1

Scale, Technique and Composition Effects in Manufacturing SO2 Emissions1

Jean-Marie Grether, University of Neuchâtel2

Nicole A. Mathys, University of Lausanne and EPFL

Jaime de Melo, University of Geneva, CERDI and CEPR

January 31, 2008

Abstract Combining two data sources on emissions with employment data, this paper constructs six data bases on sulfur dioxide (SO2) intensities that vary across countries and sectors. This allows to perform a growth decomposition exercise where the change in world manufacturing emissions is decomposed into scale, composition and technique effects. The sample covers the period 1990-2000, and includes 62 countries that account for 76% of world-wide emissions. While manufacturing employment has increased by 10% (scale effect), we estimate that emissions have fallen by about 10%, thanks to the combined influence of cleaner production techniques (a technique effect of -15%) and a shift towards cleaner industries and towards cleaner countries (i.e. between-sector and between-country composition effects of -5%). The paper also shows that these estimates are robust to changes in aggregation across entities (regions or countries) and across industries. These emission intensities do not exhibit a robust association with GDP per capita (no EKC pattern).

JEL Classification Number: F11, Q56 Key Words: manufacturing activities, bottom-up approach, SO2 emissions

1 Financial support from the Swiss National Science Foundation under research grant No 100012-109926 is gratefully acknowledged. An earlier version of this paper was presented at the ETSG 2006 conference in Vienna. We thank conference participants for comments. An Appendix describing the data and detailing the construction of the data base is available from the authors upon request. We thank Werner Antweiler and Robert Elliott for providing us with data and for very helpful comments. All remaining errors are our sole responsibility. 2 corresponding author; address: Institute for Economic Research, University of Neuchâtel, Pierre-à-Mazel, CH-2000 Neuchâtel. Tel +41 32 718 13 56, Fax: +41 32 718 14 01, e-mail: jean-marie.grether@unine.ch

2

1. Introduction It has long been recognized that our understanding of the relationship between the

economy and the environment is limited by data availability, notably with respect to

the growth-environment nexus. Although very generally accepted, in empirical and

theoretical discussions, the scale, composition and technique effects have failed so far

to materialize into robust, let alone stylized facts (see e.g. Grossman and Krueger

(1991) for an early application and the recent survey by Brock and Taylor (2005) for

extensive discussion in growth models). In the absence of emissions linked to human

activity--the so called emission intensities used in the bottom-up approach--

researchers have used average concentrations (e.g. Antweiler, Copeland and Taylor

(2001)) or total emissions per capita (Cole and Elliott (2003)) in panel aggregate

emission growth regressions. As a result, the respective role of composition effects

within and across countries identified in the literature, have been obfuscated. Thus,

whether there has been a shift towards “dirty” activities or “dirty” countries has

largely remained elusive even for SO2 emissions which have desirable characteristics

for studying compositional shifts and more generally the growth-environment nexus

(strong local effects, a by-product from goods production, and differences in per-unit

emissions across industries).

The Emission Database for Global Atmospheric Research (henceforth EDGAR) gives

information on emissions for several pollutants across sectors for three years (1990,

1995 and 2000) and provides a starting point for relating more systematically the

scale of economic activity in different sectors with emissions when it is combined

with the more complete data set of an earlier project, the Industrial Pollution and

Projection System (IPPS). We use these two data sources, and national emissions

over a large number of countries in Stern (2006) (hereafter STERN), to construct

disaggregated data bases on emission across sectors, countries and over time for

sulfur dioxide (SO2) emissions. Emission intensities by activity are then constructed

by reconciliation with and extension of a recent Trade, Production and Protection

(TPP) database (Nicita and Olarreaga, 2007) resulting in a potential set of six

databases that vary in terms of geographic (62 countries or 6 regions) or sectoral (7 or

28 sectors) disaggregation.

These data bases are then used to decompose changes in SO2 emissions into scale,

composition and technique effects over the period 1990-2000. Whatever the

database, it turns out that there has been a structural shift towards cleaner products

3

and cleaner countries during the nineties. These composition effects have been

instrumental in helping to curb down global SO2 emissions in spite of the increase in

global manufacturing activity during this period.

The remainder of the paper is organized as follows. Section 2 discusses the difference

between emissions (most suitable to measure scale, technique and composition

effects) and concentration estimates (most suitable for assessing the welfare

implications of emissions) and presents a summary comparative review of the

existing data sources for SO2 emissions that have been used in the recent literature.

Section 3 details how the different data bases were constructed. Section 4 then turns

to a decomposition of the evolution of SO2 emissions from 1990 to 2000. Section 5

concludes.

2. Comparing pollution data for SO2 at the aggregate level

Concentration data of sulfur dioxide in the air are probably the most appropriate

indicator to assess the impact of pollution on the environment or on human health.

However, they are very difficult to relate to economic activity and this for two basic

reasons. First, although distance from emission sources is a fundamental

determinant of concentration levels, the precise geographical locations of both

production and observation sites are usually either fragmentary or missing

altogether. This is one of the main obstacles to undertake convincing cross-country

studies on the basis of concentration data. Second, even with a perfect knowledge of

the location of human activities and of observation sites where concentration is

measured, the relationship between the two is still a complex one because of the

number of other factors that have to be controlled for. So, concentration levels are not

only affected by non-human and non-industrial sources (e.g. volcanic activity or

domestic heating), or by the type of measurement equipment; they also depend on

weather conditions such as the average wind, temperature or rainfall at the site

(rainfall typically reduces concentrations). Some of these effects are site-specific and

can be controlled for with dummy variables (provided the sites do not change), while

some, like weather-related effects, are time-varying and hence harder to control for.

To circumvent these difficulties, the alternative is to use emission rather than

concentration data. By definition, emissions are directly linked to specific activities,

4

and available data are suitable to analyze the production-pollution link both, at a

large scale across countries, and at a reasonably disaggregated level. As this paper

illustrates, this is particularly true regarding manufacturing emissions, which

represent a substantial part of global man-made (i.e. anthropogenic) emissions in the

case of SO2 (approximately 45%), and are especially policy-relevant since

manufacturing activities can be the target of environmental, trade and industrial

policies. However, unlike concentration data, emission data are not directly

observable, and have to be constructed. This may generate differences in estimates

between sources, depending on estimation techniques or assumptions with respect to

production or abatement technologies. The remainder of this section presents basic

stylized facts regarding concentrations and emissions, both at the global and at the

national level, and discusses in greater detail the available data sources for

manufacturing emissions.

2.1 Emissions and concentrations at the world level

As a starting point, it is instructive to examine the correlation between concentration

and emission data at the global level where it is not necessary to control for location-

specific characteristics. In the case of a lack of correlation, a focus on emissions

would loose good part of its relevance, at least for drawing the welfare implications of

emissions. Fortunately, as illustrated by figure 1, there is evidence of a strong link

between the two for a common sample of 15 countries used in two recent studies on

SO2 emissions (Antweiler, Copeland and Taylor (2003), ACT hereafter, for

concentrations, and STERN for emissions). In this sample, the authors report non-

missing values for more than 75% of the observations over the 1976-1992 period (see

table A1 in Appendix 1, which reports the composition of the various country samples

mentioned in this paper). In figure 1, global emissions are reported on the left-hand

scale, and (unweighted) average concentration on the right-hand scale. Both

measures indicate a decline in pollution for the sample across the period. The simple

correlation between the two series is 0.84, climbing up to 0.92 when the

concentration average is calculated using land surface as weights across countries

(national concentration is estimated by the unweighted average across all observation

sites).

5

How representative is this "world" sample which represented approximately 50% of

world emissions in 1975 according to the STERN data base (i.e. emissions for 152

countries with non-missing values across the 1975-2000 period), but fell to about

25% by 2000? An answer to this question is difficult because, unfortunately,

concentration data are not available for a larger sample. Moreover, as shown below

(and as expected from the above introductory discussion), the strong correlation

between emissions and concentration breaks down when the analysis is carried out

across countries.

Figure 1: Concentrations and Emissions at the world-wide level

.005

.01

.015

.02

aver

age

conc

entra

tion

in p

pm

3035

4045

50to

tal e

mis

sion

s in

Tg

1975 1980 1985 1990 1995 2000year

Stern - 15 ACT - 15, unweightedACT - 15, weighted

To get a better feel of the correlation between emissions and concentrations, figure 2

also reports total emissions for two larger groups of countries considered in this

paper that allows for comparisons across data sources. One is the 26 OECD countries

sample considered by Cole and Elliott (2003) which accounts for slightly over 50% of

world emissions, a share that--contrarily to the ACT-15 group--remains roughly

constant across the sample period. Note that for the four common years (1975, 1980,

1985, 1990), the values reported by Cole and Elliott are almost identical to those of

STERN. The other is for a group of 62 countries that is covered in the World Bank

Trade, Production and Protection database (hereafter TPP) of Nicita and Olarreaga

(2007). The TPP group represents 76% of world emissions across the sample period.

6

Until now, we referred to "world" emissions as those obtained for the 152 STERN

countries with non-missing values across the 1975-2000 period. An alternative, used

in the decompositions presented later on, is to use a sample of 233 countries

(“EDGAR countries”), with available emission data for 1990, 1995 and 2000. As

shown in figure 2, the relative importance of the various country groups is hardly

affected by this alternative benchmark, even though whatever the group, EDGAR

totals locate 10-20% above STERN estimates (see also table A1 in Appendix 1 for a

listing of 1990 individual emission shares by country). According to Stern (2006),

this gap is essentially due to a better control of abatement activities in the STERN

series insofar as the Stern data follows more closely national statistics (which may be

suspicious in some cases, e.g. China (see Stern (2006) and Grether et al. (2007) for

discussion). A second difference between the samples is the evolution of the temporal

pattern during the 90s. Although all groups exhibit a downward trend according to

STERN values, the total emissions of the middle groups (26 or 62 countries) have

been increasing according to EDGAR. This should be kept in mind as the emission

intensities estimated here are based on EDGAR data (the unique source with

disaggregated data at the sector level) for the 62 (TPP) country group.

Figure 2: Different sources of "world" emissions

2550

7510

012

515

0to

tal e

mis

sion

s in

Tg

1975 1980 1985 1990 1995 2000year

Cole - 26 Stern - 15Stern - 26 Stern - 62Stern - 152 Edgar - 15Edgar - 26 Edgar - 62Edgar - 152 Edgar - 233

7

2.2 Emissions and Concentrations: Cross country comparisons

How robust are these patterns at the country level? Table 1 reports the simple and

rank correlations between concentrations and emissions for each one of the ACT-15

countries (with a minimum of 13 yearly observations across the 1976-1992 period).

For four of the six largest economies (USA, Germany, Canada and Spain), the

correlation remains positive and significant, but for two of them (Brazil and Japan), it

turns out to be significantly negative. The remaining countries exhibit either a weakly

significant and positive link (Poland and New Zealand) or a non-significant one. The

particular cases of Brazil and Japan are opposite (emissions increase over time for

the former, while they decrease for the latter) and exemplify why concentration

measures must be interpreted with caution (see figures A1, A2 and A3 in Appendix 1).

Six out of seven Japanese stations register a suspicious sudden increase at the

beginning of the 80s and a sudden drop in 1992 (abstracting from these shifts the

trend is a downward one, as for emissions). For Brazil, the downward trend is more

common to all stations, but the data are very patchy, and may not correspond at all to

the location of the main emission sources (the number of measurement stations is

identical to that of Japan, while land area is twenty times larger).

Table 1: Correlation between concentrations and emissions at the country level (1976-1992, ACT-15 country sample)

Country Sample correlation

Rank correlation Observations

United States 0.95 ** 0.90 ** 17 Canada 0.87 ** 0.73 ** 17 Spain 0.85 ** 0.86 ** 17 Germany 0.69 * 0.52 * 17 Poland 0.58 0.57 * 15 New Zealand 0.44 0.50 * 16 Finland 0.43 0.32 14 Portugal 0.36 0.54 13 Ireland 0.26 0.27 13 Venezuela -0.11 0.00 16 Iran -0.12 -0.05 17 Thailand -0.13 -0.05 15 Hong Kong -0.16 -0.02 13 Japan -0.54 ** -0.62 ** 17 Brazil -0.66 ** -0.72 ** 17

Notes: Concentration data are from Antweiler, Copeland and Taylor (2003), emissions from Stern (2006) / **(*) significant at the 99(95)% level

8

This heterogeneity in results when comparing emissions and concentrations reflects

the large variety of factors that affect concentrations levels and illustrates the

cautiousness one would want to display when interpreting results comparing cross-

country differences in concentration levels.3

Because it may be argued that the number of observations in table 1 is too small, table

2 reports correlations over the maximum number of common observations across the

three available emission data sources and the ACT database.4 Except for the rank

correlation when using STERN data, the correlations are insignificant probably

reflecting the difficulty to control for country-specific effects. Thus a second set of

correlates is estimated considering only deviations from the country mean. This

hardly affects the picture regarding the EDGAR and Cole and Elliott databases, while

both sample and rank correlations are now significantly positive between STERN

emissions and concentration. However, note that the magnitude of the correlation

coefficient (less than 0.20) is considerably lower than at the world-wide level. This is

so because even when there is a significant relationship over time, the pattern varies a

lot across countries (i.e. small variations in concentrations are associated with large

changes in emissions in the USA, while it is the reverse for Italy or Great Britain, see

figure A4 in Appendix 1).

Table 2: Correlations between concentrations and emissions at the country level (1971-1996, all ACT countries) Levels Deviations from the mean EDGAR COLE&

ELLIOTT STERN EDGAR COLE&

ELLIOTT STERN

sample correlation -0.05 -0.01 0.00 0.38 0.12 0.15** rank correlation 0.37 0.21 0.36** 0.65 0.30 0.17** observations 17 41 475 6 37 473 **(*) significant at the 99(95)% level Since the remainder of the paper is concerned with emissions, table 3 reports the

correlations between the three different sources of emissions following the same logic

as in table 2 (calculations are based on the maximum number of common

observations). Comfortingly for what follows, whichever the case, the correlations are

significant and quite large in comparison with those in table 2. Yet, two qualifications

are in order. First, most EDGAR-level estimates lie above the corresponding

3 Our attempt to control for temperature and rainfall did not lead to any significant change of the overall pattern. 4 The original ACT database covers 43 countries. Similar calculations limited to the ACT-15 sample did not lead to any significant difference.

9

estimates for STERN (see figure A5 in the Appendix). Second, between these two

sources, the magnitude of the correlation coefficients is substantially lower when

deviations from the mean rather than levels are considered. Both qualifications are

consistent with the trends identified at the global level, namely that EDGAR figures

appear as overestimates with respect to STERN values, and that the temporal

variation across the nineties is opposite between the two sources (increasing for

EDGAR, decreasing for STERN).

Table 3: Correlations between different emission sources Levels Deviations from the mean EDGAR

vs. COLE& ELLIOTT

COLE& ELLIOTT vs. STERN

STERN vs. EDGAR

EDGAR vs. COLE& ELLIOTT

COLE& ELLIOTT vs. STERN

STERN vs. EDGAR

sample correlation 0.99** 0.99** 0.97** n.a. 0.92** 0.32** rank correlation 0.73** 0.75** 0.95** n.a. 0.73** 0.23** observations 26 104 509 n.a. 104 509 **(*) significant at the 99(95)% level n.a.: not applicable

2.3 Manufacturing emissions

EDGAR is the unique source of time and country-varying emissions at a sufficiently

disaggregated level to include several manufacturing sectors, opening the possibility

to relate emissions to trade and economic structure. According to this data source, as

shown in table 4, manufacturing emissions represent about 45% of the world total

during the 90s.

Table 4: Share of manufacturing sectors in world-wide SO2 emissions (%)

EDGAR code Name (ISIC rev. 2 codea) 1990 1995 2000

F30 Refineries, coke, gas (353, 354) 7.74 7.43 7.17 I10 Iron & steel (371) 0.57 0.59 0.66 I20 Non-ferrous metals (372) 12.26 12.92 14.94 I30 Chemicals (351, 352) 2.12 1.78 2.02 I40 Building materials – cement (369) 1.18 1.62 1.86 I50 Pulp & Paper (341) 0.58 0.71 0.69 F10+B10 Fossil fuel and biofuelb 19.97 21.19 17.15

Total manufacturing 44.42 46.25 44.49

Grand Total 100.00 100.00 100.00 Source: Authors’ calculations based on EDGAR (233) data

Notes: a the correspondence with ISIC codes is an approximation; b To be ventilated across industrial sectors according to the methodology described in section 3

10

Of the “six dirty sectors” listed in table 4, the most polluting sectors are Non Ferrous

Metals and Refineries with close to half of manufacturing emissions is due to "fossil

fuel and biofuel consumption", which is non-attributed across manufacturing sectors

in the original EDGAR database. Section 3 proposes a method to allocate these fuel-

based emissions across manufacturing sectors. Combining the resulting data with the

TPP database leads to emission intensities estimates at a disaggregated level for the

six polluting sectors identified here and a residual "clean" sector (see table A2 in

Appendix 1 for a correspondence with ISIC codes) for a sample of 62 countries.

Unfortunately, there is no alternative data base to which these constructed emissions

may be systematically compared with. The best that can be done is to compare these

estimates with those of the IPPS coefficients for the US in 1987, keeping in mind that

our US estimates are for 1990.5 This is done for emissions per dollar and for

emissions per worker. Results of this comparison are displayed in figure 3.

Figure 3: IPPS vs. EDGAR-derived emission intensities for the US

(1987 for IPPS, 1990 for EDGAR)

Petroleum

Iron&steel

NF metals

ChemicalsNM mineral

Pulp&Paper

Clean sect

510

1520

Gg

per 1

000

empl

oyee

- E

DG

AR

2 4 6 8Gg per 1000 employee - IPPS

original STERN-adjusted

Petroleum

Iron&steel

NF metals

Chemicals

NM mineral

Pulp&Paper

Clean sect

510

15kg

per

100

0 U

SD

- ED

GA

R

5 10 15kg per 1000 USD - IPPS

original STERN-adjusted

5 Although US data are probably more reliable, a similar outcome was obtained with Chinese data.

11

Whichever data source is used, both emission intensities per worker and per dollar

display enormous differences across sectors (note the use of log scales on the axis).

Also the matching between the two sets is far from perfect (the average absolute

percentage difference is 52% for original EDGAR intensities and 46% for STERN-

adjusted ones). This may be due to measurement errors, to differences in the time

period (1990 for EDGAR, 1987 for IPPS) or to differences in sector classification.

Adjusting for the above-mentioned overestimation of emissions by EDGAR data

slightly increases the correspondence since the data are usually above the 450 line.

Though not obvious from the figure because of the scaling, note also that the

matching is far better for emissions per employee than for emissions per USD since

the sample correlation between the two sources of per employee intensities is equal to

0.94 and significant at 99%, while it is non significant between per USD intensities.

Overall, we conclude that EDGAR data should preferably be adjusted downwards and

that per employee emissions tend to be more reliable than per USD emissions, at

least for the US, an observation that is also confirmed across countries and regions in

figure 4 and table 6 below. These observations are taken into account in the

construction of the data base presented below.

3 Constructing manufacturing emission intensities

We now detail the construction of a new and complete database reporting SO2

emissions for 7 manufacturing sectors in 62 developing and developed countries for

1990, 1995 and 2000 (the "base" years we refer to below). This emission data base is

then combined with production data in section 4 to carry out a growth decomposition

exercise into scale, composition and technique effects.

We proceed as follows. In constructing the data base, two data sources were available,

one disaggregated across countries (EDGAR) and one across sectors (IPPS). EDGAR

covers a large set of countries over 3 periods but only for 6 dirty industrial sectors,

while IPPS provides emission intensities for 28 ISIC 3-digit sectors but only for the

US and 1987.6 Using the assumptions described below, data on industrial output and

employment from the Trade Production and Protection (TPP) database (Nicita and

Olarreaga (2007)) were used to construct the complete data bases, imputing the 6 Note that the IPPS data is even available at the 4-digit ISIC level, giving details for more than 80 manufacturing sectors. However, given that our economic activity database is reported at the 3-digit level we had to keep this level of aggregation.

12

missing values on the basis of the procedure described in Appendix 3, and

considering 3-year moving average to control for cyclical fluctuations. These data

bases were then scaled to match the total estimates of Stern (2006). We describe first

the construction of the EDGAR-related data bases, then the IPPS-related data bases

which serve as a check.

3.1 EDGAR-related data bases

EDGAR data has two shortcomings. First, they do not take into account the fact that

"clean" sectors also generate emissions. Second, they report an awkward non-

imputed category, "fossil fuel and biofuel consumption" (F10 and B10), which

represents about 45 percent of total manufacturing emissions. Regarding emissions

by the “clean” sectors, we rely on the IPPS data base (see below). For the non-

imputed category, we carry out the following 3-step procedure:

(i) estimate the share of clean sectors in overall manufacturing emissions on the

basis of IPPS coefficients applied to TPP employment data

(ii) apply this share to the total of EDGAR-based manufacturing emissions

(imputed plus non-imputed categories), obtaining an estimate of "virtual" clean

sectors emissions

(iii) substract the virtual amount from the non-imputed amount and spread the

residual across dirty sectors according to the IPPS-derived share of each sector

in dirty emissions. If the residual is negative, all unaffected emissions are

allocated to the clean sectors.7

Steps (i) and (iii) imply that IPPS per employee intensities are assumed to be valid for

every country and year, which is inaccurate but probably closer to reality than

assuming that IPPS per dollar intensities are constant. Finally, all emissions are

scaled so that total computed manufacturing emissions match the corresponding

figure derived from Stern (2006)8 The last step to obtain intensities is to divide

7 Alternative procedures were also tested, either by using labor rather than emission shares in step (iii), or by skipping step (i) and directly splitting the non-imputed emissions among sectors. The selected procedure is the one that maximizes the sample correlation coefficient (at 0.94) between the 1987 US-IPPS intensities (our unique reference case) and the corresponding EDGAR-based intensities. 8 EDGAR-based manufacturing emissions shares are applied to Stern's 3-year moving average total estimates) For the majority of countries, the scaling factor is close to 0.9 (the interquartile range is between 0.8 and 1.0, with only a few outliers, see table A6 in Appendix 1).

13

sectoral emissions by the corresponding employment or output figures provided by

the TPP database.9

This is the first complete data base entered in the bottom left of table 5. It is labeled

EDGARda (where subscript a(d) corresponds to aggregated (disaggregated), the first

index referring to entities (countries or regions), the second to sectors) and covers 62

countries and 7 industrial sectors. For comparison purposes (see below), these

emission intensities have also been either aggregated into 6 regions (EDGARaa data

base) or disaggregated further into 28 sectors (EDGARdd data base). The aggregation

into regions is based on a definition of geo-economic regions that reflect both

geographic proximity and similarity in income per capita (see table A2 in Appendix

1). The disaggregation procedure posits that the dispersion of intensities within each

EDGAR category is identical to the one observed in the IPPS base.10 These data

sources provide alternative benchmarks to discuss measurement errors (see below).

They are also useful for applied policy analysis, which often requires working with

different classifications.

Table 5: Alternative databases on emission intensities

Sectors(a,d) Entities(a,d)

7-EDGAR sectors (a)

28 ISIC sectors (d)

6 regions (a)

IPPSaa EDGARaa

IPPSad

62 countries (d)

EDGARda

IPPSdd EDGARdd

Notes: All data bases are available for the three years (1990, 1995, 2000), and either in original levels (i.e. scaled to EDGAR totals for IPPS-based data—see text) or scaled so that total manufacturing emissions match those reported by Stern(2006). Subscript a(d) corresponds to aggregated (disaggregated), the first index referring to entities (regions or countries) the second to manufacturing sectors.

9 For a limited number of cases where countries report positive emissions but zero employment for certain sectors, aggregation over sectors was applied (see notes for table A3 in Appendix 1). 10 More precisely, it is assumed that the ratio between the emission intensity of each sector and its employment-weighted mean at the EDGAR category level is identical to the one obtained when applying US IPPS coefficients to the country’s specific employment data.

14

3.2 IPPS-related data bases

Insofar as the US IPPS emission coefficients for 1987 were carefully constructed, and

in view of the evidence reported by Hettige et al (2000) suggesting a relative

constancy of emissions per unit labor across countries and over time for one pollutant

(biological oxygen demand), it is worth checking to what extent this conjecture holds

for SO2 in the EDGAR data and, if so, construct an alternative data base using IPPS

coefficients. The boxplots in figure 4 provide a first check on the relative constancy of

emissions per unit of labor. It represents the distribution of intensities across

countries and regions for each one of the 7 EDGAR sectors, for either per unit labor

or per USD SO2 emissions (to facilitate comparisons, each intensity series has been

scaled down by the median, so that when taking logs the interquartile box is

approximately centered at zero).

Figure 4: Boxplots of EDGAR emission intensities (1990-1995-2000)

NPLMWIMWIMWIMUSMUSNPLIRLNORMUS

NOR

IDNHNDIDNIDN

NPLVENNORSWENORSWENPLCRI

NLDNLD

NORVENSWENORSWE

NPL

CHLCHLCHL

NPL

NPLKEN

SGPSGPTUNTUN

TUNTUNTUNBOL

NPL

NORNPL

NOR

HUN

NPL

HUNBLXNLDBLXBLXNPLCRIPHLPHL

HUN

BLXBLXBLXNLD

SGPSGPNPL

MACMACNPL

SGPSGP

-10

-50

5

F30 I10 I20 I30 I40 I50 OTH

(a) across countries

ln of median-scaled per employee intensityln of median-scaled per usd intensity

SAMSAMSAM

-10

-50

5

F30 I10 I20 I30 I40 I50 OTH

(b) across regions

ln of median-scaled per employee intensityln of median-scaled per usd intensity

Note: see table 4 for a description of EDGAR categories; OTH refers to all other (clean) sectors

Two stylized patterns emerge. First, emission intensities by sector are more similar

across regions than across countries. This is for both outliers, which disappear when

considering regions (apart from the South America case for non-ferrous metals,

15

which reflects the strong influence of Chile), and for the interquartile range, which is

lower across regions, in particular for emissions per unit labor. Second, based on the

interquartile range and whatever the sector, emissions per USD exhibit a larger

dispersion than emissions per unit of labor.

Because the differences between the two types of intensities is less evident when

outlier values are taken into account, table 6 uses the coefficient of variation --which

factors in outliers--as a measure of dispersion. Table 6 estimates confirm that per

unit labor intensities exhibit a smaller dispersion than intensities per USD, event

though dispersion appears to be increasing over time.

Table 6: Coefficient of Variation

(EDGAR pollution intensities)

Across countries Across regions

Gg per 1000 employee

Kg per 1000 USD

Gg per 1000 employee

Kg per 1000 USD

1990 1.28 1.54 0.53 0.99 1995 1.49 1.69 0.66 0.96 2000 1.56 1.86 0.79 0.98 All 3 years 1.47 1.70 0.64 0.97 Note: figures refer to the unweighted mean over the seven EDGAR sectors

In view of the above and of the lesser dispersion across regions than across countries,

as a check, it is arguably justifiable to apply the original IPPS emissions per employee

to each one of the 6 regions of the sample in the hope of minimizing errors.11 Then,

as before, in a second step, these original intensities are either scaled to insure that

the resulting total of manufacturing emissions per region exactly matches the one

obtained from the corresponding EDGAR-derived database (EDGARaa, see table 5)

which amounts to the totals in EDGAR or to those in Stern (2006). As a result, total

manufacturing emissions of the two versions of the IPPSad database are by

construction strictly identical to those obtained for the corresponding versions of the

EDGARaa database. The same logic is applied to construct the remaining two

databases of table 5 (IPPSaa at a more aggregated and IPPSdd at a more disaggregated

level). Even though we believe that IPPSad is the most reliable database amongst the

IPPS-related ones (and EDGARda for the EDGAR-derived ones), all databases are

used for the sensitivity analysis in section 4.

11 We also experimented with an alternative based on fitting a regression of emission intensities on a time trend and per capita and per capita GDP squared, systematically obtaining a poor fit leading us to abandon this alternative route.

16

Table 5 summarizes the characteristics of each data base. Data bases below the

diagonal reflect the between-country and between-sector variation in EDGAR data,

while those above reflect the between-sector variation in the IPPS coefficients, and

either the EDGAR or STERN-based between-country variation. Finally for both sets

of databases, the temporal variation is either based on EDGAR or STERN, depending

on which database has been used as the scaling benchmark.

4. Decomposing the Change in SO2 emissions

We now decompose the evolution of “worldwide” SO2 emissions into the scale,

composition and technique effects (see Grether, Mathys and de Melo (2007) for

further discussion). 12 To obviate the need to bring in cross-country comparisons in a

common currency, we use intensities per unit of labor rather than per unit of output.

Let then kitL represent employment in activity k in country i in year t , kitγ emission

intensity per unit of labor and E aggregate emissions at the sample-wide level. Then,

emissions at the sector, country and global levels are given by:

∑ ∑∑ === k i kitkittk kitkitktkitkitkit LE;LE;LE γγγ [1]

Let ( )titLit L/Lt ≡θ denote the share of country i in the world labor force,

( )itkitLkit L/Lit ≡θ the share of sector k in country i's labor force and ( )tit

Eit E/Et ≡θ the

corresponding emission share for country i . Then, as shown in Grether et al. (2007),

by differentiation of identity [1], SO2 growth in emissions can be decomposed into the

following expression:

( )∑ ∑∑ ∑∑ +⎟⎠⎞⎜

⎝⎛+⎟

⎠⎞⎜

⎝⎛+= i k kit

Ekiti k

Lkit

Ekitk

Lit

Eittt ˆˆˆLE titttt γθθθθθ [2]

In [2], the first term on the RHS is the scale effect. The second term measures the

between-country effect and will be positive if emissions of countries with higher per-

unit labor emissions grow faster since the sum of shares is equal to unity. Likewise,

the third term measures the between-sector effect within a country and will be 12 Our companion paper uses the preferred set of estimates from this paper and explores several counterfactuals on the role of international trade (including transport costs) and differences in techniques in the observed changes in SO2 emissions.

17

positive if dirty industries have a higher growth in employment. Finally, the fourth

term measures the technique effect. In implementing [2], interaction terms (not

shown here since they are assumed to be negligible for small changes) are attributed

proportionately across each one of the four terms. Note also the following: (i) up to

small differences reflecting the attribution of the interaction terms, the scale effect

will be the same across the data bases since the same employment data is used

throughout; (ii) likewise, by construction, total emissions are the same across sectoral

and country aggregations and for the ten year period covered here equal either to -

9.9% (Stern estimates in table 7) or to +9.2% (EDGAR estimates in table A4). Hence,

differences in estimates across data sets are attributable to differences in the relative

importance given to the two composition effects and to the technique effect, and

since, as the scale effect, the two composition effects are drawn from the same TPP

data base, the differences in total estimates (19.1%) are projected into differences in

technique effects (about 19% difference in technique effect between the two base

estimates).

4.1 Scale, Composition and Technique Effects

Table 7 concentrates on the preferred estimates coming from the EDGAR data set,

with 62 countries and 7 sectors, leaving to section 4.2 the discussion over alternatives

(IPPS-derived data sets and/or other aggregation levels). Over the entire 1990-2000

period, results are quite similar for the scale (around 10%) and the two composition

effects (above 5%), whether intensities are scaled down to STERN levels or not. The

major difference comes from the technique effect, which is negative (close to -15%)

for STERN-adjusted emissions but positive (around 5%) for non-adjusted ones. This

leads to a sharp contrast in terms of total emissions, which are decreasing in the

STERN-adjusted case and increasing in the other case.

Table 7: Scale, composition and technique effects (%) (preferred database, STERN-adjusted or Non-adjusted intensities) STERN-adjusted intensities Non-adjusted intensities 1990-1995 1995-2000 1990-2000 1990-1995 1995-2000 1990-2000

Scale effect 5.82 3.93 9.55 6.26 4.02 10.5

Betw. country -1.48 -1.89 -2.44 -1.02 -1.52 -3.06

Betw. sector -2.65 0.34 -3.03 -3.22 -1.01 -4.82

Technique -4.1 -10.04 -13.94 11.24 -5.11 6.55

Total effect -2.4 -7.64 -9.85 13.27 -3.61 9.17

18

However, differences may be smaller than they appear. When splitting the nineties

into two sub-periods, it turns out that the decomposition pattern is quite similar

between the second sub-period (1995-2000) for non-adjusted intensities and the first

sub-period (1990-1995) for STERN-adjusted ones, as if it just took a five-year lag for

the same trend to be reflected in the non-adjusted data set. This lagged response is

further confirmed by figure 5, which reports the evolution of world manufacturing

SO2 emissions when introducing one by one the four effects that compose the overall

picture.13

Figure 5: Global manufacturing SO2 emissions (1990-2000)

3540

4550

5560

Wor

ld M

anuf

actu

ring

SO2

emis

sion

s, T

g

1990 1992 1994 1996 1998 2000year

Scale Scale+techn.

Scale+techn.+b.sect. Scale+techn.+b.sect.+b.coun.

STERN-adjusted

3540

4550

5560

Wor

ld M

anuf

actu

ring

SO2

emis

sion

s, T

g

1990 1992 1994 1996 1998 2000year

Scale Scale+techn.

Scale+techn.+b.sect. Scale+techn.+b.sect.+b.coun.

EDGAR-adjusted

If only scale effects had been at work, emissions would have increased whatever the

data set. Adding the technique effect leads to lower (larger) emissions in the case of

STERN (non)-adjusted intensities, while composition effects unambiguously

decrease total emissions. Eventually, global emissions tend to decrease, whether over

the whole decade or during the second sub-period only. This trend reflects the 13 Figures between base years are obtained by linear interpolation. Total emissions can be written as

∑ ∑= i k kitLkit

Littt

ittLE γθθ . The curve representing the scale effect takes only world manufacturing

employment (Lt) into account, freezing all the other terms of the double sum at their 1990 levels. This assumption is further relaxed term by term.

19

adoption of greener production techniques in the majority of large polluting

countries. The 5-year lag between the two data sets is due to the methodology used by

Stern (2006), which allows for a better control of abatement activities in developing

countries (see discussion in Appendix 2).

In sum, the above results suggest that the nineties witnessed a structural shift

towards cleaner activities and a shift of activities towards cleaner countries (see

Grether et al (2007) for an analysis at the sector and the country level). This has been

accompanied by more abatement activities, which are better captured by the Stern-

adjusted intensities (probably the best estimates in hand), and have contributed to a

substantial decline in emissions in spite of the scale effect.

4.2 Robustness

How robust are these results when weights are altered by different choices for

aggregation across sectors and/or regions? We discuss this issue on the basis of our

preferred set of estimates, i.e. STERN-adjusted emissions, and report the results in

table 8 (similar results are obtained for EDGAR-adjusted emissions, see table A4 in

Appendix 1) Aggregation across sectors (comparing lines 1 and 2 or 5 and 6) has very

little impact suggesting that emission intensities are quite homogenous inside the

seven EDGAR sectors and disaggregating further to the 28 ISIC sectors does not

change the picture. This is not surprising given that in the EDGAR classification dirty

sectors are considered explicitly, while all clean sectors, which have smaller

differences in intensities, are lumped together into one sector.

Table 8: Scale, composition and technique effects across databases (1990-2000, STERN-adjusted total emissions)

Database 1990-2000 Growth Decomposition (%)

Entities Sectors Scale effect

Between country

Between sector Technique Total

1 EDGARdd 62 28 9.57 -2.46 -2.73 -14.25 -9.85

2 EDGARda 62 7 9.55 -2.44 -3.03 -13.94 -9.85

3 EDGARaa 6 7 9.53 -4.36 -2.39 -12.64 -9.85

4 IPPSdd 62 28 9.51 -2.43 -3.65 -13.29 -9.85

5 IPPSad 6 28 9.52 -4.41 -4.43 -10.53 -9.85

6 IPPSaa 6 7 9.53 -4.43 -4.16 -10.80 -9.85

Notes: the line in bold characters indicates the data base used in Grether, Mathys and de Melo (2007)

20

However aggregating across countries (comparing lines 2 and 3 or 4 and 5) does

increase the magnitude of the negative between-country composition effect. This

suggests that the within-region structural shift has been in the opposite direction to

the global one, i.e. intra-regional production has shifted on average towards the

dirtiest countries within each region. Thus, working at the aggregated level in terms

of geographic entities may lead to an overestimation of the composition effects.

Comparing EDGAR and IPPS data sets at the same level of aggregation (i.e. lines 1

and 4 or 3 and 6) does not alter the between-country effect but leads to a stronger

technique effect in the case of EDGAR (or equivalently a stronger between-sector

effect in the case of IPPS) which is to be expected since EDGAR is the data set that

includes temporal variation in emission-intensity data.

In sum, while there are some differences when weights are altered, these are small.

Overall, as cross-country and time variation of emission intensities is basically

derived from the EDGAR database, our preferred set of intensities is EDGARda (line

2), which matches most closely the original data source.

4.3 In search of Kuznets effects14

As a final illustrative application of the above data sources, we examine the

relationship between manufacturing emission intensities and GDP per capita15. At the

global level, considering the average over the three base years, no significant

relationship seems to emerge. This is illustrated by figure 6(a) below, where points

are far apart from the fitted curve (log scale), and basically confirmed by the

regressions reported in table A5 in the Appendix, where alternative specifications are

also tested without altering the non-significance result.

One explanation for this absence of an EKC pattern is that we are limited to

manufacturing industries, which means that we cannot capture the structural

changes that affect the whole production structure of the different countries involved.

Indeed, the only case where a significant EKC pattern is identified by the regressions,

with a turning point around 12'500 USD, is when the three base years are pooled

together. This limited inverted-u effect has probably more to do with the reversal of 14 We will only give a first insight into the relation between emission intensities and GDP per capita. A more detailed analysis is outside of the scope of this paper. For a recent EKC estimation for sulfur emissions controlling for stationarity and spatial correlation see Maddison (2007). 15 GDP PPP per capita data has been obtained from the World development indicators CD-ROM.

21

aggregate SO2 emissions over time (identified by Stern (2006) and reflected in our

intensities given the adjustment procedure) than with structural differences across

countries.

Figure 6: SO2 intensities and GDP per capita

MWI

KEN

NPLBGD

SEN

PAK

INDBOL

IDN

CHN

HND

EGYECU

MARPHL

JOR

PER

PANTUNVEN

COL

TUR

BRA

CRI

MYS

CHL

URY

MUS

MEX

POLZAF

ARG

HUN

KOR

GRC

PRTCYPNZL

MAC

KWT

SGP

ESP

ISR

IRL

AUS

FIN

DEU

HKG

ITAFRAGBRSWE

CAN

JPN

BLX

NLDISL

AUTDNK

NOR

USA

25

1020

Gg

of S

O2

per 1

000

empl

oyee

0 10 20 30GDP PPP per capita, thousand constant 2000 USD

(a) all manufacturing sectors

MWI

KEN

NPLBGD

SEN

PAK

IND

BOLIDNCHN

HNDEGY

ECU

MAR

PHL

JOR

PER

PAN

TUN

VEN

COL

TUR

BRA

CRIMYS

CHL

URY

MUS

MEX

POLZAF

ARG

HUN

KOR

GRC

PRT

CYP

NZL

MAC

KWT

SGP

ESP

ISR

IRL

AUS

FINDEUHKGITAFRAGBRSWE

CAN

JPN

BLX

NLD

ISL

AUTDNK

NOR

USA

0.0

5.1

.15

.2.2

5

0 10 20 30GDP PPP per capita, thousand constant 2000 USD

(b) clean sector

Additional evidence is provided by analyzing the relationship at the sector level. As

emission intensities and labor shares jointly determine the aggregate intensity at the

country level, both were regressed on the GDP per capita terms. It turns out that the

only significant term in these regressions is the constant (see table A6 in the

Appendix), confirming the absence of an EKC pattern detected at the country level.

The only exception is the "clean" sector, with a turning point around 12'000 USD, as

represented in figure 6(b) (arithmetic scale). In other words, for all dirty sectors, per

unit labor emission intensities appear to be roughly constant across countries, while

the inverted-u relationship identified for the "clean" category probably reflects

composition effects as this category lumps together a large number of non-dirty

sectors.

22

5. Conclusions

Taking the best out of the available information, we constructed a set of original

databases on SO2 intensities over the nineties, which is consistent with the most

recent estimates of national emissions reported by Stern(2006), and allows for the

first time to relate manufacturing activity with SO2 emissions taking composition

effects into account.

The growth decomposition exercises establish a robust pattern across all databases,

namely that polluting production at the world-wide level has shifted towards cleaner

countries and cleaner sectors during the 1990-2000 period. This has contributed to

reduce overall manufacturing emissions by more than 5%, a result that is common

across databases. The magnitude of the technique effect is more controversial, but the

two main data sources coincide in identifying a reduction in polluting intensity (since

1990 according to STERN, or 1995 according to EDGAR). Overall, these forces have

helped to contain and even reverse the scale effect at the world-wide level.

We believe these results illustrate the usefulness of the new databases developed here

which combine both trade and industrial data (including input-output tables) and

should open the field for more applied work on the growth-trade-environment nexus.

23

References Antweiler, W., B. Copeland, M.S. Taylor (2001) “Is Free Trade Good for the Environment?”, American Economic Review 91(4), 877-908. Brock, W. and M.S. Taylor (2005), “Economic Growth and the Environment: A Review of Theory and Empirics” in P. Aghion and S. Durlauf eds. Handbook of Economic Growth, vol. I(28), 1749-1821 Cole, M. and R.J.R. Elliott (2003) “Determining the Trade-environment Composition Effect: the role of Capital, labor and the Environment”, Journal of Environmental Economics and Management, 46, 363-83. Grether, J.M., N. Mathys, and J. de Melo (2007) “Global manufacturing SO2 emissions : does trade matter?”, CEPR WP 6522. Grossman, G.M. and A.B. Krueger, (1991), "Environmental Impacts of a North American Free Trade Agreement", NBER Working paper No. 3914. Hettige, M. Muthukurama, D. Wheeler (1995) “IPPS: The Industrial Pollution Project System”, World Bank PRWP 1431. Hettige, M. Muthukurama, D. Wheeler (2001) “Industrial Pollution in Economic Development: Kuznets Revisited, Journal of Development Economics, 62 445-60 Maddison D. (2007), "Modelling sulphur emissions in Europe: a spatial econometric approach", Oxford Economic Papers, Vol. 59, pp. 726-743 Nicita, A. and M. Olarreaga (2007) “Trade, Production and Protection, 1976-2004” World Bank Economic Review, Vol. 21(1) , 165-71. Olivier, J.G.J. and J.J.M. Berdowski (2001), “Global Emission Sources and Sinks”, in Berdowski, J., R. Guicherit and B. Heij eds. The Climate System, A.A. Balkema Publishers, Swets & Zeitlinger Publishers, Lisse, The Netherlands, pp. 33-78. Stern, D. I. (2006), "Reversal of the Trend in global anthropogenic sulfur emissions", Global Environmental Change, Vol. 16, pp. 207-220. World Bank (2006), World Development Indicators, CD-ROM.