865  0.973  Kyrgyz Republic  0.423  0.859  0.493  Indonesia  0.841  1.028  0.819  Cuba  0.479  1.063  0.451  Italy  0.866  0.885  0.978  Mongolia  0.485  0.919  0.528  Ethiopia  0.870  1.182  0.736  Mauritius  0.520  0.824  0.631  Pakistan  0.874  0.907  0.964  China  0.540  0.996  0.542  Finland  0.903  0.868  1.040  Albania  0.590  0.708  0.832  Kenya  0.923  1.009  0.915  Ukraine  0.601  0.730  0.823  Argentina  0.931  0.919  1.014  Honduras  0.603  1.019  0.592  Tanzania  0.945  1.110  0.851  Tajikistan  0.656  0.845  0.776  Austria  0.946  0.905  1.045  Russian Federation  0.661  0.819  0.807  Venezuela, RB  0.968  0.830  1.166  United Kingdom  0.663  0.868  0.763  Chile  0.975  0.917  1.063  Zimbabwe  0.678  0.929  0.730  Benin  0.980  1.065  0.920  Norway  0.680  1.031  0.660  Panama  1.000  1.191  0.839  Sweden  0.692  0.907  0.763  Brazil  1.013  0.766  1.323  Cameroon  0.696  1.021  0.681  Congo, Rep.  1.037  1.267  0.819  El Salvador  0.698  1.006  0.694  Dominican Rep.  1.038  0.885  1.174  Colombia  0.703  0.867  0.812  Trinidad & Tobago  1.051  1.022  1.029  Mozambique  0.707  1.071  0.660  Algeria  1.101  0.964  1.143  Denmark  0.713  0.891  0.800  Thailand  1.129  0.982  1.150  India  0.728  1.045  0.697  Ecuador  1.130  1.130  1.000  Sudan  0.731  0.597  1.225  Turkey  1.133  0.878  1.291  Ghana  0.745  1.317  0.565  Iran, Islamic Rep.  1.137  1.026  1.108  Macedonia  0.746  0.774  0.965  Senegal  1.145  1.021  1.122  France  0.753  0.884  0.852  Uruguay  1.153  0.902  1.278  Zambia  0.772  0.817  0.944  Congo, Dem. Rep.  1.157  1.094  1.057  Australia  0.773  0.914  0.845  Singapore  1.167  0.935  1.248  Botswana  0.775  0.809  0.958  Uzbekistan  1.210  0.937  1.292  Netherlands  0.787  0.827  0.951  Togo  1.228  0.918  1.337  Tunisia  0.796  0.908  0.877  Bahrain  1.317  1.007  1.307  Switzerland  0.801  0.949  0.844  Bolivia  1.408  1.001  1.407  South Africa  0.806  0.836  0.964  Bangladesh  1.416  1.070  1.324  Philippines  0.806  0.985  0.819  Saudi Arabia  1.429  1.082  1.322  Egypt, Arab Rep.  0.809  1.137  0.711  Nepal  1.586  1.005  1.578  Jordan  0.811  1.004  0.807  Malta  14.383  1.143  12.579  Mexico  0.812  1.053  0.770            21    Table 7: Determinants of Aggregate Energy Intensity Index, fixed effects estimation, Eq.(7)     (1)  (2)  (3)  VARIABLES  Full sample  Below Above             Fs  1.15***  1.06***  1.11***    (0.017)  (0.019)  (0.025)  Fi  1.11***  0.95***  1.14***    (0.003)  (0.012)  (0.003)    Observations  2,007  471  1,136  R‐squared  0.988  0.950  0.995  Standard errors in parentheses    *** p<0.01, ** p<0.05, *  p<0.1    22    Appendix  Table A1: Correlation of energy consumption per $PPP and GDP per capita PPP for 137 economies  Country  Corr.  n  Country  Corr.  n  Country  Corr.  n  Ethiopia  ‐1,000  25  New Zealand  ‐0,923  25  Montenegro  ‐0,733  10  South Sudan  ‐0,998  3  Romania  ‐0,920  25  Ukraine  ‐0,708  25  Myanmar  ‐0,997  25  Croatia  ‐0,919  20  Namibia  ‐0,687  24  Nigeria  ‐0,996  25  South Africa  ‐0,919  25  Mongolia  ‐0,684  25  Mozambique  ‐0,993  25  Denmark  ‐0,917  25  Côte d'Ivoire  ‐0,679  25  Mauritius  ‐0,993  25  Belarus  ‐0,913  25  Cyprus  ‐0,633  25  Slovak Republic  ‐0,989  23  Luxembourg  ‐0,907  25  Albania  ‐0,630  25  Germany  ‐0,987  25  Indonesia  ‐0,907  25  Georgia  ‐0,628  25  Tanzania  ‐0,985  25  Azerbaijan  ‐0,906  25  Greece  ‐0,626  25  Poland  ‐0,984  25  Cuba  ‐0,902  25  Kuwait  ‐0,619  20  India  ‐0,983  25  Vietnam  ‐0,902  25  Brunei Dar.  ‐0,606  25  Zimbabwe  ‐0,982  25  Botswana  ‐0,895  25  Argentina  ‐0,598  25  Australia  ‐0,982  25  Pakistan  ‐0,894  25  Armenia  ‐0,528  25  Congo, Dem.  ‐0,982  25  Rep.  Turkmenistan  ‐0,893  25  Tajikistan  ‐0,503  25  Latvia  ‐0,981  20  Tunisia  ‐0,892  25  Costa Rica  ‐0,485  25  Nepal  ‐0,981  25  France  ‐0,891  25  Niger  ‐0,355  15  Zambia  ‐0,981  25  Mexico  ‐0,886  25  United Arab Emir.  ‐0,337  25  Nicaragua  ‐0,980  25  Jordan  ‐0,882  25  Benin  ‐0,273  25  Bangladesh  ‐0,979  25  Netherlands  ‐0,881  25  Italy  ‐0,262  25  Uzbekistan  ‐0,979  25  Qatar  ‐0,880  15  Spain  ‐0,250  25  Sri Lanka  ‐0,977  25  Russian Federation  ‐0,877  25  Togo  ‐0,234  25  Angola  ‐0,974  25  Korea, Rep.  ‐0,873  25  Bahrain  ‐0,191  25  Sudan  ‐0,974  25  Cambodia  ‐0,865  20  Eritrea  ‐0,170  20  Finland  ‐0,973  25  Haiti  ‐0,858  19  Guatemala  ‐0,141  25  Estonia  ‐0,972  20  Macedonia  ‐0,854  25  Egypt, Arab Rep.  ‐0,135  25  Sweden  ‐0,970  25  Chile  ‐0,848  25  Lebanon  ‐0,106  25  Czech Republic  ‐0,969  25  Malta  ‐0,847  25  Brazil  ‐0,094  25  Kenya  ‐0,968  25  Turkey  ‐0,846  25  Morocco  ‐0,081  25  Philippines  ‐0,966  25  Suriname  ‐0,841  15  Ecuador  ‐0,068  25  Hungary  ‐0,955  24  Hong Kong SAR, China  ‐0,832  25  Uruguay  ‐0,054  25  China  ‐0,954  25  Kosovo  ‐0,830  15  Singapore  0,060  25  United States  ‐0,954  25  Israel  ‐0,824  25  Portugal  0,077  25  Canada  ‐0,953  25  Honduras  ‐0,822  25  Kyrgyz Republic  0,107  25  Slovenia  ‐0,950  20  Cameroon  ‐0,819  25  Malaysia  0,148  25  Lithuania  ‐0,949  20  Gabon  ‐0,811  25  Jamaica  0,210  25  Norway  ‐0,948  25  Japan  ‐0,809  25  Oman  0,305  25  Colombia  ‐0,947  25  Moldova  ‐0,800  20  Senegal  0,404  25  Switzerland  ‐0,942  25  Belgium  ‐0,792  25  Saudi Arabia  0,496  25  Iraq  ‐0,942  25  Austria  ‐0,788  25  Yemen, Rep.  0,634  25  Ghana  ‐0,941  25  El Salvador  ‐0,778  25  Bolivia  0,659  25  Peru  ‐0,935  25  Serbia  ‐0,764  20  Iran, Islamic Rep.  0,688  25  Panama  ‐0,934  25  Kazakhstan  ‐0,758  25  Algeria  0,703  25  Bulgaria  ‐0,934  25  Bosnia & Herzegovina  ‐0,752  21  Thailand  0,819  25  Libya  ‐0,933  13  Paraguay  ‐0,749  25  Trinidad & Tobago  0,820  25  Ireland  ‐0,931  25  Venezuela, RB  ‐0,747  25  Congo, Rep.  0,897  25  United Kingdom  ‐0,923  25  Dominican Republic  ‐0,745  25          23