Introduction

Over the past two decades both the world GDP and total amount of CO2 emissions have increased (see Fig. 5). With regard to GDP, Europe was surpassed by Asia in 2011 and in 2015. Overall, the Asian GDP has experienced the largest growth throughout the whole time-span (see Fig. 5). Regarding the emissions, the growth rate of Europe and North America remained close to zero until 2008 and became negative afterwards; Asia, instead, has displayed an increasing trend throughout the entire period (see Fig. 5). Although the 2008 financial crisis and the Covid-19 pandemic had clear consequences on both trends, the observations above suggest that reducing the environmental impact of economic growth remains a challenging goal.

As emissions from burning fossil fuels are the primary cause of global warming1, the last decades have witnessed significant efforts to mitigate carbon emissions. At country level, the Intergovernmental Panel on Climate Change (IPPC) accounts emissions according to a ‘production principle’, i.e. by attributing them to countries producing goods and services2,3.

Recent literature, however, has proposed to consider the adoption of a consumption-aware accounting, prescribing to focus on ‘final consumers’ as well4,5,6,7: in other terms, this stream of literature advocates for accounting the carbon embedded into imported goods and services, hence re-distributing the responsibility for these emissions between producers and users.

Studies have estimated the extent of the out-sourcing phenomenon on the basis of regressions (solely) accounting for country-specific factors such as GDP and energy efficiency8,9,10,11,12—and invoking assumptions often leading to contradictory policy recommendations13,14,15,16. Other authors have attempted to understand to what extent a ‘consumption principle’ may help reducing disparities in carbon accounting17,18: in such studies, the amount of traded carbon is estimated via Multi-Regional Input-Output (MRIO) tables19,20 that, however, are very sensitive to the accuracy of the available data on trading sectors21,22.

As noticed by Caro et al.23, data requirements can be relaxed by considering aggregate measures: specifically, the amount of carbon embedded into a country export can be simply quantified by multiplying it by the related GDP-induced carbon intensity (named National Carbon Intensity in23), defined as

$$\begin{aligned} \text {GDP-CI}_i^y=\frac{[\text {CO}_2]_i^y}{\text {GDP}_i^y} \end{aligned}$$
(1)

and measuring the kilograms of carbon, per dollar of GDP, released by country i, during a given year y (see also Appendix A): although less accurate, this method overcomes many of the aforementioned problems while keeping the uncertainty accompanying estimates in a suitable range for quantitative analysis (as the authors of23 explicitly acknowledge, their ‘[...] systemic approach, in which carbon intensity plays the role of national indicator relative to production efficiency [...] is very easy, not labor-intensive to implement and no further data is needed beyond those already available at the national level’).

Inspired by the work of Caro and co-authors, we have adopted a complex network approach24,25,26,27 and constructed a Carbon Trade Network (CTN), i.e. a graph induced by the trade-embedded carbon exchanges between world countries, with the aim of investigating its complex architecture over the past two decades. More specifically, we have focused on (1) the evolution of the GDP-CI of each country and of its nearest neighbours; (2) the geographic distribution of the differences between production and consumption-based emissions at country scale; (3) the direction and magnitude of fluxes within and between groups of countries.

Data and methods

Construction of the carbon trade network

To construct the CTN, we have combined (1) data on trade flows from UN-COMTRADE (see https://comtradeplus.un.org/), (2) data on GDPs from the World Bank (see https://data.worldbank.org/), (3) data on \(\text {CO}_2\) emissions from https://ourworldindata.org/. To consistently compare data over the years 2000–2020, we have selected a panel of 111 countries for which trade information was available for the entire period. To the best of our knowledge, the dataset employed to carry out the present study represents a quite unique example in the literature, for sample size, time span and granularity (OCSE reports usually focus on G20 countries over few years).

Following28, carbon emissions are embedded into trade exchanges by considering the yearly values of each country GDP-CI. In formulas, the CTN link weights read

$$\begin{aligned} c_{ij}^y=\text {GDP-CI}_i^y\cdot w_{ij}^y,\quad \forall \,i\ne j \end{aligned}$$
(2)

with \(w_{ij}^y\) indicating the export, in US dollars, from country i to country j, during a given year y. While the out-strength of node i, defined as \(t_i^{out}=\sum _{j(\ne i)=1}^Nc_{ij}\) (where we have dropped the y index), quantifies the total amount of its exported emissions, the in-strength of node i, defined as \(t_i^{in}=\sum _{j(\ne i)=1}^Nc_{ji}\) (where we have dropped the y index), quantifies the total amount of its imported emissions. Then, the total weight \(W=\sum _{i=1}^Nt_i^{out}=\sum _{i=1}^Nt_i^{in}\) quantifies the total amount of trade-embedded carbon emissions (see also Appendix A).

Nominal VS constant-price GDP

Evaluating the carbon intensity requires choosing a definition of GDP. Hereby, we compare the nominal with the 2015 constant-price one. As Fig. 12 shows, the differences between the nominal GDP-CIs and the 2015 constant-price GDP-CIs are larger during the first years of our dataset, while vanishing as we approach 2020 (see also below). Since employing the latter does not change the overall picture returned by our results, we follow23 and stick to the nominal definition of GDP.

Figure 1
figure 1

Left panel: the ‘economic-environmental trajectory’ of each country emerges upon scattering the tons of carbon released by it versus its GDP, in a yearly fashion. For G20 countries, two tendencies can be identified: the one characterising countries whose GDP and amount of emissions are positively correlated (i.e. Argentina, Brazil, China, India, Indonesia, Korea, the Russian Federation, Saudi Arabia, South Africa and Turkey—moving towards the top-right of the plane) and the one characterising countries whose GDP and amount of emissions are negatively correlated (i.e. France, Germany, Italy, Japan, Spain, United Kingdom and the US—moving towards the bottom-right of the plane). Right panel: scattering the weighted mean of the GDP-CIs of each country exporting partners versus its own GDP-CI reveals that many ‘net exporters’ are less economically efficient than the countries they export to; analogously, many ‘net importers’ are more economically efficient than the countries they import from. Overall, this leads us to conclude that countries whose export exceeds the import, export towards ‘cleaner’ countries; equivalently, countries whose import exceeds the export, import from ‘less clean’ countries. The size of ‘net exporter’ i is proportional to \([t^{out}]_i^y/\text {PE}_i^y\), i.e. the amount of exported emissions over the amount of produced emissions; the size of ‘net importer’ i is proportional to \([t^{in}]_i^y/\text {PE}_i^y\), i.e. the amount of imported emissions over the amount of produced emissions. Numbers are plotted on a doubly logarithmic scale. Names of countries indicate the last year covered by our dataset, i.e. 2020.

Full size image
Figure 2
figure 2

Top panel: geographic distribution of the differences between consumed and produced emissions, defined as \(\text {CE}_i^{2020}-\text {PE}_i^{2020}\equiv \Delta _i^{2020}\), \(\forall \,i\): while ‘net importers’ (or ‘consumers’) of emissions are depicted in shades of red, ‘net exporters’ (or ‘producers’) of emissions are depicted in shades of blue. Bottom-left panel: histogram of the differences between consumed and produced emissions, during the year 2020. Our analysis reveals that France, Germany, Italy, Japan, the UK and the US are among the top ten ‘net consumers’ while China, India, the Russian Federation and South Africa are among the top ten ‘net producers’—a classification that is robust across time. Bottom-right panel: histogram of the percentage differences between consumed and produced emissions, defined as \([\text {CE}_i^{2020}-\text {PE}_i^{2020}]/\text {PE}_i^{2020}\equiv \Delta _i^{2020}/\text {PE}_i^{2020}\), \(\forall \,i\). The ranking changes because of the normalisation: in this case, in fact, the top ‘net consumers’ are the countries with a low level of internal production (e.g. the islands) while the top ‘net producers’ are the countries with a low level of import.

Full size image

Results

The carbon released by a nation versus its GDP sheds light on the environmental impact of its economic development. For G20 countries, we observe two, different kinds of evolution (see Figs. 1 and 6): the first one characterises countries whose values of GDP and amount of emissions are positively correlated, i.e. Argentina, Brazil (although their trend shows an inversion in 2018 and 2014, respectively), China, India, Indonesia, the Russian Federation (although the overall amount of its emissions has risen at a much lower rate than others’), Saudi Arabia, South Africa and Turkey; the second one characterises countries whose values of GDP and amount of emissions are negatively correlated, i.e. France, Germany, Italy, Japan, Spain, United Kingdom and the US. Finally, the Australian, Canadian, Korean and Mexican GDPs have risen as well: however, the amount of Canadian emissions has remained quite constant over time; the Australian and Korean ones have risen up to 2008 and 2011, respectively, and become flat afterwards; the Mexican one has risen up to 2012 and decreased afterwards.

As Fig. 12 shows, the effect of the inflation is larger for the first years of our dataset; as we approach 2020, however, the values of the GDP-CI calculated by employing the 2015 constant-price GDP become closer to the values of the GDP-CI calculated by employing the nominal GDP. This, in turn, leads to Fig. 13, closely resembling Fig. 1. More quantitatively, the average of the relative errors

$$\begin{aligned} \delta _{\text {GDP-CI}_i^y}=\left| \frac{\text {GDP-CI}_\text {n}-\text {GDP-CI}_\text {cp}}{\text {GDP}_\text {n}}\right| _i^y\cdot 100 \end{aligned}$$
(3)

for the G20 countries, amounts at \(\lesssim 40\%\) in 2000, \(\simeq 15\%\) in 2010 and \(\simeq 8\%\) in 2020.

The analysis carried out so far merely depicts the evolution of aggregate indicators (see also Fig. 7). To disentangle the role played by trade, the network representation of the carbon embedded into the exchanges between world countries provides more insights. As Fig. 8 shows, the total weight of the CTN has decreased throughout the second half of our time span, a result indicating that, from 2011 onwards, the impact of trade on world emissions has progressively diminished: one may be, thus, tempted to conclude that the rising trend in Fig. 4 is solely due to the carbon emitted for internal production; as we will see, this is only partially true.

Figure 3
figure 3

Partitioning the group of G20 countries into ‘net exporters’ and ‘net importers’ allows the presence of a net flux of emissions, directed from (the members of) the first set towards (those of) the second, to be revealed. Although its magnitude has, overall, risen over the past twenty-one years, the amount of emissions directed from ‘net exporters’ towards ‘net exporters’ has become increasingly relevant.

Full size image

In order to unambiguously classify each country on the basis of its trading behaviour, let us follow Caro et al.23,28 and re-write the amount of consumption-based emissions (or ‘consumed emissions’—the two terms will be used interchangeably) by country i, during the year y, as \(\text {CE}_i^y=\text {PE}_i^y+\left[ t^{in}\right] _i^y- \left[ t^{out}\right] _i^y\) (see also Appendix B), where CE and PE are the acronyms for ‘consumed emissions’ and ‘produced emissions’, respectively. As a consequence, the difference between the amount of consumed and produced emissions by it is

$$\begin{aligned} \text {CE}_i^y-\text {PE}_i^y=\left[ t^{in}\right] _i^y- \left[ t^{out}\right] _i^y\equiv \Delta _i^y, \end{aligned}$$
(4)

a relationship allowing us to distinguish ‘net importers’ of emissions, characterised by \(\Delta _i^y>0\) (equivalently, \(\text {CE}_i^y>\text {PE}_i^y\bigg )\), ‘net exporters’ of emissions, characterised by \(\Delta _i^y<0\) (equivalently, \(\text {CE}_i^y<\text {PE}_i^y\bigg )\), and countries that have reached the ‘trade-embedded carbon neutrality’, characterised by \(\Delta _i^y=0\) (equivalently, \(\text {CE}_i^y=\text {PE}_i^y\bigg )\). According to Eq. (4), ‘net importers’ (‘net exporters’) can be also classified as ‘net consumers’ (‘net producers’) of emissions. Figure 2 depicts the distribution of the set of values \(\left\{ \Delta _i^{2020}\right\} \), i.e. of the differences between out-strength and in-strength for each country, during the year 2020: as our analysis reveals, France, Germany, Italy, Japan, the UK and the US are among the top ten ‘net importers’ while China, India, the Russian Federation and South Africa are among the top ten ‘net exporters’.

As an additional consistency check, the estimates of the emissions computed by using our method have been compared with the estimates of the ‘\(\text {CO}_2\) emissions embodied into the domestic, final demand’, computed by OECD as the amount of carbon that is emitted for domestic production plus the amount of carbon that is emitted abroad and embodied into imports (see https://stats.oecd.org). More quantitatively, the average of the relative errors

$$\begin{aligned} \delta _{\text {CE}_i^y}=\left| \frac{\text {CE}-\text {CE}_\text {OECD}}{\text {CE}}\right| _i^y\cdot 100 \end{aligned}$$
(5)

for the G20 countries, amounts at \(\simeq 6\%\), throughout the whole period of time considered here (see also Fig. 14).

In order to check the robustness of the aforementioned classification over time, let us consider the sign of the temporal average \(\overline{\Delta _i}\equiv \sum _{y=2000}^{2020}\Delta _i^y/21\), allowing us to distinguish the countries that have mainly served as ‘net exporters’ (i.e. for which \(\overline{\Delta _i}<0\)) from the countries that have mainly served as ‘net importers’ (i.e. for which \(\overline{\Delta _i}>0\)): within G20, Argentina, Canada, China, India, Indonesia, Korea, Mexico, the Russian Federation, Saudi Arabia and South Africa belong to the first group while Australia, Brazil, France, Germany, Italy, Japan, Spain, Turkey, the UK and the US belong to the second group (see also Fig. 9).

Let us, now, study the mutual connections between the two, aforementioned groups of countries. More specifically, let us compare the behaviour of country i (be it a ‘net exporter’ or a ‘net importer’) with that of its partners, by calculating the average value of the GDP-CIs of its neighbours. To this aim, we have distinguished the nodes pointed by it (i.e. the countries it exports to) from the nodes pointing towards it (i.e. the countries it imports from). Figure 1b shows the evolution of both kinds of trajectories. Let us focus on ‘net exporters’, first: their trajectories lie below the identity line, a result indicating that their GDP-CI is steadily larger than the one of the countries they export to; for what concerns ‘net importers’, instead, the converse is true: their trajectories lie above the identity line, a result indicating that their GDP-CI is steadily smaller than the one of the countries they import from. Overall, this suggests the presence of a flux of emissions, directed from the countries for which \(\overline{\Delta _i}<0\) towards the countries for which \(\overline{\Delta _i}>0\) (see also Fig. 11).

To gain further insight into this, let us analyse the evolution of the total amount of carbon emissions embedded into the export of ‘net-exporters’. As Fig. 3 shows, both the portion of it directed towards ‘net importers’ and the one directed towards ‘net exporters’ have increased; still, the share of emissions directed towards ‘net importers’ has diminished, a result indicating that what may be called ‘exp-to-exp’ emissions (i.e. the emissions embedded into the trading relationships directed from ‘net exporters’ towards ‘net exporters’) have become increasingly relevant (see also Fig. 10).

When considering the evolution of the GDP-CIs, this result may appear paradoxical: each country has, in fact, reduced its own GDP-induced carbon intensity, the major decrease being observable for ‘net exporters’ - specifically, the Russian Federation, South Africa (whose GDP-CI, in 2020, lies slightly above \(1\,\text {kg}/\$\)), China, India and Saudi Arabia (whose GDP-CI, in 2020, lies between \(1\,\text {kg}/\$\) and \(0.6\,\text {kg}/\$\)): as a consequence, one would expect the total amount of their trade-embedded carbon emissions to decrease as well. Its rise, only apparently contradictory, is due to an increase of the trading activity involving these countries, causing the related emissions to grow even though each actor has (individually) become more efficient.

Discussion and policy perspectives

While limiting ourselves to inspect the evolution of the GDP-CIs leads to the conclusion that each country has improved its economic efficiency, a network analysis of trade flows reveals the presence of a flux of emissions, directed from G20 countries serving as ‘net exporters’ towards G20 countries serving as ‘net importers’, whose magnitude has been rising over the past twenty-one years. In other words, our analysis reveals the presence of an unexpected, positive feedback: the rise of trading activity among countries has caused the amount of emissions to rise as well, although exchanges have (individually) become less carbon-intensive.

In a wider perspective, our results highlight the systemic dimension of the carbon leakage phenomenon, stressing the role played by the core countries of the international trade network of carbon exchanges29. Although the aggregation level of our study does not allow us to quantify the leakage rate for our panel countries, it may nonetheless provide important insights for improving general equilibrium models like GTAP-E (e.g. for what concerns the definition of unilateral carbon policies). Results also support the need to integrate embodied carbon for designing effective policy instruments, suggesting that actions to reduce emissions solely targeting production methods in a restricted number of developed countries may be ineffective; stated otherwise, policies should explicitly account for the systemicness of the carbon leakage phenomenon as well as its national specificities—thus, refining the scenario discussed by Beck et al.29, according to whom, in case of no aversion towards carbon leakage, the cost of achieving national emission targets should be minimised by adopting a uniform carbon taxation.

Our analysis of the Carbon Trade Network leads to results that are consistent with those of the econometric analysis performed by Liddle8, who identify a set of Asian countries for which exports/imports provide a significant contribution to lower/increase consumed emissions and conclude that a consumption-aware accounting may indeed be helpful to assess responsibility for climate change. Otherwise stated, a ‘production-based’ accounting criterion solely penalises the countries belonging to the group of ‘net exporters’: ‘net importers’, on the other hand, may take advantage of the current situation, by lowering their emissions as a consequence of practices such as that of off-shoring carbon-intensive productions, instead of adopting ‘environment-friendly’ technologies30,31. In this case, the global carbon footprint would be left unchanged—if not worsened—since off-shoring is typically directed towards technologically underdeveloped, hence highly polluting, countries. As signalled by several sources, this practice seems to characterise France32, the United Kingdom33 and the US34.

Coming to comment on policy instruments, consumption-based emissions may inform taxation, as a ‘consumption-aware’ criterion would burden both the ‘net exporting’ and ‘net importing’ countries, incentivising developing countries to transition towards cleaner industrial production: a measure going in this direction is the Carbon Border Adjusting Mechanism (CBAM), recently introduced by the European Union. Still, its expected effectiveness should be carefully evaluated as European firms may experience a reduction of competitiveness and developing countries with limited access to green technologies may be overburden; besides, the GDP-CIs tend to assume increasingly similar values: hence, the impact of the CBAM may reduce in the medium-long period35.

A future direction along which the present analysis could be extended concerns the possibility of employing carbon intensity in a disaggregated fashion (as the authors of23 explicitly acknowledge, ‘[...] the framework presented is less detailed than the EEIO framework, as it does not use a specific carbon intensity for each sector and does not include the indirect emissions’.). Here, however, an issue with data availability arises. As Caro and co-authors suggest, one may consider the sector-specific carbon intensities; a network approach like the one pursued here, however, would focus on the World Trade Multiplex, each layer of which corresponds to a commodity: an aggregation of commodities into sectors and a mapping between the corresponding carbon intensities would, thus, be required.