Escaping Damocles' Sword: Endogenous Climate Shocks in A Growing Economy

Environmental and Resource Economics
https://doi.org/10.1007/s10640-023-00835-w
Escaping Damocles’ Sword: Endogenous Climate Shocks

in a Growing Economy
Alexandra Brausmann1 · Lucas Bretschger2
Accepted: 19 December 2023

© The Author(s) 2024
Abstract
We consider a growing economy which is subject to recurring, random, uninsurable, and
potentially large and long-lasting climate shocks leading to destruction of infrastructure,
land degradation, collapse of ecosystems or similar loss of productive capacity. The asso-
ciated damages and the hazard rate are endogenously driven by the stock of greenhouse
gases. We highlight the important role of the relative risk aversion and provide analytical
solutions for the optimal climate policy, the growth rate and the saving propensity of the
economy. We stress the importance of jointly determining these variables, especially if the
objective is to formulate meaningful policy prescriptions. If, for example, the growth rate
or the saving rate are assumed to be exogenous, and thus independent of the characteristics
of climate shocks and economic fundamentals, then future economic developments in the
face of climate change and, consequently, the future mitigation efforts will deviate from
the optimal paths. In a quantitative assessment we show that with log-utility and under
favorable technological and climatic conditions the abatement expenditure represents only
0.5% of output, equivalent to $37 per ton carbon. Under less favorable conditions, coupled
with a relative risk aversion which exceeds unity, the abatement propensity increases to
2.9%, equivalent to $212 per ton carbon, and it jumps to a striking 16% in the pessimistic
scenario involving severe shocks and low efficiency of abatement technology.
Keywords Climate policy · Uncertainty · Natural disasters · Endogenous growth
JEL Classification O10 · Q52 · Q54
A draft version of this paper has circulated under the title "Growth and Mitigation Policies with
Uncertain Climate Damage". The paper has changed substantially compared to that draft version. We
thank Aleksey Minabutdinov, Amos Zemel, Christos Karydas, Phoebe Koundouri, and the participants
of the SURED2016 and EAERE2016 conferences for helpful comments and fruitful discussions.
* Lucas Bretschger
[email protected]
Alexandra Brausmann
[email protected]
1
Department of Economics, University of Vienna, Vienna, Austria
2
Center of Economic Research, CER‑ETH, ETH Zurich, Zurich, Switzerland
13
Vol.:(0123456789)
A. Brausmann, L. Bretschger
1 Introduction
1.1 Economics and the Climate
The search for an optimal policy response to climate change has preoccupied the minds
of economists for over three decades. While pricing carbon at its social cost, so as to limit
carbon emissions, is a generally accepted policy proposal, a global consensus on an opti-
mal carbon price or an efficient mitigation level is still lacking. It is nonetheless widely
acknowledged that a comprehensive approach should include consideration of important
risks and uncertainties associated with global warming. What is especially troublesome is
that planet’s excessive warming may lead to severe and unforeseeable natural catastrophes
(IPCC 2014) the impact of which may be long-lasting or even irreversible (Lenton and
Ciscar 2013). Such catastrophes may result in potentially huge damages to economic activ-
ity, biodiversity, eco-systems and geophysical processes, e.g. loss of productive capacity,
loss of plants and species varieties, melting of ice sheets, changes in El-Nino and Atlantic
thermohaline circulation. The focus of the present paper is on the design of climate policy
in the context of such natural disasters.
The threat of severe catastrophes puts the world economy in a situation similar to
the one of Damocles, an obsequious courtier of King Dionysius of Syracuse in the 4th
century BC. According to the moral anecdote, Dionysius offered Damocles to sit on his
throne in order to taste the fortune of a great man of power and authority, an opportu-
nity which Damocles eagerly accepted. The throne was surrounded by every luxury but the
King arranged that a sword should hang above it, held by a single hair of a horse’s tail. An
impending calamity restrained by some probability can be used as an allegory for the world
population facing the climate catastrophe. The analogy ends with an important difference
however: Damocles had no power to deal with the imminent disaster, while the planet’s
warming can be controlled by our current policies.
The policy response to climate change needs to be informed by appropriately-designed
economy-climate models. However, the complexity of both the economic and the climate
parts of the problem pose considerable modeling challenges. Many of them are being
addressed by relatively complex numerical Integrated Assessment Models (IAMs), which
have provided many useful insights with respect to the social cost of carbon (SCC) but also
produced diverging results.1 To gain further insights into the central mechanisms at work—
especially those related to the uncertainties and risks associated with climate change and
its economic impact—a framework of investigation that relies on explicit analytic solutions
and provides clear-cut implications for the optimal climate policy in a dynamic economy
appears to be desirable.
Within such a framework, a number of important questions need to be addressed. Given
the uncertain nature of environmental disasters caused by climate change, how should an
economy appropriately balance its production, consumption, investment, and reduction
of greenhouse gas (GHG) emissions? What is the optimal rate of output growth and the
optimal emissions abatement in the uncertain environment? How do these key variables
interact and respond to changes in the underlying economic and climatic fundamentals?
In the present paper we examine these questions within a model of a growing economy
1
Deterministic IAMs, such as the widely used DICE (Nordhaus 2008), produce lower estimates of SCC
than IAMs which include uncertainty, such as e.g. Lontzek et al. (2015), Cai and Lontzek (2019). Estimates
of SCC are also sensitive to the modeling of the climate system (Dietz and Venmans 2019).
13
Escaping Damocles’ Sword: Endogenous Climate Shocks in a Growing…
which features uncertainty about arrivals of climate shocks. We assume that an occurrence
of a disaster (also referred to as an "event") follows a random process, and when a disaster
strikes, some part of the economy’s productive capacity is destroyed. The magnitude of
the damage is assumed to be an increasing function of the stock of GHG. It follows that
the process of accumulation of the productive capacity is both endogenous and stochas-
tic. Natural catastrophes induced by climate change are large in scale and have a profound
negative impact on both national and global economy. Unlike in the case of relatively small
idiosyncratic shocks, the risk of such events cannot be insured.2
In our model the world neither ends nor becomes deterministic after an environmen-
tal disaster, as it is often assumed in the literature on irreversible catastrophic events (see
Sect. 1.2). We consider development with recurring shocks over time, which reflects a
likely pattern of climate-induced events in the future.3 Optimal reduction in emissions, and
the implied reduction in damages, can be achieved by appropriately balancing two types of
activities: capital accumulation and abatement. In an extension of the model we endogenize
the event hazard rate, in addition to damages, by letting it be an increasing function of pol-
lution stock (IPCC 2014).
1.2 Related Literature
There is a growing literature on random catastrophic events causing irreversible damage.4

Tsur and Zemel (1998) are the first to explicitly analyze reversible events focusing on the
optimal steady state policy and transitional dynamics of an economy which is not engaged
in any investment activity.5 Van der Ploeg (2014) analyzes the optimal carbon tax in an
economy subject to a random shock which reduces the nature’s capacity to absorb green-
house gases. Van der Ploeg and de Zeeuw (2018) examine policy adjustment with respect
to how fast a tipping point may materialize.6 Several recent contributions incorporated a
2
In a globalized world, large-scale natural disasters affect not only the economic activity of the country
where they strike but also other economies by virtue of either close geographic location or trade relations,
FDI, etc. When it comes to relatively small idiosyncratic shocks, they can be insured against by trading
insurance claims within a group of regions subject to such shocks. Our focus, however, is on a global econ-
omy where an insurance contract against a large natural catastrophe (e.g., Indian Ocean tsunami, meltdown
of Greenland ice-sheet, etc.) is challenging to design and implement. Additionally, stability of contrac-
tual arrangements on such a large scale seems to be difficult to ensure, especially in light of the long time
horizon and asymmetric probabilities of disasters across countries. Insurance companies seem to perceive
climate change and the associated disasters as a threat rather than profitable business opportunities. As a
prominent example, the strategy of one of the world’s leading reinsurance companies, Swiss Re, is to raise
awareness about climate-change risks through dialogue with clients, employees, and the public and to advo-
cate a worldwide policy framework for climate change (Swiss Re 2015, p.119,ff.).
3
Recurring climate-induced capital destruction may result in a new form of poverty traps for developing
countries. See Le Van et al. (2010) for an analysis of poverty traps where the natural environment offers
opportunities while in our case it may pose a threat to development.
4
For early theoretical contributions see Clarke and Reed (1994), Tsur and Zemel (1996) and for recent
numerical models Lemoine and Traeger (2014), Lontzek et al. (2015).
5
de Zeeuw and Zemel (2012) provide a dynamic characterization of an optimal emission policy when the
time of the regime switch from low to high damage is uncertain. One of their key findings is that, due to
precautionary reasons, emissions in the low-damage regime may be lower than in the case where the system
is already in the high-damage regime.
6
In a model with uncertain pollution stock dynamics, Athanassoglou and Xepapadeas (2012) find that
investment in damage control is increasing in the degree of uncertainty and that control might be a substi-
tute for mitigation when it is sensitive to changes in uncertainty.
13
possibility of crossing a tipping point into integrated assessment modeling (e.g. Lemoine
and Traeger 2014; Lontzek et al. 2015). Cai and Lontzek (2019) use DSICE, a stochastic
IAM, to analyze the effects on SCC of (economic) uncertainty in productivity coupled with
(climatic) uncertainty about triggering a tipping process. The results from these studies
unambiguously indicate that the current SCC should be larger by at least a factor of two,
as compared to deterministic IAMs, when the possibility of an irreversible tipping is taken
into consideration.
Our work relates to several important papers in the field of climate and growth eco-
nomics. Mueller-Fuerstenberger and Schumacher (2015) developed a decentralized ver-
sion of a neoclassical growth model where agents face stochastic extreme events which are
relatively small-scale and thus insurable. A very interesting and policy-relevant result of
their paper is that the insurance industry, being the main provider of reactive adaptation,
can induce agents to undertake more abatement by signaling the consequences of climate
change through insurance premiums.7 In a recent paper, Gerlagh and Liski (2017) intro-
duce uncertain events in a climate-economy model where the impacts of climate change
are not known but learned over time. As a consequence of the belief updating the optimal
carbon tax does not develop in lock-step with income but depends on temperature levels.
Dietz and Venmans (2019) incorporate the most recent advances in climate science with
respect to the mapping from CO2 release into temperature change. They show that mini-
mizing both the abatement costs and damages requires a much more aggressive emissions
reduction policy than what has been recently suggested by Lemoine and Rudik (2017)
based on a model with strong inertia in the climate system.8 Our paper also relates to recent
contributions on endogenous growth under environmental restrictions, in particular Peretto
and Valente (2015) and Lanz et al. (2017), where economic dynamics constitute the center
of the analysis, while we focus more on the environmental policy aspect.
Finally, we would like to highlight important links between our macroeconomic
approach and the extensive microeconometric and case study literature deriving the param-
eters we use in our quantitative analysis. In a recent empirical study on alterations in risk
aversion, Liebenehm et al. (2023) combine individual-level panel data with historical rain-
fall data for rural Thailand and Vietnam and find that rainfall shocks increase individu-
als risk aversion; this confirms the "risk vulnerability" hypothesis proposed by Gollier and
Pratt (1996), which we will adopt when modelling the effects of climate shocks in our
analysis. Brown et al. (2018) studied the effect of the 2012 cyclone on Fijian households’
risk attitudes, they equally find that being struck by an extreme event substantially changes
individuals’ risk perceptions as well as their beliefs about the frequency and magnitude
of future shocks. That risk preferences are heterogeneous among individuals, changing
over time or with policy, is a main result of Koundouri et al. (2009) who analysed Finnish
7
Ikefuji and Horii (2012) examine the optimal growth rate and the carbon tax in an economy where pri-
vate capital is subject to stochastic depreciation due to climate change. They assume, however, that these
stochastic shocks are idiosyncratic and reflect a large number of independent small climate events. Soretz
(2007) analyzes efficient pollution taxation within an endogenous growth model where environmental qual-
ity has a stochastic impact on factor productivity, which is driven by a Wiener process. Similarly to these
contributions we assume that international policy coordination is feasible, as has recently been demon-
strated by adoption of the Paris Agreement. If such a mechanism turns out to be inefficient, it has been sug-
gested to replace global carbon taxes by international trading of the rights to exploit fossil-fuel deposits, see
Harstad (2012).
8
A recent paper by Mattauch et al. (2020) presents further arguments against the conclusions of Lemoine
and Rudik (2017).
13
farm-level data. In a broad survey experiment on risk attitudes Dohmen et al. (2011) find
that gender, age, height, and parental background have a significant impact on the will-
ingness to take risk.9 Our macroeconomic model does not distinguish between different
household types as it would go beyond the scope of the present paper; but we see a promis-
ing future research avenue when applying some of our model elements in a multi-region
approach, with local differences in climate vulnerability and climate events, which would
also enable to extend regional policy analysis under uncertainty (see Brock and Xepa-
padeas 2020, and Englezos et al. 2022). Uncertainty may also affect social discounting.
Like our paper, Hepburn et al. (2009) argue that over longer horizons, as in the case of
climate change, the topic of uncertainty becomes more and more important. These authors
estimate different models of real interest rates to determine certainty-equivalent discount
rates in Australia, Canada, Germany and the UK and find that social cost-benefit analysis
should be conducted with a declining discount rate to place a substantially higher weight
on events in the distant future. Our framework argues along the same lines but uses a sim-
pler approach by assuming low social discount rates from the very beginning.
1.3 Contribution to the Literature
The present paper contributes to formulating efficient climate policies using an approach
which has been identified as an urgent priority in recent assessments of the field. Farmer
et al. (2015) conclude that the first of four major issues inadequately addressed by eco-
nomic models of climate change is uncertainty. An even more stringent view is expressed
by Stern (2007, 2015, 2016) who states that many economic models do not account for
catastrophic changes and possible tipping points which would make a crucial difference for
policy assessment in his view. Addressing these issues forms the core of our contribution.
To the best of our knowledge our paper is the first to provide closed-form solutions
for the abatement policy and the optimal growth rate of the economy subject to random
climate shocks with damages endogenously driven by the accumulation of GHG.10 Our
climate-policy instrument can be conveniently expressed as a fraction of output which
depends on the fundamental characteristics of the climate and the economy. This result
is parallel to that of Golosov et al. (2014), who derive a simple formula for the optimal
carbon tax, showing that under specific simplifying assumptions on damages and sav-
ing propensity the tax is proportional to GDP and depends on just a small number of key
parameters. In the present paper we relax some of those assumptions and, in particular,
endogenously derive the optimal saving rate of the economy to show how it is affected by
exposure to uncertainty associated with climate change.
The second contribution of the paper is to highlight the importance of joint determi-
nation of the climate policy, the growth rate and the saving rate of the economy. If, for
example, the saving rate or the growth rate are assumed to be constant and exogenous (as,
e.g., in Golosov et al. 2014 and Barro 2015, respectively), then by construction they are
independent of the economic and climatic characteristics and therefore fail to appropri-
ately reflect their optimal counterparts. When the mitigation policy relies on these assumed
9
The heterogeneity of risk attitudes is confirmed by Drichoutis and Koundouri (2012); the authors high-
light the importance of adopting a proper characterization of risk attitudes when using experimental tech-
niques.
10
Bretschger and Vinogradova (2019) analyzed the optimal abatement policy in a simplified model which
did not include any climate dynamics.
13
exogenous rates, it will deviate from the welfare-maximizing outcome. If the growth rate of
consumption can endogenously respond to changes in climatic characteristics, let’s say an
increase in the frequency of shocks, the optimal climate policy may become less stringent
than in the case with the fixed growth rate because the economy will optimally choose to
grow at a lower rate. This is because the growth rate and the associated accumulation of
productive capacity determine (i) the growth of the stock of GHG, (ii) security buffer in
the event of a climatic hazard, i.e. how much loss the economy can withstand, and (iii) how
much resources can be devoted to mitigation.
Third, and extending the contribution of Barrage (2014), we show analytically that
when preferences are logarithmic, important links to climatic characteristics, including the
arrival rate of disasters, efficiency of abatement technology and damage intensity, disap-
pear from the optimal rules.
While the closed-form solution allows us to clearly disentangle the effect of each cli-
matic and economic characteristic of the model, we are also interested in quantitative pre-
dictions with respect to the optimal policies. Numerous existing studies proposed required
emissions-mitigation schemes (or a carbon tax) to ensure that the global temperature does
not exceed a specific threshold. Other studies asked what the optimal carbon tax should be,
given the various types of uncertainty associated with climate change (Gerlagh and Liski
2017; Golosov et al. 2014; Van der Ploeg 2014; Van der Ploeg and de Zeeuw 2018; Van
Den Bremer and van der Ploeg 2018). Our investigation falls into the second category. One
of the frequent messages delivered by the recent literature is that the utility discount rate
appears to be the main driver of the optimal tax but not a possibility of a climate disaster.11
This observation motivates us to quantitatively assess the impact of possible multiple and
random climate shocks on the optimal growth rate, abatement propensity and, by exten-
sion, on the price of carbon.
We examine several scenarios based on alternative assumptions about abatement tech-
nology, risk aversion, size of damages and probability of occurrence. The gist of our find-
ings is that it matters whether disasters are low-impact (causing a loss of, say, not more
than 0.1% of GWP) occurring relatively frequently or high-impact (loss of up to 10% of
GWP) even if they occur relatively rarely, say, less than once in a 100 years. On the one
hand, low-impact shocks do not require a substantial increase in abatement efforts even
with a relatively high RRA. On the other hand, we find that it is the simultaneous reaction
of both abatement and growth which is decisive. The growth rate of the economy essen-
tially works as a stabilizer of emissions and thus to some extent substitutes for the climate
policy instrument. With a lower growth rate, less stringent abatement policy is needed.
With high-impact shocks the abatement propensity increases almost 6-fold and the growth
11
Golosov et al. (2014) calculate the optimal carbon tax to be $25.3 and $489 per ton carbon when the
realized damages amount to 0.48% and 30% of GDP (based on Nordhaus and Boyer 2000), respectively,
assuming a 1.5% discount rate and log- utility. With a much lower discount rate of 0.1% (advocated by
Stern 2007), the tax jumps up 8–9 fold to $221 and $4,263, respectively. In another important contribution,
Van der Ploeg (2014) considers a possibility of a climate catastrophe, such that a positive carbon feedback
occurs at some random (and possibly endogenous) point in time. He finds that the optimal after-catastro-
phe carbon tax is $29 and $216 per ton carbon with the discount rate of 1.5% and 0.1%, respectively. The
before-catastrophe tax is slightly lower at $27.6 and $198.6. When the hazard rate is endogenous, the tax is
only slightly higher. One exception is Van den Bijgaart et al. (2016) who show in the context of a determin-
istic IAM that damage sensitivity plays just as important role as discounting in determining the social cost
of carbon.
13
rate is cut by a factor of 3. The quantitative section also provides a comparison of our find-
ings with those in the recent studies.
The remainder of the paper is organized as follows. In order to set the stage for our
stochastic growth model with climate-driven disasters, we start by presenting in Sect. 2
a deterministic counterpart of this model and its full analytical dynamics. Then Sect. 3
develops our baseline stochastic framework. In Sect. 4, we present the main results with
respect to the optimal growth rate, abatement, and saving propensity. Section 5 provides an
extension of the baseline model by endogenizing the hazard rate. Section 6 offers quantita-
tive implications and Sect. 7 concludes.
2 Deterministic Model
We start by presenting a deterministic model, similar to the one in Sect. 6 of Michel and
Rotillon (1995), which we are going to use as our benchmark. We consider a global econ-
omy which produces a composite consumption good under constant returns to scale using
as input broadly defined capital, denoted by Kt . The production process is polluting: every
instant t a flow of greenhouse gas emissions, Et , is released into the atmosphere. The stock
of GHG, denoted by Pt , is thus augmented every instant by Et and reduced by 𝛼Pt , where
𝛼 ∈ [0, 1) represents the natural absorption rate of greenhouse gases (e.g., by biosphere or
deep oceans) and is assumed to be very small or can be set to zero.12 Cumulative emissions
cause deterioration of the natural environment and an increase in the global temperature,
resulting in a loss of welfare.13 In our model, we refer to P as pollution stock, however, it
can be interpreted more generally as the reciprocal of environmental quality. In the same
vein, we refer to E as emissions but it can be also interpreted as any type of negative exter-
nality of production which harms the environment.
The output, denoted by Yt (Kt ), can be either spent on consumption, Ct , or invested.
There are two types of non-consumption spending: (i) investment to augment the capital
stock and (ii) financing of emissions abatement. Specifically, we assume that an endog-
enous fraction, 𝜃t , of output is spent on the latter, so that abatement expenditure is given
by It = 𝜃t Yt . The remaining share (1 − 𝜃t )Yt is split between consumption and capital
accumulation. Total abatement, Z(It ), is a positive function of the abatement expenditure,
Z � (It ) > 0. The total per period emissions are then given by emissions stemming from the
12
Such parsimonious one-layer formulation of the carbon system has been widely used in the literature
(Clarke and Reed 1994; Tsur and Zemel 1998; de Zeeuw and Zemel 2012; Van der Ploeg 2014; Van den
Bijgaart et al. 2016; Van der Ploeg and de Zeeuw 2018). Integrated assessment models typically use a more
elaborate representation of the carbon cycle, with atmosphere, upper ocean layers and deep oceans as the
three main carbon reservoirs. Nordhaus and Boyer (2000) calibrate the transfer rate from the atmosphere
to the upper ocean layer as 0.333 per decade and from the upper oceans to the deep oceans as 0.115 per
decade, implying an indirect transfer rate from the atmosphere to deep oceans of approximately 0.0038 per
year, which is proxied by 𝛼 in our model. Golosov et al. (2014) argue that a one-dimensional formulation of
the carbon cycle can well approximate the cycle used in IAMs if the depreciation structure includes a term
reflecting the share of a carbon unit which stays permanently in the atmosphere and another term with a
geometric decay rate applied to the share of carbon which does not exit the atmosphere immediately.
13
We do not explicitly model temperature increase as it is now widely confirmed by the Earth system mod-
els that warming is proportional to cumulative emissions and that temperature response to a CO2 emission
is almost instantaneous and then constant as a function of time (see Dietz and Venmans 2019 and references
therein).
13
economic activity minus abatement. We assume that one unit of output causes 𝜙 units of
GHG,14 so that total emissions are given by Et = 𝜙Yt − Z(It ).
The planner’s objective is to maximize the expected discounted utility over an infinite
planning horizon with respect to consumption, Ct , and the share of output devoted to abate-
ment, 𝜃t , subject to the stochastic capital accumulation process and the dynamics of the
cumulative emissions:
{ ∞ }
∫0 (1)
−𝜌t
max U(Ct , Pt )e dt
Ct ,𝜃t
s.t.
dKt = [(1 − 𝜃t )Yt (Kt ) − Ct ]dt, K0 given, (2)
dPt = (Et − 𝛼Pt )dt, P0 given, (3)
Et = 𝜙Yt (Kt ) − Z(It ), (4)
It = 𝜃t Yt (Kt ), (5)
where 𝜌 > 0 is the constant rate of time preference. The utility function is twice con-
tinuously differentiable in both arguments. We also require that the capital stock, pollu-
tion stock, and consumption per unit of time are non-negative and 𝜃t ∈ [0, 𝜃] ̄ , where
𝜙∕𝜎 < 𝜃 < 1.
̄
2.1 Assumptions
1. Production Technology
We assume constant returns to scale in aggregate production and as capital is the only
input, output is produced with an AK technology, a frequently adopted specification in cli-
mate-economy models (see, e.g., Mueller-Fuerstenberger and Schumacher 2015). Param-
eter A denotes the constant factor productivity and Kt is interpreted as a broad measure of
capital in the economy, including physical and human capital, intangibles, etc.15
Yt = AKt . (6)
2. Abatement Technology
We also assume constant returns in abatement, i.e. total abatement is directly propor-
tional to the resources allocated to emissions control, with the proportionality parameter
𝜎 > 0 representing the efficiency of abatement technology:
Z(It ) = 𝜎It . (7)
14
With appropriate choice of units, polluting intensity 𝜙 can be restricted to lie between zero and unity.
15
Despite its formal simplicity, the AK model unites all the desirable properties of an aggregate production
function in a dynamic climate model. It generates sustained growth endogenously, results in the same impli-
cations for investment and growth as if we included different capital components such as physical, human,
and knowledge capital separately, and is consistent with the empirically observed strong positive relation-
ship between investment rates and growth rates across countries and time periods (see McGrattan 1998).
13
3. Preferences
How to represent society’s preferences over consumption and environmental quality
has long been a cornerstone question for economists. Over the last few years of research
three main trends have emerged. First, the non-expected utility and homothetic preferences
of Epstein-Zin type (Epstein and Zin 1989) have gained momentum (Barro 2015; Lem-
oine and Traeger 2014; Van der Ploeg and de Zeeuw 2018; Van Den Bremer and van der
Ploeg 2018, Douenne 2020) due to their ability to disentangle the elasticity of intertem-
poral substitution (EIS) from the relative risk aversion (RRA) coefficient.16 Second, loga-
rithmic preferences have been widely used due to their tractability (Golosov et al. 2014;
Mueller-Fuerstenberger and Schumacher 2015). The third option, adopted here (as well as
in Barrage (2014), Dietz and Venmans (2019), Van der Ploeg and de Zeeuw (2018)), is to
assume a constant relative risk aversion (CRRA) utility function, also encompassing the
logarithmic version as a special case. We assume that utility is additively-separable in con-
sumption and environmental quality, with the latter being inversely related to the stock of
GHG. Finally, the utility function is increasing and concave in consumption and decreasing
and concave in the stock of GHG, exhibiting risk aversion with respect to both arguments:
Ct1−𝜀 P1+𝛽
U(Ct , Pt ) = −𝜒 t
, 𝜀 > 0, 𝛽 > 0, 𝜒 > 0 (8)
1−𝜀 1+𝛽
where 𝜒 is the relative weight of pollution. We will show in Sects. 3 and 5 the importance
of considering a general CRRA structure, as opposed to its knife-edge case of log-utility,
which ignores important effects in a climate-growth context.
2.2 Solution
We are going to distinguish three cases depending on the initial level of P0 and K0. These
initital conditions will determine the behavior of the optimal abatement, for which we will
have a case with an interior solution and two border cases such that it is optimal to abate
nothing or the maximum amount.
The current-value Hamiltonian associated with the optimization problem is given by
H = U(Ct , Pt ) + 𝜆t [(1 − 𝜃t )AKt − Ct ] + 𝜇t [(𝜙 − 𝜎𝜃t )AKt − 𝛼Pt ]
and the associated Lagrangian by L = H + 𝜐t (𝜃̄ − 𝜃t ) + 𝜈t 𝜃t , where we explicitly attached

a non-negativity constraint on 𝜃 and a constraint specifying that 𝜃 cannot exceed some
maximally feasible abatement propensity 𝜃̄ < 1. The linearity of H in 𝜃 makes the model
analytically tractable and leads to the existence of the three above-mentioned abatement
regimes: 𝜃 = 0, 𝜃 = 𝜃̄ and an interior regime 𝜃 = 𝜃 ∗, which we descibe below. We show
that maximal/zero abatement policy is applied for a finite time only and then the dynamics
is switched to the interior regime, which is of our primal interest (we include border abate-
ment regimes for completeness in the appendix). This is why we are going to focus only on
the interior regime when we introduce our stochastic model in Sect. 3.
16
This property, however, comes at a cost. It has been recently shown that Epstein-Zin preferences are non-
monotonic with respect to the first-order stochastic dominance meaning that under some parameter constel-
lations they lead to choices of dominated strategies (Bommier et al. 2017).
13
2.2.1 The Interior Solution
The dynamics of the economy in the interior regime are characterized by the following
equations:
− r−𝜌
Pt = P0 e 𝛽
t
. (9)
( )
𝜀g
c P 0 𝛼 − 𝛽 − 𝜀g t
Kt∗ = 0 egt − ( ) e 𝛽 (10)
r−g 𝜎 r + 𝜀g 𝛽
( )
𝜀g
𝜙 − 𝛼 Pt
𝜃t∗ = +
𝛽
, (11)
𝜎 𝜎AKt∗
( 𝜎𝜒P𝛽 )−𝜀
ct = c0 egt , c0 = 0 (12)
r+𝛼
where g = 1𝜀 (r − 𝜌), and it is assumed that g < r . Equation (11) describes the dynamics of
the abatement propensity. We would like to make two comments on these dynamics. First,
the abatement propensity converges to 𝜙𝜎 in the long run. This is because pollution declines
at the rate gp ≡ 𝜀g 𝛽
over time and thus converges to zero in the long run. Second, if the natu-
ral decay rate is relatively large, in particular if it is larger than the rate of pollution decline,
i.e. 𝛼 > 𝜀g𝛽
, then the abatement propensity converges to its limit from below. The intuition
is that a large natural decay “helps” the economy to clean up emissions and while the pol-
lution stock is large, this effect is also large. However, as the pollution stock falls over time,
the natural pollution removal also becomes smaller. Thus, initially the abatement propen-
sity is below the zero-emissions threshold 𝜙𝜎 and it increases over time as pollution stock
falls. If the natural decay is relatively small, in particular it is smaller than the rate of pollu-
tion decline, i.e. 𝛼 < 𝜀g𝛽
, then the opposite occurs. The abatement propensity converges to
the threshold 𝜙𝜎 from above. The intuition is that with a small 𝛼 (think about zero natural
decay as an extreme case) all of the pollution removal takes place through abatement.
Hence, abatement in a way needs to “compensate” for the small (or absence of) natural
decay and thus initially exceeds 𝜙𝜎 : the second term in (11 is positive). As pollution stock
falls over time, however, the need for compensation falls as well and as pollution converges
to zero, so does the second term in (11). This peculiar scenario can be thought of a situa-
tion where not only all the current emissions are abated but some of the exisitng pollution
is removed as well, which would correspond to “improving” environmental quality. Given
the perculiarity of this case17 we are going to focus only on parameter constellations such
that 𝛼 > 𝜀g
𝛽
, so that abatement is initially below (or equal to) the zero-emissions threshold.
Overall, the dynamics of this economy can be summarized as follows. Consumption
grows over time at the rate g, assumed positive. Capital stock grows at the rate smaller than
17
We have not seen such a case in reality yet, although it is not unthinkable, given the developments in the
carbon capture and geoengineering technologies.
13
g but its growth rate converges to g in the long run. Pollution falls over time at the rate 𝜀g 𝛽
and coverges to zero in the long run. Note that the dynamics of pollution are governed by
the structure of preferences, which is also visible in (9). Optimal abatement propensity is
such that it converges to the zero-emissions threshold 𝜙𝜎 in the long run. The convergence
path, however, depends on the parameters of the model, in particular on the relationship
between the natural decay rate, 𝛼 , and the speed of pollution decline, 𝜀g 𝛽
. If the environment
is relatively efficient (resp., inefficient) in removing pollutants, then the optimal abatement
starts below (resp., above) the zero-emissions threshold. Note that in the interior regime,
the initial capital stock and the initial pollution stock are linked by (10). It is, of course,
only by chance that the two initial stocks align precisely to satisfy this relationship. When
they do not, the economy finds itself in one of the border-control regimes (described in
detail in the appendix) and eventually converges to the interior regime, as shown in subsec-
tion 2.3 and illustrated in Fig. 1.
Now that we have solved for the full dynamics of the economy in the interior regime, we
can derive the explicit value function associated with the interior solution. From (10), we
can express consumption as a function of the two state variables and therefore we can write
the value function as
(aK + bP)1−𝜀 P1+𝛽

V(K, P) = −x ,
1−𝜀 1+𝛽
1−𝜀 a(𝛼−gp )
where the constants are given by a = (r − g) 𝜀 , b= r+gp
, x= 𝛽𝜒
(1+𝛽)r−𝜌
. Knowing the
shape of the value function will prove particularly useful when we introduce the stochastic
case. In particular, we are going to be interested in the long-run (or asymptotic) value func-
tion such that t → ∞. We provide a derivation of this function in the appendix and proceed
by discussing the optimal trajectory under alternative values of K0 and P0.
2.3 Optimal Trajectory with Different Initial Conditions
Equations (9) and (10) define the trajectory Sint in the state-space, illustrated in Fig. 1 by
the blue solid line.
This line splits the state space in two regions. In the region above Sint , the maximum
abatement is applied, while in the region below Sint zero abatement is chosen. The ini-
tial conditions determine the initial position of the economy in the state space and conse-
quently which control from the set {0, 𝜃t∗ , 𝜃} ̄ is applied from the outset. The blue line thus
corresponds to the set of "interior initial conditions" (K0 , P0 ) = (K0 , P(K0 )) for which inte-
rior controls are applied ∀t > 0. The state vector (Kt , Pt ), with (K0 , P0 ) on Sint , at the opti-
mal trajectory (Kt , Pt , Ct , 𝜃t ), moves from left to right along the blue solid line. For the ini-
tial conditions (K0 , P0 ) not on Sint , such as those marked by (K0 , P0 )� and (K0 , P0 )��, optimal
trajectory requires switching at some future time (c-control also switches). For instance, if
the initial conditions are given by (K0 , P0 )��, the optimal trajectory is such that no abate-
ment takes place for some finite time T̃ , while pollution and capital stock both increase, as
shown by the dashed line. For (K0 , P0 ) above the blue line, such as for example (K0 , P0 )�,
the optimal trajectory, shown by the dash-dotted curve, is such that pollution declines faster
than under “interior trajectory” and it converges to the blue solid line after some time T. In
13
the long run, the economy converges to the interior path regardless of which path it started
on initially.
3 Model with Climate Shocks
3.1 Baseline Model
The stochastic model that we present below is a modification of the deterministic model of
the previous section. We continue to assume that the production process is polluting. In
addition, we are going to assume that cumulative emissions cause deterioration of the natu-
ral environment and an increase in the global temperature, leading to a random occurrence
of natural disasters.18 We assume that an arrival of a natural disaster (we shall also refer to
it as an "event") follows the Poisson process with the mean arrival rate 𝜆. In the next sec-
tion we extend our model to include endogenous arrivals, while for the moment we assume
that 𝜆 is constant. When an event occurs, an endogenously-determined amount of the exist-
ing capital stock is destroyed. We denote by 𝜔t (Pt , Kt ) ∈ (0, 1) the fraction of capital which
survives the shock. In fact, recent floods, as the one in Pakistan in 2010 or in the Philip-
pines in 2013, had a profound effect on infrastructure and the capital stock (both physical
and human). According to the predictions of climate sciences, the magnitude of the dam-
age is likely to increase in the future due to climate change and hence we model it as a
positive function of the stock of GHG or, equivalently, the higher is the pollution stock the
𝜕𝜔
lower is the after-shock share of survived capital, 𝜕Pt < 0. In addition, we assume that the
t
𝜕𝜔
damages are larger, the larger is the amount of capital exposed to destruction, i.e. 𝜕Kt < 0,
t
capturing the notion of the value at risk (Bouwer 2011). In particular, we assume that the
survived share of capital, 𝜔t , can be represented by a function of a single argument, 𝜐t ,
which depends on Pt and Kt such that 𝜐t = P𝜉t Kt𝜂 , where the constants 𝜉 > 0 and 𝜂 > 0 gov-
ern the relative importance of the climate change component and the exposure component,
respectively. We choose a negative exponential function for 𝜔, so that 𝜔(𝜐t ) = e−𝛿𝜐t ∈ (0, 1),
where the parameter 𝛿 > 0 can be interpreted as the damage intensity.19 This specification
has several desirable characteristics. First, it ensures that 𝜔 is indeed a fraction, i.e. lies
between zero and unity. Second, 𝜔 is decreasing in P and K, i.e. 𝜕𝜔 𝜕P
= 𝜉𝜔 ln 𝜔∕P < 0,
𝜕𝜔
𝜕K
= 𝜂𝜔 ln 𝜔∕K < 0 to capture the effect of climate change and the capital exposure,
respectively. In general, any function satisfying these properties would be suitable for
describing 𝜔. Note that when 𝜉 = 𝜂 = 0 the model reduces to a special case with constant
damages.
Recent empirical literature on attitude to risk in the presence of disasters has largely
confirmed the "risk vulnerability" hypothesis first proposed by Gollier and Pratt (1996).
The hypothesis states that agents operating in risky environments characterized by a pos-
sibility of a loss on average, i.e. agents who are exposed to unfair risks, tend to exhibit a
18
We do not explicitly model temperature increase as it is now widely confirmed by the Earth system mod-
els that warming is proportional to cumulative emissions and that temperature response to a CO2 emission
is almost instantaneous and then constant as a function of time (see Dietz and Venmans 2019 and references
therein).
19
Exponential damages, although without the exposure component, are commonly used in the literature,
e.g. Golosov et al. (2014), Dietz and Venmans (2019), Van der Ploeg and de Zeeuw (2018).
13
Fig. 1 Optimal paths for different initial conditions
more risk-averse behavior than otherwise (see, e.g. Harrison et al. 2007; Guiso and Paiella
2008; Cameron and Shah 2015). Moreover, risk aversion also increases in the magnitude of
damages (Cameron and Shah 2015). We include this property in our model by linking the
parameters of the damage function to the parameters governing attitude to risk in the util-
ity function: 𝜀 = 𝜀(𝜂) and 𝛽 = 𝛽(𝜉) with 𝜀� (𝜂) = 𝛽 � (𝜉) = 𝛾 > 0, i.e. the mappings are linear
with equal slope such that the relative importance of the climate change component and the
exposure component is preserved. This property will prove useful in obtaining explicit ana-
lytical solution. In the rest of the analysis we shall assume that 𝛾 is equal to unity. This is
done for simplicity, as the magnitude of this parameter does not have any qualitative effects
on the results. In the numerical part of the paper we do consider alternative values of 𝛾 as
robustness checks.
The planner’s objective is to maximize the expected discounted utility over an infinite
planning horizon with respect to consumption, Ct , and the share of output devoted to abate-
ment, 𝜃t , subject to the stochastic capital accumulation process and the dynamics of the
cumulative emissions:
{ ∞ }
∫0 (13)
−𝜌t
max 𝔼0 U(Ct , Pt )e dt
Ct ,𝜃t
s.t.
[ ]
dKt = [(1 − 𝜃t )Yt (Kt ) − Ct ]dt − 1 − 𝜔t (Pt , Kt ) Kt dqt , K0 given, (14)
dPt = (Et − 𝛼Pt )dt, P0 given, (15)
Et = 𝜙Yt (Kt ) − Z(It ), (16)
It = 𝜃t Yt (Kt ), (17)
13
where 𝔼0 is the expectations operator, dqt is an increment of the Poisson process, and 𝜌 > 0
is the constant rate of time preference. The utility function is twice continuously differenti-
able in both arguments. We also require that the capital stock, pollution stock, consumption
and emissions rates are non-negative and 𝜃t ∈ [0, 1). In what follows, we only describe the
solution on the interior path.
3.2 Solution
Denoting by V(K, P) the value function associated with the interior solution to the opti-
mization problem described in (13)–(17), the Hamilton-Jacobi-Bellman (HJB) equation
may be written as
{ [ ]}
̃ P) − V(K, P) ,
𝜌V(K, P) = max U(C, P) + Vk [(1 − 𝜃)Y − C] + Vp [(𝜙 − 𝜃𝜎)Y − 𝛼P] + 𝜆 V(K,
C,𝜃
(18)
where Vk ≡ 𝜕V(K, P)∕𝜕K , Vp ≡ 𝜕V(K, P)∕𝜕P and the new capital stock K = 𝜔(P, K)K .
̃
In what follows, the subscripts c, k, p refer to partial derivatives with respect to C, K, P ,
while we shall always use the subscript t to indicate that a variable is time-dependent. We
relegate the detailed derivations to the Appendix A.1, while focus only on the key results
in the main text. The solution method is largely based on Sennewald and Waelde (2006),
however we are going to study only the long-run behavior. The asymptotic value function
can be found explicitly as (see appendix)
𝜓 −𝜀 Kt1−𝜀 xP1+𝛽
t
V(Kt , Pt ) = − .
1−𝜀 1+𝛽
The coefficients 𝜓 and x depend on the parameters of the model and can be found by sub-
stituting optimal controls into the HJB:
( )
𝜙 [ ]
(19)
∗
𝜀𝜓 − 𝜌 + (1 − 𝜀)A 1 − + 𝜆 (𝜔(𝜓 ∗ ))1−𝜀 − 1 = 0,
𝜎
where 𝜔(𝜓, x) = e−𝛿[𝜎x𝜓 and

𝜀 ]−1∕𝛾
𝜒
x=
𝜌 + 𝛼(1 + 𝛽)
. (20)
The coefficient 𝜓 is important for our further analysis since it represents the optimal ratio
of consumption to capital.
In the next step, we obtain the stochastic consumption growth by applying the Ito’s
formula for jump processes:
{ ( ) }
dC 1 𝜙 [ ]
= A 1− − 𝜌 + 𝜆 𝜔1−𝜀 (1 + 𝜂 ln 𝜔) − 1 dt + (𝜔 − 1)dq. (21)
C 𝜀 𝜎
Given that K̃ = 𝜔K , while C = 𝜓 ∗ K and C̃ = 𝜓 ∗ K̃ , 𝜔 also represents the ratio of post-

to pre-shock consumption rates. Therefore, the last term on the RHS of (21) is negative,
reflecting the downward jump in consumption at the time of a disaster. The first term on
the RHS, multiplying dt, represents what we label as "trend" consumption growth rate, g∗,
which prevails in-between climate shocks:
13
⎧ ⎫
⎪ ⎪
⎪ � � ⎪
g ≡ ⎨A 1 −
∗ 1⎪ 𝜙 ⎪
−𝜌 + 𝜆(𝜁 − 1) ⎬, (22)
𝜀⎪ 𝜎 ⏟⏞⏟⏞⏟ ⎪
⎪⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟ precautionary ⎪
⎪ KR term ⎪
⎩ effect ⎭
where 𝜁 = 𝜔1−𝜀 (1 + 𝜂 ln 𝜔). The expression has a familiar Keynes–Ramsey form albeit
with some additional elements. The standard, deterministic Keynes–Ramsey formula states
that the growth rate of consumption equals the difference between the real interest rate
(usually the marginal product of capital) and the rate of pure time preference, adjusted by
the elasticity of intertemporal consumption substitution. First, note that in Eq. (22) the
economy’s implicit real interest rate, given by the first term inside the parentheses, is not
equal to just the marginal productivity of capital but is reduced by the emission intensity
of output, adjusted by the abatement efficiency, i.e, the term 𝜙∕𝜎 . It follows that in our
framework pollution has an unambiguously negative effect on the real interest rate. It may
be dampened by either increasing the abatement efficiency, 𝜎 , or decreasing the polluting
intensity, 𝜙.
Second, expression in (22) accounts for the effect of uncertainty, represented by the last
term 𝜆(𝜁 − 1), labeled as "precautionary effect", which includes the disaster hazard rate
(𝜆), the exposure effect (−1) and the pollution-stock effect (𝜁 ). We are especially interested
in the magnitude of 𝜁 relative to unity. If 𝜁 > 1, then the presence of uncertainty speeds
up consumption growth. In this case, the optimal stochastic consumption path is tilted
counterclockwise, as compared to the consumption path in a deterministic Keynes–Ram-
sey model. Therefore, the economy starts with a relatively low consumption rate at the
beginning of the planning horizon, which implies the presence of the precautionary-saving
motive, including saving for financing of emissions control.20 By contrast, when 𝜁 < 1, the
risky environment contributes to a growth slowdown, tilts the consumption profile clock-
wise and thus implies a precautionary dissaving motive. The peculiarity of the current set-
ting is that the gross savings are endogenously split between two purposes: capital accumu-
lation and abatement, both of which serve to protect the economy from climate disasters.
It is clear that abatement reduces emissions and therefore unambiguously contributes to a
reduction in damages. Capital accumulation, however, has a double-sided effect. On the
one hand, more capital implies more output and more emissions. On the other hand, having
more capital creates an "emergency buffer" for the rainy days—when a disaster strikes. We
discuss in more detail the economy’s optimal saving rate and how it is affected by climatic
parameters in Sect. 4.3. For the moment we shall focus our attention on the magnitude of
the precautionary effect and its relevance for the optimal growth rate.
Note that 𝜁 = 𝜔1−𝜀 (1 + 𝜂 ln 𝜔) ≷ 1 is composed of two terms. First, the ratio of post-to-
pre shock marginal utilities of consumption, U � (C)∕U
̃ �
(C) = 𝜔−𝜀, which is unambiguously
larger than unity since C̃ < C . Second, the term 𝜔(1 + 𝜂 ln 𝜔), which represents the reac-
tion of the post-shock capital stock to a change in the pre-shock capital stock.
We provide a detailed analysis in the appendix, while here we summarize the gist of
the results. If 𝜀 ∈ (0, 1), 𝜁 is unambiguously less than unity, while if 𝜀 > 1, a priori an
20
The outcome is analogous to what has been found in the literature on precautionary savings under uncer-
tainty in other contexts (see, e.g., Waelde 1999, Steger 2005).
13
ambiguity exists. We show, however, that this ambiguity can be resolved and 𝜁 is less than
unity for the relevant range of parameters (see Fig. 3 in Appendix A.2). We conclude that
(𝜁 − 1) is negative and hence the presence of uncertainty contributes to a lower consump-
tion growth rate than in the deterministic model. This finding points to the presence of
a specific type of consumption "smoothing" such that the economy aims at reducing the
slope of its consumption profile and at controlling the size of consumption jumps, in fact
the percentage drop is stabilized at a constant value in the long run. By reducing the slope
and by stabilizing the jumps, the economy achieves the optimal consumption trajectory
which is "smoothed out" as much as possible, leaving only the unpredictable component—
the timing of jumps - affect its evolution. Even though a somewhat smoother profile can be
implemented, perfect consumption smoothing is not achievable because of random timing
of events.
The smoothing effect can be confirmed by comparing the optimal consumption growth
rate of our economy with the one prevailing in a stochastic environment but with shocks of
exogenous size. Suppose that random climate disasters arrive at the same Poisson rate but
the survived share of the capital stock is exogenous and constant at 𝜔̌ ∈ (0, 1). Assume fur-
ther that the exogenous percentage drop in the capital stock is the same as the optimal one
of our baseline
{ ( model, ) so that[ 𝜔̌ = 𝜔. Then
]} the trend growth rate of consumption is given
̃
U � (C)
by ǧ = 1
𝜀
A 1− 𝜙
𝜎
−𝜌+𝜆 U � (C)
−1 , where the ratio of marginal utilities is simply
𝜔 > 1. It follows immediately that ǧ > g∗. Even though the drops in consumption are
−𝜀
identical under both scenarios, the optimal consumption path associated with ǧ is steeper
than that associated with g∗. The reason is that in the former case the economy, being una-
ble to affect the size of the damage, is forced to choose a relatively high growth rate in
order to build up an emergency buffer. In the latter case the economy endogenously con-
trols the size of the jump and simultaneously chooses a lower trend rate as a precautionary
measure. Such a growth path of the economy is also fundamentally different from its
"expected" version which may be interpreted as a hypothetical extreme case of consump-
tion smoothing. The expected growth rate in our baseline model, denoted by ge, is given by
{ ( ) }
ge ≡
d𝔼t Ct ∕dt 1 𝜙 [ ]
= A 1− − 𝜌 + 𝜆 𝜔1−𝜀 (1 + 𝜀 ln 𝜔) − 1 + 𝜀(𝜔 − 1) . (23)
Ct 𝜀 𝜎
We are now in a position to state the following
Lemma 1 The solution to the maximization problem (13) - (8) is characterized by the opti-
mal trend consumption growth rate which is (i) lower than that of the deterministic model
without climate change induced disasters; (ii) lower than that of a stochastic model with
fixed-damage disasters; but (iii) higher than the expected growth rate.
Proof Parts (i) and (ii) follow from the discussion above. Part (iii) follows from the fact that
since 𝜔 < 1, the last term inside the square brackets in (23) is negative and thus ge < g∗.
◻
As an illustration of the optimal path in the baseline model, we show in Fig. 2 the
consumption rate as a function of time. The solid line represents the stochastic path,
which exhibits a trend growth rate g∗ in the absence of climate events. At times t1 and
t2 , environmental shocks are assumed to occur causing an immediate downward jump,
followed by a subsequent period of growth at the previous rate. The dashed line shows
13
the time profile of consumption under the expected growth scenario. There is a funda-
mental difference between the dashed and the solid curves in that the former smoothes
out the jumps and discontinuities of the latter, creating an illusion of a perfect con-
sumption smoothing and thereby ignoring the crucial effects of uncertainty.
We now turn to the optimal abatement policy. How much of the current resources to
devote to emissions control is a key policy question. We show in the Appendix that the
optimal abatement propensity on the interior path is given by
𝜙 𝜉𝜆𝜔1−𝜀 ln 𝜔
𝜃∗ = − . (24)
𝜎 𝛽A
The first term on the RHS represents the zero-emissions threshold, which is the optimal
abatement propensity of the deterministic model in the long run. We see that the presence
of the stochastic climate shocks induces a larger long-run abatement propensity. The extra
abatement is reflected in the second term in (24), which is positive since ln 𝜔 < 0. Clearly,
in the absence of random disastes, i.e. if 𝜆 = 0, this term vanishes.
Proposition 1 The long-run solution of the maximization problem described by (13) - (8) is
characterized by the following:
(i) optimal consumption rate is a constant fraction of the capital stock;

(ii) optimal abatement expenditure is a constant fraction of output;
(iii) consumption, capital stock, output, and the overall abatement grow at the same
constant rate in-between climate shocks.
Proof provided in the Appendix. ◻
Corollary 1 In the case of logarithmic utility long-run consumption-to-capital ratio, 𝜓 ∗, is

independent of climate parameters.
With log utility (𝜀 = 1), the long-run value of 𝜓 ∗ can be obtained explicitly and after
substitution into (22) and (24) we obtain:
𝛿𝜆 𝜙
𝜓 ∗ = 𝜌, 𝜃 ∗ = + , g∗ = A(1 − 𝜃 ∗ ) − 𝜌.
𝜌Ax𝜎 𝜎
Since 𝜓 ∗ does not respond to any climatic parameters, the post-to-pre shock consumption
ratio, 𝜔, is independent of the hazard rate. Consequently, the optimal growth rate responds
to 𝜆 only through its direct effect, 𝜕g∕𝜕𝜆 but not through the jump effect,
(𝜕g∕𝜕𝜔)(𝜕𝜔∕𝜕𝜓)(𝜕𝜓∕𝜕𝜆). The direct effect reflects the precautionary "dissaving" motive
and is negative (equal to − x𝜎𝜌
𝛿
), contributing to a growth slowdown, while the indirect effect
may either reinforce the direct effect or counteract it, depending on the value of 𝜀. When
log utility is assumed, the latter element disappears from the general picture. This observa-
tion motivates us to depart from the log-utility case in the quantitative assessment of the
model presented in Sect. 5. One of our questions of interest will be to quantify the discrep-
ancy in the optimal policy rules under unitary and alternative values of 𝜀.
13
Fig. 2 Optimal path (solid) versus expected path (dashed)
4 Effects of Economic and Climatic Fundamentals
4.1 The Abatement Propensity
In order to study the responses of the economy’s optimal long-run abatement policy, charac-
terized by 𝜃 ∗, to changes in the fundamental parameters of the model, we totally differentiate
the system of Eqs. (19) and (24). Our key "climatic" parameters of interest are 𝜆 and 𝛿 which
reflect the expected frequency and damage intensity of climate shocks, respectively. Among
the "economic" parameters we consider 𝜎, representing the efficiency of abatement technol-
ogy, and 𝜙, representing polluting intensity of production. The system becomes
M z
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞ ⏞⏞⏞ ⎛ d𝜆 ⎞
� � � � � � � �
𝜀 1 − 𝜆(1 − 𝜀)𝜔1−𝜀 𝜓 −1 ln 𝜔 0 d𝜓 Δ𝜓𝜆 Δ𝜓𝛿 Δ𝜓𝜎 Δ𝜓𝜙 ⎜ d𝛿 ⎟
𝜀𝜔1−𝜀 𝜓 −𝜀−1 𝜆𝛿[1 + (1 − 𝜀) ln 𝜔] xA𝜎
×
d𝜃
=
Δ𝜃𝜆 Δ𝜃𝛿 Δ𝜃𝜎 Δ𝜃𝜙
× ⎜ d𝜎 ⎟ ,
⎜ ⎟
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⎝ d𝜙 ⎠
Δ ⏟⏟⏟
y
where the exact expressions for Δ’s can be found in the Appendix. We can compactly
rewrite the system as Mz = Δy. It is useful for future analysis to find the determinant of M:
[ ]
[ ] 𝜆(1 − 𝜀)𝜔1−𝜀 ln 𝜔
|M| = M11 M22 = 𝜀 1 + 𝜆(1 − 𝜀)𝜔1−𝜀 𝜓 −𝜀−1 𝛿(x𝜎)−1 xA𝜎 = 𝜀xA𝜎 1 − .
𝜓
Since M22 is unambiguously positive, the sign of |M| hinges on the sign of M11, which is in
general ambiguous. If 𝜀 ∈ (0, 1], M11 is positive, while if 𝜀 > 1, M11 may become negative
under some constellations of parameter values, in particular, if 𝜀 > 𝜀M11 ≡ 1 − 𝜓𝜔
𝜀−1
𝜆 ln 𝜔
> 1. If
𝜀 is viewed as a parameter reflecting opinions and objectives of policymakers, its estimates
may range from 1 to 3. In what follows we shall focus primarily on this empirically-rele-
vant range of 𝜀. In order to get a rough idea whether 𝜀M11 is anywhere near the empirically
relevant range, we compute it assuming a plausible range of values for 𝜆, 𝜔 and 𝜓 . It turns
out (see Appendix) that 𝜀M11 is beyond the empirically-relevant range of 𝜀 and thus we shall
exclude the case M11 < 0.
13
For the moment we are interested in the effects of the arrival rate (𝜆), damage intensity
(𝛿), abatement efficiency (𝜎 ), and polluting intensity (𝜙). The effects of the remaining
parameters can be computed in a similar manner. Applying the Cramer’s rule, we have:
|M |
d𝜃
di
= |M|𝜃i , where i = 𝜆, 𝛿, 𝜎, 𝜙, and M𝜃i is the matrix M with the second column replaced by
|M |
the column of Δ corresponding to i. Similarly, d𝜓
di
𝜓i
= |M| , where M𝜓i is the matrix M with
the first column replaced by the column of Δ corresponding to i.
Proposition 2 The long-run abatement propensity is:
(i) an increasing function of the arrival rate,

(ii) an increasing function of the damage intensity,
(iii) a decreasing function of the abatement efficiency and
(iv) an increasing function of the polluting intensity.
Corollary 2 In the case of logarithmic utility function (𝜀 = 1), the abatement propensity is
an increasing linear function of the hazard rate, of the damage intensity, and of the pollut-
ing intensity, and a decreasing convex function of the abatement efficiency.
4.2 The Growth Rate

∗ 𝜕g∗ 𝜕g∗ 𝜕𝜔 d𝜓
The effects on the growth rate are found as dgdi = 𝜕i
+ 𝜕𝜔 𝜕𝜓 di
.
Proposition 3 The optimal trend consumption growth rate is:
(i) a decreasing function of the arrival rate,

(ii) a decreasing function of the damage intensity,
(iii) an increasing function of the abatement efficiency and
(iv) a decreasing function of the polluting intensity.
While the results (i), (ii), and (iv) are self-explanatory, the statement in (iii) deserves
a short interpretation. In general, the effect of the abatement efficiency on g∗ is ambigu-
ous and consists of two terms (see Eq. (D.10)). The first term is clearly positive, while
the sign of the second term depends on 𝜀. If 𝜀 = 1 it vanishes and if 𝜀 > 1 it is unambigu-
ously positive. The fact that a higher abatement efficiency has an ambiguous bearing on
economic growth is due to two effects—the interest-rate effect and the jump-smoothing
effect—which work in opposite directions. On the one hand, an improvement in efficiency
of abatement increases the economy’s real interest rate and thus enhances the growth rate
through the first term in Eq. (22). On the other hand, it increases the post-event consump-
tion rate, shrinking the pre- to post-event consumption gap and thus contributes to a growth
13
slowdown through the last term in Eq. ( 22). It turns out that with a relatively high RRA
the latter effect is dominated by the former.
Polluting intensity also affects the growth rate through two channels. The first represents
the direct effect stemming from a decline in the real interest rate. The second, the indirect
effect, takes into account the change in the consumption jump, 𝜔, through 𝜓 . A higher pol-
luting intensity requires a higher abatement share (Proposition 2(iv)) and thus it reduces
the share of consumption in total capital. Both direct and indirect effects are negative and
contribute to a growth slowdown.
Corollary 3 In the case of logarithmic utility function the optimal trend consumption
growth rate is a decreasing linear function of the damage intensity, of the arrival rate, of
the polluting intensity, and an increasing concave function of the abatement efficiency.
4.3 Saving Propensity
We define the propensity to save, s, as the non-consumption share of output, i.e. a share of
output spent on augmenting the capital stock plus on abatement (the latter being equal to
𝜃 ∗): s∗ = 1 − 𝜓 ∗ ∕A. It follows that the effects of all our parameters of interest (𝜆, 𝛿, 𝜎, 𝜙) on
s∗ are simply the opposite of those on 𝜓 ∗, divided by A > 0.
Proposition 4 The long-run propensity to save is:
(i) an increasing function of the arrival rate,

(ii) an increasing function of the damage intensity,
(iii) a decreasing function of the abatement efficiency and
(iv) an increasing function of the polluting intensity.
Corollary 4 In the case of logarithmic utility function, the saving propensity is independent
of the arrival rate, damage intensity, abatement efficiency and polluting intensity.
Proof provided in the Appendix.
In the case of log utility, all the parameters of interest lose their relevance (the deriva-
tives of 𝜓 and thus s become zero). An increase in the abatement efficiency 𝜎 causes a
decrease in 𝜃 ∗, implying that with unchanged s∗ the share of output devoted to augmenting
the capital stock increases. A better abatement technology in an economy with logarithmic
preferences results in a smaller abatement share but a larger capital stock. On the other
hand, an increase in the polluting intensity causes 𝜃 ∗ to rise, while s∗ remains unchanged,
implying that capital investment must fall. The same is true for an increase in the arrival
rate of disasters and their damage intensity.
Note that when 𝜀 ≠ 1 these effects might be mitigated because of the impacts on s∗
which may go in the same direction as those on 𝜃 ∗. For instance, an increase in the pol-
luting intensity increases both s∗ and 𝜃 ∗. It can be shown that an increase in s∗ is smaller
13
than that in 𝜃 ∗, implying that the abatement propensity rises at the expense of the invest-
ment share, which falls. The reduction in capital investment is, however, smaller than under
logarithmic preferences.
The key endogenous variables, 𝜃 ∗, g∗ and s∗, respond non-linearly to changes in the eco-
nomic and climatic fundamentals when we depart from log utility, although the direction
of the effects is intuitively clear. A worsening in the shock’s characteristics, i.e. an increase
in expected frequency and/or damage intensity, increases 𝜃 ∗ and s∗ and reduces g∗. An
improvement in abatement efficiency or a reduction in polluting intensity decreases 𝜃 ∗, s∗,
and raises g∗. The non-linearities in responses are important for a quantitative assessment
of the optimal policy which we present in Sect. 5.
5 Endogenous Arrivals
We have mentioned in the introductory section that climate scientists have diverging opin-
ions on whether the frequency of natural disasters will increase in the future due to global
warming or not. The IPCC (2014) explicitly states that such a possibility exists. Recent
contributions by Van der Ploeg (2014), van der Ploeg and de Zeeuw (2018), and Zemel
(2015) model the hazard rate endogenously, as an increasing function of the stock of carbon
in the atmosphere.21 In this section we explore the implications of introducing an endog-
enous disaster arrival rate in our benchmark model. For the moment we shall write a gen-
eral function 𝜆 = 𝜆(Pt , Kt ) with both partial derivatives being positive: 𝜆k ≡ 𝜕𝜆∕𝜕K > 0,
𝜆p ≡ 𝜕𝜆∕𝜕P > 0 to capture the climate change and the exposure effects.22 We subsequently
specify a possible functional form for this relationship.
The HJB equation of the problem is similar to (18), except that now the arrival rate
depends on the stock of carbon and on the stock of capital. The optimality conditions with
respect to consumption and abatement propensity remain unchanged. Only the conditions
with respect to K and P are augmented by a term representing the change in the value
function due to a change in the arrival rate. The exact expressions can be found in the
Appendix.
In order to make further progress, we need to specify how the arrival rate of climatic
hazards is affected by economic activity. We assume that the arrival rate is an increas-
ing and possibly non-linear function of the capital and the pollution stocks, which capture
the exposure and the climate-change effects, respectively. Similarly to how we modeled
21
Using the Duffie-Epstein stochastic differential utility framework, van der Ploeg and de Zeeuw (2018)
find that the optimal carbon price increases in the face of a pending catastrophe to make the shock less
imminent and that adjustments to saving are needed to smooth consumption. They also confirm numerically
that assuming a RRA which is different from the inverse of the elasticity of intertemporal consumption sub-
stitution does not affect the results significantly. Zemel (2015) studies dynamic interactions between mitiga-
tion and adaptation activities, where the former reduces the risk of a harmful event, while the latter reduces
the damage inflicted in case an event occurs nonetheless.
22
We assume that the frequency of disasters may also be an increasing function of K, in addition to being
an increasing function of P, because man-made infrastructure can make disasters more likely to happen. For
example, building houses on the land which has been recovered from the sea certainly makes these houses
more exposed to flooding; in other words, these houses face a higher probability of disaster. Deforesting
a hill, e.g. for agriculture, increases the chances of a landslide. Digging for oil in the sea increases the
chances of an oil spill. Destruction of rain forests for road construction and industry raises the probability of
land soil degradation. These examples suggest that the built environment and the build up of specific infra-
structure have also an impact on the probability of adverse events, which we include in our model.
13
damages in our baseline model, we adopt a Cobb-Douglas combination of K and P such

that 𝜆 = 𝜆(𝜐𝜇 ), 𝜐 = K 𝜂 P𝜉 , as before, and 𝜇 > 0 is used to differentiate the effect of 𝜐 on the
arrival rate from the effect on damages. We relegate the detailed derivations to the Appen-
dix, while focus on the final results in the main text. We use a bar above the endogenous
variables to indicate the equilibrium values.
The long-run consumption-to-capital ratio, 𝜓̄ , is an implicit solution to
( )
𝜙 [ ]
𝜀𝜓̄ = 𝜌 − (1 − 𝜀)A 1 − ̄ 𝜐) 𝜔̄ 1−𝜀 − 1 ,
− 𝜆(̄ (25)
𝜎
where 𝜐̄ = (𝜓̄ 𝜀 x𝜎)−𝜇∕𝛾 in equilibrium. It is possible that (25) has multiple solutions for
some functional forms of 𝜆, especially non-linear.23 We discuss in the Appendix the condi-
tions under which multiple solutions occur. When the value function is increasing in 𝜓̄ , the
relevant solution is given by the largest root of (25). If the utility function is logarithmic,
however, the solution for the long-run consumption-to-capital ratio is unique, 𝜓̄ = 𝜓 ∗ = 𝜌,
and identical to the solution with exogenous arrivals.
Comparing Eq. (25) with (19), we may conclude that if 𝜆̄ ⩾ 𝜆 and 𝜀 > 1, the right-hand
side of ( 25) is unambiguously smaller than that of (19), implying that 𝜓̄ < 𝜓 ∗. Moreover,
since d𝜐̄ ∕d𝜓̄ < 0 and 𝜆̄ � (̄𝜐) > 0, a smaller consumption-capital ratio increases the arrival
rate, which reinforces our initial conclusion. Thus it appears that when disaster arrivals are
endogenous to economic and polluting activities, the optimal consumption-to-capital ratio
is reduced, suggesting that abatement is increased, which complies with the general intui-
tion. We turn next to abatement.
The optimality condition with respect to the pollution stock (see Appendix) allows us to
solve for the optimal abatement share:
[ ]
𝜙 𝜉 𝜆̄ 𝜔̄ 1−𝜀 ln 𝜔̄ 𝜐̄ 𝜆̄ � 𝜇 𝜔̄ 1−𝜀 − 1
𝜃̄ = − − . (26)
𝜎 𝛽A A 1−𝜀
Comparing Eq. (26) with Eq. (24), we note two main differences. First, in the expression
for 𝜃̄ the arrival rate depends on the consumption-to-capital ratio 𝜓̄ through 𝜐̄ , while in
the expression for 𝜃 ∗ the arrival rate is fixed. Second, the last term in (26) does not appear
in (24). Similarly to Van der Ploeg (2014), and adopting his terminology, the last term rep-
resents the "risk-averting effect," while the middle term refers to the "raising-the-stakes"
effect. To understand the intuition behind the expression for 𝜃̄ , consider first the second
term in (26) which is similar to the second term in (24), except that the function 𝜔∗ has 𝜓 ∗
as its argument, while the function 𝜔̄ has 𝜓̄ . We have established that for 𝜆̄ ⩾ 𝜆 we have
𝜓̄ < 𝜓 ∗. Moreover, d[𝜔̄ 1−𝜀 ln 𝜔]∕d
̄ 𝜓̄ > 0, so that the second term in (26) is smaller than
that of (24), working to increase 𝜃̄ compared to 𝜃 ∗. Let us now turn to the last term in (26)
involving 𝜆̄ ′. Clearly, its presence is warranted by the fact that the arrival rate is endog-
enous. Moreover, this marginal effect on the arrival rate is weighted by its contribution to
the change in the value of the program, represented by the expression in the square brack-
ets. This is given by the value of the survived unit of capital relative to the status quo, the
term (𝜔̄ 1−𝜀 − 1)∕(1 − 𝜀). Thus, the last term is negative if 𝜀 > 1 and, being subtracted from
23
Van der Ploeg and de Zeeuw (2018) use a linear and a quartic specification, while Van der Ploeg (2014)
uses a linear function and three alternatives (quadratic, cubic and quartic) to calibrate the hazard function.
The latter paper also shows that the results with quadratic, cubic and quartic specifications do not differ sig-
nificantly from the results with a simple linear function (see Table 1, p. 38).
13
the first two, it contributes to an increase in 𝜃̄ . We would thus expect 𝜃̄ to be larger than 𝜃 ∗
when 𝜆̄ is at least as large as 𝜆, which seems to be plausible in light of the predictions from
climate physicists (IPCC 2014).
We turn next to the optimal growth rate. With the optimality conditions with respect to C
and K and application of the Ito’s Lemma for jump processes we obtain
⎧ ⎫
⎪ ⎪
⎪ ⎪
⎪ � � � 1−𝜀 � ⎪
1⎪ 𝜙 � 1−𝜀 � 𝜔̄ −1 ⎪
ḡ = ⎨A 1 − − 𝜌+ ̄
𝜆 𝜔̄ (1 + 𝜀 ln 𝜔) ̄ −1 + ̄ �
𝜆 𝜀𝜇 𝜐̄ ⎬.
𝜀⎪ 𝜎 �� 1−𝜀 ⎪
⎪�� baseline
�� ⎪
⎪ KR term
additional ⎪
⎪ precautionary effect < 0
⎩ precautionary effect < 0 ⎪
⎭
(27)
( )
Comparing (27) with (22), we see that the term A 1 − 𝜎 − 𝜌, representing the standard
𝜙
Keynes–Ramsey component, is the same. The second term in (27) is similar to the respec-
tive term in (22), while the last term in (27) has no equivalent in the benchmark model.
Since 𝜓̄ < 𝜓 ∗ and 𝜔̄ < 𝜔, the second term, labeled "baseline precautionary effect", is
smaller (i.e. more negative) than the respective term in (22) and hence it contributes to a
growth slowdown. Turning to the last term in (27), which appears due to the endogeneity
of the hazard rate and has no equivalent in the benchmark model, we see that it is unam-
biguously negative for 𝜀 > 1, hence reducing the growth rate even further. It represents an
additional precautionary measure due to the positive dependence of the arrival rate on eco-
nomic activity. We summarize the results in the following
Proposition 6 When the disaster arrival rate is endogenous and is at least as large as the
exogenous rate of the benchmark model,
(i) the optimal consumption-to-capital ratio is smaller,

(ii) the saving rate is higher,
(iii) the abatement share is larger, and
(iv) the growth rate is lower.
The evidence from IPCC (2014) on rising frequencies of climate-driven natural disasters
suggests that our near future might be characterized by arrival rates 𝜆̄ which are larger than,
say, the known historical average 𝜆. If this is the case, we conclude that a larger abatement pro-
pensity is warranted and the optimal growth rate of the world economy will have to be lower.
The intuition behind is entirely driven by the precautionary considerations which dictate a
lower growth rate of polluting input and a more aggressive abatement in order to reduce the
probability of disasters and the associated damages.
13
Table 1 Benchmark values of 𝜌 𝜒 𝛽 𝛼 𝜀 A 𝜙 𝜎 𝜆 𝛿 𝛾

parameters
0.015 1 1 0.0038 1 0.05 0.0004 0.08 0.02 1e − 6 1
6 Quantitative Implications
In this Section we explore the quantitative implications of our model and compare them
with recent findings in the literature. Our overarching objective is two-fold: to quantita-
tively assess the optimal abatement propensity and the growth rate of the economy; and
to provide a sensitivity analysis with respect to the key economic and climatic character-
istics. In particular, we look at the sensitivity of the results to variations in risk aversion,
efficiency of abatement technology, and characteristics of catastrophic events. We also ask
what the implications for the abatement policy are when an event is relatively common and
low-impact vs. rare but high-impact. The latter question is to a large extent motivated by
Stern’s critique (Stern 2016) of current economy-climate models which, in his view, fail
to adequately take into account the possibility of large-scale climate shocks. In our con-
text, this task is essentially equivalent to assessing the responsiveness of the climate-policy
instrument to changes in the severity and in the frequency of natural disasters.
In a recent article developing a climate DSGE model, Golosov et al. (2014) calculate the
range of the optimal carbon tax of $56.9–$496 per ton carbon in 2010 assuming alternative
discount rates.24 For a similar range of discount rates Van der Ploeg (2014) finds that the
optimal tax should be roughly a half, between $29 and $216 per ton. There is quite some
divergence in the estimates of the optimal tax, with the main conclusion from the recent
research being that the size of economic damages induced by climate change or a possibil-
ity of crossing a tipping point play a secondary role, while the discount rate seems to be the
key driver of policy prescriptions.25 This is one of the reasons why we chose to explicitly
focus on recurring random catastrophic events as the main source of economic damages
in order to elucidate their role in shaping the abatement policy. Within our framework,
one may ask, for example, what the cost of a carbon unit would be if the world economy
were to abate all of its current emissions. This can be obtained by dividing the abatement
expenditure by total emissions in a given year.26 The abatement expenditure is obtained by
multiplying GWP with the optimal abatement propensity from Eq. (24).
6.1 Calibration
The reference unit of time is set to 1 year. In the benchmark calibration we assume the rate
of time preference of 1.5% per year, as in Dietz and Stern (2015), Golosov et al. (2014);
24
A tax range of $28–$55/tC has been obtained by Barrage (2014), see e.g. Fig. S.9 of her paper, after hav-
ing introduced TFP growth, CRRA utility function and capital depreciation in the model of Golosov et al.
(2014).
25
Dietz and Stern (2015) amend the standard DICE model with damages to the capital stock and to the
total factor productivity. They show that even with the discount rate of 1.5% the extended DICE can pro-
duce considerably higher carbon taxes than the original DICE model. See also Van den Bijgaart et al. 2016.
26
This measure of a carbon unit is not perfect since we do not know how much of emitted carbon has been
abated in a given year. Thus our measure represents an upper bound.
13
Nordhaus (2008) and Van der Ploeg (2014).27 These authors also assume a CRRA utility
function. Golosov et al. (2014) use 𝜀 = 1 (log utility), which we adopt here as a starting
point for the purpose of having a meaningful comparison. The relative weight on pollu-
tion in the utility function, 𝜒 , is set to unity. The parameter governing the curvature of the
disutility of pollution, 𝛽 , is also set to unity, which implies a quadratic disutility, often used
in the literature. The carbon absorption capacity of natural sinks is set at 𝛼 = 0.0038 (see
footnote 6).
To calibrate output emission intensity, 𝜙, we take the data from the World Bank series
"CO2 emissions per GDP" (World Bank 2016), which reports average values for the period
2011–15 ranging from 0.1 kg per dollar of GDP for Sweden and France, 0.2 for Germany
and the UK, 0.4 for the US, Canada and Brazil and up to 2.1 for China. We use a world
average value of 0.4 kg, so that 𝜙 = 0.0004 tons CO2 per dollar of GDP.
As for abatement efficiency, 𝜎 , empirical studies (Hood 2011; McKinsey 2009) show
that various abatement activities are inexpensive and thus relatively efficient; in the resi-
dential sector it applies to electronics, appliances, and insulation retrofit, in transportation,
e.g., to hybrid cars, and in agriculture to tillage management. The marginal costs of these
activities are reported to be negative or slightly positive amounting to less than $5 per ton
of CO2. Extending abatement activities through further policies, e.g. in the power sector
and with reforestation, increases the costs substantially, although Hood (2011) concludes
that "a significant level of emission abatement could be achieved with existing technologies
at carbon prices of less than $50 per ton of CO2". With the highest value of the marginal
cost curve of $50 per ton, the average value lies in the range of $10 to $15. However, for
reaching the 2 ◦ C target further emission reductions might become necessary in the future,
including carbon capturing and sequestration, whose costs are estimated to lie between $50
and $100 per ton of CO2. One should note, however, that there are considerable learning
effects in these new technologies, especially over a long time horizon. We therefore choose
an average value of the various abatement measures and aim to include dynamic effects
(learning) by setting 𝜎 = 0.08 in the benchmark calibration, which corresponds to $12.5
per ton of abated CO2. We shall also consider alternative values of $20 (𝜎 = 0.05) and $40
(𝜎 = 0.025) when technology development is viewed in a more pessimistic way. For the
total factor productivity, A, we adopt a moderate value of 5%.
Statistics for large-scale natural catastrophes over the last few decades suggest varying
arrival rates and damages for different types of disasters. The Indian Ocean Tsunami in
2004 caused at least $10 bn worth of damage and affected six countries: Indonesia, India,
Maldives, Sri Lanka, Somalia, and Thailand. The damage amounted to 0.86% of the sum
of GDPs in 2004 of the affected countries (Somalia not included due to lacking GDP data
in WDI). Hurricane Katrina in 2005 caused $108 bn damage which amounted to 0.825% of
the US GDP. Typhoon Haiyan in the Philippines in 2013 caused $2.8 bn damage, equiva-
lent to 1.05% of GDP. Cavallo et al. (2013) count 2597 natural disasters (floods and storms,
including hurricanes) worldwide during the period 1970–2008, which implies an average
annual arrival rate of 0.34. Assuming that 10 percent of the shocks are climate-related
events yields 𝜆 = 0.034. With respect to larger shocks, we refer to the IPCC to assume
that there is a 20% chance that a catastrophic-climate outcome occurs in the next 50 years,
which implies that 𝜆 = 0.004.
27
The discount rate of 1.5% is a commonly-used value in the literature. We adopt this parametrization
solely for comparison purposes and invite the reader to investigate the social and philosophical foundations
of discounting in Stern (2015).
13
In our quantitative assessment we shall consider two scenarios: (i) relatively low dam-
age intensity of disasters ("low" 𝛿) and (ii) relatively high damage intensity of disasters
("high" 𝛿).28 Within each scenario we distinguish among three arrival rates, three abate-
ment efficiencies, and two values of the relative risk aversion parameter. The arrival rates
correspond to a 20% chance of a disaster occurring in the next 50, 20, and 10 years, cor-
responding to 𝜆 = 0.004, 𝜆 = 0.01 and 𝜆 = 0.02, respectively. The abatement efficiencies
are 𝜎 = 0.08 ($12.5/t CO2 ), 𝜎 = 0.05 ($20/t CO2), and 𝜎 = 0.025 ($40/t CO2), as discussed
earlier.
The relative risk aversion (RRA) coefficient, 𝜀, is a subject of an ongoing debate in the
theoretical and empirical literature. Recall that we have assumed, following the risk-vulner-
ability literature, that preferences for risk-taking are affected by exposure to loss-generating
events. In particular, RRA is proportional to the parameter which governs the exposure
effect in the damage function, 𝜂 , with a constant proportionality coefficient 𝛾 . Instead of
calibrating 𝜂 , for which data are not readily available, we calibrate 𝜀 and then perform sen-
sitivity analysis with respect to both 𝜀 and 𝛾 . This approach also allows us to meaningfully
compare our results with those in the existing studies. We consider two calibrations for 𝜀
and check sensitivity of the results to variations in 𝛾 (in the Appendix), with the benchmark
values set to unity for both (log utility). For any given 𝜀, a value of 𝛾 larger (resp., smaller)
than unity would indicate a reduction (resp., increase) in 𝜂 and therefore a reduction (resp.,
increase) in damages. The benchmark parameter values are summarized in Table 1.
6.2 Low‑impact Events
Let us analyze the results pertaining to the first scenario (low 𝛿), summarized in Table 2.
In the left-hand panel (𝜀 = 1), with a more optimistic abatement efficiency (𝜎 = 0.08)
and a 20% chance of experiencing a climate-change driven disaster in the next 50 years
(𝜆 = 0.004), we find the optimal (worldwide) growth rate of 3.475% and the optimal abate-
ment propensity of 0.5%. To put the latter number in perspective and provide a comparison
with the existing literature, we can infer what this abatement propensity implies for the
price of 1 ton of emitted (and abated) carbon. In order to do so we multiply 𝜃 ∗ with the
world output and divide by tons of carbon emissions using the latest available data. Reuters
and World Bank report that global emissions in 2014 amounted to 10.7 bn tons of carbon
so that, with the global world output in 2014 at $78.28 trillion, the implied world carbon
price amounts to $36.6 per ton. The caveat of this approach is that we use yearly emissions,
which are in fact net emissions, that is, after some abatement has taken place in a given
year. We therefore treat our estimated carbon price as an indicative upper bound and focus
28
The intensities 𝛿 = 1e − 6 ("low") and 𝛿 = 1e − 5 ("high") correspond to damages of less than 0.1%
of gross world product (GWP) and 5 − 10% of GWP, respectively. Recall that in our model damages are
endogenous and hence depend not only on 𝛿 but on all other parameters. Calibrating 𝛿 proved to be a chal-
lenging task as there is no direct mapping to this parameter in the data. In order to circumvent this issue we
chose the values which deliver average worldwide damages from natural disasters over the last 5 years. Sev-
eral sources converge on those damages being in the range of several hundred billion USD for 2011–2015
(see http://www.economist.com/blogs/dailychart/2011/03/natural_disasters, http://www.theonebrief.com/
the-impact-of-natural-disasters-on-the-global-economy,https://www.weforum.org/agenda/2015/12/how-
much-do-natural-disasters-cost-the-world, http://www.kit.edu/kit/english/pi_2016_058_natural-disasters-
since-1900-over-8-million-deaths-and-7-trillion-us-dollars-damage.php). In particular, our lower value of
𝛿 corresponds to an average worldwide damage of approximately $370 bn which reflects the value in the
CATDAT database.
13
on the abatement propensity as our main policy variable of interest. This value is robust to
changes in the disaster arrival rate. Specifically, increasing the hazard rate from 0.004 to
0.02 has no significant impact on the optimal abatement propensity and growth. We shall
show shortly that this outcome is strictly linked to the logarithmic utility assumption and,
to some extent, to the low damage intensity of climate disasters. By contrast, changing
abatement efficiency from a relatively high value (𝜎 = 0.08) to a lower value (𝜎 = 0.05)
brings about an increase in the optimal abatement propensity from 0.5% to 0.8% and the
corresponding carbon price rises to $58.52 per ton which is comparable to the baseline
value $56.9/tC obtained by Golosov et al. (2014). Reduced abatement efficiency induces an
only slightly lower optimal growth rate of 3.46%.
Exposure to climate risks may lead to a higher degree of risk aversion according to
the risk-vulnerability hypothesis. This suggests that the unitary value of 𝜀 may be too low.
We thus consider a higher value: 𝜀 = 3. We find that under the benchmark calibration the
optimal abatement propensity increases slightly from 0.5 to 0.54%.29 The optimal growth
rate, however, drops from 3.47 to 1.16%. When abatement efficiency is less favorable, 𝜃 ∗
rises from 0.8 to 0.87% and the growth rate is significantly reduced from 3.46 to 1.15%. To
examine the sensitivity of the results to variations in 𝛾 , we replicate Table 2 for 𝛾 = 0.9 and
𝛾 = 1.1 in Appendix F.1. A higher (lower) value of 𝛾 for a given 𝜀 implies a lower (higher)
exposure to damages and hence a lower (higher) optimal abatement propensity. Comparing
the results from Table 2 with those in either Table 4 (𝛾 = 0.9) or Table 5 (𝛾 = 1.1), we find
no significant changes in either 𝜃 ∗ or g∗ for 𝜀 = 1 and only a few percentage points differ-
ences in 𝜃 ∗ (but not in g∗) for 𝜀 = 3. We anticipate that climate shocks with a larger damage
intensity may profoundly alter the optimal policy. In addition, the impact of a higher 𝜀 has
to be examined more carefully in the context of more severe disasters.
6.3 High‑impact Events
We consider next Table 3 with the same parameter constellation except for 𝛿 which is
increased 10-fold to deliver damages of up to 10% of GWP.30 First note that with log-utility
the optimal abatement propensity and the optimal growth rate are not affected in a major
way. However, when we set 𝜀 to 3, the picture changes significantly. First, looking at the
top panel (𝜆 = 0.004) and the optimistic abatement efficiency case (𝜎 = 0.08 ), we already
find that the optimal abatement propensity rises significantly from 0.541% to 0.956%,
which corresponds to an increase in the carbon price from about $40 to about $70 per ton.
The growth rate is reduced only marginally from 1.16 to 1.15%. Second, moving to the less
optimistic abatement efficiency scenario (𝜎 = 0.05) increases the abatement propensity
from 0.96 to 1.6%. Third, with more frequent disasters (𝜆 = 0.02), the abatement propen-
sity jumps 4–5 fold from 0.706 and 1.136 (for 𝜎 = 0.08 and 𝜎 = 0.05 in Table 2, resp.) to
2.9 and 5.18%, respectively. The latter implies a carbon price of $379 per ton. At the same
time the growth rate is reduced from 1.16 and 1.15 to 1.11 and 1.07%, respectively.
Finally, we can assess the difference between high-impact rare events and more common
low-impact events. To this end we compare the results from the top panel of Table 3 with the
29
This result is consistent with Barrage (2014) who finds that an increase in RRA from 1 to 2 increases the
optimal carbon tax only marginally, see Figs. S.10 and S.11 of her paper, from about $56/tC to slightly over
$60/tC.
30
Since damages are endogenous in our model, they vary between 5 and 10% of GWP, depending on the
values of other parameters.
13
Table 2 The optimal policy with low damage intensity: 𝛿 = 1e − 6 (up to 1% of GWP)
𝜀=1 𝜀=3
𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.025 𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.025
𝜆 = 0.004
𝜃∗ 0.500 0.800 1.600 0.541 0.867 1.739
g∗ 3.475 3.459 3.419 1.158 1.152 1.137
𝜆 = 0.01
𝜃∗ 0.500 0.801 1.601 0.603 0.967 1.949
g∗ 3.475 3.459 3.419 1.157 1.151 1.134
𝜆 = 0.02
𝜃∗ 0.501 0.801 1.602 0.706 1.136 2.305
g∗ 3.475 3.459 3.419 1.155 1.148 1.128
results from the bottom panel of Table 2. This corresponds to moving from a scenario with a
20% chance of experiencing a 0.05% reduction in GDP in the next 10 years to a scenario with
a 20% chance of experiencing a 5% drop in GDP in the next 50 years. Under log utility there
is almost no change in either the abatement propensity or the growth rate, regardless of the
value of 𝜎. With 𝜀 = 3 the growth rate reacts relatively moderately by falling by about half a
basis point. The abatement propensity, by contrast, reacts more strongly with an increase of 25
basis points (from 0.706 to 0.956). An even stronger increase of 128 basis points is observed
for the least optimistic abatement technology (𝜎 = 0.025). Even if we constrain the expected
damage to be exactly identical in both cases (we reduce 𝜆 down to 0.002073), the abatement
propensity is still higher (𝜃 = 1.205) and the growth rate is lower (g = 1.145) in the high-
damage scenario. We conclude that a possibility of rare but high-impact events calls for a
more stringent abatement policy as compared to relatively frequent but low-impact events
(keeping expected damages identical). Our results provide strong support of Stern’s hypothesis
that optimal climate policy becomes more stringent once rare high-impact events are taken
into consideration.
Our quantitative analysis leads to three conclusions. First, log-utility assumption in models
of climate change is innocuous only if environmental shocks are "not too severe", i.e. they
are characterized by a relatively low damage intensity. This is likely not to be the case for
natural disasters induced by climate change. Damages from such catastrophes have amounted
to approximately 1% of GDP of countries where they happened to strike. Intensity of disas-
ters will worsen even further due to the planet’s warming, according to predictions of climate
physicists. Second, efficiency of abatement technology plays an important role as well and
more so when disasters are high-impact. Decreasing abatement efficiency by about one third
from $12.5/tCO2 to $20/t CO2 requires an increase in the abatement propensity by more than
a half if shocks are low-impact and by three quarters when they are high-impact. Third, a
prospect of rare but high-impact events calls for a more stringent mitigation policy than of
low-impact frequent events with the same expected damage.
13
Table 3 The optimal policy with high damage intensity: 𝛿 = 1e − 5 (up to 10% of GWP)
𝜀=1 𝜀=3
𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.025 𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.025
𝜆 = 0.004
𝜃∗ 0.502 0.802 1.605 0.956 1.589 3.587
g∗ 3.475 3.459 3.419 1.149 1.139 1.096
𝜆 = 0.01
𝜃∗ 0.504 0.806 1.612 1.662 2.846 7.139
g∗ 3.475 3.459 3.419 1.137 1.114 1.016
𝜆 = 0.02
𝜃∗ 0.508 0.812 1.624 2.905 5.178 16.49
g∗ 3.475 3.454 3.418 1.114 1.068 0.792
7 Conclusions
An increase in the global temperature is predicted to render natural disasters, e.g. tropical
storms, hurricanes, tsunamis, floods, droughts, etc., more severe and intense. Such calami-
ties have a profound negative impact on the economy’s infrastructure, physical and human
capital, and they undoubtedly represent a set-back in terms of economic growth and devel-
opment. An efficient and timely climate policy is necessary in order to limit damages from
such devastating shocks.
In the present article we propose a model of a growing economy subject to random nat-
ural disasters, which destroy part of the economy’s productive input. An important feature
of our model is consideration of recurring shocks where the extent of the damage is endog-
enously determined through the interaction between capital accumulation process and an
appropriate emissions abatement policy. Our contribution is three-fold. First, we deliver
explicit analytical solutions for (i) the benchmark model, where the hazard rate is treated as
exogenous and only damages depend on the accumulated emissions; and (ii) an extended
model, where the hazard rate is also an increasing function of atmospheric GHG. In the
case of a unitary elasticity of marginal utility (logarithmic preferences), often used in the
literature on the grounds of better tractability, the consumption-to-capital ratio and the pro-
pensity to save are independent of the climatic parameters. Also, with log-preferences the
endogeneity of the arrival rate does not alter the optimal mitigation expenditure. In light
of these properties, policy-relevance and applicability of findings stemming from mod-
els based on the log-utility assumption seem to suffer from some limitations. Assuming a
non-unitary relative risk aversion reveals important macroeconomic interdependencies and
alters policy conclusions significantly.
Second, we emphasize the importance of simultaneously determining the optimal
dynamic behavior of the economy and the abatement of GHG, since the equilibrium
growth rate serves as a complement to the climate policy. When the economy is exposed
to stochastic climate shocks it may choose to lower its optimal rate of growth so as to limit
its GHG emissions—and thus potential damages if a shock occurs—but also to reduce its
expenditure on abatement.
Third, we provide quantitative implications by calibrating our model to the recent data on
global carbon emissions, output, frequency of large natural catastrophes and their damages.
13
With log utility, the share of output which should be devoted to emissions control is approxi-
mately 0.5% of GWP. This number is comparable to what has been found in the recent studies
which relied on log-utility assumption. However, when we use a higher value for the relative
risk aversion parameter, we find that the abatement propensity increases significantly, reach-
ing 3–5% of GWP, as we consider higher degrees of severity of environmental shocks. These
values are equivalent to a carbon price in the range $212–$370 per ton. If one takes the side
of climate physicists who believe that climate change will cause an increase in disaster fre-
quency—in addition to the damage intensity—then an even more stringent climate policy
becomes warranted.
By introducing random large climate shocks and tipping points, arriving with a known
stochastic process, our approach deals with the role of risk. In situations where objec-
tive probabilities or assessments of subjective probabilities of the underlying alternatives
are not possible, we arrive at the theory of decision-making under deep uncertainty and
ambiguity, see Bühren et al. (2021) for a recent survey. It has been shown that agents are
ambiguity-averse, which is conceptually different from being risk-averse and may decrease
the demand for self-protection, see Alary et al. (2013) and Mittelstaedt and Baumgärtner
(2023) for a theoretical foundation. The framework that we proposed can be viewed as a
situation where uncertainty becomes resolved over time due, for example, to learning, as
in Brock and Xepapadeas (2020). The latter contribution stands out as a unique study that
offers an analysis of social cost of carbon under deep uncertainty. They find that the lat-
ter contributes to more conservative climate policies and that, in addition, the economy is
willing to spend part of its output on learning which allows to partially reduce this uncer-
tainty. We conjecture that introducing deep uncertainty into our model will yield similar
results, i.e. the abatement propensity would be larger, while the optimal growth rate would
be lower.
Appendix to Section 2
Solution of the Deterministic Model
The Interior Solution
In what follows, we omit the time subscripts when there is no ambiguity and use subscripts
to denote partial derivatives. By differentiating the current-value Hamiltonian, we obtain
the following first-order conditions for the interior solution:
𝜕H
= 0 ⇔ 𝜆 = Uc� , (A.1)
𝜕C
𝜕H
= 0 ⇔ 𝜆 = −𝜇𝜎, (A.2)
𝜕𝜃
𝜕H
= 𝜌𝜇 − 𝜇̇ ⇔ 𝜒P𝛽 = 𝜇̇ − (𝜌 + 𝛼)𝜇, (A.3)
𝜕P
𝜕H
= 𝜌𝜆 − 𝜆̇ ⇔ 𝜆̇ + 𝜆[(1 − 𝜃)A − 𝜌] + 𝜇(𝜙 − 𝜃𝜎)A = 0. (A.4)
𝜕K
13
Combined with the boundary conditions they can be written as

𝜆 = −𝜇𝜎, (A.5)
𝜇̇ − (𝛼 + 𝜌)𝜇 = 𝜒P𝛽 , (A.6)
𝜆̇ + 𝜆[(1 − 𝜃)A − 𝜌] + 𝜇(𝜙 − 𝜃𝜎)A = 0, (A.7)
Ṗ = −𝛼P + (𝜙 − 𝜎𝜃)Y, (A.8)
K̇ = (1 − 𝜃)AK − C, (A.9)
Equation (A.5) combined with Eq. (A.7) gives
[ ]
1
𝜆̇ + 𝜆 (1 − 𝜃)A − 𝜌 − (𝜙 − 𝜃𝜎)A = 0
𝜎
( )
resulting in 𝜆̇ + (r − 𝜌)𝜆 = 0, where r = A 1 − 𝜙𝜎 > 0. Solving this odinary differential
equation (ODE), we obtain 𝜆t = e−(r−𝜌)t 𝜆0 . Combining the latter with (??), we get
𝜇̇ = − 𝜎1 𝜆̇ = r−𝜌
𝜎
𝜆t . After substitution into (A.6), we get
(r + 𝛼)𝜆t = 𝜎𝜒P𝛽
which can be rewritten as

[ 𝜆 (r + 𝛼) ] 1 r−𝜌
𝛽 − t
(A.10)
0
Pt = e 𝛽 .
𝜎𝜒
( ) [ ]1
− r−𝜌 𝜆0 (r+𝛼) 𝛽
Next we can rewrite Ṗ + 𝛼P as 𝛼 − r−𝜌 with . Using this in Eq.
t
P0 e 𝛽 P0 =
( ) r−𝜌
𝛽 𝜎𝜒
−
(A.8), we obtain 𝛼 − r−𝜌 P0 e 𝛽 = (𝜙 − 𝜎𝜃)AK , which implies that 𝜙 − 𝜎𝜃t → 0. Fur-
t
𝛽
̇
ther (A.8) can be rewritten as 𝜙 − 𝜎𝜃 = P+𝛼P
AK
resulting in
( )
𝜙 𝛼 − r−𝜌
𝛽
Pt
𝜃t = − .
𝜎 𝜎AKt
Since Ct = 𝜆t by (A.1), Eq. (A.9) can be rewritten as the following ODE:

−1∕𝜀
( )
( ) 𝛼 − r−𝜌
− r−𝜌 t
P0 e 𝛽
𝜙 𝛽 (r−𝜌)t
K̇ = 1 − AK + − C0 e 𝜀
𝜎 𝜎
Using the standard transversality condition, defining g = 1𝜀 (r − 𝜌), and requiring that g < r ,
we obtain
( )
c P0 𝛼 − 𝜀g 𝛽 − 𝜀g t
Kt∗ = 0 egt − ( ) e 𝛽 (A.11)
r−g 𝜎 r + 𝜀g𝛽
13
resulting in abatement propensity

( )
𝜀g
𝜙 − 𝛼 Pt
𝜃t∗ = +
𝛽
. (A.12)
𝜎 𝜎AKt∗
Now that we have solved for the full dynamics of the economy in the interior regime, we
can derive the explicit value function associated with the interior solution. From (A.11),
we can express consumption as a function of the two state variables and therefore we can
write the value function as

V(K, P) = −x ,
1−𝜀 1+𝛽
1−𝜀 a(𝛼−gp )
where the constants are given by a = (r − g) 𝜀 ,b= r+gp
,x= 𝛽𝜒
(1+𝛽)r−𝜌
. We are going to
be interested in the long-run (or asymptotic) value function such that t → ∞.
Maximum Abatement: = ̄
Suppose that the initial level of pollution P0 is relatively high, so that maximum abatement is
warranted from the outset. Now the problem is
∞
∫
max e−𝜌t U(C, P)dt
c,K,P
0
s.t.
K̇ = A(1 − 𝜃)K
̄ −C , 𝜆≥0 (A.13)
Ṗ = (𝜙 − 𝜃𝜎)AK
̄ − 𝛼P , 𝜇≤0 (A.14)
It is not difficult to see that there exists T > 0 such that for t < T the adjoint variable 𝜐t > 0,
i.e. the control 𝜃 = 𝜃̄ is applied for some finite time T only. The dynamics of the economy
are such that pollution stock declines at the rate exceeding gp because now abatement is
at its maximal value, while consumption and capital stock grow at the rate smaller than g.
The bahavior of the state variables is illustrated by the red dash-dotted line in Fig. 1.
No abatement: t ≡ 0
If the initial conditions are such that the capital stock and/or pollution stock are small
enough, then it may be optimal to do zero abatement, at least initially. Now the problem is
∞
∫
max e−𝜌t U(C, P)dt
c,K,P
0
s.t.
K̇ = AK − C , 𝜆≥0 (A.15)
13
Ṗ = −𝛼P + 𝜙AK , 𝜇≤0 (A.16)

The optimality conditions are
𝜕H
= 0 ⇔ 𝜆 = UC� (A.17)
𝜕C
𝜕H
= 𝜌𝜇 − 𝜇̇ ⇔ 𝜒P𝛽 = 𝜇̇ − (𝜌 + 𝛼)𝜇 (A.18)
𝜕P
𝜕H
= 𝜌𝜆 − 𝜆̇ ⇔ 𝜆̇ + (A − 𝜌)𝜆 + 𝜑A𝜇 = 0. (A.19)
𝜕K
with 𝜎 < −𝜇𝜆
. In the long run, pollution stock is proportional to capital stock, which implies
a smaller welfare than with the interior solution with 𝜃 > 0. Hence, the adjoint variable
𝜈t = 0 for t > T̃ , i.e. this control is applied for some finite time only. The bahavior of the
state variables is illustrated by the green dashed line in Fig. 1.
A Note on Asymptotic Value Function
The optimal pollution and capital stocks in the interior regime are given by (9) and (10).
One can show that

V(K, P) = −x ,
1−𝜀 1+𝛽
where (K, P(K)) belong to the "interior curve", which we will call curve (set) B. Curve B
is parametrized with the time-variable t, which is convenient to treat now as a parameter.
We can substitute (Kt , Pt ) into the value function. After that we can rewrite the resulting
function V(t) as follows:
Kt1−𝜀 P1+𝛽
t
V(t) = V(Kt , Pt ) = 𝜓t −x
1−𝜀 1+𝛽
(aKt +bPt )1−𝜀
where 𝜓t = Kt1−𝜀
= a1−𝜀 (1 + ) .
b Pt 1−𝜀
a Kt
Function V(t) defines lifetime-utility of an
agent with initial condition (k, p) = (Kt , Pt ) = (Kt , P(Kt )) (note that mapping t ↦ (Kt , Pt ) is
one-to-one, so choosing parameter t is equivalent to picking some point in B). In other
words, knowing function V(t) is equivalent to knowing V(K, P(K)), since one can always
change variables back.
( )(r−𝜌)
From (1) and (2) we have Kt = O exp(−(1∕𝛽 + 1∕𝜀) t) ↘ 0, when t → ∞ (or,
P
t
equivalently, when Kt → ∞ or, equivalently, when Pt → 0.
Since (1 + x)𝛼 = 1 + 𝛼x + o(x), when x → 0, 𝛼 > 0, we can write
P (P )
𝜓t = 𝜓 ∗ + (1 − 𝜀)b1−𝜀 t + o t ,
Kt Kt
where 𝜓 ∗ stands for a1−𝜀 .
13
Keeping in mind that V(t) defines the value function in the interior regime, let’s intro-
1−𝜖 1+𝛽
duce another function Ṽ ∶ ℝ+ → ℝ given by the identity V(t)
̃ = 𝜓 ∗ K − x P . We call Ṽ
1−𝜖 1+𝛽
the asymptotic (or long-run) VF.
The resulting (asymptotic) consumption C̃ and (asymptotic) abatement policy 𝜃̃ = 𝜙∕𝜎 are
slightly different from the optimal policies C∗ , 𝜃 ∗ derived above. This is because
̃ ≠ V(Kt , Pt ), but instead Ṽ t ��→
V(t) � 1.
V(Kt ,Pt ) t→∞
1−𝜖 1+𝛽
Let V(K, P) = 𝜓 ∗ K1−𝜖 − x P1+𝛽 + O(P∕K) be asymptotic expansion of function V(K, P(K)),
when K → ∞. Let C∗ , 𝜃 ∗ denote optimal policies (with smoothness assumptions). Then there
exist a1 and a2 such that for asymptotic policies C̃ , 𝜃̃ errors are bounded |C̃ − C∗ | ≤ a1 P∕K
and |𝜃̃ − 𝜃 ∗ | ≤ a2 P∕K .
Interior Solution to the Baseline Stochastic Model
Hamilton-Jacobi-Bellman equation for the problem may be written as

{
𝜌V(K, P) = max U(C, P) + Vk [(1 − 𝜃)Y − C] + Vp [(𝜙 − 𝜃𝜎)Y − 𝛼P]+
C,𝜃
[ ]} (B.1)
̃
+𝜆 V(K, P) − V(K, P) ,
and the first-order conditions with respect to the control and the state variables consist of
C∶ Uc − Vk = 0, (B.2)
𝜃∶ −Vk Y − Vp 𝜎Y = 0, (B.3)
K ∶ 𝜌Vk = Vkk [(1 − 𝜃)Y − C] + Vk (1 − 𝜃)A + Vp (𝜙 − 𝜎𝜃)A + Vpk [(𝜙 − 𝜎𝜃)Y − 𝛼P]+
( )
̃
+𝜆 V �k dK − Vk ,
dK
(B.4)
̃
P∶ �k dK − 𝛼Vp +
𝜌Vp = Up + Vpp [(𝜙 − 𝜎𝜃)Y − 𝛼P] + Vkp [(1 − 𝜃)Y − C] + 𝜆V
( ) dP
+𝜆 V�p − Vp .
(B.5)
Suppose the long-run value function is of the form
𝜓 −𝜀 K 1−𝜀 xP1+𝛽
V(K, P) = − ,
1−𝜀 1+𝛽
then from (B.2) and (B.3),
C =𝜓K (B.6)
13
(𝜓K)−𝜀 =x𝜎P𝛽 . (B.7)
Finding the Coefficients
Substituting (B.6) and (B.7) into the HJB equation we find:
𝜌𝜓 −𝜀 K 1−𝜀 𝜌xP1+𝛽 (𝜓K)1−𝜀 P1+𝛽

− = −𝜒 + (𝜓K)−𝜀 [(1 − 𝜃)A − 𝜓]K−
1−𝜀 1+𝛽 1−𝜀 1+𝛽
(𝜓K)−𝜀 ̃ 1−𝜀 𝜓 −𝜀 K 1−𝜀 ]
[ 𝜓 −𝜀 K
− (𝜙 − 𝜃𝜎)AK + 𝛼xP1+𝛽 + 𝜆 − ,
𝜎 1−𝜀 1−𝜀
To find 𝜓 and x, we match the terms with K and P, respectively, on the left and on the
right-hand sides. Collecting the terms in K we obtain:
𝜌𝜓 −𝜀 K 1−𝜀
1−𝜀
( )
(𝜓K)1−𝜀 𝜙 𝜓 −𝜀 K 1−𝜀 1−𝜀
= + 𝜓 −𝜀 [(1 − 𝜃)A − 𝜓]K 1−𝜀 − 𝜓 −𝜀 − 𝜃 AK 1−𝜀 + 𝜆 (𝜔 − 1).
1−𝜀 𝜎 1−𝜀
−𝜀 1−𝜀
Dividing both sides by 𝜓 1−𝜀
K
yields
( )
𝜙
𝜌 = 𝜀𝜓 + (1 − 𝜀)A 1 − + 𝜆(𝜔1−𝜀 − 1), (B.8)
𝜎
where 𝜔 = exp[−𝛿(𝜎x𝜓 𝜀 )−1∕𝛾 ].

Collecting the terms in P, we obtain:
𝜌xP1+𝛽 P1+𝛽
− = −𝜒 + 𝛼xP1+𝛽 .
1+𝛽 1+𝛽
1+𝛽
Dividing both sides by P1+𝛽 yields:
−𝜌x = −𝜒 + 𝛼(1 + 𝛽)x
which implies that

𝜒
x=
𝜌 + 𝛼(1 + 𝛽)
. (B.9)
Solving for Endogenous Variables
Consumption Dynamics
We solve the optimal growth rate by combining (B.2) and (B.4). First, let us divide (B.4)
throughout by Vk:
( )
Vkk Vp �k dK̃
V
𝜌= [(1 − 𝜃)AK − 𝜓K] + (1 − 𝜃)A + (𝜙 − 𝜎𝜃)A + 𝜆 −1 ,
Vk Vk Vk dK
V ̃ ̃
Using the fact that Vp = − 𝜎1 , Vk = 𝜔−𝜀, ddKK = + 𝜔, we get
V 𝜕𝜔
𝜕K
K
k k
13
Vkk 𝜙A [ ( ) ]
𝜕𝜔 K
𝜌= [(1 − 𝜃)A − 𝜓]K + (1 − 𝜃)A − + 𝜃A + 𝜆 𝜔1−𝜀 +1 −1 .
Vk 𝜎 𝜕K 𝜔
( )
V 𝜙 [ ]
𝜌 = kk [(1 − 𝜃)A − 𝜓]K + A 1 − + 𝜆 𝜔1−𝜀 (𝜂 ln(𝜔) + 1) − 1 .
Vk 𝜎
From the above equation we obtain

( )
Vkk 𝜙 [ ]
[(1 − 𝜃)A − 𝜓]K = 𝜌 − A 1 − − 𝜆 𝜔1−𝜀 (𝜂 ln(𝜔) + 1) − 1 (B.10)
Vk 𝜎
In the next step we use the Change of Variable Formula for jump process to compute the
differential of Uc. Since we know that Uc = Vk,
{ } [ ]
̃ − Uc (C) dq
dUc = Vkk [(1 − 𝜃)A − 𝜓]K dt + Uc (C)
Dividing both sides by Uc or, equivalently, by Vk , yields

{ } [ ̃ ]
dUc Vkk Uc (C)
= [(1 − 𝜃)A − 𝜓]K dt + − 1 dq,
Uc Vk Uc (C)
{ } [ ̃ ]
Vkk Uc (C)
= [(1 − 𝜃)A − 𝜓]K dt + − 1 dq
Vk Uc (C)
Using (B.10), yields

{ ( ) } [ ̃ ]
dU 𝜙 Uc (C)
− c = A 1− + 𝜆𝜔1−𝜀 [𝜂 ln(𝜔) + 1] − (𝜆 + 𝜌) dt − − 1 dq
Uc 𝜎 Uc (C)
{ ( ) }
dC 1 𝜙
= A 1− + 𝜆𝜔1−𝜀 [𝜂 ln(𝜔) + 1] − (𝜆 + 𝜌) dt + (1 − 𝜔−𝜀 )dq
C 𝜀 𝜎
The expression in the curly braces multiplying dt is the trend growth rate of consumption,
i.e. the growth rate of C in between disasters, which we denote by gc:
{ ( ) }
1 𝜙
gc = A 1− + 𝜆𝜔1−𝜀 [𝜂 ln(𝜔) + 1] − (𝜆 + 𝜌) . (B.11)
𝜀 𝜎
Since C = 𝜓K , the trend growth rate of consumption is the same as the trend growth rate
of K ( gk ) in the long run. Hence, we can write
gk = gc .
Pollution and Abatement Dynamics

For now we will only be concerned with trend growth rates and will discuss the
jumps later on. From (B.7) we know that
𝜀gk + 𝛽gp = 0.
From the dynamic motion of pollution stock we know that gp = (𝜙 − 𝜎𝜃t )A Pt − 𝛼 . Insert-
K
t
ing this equation into the above yields
13
[ ]
Kt
𝜀gk + 𝛽 (𝜙 − 𝜎𝜃t )A − 𝛼 = 0. (B.12)
Pt
This implies, in particular, that

𝜀gc
gp = −
𝛽
. (B.13)
In addition, from (B.5) we know that

𝜒 K
𝜌= + 𝛽gp − 𝜆𝜎𝜉𝜔1−𝜀 ln(𝜔) − 𝛼. (B.14)
x P
To get this latter result, it is sufficient to divide (B.5) throughout by Vp and recall that x is
constant. Solving for K/P and using the solution for x, we obtain
K 𝛽(gp + 𝛼)
= .
P 𝜆𝜎𝜉𝜔1−𝜀 ln(𝜔)
Finally, inserting this result into (B.12) yields
𝛼𝛽 − 𝜀gc (𝛼 + gp )𝜆𝜎𝜉𝜔1−𝜀 ln(𝜔) 𝜆𝜎𝜉𝜔1−𝜀 ln(𝜔)

𝜙 − 𝜎𝜃t = 𝛽(gp +𝛼)
= = . (B.15)
𝛽A 𝜆𝜎𝜉𝜔1−𝜀 ln(𝜔) 𝛽(gp + 𝛼)A 𝛽A
Hence
1−𝜀
𝜙 𝜆𝜉𝜔t ln(𝜔t )
𝜃t = − . (B.16)
𝜎 𝛽A
Disentangling the Precautionary Effect
We are interested in the magnitude of 𝜁 = 𝜔1−𝜀 (1 + 𝜀 ln 𝜔) relative to unity. The

ratio of marginal utilities, given by 𝜔−𝜀 is unambiguously larger than unity, while the
effect of K on K̃ can be larger or smaller than 1. This effect includes the direct positive
effect, given by 𝜔 > 0, and an indirect negative effect, given by a change in 𝜔 itself,
K(𝜕𝜔∕𝜕K) = 𝜔𝜀 ln 𝜔 < 0. Whether the direct effect dominates the indirect effect, depends
on the magnitude of 𝜀.31 If 𝜀 < −(ln 𝜔)−1, the direct effect dominates, implying that
̃
dK∕dK > 0. The threshold −(ln 𝜔)−1 may take values between approximately 3.47 and 99
31
With the assumption 𝜀 = 𝜂 we can also relate the magnitudes of the direct and indirect effects of K on K̃
to the parameter governing the strength of the exposure component in the damage function. Recall that if
𝜂 < 1, the survived share of capital is convex in K, i.e., as capital stock grows its marginal vulnerability to
shocks declines. Alternatively, if 𝜂 exceeds a specific threshold, the marginal vulnerability increases. The
threshold is given by (1 + ln 𝜔)−1. For relatively small damages, say 1% (𝜔 = .99), the threshold is approxi-
mately unity. For relatively large damages, say 25% (𝜔 = .75), the value is around 1.4. Given that for the
relevant range of the damage magnitude the threshold values are close to one, we shall distinguish primarily
between two possibilities, 𝜂 < 1 and 𝜂 > 1. This is also convenient for our further analysis since we need to
consider alternative values of 𝜀, with the usual cut-off value of one.
13
for 𝜔 ∈ [0.75, 0.99]. Clearly, the condition is largely satisfied for the relevant range of 𝜔
and empirically-plausible range of 𝜀 (see Fig. 3), hence dK∕dK̃ > 0.32
̃ dK̃
U � (C)
We still need to find out, however, whether 𝜁 = U � (C) dK is larger or smaller than unity.
So far we have established that an increase in K will lead to an increase in the post-event
capital stock but less than proportionally since 𝜔(1 + 𝜀 ln 𝜔) < 1 (recall that 𝜔 ∈ (0, 1) and
ln 𝜔 < 0). The valuation of this increase is given by the ratio of marginal utilities 𝜔−𝜀 > 1.
The total effect, 𝜁 = 𝜔1−𝜀 (1 + 𝜀 ln 𝜔), can thus be larger or smaller than unity, depending
on the magnitude of 𝜀.
Proof of Proposition 1 Given the value function, (i) follows directly from (B.2), so that
consumption-to-capital ratio is constant and equal to 𝜓 ∗. The statement in (ii) follows
from (24) and the fact that 𝜓 ∗ is constant. Then g∗ is the growth rate of C, K, and Y. Since
abatement share is constant, total abatement expenditure grows at the same rate as well,
which completes the proof. ◻
Non‑constant Marginal Abatement Cost
We now assume that the abatement efficiency, 𝜎 , is a function of the abatement propen-
sity, 𝜃 , such that 𝜎 = 𝜎(𝜃), 𝜎 � < 0, 𝜎 �� ≥ 0, 𝜎(0) = 0: a larger proportional expenditure
on abatement makes it harder to abate, reflecting increasing marginal abatement cost. For
practical purposes, we assume a power function: 𝜎 = 𝜎𝜃 ̄ a , a < 0. Our original model is
included as a special case a = 0. The total abatement can be written as 𝜎𝜃̄ 1+a AK .
This modification of the model leads to changes in the first-order conditions with respect
to 𝜃 , K, and P. The key difference to our original model is that the solution to the current
model is a solution to the system of 6 equations (4 FOCs plus 2 constraints), while in our
original model it was possible to separate the system into sub-components and solve them
separately. After some rearrangements, the current system becomes
𝜒
x=
𝜌 + 𝛼(1 + 𝛽) (C.1)
( )− 1
𝜃 = 𝜓 𝜀 K 𝜀 𝜎(1
̄ + a)xP𝛽 a (C.2)
[ ]
𝜙 𝜃
𝜌 =𝜀𝜓 + (1 − 𝜀)A 1 − 𝜃 − + + 𝜆(𝜔1−𝜀 − 1) (C.3)
̄ + a)𝜃 a 1 + a
𝜎(1
{ [ ] }
1 𝜙 𝜃
(C.4)
1−𝜀
gc = A 1−𝜃− + + 𝜆𝜔 [𝜂 ln 𝜔 + 1] − (𝜆 + 𝜌)
𝜀 ̄ + a)𝜃 a 1 + a
𝜎(1
32
Only for extremely large damages, exceeding 25% of GDP, the condition 𝜀 < −(ln 𝜔)−1 might be vio-
lated, implying that dK∕dK
̃ may become negative. Consequently, the term (𝜁 − 1) in (22) may become neg-
ative as well, implying a precautionary dissaving motive. The intuition here is clear, since the overall effect
of an increase in the capital stock on K̃ is negative, the optimal response of the economy is to decrease its
growth and saving rates.
13
Fig. 3 Precautionary effect and risk aversion
[ ]
AK
̄ 1+a )
0 =𝜀gk + 𝛽 (𝜙 − 𝜎𝜃 −𝛼 (C.5)
P
𝜙 (1 + a)𝜉𝜆𝜔1−𝜀 ln 𝜔
𝜃= − (C.6)
̄ a
𝜎𝜃 𝛽A
It can be easily seen that if we set a = 0 we are back to the solution of our original model.
Further, by examining the last equation, one can see that if 𝜎̄ is set equal to the abatement
efficientcy of our beaseline model, 𝜎 , then the optimal abatement propensity in the model
with increasing marginal abatement cost is smaller than in our baseline model, given
in (24). This is because the first term in (C.6) is smaller than 𝜙∕𝜎 and, given that a < 0,
the second positive term is also smaller. This result conforms with our intuition since an
increasing MAC will necessarily induce less abatement.
Comparative Statics
M x
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞ ⏞⏞⏞ ⎛ d𝜆 ⎞
� � � � � � � �
𝜀 1 − 𝜆(1 − 𝜀)𝜔1−𝜀 𝜓 −1 ln 𝜔 0 d𝜓 Δ𝜓𝜆 Δ𝜓𝛿 Δ𝜓𝜎 Δ𝜓𝜙 ⎜ d𝛿 ⎟
𝜀𝜔 𝜓 1−𝜀 −𝜀−1 𝜆𝛿[1 + (1 − 𝜀) ln 𝜔] xA𝜎
×
d𝜃
=
Δ𝜃𝜆 Δ𝜃𝛿 Δ𝜃𝜎 Δ𝜃𝜙
× ⎜ d𝜎 ⎟ ,
⎜ ⎟
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⎝ d𝜙 ⎠
Δ ⏟⏟⏟
y
where
⎛ 1 − 𝜔1−𝜀 −𝜀 1−𝜀
� 𝛿𝜓 𝜔 � ⎞
⎜ 𝜆(1 − 𝜀)𝜔1−𝜀 𝜓 −𝜀 (x𝜎)−1 𝜆𝜓 −𝜀 𝜔1−𝜀
1 − 𝛿(1 − 𝜀)𝜓 −𝜀 (x𝜎)−1 ⎟
Δ� = ⎜ 𝜀−1 �A𝜙 + 𝜆𝜔1−𝜀 𝛿𝜓 −𝜀 x−1 � 1� � � �
𝜆𝛿𝜓 −𝜀 𝜔1−𝜀 (1 − 𝜀)𝛿𝜓 −𝜀 (x𝜎)−1 − 1 − 𝜙xA ⎟
⎟
⎜ 𝜎2 𝜎
⎜ (1 − 𝜀) A𝜎 xA ⎟
⎝ ⎠
and
13
[ ]
|M| = M11 M22 = 𝜀 1 − 𝜆(1 − 𝜀)𝜔1−𝜀 𝜓 −1 ln 𝜔 xA𝜎.
Determinant of M matrix
We are interested in the minimum possible threshold and thus need to consider only
a minimum value for 𝜓 . The share of consumption in the total capital stock, 𝜓 , is com-
puted as consumption share in GDP, multiplied by the total factor productivity, assumed
to be 5%. With the consumption share as low as 60%, this gives a value of 0.03. His-
torically, high-impact events (small 𝜔 ) are associated with rare occurrence (small 𝜆 ) and
low-impact events are more frequent. We consider a range of possibilities, from fairly
common disasters to rare catastrophes. Consider first the former category. When the
maximum value of 𝜆 is equal to 0.5 and the corresponding damage is 1% ( 𝜔 = 0.99),
the threshold value of 𝜀 , denoted by 𝜀M11, is above 6.5. For a more destructive but at the
same time less frequent event ( 𝜔 = 0.95, 𝜆 = 0.1), 𝜀M11 > 5.5. In the case of very rare
and extremely damaging events the threshold remains in a similar range. For 𝜆 = 0.004
and 𝜔 = 0.7, 𝜀M11 > 5.3. For 𝜆 = 0.001 and 𝜔 = 0.5, 𝜀M11 > 4.5. All of these values are
beyond the empirically relevant range.
Proof of Proposition 2
d𝜃 ∗ |M𝜃𝜆 | { }
= = −A−1 𝜔1−𝜀 ln 𝜔 1 − 𝜆[(1 − 𝜀) ln 𝜔 + 1]𝜀𝜓 −1 (1 − 𝜔1−𝜀 )M11
−1
> 0,
d𝜆 |M|
(D.1)
d𝜃 ∗ |M𝜃𝛿 |
= = 𝜆𝜔1−𝜀 [(1 − 𝜀) ln 𝜔 + 1]𝜀𝜓 −𝜀 |M|−1 > 0, (D.2)
d𝛿 |M|
[ ]
d𝜃 ∗ |M𝜃𝜎 | 𝜙 𝜆𝜔1−𝜀 ln 𝜔 𝜀(1 − 𝜀)𝜃 ∗ A
= =− 2 + [(1 − 𝜀) ln 𝜔 + 1] 1 − < 0, (D.3)
d𝜎 |M| 𝜎 A𝜎 𝜓M11
d𝜃 ∗ |M𝜃𝜙 | 𝜀 [ ]
= = 1 + 𝜆𝜔1−𝜀 (1 − 𝜀)2 (ln 𝜔)2 𝜓 −1 > 0. (D.4)
d𝜙 |M| 𝜎M11
Proof of Corollary 2 By differentiating the explicit solution for 𝜃 ∗ when 𝜀 = 1 we find that
d𝜃 ∗ 𝛿
= > 0,
d𝜆 𝜌x𝜎A
d𝜃 ∗ 𝜆
= > 0,
d𝛿 𝜌x𝜎A
( )( )
d𝜃 ∗ 1 𝛿𝜆
= − 2 𝜙+ < 0,
d𝜎 𝜎 𝜌xA
d𝜃 ∗ 1
= > 0.
d𝜙 𝜎
13
[ ]
dg 𝜕g 𝜕g 𝜕𝜔 d𝜓 𝜕𝜔
= + + (D.5)
d𝜆 𝜕𝜆 𝜕𝜔 𝜕𝜓 d𝜆 𝜕𝜆
{ }
1
= 𝜔1−𝜀 (1 + 𝜀 ln 𝜔) − 1 − 𝜆𝜀𝜓 −1 𝜔1−𝜀 (1 − 𝜔1−𝜀 ) ln 𝜔M11
−1
Γ < 0, (D.6)
𝜀
[ ]
= + + (D.7)
d𝛿 𝜕𝛿 𝜕𝜔 𝜕𝜓 d𝛿 𝜕𝛿
( )2
= − ln 𝜔 𝜆𝜔1−𝜀 (1 − 𝜀)𝜓 −𝜀−1 A|M|−1 Γ < 0 for 𝜀 > 1, (D.8)
[ ]
= + + (D.9)
d𝜎 𝜕𝜎 𝜕𝜔 𝜕𝜓 d𝜎 𝜕𝜎
A𝜙 [ ] ( )2 (1 − 𝜀)
= 1 + 𝜆𝜔1−𝜀 (ln 𝜔)2 (1 − 𝜀)2 𝜀𝜓 −1 − 𝜆𝜔1−𝜀 ln 𝜔 Γ > 0, for 𝜀 > 1
𝜎2 𝜓𝜎
(D.10)
[ ]
dg 𝜕g 𝜕g 𝜕𝜔 d𝜓 𝜕𝜔 𝜀A [ ]
= + + =− 1 + 𝜀𝜆𝜔1−𝜀 (1 − 𝜀)2 (ln 𝜔)2 𝜓 −1 < 0,
d𝜙 𝜕𝜙 𝜕𝜔 𝜕𝜓 d𝜙 𝜕𝜙 𝜎M11
(D.11)
where Γ ≡ [(1 − 𝜀)(1 + 𝜀 ln 𝜔) + 𝜀] > 0.
Consumption-Capital Ratio
{ x𝜎𝜓 1+𝜀 𝜔𝜀−1
d𝜓 |M𝜓𝜆 | > 0, if 𝜀 ∈ (0, 1) or 𝜀 > 1 +
= = (1 − 𝜔1−𝜀 )xA𝜎|M|−1 x𝜎𝜓 1+𝜀 𝜔𝜀−1
𝜆𝛿
d𝜆 |M| ⩽ 0, if 1 ⩽ 𝜀 < 1 + .
𝜆𝛿
(D.12)
d𝜓 |M𝜓𝛿 | > 0, if 𝜀 ∈ (0, 1) or 𝜀 > 1 +
= = 𝜆𝜔1−𝜀 A(1 − 𝜀)𝜓 −𝜀 |M|−1 𝜀−1
𝜆𝛿
|M|
1+𝜀
x𝜎𝜓 𝜔
d𝛿 ⩽ 0, if 1 ⩽ 𝜀 < 1 + .
𝜆𝛿
(D.13)
d𝜓 |M𝜓𝜎 | [ ] −1 < 0, if 𝜀 ∈ (0, 1) or 𝜀 > 1 +
= = (𝜀 − 1) xA𝜙 + 𝜆𝜔1−𝜀 𝜓 −𝜀 𝛿 A(𝜎|M|) x𝜎𝜓 1+𝜀 𝜔𝜀−1
𝜆𝛿
d𝜎 |M| ⩾ 0, if 1 ⩽ 𝜀 < 1 + .
𝜆𝛿
(D.14)
d𝜓 |M𝜓𝜙 | > 0, if 𝜀 ∈ (0, 1) or 𝜀 > 1 +
= = xA2 (1 − 𝜀)|M|−1 x𝜎𝜓 1+𝜀 𝜔𝜀−1
𝜆𝛿
(D.15)
d𝜙 |M| ⩽ 0, if 1 ⩽ 𝜀 < 1 + .
𝜆𝛿
Proof of Corollary 3 By differentiating the explicit solution for g∗ when 𝜀 = 1 we find that
13
dg∗ 𝛿
=− < 0,
d𝜆 𝜌x𝜎A2
dg∗ 𝜆
=− < 0,
d𝛿 𝜌x𝜎A2
( ) ( )
dg∗ 1 𝛿𝜆
= 𝜙+ > 0,
d𝜎 𝜎2A 𝜌xA
dg∗ 1
=− < 0.
d𝜙 𝜎A
ds
= −(1 − 𝜔1−𝜀 )x𝜎|M|−1 > 0, for 𝜀 > 1,
d𝜆
ds
= −𝜆𝜔1−𝜀 (1 − 𝜀)𝜓 −𝜀 |M|−1 > 0, for 𝜀 > 1,
d𝛿
ds [ ]
= (1 − 𝜀) xA𝜙 + 𝜆𝜔1−𝜀 𝜓 −𝜀 𝛿 (𝜎|M|)−1 < 0, for 𝜀 > 1,
d𝜎
ds
= −xA(1 − 𝜀)|M|−1 > 0 for 𝜀 > 1.
d𝜙
Proof of Corollary 4 In the case of logarithmic utility function, the saving propensity
reduces to 1 − A𝜌 and is therefore independent of the arrival rate, damage intensity, abate-
ment efficiency and polluting intensity.
Endogenous Arrival
Solution of the Extended Model
The HJB equation is given by

{
𝜌V(K, P) = max U(C, P) + Vk [(1 − 𝜃)Y − C] + Vp [(𝜙 − 𝜃𝜎)Y − 𝛼P]+
[ ]}
+𝜆(K, P) V(K,̃ P) − V(K, P) ,
̃ = 𝜔K .
where K
The optimality conditions with respect to the control and the state variables consist of:
C ∶ Uc − Vk = 0, (E.1)
𝜃 ∶ − Vk Y − Vp 𝜎Y = 0, (E.2)
K ∶ 𝜌Vk = Vkk [(1 − 𝜃)Y − C] + Vk (1 − 𝜃)A + Vp (𝜙 − 𝜎𝜃)A + Vpk [(𝜙 − 𝜎𝜃)Y − 𝛼P]+
[ ]
dV(𝜔K, P)
+𝜆 − Vk + 𝜆k [V(𝜔K, P) − V(K, P)],
dK
(E.3)
13
Fig. 4 Possible solutions for 𝜓̄
P ∶ 𝜌Vp = Up + Vpp [(𝜙 − 𝜎𝜃)Y − 𝛼P] + Vkp [(1 − 𝜃)Y − C] − 𝛼Vp +

[ ]
dV(𝜔K, P) (E.4)
+𝜆 − Vp + 𝜆p [V(𝜔K, P) − V(K, P)].
dP
Solution proceeds along similar lines as in Sect. 2. First, we guess that the value function
is of the same form as in Sect. 2. Then we derive the optimal policy from the first two
optimality conditions. Recall that we also assumed 𝜆 = 𝜆(𝜐𝜇 ), where 𝜐 = K 𝜂 P𝜉 . Substitut-
ing the latter, along with the policy, into HJB equation, we find the parameters of the value
function, 𝜓̄ t and x. Condition (E.4) is then used to find the optimal 𝜃̄ , while (E.3) allows us
to obtain ḡ , following exactly the same steps as in A.1.2 (Fig. 4) .
̄
Analysis of Multiple Solutions for Ã
Multiple solutions to Eq. (25), which we reproduce here for convenience

( )
𝜙 [ ]
𝜀𝜓 = 𝜌 − (1 − 𝜀)A 1 − − 𝜆(𝜓) 𝜔1−𝜀 (𝜓) − 1 (E.5)
𝜎
are feasible when the elasticity of marginal utility is above unity (𝜀 > 1). Define the func-
tion LHS(𝜓) = 𝜀𝜓 as the left-hand side of the equation and, similarly, the function RHS(𝜓)
as the right-hand side of the equation. The LHS is a straight line with the slope equal to 𝜀.
The RHS is, in general, a non-linear function of 𝜓 . Differentiating it with respect to 𝜓 , we
obtain
dRHS d𝜆 𝜕𝜐 [ 1−𝜀 ] 𝜕𝜔
=− 𝜔 (𝜓) − 1 − 𝜆(1 − 𝜀)𝜔−𝜀 ≷ 0 ⇔ 𝜀 ≷ 1,
d𝜓 d𝜐 𝜕𝜓 𝜕𝜓
so that RHS is strictly increasing in 𝜓 for 𝜀 > 1. The second derivative is of ambigu-
ous sign which means that the function has concave and convex segments and inflection
point(s). Examining its asymptotic behavior, we see that lim𝜓→0 RHS = ±∞ ⇔ 𝜀 ≶ 1 and
lim𝜓→1 RHS is constant. Possible solutions to (E.5) are illustrated in the Fig. 4.
13
Robustness Checks
Sensitivity to
See Tables 4 and 5.
Table 4 Optimal policy with 𝛾 = 0.9 𝜀=1 𝜀=3

𝛿 = 1e − 6 (up to 1% of GWP)
and 𝛾 = 0.9 𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.08 𝜎 = 0.05
𝜆 = 0.004
𝜃 0.50015 0.80024 0.62005 1.00844
g 3.47499 3.45998 1.15589 1.14903
𝜆 = 0.01
𝜃 0.50038 0.80060 0.80158 1.32557
g 3.47497 3.45995 1.15218 1.14247
𝜆 = 0.02
𝜃 0.50075 0.80121 1.10810 1.86662
g 3.47494 3.45990 1.14593 1.13128
Table 5 Optimal policy with 𝛾 = 1.1 𝜀=1 𝜀=3

𝛿 = 1e − 6 (up to 1% of GWP)
and 𝛾 = 1.1 𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.08 𝜎 = 0.05
𝜆 = 0.004
𝜃 0.50015 0.80024 0.51710 0.82648
g 3.47499 3.45999 1.15810 1.15298
𝜆 = 0.01
𝜃 0.50038 0.80060 0.54278 0.86626
g 3.47499 3.45998 1.15776 1.15244
𝜆 = 0.02
𝜃 0.50075 0.80121 0.58565 0.93276
g 3.47497 3.45996 1.15718 1.15155
Sensitivity to ı and ˇ.
Multiplicative Utility
We have also considered an alternative preference structure with multiplicative utility of

−𝜑 1−𝜀
the form (CP1−𝜀) , which has also been used in Nordhaus’s DICE. Because of the higher
substitution with the multiplicative function, our climate policy instrument 𝜃 is indeed
lower than with the additive specification (Table 6). It is actually lower to the extent that
the optimal abatement does not fully cover the current emissions—a point raised by the
Reviewer in the later comments. Table 7 below shows the solution for our abatement pro-
pensity and the corresponding optimal growth rate with a multiplicative utility in various
13
Table 6 Other parameters are at 𝛾=1 𝛿 = 1e − 7 𝛽=2

benchmark values and 𝜀 = 3
𝜎 = 0.08 𝜎 = 0.05 𝜎 = 0.08 𝜎 = 0.05
𝜆 = 0.004
𝜃∗ 0.504 0.807 0.548 0.878
g∗ 1.158 1.153 1.157 1.152
𝜆 = 0.01
𝜃∗ 0.510 0.816 0.621 0.996
g∗ 1.158 1.153 1.156 1.150
𝜆 = 0.02
𝜃∗ 0.520 0.833 0.742 1.194
g∗ 1.158 1.153 1.154 1.147
Table 7 Optimal policy in 𝛿 = 0.1 𝛿 = 0.3

low damage and high-damage
scenario. Other parameters are at
𝜎 = 0.08 𝜃∗ 0.48 0.09
benchmark values and 𝜀 = 3
(𝜙𝜎= 0.005) g∗ 2.71 2.84
𝜎 = 0.05 𝜃∗ 0.79 0.49
(𝜙𝜎 = 0.008) g∗ 2.70 2.80
𝜎 = 0.025 𝜃∗ 1.59 1.39
(𝜙𝜎 = 0.016) g∗ 2.66 2.73
scenarios. The parameter 𝜑 in the utility function is set to unity; changing the value of 𝜑
does not bring substantial changes to the results. The calibration of the other parameters
remains as in the paper, except for the damage intensity 𝛿 which now needed to be recali-
brated for the two scenarios, i.e. low damage (<1% of GDP) and high damage (up to 10%
of GDP). Note that in all cases 𝜃 < 𝜙∕𝜎.
Funding Open access funding provided by Swiss Federal Institute of Technology Zurich.
Declarations
Conflict of interest The authors declare that they have no conflict of interest. There are no financial or non-
financial interests that are directly or indirectly related to this work.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Com-
mons licence, and indicate if changes were made. The images or other third party material in this article
are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
13
References
Alary D, Gollier C, Treich N (2013) The effect of ambiguity aversion on insurance and self-protection. Econ
J 123:1188–1202. https://doi.org/10.1111/ecoj.12035
Athanassoglou S, Xepapadeas A (2012) Pollution control with uncertain stock dynamics: when, and how, to
be precautious. J Environ Econ Manag 63:304–320
Barrage L (2014) Sensitivity analysis for Golosov, Hassler, Krusell, and Tsyvinski, (2013): optimal taxes on
fossil fuel in general equilibrium. Supp Mater Econom 82(1):41–88
Barro RJ (2015) Environmental protection, rare disasters and discount rates. Economica 82:1–23
Bommier A, Kochov A, Le Grand F (2017) On monotone recursive preferences. Econometrica
85(5):1433–1466
Bouwer LM (2011) Have disaster losses increased due to anthropogenic climate change? Bull Am Meteorol
Soc 92(1):39–46
Bretschger L, Vinogradova A (2019) Best policy response to environmental shocks: applying a stochastic
framework. J Environ Econ Manag 97:23–41
Brock W, Xepapadeas A (2020) Regional climate policy under deep uncertainty: robust control and distribu-
tional concerns. Environ Dev Econ 26(3):211–238. https://doi.org/10.1017/S1355770X20000248
Brown P, Daigneault AJ, Tjernström E, Zou W (2018) Natural disasters, social protection, and risk percep-
tions. World Dev 104:310–25
Bühren C, Meier F, Plessner M (2021) Ambiguity aversion: bibliometric analysis and literature review of
the last 60 years. Manag Rev Q 73:495–525
Cai Y, Lontzek T (2019) The social cost of carbon with economic and climate risks. J Political Econ. https://
doi.org/10.1086/701890
Cameron L, Shah M (2015) Risk-taking behavior in the wake of natural disasters. J Hum Resour 50:484–515
Cavallo E, Galiani S, Noy I, Pantano J (2013) Catastrophic natural disasters and economic growth. Rev
Econ Stat 95(5):1549–1561
Clarke HR, Reed WJ (1994) Consumption/pollution tradeoffs in an environment vulnerable to pollution-
related catastrophic collapse. J Econ Dyn Control 18:991–1010
de Zeeuw A, Zemel A (2012) Regime shifts and uncertainty in pollution control. J Econ Dyn Control
36:939–950
Dietz S, Stern N (2015) Endogenous growth, convexity of damage and climate risk: how Nordhaus’ frame-
work supports deep cuts in carbon emissions. Econ J 125:574–620
Dietz S, Venmans F (2019) Cumulative carbon emissions and economic policy: in search of general princi-
ples. J Environ Econ Manag 96:108–129
Dohmen T, Falk A, Huffman D, Sunde U, Schupp J, Wagner GG (2011) Individual risk attitudes: measure-
ment, determinants, and behavioral consequences. J Eur Econ Assoc 9(3):522–50
Douenne T (2020) Disaster risks, disaster strikes, and economic growth: the role of preferences. Rev Econ
Dyn 38:251–272
Drichoutis AC, Koundouri P (2012) Estimating risk attitudes in conventional and artefactual lab experi-
ments: the importance of the underlying assumptions. Economics 6:2012–38. https://doi.org/10.5018/
economics-ejournal.ja.2012-36
Englezos N, Kartala X, Koundouri P et al (2022) A novel hydroeconomic–econometric approach for inte-
grated transboundary water management under uncertainty. Environ Resour Econ 84:975–1030.
https://doi.org/10.1007/s10640-022-00744-4
Epstein LG, Zin SE (1989) Substitution, risk aversion, and the temporal behavior of consumption and asset
returns: a theoretical framework. Econometrica 57(4):937–969
Farmer JD, Hepburn C, Mealy P, Teytelboym A (2015) A third wave in the economics of climate change.
Environ Resour Econ 62:329–357
Gerlagh R, Liski M (2017) Carbon prices for the next hundred years. Econ J. https://doi.org/10.1111/ecoj.
12436
Gollier C, Pratt JW (1996) Risk vulnerability and the tempering effect of background risk. Econometrica
64(5):1109–1123
Golosov M, Hassler J, Krusell P, Tsyvinski A (2014) Optimal taxes on fossil fuels in general equilibrium.
Econometrica 82(1):41–88
Guiso L, Paiella M (2008) Risk aversion, wealth, and background risk. J Eur Econ Assoc 6(6):1109–1150
Harrison GW, List JA, Towe C (2007) Naturally occurring preferences and exogenous laboratory experi-
ments: a case study of risk aversion. Econometrica 75(2):433–458
Harstad B (2012) Buy coal! A case for supply-side environmental policy. J Political Econ 120(1):77–115
13
Hepburn C, Koundouri P, Panopoulou K, Pantelides T (2009) Social discounting under uncertainty: a cross
country comparison. J Environ Econ Manag 57(2):140–150. https://doi.org/10.1016/j.jeem.2008.04.
004
Hood C (2011) Summing up the parts. Combining policy instruments for least-cost climate mitigation strat-
egies. International Energy Agency, Information Paper, Paris
Ikefuji M, Horii R (2012) Natural disasters in a two-sector model of endogenous growth. J Public Econ
96:784–796
IPCC (2014) Fifth Assessment Report, contribution of working group II, climate change 2014: impacts,
adaptation, and vulnerability. http://www.ipcc.ch/report/ar5/wg2/
Koundouri P, Laukkanen M, Myyrä S, Nauges C (2009) The effects of EU agricultural policy changes on
farmers’ risk attitudes. Eur Rev Agric Econ 36(1):53–77
Lanz B, Dietz S, Swanson T (2017) Global population growth, technology and Malthusian constraints: a
quantitative growth theoretic perspective. Int Econ Rev 58:973–1006
Le Van C, Schubert K, Nguyen TA (2010) With exhaustible resources, can a developing country escape
from the poverty trap? J Econ Theory 145:2435–2447
Lemoine D, Rudik Y (2017) Steering the climate system: using inertia to lower the cost of policy. Am Econ
Rev 107(10):2947–2957
Lemoine D, Traeger C (2014) Watch your step: optimal policy in a tipping climate. Am Econ J Econ Pol
6(1):137–166
Lenton TM, Ciscar J-C (2013) Integrating tipping points into climate impact assessments. Clim Change
117:585–597
Liebenehm S, Schumacher I, Strobl E (2023) Rainfall shocks and risk aversion: evidence from Southeast
Asia. Am J Agric Econ. https://doi.org/10.1111/ajae.12403
Lontzek TS, Cai Y, Judd KL, Lenton TM (2015) Stochastic integrated assessment of climate tipping points
indicates the need for strict climate policy. Nat Clim Chang 5:441–444
Mattauch L, Matthews HD, Millar R, Rezai A, Solomon S, Venmans F (2020) Steering the climate system:
using inertia to lower the cost of policy: comment. Am Econ Rev 110(4):1231–1237
McGrattan E (1998) A defense of AK growth models. Fed Reserve Bank Minneap Q Rev 22(4):13–27
McKinsey (2009) Pathways to a low carbon economy, version 2 of the global greenhouse gas abatement cost
curve. Mc Kinsey and Company
Michel P, Rotillon G (1995) Disutility of pollution and endogenous growth. Environ Resour Econ 6:279–300
Mittelstaedt C, Baumgärtner S (2023) Preference functions for knightian uncertainty, SSRN Working Paper,
https://ssrn.com/abstract=4306869 or http://dx.doi.org/10.2139/ssrn.4306869
Mueller-Fuerstenberger G, Schumacher I (2015) Insurance and climate-driven extreme events. J Econ Dyn
Control 54:59–73
Nordhaus WD (2008) A question of balance: weighing the options on global warming policies. Yale Univer-
sity Press, New Haven
Nordhaus WD, Boyer J (2000) Warming the world. MIT Press, Cambridge
Peretto P, Valente S (2015) Growth on a finite planet: resources, technology and population in the long run.
J Econ Growth 20(3):305–331
Sennewald K, Waelde K (2006) "Ito’s Lemma’’ and the Bellman equation for Poisson process: an applied
view. J Econ 89:1–36
Soretz S (2007) Efficient dynamic pollution taxation in an uncertain environment. Environ Resour Econ
36:57–84
Steger TM (2005) Stochastic growth under wiener and poisson uncertainty. Econ Lett 86:311–316
Stern N (2007) The economics of climate change: the stern review. Cambridge University Press, Cam-
bridge, UK
Stern N (2015) Why are we waiting? The logic, urgency, and promise of tackling climate change. MIT
Press, Cambridge
Stern N (2016) Current climate models are grossly misleading. Nature 530:407–409
Swiss Re (2015) Financial Report 2015, Zuerich
Tsur Y, Zemel A (1996) Accounting for global warming risks: resource management under event uncer-
tainty. J Econ Dyn Control 20:289–1305
Tsur Y, Zemel A (1998) Pollution control in an uncertain environment. J Econ Dyn Control 22:967–975
Van den Bijgaart I, Gerlagh R, Liski M (2016) Accounting for global warming risks: resource management
under event uncertainty. J Environ Econ Manag 77:75–94
Van Den Bremer T, van der Ploeg F (2018) Pricing carbon under economic and climatic risks: leading-order
results from asymptotic analysis. CEPR Discussion Paper No. DP12642. Available at SSRN: https://
ssrn.com/abstract=3112266
Van der Ploeg F (2014) Abrupt positive feedback and the social cost of carbon. Eur Econ Rev 67:28–41
13
Van der Ploeg F, de Zeeuw A (2018) Climate tipping and economic growth: precautionary capital and the
price of carbon. J Eur Econ Assoc 16(5):1577–1617
Waelde K (1999) Optimal saving under poisson uncertainty. J Econ Theory 87:194–217
World Bank (2016) World Development Indicators. http://data.worldbank.org/indicator/ EN.ATM.CO2E.
KD.GD
Zemel A (2015) Adaptation, mitigation and risk: an analytic approach. J Econ Dyn Control 51:133–147
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
13

Escaping Damocles' Sword: Endogenous Climate Shocks in A Growing Economy

Uploaded by

Copyright:

Available Formats

Escaping Damocles' Sword: Endogenous Climate Shocks in A Growing Economy

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Escaping Damocles' Sword: Endogenous Climate Shocks in A Growing Economy

Uploaded by

Copyright:

Available Formats

Environmental and Resource Economics

Escaping Damocles’ Sword: Endogenous Climate Shocks

Alexandra Brausmann1 · Lucas Bretschger2

Accepted: 19 December 2023

Keywords Climate policy · Uncertainty · Natural disasters · Endogenous growth

JEL Classification O10 · Q52 · Q54

1.1 Economics and the Climate

There is a growing literature on random catastrophic events causing irreversible damage.4

1.3 Contribution to the Literature

dPt = (Et − 𝛼Pt )dt, P0 given, (3)

Et = 𝜙Yt (Kt ) − Z(It ), (4)

and the associated Lagrangian by L = H + 𝜐t (𝜃̄ − 𝜃t ) + 𝜈t 𝜃t , where we explicitly attached

2.2.1 The Interior Solution

(aK + bP)1−𝜀 P1+𝛽

2.3 Optimal Trajectory with Different Initial Conditions

3 Model with Climate Shocks

Fig. 1 Optimal paths for different initial conditions

dPt = (Et − 𝛼Pt )dt, P0 given, (15)

Et = 𝜙Yt (Kt ) − Z(It ), (16)

where 𝜔(𝜓, x) = e−𝛿[𝜎x𝜓 and

Given that K̃ = 𝜔K , while C = 𝜓 ∗ K and C̃ = 𝜓 ∗ K̃ , 𝜔 also represents the ratio of post-

We are now in a position to state the following

(i) optimal consumption rate is a constant fraction of the capital stock;

Proof provided in the Appendix.  ◻

Corollary 1 In the case of logarithmic utility long-run consumption-to-capital ratio, 𝜓 ∗, is

Fig. 2 Optimal path (solid) versus expected path (dashed)

4 Effects of Economic and Climatic Fundamentals

4.1 The Abatement Propensity

Proposition 2 The long-run abatement propensity is:

(i) an increasing function of the arrival rate,

Proof provided in the Appendix.  ◻

Proof provided in the Appendix.  ◻

4.2 The Growth Rate

Proposition 3 The optimal trend consumption growth rate is:

(i) a decreasing function of the arrival rate,

Proof provided in the Appendix.  ◻

Proof provided in the Appendix.  ◻

Proposition 4 The long-run propensity to save is:

(i) an increasing function of the arrival rate,

Proof provided in the Appendix.  ◻

Proof provided in the Appendix.

damages in our baseline model, we adopt a Cobb-Douglas combination of K and P such

(i) the optimal consumption-to-capital ratio is smaller,

Table 1 Benchmark values of 𝜌 𝜒 𝛽 𝛼 𝜀 A 𝜙 𝜎 𝜆 𝛿 𝛾

Solution of the Deterministic Model

The Interior Solution

Combined with the boundary conditions they can be written as

𝜇̇ − (𝛼 + 𝜌)𝜇 = 𝜒P𝛽 , (A.6)

𝜆̇ + 𝜆[(1 − 𝜃)A − 𝜌] + 𝜇(𝜙 − 𝜃𝜎)A = 0, (A.7)

Ṗ = −𝛼P + (𝜙 − 𝜎𝜃)Y, (A.8)

which can be rewritten as

Since Ct = 𝜆t by (A.1), Eq. (A.9) can be rewritten as the following ODE:

resulting in abatement propensity

(aK + bP)1−𝜀 P1+𝛽

Ṗ = −𝛼P + 𝜙AK , 𝜇≤0 (A.16)

A Note on Asymptotic Value Function

(aK + bP)1−𝜀 P1+𝛽

where 𝜓 ∗ stands for a1−𝜀 .

Interior Solution to the Baseline Stochastic Model

1.1 Economics and the Climate

1.3 Contribution to the Literature

2.2.1 The Interior Solution

2.3 Optimal Trajectory with Different Initial Conditions

3 Model with Climate Shocks

Proof provided in the Appendix. ◻

4 Effects of Economic and Climatic Fundamentals

4.1 The Abatement Propensity

Proof provided in the Appendix. ◻

Proof provided in the Appendix. ◻

4.2 The Growth Rate

Proof provided in the Appendix. ◻

Proof provided in the Appendix. ◻

Proof provided in the Appendix. ◻