WORLD CARBON DIOXIDE EMISSIONS: 1950–2050

Richard Schmalensee, Thomas M. Stoker, and Ruth A. Judson*

Abstract —Emissions of carbon dioxide from the combustion of fossil future warming.3 This paper focuses on the roughly 80% of
fuels, which may contribute to long-term climate change, are projected
through 2050 using reduced-form models estimated with national-level
anthropogenic CO2 emissions currently produced by the
panel data for the period of 1950–1990. Using the same set of income and combustion of fossil fuels.4 The literature contains many
population growth assumptions as the Intergovernmental Panel on Climate long-run forecasts of these emissions; see Alcamo et al.
Change (IPCC), we find that the IPCC’s widely used emissions growth
projections exhibit significant and substantial departures from the implica-
(1995) for a recent survey produced as part of the IPCC
tions of historical experience. Our model employs a flexible form for process. Almost all of these have been produced using
income effects, along with fixed time and country effects, and we handle structural models in which parameter values have been fixed
forecast uncertainty explicitly. We find clear evidence of an ‘‘inverse U’’
relation with a within-sample peak between carbon dioxide emissions (and
by a mix of judgment and calibration. Fewer than a handful
energy use) per capita and per-capita income. of these studies consider the implications of the (subjective)
uncertainty attaching to key parameter values.5
I. Introduction In this paper we use an econometric model to construct
projections of CO2 emissions from fossil fuel combustion
M OST scientists consider it likely that if the atmospheric
concentrations of carbon dioxide (CO2 ) and other
so-called greenhouse gases continue to rise, the earth’s
through 2050. We compare our projections to those of the
IPCC in order to assess whether the IPCC projections are
consistent with historical experience. We also analyze the
climate will become warmer.1 While relatively little is precision of our projections, to judge whether differences we
known about the likely costs and benefits of such warming, note are strongly supported by the data evidence. We utilize
it seems clear that both depend critically on the rate at which the same scenarios for population and economic growth as
warming occurs. The rate of future warming depends, in those used by the IPCC, but find that our projections are
turn, on a number of poorly understood natural processes appreciably higher in all but one scenario. Moreover, the
and on future emissions of greenhouse gases. Key climate confidence intervals around our emissions projections are
processes (in particular, warming the deep ocean) involve generally above the IPCC projections in these cases.
long lags, and important greenhouse gases (in particular, Our projections are derived from a reduced-form econo-
CO2 ) remain in the atmosphere for many years after they are metric model estimated with a large national-level panel
emitted. Accordingly, climate change analyses necessarily data set covering the years from 1950 to 1990. Our approach
involve emissions forecasts spanning several decades and is to model similarities across countries in the growth of CO2
often a century or more. emissions with economic development, together with fixed
The Intergovernmental Panel on Climate Change (IPCC) (scale) effects for the overall levels of emissions across
was established in 1988 to inform international negotiations countries and time periods.
on climate change. Among the most visible of the IPCC’s Figure 1 illustrates two of the typical patterns that our
activities has been the generation of scenarios of future model captures. First, for developing countries (here Korea
greenhouse gas emissions extending to the year 2100.2 A and India) there is continuous or even accelerating growth of
Framework Convention on Climate Change was signed by per-capita CO2 emissions with per-capita gross domestic
the United States and other nations at Rio de Janeiro in product (GDP). Second, for highly developed countries
August 1992; it entered into force in March 1994. The (here the United States and Japan) the growth of per-capita
convention’s stated long-run objective is mitigating emis- CO2 emissions with per-capita GDP flattens and may even
sions of greenhouse gases to permit ultimate stabilization of reverse at higher levels of economic development. This
their atmospheric concentrations. latter pattern, referred to as an ‘‘inverted U’’ relation, has
Emissions of CO2 caused by human activity are generally been noted by other researchers for various air pollutants.6
considered the most important single source of potential However, CO2 was not regarded as a pollutant until the late
3 Because greenhouse gases’ atmospheric lifetimes differ substantially

Received for publication February 14, 1996. Revision accepted for pub- and the relevant chemical processes are complex and nonlinear, assessing
lication February 10, 1997. the relative importance of greenhouse gases for policy purposes is not
* Massachusetts Institute of Technology, Massachusetts Institute of trivial; see Schmalensee (1993). A few years ago the IPCC (1990)
Technology, and U.S. Federal Reserve Board, respectively. estimated that CO2 alone accounted for about 55% of the increase in
The authors are indebted to the U.S. Department of Energy, the MIT radiative forcing (net solar radiation retention by the earth) during the
Center for Energy and Environmental Policy Research, the MIT Joint 1980s. No other single gas was estimated to account for more than 15%.
Program on the Science and Policy of Global Change, and the National Chlorofluorocarbons (CFCs) were estimated to account in aggregate for
Science Foundation for financial support. They are grateful to Taejong Kim about 24%. Recent research (see IPCC (1992, p. 14)) has shown that this
for diligent research assistance and to Tom Boden, Richard Eckaus, earlier work overstated the effect of the CFCs, so that CO2 likely accounted
Howard Gruenspecht, Eric Haites, Alan Manne, William Nordhaus, Joel for well over 55% of the increase in radiative forcing during the 1980s.
Schiraga, and participants in the MIT Global Change Forum VII for help 4 Pepper et al. (1992, p. 101) provide the following breakdown of 1990

and advice. Errors made and opinions expressed in this paper are solely the anthropogenic CO2 emissions: fossil fuel combustion 80%, deforestation
authors’ responsibility. 17%, and cement production 3%.
1 For general discussions of climate change, see ‘‘Symposium on Global 5 See Alcamo et al. (1995). Manne and Richels (1992, 1994) and

Climate Change’’ (1993), Cline (1992), and Nordhaus (1994). Nordhaus (1994) are notable examples.
2 See IPCC (1990, 1992, 1996) and Alcamo et al. (1995). 6 See, for instance, Selden and Song (1994).

r 1998 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology [ 15 ]
16 THE REVIEW OF ECONOMICS AND STATISTICS

FIGURE 1.—HISTORICAL PATTERNS OF CO2 EMISSIONS AND GDP GROWTH

1980s, and there were no significant policies aimed at forecasting. Not only is it extremely difficult to forecast
restraining CO2 emission in effect within our sample period. environmental policies and other exogenous variables de-
We use a flexible representation of the per-capita cades into the future, it is even more difficult to quantify the
CO2–GDP relationship, along with time and country fixed uncertainty attached to such forecasts.8
effects, and we handle forecast uncertainty explicitly in our Nevertheless, our estimates provide a (historical) bench-
projections. This reduced-form approach of estimation and mark for the construction of simulation models, and our
projection of historical trends amounts to forecasting by projections provide a check of simulation-based forecasts.
‘‘sighting along the data.’’ Our estimates thus reflect any The simplicity of our model contrasts with the ‘‘black box’’
relevant historical tightening of environmental standards, for character of many structural simulation models. For in-
example, and our projections reflect the (implicit) assump- stance, the distributional underpinnings of our projections of
tion that such standards would continue to be tightened at CO2 emissions are easy to trace. For evaluating the IPCC
roughly the historical pace.7 forecasts, we use the same basic data as used by the IPCC,
Our approach represents more of a ‘‘change as usual’’ and the substantial differences noted between the projections
than a ‘‘business as usual’’ approach. If one actually knew summarize how the IPCC’s forecasts depart from past
how environmental standards would change over time trends. Moreover, since our approach permits an explicit
around the world, one could enhance forecast accuracy by analysis of the forecast uncertainty implied by the historical
exploiting that information in a structural model. Unfortu- record, at the very least our approach should serve to inform
nately, available data do not permit econometric estimation
of a global structural model suitable for long-term emissions 8 Prior analyses of forecast uncertainty in this context have apparently

relied exclusively on subjective estimates; see Alcamo et al. (1995).

7 It is useful to note that Holtz-Eakin and Selden (1995) have previously Unfortunately, numerous experiments have established that experts tend to
estimated similar models. They use a significantly less flexible income underestimate uncertainty substantially in their domain of expertise.
specification, however, and they do not compare their projections with Lichtenstein et al. (1982) survey the voluminous literature on this
those of the IPCC. overconfidence bias.
 WORLD CARBON DIOXIDE EMISSIONS: 1950–2050 17

judgments regarding the uncertainty of the structure of Flaring is more closely related to energy production than to
simulation models. energy consumption, and variations in flaring over time
The paper is organized as follows. Section II describes our seem unlikely to reflect the same forces that drive energy
data and model specifications, and section III presents our consumption and carbon emissions decisions.
estimation results. Section IV outlines the methods used to In part as a consequence of these exclusions, even though
project CO2 emissions through 2050, and section V de- our data omit countries with only about 1.4% of the world’s
scribes the resulting projections. Methodological and substan- population in 1990,11 our total CO2 emissions are 7.1%
tive conclusions are outlined in section VI. below the 1990 total used by the IPCC (Pepper et al. (1992)).
Our 1990 total energy consumption is 6.1% below the
II. Empirical Specifications corresponding IPCC total.12 Because of these differences in
base year totals, we focus on comparing our projections of
A. Data post-1990 growth with those of the IPCC, not on comparing
This study is based on national-level panel data on the projections of absolute levels.
following variables for the period 1950–1990: Data on Y and N were primarily taken from the Penn
C 5 CO2 emissions from energy consumption, millions World Table, Mark 5.5 (see Summers and Heston (1991)).
of metric tons (tonnes) of carbon We employed the RGDPCH series for Y, which is based on a
Y 5 GDP, millions of 1985 U.S. dollars chain index of prices in each country and estimates of
N 5 population, millions of persons purchasing-power-parity exchange rates in 1985. Because
E 5 energy consumption, millions of Btu. our sample coverage was constrained by the coverage of the
Penn World Table, and because it seemed important to have
Our data set contains 4018 observations on these vari- comprehensive geographic coverage in 1990, the base year
ables. In 1990 it covers 141 countries, which account for for our projections, we employed other standard sources of
98.6% of the world’s population. The geographic coverage income and population data to add 92 post-1984 observa-
of these data increases sharply in 1970, and 2620 observa- tions on 48 countries to our samples.13
tions (65.2%) are from the 1970–1990 period. We have data
on 47 nations for the entire 1950–1990 period. (Holtz-Eakin 11 This is based on the figure for world population in 1990, given on p.
and Selden (1995), HES hereafter, use earlier versions of our 219 of the World Bank’s World Development Report, 1992. The following
primary data sources and do not employ supplemental countries are excluded entirely from our data set: Afghanistan, Albania,
sources of information on Y and N. Their data set has 3754 Bermuda, Burkina Faso (Upper Volta), Khmer (Cambodia), Dominica,
French Guyana, Lebanon, Liberia, Libya, Macau, Solomon Islands, Tonga,
observations over the period of 1951–1986.) and North and South Yemen.
Data on C, which will simply be referred to as CO2 or 12 We excluded consumption of ‘‘traditional fuels,’’ which include wood,

carbon emissions in what follows, and E were provided by charcoal, and peat, from our measure of E. Because these fuels are treated
as renewable, their consumption is treated in the Oak Ridge data and by the
the Carbon Dioxide Information Analysis Center of the Oak IPCC as not increasing C. In addition, national-level data on the
Ridge National Laboratory.9 These data are based on United consumption of traditional fuels are both incomplete and unreliable. The
Nations estimates of national energy consumption; see IPCC includes traditional fuels in their energy sector analysis (as
noncommercial biomass). Our 1990 total energy consumption is 13.1%
Marland et al. (1989).10 The United Nations data exclude below theirs, but excluding traditional fuels from their total reduces the
bunker fuel, which cannot be allocated unambiguously to gap to 6.1%.
13 We employed various editions of World Factbook (CIA), World
particular nations, and the associated carbon emissions. In
Development Report (World Bank), and International Financial Statistics
addition, following HES, we have excluded gas flaring and (IMF), along with Trends in Developing Economies 1992 (World Bank).
the associated CO2 emissions (which amounted to about For almost all added observations, growth rates in population and/or real
0.9% of total energy-related emissions in the mid-1980s). GDP from these sources were used to extend the coverage of the Penn
World Table forward in time. A single observation for 1990 was added in
this fashion for the following 29 countries: Angola, Barbados, Botswana,
9 These data generally reflect national boundaries in each year for which Burma, Cape Verde, Sri Lanka, Zaire, Benin, Ghana, Guinea, Haiti, Iran,
data are presented. Thus the USSR is a single nation in all years, for Jamaica, South Korea, Kuwait, Malta, Oman, Niger, Puerto Rico, Qatar,
instance, while Germany is two nations. The following adjustments were Saudi Arabia, Seychelles, Somalia, Suriname, Swaziland, United Arab
made for border changes during the sample period. For 1957–1969, Sabah Emirates, Uganda, USSR, and Vanuatu. For the following 15 countries,
and Sarawak were added to Malaysia. For 1950–1979, the Panama Canal Penn World Table coverage ended before 1989, and multiple observations
Zone was added to Panama. For 1972–1990, Bangladesh and Pakistan (54 in total) were added to extend coverage to 1990: Bahamas (1988–
were combined. For 1950–1972, the Ryukyu Islands (Okinawa) were 1990), Bahrain (1989–1990), Bhutan (1986–1990), Belize (1986–1990),
added to Japan. For 1964–1990, the period for which data are available, Comoros (1988–1990), Ethiopia (1987–1990), Djibouti (1988–1990), Iraq
Malawi, Zambia, and Zimbabwe are combined. For 1962–1990, the period (1988–1990), Nepal (1987–1990), Nicaragua (1988–1990), Reunion (1989–
for which data are available, Rwanda and Burundi were combined. For 1990), Romania (1986–1990), Saint Lucia (1986–1990), Saint Vincent and
1950–1969, Tanganyika and Zanzibar were combined. For 1950–1969, the Grenadines (1986–1990), and Tanzania (1989–1990). Finally, popula-
North and South Vietnam were combined. tion and income data from the World Factbook were added for four
10 Energy consumption estimates by fuel types were derived as the countries not covered at all in the Penn World Table: Cuba (1990), East
difference between (production 1 imports) and (exports 1 bunker Germany (1985–1990), North Korea (1990), and Vietnam (1990). The
fuel 1 increases in stocks). Carbon emissions were calculated from the Factbook asserts that the real GDPs for East Germany and North Korea
consumption figures using standard conversion factors. Apparent data were computed using purchasing-power-parity exchange rates. In estima-
errors produced 14 negative carbon emissions estimates (out of well over tion, using market instead of purchasing-power-parity exchange rates for
4000 total observations on E and C ); the corresponding observations were Cuba and Vietnam affects only estimates of the corresponding country
dropped. fixed effects.
18 THE REVIEW OF ECONOMICS AND STATISTICS

B. The Econometric Model TABLE 1.—FRACTIONS OF VARIANCE EXPLAINED

ln of ln of
Using i to refer to countries and t to refer to years, the Per-Capita Per-Capita
analysis that follows employs equations of the following Carbon Energy
general form: Model Emissions Consumption

Full model (10 income segments, time

ln (cit ) 5 ai 1 bt 1 F[ln (yit )] 1 eit (1) fixed effects) 0.9760 0.9784
Country effects only 0.9424 0.9380
Income effects only 0.8308 0.8482
where c 5 C/N, y 5 Y/N, the ai and bt are country and year Time effects only 0.0113 0.0141
fixed effects, respectively, F is some reasonably flexible Percentage of within-country variation
function, and eit is the error term. We employ per-capita explained
quantities because we see no reason why national population Income effects only (10 segments) 0.5277 0.5769
Time effects only 0.5054 0.5322
should affect average behavior. Log-linear specifications are Income and time effects 0.5836 0.6511
attractive primarily because multiplicative country and year
Country effects and 10 income segments
fixed effects seem more plausible than additive effects, given Time spline (model 10S) 0.9756 0.9779
the vast differences among nations in our data. In addition, Log trend (model 10L) 0.9754 0.9777
HES examine both linear and log-linear models of this Country effects and 8 income segments
general sort and report no large differences. In our results we Time spline (model 8S) 0.9751 0.9778
also include some estimates of the structure of per-capita Log trend (model 8L) 0.9749 0.9775
energy usage eit 5 Eit/Nit for comparison. Namely, we Notes: Except for lines 5–7, the numbers shown are R 2 statistics. Lines 2–4 are taken from regressions
in which only the effects indicated are present. Lines 5–7 show the fractions by which the residual sums of
estimate equations of the form (1) with ln (eit ) as the squares from the ‘‘country effects only’’ regressions are reduced by adding the effects indicated. The last
four lines show the effects of replacing time fixed effects by the two simple time effect representations
dependent variable. discussed in the text; these are the models developed in section IV and used for projections in section V.
Reduced-form equations of this sort necessarily reflect
both production and demand relationships. Data on domestic
prices and relevant policy variables, even if available, would We began econometric analysis of log per-capita CO2
not alter this. The ai terms in equation (1) reflect persistent emissions, ln (c), with 20- and 24-segment splines, with
differences across countries in fossil fuel availabilities and each segment containing the same number of data points,
prices, output mixes, regulatory structures, tax and subsidy and considered simplifications that preserved this symmetry.
policies, and tastes. The bt terms reflect changes over time in Using the 0.05 significant level, simplification to 10 or 12
domestic prices, for which historical data are not available segments could not be rejected, but further simplifications
outside the OECD and which, because of the importance of could be. As the 10-segment and 12-segment specifications
taxes and subsidies, must be considered endogenous in the were nearly identical, we retained the former on grounds of
long run. While bt includes the impacts of world oil prices, it simplicity. We tested for shifts in the spline coefficients over
is worth noting that nonoil energy sources such as coal are time and for differences between planned and market
more important to CO2 emissions. In addition, the bt reflect economies. In both cases statistically significant differences
changes in technologies in use, environmental policies and were detected, but the differences were small and without
standards, and relevant taxes and subsidies, as well as obvious pattern. Accordingly, we retained the null hypoth-
changes in tastes unrelated to income levels. Future projec- esis in both cases.14 We obtained similar results for log
tion of these effects is an important part of the forecasting per-capita energy consumption ln (e).
process, and we discuss this in detail in section IV. Table 1 shows that equation (1) with a 10-segment spline
for F explained 97.6% of the sample variance in ln (c) and
97.8% of the sample variance in ln (e). The slightly lower R 2
III. Estimation Results for ln (c) presumably reflects the effects of idiosyncratic
Following HES, who report results for quadratic and
14 We also tested for heteroskedasticity. Regression analysis revealed that
cubic specifications, we initially approximated the function
squared residuals were significantly smaller on average for countries with
F by polynomials. Sixth-order functions had all significant larger sample-average real GDP and, to a lesser extent, for those with
coefficients and fit the data well. In part because we have a larger sample-average population. Because of sample sizes, these regres-
large sample, however, lower order polynomials fit the data sions decisively rejected the null hypothesis with R 2 values of only about
0.03. GLS estimation of equations (1) produced results quite similar to
nearly as well, and polynomials with essentially identical fits those reported in the text. The top-decile elasticity for c was smaller in
and in-sample shapes implied wildly different predictions absolute value (20.18) than the ordinary least-squares (OLS) estimate
for income levels above the sample range. In order to avoid shown in table 2 but remained significant. (The corresponding top-decile
elasticity for e was both small (20.06) and insignificant.) The 10-segment
this problem, we took F to be a spline (piecewise linear) energy consumption and carbon emissions forecasts generated from
function. Our forecasts thus involve the assumption that the weighted regressions were somewhat higher than those reported in the
income elasticity estimated for the sample observations with text. Since weighting to correct for heteroskedasticity did not materially
change our main results, and since the weighted analysis is considerably
the highest levels of per-capita GDP will also apply at all more complex, we present only OLS estimates and the corresponding
higher income levels. projections in the text.
 WORLD CARBON DIOXIDE EMISSIONS: 1950–2050 19

changes in nation-specific circumstances affecting the car- TABLE 2.—ESTIMATED INCOME ELASTICITIES FROM 10-SEGMENT SPLINES
WITH TIME AND COUNTRY EFFECTS
bon intensity of energy consumption. Coefficient estimates
and other results for these two dependent variables were Carbon Emissions Energy Consumption
always very similar. While this reflects the high sample GDP Range Elasticity t-Statistic Elasticity t-Statistic
correlation (r 5 0.9974) between ln (c) and ln (e), the (1985$/capita) (Std. Error) on Difference (Std. Error) on Difference
strength of this correlation is somewhat surprising in light of 200–629 20.28 20.13
the significant differences in the carbon intensities of various (0.10) (0.09)
3.82 2.86
countries’ fuel mixes. In any case, in the following we focus 629–932 0.31 0.28
on carbon emissions because of their greater policy interest. (0.10) (0.09)
Table 1 also provides information on the relative impor- 5.54 5.38
932–1283 1.29 1.18
tances of country, income, and time effects in these data. (0.12) (0.11)
Even though our sample spans four decades, differences 22.68 22.49
between countries are more important than changes within 1283–1728 0.79 0.75
(0.11) (0.10)
countries over time: about 94% of the variance of each of the 1.71 2.08
dependent variables is accounted for by country fixed 1728–2352 1.10 1.09
effects. Time effects and differences in income over time (0.10) (0.10)
22.34 22.58
have roughly equal power in explaining the remaining 2352–3190 0.66 0.65
within-country variance. Note that country fixed effects are (0.11) (0.10)
slightly less important for energy consumption than for 20.71 20.69
3190–4467 0.54 0.53
carbon emissions, while the reverse holds for income and (0.10) (0.09)
time effects. This is consistent with country-specific factors, 1.08 1.01
such as fossil fuel reserves, playing a relatively greater role 4467–6598 0.71 0.68
(0.09) (0.08)
in carbon emissions per unit of energy than in the relation 24.37 23.24
between energy and economic activity.15 6598–9799 0.07 0.23
Some patterns are apparent in the estimated country fixed (0.09) (0.08)
22.46 23.20
effects, but a detailed analysis would be beyond the scope of 9799–19,627 20.30 20.22
this paper. The estimated ai for the United States is relatively (0.09) (0.09)
large: the United States ranks fifth for ln (c). Other countries Note: Estimated income elasticities are shown for each sample decile, along with t-statistics for
differences between elasticities in adjacent ranges.
with relatively large estimated fixed effects are oil exporters
(Qatar, United Arab Emirates, Bahrain, Kuwait) countries
that had centrally planned economies in the sample period
It is clear that the pattern of estimated income effects is
(Czechoslovakia, USSR, East Germany, Bulgaria), and
consistent with the data for the countries illustrated in figure
some OECD members (Luxembourg, Canada, Belgium,
1. The developing countries, with lower values of GDP per
West Germany). Countries with low estimates of ai are
capita, experience continued rapid carbon emissions growth,
generally poor countries where real GDP measurement is
even through the period of oil shocks of the 1970s. The more
relatively difficult:16 the lowest five ai were for Nepal, Laos,
highly developed countries showed a clear change in carbon
Ethiopia, Rwanda and Burundi, and Chad.
emissions in the 1970s from growth to either stability or
Table 2 shows the estimated income elasticities for carbon
decline. As table 3 shows, both carbon emissions per capita
emissions and energy consumption. The corresponding
and energy consumption per capita peaked during the 1970s
income–emissions relation for CO2, normalized for the
for other leading OECD nations.18
United States in 1990, is graphed in figure 2. The negative
This raises the further issue of whether the same income
estimated elasticities for the lowest sample decile do not
structure F applies to developing countries and developed
have a material effect on our out-of-sample projections
countries. As a statistical matter, the null hypothesis that the
because only a small and declining fraction of the future
parameters of the income function F are the same for OECD
world population is assumed to have incomes in this range.
and non-OECD nations was decisively rejected. The esti-
The negative and significant elasticity estimates for the
mated differences were small and nonsystematic, however,
highest decile do have an important impact on our projec-
and we elected to retain the null hypothesis.19 There is
tions, however.17
something of an identification problem here, since there is
15 Other entries in table 1 refer to specializations of the basic model used relatively little overlap between the per-capita income
for the projections (time trend models and refinements of the income
structure), as discussed in section IV. 18 It is also worth noting that except for West Germany, energy
16 In addition, traditional fuels (or noncommercial biomass) are relatively consumption peaks with or after carbon emissions. This is consistent with a
important in low-income countries; see footnote 10. shift toward gas and nuclear power in Europe and away from coal
17 HES also find evidence for negative elasticities at high income levels. generally (with Germany the exception).
Perhaps because they employ more restrictive representations of the 19 For exactly the same reason, we retained the null hypothesis that the

income function F, however, their estimates imply positive elasticities income function coefficients were the same for nations with centrally
until well above the sample range. planned economies as for other nations.
20 THE REVIEW OF ECONOMICS AND STATISTICS

FIGURE 2.—INCOME EFFECTS FROM 10-SEGMENT CO2 REGRESSION: UNITED STATES, 1990

TABLE 3.—OECD COUNTRIES WITH PRE-1985 PEAKS IN PER-CAPITA rity concerns, and shifts away from heavy manufacturing—
CARBON EMISSIONS OR ENERGY CONSUMPTION
all of which are income related in the medium or long term
Year of Peak as an empirical matter.21
Per-Capita Per-Capita The estimated time effects for carbon emissions are
Country Carbon Emissions Energy Consumption shown in figure 3. These effects grow significantly through-
Austria 1979 1979 out the majority of the sample. They exhibit a slowdown in
Belgium 1973 1979 the latter part of the sample but do not exhibit a negative
Canada 1979 —
Denmark 1979 1979 trend. With reference to earlier discussion, it is clear that the
Finland 1980 — estimates of bt pick up a general upward but possibly
France 1973 1979 slowing trend, and do not appear to be dominated by any
West Germany — 1979
Japan 1973 — specific influence such as the variations in world oil prices of
Luxembourg 1974 1974 the 1970s.
Netherlands 1979 1979
Sweden 1970 1976
Switzerland 1973 — IV. Projection Methods
United Kingdom 1970 1979
United States 1973 1973 In order to see whether the emissions projections of the
IPCC are consistent with the historical record, we used our
estimates of equations (1) and the income and population
growth assumptions employed by the IPCC to generate
distributions of the two groups of nations.20 Whatever way
unbiased forecasts of C over the 1990–2050 period. The
one wants to interpret our (reduced-form) estimates, it is
IPCC itself has done projections to 2100, but we felt this was
clear that the world oil price is not the only important factor
that has varied over time in our sample period. The
21 A more serious question is whether the relation between these factors
difference between OECD and non-OECD behavior points
and per-capita income is likely to be the same in the future as in the recent
up the importance of environmental policies, national secu- past, since future decisions in all nations will be made with different
technological and environmental information than past decisions. Greater
20 See U.S. Energy Information Administration (1994, p. 11) for other knowledge of environmental risks may or may not offset advances in
information on the differences between OECD and non-OECD patterns of energy-using technologies. At any rate, our methods allow us to extrapo-
energy consumption and carbon emissions. late history, not to consider these or related structural changes.
 WORLD CARBON DIOXIDE EMISSIONS: 1950–2050 21

FIGURE 3.—TIME EFFECTS FROM 10-SEGMENT CO2 REGRESSION

beyond the period for which historical experience could for scenarios A and B, and we use the average of the IPCC’s
provide a useful benchmark. projections for comparison purposes.
The IPCC’s assumptions are summarized in table 4 and in As Eckaus (1994) and others have noted, the IPCC’s
Pepper et al. (1992). We obtained the five-year regional growth assumptions are generally conservative in light of
growth assumptions employed by the IPCC on floppy disk recent experience. Also, as Nordhaus (1994, pp. 13–14)
from participants in the IPCC process. The IPCC used the points out, there is no historical basis for the common
same income and population growth assumptions for its assumption, made by the IPCC in all scenarios, that per-
scenarios A and B. These scenarios differ in other exogenous capita income growth slows over time. Because we are not
variables that we do not employ and produce very similar persuaded that the IPCC assumptions are a fair representa-
projections. We use scenario A/B to denote projections made tion of the distribution of plausible future growth outcomes,
using the IPCC income and population growth assumptions we view the absolute levels of the projections discussed in
this paper as primarily illustrative. Our primary focus is on
comparisons of our projections with the IPCC’s projections.
TABLE 4.—SUMMARY OF IPCC POPULATION AND GDP GROWTH ASSUMPTIONS
In order to calculate projections, we need to address two
Average Annual Scenario Scenario Scenario Scenario Scenario
Growth Rate A/B C D E F methodological questions. First, should the negative elastic-
ity for the top income segment be taken at face value, or
Population
1990–2025 1.35 1.05 1.05 1.35 1.68
should it be dismissed as an artifact of the timing of policy
2025–2050 0.70 0.12 0.12 0.70 1.12 changes, oil shocks, and other factors of the 1970s and
1990–2050 1.08 0.66 0.66 1.08 1.44 1980s? Second, how should the time effects be projected?
GDP per capita We take up these questions in turn.
1990–2025 1.51 0.85 1.66 2.20 1.31 The question of whether one should use a negative
2025–2050 1.40 0.77 1.71 2.05 1.19
1990–2050 1.46 0.82 1.68 2.14 1.26 income elasticity for countries with high per-capita GDP is
GDP
an important question. In 1990 about 17% of the sample
1990–2025 2.86 1.91 2.71 3.55 2.98 population has y in the top segment, but under the IPCC
2025–2050 2.10 0.89 1.82 2.75 2.31 growth assumptions this percentage rises to at least 47% by
1990–2050 2.54 1.48 2.34 3.22 2.70
2025 and to at least 73% by 2050.
22 THE REVIEW OF ECONOMICS AND STATISTICS

TABLE 5.—INCOME AND TIME EFFECT SPECIFICATIONS In terms of the estimates for carbon emissions, the
Spline Model Log Model estimated annual trend increase was roughly the same in
Income (Constant Trend) (Flattening Trend) 1990 for model 10L as for model 10S (0.70 versus 0.73%)
10-segment (negative top elasticity) 10S 10L and for 8L as for 8S (0.53 versus 0.59%). (The difference
between the 10-segment and 8-segment specifications re-
8-segment (all positive elasticities) 8S 8L
flects the negative income effects estimated for some
countries in the former.) In the log-trend models the
estimated increase falls over time, to 0.25% per annum by
2050 under model 10L and to 0.002% under model 8L.
The correct answer to this question does not seem
obvious, and so we present projections with two alternative
V. Projection Results and Comparisons
approaches to the top segment elasticity. The first approach
is to take the negative top-segment elasticities at face value Figure 4 shows carbon emissions projections relative to
and employ our 10-segment estimates. The second approach predicted emissions in 1990 from our four models and from
is to use a model without a negative top-segment elasticity, the IPCC for the central case of scenario A/B.24 Our
obtained from the original model by constraining the top projections match the IPCC in 1990 by construction,25 but
three segments to have the same elasticity. Estimates of the are higher in every period after 1990. All of our models
resulting 8-segment model had all income elasticities posi- project more rapid growth than the IPCC through 2025; all
tive and had an R 2 value that was 0.0005 lower than the but one (10L) also show more rapid growth from 2025 to
original model’s R 2. Time and country fixed effects were not 2050.
changed substantially by these modifications, though, as one Note that models 8L and 8S predict more growth than 10L
might expect, time effect growth is slower after 1970 in the and 10S, respectively, because of the negative top-segment
8-segment model than in the 10-segment model. income elasticities in the latter specifications. Similarly,
We now turn to the question of how to project the time models 10S and 8S predict more growth than 10L and 8L,
effects.22 At issue is the best way to extrapolate the pattern of respectively, because of the slowdown in time-effect growth
time effects shown in figure 3. We examined several models built into the latter two models. While the differences among
of the time effects and report projections based on fitting the our growth projections are substantial, at least through 2025
model with two alternative specifications of the time effects. they are less important than the difference between our
First, we specified a linear spline model (model S) with projections on the one hand, and those of the IPCC on the
different growth rates prior to 1970 and after 197023: other.
We compute standard errors of our forecasts in order to
bt 5 b0 1 b1t 1 b2 (t 2 1970) · 1[t $ 1970]. (2) assess the statistical significance of these differences. The
appendix describes the calculation of the standard errors,
Second, we specified a nonlinear trend model including a which reflect a standard calculation capturing the effect of
logarithmic term (model L): uncertainty due to future disturbances as well as the use of
estimated parameters. Some caveats are in order: our
bt 5 b0* 1 b1*t 1 b2* ln (t 2 1940). (3) standard errors condition on the specification of the time
effect used, and our calculations do not explicitly take into
When the original model was estimated, these two alterna- account either spatial correlation in errors across countries,
tive time effect specifications had essentially the same or serial correlation for disturbances within countries.26
goodness-of-fit performance, as shown in the bottom half of Problems along these lines could give rise to larger standard
table 1. error estimates.
However, the two specifications represent different speci- Figure 5 provides comparisons of our growth projections
fications for out-of-sample projections. The spline model S and those of the IPCC for all five scenarios for 2050, along
(equation (2)) projects the time effects by continuing the with approximate 95% confidence intervals27 for our projec-
estimated 1970–1990 trend to 2050. The log model L
(equation (3)) projects a flattening trend consistent with the 24 We cannot usefully compare our projections with those of HES, since
trend deceleration from 1950 to 1990 shown in figure 3. We they develop and employ their own projections of growth in per-capita
know of no a priori basis for preferring one of these time income.
25 All our models’ initial 1990 predictions exceed actual 1990 emissions.
effect specifications to the other. Thus, we present results
We scaled the ai estimates to match actual 1990 emissions for each
from projecting four different models, which are denoted as country, so that we can focus on long-term growth projections.
in table 5. 26 We examined residuals for first-order serial correlation; and found that

there was some mild estimated dependence (for example, a first-order

autocorrelation parameter of 0.4) for some of the countries. However, we
22 For their main case, HES simply set the time effect at its value in the did not carry out a country-by-country time-series analysis for correcting
last year in their sample. the overall standard errors of our forecasts.
23 1[t $ 1970] denotes a dummy variable, which is 1 when t $ 1970 and 27 These intervals are defined as the projected value 61.96 times the

0 otherwise. estimated standard error.

 WORLD CARBON DIOXIDE EMISSIONS: 1950–2050 23

FIGURE 4.—COMPARISON OF CO2 EMISSIONS FORECASTS FOR SCENARIO A/B

tions. Our projections are higher for all models for all but years of the period studied.29 On the one hand, one might
one scenario. For the slow-growth scenarios C and D, all expect that forecasts 60 years into the future would be so far
differences are statistically significant.28 For scenarios A/B out of sample as to contain little useful information. On the
and F, the S specifications (which continue the time trend of other hand, under the IPCC scenarios, most of the world’s
1970–1990) are significantly different at the 5% level. For population is projected to have per-capita income levels
scenario E, our projections are similar or, if anything, within the sample range for most of the forecast period. In all
somewhat below those of the IPCC. In summary, our results scenarios at least 89% of the world’s population is projected
are substantially and significantly above the IPCC results for to live in countries with y within the sample range in 2025;
the two slow-growth scenarios C and D, well (and signifi- by 2050 this lower bound falls to only 69%.
cantly for half the models) above the IPCC’s for the We computed the approximate distributions of differences
moderate-growth scenarios A/B and F, and, if anything, between forecasts under different scenarios (also described
somewhat below the IPCC’s for the rapid-growth scenario in the appendix), and used those distributions to test the null
E. hypotheses that the observed differences were drawn from
Though the IPCC projects the highest emissions in distributions with zero means. With a very few exceptions
scenario E, we project higher emissions in scenario F. As
(which occur early in the forecast period and reflect absolute
table 4 shows, scenario F has more rapid population growth
small differences in assumed population and income levels),
than scenario E, and all our models embody a unitary
all these null hypotheses were rejected at well below the 1%
elasticity of emissions with respect to population. Scenario
level. Thus as a statistical matter at least, our projection
E has more rapid growth in per-capita income, but all our
models have per-capita income elasticities substantially process provides useful information about differences be-
below unity over much of the relevant range. A comparison tween scenarios throughout the period analyzed.
of these two scenarios also reveals the negative impact of A second question raised by figure 5 (and the analogous
high per-capita income growth in models 10L and 10S. results for earlier years) is why the IPCC’s projections under
Figure 5 also raises the question of whether the differ- scenarios C and D are so low relative to our extrapolation of
ences between our projections under the various IPCC
scenarios are statistically significant, particularly in the later 29 A conceptually harder question, which we do not attempt to answer

here, is whether the projections from different models are statistically

28 For model 10S in scenario D, the difference has a p-value of 6.1%. distinct.
24 THE REVIEW OF ECONOMICS AND STATISTICS

FIGURE 5.—2050 EMISSIONS AS A PERCENTAGE OF 1990 EMISSIONS WITH 95% CONFIDENCE INTERVALS

historical experience. Leggett et al. (1992) list a number of The contrast between projections for the OECD on one
assumptions for each scenario in addition to those regarding hand, and for China and India on the other is striking.
income and population growth, but it is unclear what effect Together, China and India account for 14.8% of 1990 carbon
those assumptions have on the results. It does seem clear that emissions in our data. By 2050 we project these two nations
drastic emissions controls are not being assumed, and one to account for between 27 and 30% of emissions. Perhaps
could argue that such controls would be politically unlikely more important, we project them to account for between 31
anyway under such slow growth in living standards. Analy- and 44% of emissions growth over the 1990–2050 period.
sis of forecast output does suggest two partial answers. These percentages would be even more impressive, of
The first answer involves differences in projections in course, under income growth assumptions more in line with
carbon intensity. For scenarios C and D the IPCC projects a recent experience in China and India. Even under the IPCC’s
much higher fall in carbon intensity than our models do. Our assumptions, as many observers have argued, carbon emis-
models give similar results for carbon intensity for all sions growth in China and India must be controlled if global
scenarios, which are comparable to the IPCC results for emissions growth is to be slowed relative to historical trends.
scenarios A/B, E, and F.30 The second answer is based on A final question that arises in this context is how to
regional differences. The OECD accounted for about 46% of summarize the projection uncertainty induced by the varia-
emissions in 1990 in both our and the IPCC’s data. Across tion in growth assumptions across IPCC scenarios. In its
the various scenarios, the IPCC projects that this share will recent review (Alcamo et al. (1995)), the IPCC uses the ratio
decline to between 26 and 31% by 2050; this is between the of maximum to minimum projections as a measure of
shares projected by our 10-segment (19–22%) and 8- uncertainty.31 By this measure, the IPCC’s work implies
segment (29–32%) models. However, further analysis shows greater uncertainty than any of our models (see table 6).
that we generally project the OECD to account for smaller An advantage of the econometric approach employed here
fractions of emissions growth over the 1990–2050 period, is that we can go beyond ad hoc comparisons of point
and that the IPCC projects declines in OECD emissions in forecasts to systematic analysis of forecast distributions. We
both scenarios C and D that are out of line with our attached a subjective probability of 1/3 to scenario A/B,
extrapolation of historical experience.
31 In fact, the IPCC uses the ratio of maximum to minimum published
30 Alcamo et al. (1995, fig. 6.6) shows that the IPCC’s carbon intensity projections, so that authors’ and editorial boards’ collective willingness to
projections in scenarios C and D are also outliers in the set of published publish outliers is used to calibrate judgments regarding forecast uncer-
projections. tainty. It is hard to imagine any persuasive rationale for this approach.
 WORLD CARBON DIOXIDE EMISSIONS: 1950–2050 25

TABLE 6.—RATIOS OF MAXIMUM TO MINIMUM FORECASTS IPCC range is also somewhat hard to interpret (and too
AND UPPER TO LOWER CONFIDENCE INTERVAL BOUNDS
small) because of their neglect of forecast uncertainty.
Maximum/Minimum Upper/Lower Confidence
Forecasts Interval Bounds
Model 2025 2050 2025 2050 VI. Concluding Observations
IPCC 1.82 2.86 — — Most of the more vexing issues facing policymakers
10L 1.27 1.59 1.71 2.15
8L 1.30 1.63 1.74 2.23
involve the results of complex systems, and the impact of
10S 1.26 1.59 1.66 2.04 carbon emissions on the global environment is no exception.
8S 1.29 1.63 1.69 2.10 To understand what generates carbon emissions and how
Notes: The first and second columns give the ratio of the highest forecast for 2025 and 2050 to the those emissions affect global warming, one needs an under-
lowest forecasts for that year. (For the IPCC, this is the ratio of the forecast for scenario E to that for
scenario C. For our models this is the ratio of the forecast for scenario F to that for scenario C.) The third standing of the process of global economic development, as
and fourth columns give the ratios of the upper bound of the relevant 95% confidence interval (discussed
in the text) to the lower bound of that interval. well as an understanding of the physical and ecological
systems that translate emissions into levels of greenhouse
gases and then into global warming. Structural simulation
which combined two of the original IPCC scenarios, and 1/6 models of these systems are necessarily complex because
to each of the other four scenarios. Then, as discussed in the the systems they represent are complex. Because of limita-
appendix, treating the five scenario-specific forecast distribu- tions on the availability of detailed data on all relevant
tions as conditional distributions yields a set of model- and aspects of such systems, structural simulation models are
year-specific confidence intervals. As the last two columns typically based on very little observed data. For instance,
in table 6 indicate, the widths of these intervals are they are usually calibrated or benchmarked to a set of
comparable to the ranges of IPCC point forecasts. observations from a single year.
Figure 6, which is representative of all four models, The importance of the issues faced by policymakers, as
shows that our analysis places the range of likely outcomes well as their needs for structural simulation models as tools
appreciably above the range found by the IPCC. The IPCC for making informed decisions, bring to the fore the need for
range is pulled down at the bottom by inclusion of their the validation of such structural models. And since the only
projections for scenarios C and D, which, as we have source of information on the operation of the complex
discussed, depart downward from historical trends. The systems under study comes from history, validation must

FIGURE 6.—IPCC SCENARIOS C AND E (SOLID) AND 95% CONFIDENCE INTERVAL BOUNDS FROM MODEL 10S (DASHED)
26 THE REVIEW OF ECONOMICS AND STATISTICS

consider how well the structural model reflects historical sumption are negative at high income levels raises a host of
data. For instance, one can carry out historical simulation by research issues. For instance, future research is necessary to
calibrating the model at a past time period and comparing determine whether the reversal of the growth of carbon
simulations to the observed historical data patterns. How- emissions with higher economic activity has been due to a
ever, to evaluate the results of forecasts made from a heightened environmental awareness and a tighter regula-
simulation model, some method of comparing historical data tory regime in industrialized countries, or whether there are
patterns to the forecasts is necessary for validation. fundamental economic trends in development that lead
This paper uses an econometric approach to compare the countries away from carbon-intensive energy sources.
forecasts of carbon emissions by the IPCC model with the However, even allowing for this decline in carbon emis-
recent historical record. Our reduced-form econometric sions, it is worth stressing our finding that the IPCC’s
approach permits the systematic distillation of decades of low-growth emissions projections are way too low to be
worldwide experience. As such, our approach permits a clear consistent with the historical experience, while their high-
translation of historical data patterns that can be applied to growth scenarios are more consistent with our own projec-
the IPCC’s income and population scenarios to give the tions. While one can easily list reasons why the future might
carbon emission paths comparable with the patterns of depart from the past in this regard, not all such reasons imply
recent history. And those recent patterns, as summarized by lower carbon emissions. In any case, we feel our findings
our model, give much higher worldwide CO2 emissions than can play an important role in forming judgments of a the
the IPCC forecasts in all but one population and income nature and quality of the emissions forecasts, which are
growth scenario, with many of the differences statistically necessary components of global environmental planning.
significant.
In addition, our econometric approach permits judgments REFERENCES
on the precision of forecasts that are based directly on the Alcamo, J., A. Bouwman, J. Edmonds, A. Grübler, T. Morita, and A.
goodness of fit of our reduced-form model to the historical Sugardhy, ‘‘An Evaluation of the IPCC IS92 Emission Scenarios,’’
data. In other words, the quality of the adherence of our in J. T. Houghton, L. G. Meira Filho, J. Bruce, Hoesung Lee, B. A.
Callander, E. Haites, N. Harris, and K. Maskell (eds.), Climate
reduced-form model to the observed data series plays a key Change 1994 (Cambridge, UK: Cambridge University Press,
role in the evaluation of the forecasts. And, under the 1995).
standard (95%) criteria for constructing confidence inter- Cline, William, The Economics of Global Warming (Washington, DC:
Institute for International Economics, 1992).
vals, many of our reduced-form forecasts of carbon emis- Eckaus, Richard S., ‘‘Potential CO2 Emissions: Alternatives to the IPCC
sions are significantly different than the IPCC forecasts. This Projections,’’ Mimeo, MIT Joint Program on the Science and Policy
constitutes strong evidence that the corresponding IPCC of Global Change (Aug. 1994).
Holtz-Eakin, Douglas, and Thomas M. Selden, ‘‘Stoking the Fires? CO2
forecasts are considerably below what one would anticipate Emissions and Economic Growth,’’ Journal of Public Economics
from recent historical experience. For policymakers to make 57 (May 1995), 85–101.
intelligent use of the IPCC forecasts, our evidence under- Intergovernmental Panel on Climate Change (IPCC), Climate Change:
The IPCC Scientific Assessment (Cambridge, UK: Cambridge
lines the need to understand what parts of the structure of the University Press, 1990).
IPCC model cause such strong departures from historical ——— Climate Change 1992: The Supplementary Report to the IPCC
trends. Scientific Assessment (Cambridge, UK: Cambridge University
Press, 1992).
In terms of evaluating our forecasts on a stand-alone ——— Climate Change 1995, 3 vols. (Cambridge, UK: Cambridge
basis, there are at least two limitations. First, we have not University Press, 1996).
undertaken the development of our own projections of GDP Leggett, Jane, W. J. Pepper, and R. J. Swarf, ‘‘Emissions Scenarios for the
IPCC, an Update,’’ in IPCC (1992), chap. A3.
or population growth for each country, nor have we at- Lichtenstein, Sarah, Baruch Fischoff, and Lawrence D. Phillips, ‘‘Calibra-
tempted to justify whether the IPCC projections of economic tion of Probabilities: The State of the Art to 1980,’’ in D.
growth are particularly good or bad. Second, the explicit Kahneman, P. Slovic, and A. Tversky (eds.), Judgement under
Uncertainty: Heuristics and Biases (Cambridge, UK: Cambridge
reliance of our approach on recent history can be a limitation University Press, 1982), 306–334.
for creating forecasts, as our reduced-form model will not Manne, Alan S., and Richard G. Richels, Buying Greenhouse Insurance:
account for dramatic innovations or alternative structural The Economic Costs of CO2 Emissions Limits (Cambridge, MA:
MIT Press, 1992).
changes that are away from historical trends. Because ——— ‘‘The Costs of Stabilizing Global CO2 Emissions: A Probabilistic
radical innovations and structural changes become more Model Based on Expert Judgements,’’ Energy Journal 15 (1994),
likely the farther one looks into the future, and because 31–56.
Marland, Gregg, T. A. Boden, R. C. Griffin, S. F. Huang, P. Kanciruk, and
forecast uncertainty rises over time, we feel that it is T. R. Nelson, ‘‘Estimates of CO2 Emissions from Fossil Fuel
doubtful that our approach can provide useful projections Burning and Cement Manufacturing, Based on the United Nations
beyond about 2050 in any case. For much climate change Energy Statistics and the U.S. Bureau of Mines Cement Manufactur-
ing Data,’’ ORNL/CDIAC-25, NDP-030, U.S. Department of
analysis, longer horizons are often relevant. Energy, Oak Ridge National Laboratory, Oak Ridge, TN (May
Beyond providing evidence to evaluate the IPCC fore- 1989).
casts of carbon emissions, our empirical results have substan- Nordhaus, William D., Managing the Global Commons: The Economics of
Climate Change (Cambridge, MA: MIT Press, 1994).
tive implications as well. The finding that the income Pepper, William, J. A. Leggett, R. J. Swarf, J. Wasson, J. Edmonds, and I.
elasticities of per-capita carbon emissions and energy con- Mintzer, ‘‘Emission Scenarios for the IPCC, an Update: Assump-
 WORLD CARBON DIOXIDE EMISSIONS: 1950–2050 27

tions, Methodology, and Results,’’ Mimeo, U.S. Environmental becomes

Protection Agency (May 1992).
Schmalensee, Richard, ‘‘Comparing Greenhouse Gases for Policy Pur- E [(Yt 2 Pt )2 ] > [exp (s2 ) 2 1] o (N ) f (b, s )
i
it
2
it
2 2

poses,’’ Energy Journal 14 (1993), 245–255.

­fit
3o 4
Selden, Thomas M., and Daqing Song, ‘‘Environmental Quality and
Development: Is There a Kuznets Curve for Air Pollution Emis- 1 Nit (A.5)
i ­(b, s2 )
sions?’’ Journal of Environmental Economics and Management 27
­fit
3o 4
(Sept. 1994), 147–162.
Summers, Robert, and Alan Heston, ‘‘The Penn World Table (Mark 5): An 3 var (b, s 2 ) Nit .
i ­(b, s2 )
Expanded Set of International Comparisons, 1950–1988,’’ Quar-
terly Journal of Economics 106 (May 1991), 327–368. var (b, s 2 ) is block diagonal, with the upper block equal to the estimated
‘‘Symposium on Global Climate Change,’’ Journal of Economic Perspec- covariance matrix of b and a scalar lower block equal to var (s 2 ). If the
tives 7 (Fall 1993), 3–86. regression has M degrees of freedom, the assumption of normality implies
U.S. Energy Information Administration, ‘‘Energy Use and Carbon that Ms 2/s2 is distributed as x2 (M ). Since the variance of this random
Emissions: Some International Comparisons,’’ DOE/EIA-0579, variable is 2M, var (s 2 ) 5 2M(s4/M 2 ) 5 2s4/M.
Energy Information Administration, Washington DC, (Mar. 1994). 2. To test the significance of differences between forecasts conditional
on the inputs from different scenarios, we compute standard errors for
these differences under the assumption that the disturbances are the same
across scenarios.32 Using the notation above, let P 1t be the forecast for
APPENDIX some year t under scenario 1, and let P 2i be the forecast under scenario 2.
The basic models are
In this appendix we show (1) how standard errors of forecasts were
1 2 1 2
Y 1it Y 2it
computed, (2) how tests for differences between forecasts were performed, and ln 5 X 1itb 1 eit and ln 5 X 2itb 1 eit (A.6)
(3) how multiscenario confidence intervals were computed. Here let Yit equal N 1it N 2it
total carbon emissions in country i during year t. Then, following equation (1),
the models used in forecasting can be written as where X 1it and X 2it include country and time effects as well as scenario-
specific income and population inputs. Equations (A.6) give the true
aggregate values as
1N 2 5 X b 1 e
Yit

o N f (b, s )u o N f (b, s )u
ln it it (A.1)
it Y 1t 5 1
it
1
it
2
it and Y 2t 5 2
it
2
it
2
it (A.7)
i i

where Xit includes country, time, and income effects, and eit is assumed
where, as before, uit ; exp [eit 2 (s2/2)] and fjit (b, s2 ) ;
normal with mean zero and variance s2. Total global emissions or
exp [X jitb 1 (s2/2)] for j 5 1, 2.
consumption in year t is then given by
The error in the difference between forecasts is then given by

Yt ; o Y 5 o N f (b, s )u
it it it
2
it (A.2) (Y 1t 2 Y 2t ) 2 (P 1t 2 P 2t ) 5 (Y 1t 2 P 1t ) 2 (Y 2t 2 P 2t )

o [N f (b, s ) 2 N f (b, s )](u

i i
5 1
it
1
it
2 2
it
2
it
2
it 2 1)
i
where (A.8)
2 o 5N [f (b, s ) 2 f (b, s )]
1
it
1
it
2 1
it
2

fit (b, s2 ) ; exp [Xitb 1 (s2/2)] i

2 N 2it [f2it (b, s 2 ) 2 f2it (b, s2 )]6.

uit ; exp [eit 2 (s2/2)]
Consequently, using the same approach that led to equation (A.5), we
have
and the summation is over all countries.
1. Since [eit 2 (s2/2)] is normal with mean 2s2/2 and variance s2, uit is E 5[(Y 1t 2 Y 2t ) 2 (P 1t 2 P 2t )]2 6
lognormal with E(uit ) 5 1 and E (u 2it ) 5 exp(s2 ). If b is the least-squares
estimate of b, s 2 is the usual estimate of s2, and Pt is the unbiased forecast > [exp (s2 ) 2 1] o [N f (b, s ) 2 N f (b, s )]
i
1
it
1
it
2 2
it
2
it
2 2

of Yt, the foregoing implies

­(N 1itf1it 2 N 2itf2it )
Yt 2 Pt 5 o i
Nitfit (b, s )(uit 2 1) 2 1 3o i ­(b, s ) 2 4 8
(A.9)

(A.3) ­(N 1itf1it 2 N 2itf2it )

2 o
i
Nit [fit (b, s 2 ) 2 fit (b, s2 )]. 3 var (b, s 2 )
­(b, s2 ) 3o .
i
4
The various terms in this equation are evaluated as before.
Using the usual first-order approximation, we have 3. Finally, the multiscenario confidence intervals discussed at the end
of section V were calculated as follows. Suppose that there are J scenarios,
E [(Yt 2 Pt )2 ] > o (N ) f (b, s ) E [(u
i
it
2
it
2 2
it 2 1)2 ] with the probability of scenario j occurring being pj. Suppose also that
conditional on scenario j occurring, the analysis of forecast errors implies
that Y is approximately normally distributed with mean Pj and standard
­fit
1 3o i
Nit
­(b, s2 ) 4 (A.4)
deviation hj. Then if F is the standard normal distribution function, the
probability that Y is less than K conditional on scenario j occurring is
F[(K 2 Pj )/hj ]. The unconditional probability that Y is less than K is then
­fit C(K) 5 Sj 5pj F[(K 2 Pj )/hj ]6. Lower and upper confidence bounds are
3 var (b, s 2 ) 3o
i
Nit
­(b, s2 ) 4 obtained by numerical solution of C(Yl ) 5 0.025 and C(Yu ) 5 0.975,
respectively.

where var (b, s2 ) is the covariance matrix of the estimated parameters, and 32 If the disturbances across scenarios were independent, the standard
[Si Nit­f it/­(b, s2 )] is a column vector of derivatives with respect to errors of differences between forecasts would be larger than shown in what
those parameters. Since E [(uit 2 1)2 ] 5 E (u 2it ) 2 1, equation (A.4) follows.