Endogenous Growth, Convexity of Damage and Climate Risk: How Nordhaus' Framework Supports Deep Cuts in Carbon Emissions
Endogenous Growth, Convexity of Damage and Climate Risk: How Nordhaus' Framework Supports Deep Cuts in Carbon Emissions
Endogenous Growth, Convexity of Damage and Climate Risk: How Nordhaus' Framework Supports Deep Cuts in Carbon Emissions
The Economic Journal published by John Wiley & Sons Ltd on behalf of
Royal Economic Society. Published by John Wiley & Sons, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA.
* Corresponding author: Simon Dietz, Department of Economics, London School of Economics and
Political Science, Houghton Street, London WC2A 2AE, UK. Email: s.dietz@lse.ac.uk.
We thank Emanuele Campiglio, Antoine Dechezlepr^etre, Baran Doda, David Greenaway, Tom McDer-
mott, Elisabeth Moyer, Antony Millner, Bill Nordhaus and Bob Pindyck for helpful comments and
discussions, and the editor, Rachel Griffiths. We also acknowledge the financial support of the Grantham
Foundation for the Protection of the Environment and the Economic and Social Research Council. We alone
are responsible for the content.
This is an open access article under the terms of the Creative Commons Attribution License, which permits
use, distribution and reproduction in any medium, provided the original work is properly cited.
1
Shortly afterwards Bill Cline (1992) published what is generally considered to be the other foundational
analysis of climate mitigation benefits and costs.
2
Both accessed on 24 March 2014. However, these citation counts likely understate the paper’s legacy
considerably, since many will instead cite later work that is based on it (see Section 1).
[ 574 ]
[MARCH 2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 575
By the late 1980s, however, climate change was becoming both a policy issue and
increasingly political. In 1988, the Intergovernmental Panel on Climate Change
(IPCC) was established and in 1990 it published the first of its regular and
influential Assessment Reports to member governments. In 1989, the first meeting
of (22) Heads of State to discuss climate change was held in the Netherlands and
various other major international summits that year also put it on the agenda. Most
OECD countries already had their first climate-change targets by 1990 (Gupta,
2010), for instance the European Community, as it was then, had pledged to
3
There is a problem in using the language of benefit–cost analysis, if it is interpreted in its common and
narrow, marginal, fairly undynamic way and where risk is also treated narrowly. Climate-change policy raises
major questions of the strategic management of potentially immense risks and where different paths will have
different endogenous learning and discovery. This broader perspective is a major focus of this study and
should be central to economic research on the topic.
4
To get an idea of the simplicity of the modelling framework, especially the science module, note that a
fully fledged atmosphere–ocean general circulation model such as that of the UK Hadley Centre would
comprise hundreds of thousands of equations.
5
There is little plausibility in moral philosophy for a social welfare function that is the sum across
generations of the (discounted) utility per capita of each generation, irrespective of the number of people in a
generation, unless population is constant. Adding the (undiscounted) total utility of each generation is
essentially utilitarian. Pure-time discounting can be given a utilitarian interpretation if the discounting is
based on the probability of existence as a function of time, and that becomes an exponential function in
continuous time if the end of the world is the first event in a Poisson process.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
576 THE ECONOMIC JOURNAL [MARCH
(v) consumption per capita is given by (exogenous) output, less the total cost of
abating emissions, and the total cost of climate change;
(vi) a reduced-form abatement cost function, in which the total cost of abatement
depends on global aggregate emissions and emissions abatement; and
(vii) reduced-form damage, in which the total cost of climate change depends on
global mean temperature but where global mean temperature is an index of
a wider set of climatic changes including changes in precipitation and sea
level.
6
A careful exploration of the strong basis in moral philosophy for low pure-time discounting is provided
in Stern (2014a,b). In many IAM studies, high pure-time discounting is introduced without much discussion.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
578 THE ECONOMIC JOURNAL [MARCH
source of growth per capita in the long run is exogenous improvements in productivity,
but where climate change only impacts on current output.7 There are compelling
reasons for thinking that climate change could have long-lasting impacts on growth
(Stern, 2013) and there is now an emerging body of empirical evidence pointing in this
direction (Dell et al., 2012), even though climatic conditions in the recent past have
been relatively stable compared with what we now have to contemplate.
Second, we assume that the damage function linking the increase in global mean
temperature with the instantaneous reduction in output is highly convex at some
7
While a reduction in current output may impact future growth via reduced savings – for a given savings
rate – we hypothesise that this effect is weak compared with direct reductions in the capital stock and
reductions in productivity. Fankhauser and Tol (2005) also find a weak impact of climate change on growth
via savings, using DICE. They did not, however, consider that climate damage could work on the capital stock
or on productivity.
8
Nordhaus sees the specification of the damage function for warming above 3°C as a ‘placeholder’ (Stern,
2013) but it is a placeholder that can have a powerful effect on the conclusions as we will see below.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 579
3. Extending DICE
3.1. Endogenous Growth
In standard DICE, the production function is:
Yt ¼ F ðKt ; Lt Þ ¼ ð1 D^t Þð1 Kt ÞAt Kta Lt1a ; (1)
where At is the exogenous element of total factor productivity (TFP) at time t, K is
capital, L is labour and a 2 (0, 1) is the capital exponent. D^ is the standard DICE
3.1.1. A model of capital damage, and knowledge proportional to the capital stock
Our first growth model incorporates knowledge spillovers via the capital stock in the
tradition of Arrow (1962), Romer (1986) and others. We combine this formulation
with a partitioning of the damage multiplier between output and capital. The
production function becomes
Yt ¼ ð1 DtY Þð1 Kt ÞAt Ktaþb Lt1a ; (1.K)
Y
where D now denotes the damage that directly reduce annual output. In this model,
we think of the economy as being composed of a number of firms, each making
investments. Growth is driven in part by learning-by-doing, which in turn depends on
each firm’s net investment, so that when the firm’s capital stock increases, so does
economy-wide productivity. We also make the standard assumption in this tradition
that knowledge is a pure public good. The elasticity of output with respect to
knowledge is b > 0, so that the knowledge process has a productivity factor Kb. These
assumptions have the effect of increasing the overall capital exponent to a + b. We
continue to assume an exogenous element of TFP A. This could be taken to represent
elements of productivity not captured in knowledge spillovers but we use it here
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
580 THE ECONOMIC JOURNAL [MARCH
principally for the narrower, instrumental purpose of calibrating (1.K) on (1) in the
absence of climate-change damage and emissions abatement costs, thus achieving a
controlled comparison of different production specifications.
We suppose there is further damage from climate change that reduces the
capital stock, which we label D K , so we obtain the following equation of motion of
capital:
Ktþ1 ¼ ð1 DtK Þð1 dK ÞKt þ It ; (2)
savings rate s (see Appendix A). In specifying D K we have in mind the representation of
two phenomena. First, D K includes permanent, direct climate damage to the capital
stock, for example, if climate change increases the likelihood of storms and those
storms damage infrastructure, or the abandonment of capital in coastal areas due to
sea-level rise. Second, D K could indirectly include broader impacts of climate change
on productivity via the endogenous growth mechanism (1.K). One effect it could pick
up is of a changing climate on the productivity of capital stocks, accumulated during a
different and more stable climatic regime. For example, water supply infrastructure
may become less productive, given a long-run change in precipitation. Another could
be that, if investment is increasingly diverted towards repair and replacement of capital
damaged by extreme weather, it may produce fewer knowledge spillovers. Appendix A
contains further details of how, for our simulation work, we partition damage D
between D Y and D K .
In sum, according to this model of growth and climate damage, some part of the
instantaneous impacts of climate change falls on capital rather than output, so that this
type of damage represents a permanent reduction in output possibilities in the future.
Moreover since the economy’s stock of knowledge is proportional to its stock of capital,
the negative effect on future output possibilities is magnified.
Ktþ1 ¼ ð1 dK ÞKt þ It : (2 0 )
Notice that in this specification, we do not allow climate damage to impact the
capital stock, although doing so would be straightforward by reverting to (2). The
equation of motion of TFP is given by
Atþ1 ¼ ð1 DtA Þð1 dAt ÞAt þ aðIt Þ; (3)
A
where dA is the net depreciation rate for productivity. We can think of d as
9
Whether such concavity is theoretically or empirically plausible is not for this study.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
582 THE ECONOMIC JOURNAL [MARCH
for example, studies of crop losses and changing energy demand for space cooling
and heating.10 We should recognise, however, that these are ‘quasi’ data points,
since T = 3°C has not been seen on the planet for around 3 million years and
might lead to radical transformations in global climatic patterns. Making assump-
tions about the form of (2) is made still more difficult by the complete absence of
evidence on aggregate impacts for T ≥ 3°C. The quadratic form was originally
selected largely for convenience11 but it results in implausibly low damage at high
temperatures (Stern, 2008; Weitzman, 2012). This has prompted Weitzman (2012)
where the coefficient p3 and its corresponding exponent are together used to satisfy
the assumption that, at T = 6, 50% of output is lost.12 This is the functional form
we use in this study13 but, in addition to Weitzman’s calibration of p3, we offer a
second, alternative calibration such that Dt = 0.5 when T = 4. Science and impact
studies tell us that, not only could we cross several key physical tipping points in the
climate system by the time the 4°C mark is reached (Lenton et al., 2008), the
impacts of such warming on the natural environment, economies and societies
could be severe, with reason to believe in the risk of vast movements of population
and associated conflict, unrest and loss of life (Stern, 2013). Global mean
temperatures regularly exceeding 4°C above pre-industrial have probably not been
seen for at least 10 million years (Zachos et al., 2008) and are within the range of
difference between today and the peak of the last Ice Age, when large ice sheets
covered northern Europe and North America (IPCC, 2013), radically influencing
where people could be. Given the potential magnitude of transformation illustrated
by this example, the assumption that Dt = 0.5 when T = 4 may be no less plausible,
to put it cautiously, than assuming, as (2) does with the standard parameterisation,
that Dt = 0.04 when T = 4, i.e. only 4% of output is lost as a result of temperatures
not seen for 10 million plus years.
In our first growth model, we partition damage as expressed in (40 ) between
damage affecting output D Y and those affecting capital D K , while in our second
model damage are partitioned between output and TFP as in (3). We do so in a
similar way to Moyer et al. (forthcoming) and the procedure is described in detail
in Appendix A.
10
Note that within this set of studies are some estimates of the money value of direct welfare losses due to
climate change, e.g. impacts on health and the amenity value of the environment.
11
Which is why Nordhaus himself describes such functions and the assumptions they embody about
damage at different temperatures as ‘placeholders subject to further research’ (Stern, 2013). However, we
will see data points of 4, 5 or 6°C, if we are negligent and unlucky, within decades. Hence, it makes sense to
try different formulations as representing different possibilities, including of the extremely damaging
circumstances the science suggests as possible.
12
A quadratic function could not be made to simultaneously fit the existing data, while satisfying this
additional assumption; it would give excessive damage for smaller temperature increases.
13
Elsewhere Dietz et al. (2007a,b,c); Stern (2007, 2008) we investigated models based on the PAGE IAM, in
which damage was a power function of temperature. We examined the sensitivity of damage to the exponent
of the power function up to a value of three.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 583
3.3. Climate Risk
Our last extension to the basic framework involves the climate sensitivity parameter.
We take two approaches here. First, we explore high values of this parameter in
sensitivity analysis. Second, we replace its sure value with a probability density function
(pdf). Climate sensitivity is a key factor in driving the change in temperature in DICE,
as it is in many other simple climate models. Thus, it is a natural example of large-scale
risk. Others would be relevant too, such as the scale of damage for a given temperature
14
Where the density in the upper tail approaches zero more slowly than the exponential distribution.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
584 THE ECONOMIC JOURNAL [MARCH
It is worth emphasising, before moving on to the results, that there are other
potentially significant sources of risk attending to the impacts of greenhouse gas
emissions. Some of these are in the climate system – for instance the effective heat
capacity of the oceans (Calel et al., 2014) – yet a focus on S captures the essence of
physical climate risk in a clear and simple way. Other sources of risk relate to damage
for any given temperature and could also be modelled with probabilities, were the
evidence to justify doing so. However, as we have argued, the damage functional form
and parameterisation are currently very poorly constrained by evidence, and therefore
4. Results
4.1. Baseline
At the heart of this exercise is an investigation into the prospects for growth and damage
in a changing climate. Figure 1 plots baseline consumption per capita – i.e. in the
absence of controls on carbon dioxide emissions imposed by a social planner – under
various scenarios over the next two centuries. The upper panel plots the forecasts of the
model with production (1.K) and damage from climate change on the capital stock,
while the lower panel plots the forecasts of the model with production (1.TFP), where
TFP growth is endogenous and where climate change reduces TFP.
The ‘standard’ trajectory represents the forecast of the standard DICE model
without the various extensions we are considering in this article. The starting year is
2005. It is of course the same in both panels and notice immediately by how much
consumption per capita increases in it, powered largely by exogenous productivity
growth15 – in 2205 it is more than 15 times the 2005 level. This is despite a large
increase in the atmospheric stock of carbon dioxide and in the global mean
temperature (discussed below). Without large assumed improvements in the exoge-
nous element of TFP, the increase in per capita consumption would be much smaller.
Changing the model of growth begins to yield more pessimistic forecasts, although it
does not by itself qualitatively alter the tendency for the future to be much better off
than the present. Under the model with capital damage, consumption/head in 2205 is
13.3 times higher than in 2005, while under the model of productivity damage it is 11.4
times higher. Since the total damage multiplier Dt in (4) is the same in the two models,
simply being partitioned differently between damage on output, capital and TFP (see
Appendix A), the larger effect in the model of productivity damage partly reflects the
longer lasting impact of climate change in this model, where depreciation of
productivity is slow compared with capital.
The divergence in forecasts is much more marked, however, when we layer on
greater convexity of damage as in (40 ). With Weitzman’s (2012) calibration,
consumption per capita grows much more slowly after 2150 in the model of capital
15
With no growth in labour, the long-run output growth rate implied by (1) is simply that of exogenous
TFP.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 585
100 K Model
Standard
90 S = 3, Quardratic Damage
S = 3, Weitzman Damage
Consumption Per Capita (US $2005)
80 S = 3, High Damage
S = 6, Weitzman Damage
70
60
40
30
20
10
0
2000 2050 2100 2150 2200
90
Consumption Per Capita (US $2005)
80
70
60
50
40
30
20
10
0
2000 2050 2100 2150 2200
damage, while in the model of TFP damage it peaks around 2150 before actually falling
thereafter. By 2205, it is only 8.3 and 5.8 times higher respectively than today. If the
damage function is set such that damage equivalent to 50% of global output are
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
586 THE ECONOMIC JOURNAL [MARCH
assumed to occur upon 4°C warming, the collapse in living standards is much stronger,
with consumption/head peaking before the end of this century and ending up in both
models around or below the present level in real terms. A similar forecast is generated
by Weitzman damage, when we instead increase the climate sensitivity parameter S to
6°C, which has a probability, as described above, of up to 0.1 according to IPCC. The
two growth models yield similar forecasts in these cases, demonstrating the diminished
importance of growth assumptions when instantaneous damage are severe and or
warming is very rapid.
16
Where the optimal carbon price is defined as the marginal cost of abatement at the optimal emissions
level calculated. Whether it is reasonable to interpret this as a price depends on the convexity of the
abatement cost curve, i.e. it depends on there being rising marginal costs. It has been contended that
marginal costs do not rise but these are issues for another paper.
17
World Bank data on GDP deflator, from http://data.worldbank.org/indicator/ NY.GDP.DEFL.KD.ZG,
retreived on 22 November 2013.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 587
Table 1
Optimal Emissions Control Rate, 2015–2105. S = 3 Unless Otherwise Indicated
2015 2025 2035 2045 2055 2065 2075 2085 2095 2105
Standard 0.158 0.184 0.211 0.240 0.270 0.302 0.335 0.370 0.407 0.446
Capital models
Quadratic damage 0.213 0.289 0.356 0.424 0.495 0.565 0.636 0.706 0.777 0.848
Weitzman damage 0.235 0.322 0.401 0.484 0.568 0.650 0.730 0.805 0.875 0.944
Weitzman damage, S = 6 0.360 0.494 0.619 0.751 0.883 1 1 1 1 1
Table 2
Optimal Carbon Prices (2005 US$/tC), 2015–2105. S = 3 Unless Otherwise Indicated
2015 2025 2035 2045 2055 2065 2075 2085 2095 2105
Standard 44.4 57.0 71.2 87.8 106.2 127.1 150.0 175.8 204.6 236.6
Capital models
Quadratic damage 76 129 182 245 316 393 476 563 656 752
Weitzman damage 91 156 226 310 405 506 609 711 812 912
Weitzman damage, S = 6 196 337 495 684 895 1097 1074 1052 1032 1012
High damage 178 309 455 617 774 909 1017 1052 1032 1012
Productivity models
Quadratic damage 118 196 272 363 466 580 705 840 984 1012
Weitzman damage 133 222 313 420 541 670 806 945 1032 1012
Weitzman damage, S = 6 271 456 653 888 1121 1097 1074 1052 1032 1012
High damage 233 393 559 738 911 1066 1074 1052 1032 1012
growth to (1.K) with capital damage, the optimal emissions control rate rises to
0.213 (optimal carbon price = $76/tC). Further extending this model to
incorporate highly convex damage with Weitzman’s (2012) parameterisation, it
rises to 0.235 (optimal price = $91/tC), while with our high damage function
scenario it is 0.342 (optimal price = $178/tC). When Weitzman damage are
combined with a high climate sensitivity, the optimal control rate is 0.36, brought
about by an optimal price levied at $196/tC. Some caution should be exercised,
however, in interpreting the relevance of these strong initial control rates and
prices, because DICE, as a model of medium and long-run dynamics, lacks
adjustment costs, which could render such a rapid decarbonisation infeasible.
In the endogenous growth model (1.TFP) where instantaneous climate damage
work on TFP as well as output, the increase in the controls is even stronger. With
quadratic damage, the optimal control rate on emissions is 0.272 with an associated
carbon price of $118/tC. Moving to Weitzman damage increases this to 0.29
(optimal carbon price = $133/tC), while with our high damage function scenario
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
588 THE ECONOMIC JOURNAL [MARCH
the controls are respectively 0.396 and $233/tC. When Weitzman damage is
combined with a high climate sensitivity they are respectively 0.432 and $271/tC.
Notice for both growth models the marked rise in the carbon price when we
move from Weitzman to high damage or from S = 3 to S = 6, which reflects
convexity in the marginal abatement cost function. Nonetheless, the same remarks
regarding adjustment costs and their potential effect on the optimal controls apply
here.
Figures B3 and B4 in Appendix B show the consequences of the optimal controls
18
In line with much of the literature, we simplify the problem by omitting the possibility of learning about
the climate sensitivity from observations obtained after the first period has commenced. So the planner must
stick to optimal controls computed at the outset, a so-called open-loop control. Were it possible to learn
about climate sensitivity from observations and to change policy settings in response – a closed-loop policy –
the planner could of course achieve at least as high a level of social welfare, most probably much higher.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 589
Table 3
Optimal Emissions Control Rate Under Random S, 2015–2105
2015 2025 2035 2045 2055 2065 2075 2085 2095 2105
Standard 0.158 0.184 0.211 0.240 0.270 0.302 0.335 0.370 0.407 0.446
Capital models
Quadratic damage 0.204 0.277 0.341 0.406 0.474 0.543 0.611 0.681 0.750 0.819
Weitzman damage 0.250 0.344 0.434 0.532 0.631 0.732 0.831 0.925 0.993 0.993
High damage 0.393 0.542 0.688 0.841 0.986 0.998 0.998 0.998 0.999 0.999
Table 4
Optimal Carbon Prices (2005 US$/tC) Under Random S, 2015–2105
2015 2025 2035 2045 2055 2065 2075 2085 2095 2105
Capital models
Quadratic damage 70 119 169 226 293 365 442 528 614 707
Weitzman damage 101 176 261 368 490 625 769 914 1019 999
High damage 229 399 598 838 1093 1092 1071 1050 1030 1010
Productivity models
Quadratic damage 110 181 253 338 435 543 664 792 930 1012
Weitzman damage 147 248 359 494 657 839 1040 1050 1032 1012
High damage 329 563 830 1146 1121 1097 1074 1053 1032 1012
ultimately matters is the pdf of consumption per capita that results from f (S), it should
start to become clear that, when the damage function has modest curvature, the effect
of randomising S on the optimal controls can be to lower them, but when the damage
function has strong curvature the opposite is true, because the tail of high
temperatures exerts an ever larger relative effect on consumption per capita, utility
and social welfare.
Figures B5 and B6 in Appendix B show the consequences of the optimal controls
for the atmospheric stock of CO2 and global mean temperature respectively.19
Figure B5 shows that the optimal mean stock of atmospheric CO2 peaks in our
endogenous growth models at no more than about 500 ppm, and as little as 420
ppm, depending on the growth model and damage function. These stock levels are
well below those in the standard DICE model. Those combinations of growth model
19
Since the climate sensitivity is uncertain, so, obviously, is the change in the global mean temperature,
and since this goes on to affect emissions via damage, there is also some uncertainty in the longer run about
the atmospheric stock of CO2. Therefore, both figures report mean values from the Monte Carlo simulation.
In the case of the atmospheric stock of CO2, the uncertainty is very small (no more than 1 ppm), but in the
case of global mean temperature it is considerably larger. Therefore, in the latter case we also show the 90%
confidence interval, in 2205, from the Monte Carlo simulation to the right of the main chart.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
590 THE ECONOMIC JOURNAL [MARCH
and damage function yielding higher climate impacts support a lower optimal stock.
Compared with Figure B3 we can see that the optimal stock is lower under random
S than when S = 3. Figure B6 shows that mean temperature is kept to a maximum
of around 2°C except in two cases. First, in model (1.K) with capital damage, when
the damage function is quadratic, mean warming peaks at around 2.5°C early next
century. Second, in standard DICE mean warming peaks at c. 3.5°C. Notice the
spreads around mean warming and in particular the very large 90% confidence
interval around warming in standard DICE, where the 90th percentile reaches as
5. Conclusions
‘To slow or not to slow’ (Nordhaus, 1991) and its subsequent development into the
dynamic DICE model have given us what seems to be a coherent and powerful
framework for assessing the costs and benefits of climate-change mitigation. But it
has in-built assumptions on growth, damage and risk, which together result in gross
underassessment of the overall scale of the risks from unmanaged climate change
(Stern, 2013). This criticism applies with just as much force to most of the other
IAMs that DICE has inspired. The purpose of this article has been to show how
these unrealistic assumptions might be relaxed and what would be the conse-
quences of doing so, in terms of optimal emissions reductions and carbon prices,
atmospheric concentrations of carbon dioxide and global mean temperature.
The first assumption we have relaxed is that the underlying drivers of economic
growth are exogenous and unaffected by climate change. Instead we look at two
models of endogenous growth, in which the damage from climate change affect the
drivers of long-run growth, not just current output. The second assumption we have
relaxed is that the damage function relating instantaneous climate damage to the
increase in global mean temperature is only weakly convex. Instead, we allow for
the possibility that instantaneous damage increase rapidly, particularly once the
global mean temperature reaches 4–6°C above the pre-industrial level. We
suggest this representation is more plausible, given the scale of change that such
warming could bring; at the very least, simulations based on weak convexity should
not dominate our attention as they have come to do. The third assumption we have
relaxed is that the climatic response to greenhouse gas emissions is moderate and
moreover is precisely understood. Very few, if any, commentators would
explicitly claim that climate sensitivity is precisely understood, of course.
Nonetheless, most economic modelling is undertaken using only a single, central
estimate of the climate sensitivity parameter, fixed in the centre of the distribu-
tion of available estimates from the science. We explore risk in this crucial
parameter.
Overall, the scale of the risks from unmanaged climate change in this modelling
framework is the convolution of these three extensions. We show that, with the
models extended in this way, business-as-usual trajectories of greenhouse gas
emissions give rise to potentially large impacts on growth and prosperity in the
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 591
future, especially after 2100. Indeed, these impacts are large enough to feed back
into future emissions via reduced activity but the feedback is too small and too late
for the system to self-regulate. Thus, optimal emissions control is strong and
strongly increasing. As a guide, we find that these models suggest the carbon price
in a setting of globally coordinated policy, such as a cap-and-trade regime or a
system of harmonised domestic carbon taxes, should be in the range $32–103/tCO2
(2012 prices) in 2015. It must be remembered that the DICE model lacks
adjustment costs, so the high end of the range should be interpreted cautiously. On
X
Tmax
max W ¼ uðct ÞLt ð1 þ qÞt ;
flt gTt¼1
max
t¼0
where l 2 [0, 1] is the emissions control rate, u(ct) is the instantaneous social utility of
consumption per capita at time t and q = 0.015 is the utility discount rate. Note that c is not only
time-dependent as the above equation implies, it is also state-dependent when we undertake
stochastic modelling. We suppress notation of state-dependence for simplicity; bear in mind that,
when running the model with a random parameter (the climate sensitivity S in (3)), we take the
expectation of social welfare. Tmax is the terminal period, which is 2595. The model proceeds in
time steps of ten years from 2005, so appropriate interpretations must be made in considering
the various equations of motion. Notice that, since 2005 is in the past, our first control period is
t = 1, i.e. 2015.21
The utility function is iso-elastic,
ct1g
uðct Þ ¼ ;
1g
where g is the elasticity of marginal social utility of consumption and is set to 1.5 to allow
comparison with standard DICE.
As set out in the main body of the article, we explore two alternative production functions:
20
For example, the paradoxes of social choice theory can be better understood by broadening the
philosophical perspective and are not easily resolved within the standard framework (Sen, 2009).
21
In fact, we need only solve lt from 2015 to 2245 inclusive, since DICE assumes that from 2255 onwards
lt = 1, because a zero-emissions backstop energy technology becomes competitive.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
2015] GROWTH, CONVEX DAMAGES AND CLIMATE RISK 593
The capital elasticity a = 0.3, while the elasticity of output with respect to knowledge, b = 0.3
(Mankiw et al., 1995).
The equation of motion of capital in model (1.K) is
where Ct = ctLt is aggregate consumption and It = sYt, where s = 0.23 is the savings rate
(calibrated to long-run average optimal savings in standard DICE, without climate damage and
emissions abatement costs and, in principle, some private inter-temporal objectives). We specify
exogenous, constant savings in order to capture in a simple way the second-best context implied
by fitting our models of endogenous growth to current macroeconomic data. In growth models
with knowledge spillovers, the savings rate chosen by a planner will be greater than the savings
rate emerging from a decentralised equilibrium of firms and households, because the marginal
private return to investment does not include the spillovers.
A more elaborate analysis would permit households to choose their optimal savings rate in
equilibrium with firms’ private marginal product of capital (in response to the planner’s
emissions controls), but it is worth noting that, in standard DICE, endogenising the savings rate
has been shown to make little difference to the optimal policy (Fankhauser and Tol, 2005; and
Nordhaus’s laboratory notes on DICE), so our simplification is unlikely to matter.22 In any case,
whether households are currently taking into account the effects of climate policy on future
consumption prospects when choosing how much to save is unclear.
In model (1.TFP), productivity is endogenous and its equation of motion is
where dA = 0.01 is the rate of depreciation of the stock of TFP, while c1 0.0003 and
c2 0:373 are parameters of the spillovers function. c1 and c2 are calibrated so that output
in (1.TFP), in the absence of climate damage and emissions abatement costs, is the
same as in standard DICE. In model (1.K), TFP is an exogenous time series, so (3) does not
apply.
The climate damage function is
where p1 = 0 and p2 0.00284 throughout. p3 = 0 when we compute results for the standard
setting (i.e. Dt ¼ D^t ), 5.07 9 106 when we use Weitzman’s parameterisation, or
8.199105 according to our high damage specification, where Dt is assumed to be equal to
0.5 when the atmospheric temperature is 4°C above the pre-industrial level.
Damage is then partitioned between output and capital, or output and TFP, depending on the
growth model:
Dti ¼ f i Dt ;
ð1 Dt Þ
DtY ¼ 1
ð1 Dti Þ
22
See also Mirrlees and Stern (1972), who first illustrated this feature in simple optimal growth models.
© 2015 The Authors.
The Economic Journal published by John Wiley & Sons Ltd on behalf of Royal Economic Society.
594 THE ECONOMIC JOURNAL [MARCH
where f is the share of damage to i = A or i = K. IAMs do not in general explicitly address the
allocation of damage between capital and output, and vary widely in what they implicitly assume
about it. Nordhaus and Boyer (2000) analysis might be read to suggest that f K is in the region of
1/3, so 0.3 is the value we choose. The calibration problem is even more acute in the case of
allocating damage between output and TFP – there are, as Moyer et al. (forthcoming) also point
out, currently severe modelling, data and estimation problems in carrying out such an allocation.
Moyer et al. (forthcoming) consequently explore a range of values of f A between 1% and 100%.
We make the relatively conservative assumption that f A = 0.05.
The total abatement cost function is
where h1,t is a time-varying coefficient and h2 ¼ 2:8, hence marginal abatement costs are
increasing in emissions control.
Cumulative industrial carbon dioxide emissions are constrained by remaining fossil fuel
reserves,
X
Tmax
EIND
t C Cum;
t¼0
where C Cum = 6000 gigatonnes of carbon is the constraint, and total emissions of carbon are
the sum of industrial emissions of carbon dioxide and exogenous emissions of carbon dioxide
from land use:
Et ¼ EIND
t þ ELAND
t :
Industrial carbon dioxide emissions at time t are proportional to gross output in the same
period, hence there is a different function depending on the growth model:
EIND
t ¼ rt ð1 lt ÞAt Ktaþb Lt1a ; or
¼ rt ð1 lt ÞAt K a L 1a ;
t t
Cycling is determined by a set of coefficients /jk that govern the rate of transport from
reservoir j to k per unit of time.
The change in the atmospheric stock of carbon from the pre-industrial level determines
radiative forcing,
M AT
Ft ¼ F2CO2 log2 t þ FtEX ;
Md AT
1750) and FtEX is exogenous radiative forcing (capturing among other things the forcing due to
greenhouse gases other than carbon dioxide) and is time-dependent. The equation of motion of
temperature is given by:
F2CO2
Tt ¼ Tt1 þ j1 Ft ðTt1 Þ j2 Tt1 Tt1
LO
; (5)
S
where T LO is the temperature of the lower oceans and evolves according to:
1200
Atmospheric Concentration of Carbon Dioxide (ppm)
1000
800
Standard
600 TFP Model; Weitzmen Damage
TFP Model; High Damage
Capital Model; Weitzman Damage
400 Capital Model; High Damage
200
0
2000 2050 2100 2150 2200
5.0
3.0
2.0
1.0
0.0
2000 2050 2100 2150 2200
6.0
5.0
Above Pre-industrial)
4.0
3.0
2.0
1.0
0.0
2000 2050 2100 2150 2200
Fig. B2. Baseline Global Mean Temperature (Degrees Centigrade Above Pre-industrial), 2005–2205
Notes. The upper panel corresponds with the model of capital damage, while the lower panel
corresponds with the model of TFP damage. The damage function calibration is ‘Weitzman’
unless otherwise indicated.
600
500
300
TFP Model; Weitzman Damage
200 TFP Model; High Damage
Capital Model; Weitzman Damage
Capital Model; High Damage
100 Standard
0
2000 2050 2100 2150 2200
2.5
1.5
1.0
0.5
0.0
2000 2050 2100 2150 2200
3.5
3.0
Above Pre-industrial)
2.5
2.0
1.5
1.0
0.5
0.0
2000 2050 2100 2150 2200
Fig. B4. Optimal Global Mean Temperature (Degrees Centigrade Above Pre-industrial), 2005–2205
Notes. The upper panel corresponds with the model of capital damage, while the lower panel
corresponds with the model of TFP damage. The damage function calibration is ‘Weitzman’
unless otherwise indicated.
600
500
0
2000 2050 2100 2150 2200
Fig. B5. Optimal Atmospheric Stock of Carbon Dioxide, 2005–2205, Mean Over Random S
0
2000 2050 2100 2150 2200
Atmospheric Temperature (Degrees Above Pre-industrial)
6
TFP Model
0
2000 2050 2100 2150 2200
Fig. B6. Optimal Global Mean Temperature (Degrees Centigrade Above Pre-industrial), 2005–2205, Mean
Over Random S
Notes. The upper panel corresponds with the model of capital damage, while the lower panel
corresponds with the model of TFP damage. The bars on the right-hand side give the 90%
confidence interval in 2205.
References
Agrawala, S. (1998). ‘Context and early origins of the intergovernmental panel on climate change’, Climatic
Change, vol. 39(4), pp. 605–20.
Arrow, K.J. (1962). ‘The economic implications of learning by doing’, Review of Economic Studies, vol. 29(3),
pp. 155–73.
Calel, R., Stainforth, D.A. and Dietz, S. (2014). ‘Tall tales and fat tails: the science and economics of extreme
Appendix C. Nordhaus, W.D. (1991). ‘To slow or not to slow: the economics of the
greenhouse effect’, ECONOMIC JOURNAL, vol. 101(407), pp. 920–37.