Data
Our empirical analysis makes use of data on production, trade, consumption, sectoral CO2 emissions, and carbon footprints from the World Input-Output Database (WIOD) (Dietzenbacher et al. 2013b; Timmer 2012), which pertains to a new generation of global trade databases for tracing flows of carbon embodied in trade along the whole value chain. WIOD was chosen over the EXIOBASE (Tukker et al. 2013), Eora (Lenzen et al. 2012; Lenzen et al. 2013; Kan et al. 2020), and GTAP (Andrew and Peters 2013; Narayanan et al. 2012) because of its homogenous sector classification and its sectoral, spatial, and temporal detail and coverage. The time span 1995–2009 is of interest since developing countries lowered significantly their average tariffs. For a discussion of the relative strengths and weaknesses of these databases, see Dietzenbacher et al. (2013a), Owen et al. (2014), and Tukker and Dietzenbacher (2013). WIOD covers 41 countries (listed in Appendix 33) each containing 35 sectors (listed in Appendix Table 4) over the period 1995–2009. Deducting a few missing observations, this leads to a dataset of roughly 20,000 observations when the sectoral dimension is used and roughly 600 observations when sectors are aggregated at the country level. Our research questions focus on the country (and not the country-pair) perspective. Only Eqs. (8), (11), and (12) in the “Determinants of emission intensities” section, and the corresponding results in the Appendix are performed at the country-pair level.
Income per capita is taken directly from WIOD, and further variables from other sources are used to complement the database. Income, population, and natural resource rents are taken from the WIOD database and the World Development Indicators (World Bank). A dummy variable is also used to indicate whether or not a country has ratified the Kyoto Protocol in a given year. As an alternative to the latter indicator, a CO2 stringency index is borrowed from Sauter (2014) and constructed by counting supra-national, national, and sub-national laws, which explicitly refer to the goal of reducing CO2 emissions.
Empirical methodology
Our empirical methodology derives from a standard input-output analysis (see, e.g., Miller and Blair 2009 for an extensive presentation). In this framework, CO2 emissions from sector s of country i can be expressed as territorial emissions T (also known as production-based) or consumption-based emissions C as follows:Footnote 2
$$ {T}_{is}={e}_{is}{x}_{is}={\varepsilon}_{is}{z}_{is} $$
(1)
$$ {C}_{is}={\varphi}_{is}{y}_{is} $$
(2)
where e represents emission intensity of output, i.e., the quantity of CO2 emitted per unit of output, x represents output, ε represents emission intensity of value added, z represents value added, φ is emission intensity of final demand including embodied carbon emissions (both from national and international intermediate goods), and y is final demand. Note that \( {\sum}_{is}{T}_{is}={\sum}_{is}{C}_{is} \) by definition.
Fig. 2 plots emission intensities of a typical sector in a typical country in 2009, the most recent year available in the dataset. These values were obtained by regressing emission intensities on time fixed effects, country-time fixed effects normalized on average to zero in each year, and sector-time fixed effects normalized on average to zero in each year. Dark labels indicate trade-intensive sectors (i.e., sectors with exports above average), while light labels indicate sheltered sectors (i.e., sectors with exports below average). Emission intensities of value added (ε) are shown on the horizontal axis, while emission intensities of demand (φ) are presented on the vertical axis (both axes in logarithmic scale). We observe that a few sectors are much more emission-intensive than all others. In particular, “Electricity, Gas and Water Supply” (ELCT), “Air Transport” (AIR), “Other non-metallic minerals” (MRLS), and “Water Transport” (WTR)Footnote 3 are classified as the most emission-intensive sectors, both in terms of value added and consumption. No clear-cut picture emerges at this stage concerning the degree of trade exposure and emission intensity. These results do not seem to be driven by the 2008–2009 economic crisis as they also hold for the other years of the observation period.
At first glance, this finding might seem at odd with Fig. 1, which shows that exports are more emission-intensive than final demand. This apparent contradiction is explained as follows: though the most emission-intensive sectors are sheltered, they are small compared to the next group of emission-intensive traded sectors. Specifically, the top 4 sectors in terms of emission intensities (ELCT, AIR, MRLS, WTR) make up only 3.5% of total worldwide final demand in 2009, and thereby contribute a limited amount to the average emission intensity of final demand. Among the trade-intensive sectors, “Coke, Refined Petroleum and Nuclear Fuel” (PTR), “Basic Metals and Fabricated Metal” (MTL), and “Chemicals and Chemical Products” (CHM) are the most emission-intensive sectors, making up 19.3% of total worldwide exports. Hence, these sectors are relatively exposed to trade, large and relatively emission-intensive.
At the world level, T = C by definition, but the two measures differ when for individual countries and individual sectors. For each country i, net CO2 exports (NCO2XP) can then be expressed as:Footnote 4
$$ {\displaystyle \begin{array}{c} NCO2{XP}_i={T}_i-{C}_i={\boldsymbol{e}}_{\boldsymbol{i}}^{\prime }{\boldsymbol{x}}_{\boldsymbol{i}}-{\boldsymbol{\varphi}}_{\boldsymbol{i}}^{\prime }{\boldsymbol{y}}_{\boldsymbol{i}}={\boldsymbol{\varphi}}_{\boldsymbol{i}}^{\prime}\left(\boldsymbol{I}-{\boldsymbol{A}}_{\boldsymbol{i}}\right){\boldsymbol{x}}_{\boldsymbol{i}}-{\boldsymbol{\varphi}}_{\boldsymbol{i}}^{\prime }{\boldsymbol{y}}_{\boldsymbol{i}}\\ {}={\boldsymbol{\varphi}}_{\boldsymbol{i}}^{\prime}\left[\left(\boldsymbol{I}-{\boldsymbol{A}}_{\boldsymbol{i}}\right){\boldsymbol{x}}_{\boldsymbol{i}}-{\boldsymbol{y}}_{\boldsymbol{i}}\right]={\boldsymbol{\varphi}}_{\boldsymbol{i}}^{\prime}\left({\boldsymbol{XP}}_{\boldsymbol{i}}-{\boldsymbol{MP}}_{\boldsymbol{i}}\right)\end{array}} $$
(3)
where I is an identity matrix, Ai is the input-output coefficients matrix, i.e., a matrix where each column indicates the inputs from all sectors needed to produce one unit of output in a given sector, XPi are exports from country i and MPi are imports including imports of intermediate goods to country i. We decompose net CO2 exports (adapting the decomposition by Grossman and Krueger 1993, to the net emission content of trade) into the following three components: trade balance (reflecting the scale effect), sector specialization (reflecting the composition effect), and country-specific emission intensities (reflecting the technique effect):
$$ NCO2{XP}_i=\overline{\varphi}{\boldsymbol{u}}^{\prime}\left({\boldsymbol{XP}}_{\boldsymbol{i}}-{\boldsymbol{MP}}_{\boldsymbol{i}}\right)+\left({\overline{\boldsymbol{\varphi}}}_{\boldsymbol{s}}^{\prime }-\overline{\varphi}{\boldsymbol{u}}^{\prime}\right)\left({\boldsymbol{XP}}_{\boldsymbol{i}}-{\boldsymbol{MP}}_{\boldsymbol{i}}\right)+\left({\boldsymbol{\varphi}}_{\boldsymbol{i}}^{\prime }-{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{s}}^{\prime}\right)\left({\boldsymbol{XP}}_{\boldsymbol{i}}-{\boldsymbol{MP}}_{\boldsymbol{i}}\right) $$
(4)
where φi is the vector of sectoral emission intensities of demand in country i (this is also known as the Leontief multiplier or embodied emissions intensity), \( {\overline{\boldsymbol{\varphi}}}_{\boldsymbol{s}} \) is the vector of world average emission intensities per sector, \( \overline{\varphi} \) is the average emission intensity over all sectors and all countries (i.e., a scalar), and u is a vector of ones.
The first term on the right-hand side of (4) represents the net CO2 trade related to the economic trade balance. This term uses a worldwide average emission intensity of goods. Countries exporting much more than they import, such as China, tend to have a positive first term.
The second term represents the net CO2 trade position related to the sector structure of exports and imports. The term is positive if a country exports in sectors that tend to be emission-intensive and/or it imports in sectors that tend to have low associated emissions. The second term is closely related to the pollution haven debate (Duarte et al. 2018).
The third term represents the net CO2 trade related to differences in the emission intensities between the (exporting) country i and its importing partners. The term is positive if domestic emission intensities exceed the sector world average and/or if the foreign emission intensities from which the country imports are below the sector world average. This term is thus expected to be positive for countries with “inefficient” domestic production, and for countries whose trade partners are emission-efficient. This term measures overall production efficiency of a country relative to its trading partners. A country such as the USA may be emission-intensive compared to the EU, but if it trades more intensely with China, then its relative performance to China matters more for its net trade in CO2 position.
We consider the decomposition in (4) over time in order to identify how the contributions of the three factors evolve. Moreover, looking at the correlations between the different components and their evolution over time indicates whether trade tends to increase or decrease worldwide emissions. For example, a positive correlation between sector specialization and emission intensities (second and third terms) would imply that CO2-intensive countries specialize in CO2-intensive sectors, and more trade is then accompanied by more emissions. Also, if emission-intensive countries tend to exhibit a trade surplus, worldwide emissions would increase with trade, everything else equal. The results of this decomposition are reported and discussed in the “Decomposing CO2 embodied in trade” section below.
Determinants of emission intensities
Fig. 3 displays the relation between income and emission intensity of value added. It shows that production in high-income countries tends to be more emission-efficient compared to that in low-income countries. However, for a given income level, there is wide variability in the emission intensity of production.
Fig. 4 displays the evolution of emission intensities for some large countries. While emission intensities increase and then decrease over the years for Russia and Brazil, they increase (almost) continuously for India and Japan, and decrease (almost) continuously for China.Footnote 5 The USA does not show any significant change in emission intensities. While income is clearly negatively correlated with the level of emission intensities across countries (Fig. 3), the evolution of this relationship within countries over time is much less obvious (Fig. 4).
In order to investigate if and how income, fuel markets, climate policies, and trade opportunities drive changes in emission intensities and in trade patterns, we use the following specifications:
$$ \mathrm{EIVA}:\ln \left({\varepsilon}_{ist}\right)={\beta}^{VA}{Z}_{it}+{\gamma}_i+{\delta}_{st}+{\mu}_{ist} $$
(5)
$$ \mathrm{EID}:\ln \left({\varphi}_{ist}\right)={\beta}^D{Z}_{it}+{\gamma}_i+{\delta}_{st}+{\mu}_{ist} $$
(6)
where εist is emission intensity of value added (EIVA) in sector s of country i at time t, φist is emission intensity of demand inclusive of embodied emissions (EID), Zit includes country variables such as income, fossil fuel income shares, and policies. The effect of these variables is identified through different trends between countries, as time fixed effects are absorbed by the sector-time fixed effects δst, and time-invariant country characteristics are absorbed through country fixed effects γi, while μist is the remaining idiosyncratic noise. Depending on the variables included in Zit, the estimated coefficients β can answer questions such as whether domestic fossil fuel abundance, Kyoto policies, and trade opportunities tend to increase or decrease emission intensities.
To gain further insights, we test alternative measures of emission intensity that are relevant in the context of trade:
$$ \ln \left(\frac{{\boldsymbol{\varphi}}_{\boldsymbol{it}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{it}}}{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{it}}}\right)={\beta}_1^X\ {Z}_{it}+{\gamma}_i+{\delta}_t+{\mu}_{it} $$
(7)
$$ \ln \left(\frac{{\boldsymbol{\varphi}}_{\boldsymbol{it}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{ijt}}}{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{ijt}}}\right)={\beta}_2^X\ {Z}_{it}+{\gamma}_i+{\delta}_{jt}+{\mu}_{ijt} $$
(8)
The left-hand side variable in (7) measures emissions of country i exports (computed as the product between the (transposed) vector of country-sector emission intensities with the corresponding export flows), relative to emissions for an average country (i.e., using the world average vector for emission intensity) with the same sector structure of exports (i.e., multiplying with country i’s export structure). The dependent variable in (8) is similar, but specified for each bilateral country-pair: XPijt represents exports from country i to country j during year t. In this case, we control for country and partner-year fixed effects. These two dependent variables are closely related to the third term of (4) and these two equations will give insights in the factors explaining country-specific emission intensities.
We then investigate the sectoral structure of trade by estimating the following four equations:
$$ \ln \left(\frac{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{it}}}{{\overline{\varphi}}_t{\boldsymbol{u}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{it}}}\right)={\beta}_1\ {Z}_{it}+{\gamma}_i+{\delta}_t+{\mu}_{it} $$
(9)
$$ \ln \left(\frac{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{MP}}_{\boldsymbol{it}}}{{\overline{\varphi}}_t{\boldsymbol{u}}^{\prime }{\boldsymbol{MP}}_{\boldsymbol{it}}}\right)={\beta}_2\ {Z}_{it}+{\gamma}_i+{\delta}_t+{\mu}_{it} $$
(10)
$$ \ln \left(\frac{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{ijt}}}{{\overline{\varphi}}_t{\boldsymbol{u}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{ijt}}}\right)={\beta}_3\ {Z}_{it}+{\beta}_4{Z}_{jt}+{\gamma}_i+{\lambda}_j+{\delta}_t+{\mu}_{ijt} $$
(11)
$$ \ln \left(\frac{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{ijt}}}{{\overline{\varphi}}_t{\boldsymbol{u}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{ijt}}}\right)-\ln \left(\frac{{\overline{\boldsymbol{\varphi}}}_{\boldsymbol{st}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{jit}}}{{\overline{\varphi}}_t{\boldsymbol{u}}^{\prime }{\boldsymbol{XP}}_{\boldsymbol{jit}}}\right)={\beta}_5\ \left({Z}_{it}-{Z}_{jt}\right)+\left({\gamma}_i-{\gamma}_j\right)+{\mu}_{ijt} $$
(12)
All dependent variables in these equations are measures of sector structure and are linked to the second term in (4). The dependent variable in (9) measures the sector bias of exports towards emission-intensive sectors, i.e., how the export structure of country i causes its emission intensity to differ from the average. In (10), we consider an equivalent variable for imports. In (11), the dependent variable measures the sector bias for all country-pairs of bilateral trade, considering each country-pair in both ways (i is both an exporter to j and an importer from j). Coefficient λj represents partner fixed effects. In (12), we subtract exports from country j to country i (imports in country i from country j) from exports from country i to country j to obtain a symmetric indicator equivalent to (11) for relative emission intensities in net exports (export-imports). We expect β5 to be about equal to β3 − β4. Note that the country-partner fixed effects in (12) are structured so that their number is equal to the number of countries, and not to the number of country-partner pairs.
Controlling for unobserved endogeneity and weighting observations
The objective of our analysis is to investigate whether an increase in fossil fuel rents (e.g., coal) tends to increase or decrease the emission intensity of production (5), consumption (6), and exports (7)-(8), and whether it alters the sector structure of trade (9)-(12). However, reverse causality could also arise: an increased demand for emission-intensive sectors leads to higher fossil fuel prices, and thus to higher fossil fuel rents. Therefore, we implement a “shift-share” approach similar to that in Allcott and Keniston (2018) and Bartik (1991). For each country, we calculate the share of that country i, over the entire period, in worldwide fuel rents: \( {s}_i^c \). In addition, for each year t, we calculate the global fuel rents as a share of world GDP: \( {R}_t^g \). The interaction between the country’s share and the world fuel rents is used as an independent variable instead of the country’s fossil fuel rent:
$$ {R}_{it}={s}_i^c{R}_t^g $$
(13)
By construction and assuming that country i’s influence on total world resource rents is sufficiently small, this interaction cannot suffer from reverse causality: an increase in fossil fuel demand in one country in 1 year will have no effect on the interaction term for that country in that year. This seems a plausible assumption to the best of our knowledge.
We also use trade openness as an independent variable in our estimations. Similarly, to avoid endogeneity, we consider openness in our estimations through the interaction between a country’s average openness over the entire period and the world trade share in world GDP, for each year.
The approach outlined above, inspired by Allcott and Keniston (2018), uses the interactions between the country’s share and the world fuel rents directly in the equation of interest. This methodology relies on a single-equation methodology, and therefore avoids any cross-influences of the various instruments on the endogenous variables taking place in a standard two-stage approach.Footnote 6
Depending on the dimensions of the dependent variable (country-sector-year-partner), we use the corresponding fixed effects in order to control for unobserved heterogeneity (see results’ tables for details). We do not include scaling variables such as population or GDP since our dependent variable reflects emission intensities and not emissions.
We conduct both weighted and unweighted regressions. Weighting is warranted if we expect observations concerning large trade flows to have better quality, in relative terms, compared to observations concerning small trade flows. Another way to interpret differences between weighted and unweighted estimations is that the former indicates marginal effects for the weighted average observation, while the latter applies to the unweighted average observation. The two outcomes will differ when large countries behave systematically differently compared to smaller ones.