Environ Resource Econ

DOI 10.1007/s10640-007-9130-7
The stability likelihood of an international climate
agreement
Rob Dellink · Michael Finus · Niels Olieman
Received:15 February 2006/Accepted: 19 April 2007
© Springer Science+Business Media B.V. 2007
Abstract Results derived from empirical analyses on the stability of climate coalitions are
usually very sensitive to the large uncertainties associated with the benefits and costs of cli-
mate policies. This paper provides the methodology of Stability Likelihood (SL) that links
uncertainties about benefits and costs of climate change to the stability of coalitions. We show
that the concept of SL improves upon the robustness and interpretation of stability analyses.
Moreover, our numerical application qualifies conclusions from a recent strand of literature
based on stylised models with ex-ante symmetric players that learning has a negative impact
on the success of coalition formation in context of uncertainty.
Keywords Climate change modelling · International environmental agreements ·
Non-cooperative game theory · Uncertainty
JEL Classification C79 · H87 · Q54
R. Dellink (B)
Environmental Economics and Natural Resources Group, Wageningen University, Hollandseweg 1,
Wageningen 6706 KN, The Netherlands
e-mail: rob.dellink@wur.nl
R. Dellink
Institute for Environmental Studies, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands
M. Finus
Institute of Economic Theory, Department of Economics, University of Hagen, Hagen, The Netherlands
N. Olieman
Operations Research and Logistics Group, Wageningen University, Wageningen, The Netherlands
N. Olieman
Risk Model Validation and Methodologies Department, Rabobank Netherlands, Rabobank,
The Netherlands
123
R. Dellink et al.
1 Introduction
There are many obstacles that prevent the formation of effective international environmental
agreements (IEAs). Two of the most important obstacles are free-riding and uncertainty.
First, IEAs aim at providing a public good and therefore countries face free-rider incen-
tives. Typically it is beneficial for a country when other regions increase their abatement
efforts. This free-riding problem has been studied extensively in theoretical models, as for
instance in Hoel (1992), Carraro and Siniscalco (1993) and Barrett (1994) and in empirical
climate models, as for instance in Bosello et al. (2003), Botteon and Carraro (1997), Tol
(2001), Eyckmans and Finus (2006), and Finus et al. (2006). A main conclusion of this lit-
erature is that stable coalitions improve only marginally upon the non-cooperative outcome
because either stable coalitions are small or the global abatement level that can be sustained
in larger coalitions is very low. However, regardless of the specific model, all of these papers
determine stable coalitions in a deterministic setting.
Second, the uncertainties surrounding the future impacts of climate change as well as
abatement costs are large. In particular, the benefits from abatement that arise in the distant
future are highly uncertain, but also future abatement costs are not fully known. These issues
are highlighted for instance in Roughgarden and Schneider (1999) and Tol (2002, 2005).
Depending on the model specification and assumptions, uncertainty and learning may either
call for laxer environmental standards today or imply that countries follow the precaution-
ary principle by increasing their short-term abatement efforts (Kolstad 1996a, b; Ulph and
Maddison 1997; Ulph and Ulph 1997). However, regardless of how uncertainty and learning
is captured in these papers and what this means for optimal abatement strategies, success
can only be evaluated if free-rider incentives are also included in the analysis. This implies
that models with a single global decision-maker cannot provide full insight into the regional
impacts of uncertainty and learning. This drawback is common to most integrated assess-
ment models that are used to analyse optimal abatement strategies, including for instance
DICE/RICE (Nordhaus and Boyer 2000), MERGE (Manne et al. 1995) and EPPA (Babiker
et al. 2001). This also applies to some extent to models with two countries that compare
non-cooperative with cooperative behaviour and hence coalition formation is not really an
issue in these models (e.g. Ulph and Maddison 1997; Ulph and Ulph 1997).
Recently, some papers have appeared that combine coalition formation with uncertainty
and learning (Na and Shin 1998; Ulph 2004; Kolstad and Ulph 2006; Kolstad 2007). They
assume uncertainty about parameter values of the payoff function where probability distri-
bution functions are known to all countries. In the context of coalition formation, in which
countries choose their membership in the first stage and their abatement strategies in the
second stage, this allows to distinguish three cases according to Kolstad and Ulph (2006)
and Kolstad (2007): (1) Uncertainty is resolved before the first stage. This corresponds to the
case of full learning. (2) Uncertainty is resolved after the first stage but before the second
stage. This corresponds to the case of partial learning. (3) Uncertainty is not resolved. This
is the case with no learning. In cases 1 and 2 where there is learning, it takes the form of
perfect learning, i.e. all players learn the true values of all uncertain parameters.
All these papers are stylized models with ex-ante identical players and mostly come to
the conclusion that learning has a negative impact on the success of coalition formation. Na
and Shin (1998) compare case 2 and 3, but their model is essentially static (flow pollutant)
and restricted to only three countries. The first restriction is removed in Ulph (2004), who
compares all three cases and considers a two period model with a stock pollutant. Due to
this complication, results are based on simulations and conclusions are not always clear-cut.
Moreover, he assumes that countries have only two abatement strategies (no abatement versus
123
The stability likelihood of an international climate agreement
full abatement), the only uncertainty concerns the benefits from global abatement, with only
two states (low and high benefits) and uncertainty is correlated between players. This means
that in the case of learning all countries are also ex-post identical; either all countries are
low or high benefit countries. These assumptions are also made in Kolstad (2007). Because
of the assumption of a flow pollutant, he unambiguously shows that when comparing case 1
and 3, learning increases the size of a stable IEA but has a negative impact on global welfare,
as proven in Kolstad and Ulph (2006). Case 2 is ambiguous; learning may have a positive or
negative impact on participation and welfare. Finally, in the case of uncorrelated uncertainty,
Kolstad and Ulph (2006) confirm the negative impact of learning, but again have to resort to
simulation, despite the assumption that all countries are ex-ante identical and that there are
only two states in the world.
Because of the difficulties of deriving analytical results even in a simple model, we con-
sider uncertainty, learning and coalition formation based on an empirical climate model with
twelve heterogeneous world regions. In the first stage, regions can only choose between
becoming a signatory or to remain an outsider as assumed in the papers above. However, in
the second stage, regions can choose from a continuum of abatement strategies, while almost
all model parameters are assumed to be uncertain with (uncorrelated) probability distribution
functions. We restrict ourselves to the analysis of cases 1 and 3, namely no and full resolution
of uncertainty. Case 1 is relatively straightforward. Because we assume that utility is linear
in payoffs as in Kolstad and Ulph (2006), we can also apply the certainty equivalence prin-
ciple. That is, the case of no resolution of uncertainty is identical to the deterministic case
if the mean or expected value of a parameter corresponds to the value in the deterministic
case. Together with case 3, which is the focus of this analysis, case 1 allows for a direct
comparison with the conclusions derived in Kolstad and Ulph (2006). Case 3 implies that
a coalition cannot be called stable or not stable any longer (based on certain payoffs in the
deterministic case or expected payoffs in the case of no resolution of uncertainty). For this,
we introduce our concept of Stability Likelihood (SL), which determines the likelihood that
a coalition structure is stable. Compared to Kolstad and Ulph (2006) and the other papers
above, we find that learning does not always have a negative impact on the success of coalition
formation. Moreover, we show that the concept of SL has advantages over a deterministic
analysis. Firstly, in a deterministic setting, coalitions can only be stable or unstable. This may
imply in some cases that small perturbations of parameters are sufficient to change the set of
stable coalitions, invalidating conclusions. Secondly, in a deterministic model, uncertainty
of parameter values can only be indirectly accounted for through sensitivity analyses where
each analysis is subject to the first problem mentioned above. Hence, the concept of SL can
be interpreted as a more sophisticated sensitivity analysis than is traditionally done in the
context of coalition formation with empirical data, as for instance in Eyckmans and Finus
(2006) and Finus et al. (2006).
In the following, we present our model of coalition formation and introduce the concept
of SL in Sect. 2. Section 3 discusses the calibration of the model, including the distribution
functions used for the uncertain model parameters. Section 4 presents and discusses result
of our stability analysis. Section 5 concludes.
2 The model of coalition formation
Consider a set of N heterogeneous players, each representing a region of the world. In the
first stage, regions decide whether to become a member of IEA or to remain an outsider.
Announcement c
i
= 1 means “Region i joins the coalition” and announcement c
i
= 0
123
R. Dellink et al.
“Region i remains an outsider”; a coalition structure c is described by announcement vector
c = (c
1
, . . . , c
N
). The set of players that announce 1 are coalition members and is denoted as
k = {i | c
i
= 1, ∀i = 1, . . . , N}. Thus, in this simple setting, a coalition structure is defined
by coalition k. Hence, we can use the term coalition structure and coalition interchangeably.
Note that there are 2
N
announcement vectors, but essentially there are only 2
N
−N coalition
structures. The reason is that regardless whether all regions announce c
i
= 0 or only one
region announces c
i
= 1 and all other announce c
i
= 0, this will result in the “All-Singletons
Coalition Structure”.
In the second stage, regions choose their abatement levels. This leads to abatement vector
q = (q
1
, . . . , q
N
). The payoff of an individual region i, π
i
(q, b) depends on abatement vec-
tor q (and hence on the strategy of all regions, due to the public good nature of abatement)
and on a vector of model parameters b.
We solve the game backwards assuming that strategies in each stage must form a Nash
equilibrium. For the second stage, this entails that abatement strategies form a Nash equilib-
rium between coalition k and the remaining non-signatories. That is,

i ∈k
π
i
(q

k
, q

−k
, b) ≥

i ∈k
π
i
(q
k
, q

−k
, b)∀q
k
and
π
i
(q

k
, q

i
, q

−i
, b) ≥ π
i
(q

k
, q
i
, q

−i
, b)∀q
i
(1)
where q
k
is the abatement strategy vector of coalition k, q
−k
the vector of all regions not
belonging to k, q
i
the strategy of non-signatory i , and q
−i
the strategy vector of all other
non-signatories except i . An asterisk denotes equilibrium strategies. Computationally, this
implies that non-signatories (i / ∈ k) that announced c
i
= 0 will choose their abatement strat-
egies so as to maximize their individual payoff π
i
(q, b), whereas all signatories (i ∈ k) that
announced c
i
= 1 jointly maximize the aggregate payoff of their coalition

i ∈k
π
i
(q, b),
taking the abatement strategies of all other regions as given. Strategically, this means that the
behaviour of non-signatories towards all other regions is selfish and non-cooperative; signa-
tories’ behaviour is cooperative towards their fellow members, but non-cooperative towards
outsiders. Economically, this means strategies are group (but not globally) efficient within
coalition k. Hence, the equilibrium economic strategy vector q

corresponds to the classical
“social or global optimum” if coalition k comprises all countries, i.e. the grand coalition
forms, and corresponds to the classical “Nash equilibrium” if coalition k comprises only
one member or is empty. Thus, any inefficiency stems from the fact that k is not the grand
coalition.
Since in the context of our empirical model the equilibrium abatement vector q

is unique
for every coalition structure c and a given vector of parameters b (see the proof in Olieman and
Hendrix 2006), there is a unique vector of equilibrium payoffs for every coalition structure c:
v
i
(c, b) ≡ π
i
(q

(c), b).
We now turn to the first stage. Also in the first stage, strategies form a Nash equilibrium.
That is, no signatory that announced c
i
= 1 should have an incentive to change its announce-
ment to ˜ c
i
= 0 (internal stability) and no non-signatory that announced c
i
= 0 should want
to announce ˜ c
i
= 1 (external stability), given the announcement of other players c
−i
. For
our purposes, this condition can be summarised compactly by the payoff stability function
f (c, b), which assigns the value 1 to a stable announcement vector (i.e. stable coalition) and
the value 0 to an unstable announcement vector (i.e. unstable coalition):
123
The stability likelihood of an international climate agreement
f (c, b) =



1 if (v
i
(c
i
, c
−i
, b) − v
i
( ˜ c
i
, c
−i
, b) ≥ 0) for ˜ c
i
= 1 − c
i
and for all
i = 1, . . . , N
0 else
(2)
The function f (c, b) will take on the value 1 if for all regions the payoff from announcement
c is higher than, or equal to, the payoff associated with alternative announcements ˜ c. These
alternatives are constructed by changing the announcement of one player at a time. It is worth
noting that for any given set of parameters b, this function allows for either a unique stable
coalition, multiple stable coalitions or no non-trivial stable coalition.
In a deterministic model, the vector of parameters b may be based on estimates but is
treated as given. In contrast, in a stochastic model the deterministic vector of parameters b
is replaced by the stochastic (vector) variable b which is characterised by probability space
(B, B, Pr) and probability density function g (b). In the case of no resolution of uncertainty,
the decision on membership in the first stage will be based on expected payoffs using the
mean values of b. If the vector of mean values of b is b, this corresponds to the deterministic
setting. However, in the case of full resolution, decisions are based on the stochastic variable
b and as function f (c, b) becomes a Bernoulli variable, we can assign a likelihood to the
event f (c, b) = 1, i.e. to a stable coalition.
The SL of coalition structure c is therefore defined as SL ≡ Pr { f (c, b) = 1}, which
equals
_
B
f (c, b) · g(b)db. Assuming that f (.) is a correct representation of reality, we can
interpret SL as the probability that coalition structure c is stable.
The heterogeneity of players with payoff functions based on the empirical model STACO
(see Sect. 3), requires us to resort to numerical calculations for the SL of various coalitions.
1
The SL is calculated using a Monte Carlo sampling technique: we generate M samples b
m
from b; based on these samples we can estimate the SL with
¨
SL =
1
M

M
m=1
f (c, b
m
) that
has an estimated variance of
1
M−1
_
¨
SL
_
1 −
¨
SL
__
.
A more detailed discussion of the SL concept and computation methods can be found in
Olieman and Hendrix (2006).
3 Calibration of the model
3.1 Introduction
In this section, the calibration of the empirical model, called STAbility of COalitions (STA-
CO) is described. For more detailed information on the model and calibration procedure, see
Dellink et al. (2004) and Finus et al. (2006). The model comprises benefit and cost functions
of twelve world regions: USA (USA), Japan (JPN), European Union (EEC), other OECD
countries (OOE), Eastern European countries (EET), former Soviet Union (FSU), energy
exporting countries (EEX), China (CHN), India (IND), dynamic Asian economies (DAE),
Brazil (BRA) and “rest of the world” (ROW). The philosophy behind the construction of the
model comprises two items. First, the model must be simple enough to allow for sufficient
samples to be calculated within reasonable time.
2
Second, the model should reflect important
results and features of integrated assessment models in terms of overall magnitudes of global
1
We are not aware of any paper that provides analytical solutions of stable coalitions in the context of
heterogeneous players even in the absence of uncertainty.
2
We need around 40,000 samples to get sufficient accuracy in the estimates of SL. This implies around 2
billion calculations of optimal regional abatement and payoff levels (40,000 times 12 regions times 4096
membership announcement vectors).
123
R. Dellink et al.
emissions and concentration, abatement costs from regional abatement and benefits from
global abatement over some time period. Therefore, the model focuses on carbon dioxide,
but the exogenous level of other greenhouse gases is included in the calibration of the benefit
function (cf. Nordhaus 1994). The analysis is based on the net present value of a stream of
net benefits starting in 2010 and covering a time period of 100years in order to capture the
long-run effects of global warming. They are computed for every possible coalition structure
and then stability is computed according to the procedure explained in Sect. 2. The price
for this complexity (see footnote 2) is that we assume constant abatement strategies (cf.
Sect. 3.2), rendering the model essentially static as in Na and Shin (1998), Kolstad and Ulph
(2006) and Kolstad (2007). Therefore, numerical results should be interpreted with caution.
However, this allows for a direct comparision with the papers mentioned above about the role
of learning. Moreover, we believe that our simulations illustrate the usefulness of SL as an
indicator, as we include the best available information on the probability density functions
of the model parameters.
3.2 Elements of the empirical model
The payoff function of region i is given by:
π
i
=
T

t =1
(1 +r)
−t
(B
i t
(q
t
) − AC
i t
(q
i t
)) (3a)
where T denotes the time horizon, t = 2011, . . . , 2110; r is the discount rate; B
it
are ben-
efits from global abatement (q
t
=

N
i =1
q
it
); and AC
it
are abatement costs from regional
abatement q
it
.
The payoff function is calculated as the net present value of a stream of net benefits from
abatement between 2010 and 2110 and thus reflects discounted avoided damages minus dis-
counted abatement costs. Regional abatement costs are a function of the level of individual
abatement by region i , while regional benefits depend on global abatement efforts.
Assuming constant abatement strategies (q
it
= q
i
/100), benefits in year t can be expressed
as a function of global abatement over the entire period. Furthermore, we consider that the
stock of CO
2
can be approximated by a linear function of emissions and that damages are
linear in the stock of CO
2
.
3
It follows that annual global benefits from reduced emissions are
also linear in the level of abatement:
B
t
(q) = ϕ
t
· q (4a)
where ϕ
t
denotes marginal benefits in period t from total abatement over the entire period.
This parameter also includes the effects of limited retention of GHGs in the atmosphere and
the decay of the stock of GHGs over time. As benefits in period t result from the stream of
abatement in all periods until t , these marginal benefits are increasing over time.
Discounted global or total benefits, T B(q) can then be expressed as
T B(q) = γ · q (4b)
where γ represents discounted marginal benefits in $/ton carbon equivalents (tC) and is
calculated as
γ =
2110

t =2011
(1 +r)
−t
· ϕ
t
. (5)
3
The details of this approximation are given in Dellink et al. (2004).
123
The stability likelihood of an international climate agreement
Regional discounted benefits are assumed to be a share of discounted global benefits from
abatement:
T B
i
(q) = s
i
· T B(q) = s
i
· γ · q (4c)
where s
i
is the share of global benefits for region i and

N
i =1
s
i
= 1.
For the specification of the abatement cost functions, estimates of the EPPA model are
used (Ellerman and Decaux 1998). It assumes an annual abatement cost function of region i
of the following form:
AC
i t
(q
i t
) =
1
3
· α
i
· (q
i t
)
3
+
1
2
· β
i
· (q
i t
)
2
(6a)
with α
i
and β
i
regional cost parameters. Total abatement cost of region i, T AC
i
(q
i
), is the
sum of annual abatement cost over the entire time horizon. Assuming constant strategies and
a constant discount rate, as in the case of benefits, the total abatement cost function becomes
T AC
i
(q
i
) = ρ ·
_
1
3
· α
i
· (q
i
)
3
+
1
2
· β
i
· (q
i
)
2
_
(6b)
with
ρ =
2110

t =2011
(1 +r)
−t
. (7)
Taken together, the payoff function of region i can be written as:
4
π
i
= T B
i
(q) − T AC
i
(q
i
). (3b)
Generally, all model parameters are uncertain. However, we have insufficient information
to fully estimate the probability density functions of all parameters at a regional level. A
more advanced meta-analysis of published and unpublished estimates would require a study
of its own and therefore is beyond the scope of the current paper. Hence, we consider only
uncertainty of the benefit and cost parameters γ, s
i
, α
i
and β
i
. Other parameters, such as
emissions in 2010, the decay rate of greenhouse gases and the discount rate are assumed to
be known with certainty. Calibration of emissions and concentrations is based on the widely
known DICE-model by Nordhaus (1994).
The discount rate r is set at 2% per year. This is in line with Weitzman (1998, 2001) who
suggests that in the case of climate change long-terms effects should be discounted with a
low discount rate and suggests that 2% is appropriate for the “medium future”. In Sect. 4.3,
we also consider the possibility of a higher discount rate, though this done indirectly by low-
ering the mean value of the global benefit parameter γ . We want to stress that the discount
rate only affects the aggregation of costs and benefits over time, as marginal benefits are
4
This means that we base our stability analysis on the net present value of benefits and abatement costs (i.e.
discounted payoffs). The assumption that the abatement paths are constant over time simplifies the calcula-
tions, but is not essential. The nature of our analysis requires only that the overall magnitude of concentrations
and emissions mimics that of more elaborate integrated assessment models (because decisions whether to
join a coalition are based on discounted payoffs). The exact shape of the equilibrium emission or abatement
paths are not important for our analysis. A similar stability analysis in a deterministic setting with dynamic
equilibriumabatement paths, as reported in Nagashima et al. (2006), reveals that the impact of our simplifying
assumption is limited: while levels of payoffs are somewhat affected, the stability of coalitions is robust in
this respect.
123
R. Dellink et al.
Table 1 Characteristics of the
2-sided exponential distribution
function of the global benefit
parameter γ
Value
5% Density −9$/tC
Mode 5 $/tC
Density at mode 13%
95% density 245 $/tC
Mean 77 $/tC
constant (though uncertain). Note also that our assumption of a constant (money) discount
rate is consistent with a low, but positive, pure rate of time preference.
5
3.3 Global benefit function
The distribution function of the global benefit parameter γ is based on a recent study by
Tol (2005). The probability density function of peer-reviewed studies as presented by Tol
can be closely mimicked by a 2-sided exponential function. This function is described by
four pieces of information: (i) the 5 percentile point; (ii) the value where the two sides of
the exponential function are separated (assuming this is above the 5 percentile point); (iii)
the cumulative probability density at this separation value; and (iv) the 95 percentile point
(assuming this is above the separation value). The first piece of information is given by the
value that corresponds to a cumulative probability of 5%; this value equals −9 $/tC, implying
a strictly positive probability that benefits from abatement are negative. The point of separa-
tion between both sides of the function is given by the mode and equals 5 $/tC; the associated
cumulative probability density is 13%. The point on the right side of the function is given by
the 95% cumulative probability density and equals 245 $/tC. The mean value of this function
is 77 $/tC. These numbers are summarised in Table 1 and the corresponding histogram of
all drawn samples is shown in Appendix I. This distribution function implies that there is
a probability of around 9% that the benefits from global abatement are zero or negative. In
this case, optimal abatement is zero, and regions are indifferent between no, partial and full
cooperation.
3.4 Regional benefit shares
The distribution function of regional benefit shares (s
i
) is based on insights from a study by
Tol (2002). However, this source does not provide sufficient information on the functional
form of the distribution function of these shares. Therefore, we opt for a gamma distribution
that can handle the lower bound of the shares such that no region receives a negative share
of the benefits. Moreover, as the shares can vary between zero and one and the mean value
is typically close to zero, the distribution function should be skewed, as this is the case for a
gamma distribution. However, the choice of the gamma distribution function will be subject
to a sensitivity analysis in Sect. 4.4.
The mean values used for the regional shares are the values reported in the deterministic
setting of the STACO model, which are in turn based on Fankhauser (1995) and Tol (1997).
Standard deviations are based on expert judgement, using in particular Tol (2002). Typically,
5
Alternatively, we could assume a constant pure rate of time preference of 0.5% and use the Ramsey rule
to derive a declining discount rate; as only net present values matter in our model, these two alternatives are
equivalent for our analysis.
123
The stability likelihood of an international climate agreement
Table 2 Characteristics of the
gamma distribution function of
the regional parameter s
i
(share
of global benefits)
Region Lower bound Mean Standard deviation
USA 0 0.2263 0.1414
JPN 0 0.1725 0.1078
EEC 0 0.2360 0.1475
OOE 0 0.0345 0.0216
EET 0 0.0130 0.0130
FSU 0 0.0675 0.0675
EEX 0 0.0300 0.0300
CHN 0 0.0620 0.0620
IND 0 0.0500 0.1000
DAE 0 0.0249 0.0498
BRA 0 0.0153 0.0306
ROW 0 0.0680 0.1360
standard deviations are of the same order of magnitude as mean values, and are lower for
OECD countries than for the other regions. The regional numbers are represented in Table 2.
For a better understanding, the corresponding histograms for the USA and China are shown
in Appendix I.
After the samples of the regional shares are drawn, all shares are scaled up or down to
force the sum of shares to unity. This implies that the regional shares do not strictly follow
the gamma distribution and the resulting variance in the shares is smaller than in the original
gamma distribution.
6
Though rescaling is not a necessary condition for the working of the
model, the restriction on the sum of shares is enforced to avoid an impact of variations in
this parameter on global benefits. The consequences of this assumption will be investigated
in a sensitivity analysis in Sect. 4.4
3.5 Abatement cost functions
The two regional abatement cost parameters (α
i
and β
i
) cannot be assumed to be indepen-
dent. Typically, empirical studies report only variations in marginal abatement costs without
providing information on how the slopes and curvatures of these functions vary. Therefore,
we assume the functional form of the abatement cost curves to be given, and vary the level
of marginal abatement costs per region in the simulations. This implies that both parameters
in the abatement cost functions move simultaneously and in the same direction.
7
We assume a normal distribution function; the mean is based on the deterministic version
of the STACO model, which is in turn based on Ellerman and Decaux (1998). The normal
distribution is chosen, as a priori, there seems to be no reason to assume a skewed distribution.
The standard deviation of the abatement cost parameters are based on information provided
by IPCC (Metz et al. 2001). Due to lack of regional information, the standard deviation is
6
More information on the impact of this rescaling on the probability density function can be obtained from
the authors upon request.
7
Effectively, we replace equation (6b) by T AC
i
(q
i
) = ψ
i
· ρ ·
_
1
3
· α
i
· (q
i
)
3
+
1
2
· β
i
· (q
i
)
2
_
, where ψ
i
is stochastic with mean value 1 and α
i
and β
i
are deterministic. However, we interpret the model as if the
variation occurs in α
i
and β
i
.
123
R. Dellink et al.
Table 3 Characteristics of the normal distribution of the abatement cost parameters α
i
and β
i
Region α
i
β
i
Mean Standard deviation Mean Standard deviation
USA 0.00050 0.00006 0.00398 0.00050
JPN 0.01550 0.00194 0.18160 0.02270
EEC 0.00240 0.00030 0.01503 0.00188
OOE 0.00830 0.00104 0.00000 0.00000
EET 0.00790 0.00198 0.00486 0.00122
FSU 0.00230 0.00058 0.00042 0.00011
EEX 0.00320 0.00080 0.03029 0.00757
CHN 0.00007 0.00002 0.00239 0.00060
IND 0.00150 0.00038 0.00787 0.00197
DAE 0.00470 0.00118 0.03774 0.00944
BRA 0.56120 0.14030 0.84974 0.21244
ROW 0.00210 0.00053 0.00805 0.00201
calibrated to 12.5 % of the mean for OECD regions and 25% for non-OECD regions. Hence,
the variation in abatement costs is typically much smaller than the variation in benefits for
most regions. The key inputs for the normal distribution are shown in Table 3 and for USA
and China the corresponding probability density functions are illustrated in Appendix I.
4 Results
Using the confidence intervals and distribution functions described above, we calculate the
SL of all possible 4084 coalition structures. Results of this simulation are reported and dis-
cussed in Sect. 4.1. Section 4.2 identifies, which uncertainties are the main determinants
of the SL for selected coalitions using a meta-modelling approach. Section 4.3 investigates
an alternative scenario with lower probability of extreme observations for the benefits from
abatement, by adopting much stricter bounds on the distribution function of the global benefit
parameter. These stricter bounds imply a lower mean value of global benefits. Section 4.4
briefly discusses the results of some sensitivity analyses on the uncertainty of the various
parameters and the associated distribution functions. All simulations are carried out with a
precision of 95% on the second digit of SL (σ
SL
= 0.0025), implying that we keep drawing
samples until we have sufficient accuracy on SL to present two decimals.
4.1 Results for the base specification
Table 4 presents the main results for those coalition structures that have the highest SL and
for the Grand Coalition. Note that the stability of the All-Singletons Coalition Structure is
SL=1 by definition. This coalition structure can always be supported as an equilibrium if all
regions announce c
i
= 0. In the All-Singletons Coalition Structure abatement efforts amount
to almost 70 GtC per century, or an annual abatement of 5.8% of global emissions in 2010,
which leads to a discounted payoff of almost 8.8 trillion US$ over the century. The relatively
high abatement level is due to a high value of the global benefit parameter γ in the study
123
The stability likelihood of an international climate agreement
Table 4 Results for selected coalitions (mean values)
Coalition SL (fraction) Strict SL Total abatement Abatement Total
(fraction) (GtC) (% of emissions) payoff (bln $)
All-Singletons 1 1 69.4 5.8% 8775
JPN,EEC 0.24 0.15 74.5 6.2% 9251
USA,JPN 0.15 0.06 79.1 6.6% 9663
USA,EEC 0.14 0.05 84.7 7.1% 10090
JPN,BRA 0.12 0.02 69.7 5.8% 8824
JPN,ROW 0.11 0.02 76.7 6.4% 9613
EEC,ROW 0.11 0.01 79.2 6.6% 9835
JPN,IND 0.10 0.01 78.0 6.5% 9773
JPN,FSU 0.10 0.01 75.8 6.3% 9475
EEC,BRA 0.10 0.01 70.1 5.9% 8857
EEC,FSU 0.10 0.01 78.4 6.6% 9699
Grand Coalition 0.09 0 326.3 27.3% 27360
of Tol (2005), as discussed in Sect. 3.3, and due to our assumption of a relatively moderate
discount rate. A major part of the global abatement effort is carried out by the two regions
that have very flat marginal abatement cost curves, the USA and China. For these regions,
equating marginal abatement costs to marginal benefits implies substantial abatement levels.
In contrast, Brazil, which is characterised by a very steep marginal abatement cost curve and
low marginal benefits, hardly undertakes any abatement alone.
The relationship between the samples of global marginal benefits γ and the corresponding
global abatement levels is illustrated in Fig. 1 for the All-Singletons Coalition Structure. For
samples with negative marginal benefits, optimal global abatement is zero. For positive mar-
ginal benefits, there is a clear positive correlation between the marginal benefits and global
abatement, but the relation is not linear. As marginal abatement costs increase quadratic in
abatement, higher marginal benefits lead to a less than proportional increase in abatement
levels.
The graph clearly shows the dispersion of global marginal benefits as generated by the
distribution function and the resulting dispersion of optimal global abatement levels. The bulk
of samples produces global marginal benefits between 0 and 200 $/tC, with corresponding
optimal global abatement of less than 200 GtC over the century, though outliers may have
marginal benefits of more than 800 $/tC, while global abatement may be as high as 400 GtC.
The Grand Coalition would lead to a much larger global abatement effort and a higher
global payoff than no cooperation, as the gains from cooperation are fully reaped. However,
the Grand Coalition has a low SL. Though it is externally stable by definition, the SL for
internal stability is very low. In fact, the Grand Coalition is only stable when there are no
benefits to be reaped from abatement, i.e. when the global benefit parameter γ is zero or
negative. In such cases, optimal abatement levels are zero, and hence regions have no incen-
tive to leave this coalition. In the calculation of SL, this is interpreted as stable. Therefore,
Table 4 also reports “Strict SL”, which equals SL minus the “probability of indifference”.
Only a fewnon-trivial coalitions (i.e. coalitions with at least two members) have a positive
strict SL. All these coalitions are small and improve only little over the All-Singletons Coa-
lition Structure. The coalition of Japan and the European Union (EEC) has the highest SL.
123
R. Dellink et al.
0
50
100
150
200
250
300
-
1
0
0
-
5
0 0
5
0
1
0
0
1
5
0
2
0
0
2
5
0
3
0
0
3
5
0
4
0
0
4
5
0
5
0
0
Global marginal benefits ($/tC)
G
l
o
b
a
l

a
b
a
t
e
m
e
n
t
(
G
t
C
)
Fig. 1 Relationship between global marginal benefits and global abatement levels in the All-singletons coa-
lition structure
Therefore, we have a closer look at this coalition in Table 5, which provides regional data.
Though the standard deviations of the abatement levels are rather high, there is a probabil-
ity of more than 90% that abatement levels are strictly positive, implying that the implicit
distribution function for abatement levels is right-skewed. Though they are not part of the
coalition, the largest abatement efforts are carried out by China and the USA. These regions
have strong incentives to unilaterally reduce their emissions because of flat marginal abate-
ment cost functions compared to their high constant marginal benefits.
8
For Japan and the
European Union, marginal abatement costs are equal to the sum of marginal benefits of
this coalition, reflecting the first order condition implied by the assumption of joint welfare
maximisation of coalitions as explained in Section 2.
9
The last column in Table 5 shows the mean incentive to change membership. For exam-
ple, if the USA were to join the coalition of Japan and European Union, its mean payoff
would be reduced by 247 bln$. Similarly, Japan and the European Union would have a lower
payoff of 79 and 12bln$, respectively, if they were to leave the coalition. It turns out that
the mean values for the incentive to change membership are negative for all regions. Thus,
a deterministic model assuming mean parameter values, would come to the conclusion that
this coalition is stable. However, the corresponding SL is only 24%, implying that in three
quarter of all samples the coalition is unstable. This clearly illustrates the importance of using
a more sophisticated stability indicator, such as SL, when the probability density functions
of the model parameters are strongly skewed. Moreover, in a deterministic setting, no other
coalition would be stable, even though there are some other coalitions with a positive SL.
8
This is in line with the arguments put forward by the Bush administration: the USA will not ratify the Kyoto
protocol, but will carry out abatement efforts in its own interest.
9
For singletons, the equality between regional marginal abatement costs and regional marginal benefits holds
for every sample if marginal benefits are non-negative. However, this must not necessarily hold for mean
values. Similarly, for coalition members regional marginal abatement costs equal the sum of marginal benefits
of coalition members at the level of individual samples, but not necessarily for mean values.
123
The stability likelihood of an international climate agreement
Table 5 Regional results of the coalition Japan (JPN) and the European Union (EEC)
Region Total abatement Abatement Total payoff MAC
i
($/tC) MB
i
($/tC) Incentive
(GtC) (% emission) (bln $) (bln $)
Mean Standard Mean Mean Standard Mean Mean Mean
deviation deviation
USA 20.31 16.01 8.4% 2050 4313 17.6 17.3 −247
JPN 2.83 2.68 5.1% 1650 3525 31.7 13.3 −79
EEC 12.29 9.15 8.8% 2043 4213 31.7 17.9 −12
OOE 2.30 1.71 3.7% 369 817 2.9 2.8 −127
EET 1.14 1.20 2.2% 141 406 1.1 1.1 −145
FSU 5.64 5.11 5.6% 670 1712 5.4 5.3 −221
EEX 1.40 2.01 1.2% 312 850 2.5 2.4 −206
CHN 21.02 24.19 8.9% 611 1739 4.9 4.8 −1282
IND 3.30 5.59 5.2% 441 1647 3.6 3.5 −289
DAE 0.81 1.80 2.0% 245 1008 1.9 1.9 −167
BRA 0.03 0.08 0.2% 151 636 1.2 1.2 −11
ROW 3.42 5.54 4.9% 568 2117 4.7 4.6 −227
Note: MAC
i
= regional marginal abatement costs; MB
i
= regional marginal benefits, Incentive = change in net
benefits if one region changes its announcement of membership, given the announcement of all other regions
where initially all regions anncounce c
i
= 0, except JPN and EEC that announce c
i
= 1
As discussed in the introduction, we can also relate our results to the cases of no and full
learning in the presence of uncertainty (i.e. no and full resolution of uncertainty) as described
in Kolstad and Ulph (2006). The case of no learning corresponds to the deterministic set-
ting. Hence, the expected total payoff in case of no learning is 9251bln$, resulting from the
only stable coalition between Japan and the European Union. In the case of full learning,
the expected payoff is the average payoff of all stable coalitions for each parameter sample.
It seems plausible to assume that if there are multiple stable coalitions for a sample, each
equilibriumis equally likely. It turns out that expected payoff is 9105bln$. Hence, the impact
of learning on the global payoff is negative in this example, confirming the conclusion of
Kolstad and Ulph (2006).
4.2 Ranking the uncertainties
Based on the information on individual samples, we can investigate which uncertainties are
significant in influencing the stability of a coalition. Unfortunately, we cannot derive a ranking
of the uncertainties directly from the SL estimator. The fact that not all stochastic parameters
are normally distributed requires that we have to resort to numerical approximations. Using
a meta-modelling approach, we construct a Probit model to identify which of the uncertain
parameters influence the stability of the coalition between Japan and the European Union. To
allow for a direct interpretation of the coefficients, we standardise all inputs such that they
lie in the interval [0, 1]. The results are depicted in Table 6.
The main determinant of stability of the coalition is the level of global benefits: higher
benefit levels have a negative impact on the stability of this coalition. As expected, there is
a positive impact of the benefit share of the coalition member European Union (EEC), while
the benefit shares of other regions mostly have a negative or insignificant impact on stability.
123
R. Dellink et al.
Table 6 Results of the meta-analysis of the coalition between Japan and the European Union
Parameter Coefficient Standard deviation Parameter Coefficient Standard deviation
Global benefits −3.64 0.10** Intercept −0.15 0.11
Benefit shares Abatement cost parameter
USA −1.50 0.08** USA −0.10 0.06
JPN 0.03 0.07 JPN −0.05 0.06
EEC 2.54 0.06** EEC 0.12 0.06**
OOE −0.01 0.07 OOE −0.00 0.06
EET −0.04 0.09 EET 0.00 0.06
FSU −0.49 0.09** FSU −0.08 0.06
EEX 0.01 0.09 EEX 0.07 0.06
CHN −0.04 0.08 CHN 0.05 0.07
IND −1.50 0.16** IND 0.05 0.06
DAE −0.72 0.10** DAE 0.08 0.06
BRA −2.37 0.16** BRA −0.10 0.06
ROW −1.94 0.14** ROW −0.05 0.06
Note: ** indicates significance at the 5% level; pseudo R
2
equals 0.09
Surprisingly, the benefit share of Japan is also insignificant. This means that Japan has a large
incentive to collaborate, even when its regional benefit level turns out to be relatively low.
This can be explained by noting that Japan has to contribute little to cooperation because of
its steep marginal abatement cost function. The large coefficient of the benefit share of the
European Union confirms the importance of internal stability. The significant and negative
coefficients of the benefit shares of the USA, former Soviet Union, India, dynamic Asian
economies, Brazil and the Rest of the World indicate that when these regions have high
benefit shares, they will want to join the coalition of Japan and the European Union, thereby
making it externally unstable.
Table 6 confirms that marginal abatement costs are much less important; this is related to
the assumed smaller variance of the marginal abatement cost parameters (see Sect. 3). The
only significant parameter for abatement costs is that of the European Union. As the positive
coefficient shows, higher marginal abatement costs in the European Union have a positive
impact on stability of the coalition with Japan. The reason is the same as mentioned for Japan
above. A steep marginal abatement cost curve implies for a coalition member a small share
in coalitional abatement efforts and hence a lower incentive to free-ride.
A similar probit analysis is carried out for the coalition between the USA and Japan; the
results are depicted in Table 7.
The significant inputs for the stability of this coalition are the same as for the coalition
between Japan and the European Union; additionally the benefit shares of Japan (negative
effect) and Other OECD(positive effect) are important. The ranking of the main uncertainties
is also the same, except for the higher ranking of the benefit share of the USA. As the USA is
now a coalition member and the European Union is not, the signs of their benefit coefficients
reverse. The high ranking of the European Union (3rd) shows that external stability can pose
a problem in some cases: a high marginal benefit share in the European Union induces this
region to join the coalition of the USA and Japan, thereby making it externally unstable.
Though the benefit share of Japan is significant in explaining the SL of the coalition between
123
The stability likelihood of an international climate agreement
Table 7 Results of the meta-analysis of the coalition between the USA and Japan
Parameter Coefficient Standard deviation Parameter Coefficient Standard deviation
Global benefits −8.24 0.16** Intercept 0.26 0.13
Benefit shares Abatement cost parameter
USA 3.08 0.08** USA 0.09 0.07
JPN −0.36 0.09** JPN −0.09 0.07
EEC −1.95 0.09** EEC −0.12 0.07
OOE 0.17 0.08** OOE −0.02 0.07
EET 0.04 0.11 EET 0.10 0.07
FSU −0.60 0.10** FSU −0.08 0.07
EEX −0.05 0.10 EEX −0.05 0.07
CHN 0.01 0.09 CHN −0.08 0.08
IND −1.18 0.19** IND 0.07 0.07
DAE −0.47 0.11** DAE 0.07 0.07
BRA −1.65 0.18** BRA −0.06 0.07
ROW −1.32 0.16** ROW −0.06 0.13
Note: ** indicates significance at the 5% level; pseudo R
2
equals 0.17
the USA and Japan, its ranking is rather low, confirming the finding above that the incentive
for Japan to leave the coalition is in most circumstances (strongly) negative and thus cannot
explain the variation in stability.
4.3 Results for the alternative specification with lower mean global benefits
We expect that the probability density function of the global benefit parameter γ is crucial
for the analysis. Moreover, estimates of future benefits from abatement are widely acknowl-
edged to be highly uncertain. Therefore, we carry out an alternative simulation with a different
assumption on the global benefit parameter γ . Table 8 shows results assuming that the 5 per-
centile point is set to 0 $/tC and the 95 percentile point is set to 110 $/tC. This implies that
extreme values are rare and that the mean value of the global benefit parameter γ is reduced
from 77 to 37 $/tC. Note that in the STACO model, a lower mean value of γ has a similar
impact as raising the discount rate as explained in Weikard et al. (2006) : as benefits from
abatement will occur later than the associated costs, a higher discount rate implies a smaller
weight of (future) benefits and vice versa; see also the discussion of equation [4a]. Thus,
the discount rate is not explicitly subjected to a sensitivity analysis, but only implicitly. The
specific values considered here correspond to raising the discount rate from 2 to 7 percent.
10
For all coalitions, the SL is lower than in the base specification. This is partly due to the
lower probability of negative marginal benefits (this probability is reduced to 5%) and partly
due to the fact that mostly positive outliers of marginal benefits are removed from the model.
The ranking (in terms of SL) of different coalitions is largely unchanged, implying that only
the quantitative but not the qualitative conclusions are affected.
The lower mean value of γ reduces equilibriumabatement. In the All-Singletons Coalition
Structure, global abatement drops from around 70 GtC per century to 45 GtC. The reduction
10
Analogous to the procedure explained in note 5, we can also mimic these results using the Ramsey rule;
the associated pure rate of time preference equals 5.2%.
123
R. Dellink et al.
Table 8 Results for selected
coalitions assuming smaller
bounds on the global benefit
parameter γ
Coalition SL (fraction) Strict SL Mean total Mean total
(fraction) abatement payoff
(GtC) (bln $)
All-Singletons 1 1 44.5 2520
JPN,EEC 0.20 0.15 47.9 2660
USA,JPN 0.11 0.06 51.2 2790
USA,EEC 0.10 0.05 55.2 2929
JPN,BRA 0.07 0.02 44.7 2532
JPN,ROW 0.07 0.02 49.5 2774
EEC,ROW 0.06 0.01 51.2 2844
JPN,IND 0.06 0.01 50.3 2819
EEC,BRA 0.06 0.01 45.0 2543
JPN,FSU 0.06 0.01 49.1 2738
EEC,IND 0.06 0.01 52.1 2894
Grand Coalition 0.05 0.00 225.8 8353
in mean marginal benefits does not translate into a proportional reduction in mean abatement
levels. This is due to the non-linear nature of the model: the quadratic marginal abatement
cost functions imply that lower constant marginal benefits lead to less than proportionately
lower abatement levels (cf. Fig. 1). Lower benefits and abatement levels also lead to lower
payoffs (2.5 trillion $ versus 8.8 trillion$) compared to the base specification.
It is interesting to note that in a deterministic setting, with γ = 37 $/tC, no coalition
except the All-Singletons Coalition Structure would be stable. Only if this value is raised
by more than 20% does the coalition between Japan and the European Union emerge as
stable coalition. In contrast, the SL-value changes only from 20% to 24% if the mean value
of γ is raised from 37 to 77 $/tC. Once more, this highlights that the SL-concept is useful
when conducting an analysis of stability of coalitions when parameter values are not exactly
known. Again, we can relate our results to the cases of no and full learning in Kolstad and
Ulph (2006). In the case of no learning, the expected total payoff is 2520 bln $, resulting from
the All-Singleton Coalition Structure. For the expected payoff in the case of learning we find,
according to the procedure outlined in Sect. 4.1, a value of 2724 bln $. Thus, now learning
has a positive impact on the global payoff, in contrast with the finding in Kolstad and Ulph
(2006). However, it has to be pointed out that Kolstad and Ulph’s assumptions imply that
in the case of no learning at least a coalition of two countries is stable. Hence, our different
conclusion may rely exactly on this fact.
4.4 Results of the sensitivity analyses
In order to shed light on the robustness of our previous conclusions, a series of sensitivity
analyses are carried out. The main results of alternative assumptions regarding the uncertainty
of the main model parameters are reported in Table 9. A second set of sensitivity analyses,
which is reported in Table 10, concerns the functional forms of the distribution of uncertain
parameters.
A general conclusion that emerges from Tables 9 and 10 is that although the regional
abatement levels and payoffs depend on several model parameters, the sensitivity of SL is
123
The stability likelihood of an international climate agreement
Table 9 Stability Likelihood of selected coalitions under different model specifications
Coalition Base No uncertainty No uncertainty No uncertainty
specification w.r.t. s
i
w.r.t. γ w.r.t. α
i
, β
i
JPN,EEC 0.24 0.63 0.17 0.24
USA,JPN 0.15 0.09 0.07 0.15
USA,EEC 0.14 0.09 0.06 0.14
JPN,BRA 0.12 0.09 0.02 0.11
JPN,ROW 0.11 0.09 0.02 0.11
EEC,ROW 0.11 0.09 0.02 0.10
JPN,IND 0.10 0.09 0.01 0.10
JPN,FSU 0.10 0.09 0.01 0.10
EEC,BRA 0.10 0.09 0.01 0.10
EEC,FSU 0.10 0.09 0.01 0.10
Grand Coalition 0.09 0.09 0 0.09
Table 10 Stability Likelihood of selected coalitions under different model specifications
Coalition Base specification Normal distribution s
i
Normal Gamma

s
i
distribution distribution varies
γ α
i
, β
i
from unity
JPN,EEC 0.24 0.26 0.32 0.24 0.24
USA,JPN 0.15 0.13 0.24 0.15 0.15
USA,EEC 0.14 0.14 0.23 0.14 0.14
JPN,BRA 0.12 0.09 0.20 0.11 0.11
JPN,ROW 0.11 0.09 0.20 0.11 0.11
EEC,ROW 0.11 0.09 0.20 0.10 0.11
JPN,IND 0.10 0.06 0.20 0.10 0.10
JPN,FSU 0.10 0.07 0.20 0.10 0.10
EEC,BRA 0.10 0.08 0.19 0.10 0.10
EEC,FSU 0.10 0.08 0.19 0.10 0.10
Grand Coalition 0.09 0.01 0.19 0.09 0.09
much smaller. Essentially, the value of SL depends crucially on two elements: (i) the variance
in regional benefit shares (s
i
), and (ii) the variance and level of the global benefit parameter
(γ ). Reducing the uncertainty with respect to the regional benefit shares from the model
greatly reduces the variability of stability over the different coalitions. Thus, when regional
benefit shares are known, it becomes much easier to predict which regions have incentives
to collaborate and which do not. Then, only the coalition between Japan and the European
Union has a SL that is significantly higher than the probability of indifference. Removing
the uncertainty in the global benefit parameter implies that the probability of indifference is
zero, and this affects the SL of all coalitions.
The ranking of different coalitions in terms of their SL is very robust: in all simulations,
the same ranking emerges; the coalition between Japan and the European Union has the
highest SL, followed by the coalition between the USA and Japan and the coalition between
123
R. Dellink et al.
the USA and the European Union. Thus, we can conclude that SL provides a highly robust
indicator of the stability of climate coalitions.
This is also confirmed in the last column in Table 10 where the assumption is given up that
the sum of regional benefit shares is one. Thus, global benefit levels vary not only with the
variations in the global benefit parameter, but also with the sumof the regional shares. It turns
out that this assumption was not important: results are very similar to the base specification.
5 Discussion
This paper investigated the formation and stability of coalitions to reduce the negative impacts
of climate change. We introduced a methodology to calculate the stability of all possible coa-
lition structures in a stochastic, empirical setting. The concept of SL provided a notion of
stability that conveyed much more relevant information than the ordinary binary outcome:
stable/unstable. The numerical application showed that the SL-concept contributes to a better
understanding of stability of coalitions when there is uncertainty about model parameters: it
constitutes a superior alternative to running numerous sensitivity analyses in a deterministic
setting.
The results suggested that for most possible climate coalition there is at least one region
that is better off by changing its actions, thereby rendering coalitions unstable. This included
important coalitions such as the coalition of industrialised countries and the Grand Coalition.
The lack of stability was very robust and held for a wide range of possible values for regional
abatement costs and benefits, as described by the respective probability distribution func-
tions. There were only a small number of coalitions with a (strict) SL significantly different
from zero. Only when benefits from abatement are zero or negative (calibrated to be around
9 percent in the base specification), i.e. when climate change does not pose a problem, will
regions be indifferent between cooperation and no cooperation. However, in such cases, the
optimal abatement levels were zero, regardless of which coalitions formed.
The coalition with the highest SL was the coalition of Japan and the European Union, with
an SL equal to 24%in our base case. This relatively lowpercentage stressed the difficulties in
striking an international environmental agreement. This coalition but also the other coalitions
with a positive SL hardly improved over the All-Singletons Coalition Structure, correspond-
ing to the case of no cooperation: global abatement and global welfare were only slightly
higher than when there was no international collaboration at all, confirming a conclusion in
the literature derived with deterministic models.
The calculated SL of different coalitions was robust in terms of the considered distribution
functions of parameter values, the variance in regional abatement costs and the global level
of benefits from global abatement, but was sensitive to the variance in the regional distri-
bution of benefits. Unfortunately, regional benefits from abatement are very uncertain, and
international research on adaptation and damage costs is still in its infancy. Therefore, it is
of utmost importance to get better information on regional benefits.
We also related our results to those derived from stylized models with ex-ante symmet-
ric players which conclude that learning has a negative impact on the outcome of coalition
formation. It became apparent that our approach corresponds to the case of full resolution
of uncorrelated uncertainty with perfect learning, according to the classification in Kols-
tad and Ulph (2006). This case could be compared to the case of no resolution of uncer-
tainty where decisions on coalition membership have to be taken on the basis of expected
values. We showed that learning may not always have a negative impact on global wel-
fare. This suggested that the conclusions derived from stylized models as for instance in
123
The stability likelihood of an international climate agreement
Na and Shin (1998), Kolstad and Ulph (2006) and Kolstad (2007) are sensitive to the under-
lying assumptions.
There are many ways in which the analysis above can be ameliorated. Firstly, better esti-
mates of the variance on regional benefits may be obtained via a meta-analysis of the existing
literature, similar to what Tol (2005) did for global benefits. Secondly, the empirical model
used to calculate the regional payoffs can be extended to a fully-fledged computable general
equilibrium model. Together, we expect that these two extensions will improve the empirical
validity of the numerical example. Thirdly, it would be interesting to consider the case of
partial learning as analysed in Kolstad and Ulph (2006) in order to better understand the role
of learning. Finally, the analysis could be extended to decision making under uncertainty by
including risk-aversion of players as for instance considered in Endres and Ohl (2003).
Appendix 1. Histograms of the uncertain model parameters
Fig. A.1 Histogram of the global benefit parameter (γ )
Fig. A.2 Histogram of the regional benefit parameters (s
i
) for USA and China, respectively
123
R. Dellink et al.
Fig. A.3 Histogram of the regional abatement cost parameter (α
i
) for USA and China, respectively
Acknowledgements Special thanks to Ivo Mulder and Kelly de Bruin for their research assistance and to
Ekko van Ierland, Eligius Hendrix, John Janmaat and two anonymous referees for their comments; the usual
disclaimer applies.
References
Babiker MH, Reilly JM, Mayer M, Eckaus RS, SueWing I, Hyman RC (2001) The MIT emissions prediction
and policy analysis (EPPA) model: revisions, sensitivities and comparison of results. Joint program on
the science and policy of global change Report, 71. MIT, Cambridge
Barrett S(1994) Self-enforcing international environmental agreements. Oxford Economic Papers 46:878–894
Bosello F, Buchner B, Carraro C(2003) Equity, development, and climate change control. J Euro Econ Assoc
1:601–611
Botteon M, Carraro C(1997) Burden-sharing and coalition stability in environmental negotiations with asym-
metric countries. In: Carraro C (ed) International environmental negotiations: strategic policy issues.
Edward Elgar, Cheltenham 26–55
Carraro C, Siniscalco D(1993) Strategies for the iternational potection of the environment. J Publ Econ 52:309–
328
Dellink RB, Finus M, van Ierland EC, Altamirano J-C (2004) Empirical background paper of the STACO
model, available on the STACO website http://www.enr.wur.nl/uk/staco, Wageningen University
Ellerman AD, Decaux A (1998) Analysis of post-Kyoto CO
2
emissions trading using marginal abatement
curves. Joint program on the science and policy of global change Report, vol 40. MIT, Cambridge
Endres A, Ohl C(2003) International environmental cooperation with risk aversion. Int J Sustain Develop
6:378–392
Eyckmans J, Finus M(2006) Coalition formation in a global warming game: how the design of protocols
affects the success of environmental treaty-making. Nat Resour Model 19:323–358
Fankhauser S(1995) Valuing climate change. Earthscan, London
Finus M, van Ierland EC, Dellink RB(2006) Stability of climate coalitions in a cartel formation game. Econ
Govern 7:271–291
Hoel M(1992) International environment conventions: the case of uniform reductions of emissions. Environ
Resour Econ 2:141–159
Kolstad CD(1996) Fundamental irreversibilities in stock externalities. J Publ Econ 60:221–233
Kolstad CD(1996) Learning and stock effects in environmental regulation: the case of greenhouse gas emis-
sions. J Environ Econ Manage 31:1–18
Kolstad CD(2007) Systematic uncertainty in self-enforcing international environmental agreements. J Environ
Econ Manage 53:68–79
Kolstad CD, Ulph A (2006) International environmental agreements under uncertainty when learning reveals
differences between countries, paper presented at the conference “Learning and climate change”. IIASA,
Laxenburg, Austria
123
The stability likelihood of an international climate agreement
Manne AS, Mendelsohn R, Richels RG(1995) Merge:a model for evaluating regional and global effects of
GHG reduction policies. Energy Policy 23:17–34
Metz B, Davidson O, Swart R, Pan J (eds) (2001) Climate change 2001: mitigation. Cambridge University
Press, Cambridge
Na S, Shin HS(1998) International environmental agreements under uncertainty. Oxford Economic Papers
50:173–185
Nagashima M, Dellink RB, van Ierland EC(2006) Dynamic transfer schemes and stability of international
climate coalitions. Mansholt Discussion Paper 23, Wageningen
Nordhaus WD(1994) Managing the global commons. MIT Press, Cambridge
Nordhaus WD, Boyer J (2000) Warming the world: economic models of global warming. MIT Press, Cam-
bridge
Olieman NJ, Hendrix EMT(2006) Stability likelihood of coalitions in a two-stage cartelgame: an estimation
method. Euro J Oper Res 174:333–348
Roughgarden T, Schneider SH(1999) Climate change policy: quantifying uncertainties for damages and opti-
mal carbon taxes. Energy Policy 27:415–429
Tol RSJ (1997) Adecision-analytic treatise of the enhanced greenhouse effect. PhD. Thesis. Vrije Universiteit,
Amsterdam
Tol RSJ (2001) Climate coalitions in an integrated assessment model. Computat Econ 18:159–172
Tol RSJ (2002) Estimates of the damage costs of climate change: part 1: benchmark estimates. Environ Resour
Econ 21:47–73
Tol RSJ (2005) The marginal damage costs of carbon dioxide emissions: an assessment of the uncertainties.
Energy Policy 33:2064–2074
Ulph A(2004) Stable international environmental agreements with a stock pollutant, uncertainty and learning.
J Risk Uncertain 29:53–73
Ulph A, Maddison D(1997) Uncertainty, learning and international environmental policy coordination. Envi-
ron Resour Econ 9:451–466
Ulph A, Ulph D(1997) Global warming, irreversibility and learning. Econ J 107:636–650
Weikard H-P, Finus M, Altamirano-Cabrera JC(2006) The impact of surplus sharing on the stability of inter-
national climate coalitions. Oxford Econ Papers 58:209–232
Weitzman ML(1998) Why the far-distant future should be discounted at its lowest possible rate. J Environ
Econ Manage 36:201–208
Weitzman ML(2001) Gamma discounting. Am Econ Rev 91:260–71
123

R. Dellink et al.

1 Introduction There are many obstacles that prevent the formation of effective international environmental agreements (IEAs). Two of the most important obstacles are free-riding and uncertainty. First, IEAs aim at providing a public good and therefore countries face free-rider incentives. Typically it is beneficial for a country when other regions increase their abatement efforts. This free-riding problem has been studied extensively in theoretical models, as for instance in Hoel (1992), Carraro and Siniscalco (1993) and Barrett (1994) and in empirical climate models, as for instance in Bosello et al. (2003), Botteon and Carraro (1997), Tol (2001), Eyckmans and Finus (2006), and Finus et al. (2006). A main conclusion of this literature is that stable coalitions improve only marginally upon the non-cooperative outcome because either stable coalitions are small or the global abatement level that can be sustained in larger coalitions is very low. However, regardless of the specific model, all of these papers determine stable coalitions in a deterministic setting. Second, the uncertainties surrounding the future impacts of climate change as well as abatement costs are large. In particular, the benefits from abatement that arise in the distant future are highly uncertain, but also future abatement costs are not fully known. These issues are highlighted for instance in Roughgarden and Schneider (1999) and Tol (2002, 2005). Depending on the model specification and assumptions, uncertainty and learning may either call for laxer environmental standards today or imply that countries follow the precautionary principle by increasing their short-term abatement efforts (Kolstad 1996a, b; Ulph and Maddison 1997; Ulph and Ulph 1997). However, regardless of how uncertainty and learning is captured in these papers and what this means for optimal abatement strategies, success can only be evaluated if free-rider incentives are also included in the analysis. This implies that models with a single global decision-maker cannot provide full insight into the regional impacts of uncertainty and learning. This drawback is common to most integrated assessment models that are used to analyse optimal abatement strategies, including for instance DICE/RICE (Nordhaus and Boyer 2000), MERGE (Manne et al. 1995) and EPPA (Babiker et al. 2001). This also applies to some extent to models with two countries that compare non-cooperative with cooperative behaviour and hence coalition formation is not really an issue in these models (e.g. Ulph and Maddison 1997; Ulph and Ulph 1997). Recently, some papers have appeared that combine coalition formation with uncertainty and learning (Na and Shin 1998; Ulph 2004; Kolstad and Ulph 2006; Kolstad 2007). They assume uncertainty about parameter values of the payoff function where probability distribution functions are known to all countries. In the context of coalition formation, in which countries choose their membership in the first stage and their abatement strategies in the second stage, this allows to distinguish three cases according to Kolstad and Ulph (2006) and Kolstad (2007): (1) Uncertainty is resolved before the first stage. This corresponds to the case of full learning. (2) Uncertainty is resolved after the first stage but before the second stage. This corresponds to the case of partial learning. (3) Uncertainty is not resolved. This is the case with no learning. In cases 1 and 2 where there is learning, it takes the form of perfect learning, i.e. all players learn the true values of all uncertain parameters. All these papers are stylized models with ex-ante identical players and mostly come to the conclusion that learning has a negative impact on the success of coalition formation. Na and Shin (1998) compare case 2 and 3, but their model is essentially static (flow pollutant) and restricted to only three countries. The first restriction is removed in Ulph (2004), who compares all three cases and considers a two period model with a stock pollutant. Due to this complication, results are based on simulations and conclusions are not always clear-cut. Moreover, he assumes that countries have only two abatement strategies (no abatement versus

123

in a deterministic model. Together with case 3. (2006). Section 3 discusses the calibration of the model. while almost all model parameters are assumed to be uncertain with (uncorrelated) probability distribution functions. we introduce our concept of Stability Likelihood (SL). Section 4 presents and discusses result of our stability analysis. In the first stage. Case 2 is ambiguous. In the first stage. Hence. Case 3 implies that a coalition cannot be called stable or not stable any longer (based on certain payoffs in the deterministic case or expected payoffs in the case of no resolution of uncertainty). That is.The stability likelihood of an international climate agreement full abatement). either all countries are low or high benefit countries. learning may have a positive or negative impact on participation and welfare. as proven in Kolstad and Ulph (2006). despite the assumption that all countries are ex-ante identical and that there are only two states in the world. in a deterministic setting. This means that in the case of learning all countries are also ex-post identical. regions can choose from a continuum of abatement strategies. we find that learning does not always have a negative impact on the success of coalition formation. These assumptions are also made in Kolstad (2007). in the case of uncorrelated uncertainty. including the distribution functions used for the uncertain model parameters. case 1 allows for a direct comparison with the conclusions derived in Kolstad and Ulph (2006). regions decide whether to become a member of IEA or to remain an outsider. Section 5 concludes. Secondly. For this. namely no and full resolution of uncertainty. Because of the difficulties of deriving analytical results even in a simple model. Case 1 is relatively straightforward. as for instance in Eyckmans and Finus (2006) and Finus et al. we show that the concept of SL has advantages over a deterministic analysis. he unambiguously shows that when comparing case 1 and 3. Announcement ci = 1 means “Region i joins the coalition” and announcement ci = 0 123 . Because we assume that utility is linear in payoffs as in Kolstad and Ulph (2006). we can also apply the certainty equivalence principle. we consider uncertainty. Finally. 2. We restrict ourselves to the analysis of cases 1 and 3. Compared to Kolstad and Ulph (2006) and the other papers above. the case of no resolution of uncertainty is identical to the deterministic case if the mean or expected value of a parameter corresponds to the value in the deterministic case. in the second stage. coalitions can only be stable or unstable. Moreover. but again have to resort to simulation. In the following. which determines the likelihood that a coalition structure is stable. each representing a region of the world. learning and coalition formation based on an empirical climate model with twelve heterogeneous world regions. learning increases the size of a stable IEA but has a negative impact on global welfare. uncertainty of parameter values can only be indirectly accounted for through sensitivity analyses where each analysis is subject to the first problem mentioned above. However. we present our model of coalition formation and introduce the concept of SL in Sect. Because of the assumption of a flow pollutant. invalidating conclusions. This may imply in some cases that small perturbations of parameters are sufficient to change the set of stable coalitions. Kolstad and Ulph (2006) confirm the negative impact of learning. which is the focus of this analysis. with only two states (low and high benefits) and uncertainty is correlated between players. Firstly. the concept of SL can be interpreted as a more sophisticated sensitivity analysis than is traditionally done in the context of coalition formation with empirical data. the only uncertainty concerns the benefits from global abatement. regions can only choose between becoming a signatory or to remain an outsider as assumed in the papers above. 2 The model of coalition formation Consider a set of N heterogeneous players.

b). Hence. a coalition structure c is described by announcement vector c = (c1 . i. Also in the first stage. due to the public good nature of abatement) and on a vector of model parameters b. b)∀qk and (1) where qk is the abatement strategy vector of coalition k. qi∗ .e. For ˜ our purposes. the grand coalition forms. The reason is that regardless whether all regions announce ci = 0 or only one region announces ci = 1 and all other announce ci = 0. but non-cooperative towards outsiders. Economically. Note that there are 2 N announcement vectors. this means strategies are group (but not globally) efficient within coalition k. q−k . b) ≡ πi (q ∗ (c). qi the strategy of non-signatory i. “Region i remains an outsider”. . c N ). this implies that non-signatories (i ∈ k) that announced ci = 0 will choose their abatement strat/ egies so as to maximize their individual payoff πi (q. q−k the vector of all regions not belonging to k. and q−i the strategy vector of all other non-signatories except i. a coalition structure is defined by coalition k. . b). . qN ). b). taking the abatement strategies of all other regions as given. .R. Computationally. b) depends on abatement vector q (and hence on the strategy of all regions. πi (q. q−i . b) ≥ i∈k i∈k ∗ ∗ ∗ ∗ πi (qk . ∀i = 1. whereas all signatories (i ∈ k) that announced ci = 1 jointly maximize the aggregate payoff of their coalition i∈k πi (q. Thus. signatories’ behaviour is cooperative towards their fellow members. N }. b). any inefficiency stems from the fact that k is not the grand coalition. We solve the game backwards assuming that strategies in each stage must form a Nash equilibrium. given the announcement of other players c−i .e. this means that the behaviour of non-signatories towards all other regions is selfish and non-cooperative. q−i . which assigns the value 1 to a stable announcement vector (i. the equilibrium economic strategy vector q ∗ corresponds to the classical “social or global optimum” if coalition k comprises all countries. .e. we can use the term coalition structure and coalition interchangeably. unstable coalition): 123 . For the second stage. qi . this will result in the “All-Singletons Coalition Structure”. Thus. . The set of players that announce 1 are coalition members and is denoted as k = {i | ci = 1. no signatory that announced ci = 1 should have an incentive to change its announcement to ci = 0 (internal stability) and no non-signatory that announced ci = 0 should want ˜ to announce ci = 1 (external stability). Since in the context of our empirical model the equilibrium abatement vector q ∗ is unique for every coalition structure c and a given vector of parameters b (see the proof in Olieman and Hendrix 2006). In the second stage. and corresponds to the classical “Nash equilibrium” if coalition k comprises only one member or is empty. . q−k . We now turn to the first stage. ∗ ∗ πi (qk . Strategically. Dellink et al. That is. . . strategies form a Nash equilibrium. . this condition can be summarised compactly by the payoff stability function f (c. The payoff of an individual region i. this entails that abatement strategies form a Nash equilibrium between coalition k and the remaining non-signatories. regions choose their abatement levels. there is a unique vector of equilibrium payoffs for every coalition structure c: vi (c. . This leads to abatement vector q = (q1 . in this simple setting. That is. but essentially there are only 2 N − N coalition structures. b) ≥ πi (qk . stable coalition) and the value 0 to an unstable announcement vector (i. b)∀qi ∗ πi (qk . Hence. . An asterisk denotes equilibrium strategies.

dynamic Asian economies (DAE). If the vector of mean values of b is b. called STAbility of COalitions (STACO) is described. (2004) and Finus et al. b) will take on the value 1 if for all regions the payoff from announcement c is higher than. These ˜ alternatives are constructed by changing the announcement of one player at a time. 3). This implies around 2 billion calculations of optimal regional abatement and payoff levels (40. the model should reflect important results and features of integrated assessment models in terms of overall magnitudes of global 1 We are not aware of any paper that provides analytical solutions of stable coalitions in the context of heterogeneous players even in the absence of uncertainty. b) ≥ 0) for ci = 1 − ci and for all i = 1. b) = ⎩ 0 else (2) The function f (c. Japan (JPN). we can assign a likelihood to the event f (c. 2 We need around 40. energy exporting countries (EEX). . The philosophy behind the construction of the model comprises two items. former Soviet Union (FSU). India (IND).1 The SL is calculated using a Monte Carlo sampling technique: we generate M samples bm 1 from b. However. 123 . or equal to. Brazil (BRA) and “rest of the world” (ROW). Pr) and probability density function g (b). B. b) = 1}. see Dellink et al. 3 Calibration of the model 3. .2 Second. In the case of no resolution of uncertainty. The heterogeneity of players with payoff functions based on the empirical model STACO (see Sect.e. N f (c. requires us to resort to numerical calculations for the SL of various coalitions. other OECD countries (OOE). c−i . A more detailed discussion of the SL concept and computation methods can be found in Olieman and Hendrix (2006). c−i . European Union (EEC). the model must be simple enough to allow for sufficient samples to be calculated within reasonable time. For more detailed information on the model and calibration procedure. the decision on membership in the first stage will be based on expected payoffs using the mean values of b. i. bm ) that m=1 1 has an estimated variance of M−1 S L 1 − S L . this corresponds to the deterministic setting. (2006). in the case of full resolution. based on these samples we can estimate the SL with S L = M M f (c. b) · g(b)db.000 times 12 regions times 4096 membership announcement vectors).000 samples to get sufficient accuracy in the estimates of SL. to a stable coalition. It is worth noting that for any given set of parameters b. we can interpret SL as the probability that coalition structure c is stable. The SL of coalition structure c is therefore defined as SL ≡ Pr { f (c. b) = 1. this function allows for either a unique stable coalition. which equals B f (c. the payoff associated with alternative announcements c. b) becomes a Bernoulli variable. multiple stable coalitions or no non-trivial stable coalition. b) − vi (ci . . First.The stability likelihood of an international climate agreement ⎧ ˜ ˜ ⎨ 1 if (vi (ci . in a stochastic model the deterministic vector of parameters b is replaced by the stochastic (vector) variable b which is characterised by probability space (B.1 Introduction In this section. .) is a correct representation of reality. China (CHN). the calibration of the empirical model. The model comprises benefit and cost functions of twelve world regions: USA (USA). In a deterministic model. the vector of parameters b may be based on estimates but is treated as given. Eastern European countries (EET). Assuming that f (. decisions are based on the stochastic variable b and as function f (c. In contrast.

the model focuses on carbon dioxide. As benefits in period t result from the stream of abatement in all periods until t. but the exogenous level of other greenhouse gases is included in the calibration of the benefit function (cf. these marginal benefits are increasing over time. T B(q) can then be expressed as T B(q) = γ · q (4b) where γ represents discounted marginal benefits in $/ton carbon equivalents (tC) and is calculated as 2110 γ = t=2011 (1 + r )−t · ϕt . However. .R. Therefore. emissions and concentration. numerical results should be interpreted with caution. this allows for a direct comparision with the papers mentioned above about the role of learning. Dellink et al. we believe that our simulations illustrate the usefulness of SL as an indicator. Bit are benN efits from global abatement (qt = i=1 qit ). Furthermore. 2110. The analysis is based on the net present value of a stream of net benefits starting in 2010 and covering a time period of 100 years in order to capture the long-run effects of global warming. Discounted global or total benefits. we consider that the stock of CO2 can be approximated by a linear function of emissions and that damages are linear in the stock of CO2 . They are computed for every possible coalition structure and then stability is computed according to the procedure explained in Sect. . The payoff function is calculated as the net present value of a stream of net benefits from abatement between 2010 and 2110 and thus reflects discounted avoided damages minus discounted abatement costs. 2. (2004). 123 . 3.2 Elements of the empirical model The payoff function of region i is given by: T πi = t=1 (1 + r )−t (Bit (qt ) − ACit (qit )) (3a) where T denotes the time horizon. 3. Moreover. Sect. Nordhaus 1994). Kolstad and Ulph (2006) and Kolstad (2007). r is the discount rate. as we include the best available information on the probability density functions of the model parameters. Therefore. Assuming constant abatement strategies (qit = qi /100).2). while regional benefits depend on global abatement efforts. and ACit are abatement costs from regional abatement qit . The price for this complexity (see footnote 2) is that we assume constant abatement strategies (cf. t = 2011. . .3 It follows that annual global benefits from reduced emissions are also linear in the level of abatement: Bt (q) = ϕt · q (4a) where ϕt denotes marginal benefits in period t from total abatement over the entire period. abatement costs from regional abatement and benefits from global abatement over some time period. (5) 3 The details of this approximation are given in Dellink et al. benefits in year t can be expressed as a function of global abatement over the entire period. rendering the model essentially static as in Na and Shin (1998). This parameter also includes the effects of limited retention of GHGs in the atmosphere and the decay of the stock of GHGs over time. Regional abatement costs are a function of the level of individual abatement by region i.

we have insufficient information to fully estimate the probability density functions of all parameters at a regional level. (7) Taken together. For the specification of the abatement cost functions. T ACi (qi ). Assuming constant strategies and a constant discount rate. estimates of the EPPA model are used (Ellerman and Decaux 1998). αi and βi . (2006). The exact shape of the equilibrium emission or abatement paths are not important for our analysis. The assumption that the abatement paths are constant over time simplifies the calculations. discounted payoffs). the decay rate of greenhouse gases and the discount rate are assumed to be known with certainty. It assumes an annual abatement cost function of region i of the following form: ACit (qit ) = 1 1 · αi · (qit )3 + · βi · (qit )2 3 2 (6a) with αi and βi regional cost parameters. we also consider the possibility of a higher discount rate. the payoff function of region i can be written as:4 πi = T Bi (q) − T ACi (qi ). the stability of coalitions is robust in this respect. We want to stress that the discount rate only affects the aggregation of costs and benefits over time. The discount rate r is set at 2% per year.The stability likelihood of an international climate agreement Regional discounted benefits are assumed to be a share of discounted global benefits from abatement: T Bi (q) = si · T B(q) = si · γ · q (4c) N where si is the share of global benefits for region i and i=1 si = 1. Other parameters. but is not essential. 123 . The nature of our analysis requires only that the overall magnitude of concentrations and emissions mimics that of more elaborate integrated assessment models (because decisions whether to join a coalition are based on discounted payoffs). (3b) Generally. Hence. 4. the total abatement cost function becomes T ACi (qi ) = ρ · with 2110 1 1 · αi · (qi )3 + · βi · (qi )2 3 2 (6b) ρ= t=2011 (1 + r )−t . This is in line with Weitzman (1998. A similar stability analysis in a deterministic setting with dynamic equilibrium abatement paths. is the sum of annual abatement cost over the entire time horizon.e. all model parameters are uncertain. reveals that the impact of our simplifying assumption is limited: while levels of payoffs are somewhat affected. A more advanced meta-analysis of published and unpublished estimates would require a study of its own and therefore is beyond the scope of the current paper. though this done indirectly by lowering the mean value of the global benefit parameter γ . we consider only uncertainty of the benefit and cost parameters γ . Total abatement cost of region i. Calibration of emissions and concentrations is based on the widely known DICE-model by Nordhaus (1994). 2001) who suggests that in the case of climate change long-terms effects should be discounted with a low discount rate and suggests that 2% is appropriate for the “medium future”.3. However. such as emissions in 2010. as in the case of benefits. In Sect. as reported in Nagashima et al. si . as marginal benefits are 4 This means that we base our stability analysis on the net present value of benefits and abatement costs (i.

(iii) the cumulative probability density at this separation value. these two alternatives are equivalent for our analysis. we opt for a gamma distribution that can handle the lower bound of the shares such that no region receives a negative share of the benefits. Typically. 5 Alternatively. The mean value of this function is 77 $/tC.3 Global benefit function The distribution function of the global benefit parameter γ is based on a recent study by Tol (2005). This function is described by four pieces of information: (i) the 5 percentile point.5% and use the Ramsey rule to derive a declining discount rate. The point of separation between both sides of the function is given by the mode and equals 5 $/tC. These numbers are summarised in Table 1 and the corresponding histogram of all drawn samples is shown in Appendix I. this value equals −9 $/tC. Dellink et al. this source does not provide sufficient information on the functional form of the distribution function of these shares. which are in turn based on Fankhauser (1995) and Tol (1997). In this case.4 Regional benefit shares The distribution function of regional benefit shares (si ) is based on insights from a study by Tol (2002). However. The probability density function of peer-reviewed studies as presented by Tol can be closely mimicked by a 2-sided exponential function. Note also that our assumption of a constant (money) discount rate is consistent with a low. The mean values used for the regional shares are the values reported in the deterministic setting of the STACO model. as only net present values matter in our model. 4. optimal abatement is zero. The point on the right side of the function is given by the 95% cumulative probability density and equals 245 $/tC. and (iv) the 95 percentile point (assuming this is above the separation value). (ii) the value where the two sides of the exponential function are separated (assuming this is above the 5 percentile point). as the shares can vary between zero and one and the mean value is typically close to zero.5 3. but positive. Table 1 Characteristics of the 2-sided exponential distribution function of the global benefit parameter γ Value 5% Density Mode Density at mode 95% density Mean −9 $/tC 5 $/tC 13% 245 $/tC 77 $/tC constant (though uncertain). 123 . Standard deviations are based on expert judgement. the distribution function should be skewed. The first piece of information is given by the value that corresponds to a cumulative probability of 5%. However. partial and full cooperation.4. the choice of the gamma distribution function will be subject to a sensitivity analysis in Sect. pure rate of time preference. Therefore.R. the associated cumulative probability density is 13%. Moreover. and regions are indifferent between no. 3. we could assume a constant pure rate of time preference of 0. using in particular Tol (2002). as this is the case for a gamma distribution. This distribution function implies that there is a probability of around 9% that the benefits from global abatement are zero or negative. implying a strictly positive probability that benefits from abatement are negative.

The consequences of this assumption will be investigated in a sensitivity analysis in Sect.0153 0.2360 0.The stability likelihood of an international climate agreement Table 2 Characteristics of the gamma distribution function of the regional parameter si (share of global benefits) Region USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW Lower bound 0 0 0 0 0 0 0 0 0 0 0 0 Mean 0.0675 0.0345 0. Due to lack of regional information. and are lower for OECD countries than for the other regions. which is in turn based on Ellerman and Decaux (1998).0249 0. the mean is based on the deterministic version of the STACO model.0216 0.0300 0.0498 0. the standard deviation is 6 More information on the impact of this rescaling on the probability density function can be obtained from the authors upon request.1475 0. Therefore.2263 0.0500 0. the restriction on the sum of shares is enforced to avoid an impact of variations in this parameter on global benefits. The regional numbers are represented in Table 2.1000 0. all shares are scaled up or down to force the sum of shares to unity.4 3. The normal distribution is chosen. where ψ i i i i i 3 2 123 . For a better understanding.1360 standard deviations are of the same order of magnitude as mean values.6 Though rescaling is not a necessary condition for the working of the model. 1 · α · (q )3 + 1 · β · (q )2 . and vary the level of marginal abatement costs per region in the simulations. The standard deviation of the abatement cost parameters are based on information provided by IPCC (Metz et al. as a priori.0620 0.1725 0.1414 0.0306 0. This implies that both parameters in the abatement cost functions move simultaneously and in the same direction.0675 0. we replace equation (6b) by T AC (q ) = ψ · ρ · i i i is stochastic with mean value 1 and αi and βi are deterministic. empirical studies report only variations in marginal abatement costs without providing information on how the slopes and curvatures of these functions vary.0300 0. 7 Effectively.1078 0.0680 Standard deviation 0. This implies that the regional shares do not strictly follow the gamma distribution and the resulting variance in the shares is smaller than in the original gamma distribution. Typically. there seems to be no reason to assume a skewed distribution. we interpret the model as if the variation occurs in αi and βi . However.7 We assume a normal distribution function. the corresponding histograms for the USA and China are shown in Appendix I. we assume the functional form of the abatement cost curves to be given. After the samples of the regional shares are drawn.0130 0.5 Abatement cost functions The two regional abatement cost parameters (αi and βi ) cannot be assumed to be independent. 4.0620 0.0130 0. 2001).

00787 0. Hence.00118 0. Section 4. which uncertainties are the main determinants of the SL for selected coalitions using a meta-modelling approach.00470 0. These stricter bounds imply a lower mean value of global benefits.00944 0.00060 0.00011 0.2 identifies.00757 0.01550 0.00398 0.00006 0. Results of this simulation are reported and discussed in Sect.1 Results for the base specification Table 4 presents the main results for those coalition structures that have the highest SL and for the Grand Coalition.00007 0.03774 0.00080 0.18160 0.56120 0.00830 0. Dellink et al.00058 0. 4.00000 0.02270 0. Note that the stability of the All-Singletons Coalition Structure is SL = 1 by definition.R.00194 0.84974 0.8 trillion US$ over the century.00122 0. or an annual abatement of 5. implying that we keep drawing samples until we have sufficient accuracy on SL to present two decimals. which leads to a discounted payoff of almost 8.00050 0.00053 βi Mean 0.00320 0. Table 3 Characteristics of the normal distribution of the abatement cost parameters αi and βi Region αi Mean USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW 0.00104 0.8% of global emissions in 2010.00805 Standard deviation 0. the variation in abatement costs is typically much smaller than the variation in benefits for most regions.00188 0. we calculate the SL of all possible 4084 coalition structures.00197 0.14030 0.00038 0. In the All-Singletons Coalition Structure abatement efforts amount to almost 70 GtC per century.00042 0.00002 0. 4.3 investigates an alternative scenario with lower probability of extreme observations for the benefits from abatement.00000 0.00198 0. The relatively high abatement level is due to a high value of the global benefit parameter γ in the study 123 .00230 0.00486 0.5 % of the mean for OECD regions and 25% for non-OECD regions.03029 0.00790 0.4 briefly discusses the results of some sensitivity analyses on the uncertainty of the various parameters and the associated distribution functions.0025).00239 0.01503 0. by adopting much stricter bounds on the distribution function of the global benefit parameter.00050 0.21244 0.00210 Standard deviation 0. Section 4. 4 Results Using the confidence intervals and distribution functions described above.00150 0. This coalition structure can always be supported as an equilibrium if all regions announce ci = 0. Section 4.00030 0.00240 0. All simulations are carried out with a precision of 95% on the second digit of SL (σ S L = 0. The key inputs for the normal distribution are shown in Table 3 and for USA and China the corresponding probability density functions are illustrated in Appendix I.00201 calibrated to 12.1.

and hence regions have no incentive to leave this coalition.12 0.BRA JPN.10 0. there is a clear positive correlation between the marginal benefits and global abatement.0 75.4% 6. which is characterised by a very steep marginal abatement cost curve and low marginal benefits.02 0.4 326.3% 5.8% 6. which equals SL minus the “probability of indifference”. the USA and China.e.EEC JPN. As marginal abatement costs increase quadratic in abatement.6% 27. The relationship between the samples of global marginal benefits γ and the corresponding global abatement levels is illustrated in Fig.5 79.02 0. The bulk of samples produces global marginal benefits between 0 and 200 $/tC.06 0.e.3% Total payoff (bln $) 8775 9251 9663 10090 8824 9613 9835 9773 9475 8857 9699 27360 All-Singletons JPN. optimal abatement levels are zero. 3. The coalition of Japan and the European Union (EEC) has the highest SL. when the global benefit parameter γ is zero or negative.09 of Tol (2005). higher marginal benefits lead to a less than proportional increase in abatement levels. Table 4 also reports “Strict SL”.FSU EEC. i. as discussed in Sect.01 0. In fact.1% 5.01 0.3. The Grand Coalition would lead to a much larger global abatement effort and a higher global payoff than no cooperation.7 69.IND JPN. with corresponding optimal global abatement of less than 200 GtC over the century.4 74. 1 for the All-Singletons Coalition Structure. while global abatement may be as high as 400 GtC. though outliers may have marginal benefits of more than 800 $/tC.2% 6.11 0.ROW JPN.EEC USA. optimal global abatement is zero.05 0.BRA EEC. the Grand Coalition is only stable when there are no benefits to be reaped from abatement. The graph clearly shows the dispersion of global marginal benefits as generated by the distribution function and the resulting dispersion of optimal global abatement levels. Brazil. but the relation is not linear.9% 6. Therefore. In contrast.1 84.6% 6.8 70.14 0.01 0. A major part of the global abatement effort is carried out by the two regions that have very flat marginal abatement cost curves.15 0.3 Abatement (% of emissions) 5.ROW EEC.1 78.7 76.11 0. All these coalitions are small and improve only little over the All-Singletons Coalition Structure. For samples with negative marginal benefits.7 79. hardly undertakes any abatement alone.15 0.6% 7. and due to our assumption of a relatively moderate discount rate. 123 . Though it is externally stable by definition. as the gains from cooperation are fully reaped. equating marginal abatement costs to marginal benefits implies substantial abatement levels.The stability likelihood of an international climate agreement Table 4 Results for selected coalitions (mean values) Coalition SL (fraction) Strict SL (fraction) 1 0.10 0. In such cases. However. In the calculation of SL. the Grand Coalition has a low SL. coalitions with at least two members) have a positive strict SL.01 0.24 0.FSU Grand Coalition 1 0.JPN USA.10 0.8% 6.2 78. the SL for internal stability is very low. For these regions.5% 6.01 0 Total abatement (GtC) 69.10 0. this is interpreted as stable. Only a few non-trivial coalitions (i. For positive marginal benefits.

there is a probability of more than 90% that abatement levels are strictly positive. This clearly illustrates the importance of using a more sophisticated stability indicator. would come to the conclusion that this coalition is stable. These regions have strong incentives to unilaterally reduce their emissions because of flat marginal abatement cost functions compared to their high constant marginal benefits. we have a closer look at this coalition in Table 5. for coalition members regional marginal abatement costs equal the sum of marginal benefits of coalition members at the level of individual samples. Dellink et al. 9 For singletons. Moreover. such as SL. which provides regional data. It turns out that the mean values for the incentive to change membership are negative for all regions. but will carry out abatement efforts in its own interest. implying that the implicit distribution function for abatement levels is right-skewed.8 For Japan and the European Union. However. no other coalition would be stable. a deterministic model assuming mean parameter values. respectively. Thus. even though there are some other coalitions with a positive SL. Though the standard deviations of the abatement levels are rather high. its mean payoff would be reduced by 247 bln $. this must not necessarily hold for mean values. However. 8 This is in line with the arguments put forward by the Bush administration: the USA will not ratify the Kyoto protocol. marginal abatement costs are equal to the sum of marginal benefits of this coalition. 123 500 0 -50 50 0 . when the probability density functions of the model parameters are strongly skewed. 300 250 Global abatement ( GtC) 200 150 100 50 -100 100 150 200 250 300 350 400 450 Global marginal benefits ($/tC) Fig. Similarly. the corresponding SL is only 24%. implying that in three quarter of all samples the coalition is unstable. if they were to leave the coalition. if the USA were to join the coalition of Japan and European Union. Though they are not part of the coalition.R. the largest abatement efforts are carried out by China and the USA. Japan and the European Union would have a lower payoff of 79 and 12 bln $. in a deterministic setting. reflecting the first order condition implied by the assumption of joint welfare maximisation of coalitions as explained in Section 2. the equality between regional marginal abatement costs and regional marginal benefits holds for every sample if marginal benefits are non-negative. For example. Similarly. but not necessarily for mean values.9 The last column in Table 5 shows the mean incentive to change membership. 1 Relationship between global marginal benefits and global abatement levels in the All-singletons coalition structure Therefore.

40 21.8 3.01 2.9 1.e. Hence.5 1.08 5. Hence.2% 4.9 1.6% 1.3 13.20 5. confirming the conclusion of Kolstad and Ulph (2006). no and full resolution of uncertainty) as described in Kolstad and Ulph (2006). the expected total payoff in case of no learning is 9251 bln $. we can investigate which uncertainties are significant in influencing the stability of a coalition.0% 0.4% 5.8% 3.9% 2050 1650 2043 369 141 670 312 611 441 245 151 568 17.9 2. To allow for a direct interpretation of the coefficients.83 12.8 1.31 2. except JPN and EEC that announce ci = 1 As discussed in the introduction.3 2.2 Ranking the uncertainties Based on the information on individual samples. 1]. the expected payoff is the average payoff of all stable coalitions for each parameter sample. Using a meta-modelling approach.6 1.5 4.4 4.64 1.7 31.2 4.2% 8.2 4.03 3. while the benefit shares of other regions mostly have a negative or insignificant impact on stability.The stability likelihood of an international climate agreement Table 5 Regional results of the coalition Japan (JPN) and the European Union (EEC) Region Total abatement (GtC) Mean Standard deviation 16.6 31.19 5. The fact that not all stochastic parameters are normally distributed requires that we have to resort to numerical approximations. the impact of learning on the global payoff is negative in this example. 123 . The main determinant of stability of the coalition is the level of global benefits: higher benefit levels have a negative impact on the stability of this coalition.4 2.7 17.02 3. MBi = regional marginal benefits. It seems plausible to assume that if there are multiple stable coalitions for a sample.81 0. It turns out that expected payoff is 9105 bln $.1 5.42 8.1 5.7 2.2% 5. we standardise all inputs such that they lie in the interval [0. we construct a Probit model to identify which of the uncertain parameters influence the stability of the coalition between Japan and the European Union.80 0.68 9. The case of no learning corresponds to the deterministic setting.7% 2. we cannot derive a ranking of the uncertainties directly from the SL estimator.9 3.30 1. resulting from the only stable coalition between Japan and the European Union. there is a positive impact of the benefit share of the coalition member European Union (EEC). each equilibrium is equally likely. we can also relate our results to the cases of no and full learning in the presence of uncertainty (i. Incentive = change in net benefits if one region changes its announcement of membership. 4.9 1.30 0.6 Note: MACi = regional marginal abatement costs.29 2. In the case of full learning.1% 8. Unfortunately.11 2.14 5.2% 2.54 Abatement (% emission) Mean Total payoff (bln $) Mean Standard deviation 4313 3525 4213 817 406 1712 850 1739 1647 1008 636 2117 MACi ($/tC) Mean MBi ($/tC) Mean Incentive (bln $) Mean −247 −79 −12 −127 −145 −221 −206 −1282 −289 −167 −11 −227 USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW 20.9% 5.3 17. The results are depicted in Table 6.01 24. given the announcement of all other regions where initially all regions anncounce ci = 0.59 1. As expected.15 1.71 1.

even when its regional benefit level turns out to be relatively low. India. The large coefficient of the benefit share of the European Union confirms the importance of internal stability. Brazil and the Rest of the World indicate that when these regions have high benefit shares.06** 0.04 −0.15 −0. Dellink et al. Table 6 confirms that marginal abatement costs are much less important. higher marginal abatement costs in the European Union have a positive impact on stability of the coalition with Japan. The reason is the same as mentioned for Japan above.50 0. the benefit share of Japan is also insignificant. dynamic Asian economies. Table 6 Results of the meta-analysis of the coalition between Japan and the European Union Parameter Global benefits USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW Coefficient −3.12 −0.07 0.06 0.64 Benefit shares −1. this is related to the assumed smaller variance of the marginal abatement cost parameters (see Sect. pseudo R 2 equals 0. This can be explained by noting that Japan has to contribute little to cooperation because of its steep marginal abatement cost function.05 0.09 0. additionally the benefit shares of Japan (negative effect) and Other OECD (positive effect) are important.06 0.01 −0.06 0.01 −0.08 0.04 −1.03 2.08 0. thereby making it externally unstable.94 0. A similar probit analysis is carried out for the coalition between the USA and Japan.37 −1.54 −0.11 0.16** 0.06 Abatement cost parameter Note: ** indicates significance at the 5% level.07 0.10 −0.00 0.09** 0. thereby making it externally unstable.08 −0.06 0.10** 0.09 0.09 Surprisingly. The ranking of the main uncertainties is also the same. A steep marginal abatement cost curve implies for a coalition member a small share in coalitional abatement efforts and hence a lower incentive to free-ride. As the positive coefficient shows. the results are depicted in Table 7. Though the benefit share of Japan is significant in explaining the SL of the coalition between 123 .R.10 −0. the signs of their benefit coefficients reverse. As the USA is now a coalition member and the European Union is not. The significant and negative coefficients of the benefit shares of the USA. 3). former Soviet Union.06 0.14** USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW Standard deviation 0.10** Parameter Intercept Coefficient −0. they will want to join the coalition of Japan and the European Union.07 0. The significant inputs for the stability of this coalition are the same as for the coalition between Japan and the European Union.05 0.05 0.06 0.00 −0. except for the higher ranking of the benefit share of the USA. The high ranking of the European Union (3rd) shows that external stability can pose a problem in some cases: a high marginal benefit share in the European Union induces this region to join the coalition of the USA and Japan.16** 0. This means that Japan has a large incentive to collaborate.06 0.50 −0.06 0.06 0.06** 0.72 −2.49 0.07 0.05 Standard deviation 0. The only significant parameter for abatement costs is that of the European Union.08** 0.

32 0. confirming the finding above that the incentive for Japan to leave the coalition is in most circumstances (strongly) negative and thus cannot explain the variation in stability. Moreover.08 0.07 0.09 0. its ranking is rather low.09 −0.10** 0.07 0. (2006) : as benefits from abatement will occur later than the associated costs.13 Abatement cost parameter Note: ** indicates significance at the 5% level.07 −0.11** 0.13 0.08 0. The lower mean value of γ reduces equilibrium abatement.08 −0.06 Standard deviation 0. Therefore. In the All-Singletons Coalition Structure. Table 8 shows results assuming that the 5 percentile point is set to 0 $/tC and the 95 percentile point is set to 110 $/tC.08 −0. a lower mean value of γ has a similar impact as raising the discount rate as explained in Weikard et al.36 −1.07 0.05 0.09 −0.2%.07 0.18** 0.07 0. the SL is lower than in the base specification.08** 0.47 −1.09** 0. a higher discount rate implies a smaller weight of (future) benefits and vice versa.95 0.10 For all coalitions.07 0.17 0.07 0. Thus.24 Benefit shares 3. we can also mimic these results using the Ramsey rule. but only implicitly. the associated pure rate of time preference equals 5. estimates of future benefits from abatement are widely acknowledged to be highly uncertain.26 0.10 −0.18 −0. The specific values considered here correspond to raising the discount rate from 2 to 7 percent.3 Results for the alternative specification with lower mean global benefits We expect that the probability density function of the global benefit parameter γ is crucial for the analysis. pseudo R 2 equals 0.16** USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW Standard deviation 0. This is partly due to the lower probability of negative marginal benefits (this probability is reduced to 5%) and partly due to the fact that mostly positive outliers of marginal benefits are removed from the model.16** Parameter Intercept Coefficient 0.The stability likelihood of an international climate agreement Table 7 Results of the meta-analysis of the coalition between the USA and Japan Parameter Global benefits USA JPN EEC OOE EET FSU EEX CHN IND DAE BRA ROW Coefficient −8.06 −0.08** 0.07 0. global abatement drops from around 70 GtC per century to 45 GtC. 4. 123 . implying that only the quantitative but not the qualitative conclusions are affected. This implies that extreme values are rare and that the mean value of the global benefit parameter γ is reduced from 77 to 37 $/tC. The reduction 10 Analogous to the procedure explained in note 5.12 −0.11 0. the discount rate is not explicitly subjected to a sensitivity analysis.65 −1.10 0.07 0.05 −0.01 −1. we carry out an alternative simulation with a different assumption on the global benefit parameter γ . see also the discussion of equation [4a].02 0.09** 0.17 the USA and Japan.19** 0.60 −0. The ranking (in terms of SL) of different coalitions is largely unchanged.07 0. Note that in the STACO model.04 −0.07 0.

4.07 0.5 trillion $ versus 8.3 45.BRA JPN.05 1 0.5 47.8 Mean total payoff (bln $) 2520 2660 2790 2929 2532 2774 2844 2819 2543 2738 2894 8353 All-Singletons JPN. For the expected payoff in the case of learning we find.07 0. Hence. Dellink et al. In contrast. However.06 0.5 51. we can relate our results to the cases of no and full learning in Kolstad and Ulph (2006). resulting from the All-Singleton Coalition Structure.06 0.06 0. no coalition except the All-Singletons Coalition Structure would be stable.2 44. Once more.IND Grand Coalition 1 0. Lower benefits and abatement levels also lead to lower payoffs (2. a value of 2724 bln $. concerns the functional forms of the distribution of uncertain parameters. our different conclusion may rely exactly on this fact.9 51.11 0. Fig.15 0.EEC JPN.06 0. it has to be pointed out that Kolstad and Ulph’s assumptions imply that in the case of no learning at least a coalition of two countries is stable.01 0. Only if this value is raised by more than 20% does the coalition between Japan and the European Union emerge as stable coalition.01 0.BRA JPN. Table 8 Results for selected coalitions assuming smaller bounds on the global benefit parameter γ Coalition SL (fraction) Strict SL (fraction) Mean total abatement (GtC) 44. A second set of sensitivity analyses.02 0.EEC USA. 4.01 0. It is interesting to note that in a deterministic setting.00 in mean marginal benefits does not translate into a proportional reduction in mean abatement levels.1 52. the SL-value changes only from 20% to 24% if the mean value of γ is raised from 37 to 77 $/tC.JPN USA.8 trillion $) compared to the base specification. which is reported in Table 10. in contrast with the finding in Kolstad and Ulph (2006). A general conclusion that emerges from Tables 9 and 10 is that although the regional abatement levels and payoffs depend on several model parameters.06 0.R.1.ROW EEC. In the case of no learning.10 0.FSU EEC.ROW JPN. now learning has a positive impact on the global payoff. according to the procedure outlined in Sect. this highlights that the SL-concept is useful when conducting an analysis of stability of coalitions when parameter values are not exactly known.01 0.05 0. 1).20 0.1 225.01 0. the expected total payoff is 2520 bln $. This is due to the non-linear nature of the model: the quadratic marginal abatement cost functions imply that lower constant marginal benefits lead to less than proportionately lower abatement levels (cf. Again.06 0.0 49.02 0.2 50.4 Results of the sensitivity analyses In order to shed light on the robustness of our previous conclusions. The main results of alternative assumptions regarding the uncertainty of the main model parameters are reported in Table 9. with γ = 37 $/tC. the sensitivity of SL is 123 .IND EEC.7 49.2 55. a series of sensitivity analyses are carried out. Thus.

24 0.09 0.ROW EEC.BRA EEC.63 0. only the coalition between Japan and the European Union has a SL that is significantly higher than the probability of indifference.11 0.10 0.r.11 0.14 0.BRA JPN. Thus.20 0.09 0. when regional benefit shares are known. Essentially.10 0.10 0.t.10 0.10 0.ROW JPN.20 0.11 0.19 Gamma distribution αi .20 0.09 JPN.10 0.09 0.24 0.01 0 No uncertainty w.10 0.11 0.r. the same ranking emerges.01 0.11 0. Reducing the uncertainty with respect to the regional benefit shares from the model greatly reduces the variability of stability over the different coalitions.FSU EEC.09 0.10 0.10 0.10 0.20 0.09 0.14 0.01 0.10 0.20 0.06 0.JPN USA.32 0.The stability likelihood of an international climate agreement Table 9 Stability Likelihood of selected coalitions under different model specifications Coalition Base specification 0.06 0.EEC JPN.09 0.IND JPN.15 0.10 0. βi 0. and (ii) the variance and level of the global benefit parameter (γ ). si 0.EEC JPN.EEC USA.09 0.11 0.10 0.09 0.02 0.07 0. and this affects the SL of all coalitions.24 0.09 0.02 0.08 0.r.15 0.15 0. Then.26 0.10 0.17 0.13 0.19 0.t.02 0. βi 0.01 0.10 0.FSU EEC.JPN USA.09 No uncertainty w.11 0.24 0.10 0.07 0. followed by the coalition between the USA and Japan and the coalition between 123 .11 0.ROW JPN.09 0. γ 0.12 0.10 0.14 0.14 0.t.12 0. The ranking of different coalitions in terms of their SL is very robust: in all simulations. αi .24 0. Removing the uncertainty in the global benefit parameter implies that the probability of indifference is zero.FSU Grand Coalition Table 10 Stability Likelihood of selected coalitions under different model specifications Coalition Base specification Normal distribution si Normal distribution γ 0.09 0.14 0.09 0.09 JPN.01 much smaller.10 0.14 0.09 0.BRA EEC. the coalition between Japan and the European Union has the highest SL.15 0.10 0.24 0.09 si varies from unity 0.23 0.11 0.10 0.FSU Grand Coalition 0.08 0.11 0.IND JPN.10 0.19 0.ROW EEC.15 0. the value of SL depends crucially on two elements: (i) the variance in regional benefit shares (si ).10 0.11 0. it becomes much easier to predict which regions have incentives to collaborate and which do not.EEC USA.BRA JPN.09 No uncertainty w.

This included important coalitions such as the coalition of industrialised countries and the Grand Coalition. It became apparent that our approach corresponds to the case of full resolution of uncorrelated uncertainty with perfect learning. It turns out that this assumption was not important: results are very similar to the base specification. but was sensitive to the variance in the regional distribution of benefits. Therefore. and international research on adaptation and damage costs is still in its infancy. i. This is also confirmed in the last column in Table 10 where the assumption is given up that the sum of regional benefit shares is one. The lack of stability was very robust and held for a wide range of possible values for regional abatement costs and benefits. regional benefits from abatement are very uncertain.e. the USA and the European Union. in such cases.R. according to the classification in Kolstad and Ulph (2006). thereby rendering coalitions unstable. The coalition with the highest SL was the coalition of Japan and the European Union. The calculated SL of different coalitions was robust in terms of the considered distribution functions of parameter values. with an SL equal to 24% in our base case. We showed that learning may not always have a negative impact on global welfare. corresponding to the case of no cooperation: global abatement and global welfare were only slightly higher than when there was no international collaboration at all. This case could be compared to the case of no resolution of uncertainty where decisions on coalition membership have to be taken on the basis of expected values. Unfortunately. We also related our results to those derived from stylized models with ex-ante symmetric players which conclude that learning has a negative impact on the outcome of coalition formation. when climate change does not pose a problem. we can conclude that SL provides a highly robust indicator of the stability of climate coalitions. We introduced a methodology to calculate the stability of all possible coalition structures in a stochastic. Dellink et al. as described by the respective probability distribution functions. 5 Discussion This paper investigated the formation and stability of coalitions to reduce the negative impacts of climate change. will regions be indifferent between cooperation and no cooperation. This suggested that the conclusions derived from stylized models as for instance in 123 . but also with the sum of the regional shares. The concept of SL provided a notion of stability that conveyed much more relevant information than the ordinary binary outcome: stable/unstable. Thus. empirical setting. confirming a conclusion in the literature derived with deterministic models. There were only a small number of coalitions with a (strict) SL significantly different from zero. The numerical application showed that the SL-concept contributes to a better understanding of stability of coalitions when there is uncertainty about model parameters: it constitutes a superior alternative to running numerous sensitivity analyses in a deterministic setting. Thus. Only when benefits from abatement are zero or negative (calibrated to be around 9 percent in the base specification). However. it is of utmost importance to get better information on regional benefits. This coalition but also the other coalitions with a positive SL hardly improved over the All-Singletons Coalition Structure. the variance in regional abatement costs and the global level of benefits from global abatement. This relatively low percentage stressed the difficulties in striking an international environmental agreement. regardless of which coalitions formed. global benefit levels vary not only with the variations in the global benefit parameter. The results suggested that for most possible climate coalition there is at least one region that is better off by changing its actions. the optimal abatement levels were zero.

Kolstad and Ulph (2006) and Kolstad (2007) are sensitive to the underlying assumptions. Firstly. Histograms of the uncertain model parameters Fig. Thirdly.The stability likelihood of an international climate agreement Na and Shin (1998).2 Histogram of the regional benefit parameters (si ) for USA and China. respectively 123 . Finally. A. the empirical model used to calculate the regional payoffs can be extended to a fully-fledged computable general equilibrium model. Appendix 1. we expect that these two extensions will improve the empirical validity of the numerical example. Secondly. There are many ways in which the analysis above can be ameliorated. better estimates of the variance on regional benefits may be obtained via a meta-analysis of the existing literature. Together. the analysis could be extended to decision making under uncertainty by including risk-aversion of players as for instance considered in Endres and Ohl (2003). it would be interesting to consider the case of partial learning as analysed in Kolstad and Ulph (2006) in order to better understand the role of learning. similar to what Tol (2005) did for global benefits. A.1 Histogram of the global benefit parameter (γ ) Fig.

wur. London Finus M. van Ierland EC. available on the STACO website http://www. Mayer M. MIT. SueWing I. Buchner B. 71. J Environ Econ Manage 53:68–79 Kolstad CD. the usual disclaimer applies. Austria 123 .enr.nl/uk/staco. Finus M. Joint program on the science and policy of global change Report. Wageningen University Ellerman AD. paper presented at the conference “Learning and climate change”. Econ Govern 7:271–291 Hoel M (1992) International environment conventions: the case of uniform reductions of emissions. Reilly JM. van Ierland EC. Joint program on the science and policy of global change Report. Altamirano J-C (2004) Empirical background paper of the STACO model. respectively Acknowledgements Special thanks to Ivo Mulder and Kelly de Bruin for their research assistance and to Ekko van Ierland. Carraro C (2003) Equity. John Janmaat and two anonymous referees for their comments. A. Cheltenham 26–55 Carraro C. Environ Resour Econ 2:141–159 Kolstad CD (1996) Fundamental irreversibilities in stock externalities. References Babiker MH. Nat Resour Model 19:323–358 Fankhauser S (1995) Valuing climate change. Carraro C (1997) Burden-sharing and coalition stability in environmental negotiations with asymmetric countries. MIT. IIASA. development. Dellink et al. sensitivities and comparison of results. Oxford Economic Papers 46:878–894 Bosello F. Fig.R. Finus M (2006) Coalition formation in a global warming game: how the design of protocols affects the success of environmental treaty-making. In: Carraro C (ed) International environmental negotiations: strategic policy issues. Dellink RB (2006) Stability of climate coalitions in a cartel formation game. Cambridge Barrett S (1994) Self-enforcing international environmental agreements. J Environ Econ Manage 31:1–18 Kolstad CD (2007) Systematic uncertainty in self-enforcing international environmental agreements. Eligius Hendrix. Ohl C (2003) International environmental cooperation with risk aversion. Ulph A (2006) International environmental agreements under uncertainty when learning reveals differences between countries. Hyman RC (2001) The MIT emissions prediction and policy analysis (EPPA) model: revisions. vol 40. J Publ Econ 60:221–233 Kolstad CD (1996) Learning and stock effects in environmental regulation: the case of greenhouse gas emissions. Earthscan.3 Histogram of the regional abatement cost parameter (αi ) for USA and China. J Euro Econ Assoc 1:601–611 Botteon M. Siniscalco D (1993) Strategies for the iternational potection of the environment. Cambridge Endres A. Eckaus RS. J Publ Econ 52:309– 328 Dellink RB. Decaux A (1998) Analysis of post-Kyoto CO2 emissions trading using marginal abatement curves. Int J Sustain Develop 6:378–392 Eyckmans J. and climate change control. Edward Elgar. Laxenburg.

Mansholt Discussion Paper 23. Swart R. J Risk Uncertain 29:53–73 Ulph A. Altamirano-Cabrera JC (2006) The impact of surplus sharing on the stability of international climate coalitions. Boyer J (2000) Warming the world: economic models of global warming. Dellink RB. PhD. Oxford Econ Papers 58:209–232 Weitzman ML (1998) Why the far-distant future should be discounted at its lowest possible rate. Richels RG (1995) Merge:a model for evaluating regional and global effects of GHG reduction policies. learning and international environmental policy coordination. MIT Press. Energy Policy 27:415–429 Tol RSJ (1997) A decision-analytic treatise of the enhanced greenhouse effect. Cambridge Olieman NJ. Maddison D (1997) Uncertainty. Oxford Economic Papers 50:173–185 Nagashima M. Energy Policy 23:17–34 Metz B. Shin HS (1998) International environmental agreements under uncertainty. irreversibility and learning. Econ J 107:636–650 Weikard H-P.The stability likelihood of an international climate agreement Manne AS. Environ Resour Econ 9:451–466 Ulph A. Cambridge Na S. Pan J (eds) (2001) Climate change 2001: mitigation. Environ Resour Econ 21:47–73 Tol RSJ (2005) The marginal damage costs of carbon dioxide emissions: an assessment of the uncertainties. Finus M. Vrije Universiteit. MIT Press. Computat Econ 18:159–172 Tol RSJ (2002) Estimates of the damage costs of climate change: part 1: benchmark estimates. Cambridge University Press. Thesis. J Environ Econ Manage 36:201–208 Weitzman ML (2001) Gamma discounting. Wageningen Nordhaus WD (1994) Managing the global commons. van Ierland EC (2006) Dynamic transfer schemes and stability of international climate coalitions. Amsterdam Tol RSJ (2001) Climate coalitions in an integrated assessment model. Ulph D (1997) Global warming. Energy Policy 33:2064–2074 Ulph A (2004) Stable international environmental agreements with a stock pollutant. Euro J Oper Res 174:333–348 Roughgarden T. Am Econ Rev 91:260–71 123 . uncertainty and learning. Hendrix EMT (2006) Stability likelihood of coalitions in a two-stage cartelgame: an estimation method. Schneider SH (1999) Climate change policy: quantifying uncertainties for damages and optimal carbon taxes. Mendelsohn R. Davidson O. Cambridge Nordhaus WD.

Sign up to vote on this title
UsefulNot useful