Energy Economics 104 (2021) 105624

Contents lists available at ScienceDirect

Energy Economics
journal homepage: www.elsevier.com/locate/eneeco

Long-term macroeconomic effects of climate change: A cross-country

analysis✩
Matthew E. Kahn a ,1 , Kamiar Mohaddes b ,∗,1 , Ryan N.C. Ng c ,1 , M. Hashem Pesaran a,d ,1 ,
Mehdi Raissi e ,1 , Jui-Chung Yang f ,1
a
Department of Economics, University of Southern California, USA
b Judge Business School & King’s College, University of Cambridge, UK
c Faculty of Economics, University of Cambridge, UK
d Trinity College, Cambridge, UK
e
International Monetary Fund, Washington DC, USA
f
Department of Economics, National Taiwan University, Taiwan

ARTICLE INFO ABSTRACT

Dataset link: http://dx.doi.org/10.17632/hytzz We study the long-term impact of climate change on economic activity across countries, using a stochastic
8wftw growth model where productivity is affected by deviations of temperature and precipitation from their long-
JEL classification: term moving average historical norms. Using a panel data set of 174 countries over the years 1960 to 2014,
C33 we find that per-capita real output growth is adversely affected by persistent changes in the temperature
O40 above or below its historical norm, but we do not obtain any statistically significant effects for changes in
O44 precipitation. We also show that the marginal effects of temperature shocks vary across climates and income
O51 groups. Our counterfactual analysis suggests that a persistent increase in average global temperature by 0.04 ◦ C
Q51 per year, in the absence of mitigation policies, reduces world real GDP per capita by more than 7 percent by
Q54 2100. On the other hand, abiding by the Paris Agreement goals, thereby limiting the temperature increase to
Keywords: 0.01 ◦ C per annum, reduces the loss substantially to about 1 percent. These effects vary significantly across
Climate change countries depending on the pace of temperature increases and variability of climate conditions. The estimated
Economic growth losses would increase to 13 percent globally if country-specific variability of climate conditions were to rise
Adaptation commensurate with annual temperature increases of 0.04 ◦ C.
Counterfactual analysis

1. Introduction the distribution of weather patterns (i.e., climate change2 ) are not
only affecting low-income countries and emerging markets, but also
Global temperatures have increased significantly in the past half advanced economies. A persistent rise in temperatures, changes in pre-
century possibly causing a wide range of impacts, including cold snaps cipitation patterns and/or more volatile weather events can have long-
and heat waves, droughts and floods, hurricanes, higher sea levels, term macroeconomic effects by adversely affecting labour productivity,
and weather whiplash; see IPCC (2021) for details. These changes in slowing investment and damaging human health.

✩ We are grateful to Tiago Cavalcanti, Francis X. Diebold, Christopher Hajzler, Stephane Hallegatte, Zeina Hasna, John Hassler, Per Krusell, Miguel Molico, Peter
Phillips, Margit Reischer, Ron Smith, Richard Tol, Carolyn A. Wilkins and seminar participants at the International Monetary Fund (IMF), Bank of Lithuania, Bank
of Canada, EPRG, Cambridge Judge Business School, the ERF 24th Annual Conference, the 2018 MIT CEEPR Research Workshop, the 2019 Keynes Fund Research
Day, National Institute of Economic and Social Research, Copenhagen Business School, Bank of England, Federal Reserve Bank of San Francisco, London School
of Economics, European Central Bank, and RES 2021 Annual Conference for comments and suggestions. We would also like to thank the editor in charge of our
paper and five anonymous referees for helpful suggestions and Matthew Norris for help with constructing the global climate dataset. We gratefully acknowledge
financial support from the Keynes Fund. The views expressed in this paper are those of the authors and do not necessarily represent those of the IMF or its policy.
∗ Corresponding author.
E-mail address: km418@cam.ac.uk (K. Mohaddes).
1
All co-authors of this manuscript have contributed equally to every aspect of the work.
2
Weather refers to atmospheric conditions over short periods of time (e.g., temperature and precipitation). Climate refers to the long-term average and
variability of weather. Climate change is a shift "in the state of the climate that can be identified (e.g., via statistical tests) by changes in the mean and/or the
variability of its properties, and that persists for an extended period, typically decades or longer” (IPCC, 2014).

https://doi.org/10.1016/j.eneco.2021.105624
Received 10 January 2021; Received in revised form 20 September 2021; Accepted 2 October 2021
Available online 15 October 2021
0140-9883/© 2021 Elsevier B.V. All rights reserved.
M.E. Kahn et al. Energy Economics 104 (2021) 105624

This paper investigates the long-term macroeconomic effects of average to estimate unbiased weather effects in panel data studies.
weather patterns transformed by climate change across 174 countries As well, this transformation would allow for an implicit model of
over the period 1960 to 2014. While weather could affect the level of adaptation. Also, current panel models do not explicitly model climate
output across climates, for example, by changing agricultural yields, variability in the estimation of long-term damage functions.
climate change, by shifting the long-term average and variability of Thirdly, the fixed effects (FE) estimators used in panel-data studies
weather, could impact an economy’s ability to grow in the long-term, assume that climate variables are strictly exogenous. At the heart of
through reduced investment and lower labour productivity. We focus the Dynamic Integrated Climate-Economy (DICE) model of Nordhaus is
on both of these issues and develop a theoretical growth model that the need to account for bi-directional feedback effects between growth
links deviations of temperature and precipitation (weather) from their and climate change (see Nordhaus, 1992). In his work, Nordhaus
long-term moving-average historical norms (climate) to per capita real accounts for the fact that faster economic activity increases the stock of
output growth (Appendix A.1). greenhouse gas (GHG) emissions and thereby the average temperature
In our empirical application, we allow for dynamics and feedback ef- (possibly with a long lag). At the same time, rising average temperature
fects in the interconnections of climatic and macroeconomic variables, could reduce real economic activity. Consequently, when estimating the
distinguish between level and growth effects – including for long- impact of temperature on economic growth, 𝑇𝑖𝑡 may not be considered
term –, consider asymmetric weather effects, and test for differential as strictly exogenous, but merely weakly exogenous/predetermined to
impact of weather shocks across climates. Also, by using deviations of income growth; in other words economic growth in the past might
temperature and precipitation from their respective historical norms, have feedback effects on future temperature. While it is well known
while allowing for nonlinearity3 and an implicit model for adaptation, that the FE estimator suffers from small-𝑇 bias in dynamic panels (see
we avoid the econometric pitfalls associated with the use of trended Nickell, 1981) with 𝑁 (the cross-section dimension) larger than 𝑇 (the
variables, such as temperature, in output growth equations. As it is well time series dimension), Chudik et al. (2018) show that this bias exists
known, and is also documented in our paper, temperature has been regardless of whether the lags of the dependent variable are included
trending upward strongly in almost all countries in the world, and its or not, so long as one or more regressors are not strictly exogenous. In
use as a regressor in growth regressions can lead to spurious results. such cases, inference based on the standard FE estimator will be invalid
A detailed analysis of how trends in temperature can lead to spurious and can result in large size distortions unless 𝑁∕𝑇 → 0, as 𝑁, 𝑇 → ∞
trends in output growth in regressions used in the literature is provided jointly.
in Appendix A.2. We contribute to the literature along the following dimensions.
The literature which attempts to quantify the effects of weather Firstly, we explicitly model and test for level or growth effects of
and/or climate on economic performance (agricultural production, weather shocks and estimate the long-term macroeconomic impact of
labour productivity, commodity prices, health, conflict, and economic persistent increases in temperature. Secondly, we use the half-panel
growth) is growing fast—see Stern (2007), IPCC (2014), Hsiang (2016), Jackknife FE (HPJ-FE) estimator proposed in Chudik et al. (2018)
Cashin et al. (2017), Letta and Tol (2019), Henseler and Schumacher to deal with the possible bias and size distortion of the commonly-
(2019), and recent surveys by Tol (2009), Dell et al. (2014), and Tol used FE estimator (given that 𝑇𝑖𝑡 is weakly exogenous). When the
(2018). There are a number of grounds on which the econometric time dimension of the panel is moderate relative to 𝑁, the HPJ-FE
evidence of climate impacts on the economy may be questioned. Firstly, estimator effectively corrects the Nickel-type bias if regressors are
the literature that relies on the cross-sectional approach (e.g., Sachs and weakly exogenous, and is robust to possible feedback effects from
Warner, 1997, Gallup et al., 1999, Nordhaus, 2006, and Kalkuhl and aggregate economic activity to the climate variables. Thirdly, we test
Wenz, 2020) is hindered by the temporal invariance of climate over the predictions of our theoretical growth model using cross-country
the studied time-frames and by important omitted variables that affect data on per-capita GDP growth and deviations of temperature and
economic performance (e.g., institutions). The more recent literature precipitation from their moving average historical norms over the past
largely uses panel data models to estimate the economic effects of fifty-five years (1960–2014). Our focus on ‘‘deviations’’ is a departure
weather shocks. See, for example, Burke et al. (2015), Dell et al. (2009, from the literature, as changes in the distribution of weather patterns
2012, 2014), and Hsiang (2016). There is, however, some disagreement (not only averages of temperature and precipitation but also their
in the literature as to whether temperature affects the level of economic variability) are modelled explicitly; an implicit model of adaptation is
output or its growth. See Schlenker and Auffhammer (2018) and Newell introduced; and the econometric pitfalls of including trended variables
et al. (2021) for a discussion. (that is, 𝑇𝑖𝑡 ) in growth regressions are avoided (see Appendix A.2 for
Secondly, econometric specifications of the weather– details). Moreover, rather than assuming a common climate threshold
macroeconomic relation are often written in terms of GDP per capita across countries, we allow for country-specific and time-varying climate
growth and the level of temperature, 𝑇𝑖𝑡 , and in some cases also 𝑇𝑖𝑡2 ;4 thresholds and also test for asymmetric effects.5 Finally, we estimate
see, for instance, Dell et al. (2012), Burke et al. (2015) , and Kalkuhl the differential impact of weather shocks across climates (e.g., hot and
and Wenz (2020). But if 𝑇𝑖𝑡 is trended, which is the case in almost cold) and income groups (rich and poor) using a heterogeneous panel
all countries in the world (see Appendix A.3), its inclusion in the data model.
regression will introduce a linear trend in per capita output growth Our results suggest that a series of positive (or negative) weather
which is spurious and is not supported by the data (see Table A.1), and shocks has a long-term negative effect on per capita GDP growth.
can in turn lead to biased estimates. The prevalence of this issue in Since we are measuring an integral of marginal weather effects in our
the econometric specifications used in the literature is demonstrated
in Appendix A.2. Indeed, Mendelsohn (2016) and Tol (2021) argue
that researchers should focus on the deviation of 𝑇𝑖𝑡 from its long-term 5
Assuming common climate thresholds, as is done in the literature, leads to
important oddities in individual country estimates. For example, Burke et al.
(2015) estimate that per capita GDP will be 63, 210, 247, 419, 516, 1413
3 percent larger in Germany, Sweden, Canada, Russia, Finland, and Mongolia
Non-linearity arises because growth is only affected when temperature
(or precipitation) goes above or below a time-varying and country-specific as a result of climate change by 2100. Similarly, it is estimated that many
historical threshold (i.e., the norm). It is due to this feature that future growth countries (including Brazil, India, and most African and South East Asian
is affected not only by warming (or cooling if that was the case) but also by countries) will experience per capita GDP losses of more than 80 percent
its variability. which is hard to imagine barring climate disasters (which cannot be modelled
4
It is argued that this quadratic specification would account for the within a stochastic growth framework as we document in Appendix A.1).
global nonlinear relationship between temperature and growth; i.e., a common See https://web.stanford.edu/\char126\relaxmburke/climate/map.php for the
temperature threshold. mentioned individual country results.

2
M.E. Kahn et al. Energy Economics 104 (2021) 105624

regressions, we can cautiously link them to climate change. Specifically,

we show that if temperature rises (falls) above (below) its historical
norm by 0.01 ◦ C annually for a long period of time, income growth
will be lower by 0.0543 percentage points per year. We could not detect
any significant evidence of an asymmetric long-term growth impact
from persistent positive and negative deviations of temperature from
its norms. Furthermore, we show that our empirical findings pertain
to poor or rich, and hot or cold countries alike (albeit to varying
degrees) as economic growth is affected not only by persistent increases
in temperatures (and the pace with which they are rising) but also
by the degree of climate variability.6 One of the reasons that cold
countries are also affected by climate change is the faster pace with
which temperatures are rising in these regions than in hot countries.
Suppose that the pace of temperature increases was the same across
hot and cold climates, then our heterogeneous panel estimations would
suggest a smaller, but still negative, marginal weather effect in cold
countries. Most papers in the literature find that temperature increases Fig. 1. Global Temperature Projections (Deviations from 1984–2014). Notes: The thin
lines represent each of the 40 models in the IPCC WG1 AR5 Annex I Atlas. The thick
have had uneven macroeconomic effects, with adverse consequences
lines represent the multimodel mean. Representative Concentration Pathways (RCP)
only in countries with hot climates or low-income countries; see, for are scenarios of greenhouse gas concentrations, constructed by the IPCC. RCP 2.6
instance, Sachs and Warner (1997), Jones and Olken (2010), Dell et al. corresponds to the Paris Agreement which aims to hold the increase in the global
(2012), International Monetary Fund (2017), and Mejia et al. (2018). average temperature to below 2 degrees Celsius above pre-industrial levels. RCP 8.5
is an unmitigated scenario in which emissions continue to rise throughout the 21st
We estimate that the marginal effects of weather shocks are larger in
century.
low-income countries because they have lower capacity to deal with Source: Intergovernmental Panel on Climate Change (IPCC) Coupled Model
the consequences of climate change. However, this does not mean that Intercomparison Project Phase Five AR5 Atlas Subset.
rich nations are immune from the effects of climate change.
To contribute to climate change policy discussions, we perform a
number of counterfactual exercises where we investigate the cumu- its average temperatures), but would be limited to 1.88 percent under
lative income effects of annual increases in temperatures over the the Paris Agreement objective. Moreover, the speed with which the
period 2015–2100 (when compared to a baseline scenario under which historical norms change (20-, 30-, or 40 year moving averages)– that is
temperature in each country increases according to its historical trend how fast countries adapt to global warming or new climate conditions
of 1960–2014). We show that an increase in average global temperature – affects the size of income losses.8 Overall, while adaptation to climate
of 0.04 ◦ C per year – corresponding to the Representative Concentration change can reduce these negative long-run growth effects, it is highly
Pathway (RCP) 8.5 scenario (see Fig. 1), which assumes higher green- unlikely to offset them entirely.
house gas emissions in the absence of mitigation policies – reduces The rest of the paper is organized as follows. Section 2 discusses
world’s real GDP per capita by 7.22 percent by 2100. The estimated the long-run macroeconomic effects of weather patterns transformed by
losses under the RCP 8.5 scenario would almost double (to 13.11 climate change. Counterfactuals in Section 3 investigate the cumulative
percent globally by 2100) if country-specific variability of climate income effects of annual increases in temperatures under an unmiti-
conditions were to rise commensurate to temperature increases (see gated path as well as the Paris Agreement objective up to the year 2100.
Fig. 2 and Table 7). Limiting the increase to 0.01 ◦ C per annum, which Section 4 concludes. The paper also contains four appendices. Appendix
corresponds to the December 2015 Paris Agreement objective, reduces A.1 develops a multi-country stochastic growth model with weather
the output loss substantially to 1.07 percent.7 and climate effects. Appendix A.2 discusses a number of key growth
To put our results into perspective, Fig. 2 compares our economic regressions used in macroeconomy-climate research, and how they
loss estimates with those from select papers in the literature. Our coun- relate to our approach. Appendix A.3 provides detailed evidence on
terfactual estimates are relatively large. They suggest that all countries the historical patterns of climate change across 174 countries. Finally,
would experience a fall in GDP per capita by 2100 in the absence of Appendix A.4 provides additional empirical results.
climate change policies (i.e., under a high-emission scenario or RCP
8.5). However, the size of these income effects varies across countries
2. Empirical results
and regions depending on the pace with which temperatures increase
over the century and the historical variability of climate conditions in
In the empirical application, we use annual population-weighted
each country and their evolution going forward (see Figs. 3, 6 and 7);
climate data and real GDP per capita. For the climate variables we con-
for instance, for the U.S. the losses are relatively large at 10.52 percent
sider temperature (measured in degrees Celsius, ◦ C) and precipitation
under the RCP 8.5 scenario in year 2100 (reflecting a sharp increase in
(measured in metres). We construct population-weighted climate data
for each country and year between 1900 and 2014 using the terrestrial
6 air temperature and precipitation observations from Matsuura and Will-
For example, while the level of temperature in Canada is low, the
country is warming up twice as fast as the rest of the world and therefore mott (2015) (containing 0.5 degree gridded monthly time series), and
is being affected by climate change (including from damages to its physical the gridded population of the world collection from CIESIN (2016), for
infrastructure, coastal and northern communities, human health and wellness, which we use the population density in 2010.9 We obtain the real GDP
ecosystems and fisheries).
7
The Paris Agreement, reached within the United Nations Framework
Convention on Climate Change (UNFCCC), aims to keep the increase in the 8
Another way to assess adaptation is to test how the elasticity of per capita
global average temperature to below 2 degrees Celsius above pre-industrial GDP to climate variables evolves over time.
levels over the 21st century. The average global temperature is already 1 ◦ C 9
We follow the literature in applying constant weights to create population-
above the pre-industrial levels. For most countries, the Nationally Determined weighted temperatures, therefore abstracting from changes in the distribution
Contributions pledged under the Paris Agreement are deemed insufficient to of population over time and/or population trends (see Tol, 2017 for details).
meet either the 1.5 ◦ C or the 2 ◦ C target, and, judging by current policies, Nonetheless, in our empirical analysis we use deviations of temperature and
unlikely to be met in the first place. precipitation from their historical norms, which are not trended.

3
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Fig. 2. GDP Impact of Increases in Temperature. Notes: Projected GDP impact is for some future year, typically 2100. The shaded area represents the GDP per capita losses from
our counterfactual exercise in Section 3 with the upper bound based on 𝑚 = 20 and the lower bound based on 𝑚 = 40 (with increased climate variability). See Tables 6 and 7 for
details.
Source: Tol (2009, 2014), Burke et al. (2015), International Monetary Fund (2017) and authors’ estimates (shown as the grey area in the chart).

( )[ ] ( )
2 2
𝑃̃𝑖𝑡 (𝑚)− ]′ , 𝑇̃𝑖𝑡 (𝑚) = ∗ (𝑚)
𝑇𝑖𝑡 − 𝑇𝑖,𝑡−1 and 𝑃̃𝑖𝑡 (𝑚) =
[ ] 𝑚+1 𝑚+1
∗ (𝑚) are measures of temperature and precipitation relative
𝑃𝑖𝑡 − 𝑃𝑖,𝑡−1
to their historical norms per annum, 𝑇𝑖𝑡 and 𝑃𝑖𝑡 are the population-
weighted average temperature and precipitation of country 𝑖 in year
∗ (𝑚) = 1 ∑𝑚 𝑇
𝑡, and 𝑇𝑖,𝑡−1 ∗ 1 ∑𝑚
𝑚 𝓁=1 𝑖,𝑡−𝓁 and 𝑃𝑖,𝑡−1 (𝑚) = 𝑚 𝓁=1 𝑃𝑖,𝑡−𝓁 are the
time-varying historical norms of temperature and precipitation over
the preceding 𝑚 years in each 𝑡. Climate norms are typically computed
using 30 year moving averages (see, for instance, Arguez et al., 2012
and Vose et al., 2014), but to check the robustness of our results, we
also consider historical norms computed using moving averages with
𝑚 = 20 and 40.10 With 𝑇̃𝑖𝑡 (𝑚) and 𝑃̃𝑖𝑡 (𝑚) separated into positive and
negative values, we account for the potential asymmetrical effects of
climate change on growth around the threshold. The (average) long-
run effects, 𝜃, are calculated from the OLS estimates of the short-run
∑ ∑
coefficients in Eq. (1): 𝜽 = 𝜙−1 𝑝𝓁=0 𝜷 𝓁 , where 𝜙 = 1 − 𝑝𝓁=1 𝜑𝓁 .
The reasons for using ARDL growth regressions in deviations form
(i.e., temperature and precipitation relative to their long-term moving
Fig. 3. GDP Per Capita Losses from Increases in Temperature: Cold vs. Hot. Notes: average historical norms), rather than in levels and/or squares of
GDP per capita losses by 2100 from our baseline counterfactual exercise in Section 4 climate variables, are discussed in some detail in Appendix A.2, where
for hot (on left axis and in red) and cold (on right axis and in blue) countries. (For
interpretation of the references to colour in this figure legend, the reader is referred
it is shown that including 𝑇𝑖𝑡 and 𝑇𝑖𝑡2 will introduce trends in 𝛥𝑦𝑖𝑡 , which
to the web version of this article.) is not present in the data. As documented in Table A.1, we find that at
the 5% significance level, output growth is upward trended in only 21
countries out of 174 under consideration, and in fact 9 (174 × 0.05) of
per capita data between 1960 and 2014 from the World Development the 21 countries with statistically significant trend coefficients could
Indicators database of the World Bank. Combining the GDP per capita have arisen by pure chance given the large number of multiple tests
and the climate data, we end up with an unbalanced panel, which being carried out.
is very rich both in terms of the time dimension (𝑇 ), with maximum Other important econometric considerations behind the use of ARDL
𝑇 = 55 and average 𝑇 ≈ 39, and the cross-sectional dimension (𝑁), regressions are set out in Pesaran and Smith (1995), Pesaran (1997),
containing 174 countries. and Pesaran and Shin (1999) who show that the traditional ARDL
approach can be used for long-run analysis; it is valid regardless of
2.1. Long-term impact of climate change on economic growth whether the underlying variables are 𝐼 (0) or 𝐼 (1); and it is robust
to omitted variables bias and bi-directional feedback effects between
Considering strong evidence of an upward trend in temperatures economic growth and its determinants. These features of the panel
worldwide (see Appendix A.3), and guided by the theoretical growth ARDL approach are clearly appealing in our empirical application. For
model with weather and climate variables in Appendix A.1, we base validity of this technique, however, the dynamic specification of the
our empirical analysis on the following panel ARDL model: model needs to be augmented with a sufficient number of lagged effects
∑
𝑝 ∑
𝑝
′
so that regressors become weakly exogenous. Specifically, Chudik et al.
𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝜑𝓁 𝛥𝑦𝑖,𝑡−𝓁 + 𝜷 𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 (𝑚) + 𝜀𝑖𝑡 , (1)
𝓁=1 𝓁=0
where 𝑦𝑖𝑡 is the log of real GDP per capita of country 𝑖 in year 𝑡, 𝑎𝑖 10
𝑚 = 30 also corresponds to the official World Meteorological Organization
is the country-specific fixed effect, 𝐱̃ 𝑖𝑡 (𝑚) = [𝑇̃𝑖𝑡 (𝑚)+ , 𝑇̃𝑖𝑡 (𝑚)− , 𝑃̃𝑖𝑡 (𝑚)+ , definition of climate (i.e., norm).

4
M.E. Kahn et al. Energy Economics 104 (2021) 105624
( )
2
(2016), show that sufficiently long lags are necessary for the consis- level)—calculated as −0.806 × 𝑚+1 . To make sure that our results
tency of the panel ARDL approach.11 Since we are interested in studying are robust to the choice of historical norms, Table 2 also reports the
the growth effects of climate change (a long-term phenomenon), the estimation results with climate norms constructed as moving averages
lag order should be long enough, and as such we set 𝑝 = 4 for all the of the past 20 (𝑚 = 20) and 40 (𝑚 = 40) years, respectively. As
variables/countries. Using the same lag order across all the variables in the case with 𝑚 = 30, we note that the estimated coefficients of
and countries help reduce the possible adverse effects of data mining the precipitation variables, 𝜃̂𝛥𝑃̃𝑖𝑡 (𝑚)+ and 𝜃̂𝛥𝑃̃𝑖𝑡 (𝑚)− , are not statistically
that could accompany the use of country and variable specific lag order significant (specification 1). However, the estimated coefficients of
selection procedures such as Akaike or Schwarz criteria. Note also that the deviations of temperature from its historical norm are statistically
our primary focus here is on the long-run estimates rather than the significant in both specifications. The speed of adjustment to long-run
specific dynamics that might be relevant for a particular country. equilibrium (𝜙) ̂ is quick in both specifications and for different values
Table 1 presents the estimation results for two specifications of the of 𝑚. However, this does not mean that the effects of changes in 𝑇̃𝑖𝑡 (𝑚)+
panel ARDL regression in (1) and different adaptation speeds (𝑚 = and 𝑇̃𝑖𝑡 (𝑚)− are short lived.
20, 30 and 40). We report the fixed effects (FE) estimates of the long- As discussed above, estimates of the coefficients of 𝑇̃𝑖𝑡 (𝑚)+ and
run impact of changes in temperature and precipitation variables on 𝑇̃𝑖𝑡 (𝑚)− are very similar in magnitude. There is, therefore, little ev-
GDP per capita growth (𝜃), ̂ and the estimated coefficients of the error idence of asymmetry in the long-run relationship between output
correction term (𝜙) ̂ in columns (a). When the cross-sectional dimension growth and positive or negative deviations of temperature from its
of the panel is larger than the time dimension (in our panel, 𝑁 = 174 historical norm (or the country-specific threshold). This lack of asym-
and the average 𝑇 ≈ 38, see Table 1), the standard FE estimator suffers metry suggests that a simpler specification might be preferred and we
from small-𝑇 bias regardless of whether the lags of the dependent therefore re-estimate Eq. (1) by replacing 𝐱̃ 𝑖𝑡 (𝑚) = [𝑇̃𝑖𝑡 (𝑚)+ , 𝑇̃𝑖𝑡 (𝑚)− ,
( )′
variable are included or not, so long as one or more of the regressors 𝑃̃𝑖𝑡 (𝑚)+ , 𝑃̃𝑖𝑡 (𝑚)− ]′ with 𝐱̃ 𝑖𝑡 (𝑚) = ||𝑇̃𝑖𝑡 (𝑚)|| , ||𝑃̃𝑖𝑡 (𝑚)|| . The FE and HPJ-
are not strictly exogenous (see Chudik et al., 2018). Since the lagged FE results are reported in Table 2. Like our earlier results, permanent
values of growth and temperature/precipitation can be correlated with deviations of precipitation from their historical norms do not affect
the lagged values of the error term 𝜀𝑖𝑡 , the regressors (climate variables) long-term growth, but permanent deviations of temperature from their
are weakly exogenous, and hence, inference based on the standard FE time-varying historical norms have a negative effect on long-run GDP
estimator is invalid and can result in large size distortions. To deal with growth, with the magnitudes of the coefficient of ||𝑇̃𝑖𝑡 (𝑚)|| being similar
these issues, we use the half-panel Jackknife FE (HPJ-FE) estimator to those reported for 𝑇̃𝑖𝑡 (𝑚)+ and 𝑇̃𝑖𝑡 (𝑚)− in Table 1. Focusing on
of Chudik et al. (2018) and report the results in columns (b) of Table 1 Specification 2 with 𝑥̃ 𝑖𝑡 (𝑚) = ||𝑇̃𝑖𝑡 (𝑚)|| and the HPJ-FE estimates (our
alongside the estimated coefficients of the error correction term (𝜙). ̂ preferred model and estimator), we observe that 𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| is robust to
alternative ways of measuring 𝑇𝑖,𝑡−1 ∗ (𝑚).
The jackknife bias correction requires 𝑁, 𝑇 → ∞, but it allows 𝑇 to
rise at a much slower rate than 𝑁. To put our results into perspective, note that models that relate tem-
Specification 1 of Table 1 for 𝑚 = 30 reports the baseline results. perature to GDP levels yield income loss estimates that are relatively
The FE and HPJ-FE estimated coefficients of the precipitation variables, small—consistent with damage functions embedded in major integrated
𝜃̂𝛥𝑃̃𝑖𝑡 (𝑚)+ and 𝜃̂𝛥𝑃̃𝑖𝑡 (𝑚)− , are not statistically significant. However, long-run assessment models (IAMs). Specifically, most such models find that
economic growth is adversely affected when temperature deviates from when a poor (hot) country gets 1 ◦ C warmer, the level of its GDP
its time-varying historical norm persistently, as 𝜃̂𝛥𝑇̃𝑖𝑡 (𝑚)+ and 𝜃̂𝛥𝑇̃𝑖𝑡 (𝑚)− per capita falls by 1–3 percent; (ii) when a rich (temperate) country
are both statistically significant. The HPJ-FE estimates suggest that a gets 1 ◦ C warmer, there is little impact on its economic activity. The
0.01 ◦ C annual increase in the temperature above its historical norm IAMs have been extensively used in the past few decades to investigate
reduces real GDP per capita growth ( ) by 0.0577 percentage points per the welfare effects of temperature increases by relying on aggregation
2
year – calculated as −0.894 × 𝑚+1 – and a 0.01 ◦ C annual decrease of sector-specific effects, see Tol (2014); they have also been used
in the temperature below its historical norm reduces real GDP per as tools for policy analyses (including by the Obama administration,
capita growth see Obama, 2017, and at international forums). More recent studies,
( )by 0.0505 percentage points per year—calculated as
−0.783 × 𝑚+1 2
. As expected, the FE estimates (which are widely that relate temperature to GDP growth (possibly nonlinearly), arguably
used in the literature) are smaller than their HPJ-FE counterparts in show that a shift to a higher (but nonincreasing) temperature level
absolute values.12 Therefore, bias correction is important, including reduces per capita output growth substantially (with compounding
for the counterfactual exercises in Section 3; otherwise the cumulative effects over time). For example, Burke et al. (2015) consider a panel
effects of climate change could be underestimated. specification that includes quadratic climate variables in regressions
Since the baseline estimates of deviations of precipitation variables and detect: (i) non-linearity in the relationship with a universal optimal
from their historical norms (both above and below) are not statistically temperature level of 13 ◦ C; (ii) differential impact on hot versus cold
significant for 𝑚 = 30, we re-estimate Eq. (1) without them; setting countries with opposite sign; and (iii) weak lagged effects—their higher
𝐱̃ 𝑖𝑡 (𝑚) = [𝑇̃𝑖𝑡 (𝑚)+ , 𝑇̃𝑖𝑡 (𝑚)− ]′ in specification 2. The results show that lag order (between 1 and 5) estimates reported in Supplementary Table
persistent deviations of temperature above or below its historical norm, S2, show that only 3 out of 18 estimates are statistically significant.
𝑇̃𝑖𝑡 (𝑚)+ or 𝑇̃𝑖𝑡 (𝑚)− , have negative effects on long-run economic growth. However, our results show that an increase in temperature above its
Specifically, the HPJ-FE estimates suggest that a persistent 0.01 ◦ C historical norm for an extended period of time is associated with
increase in the temperature above its historical norm reduces real GDP lower economic growth in the long run – suggesting that a temporary
per capita growth by 0.0586 percentage points per annum in the long temperature shock will only have short-term growth effects but climate
run (being change – by shifting the long-term average and variability of weather—
( statistically
) significant at the 1% level) – calculated as
2 could impact an economy’s ability to grow in the long-term. Moreover,
−0.908 × 𝑚+1 – and a 0.01 ◦ C annual decrease in the temperature
the marginal impact of weather shocks are estimated to be larger than
below its historical norm reduces real GDP per capita growth by 0.0520
most papers in the literature and vary across hot and cold climates.
percentage points per year (being statistically significant at the 5%
Therefore, our findings call for a more forceful policy response to
climate change.
11
See also Chudik et al. (2013) and Chudik et al. (2017). If the world economy were adapting to climate change, ceteris
12
Since the half-panel Jackknife procedure splits the data set into two paribus, should we not expect the impact of temperature increases
halves, for countries with an odd number of time observations, we drop the to be shrinking over time? To investigate this hypothesis, we re-
first observation. Thus, the number of observations in Columns (a) and (b) are estimate our preferred model (with 𝑚 = 30 and 𝑥̃ 𝑖𝑡 (𝑚) = ||𝑇̃𝑖𝑡 (𝑚)||)
somewhat different. over different time windows using real GDP per capita growth as the

5
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Table 1
Long-run effects of climate change on per capita real GDP growth, 1960–2014.
Specification 1 Specification 2
𝑚 = 20 𝑚 = 30 𝑚 = 40 𝑚 = 20 𝑚 = 30 𝑚 = 40
(a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE
𝜃̂𝛥𝑇̃𝑖𝑡 (𝑚)+ −0.373*** −0.566*** −0.583*** −0.894*** −0.701*** −1.072*** −0.378*** −0.572*** −0.586*** −0.908*** −0.709*** −1.105***
(0.141) (0.209) (0.195) (0.291) (0.248) (0.373) (0.141) (0.208) (0.196) (0.290) (0.249) (0.372)
𝜃̂𝛥𝑇̃𝑖𝑡 (𝑚)− −0.441** −0.500** −0.699** −0.783** −0.834* −0.909* −0.451** −0.508** −0.712** −0.806** −0.851* −0.954**
(0.217) (0.249) (0.346) (0.380) (0.445) (0.485) (0.217) (0.249) (0.346) (0.380) (0.446) (0.485)
𝜃̂𝛥𝑃̃𝑖𝑡 (𝑚)+ −0.044 −0.031 0.104 0.122 −0.058 −0.005 – – – – – –
(0.289) (0.357) (0.485) (0.556) (0.684) (0.766)
𝜃̂𝛥𝑃̃𝑖𝑡 (𝑚)− −0.072 −0.175 −0.132 −0.320 −0.382 −0.595 – – – – – –
(0.323) (0.431) (0.576) (0.660) (0.754) (0.857)
𝜙̂ 0.671*** 0.603*** 0.671*** 0.603*** 0.671*** 0.602*** 0.672*** 0.604*** 0.671*** 0.604*** 0.671*** 0.604***
(0.049) (0.045) (0.049) (0.045) (0.049) (0.045) (0.049) (0.045) (0.049) (0.045) (0.049) (0.045)
𝑁 174 174 174 174 174 174 174 174 174 174 174 174
max 𝑇 50 50 50 50 50 50 50 50 50 50 50 50
avg 𝑇 38.59 38.36 38.59 38.36 38.59 38.36 38.59 38.36 38.59 38.36 38.59 38.36
min 𝑇 2 2 2 2 2 2 2 2 2 2 2 2
𝑁 ×𝑇 6714 6674 6714 6674 6714 6674 6714 6674 6714 6674 6714 6674
∑𝑝 ∑𝑝 ′
Notes: Specification 1 (the baseline) is given by 𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝓁=1 𝜑𝓁 𝛥𝑦𝑖,𝑡−𝓁 + 𝓁=0 𝛽𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 + 𝜀𝑖𝑡 , where 𝑦𝑖𝑡 is the log of real GDP per capita of country 𝑖 in year 𝑡,
( )[ ] ( )[ ]
2 2
𝐱̃ 𝑖𝑡 (𝑚) = [𝑇̃𝑖𝑡 (𝑚)+ , 𝑇̃𝑖𝑡 (𝑚)− , 𝑃̃𝑖𝑡 (𝑚)+ , 𝑃̃𝑖𝑡 (𝑚)− ]′ , 𝑇̃𝑖𝑡 (𝑚) = 𝑚+1 ∗
𝑇𝑖𝑡 − 𝑇𝑖,𝑡−1 (𝑚) and 𝑃̃𝑖𝑡 (𝑚) = 𝑚+1 ∗
𝑃𝑖𝑡 − 𝑃𝑖,𝑡−1 (𝑚) are measures of temperature and precipitation relative to their historical
∑𝑚 ∑𝑚
norms per annum, 𝑇𝑖𝑡 and 𝑃𝑖𝑡 are the population-weighted average temperature and of precipitation country 𝑖 in year 𝑡, and 𝑇𝑖,𝑡−1 ∗
(𝑚) = 𝑚1 𝓁=1 𝑇𝑖,𝑡−𝓁 and 𝑃𝑖,𝑡−1
∗
(𝑚) = 𝑚1 𝓁=1 𝑃𝑖,𝑡−𝓁
are the time-varying historical norms of temperature and precipitation over the preceding 𝑚 years. 𝑧+ = 𝑧𝐼(𝑧 ≥ 0), and 𝑧− = −𝑧𝐼(𝑧 < 0). The long-run effects, 𝜃𝑖 , are calculated
∑𝑝 ∑𝑝
from the OLS estimates of the short-run coefficients in Eq. (1): 𝜃 = 𝜙−1 𝓁=0 𝛽𝓁 , where 𝜙 = 1 − 𝓁=1 𝜑𝓁 . Specification 2 drops the precipitation variables from the baseline model:
[ + −
]′
𝐱̃ 𝑖𝑡 (𝑚) = 𝑇̃𝑖𝑡 (𝑚) , 𝑇̃𝑖𝑡 (𝑚) . Columns labelled (𝑎) report the FE estimates and columns labelled (𝑏) report the half-panel Jackknife FE (HPJ-FE) estimates, which corrects the bias in
columns (𝑎). The standard errors are estimated by the estimator proposed in Proposition 4 of Chudik et al. (2018). Asterisks indicate statistical significance at the 1% (***), 5%
(**), and 10% (*) levels.

Table 2
Long-run effects of climate change on per capita real GDP growth, 1960–2014 (using absolute value of deviations of climate variables from their historical norm).
Specification 1 Specification 2
𝑚 = 20 𝑚 = 30 𝑚 = 40 𝑚 = 20 𝑚 = 30 𝑚 = 40
(a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.375*** −0.523*** −0.582*** −0.836*** −0.702*** −0.981*** −0.379*** −0.529*** −0.583*** −0.841*** −0.706*** −0.996***
(0.142) (0.201) (0.199) (0.284) (0.252) (0.361) (0.142) (0.201) (0.199) (0.284) (0.253) (0.361)
𝜃̂𝛥|𝑃̃𝑖𝑡 (𝑚)| −0.070 −0.125 −0.032 −0.131 −0.259 −0.404 – – – – – –
(0.237) (0.335) (0.473) (0.527) (0.646) (0.709) – – – – – –
𝜙̂ 0.671*** 0.604*** 0.671*** 0.604*** 0.671*** 0.603*** 0.672*** 0.604*** 0.672*** 0.604*** 0.672*** 0.604***
(0.049) (0.045) (0.049) (0.045) (0.049) (0.045) (0.049) (0.045) (0.049) (0.045) (0.049) (0.045)
𝑁 174 174 174 174 174 174 174 174 174 174 174 174
max 𝑇 50 50 50 50 50 50 50 50 50 50 50 50
avg 𝑇 38.36 38.36 38.36 38.36 38.36 38.36 38.36 38.36 38.36 38.36 38.36 38.36
min 𝑇 2 2 2 2 2 2 2 2 2 2 2 2
𝑁 ×𝑇 6714 6674 6714 6674 6714 6674 6714 6674 6714 6674 6714 6674
∑𝑝 ∑𝑝
Notes: Specification 1 (the baseline) is given by 𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝓁=1 𝜑𝓁 𝛥𝑦𝑖,𝑡−𝓁 + 𝓁=0 𝛽𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 +𝜀𝑖𝑡 , where 𝑦𝑖𝑡 is the log of real GDP per capita of country 𝑖 in year 𝑡, 𝐱̃ 𝑖𝑡 (𝑚) = [||𝑇̃𝑖𝑡 (𝑚)|| , ||𝑃̃𝑖𝑡 (𝑚)||] ,
′ ′
( )[ ] ( )[ ]
2 2
𝑇̃𝑖𝑡 (𝑚) = 𝑚+1 ∗
𝑇𝑖𝑡 − 𝑇𝑖,𝑡−1 (𝑚) and 𝑃̃𝑖𝑡 (𝑚) = 𝑚+1 ∗
𝑃𝑖𝑡 − 𝑃𝑖,𝑡−1 (𝑚) are measures of temperature and precipitation relative to their historical norms per annum, 𝑇𝑖𝑡 and 𝑃𝑖𝑡 are the
∑𝑚 ∑𝑚
population-weighted average temperature and of precipitation country 𝑖 in year 𝑡, and 𝑇𝑖,𝑡−1 ∗
(𝑚) = 𝑚1 𝓁=1 𝑇𝑖,𝑡−𝓁 and 𝑃𝑖,𝑡−1
∗
(𝑚) = 𝑚1 𝓁=1 𝑃𝑖,𝑡−𝓁 are the time-varying historical norms
of temperature and precipitation over the preceding 𝑚 years in each 𝑡. The long-run effects, 𝜃𝑖 , are calculated from the OLS estimates of the short-run coefficients in Eq. (1):
∑𝑝 ∑𝑝
𝜃 = 𝜙−1 𝓁=0 𝛽𝓁 , where 𝜙 = 1 − 𝓁=1 𝜑𝓁 . Specification 2 drops the precipitation variable from the baseline model. The standard errors are estimated by the estimator proposed in
Proposition 4 of Chudik et al. (2018). Asterisks indicate statistical significance at the 1% (***), 5% (**), and 10% (*) levels.

dependent variable. We start with the full sample, 1960–2014, and added to sectors that are more exposed to climate change). Fourth,
then drop a year at a time (with the last estimation being carried if firms underestimate the likelihood or severity of future weather
out for the sub-sample 1983–2014). The results are plotted in Fig. 4, events, they may not adapt sufficiently; i.e. adaptation technologies are
showing that the estimated coefficients on 𝛥 ||𝑇̃𝑖𝑡−𝓁 (𝑚)|| are becoming readily available but the take-up so far has been limited by firms. In a
larger (in absolute value) over time. Do these results cast doubt on the survey of private sector organizations across multiple industries within
efficacy of adaptation efforts over the last five decades? Ceteris paribus, the Organization for Economic Cooperation and Development (OECD)
while it is expected that adaptation weakens the relationship between countries, Agrawala et al. (2011) find that only few firms have taken
temperature and economic growth over time, we cannot conclude that sufficient steps to assess and manage the risks from climate change.
the world economy has not been adapting to climate change based Fifth, according to Deryugina and Hsiang (2014) firms tend to under-
on Fig. 4. First, adaptation efforts might be concentrated in certain invest in adaptation owing to its high cost.13 Overall, the evidence
countries (typically advanced economies) and certain sectors. Second, appears to suggest that (at least for now) adaptation has so far had
it may be the case that adaptation is not keeping pace with the climate
change; i.e., global temperatures have increased at an unprecedented
pace over the past 40 years. Third, the effects of adaptation might have 13
Other reasons for underinvestment include knowledge spillovers and
been offset by structural changes to the economy (that is a shift of value networks externalities.

6
M.E. Kahn et al. Energy Economics 104 (2021) 105624

economies; they are more urbanized and much of the economic activity
takes place indoors. For instance, Singapore has attempted to insulate
its economy from the heat by extensively engaging in economic activity
in places with air conditioning. Therefore, if individuals are aware of
how extreme heat affects their economic performance, they can invest
in self protection to reduce their exposure to such risks.14 Mendelsohn
(2016) also argues that economic effects of weather shocks are likely
to be very different in cold versus hot climates.
Given our heterogeneous sample of 174 countries and motivated by
above studies, an immediate question is whether the estimated adverse
long-run growth effects of weather shocks in Specifications 1 and 2 of
Table 2 are driven by poor countries. We, therefore, follow Dell et al.
(2012) and Burke et al. (2015) and augment Specification 2 with an
interactive term, 𝛥̃𝐱𝑖,𝑡−𝓁 (𝑚) × I (country 𝑖 is poor), to capture any possi-
ble differential effects of temperature changes from the moving-average
norm for the rich and poor countries:
∑
𝑝 ∑
𝑝
′
𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝜑𝓁 𝛥𝑦𝑖,𝑡−𝓁 + 𝜷 𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 (𝑚)
𝓁=1 𝓁=0
(2)
∑
𝑝
+ 𝜻 ′𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 (𝑚) × I (country 𝑖 is poor) + 𝜀𝑖𝑡 ,
𝓁=0
Fig. 4. Rolling Estimates of the Long-Run Effects of Temperature Increases on per where, as in Burke et al. (2015), we define country 𝑖 as poor (rich) if
capita Real GDP Growth. Notes: Figure shows the long-run effects (and their 95%
its purchasing-power-parity-adjusted (PPP) GDP per capita was below
standard error bands) of temperature increases on per capita real GDP growth over
different time windows, using the ARDL specification in (1). We start the estimation (above) the global median in 1980. Moreover, to investigate whether
with the full sample (1960–2014) and then drop one year at a time, ending with the temperature increases affect hotter countries more than colder ones, we
final estimates based on the 1983–2014 sub-sample. estimated the following panel data model
∑
𝑝 ∑
𝑝
′
𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝜑𝓁 𝛥𝑦𝑖,𝑡−𝓁 + 𝜷 𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 (𝑚)
𝓁=1 𝓁=0
(3)
∑
𝑝
+ 𝝃 ′𝓁 𝛥̃𝐱𝑖,𝑡−𝓁 (𝑚) × I (country 𝑖 is hot) + 𝜀𝑖𝑡 ,
𝓁=0

where a country is defined as cold (hot) if its historical average temper-

ature is below (above) the global median. The results from estimating
specifications (2) and (3) are reported in Table 3 where the estimated
coefficients of the interactive terms are not statistically significant—
hence, we cannot reject the hypothesis that there are no differential
effects of climate change on poor versus rich nations or hot versus cold
countries.
The statistical insignificance of estimated coefficients of the inter-
active terms in Table 3 may be due to lack of statistical power. In
what follows, we attempt to explore the heterogeneity issue further by
relying on the half-panel Jackknife Mean Group (HPJ-MG) estimator in
the context of the following heterogeneous panel data model:
Fig. 5. {𝜓𝑗 } for 𝑗 = 0, 1, 2, … , 20.
𝑝𝑇̃
∑
𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝜔𝑖 𝛥𝑦̄𝑤,𝑡−1 + 𝜑𝑖 𝛥𝑦𝑖,𝑡−𝓁 + 𝛽𝑖𝓁 𝛥 ||𝑇̃𝑖𝑡−𝓁 (𝑚)|| + 𝜀𝑖𝑡 . (4)
𝓁=0
limited impact in dampening the negative effects of climate change Unlike pooled estimation techniques such as FE where only intercept
globally. But it is possible that with greater public awareness and heterogeneity is taken into account, the above specification allows for
government efforts, we will be seeing a much faster rate of adaptation the marginal effects of weather shocks to vary across countries or
in the future. Our analysis is counterfactual given the current state sub-group of countries. Under this more general specification, country-
of the world, and outcomes could, and hopefully will, deviate from specific marginal effects can be estimated by running least squares
our counterfactual with better and more forceful environmental polices regressions for each country 𝑖 separately, and then considering averages
(both mitigation and adaptation). or medians of the estimated coefficients across countries or regions,
for example cold versus hot climates, or rich versus poor countries.
2.2. Weather effects across climates and income groups Pesaran and Smith (1995) show that simple averages of the estimated
coefficients (known as mean group, MG, estimates) result in consistent
The literature provides evidence for uneven effects of temperature estimates of the underlying population means of the parameters when
shocks, with worse adverse consequences in economies with hot cli- the time-series dimension of the data is sufficiently large. Whilst it is
mates and/or in low-income countries; see, for instance, Sachs and not possible to be sure about when 𝑇 is sufficiently large, Monte Carlo
Warner (1997), Jones and Olken (2010), Dell et al. (2012), Burke evidence suggests that reliable estimates can be obtained with 𝑇 ≥ 30
et al. (2015) and Mejia et al. (2018). In other words, when a rich
(temperate) country gets warmer, there will be little impact on its eco-
nomic activity. There are intuitive reasons and anecdotal evidence for 14
For a survey of the literature on heat and productivity, see Heal and Park
this, including adaptation that has taken place particularly in advanced (2016).

7
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Table 3 3. Counterfactual analysis

Long-run effects of climate change on per capita real GDP growth of poor and hot
countries, 1960–2014 (using absolute value of deviations of climate variables from
their historical norm). We perform a number of counterfactual exercises to measure the
Specification 3 Specification 4
cumulative output per capita effects of persistent increases in annual
temperatures above their norms (or thresholds) over the period 2015–
𝑚 = 30 (a) FE (b) HPJ-FE (a) FE (b) HPJ-FE
2100. We carry out this analysis using the HPJ-FE estimates based on
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.551** −0.836** −0.754*** −1.029***
the ARDL specification given by (1), which we write equivalently as
(0.235) (0.368) (0.200) (0.287)
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)|×I(𝑖 is poor) −0.156 −0.137 – – 𝜑 (𝐿) 𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝛽(𝐿)𝛥𝑥𝑖𝑡 (𝑚) + 𝜀𝑖𝑡 ,
| ∗ (𝑚)|, 𝜑 (𝐿) = 1− ∑4 ∑4
(0.396) (0.586) – –
where 𝑥𝑖𝑡 (𝑚) = |𝑇𝑖𝑡 − 𝑇𝑖𝑡−1 | 𝑙
𝓁=1 𝜑𝓁 𝐿 , 𝛽(𝐿) =
𝑙
𝓁=0 𝛽𝓁 𝐿 ,
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)|×I(𝑖 is hot) – – 0.496 0.562 | |
and 𝐿 is the lag operator. Pre-multiplying both sides of the above
– – (0.420) (0.656)
equation by the inverse of 𝜑 (𝐿) yields
𝜙̂ 0.661*** 0.596*** 0.672*** 0.605***
(0.050) (0.047) (0.049) (0.045) 𝛥𝑦𝑖𝑡 = 𝑎̃𝑖 + 𝜓(𝐿)𝛥𝑥𝑖𝑡 + 𝜗(𝐿)𝜀𝑖𝑡 , (5)
𝑁 165 165 174 174
max 𝑇 50 50 50 50 where 𝑎̃𝑖 = 𝜑(1)−1 𝑎
𝑖 , 𝜗(𝐿) = 𝜗0 +𝜗1 𝐿+𝜗2 𝐿2 +⋯ and 𝜓(𝐿) = 𝜑(𝐿)−1 𝛽(𝐿) =
avg 𝑇 38.76 38.76 38.36 38.36 𝜓0 + 𝜓1 𝐿 + 𝜓2 𝐿2 + ⋯ for 𝑗 = 0, 1, 2, ….15
min 𝑇 8 8 2 2 The counterfactual effects of climate change can now be derived
𝑁 ×𝑇 6431 6396 6714 6674
by comparing the output trajectory of country 𝑖 over the period 𝑇 + 1
Notes: See notes to Table 2. Specifications 3 and 4 interact the temperature variables to 𝑇 + ℎ under the no change scenario denoted by 𝑏0𝑇 and 𝜎𝑇0 , with
with dummies for poor and hot countries, respectively (see Eqs. (2) and (3)). The 𝑖 𝑖
an alternative expected trajectory having the counterfactual values of
standard errors are estimated by the estimator proposed in Proposition 4 of Chudik
𝑏1𝑇 and 𝜎𝑇1 . Denoting the values of 𝑥𝑖𝑡 for 𝑡 = 𝑇 + 1, 𝑇 + 2, … , 𝑇 + ℎ
et al. (2018). Asterisks indicate statistical significance at the 1% (***), 5% (**), and 𝑖 𝑖 { }
10% (*) levels. under these two scenarios by 𝐱𝑖,𝑇 0 = 𝑥0𝑖,𝑇 +1 , 𝑥0𝑖,𝑇 +2 , … , 𝑥0𝑖,𝑇 +ℎ ,
{ +1,𝑇 +ℎ }
and 𝐱𝑖,𝑇1 = 𝑥1𝑖,𝑇 +1 , 𝑥1𝑖,𝑇 +2 , … , 𝑥1𝑖,𝑇 +ℎ , the counterfactual output
+1,𝑇 +ℎ
change can be written as
and 𝑁 ≥ 20, when output growth is not very persistent, which is the ( ) ( )
| 1 | 0
case in our applications. We, therefore, select countries for which we 𝜉𝑖,𝑇 +ℎ = E 𝑦𝑖,𝑇 +ℎ |ϝ𝑖,𝑇 , 𝐱𝑖,𝑇 − E 𝑦𝑖,𝑇 +ℎ |ϝ𝑖,𝑇 , 𝐱𝑖,𝑇 ,
| +1,𝑇 +ℎ | +1,𝑇 +ℎ
have at least 30 years of observations for GDP growth, resulting in a
sample of 130 countries. To explore heterogeneous responses across where ϝ𝑖𝑇 = (𝑦𝑖𝑇 , 𝑦𝑖,𝑇 −1 , 𝑦𝑖,𝑇 −2 , … ; 𝑥𝑖𝑇 , 𝑥𝑖,𝑇 −1 , 𝑥𝑖,𝑇 −2 , …). Cumulating
regions, we define a country as cold if its historical average temper- both sides of (5) from 𝑡 = 𝑇 + 1 to 𝑇 + ℎ and taking conditional
ature is among the bottom third of the temperature distribution. Other expectations under the two scenarios we have
countries fall into temperate or hot climates. Poor and rich countries ∑
ℎ ( )
are selected based on International Monetary Fund’s classifications. The 𝜉𝑖,𝑇 +ℎ = 𝜓ℎ−𝑗 𝑥1𝑖,𝑇 +𝑗 − 𝑥0𝑖,𝑇 +𝑗 , (6)
results from estimating equation (4) with and without lagged world 𝑗=1

output growth, 𝛥𝑦̄𝑤,𝑡−1 , are reported in Table 4. The inclusion of 𝛥𝑦̄𝑤,𝑡−1 The impact of climate change clearly depends on the magnitude of
serves two purposes: (1) it accounts for unobserved global factors, and 𝑥1𝑖,𝑇 +𝑗 − 𝑥0𝑖,𝑇 +𝑗 .
(2) it renders the errors of the regressions across countries weakly We consider the output effects of country-specific average annual
(rather than strongly) correlated. increases in temperatures over the period 2015–2100 as predicted
Key findings are as follows: First, the HPJ-MG estimation results for under RCP 2.6 and RCP 8.5 scenarios, and compare them with a
the sample of all 130 countries are similar (in sign and statistical signifi- baseline scenario under which temperature in each country increases
cance) to those reported in Table 2. Specifically, persistent temperature according to its historical trend of 1960–2014.16 However, owing to the
deviations from their historical norms (owing to climate change) are non-linear nature of our output-growth specification, changes in trend
estimated to have a negative effect on long-run per capita GDP growth temperature do not translate on a one-to-one basis to absolute changes
(especially when 𝛥𝑦̄𝑤,𝑡−1 is included as an additional regressor). The in temperature. In line with (A.34), future temperature changes over
mean group estimates for all 130 countries can be viewed as the the counterfactual horizon, 𝑇 + 𝑗, 𝑗 = 1, 2, …. can be represented by
weighted average of the estimates for cold and hot countries, and
the weighted average of the estimates for poor and rich economies. 𝑇𝑖,𝑇 +𝑗 = 𝑎𝑇 𝑖 + 𝑏𝑇 𝑖,𝑗 (𝑇 + 𝑗) + 𝑣𝑇 𝑖,𝑇 +𝑗 , for 𝑗 = 1, 2, … , (7)
Second, there is some evidence that negative growth effects of weather where we allow for the trend change in the temperature to vary over
shocks are less severe in cold climates. For example, using the reported time. The above equation reduces to (A.34) if we set 𝑏𝑇 𝑖,𝑗 = 𝑏𝑇 𝑖 for all
standard errors for the average estimated coefficients on cold and hot 𝑗. Suppose also that, as before, the historical norm variable associated
countries in regressions featuring 𝛥𝑦̄𝑤,𝑡−1 and 𝑚 = 30 (see Column 6 of with 𝑇𝑖,𝑇 +𝑗 , namely 𝑇𝑖,𝑇 ∗ (𝑚), is constructed using the past 𝑚 years.
+𝑗−1
Table 4), the 95% range estimates can be calculated as (−0.342 − 1.96 ∗ Then it is easy to show that
0.1509 = −0.64 to −0.342 + 1.96 ∗ 0.1509 = −0.05) for cold countries ( )
∗ 𝑚+1
and (−1.180 − 1.96 ∗ 0.3713 = −1.91 to −1.180 + 1.96 ∗ 0.3713 = −0.45) 𝑇𝑖,𝑇 +𝑗 − 𝑇𝑖,𝑇 +𝑗−1
(𝑚) = 𝑏𝑇 𝑖,𝑗
for hot countries. As can be seen, the 95% interval for cold climates (2 ) (8)
+ 𝑣𝑇 𝑖,𝑇 +𝑗 − 𝑣̄ 𝑇 𝑖,𝑇 +𝑗−1,𝑚 , 𝑗 = 1, 2, … , ℎ,
largely falls outside the 95% interval for hot climates, suggesting more
∑
severe growth effects of weather shocks in hot countries. Nevertheless, where 𝑣̄ 𝑇 𝑖,𝑇 +𝑗−1,𝑚 = 𝑚−1 𝑚 𝑠=1 𝑣𝑇 𝑖,𝑇 +𝑗−𝑠 . The realized values of
the impact of persistent changes in ||𝑇̃𝑖𝑡 (𝑚)|| on GDP growth in cold | ∗ |
|𝑇𝑖,𝑇 +𝑗 − 𝑇𝑖,𝑇 +𝑗−1 (𝑚)| depend on the probability distribution of weather
climates is still negative, statistically significant, and increasing with | |
shocks, 𝑣𝑇 𝑖,𝑇 +𝑗 , as well as the trend change in temperature, given by
𝑚 (namely depends on how fast adaptation is taking place). Third, 𝑏𝑇 𝑖,𝑗 . As a first order approximation, and in order to obtain analytic
while poor countries are found to be disproportionately affected by
weather shocks, rich countries are by no means immune to climate
change. Note that lagged world output growth, 𝛥𝑦̄𝑤,𝑡−1 , plays a crucial 15
We are suppressing the dependence of 𝑥𝑖𝑡 on 𝑚 to simplify the exposition.
role in accounting for global output trends that likely interact with 16
A similar analysis can also be carried out in terms of changes in precip-
global climate conditions. The weather effects are generally weaker itation. For brevity and given the empirical results in Section 2, we focus on
when 𝛥𝑦̄𝑤,𝑡−1 is excluded from regressions. the counterfactual effects of changes in temperature only.

8
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Table 4
Mean group estimates of the long-run effects of climate change on per capita real GDP growth, 1960–2014.
Excluding 𝛥𝑦̄𝑤,𝑡−1 Including 𝛥𝑦̄𝑤,𝑡−1
Historical norm: 𝑚 = 20 𝑚 = 30 𝑚 = 40 𝑚 = 20 𝑚 = 30 𝑚 = 40
(a) All 130 countries
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.447* −0.487 −0.521 −0.706*** −0.918** −1.051**
(0.234) (0.367) (0.473) (0.237) (0.393) (0.519)
𝑁 ×𝑇 6198 6198 6198 6020 6020 6020
(b) Cold (𝑇̄𝑖 < 33𝑡ℎ 𝑃 𝑒𝑟𝑐𝑒𝑛𝑡𝑖𝑙𝑒)
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.227** −0.230* −0.198 −0.238** −0.342** −0.457***
(0.101) (0.128) (0.175) (0.105) (0.151) (0.169)
𝑁 ×𝑇 2090 2090 2090 1964 1964 1964
(c) Temperate or hot (𝑇̄𝑖 ≥ 33𝑡ℎ 𝑃 𝑒𝑟𝑐𝑒𝑛𝑡𝑖𝑙𝑒)
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.665*** −0.780*** −0.613 −0.842*** −1.180*** −1.212**
(0.193) (0.302) (0.431) (0.222) (0.371) (0.504)
𝑁 ×𝑇 4108 4108 4108 3990 3990 3990
(d) Poor (Low-income developing countries)
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.603** −0.759* −0.855* −1.020*** −1.463*** −1.703***
(0.270) (0.406) (0.488) (0.262) (0.429) (0.547)
𝑁 ×𝑇 3140 3140 3140 3048 3048 3048
(e) Rich (Advanced economies and G20 emerging markets)
𝜃̂𝛥|𝑇̃𝑖𝑡 (𝑚)| −0.586*** −0.849*** −1.047*** −0.587*** −1.003*** −1.280***
(0.195) (0.272) (0.373) (0.209) (0.310) (0.392)
𝑁 ×𝑇 1794 1794 1794 1734 1734 1734
∑𝑝𝑇̃
Notes: Specification 1 is given by 𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝜑𝑖 𝛥𝑦𝑖,𝑡−𝓁 + 𝓁=0 𝛽𝑖𝓁 𝛥 ||𝑇̃𝑖𝑡−𝓁 (𝑚)|| + 𝜀𝑖𝑡 , where 𝑦𝑖𝑡 is the log of real GDP per
( )[ ]
2
capita of country 𝑖 in year 𝑡, 𝑇̃𝑖𝑡 (𝑚) = 𝑚+1 ∗
𝑇𝑖𝑡 − 𝑇𝑖,𝑡−1 (𝑚) is a measure of temperature relative to its historical norm
∑𝑚
per annum, 𝑇𝑖𝑡 is the population-weighted average temperature of country 𝑖 in year 𝑡, and 𝑇𝑖,𝑡−1 ∗
(𝑚) = 𝑚1 𝓁=1 𝑇𝑖,𝑡−𝓁 is
the time-varying historical norm of temperature over the preceding 𝑚 years in each 𝑡. Specification 2 is given by 𝛥𝑦𝑖𝑡 =
∑𝑝𝑇̃
𝑎𝑖 + 𝜔𝑖 𝛥𝑦̄𝑤,𝑡−1 + 𝜑𝑖 𝛥𝑦𝑖,𝑡−𝓁 + 𝓁=0 𝛽𝑖𝓁 𝛥 ||𝑇̃𝑖𝑡−𝓁 (𝑚)|| + 𝜀𝑖𝑡 , where 𝑦̄𝑤𝑡 is the log of world’s real GDP per capita in year 𝑡 and the other
variables are as before. The models are estimated using the half-panel Jackknife mean-group estimator. Asterisks indicate
statistical significance at the 1% (***), 5% (**), and 10% (*) levels.

expressions, we assume that temperature shocks, 𝑣𝑇 𝑖,𝑇 +𝑗 , over 𝑗 = where 𝑇𝑖,𝑇 +𝐻 denotes the level of temperature at the end of the coun-
1, 2, …, are serially uncorrelated, Gaussian random variables with zero terfactual period. Averaging (10) over 𝑗 we have
means and variances, 𝜎𝑇2 𝑖 . Under these assumptions and using the ( )
results in Lemma 3.1 of Dhyne et al. (2011), we have 2 𝑏1𝑇 𝑖 − 𝑏0𝑇 𝑖
[ ( ) ( )] 𝑑𝑖 = . (12)
| | 𝜇𝑇 𝑖,𝑗 −𝜇𝑇 𝑖,𝑗 𝐻 +1
∗
E |𝑇𝑖,𝑇 +𝑗 − 𝑇𝑖,𝑇 (𝑚)| = 𝜇𝑇 𝑖,𝑗 𝛷 −𝛷
| +𝑗−1 | 𝜔 𝜔𝑇 𝑖 In our empirical application, we set 𝑇𝑖,𝑇 +𝐻 = 𝑇𝑖,2099 and 𝑇𝑖,𝑇 +1 = 𝑇𝑖,2015 ,
( 𝑇𝑖 ) (9)
𝜇𝑇 𝑖,𝑗 with implied 𝐻 = 85. For 𝑇𝑖,2099 , for 𝑖 = 1, 2, … , 𝑁, we consider two sets
+ 2𝜔𝑇 𝑖 𝜙 = 𝑔𝑇 𝑖 (𝑚, 𝑏𝑇 𝑖,𝑗 , 𝜎𝑇 𝑖 )
𝜔𝑇 𝑖 of values based on IPCC’s projections under the RCP 2.6 and RCP 8.5
where 𝛷(⋅) and 𝜙(⋅) are the cumulative and density distribution func- scenarios (see Table A.7). In effect, this specification assumes that over
tions of a standard Normal variate, respectively, and the counterfactual period temperature in country 𝑖 increases by 𝑗𝑑𝑖 per
( ) ( )
𝑚+1 1 annum over the period 𝑇 + 1 to 𝑇 + 𝑗 , relative to its historical trend
𝜇𝑇 𝑖,𝑗 = 𝑏𝑇 𝑖,𝑗 , and 𝜔2𝑇 𝑖 = 𝜎𝑇2 𝑖 1 + .
2 𝑚 value of 𝑏0𝑇 𝑖 .
It is clear from the above expressions that the responses of our climate
We also assume that the postulated trend rise in temperature,
variables to a postulated rise in temperature most crucially depend
on the volatility of temperature around its trend, 𝜎𝑇 𝑖 , which differs specified in (10), does not affect the volatility of temperature shocks,
markedly across countries.17 and set 𝜎𝑇1 𝑖 to its pre-counterfactual value of 𝜎𝑇0 𝑖 . This is a conservative
For the baseline scenario, we set 𝑚 = 30 and consider the following assumption and most likely will result in an under-estimation of the
counterfactual country-specific changes in the trend temperature over adverse effects of temperature increases, since one would expect rising
the period 𝑇 + 𝑗, for 𝑗 = 1, 2, … , 𝐻, as compared to the historical trend temperature to be associated with an increase in volatility.18 With these
rise in temperature (namely 𝑏0𝑇 𝑖 ):
considerations in mind, and using (6), the mean counterfactual impact
𝑏1𝑇 𝑖,𝑗 = 𝑇𝑖,𝑇 +𝑗 − 𝑇𝑖,𝑇 +𝑗−1 = 𝑏0𝑇 𝑖 + 𝑗𝑑𝑖 , for all 𝑗 = 1, 2..., 𝐻, (10) of the temperature change on output is given by
( ) ( ) ( )
where 𝑑𝑖 is the average incremental change in the trend rise in tem-
𝛥𝑖ℎ 𝑑𝑖 = E 𝑦1𝑖,𝑇 +ℎ ||ϝ𝑖,𝑇 − E 𝑦0𝑖,𝑇 +ℎ ||ϝ𝑖,𝑇
perature for country 𝑖. We set 𝑑𝑖 to ensure that the average rise in
temperature over the counterfactual period in country 𝑖 is equal to the ∑
ℎ
[ ]
hypothesized value of 𝑏1𝑇 𝑖 , and note that = 𝜓ℎ−𝑗 𝑔𝑇 𝑖 (𝑚, 𝑏0𝑇 𝑖 + 𝑗𝑑𝑖 , 𝜎𝑇0 𝑖 ) − 𝑔𝑇 𝑖 (𝑚, 𝑏0𝑇 𝑖 , 𝜎𝑇0 𝑖 ) , (13)
𝑗=1
∑
𝐻 ∑𝐻
( ) 𝑇𝑖,𝑇 +𝐻 − 𝑇𝑖,𝑇
𝑏1𝑇 𝑖 = 𝐻 −1 𝑏1𝑇 𝑖,𝑗 = 𝐻 −1 𝑇𝑖,𝑇 +𝑗 − 𝑇𝑖,𝑇 +𝑗−1 = , (11)
𝑗=1 𝑗=1
𝐻

18
Moreover, accounting for international spillover effects of climate change,
17
For estimates of 𝜎𝑇 𝑖 across countries see Table A.7. individual countries’ long-term growth effects could be larger.

9
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Table 5 Averaging the losses across countries, using PPP-GDP weights, we

Effects of climate change on per capita real GDP growth, 1960–2014.
report the global income effects of climate change under the RCP
𝛽̂0 −0.0038* 𝜑
̂1 0.2643*** No. of countries (𝑁) 174 2.6 and RCP 8.5 scenarios in Table 6. Under the Paris agreement
(0.0021) (0.0500) max 𝑇 50
objective, and assuming 𝑚 = 30, our results indicate that the world
𝛽̂1 −0.0056* 𝜑
̂2 0.0785*** avg 𝑇 38.36
(0.0029) (0.0266) min 𝑇 2 could actually benefit from mitigation policies in year 2030 (compared
𝛽̂2 −0.0084*** 𝜑
̂3 0.0547** No. of obs. (𝑁 × 𝑇 ) 6674 to a reference case in which temperatures increase according to their
(0.0031) (0.0216) historical trends of 1960–2014), while limiting the economic losses of
𝛽̂3 −0.0090*** 𝜑
̂4 −0.0016 climate change to 0.11 and 1.07 percent over the next 36 and 86 years,
(0.0026) (0.0327)
respectively. However, a persistent above-norm increase in average
𝛽̂4 −0.0060***
(0.0021) global temperature by 0.04 ◦ C per year (based on RCP 8.5) leads to
∑4 ∑4 ′
substantial output losses, reducing real per capita output by 0.80, 2.51
Notes: Estimates are based on 𝛥𝑦𝑖𝑡 = 𝑎𝑖 + 𝜑 𝛥𝑦𝑖,𝑡−𝓁 + 𝓁=0 𝛽𝓁 𝛥𝑥𝑖,𝑡−𝓁 (𝑚) + 𝜀𝑖𝑡 , where
𝓁=1 𝓁
| |
and 7.22 percent in 2030, 2050 and 2100, respectively. Overall these
𝑦𝑖𝑡 is the log of real GDP per capita of country 𝑖 in year 𝑡, 𝑥𝑖𝑡 (𝑚) = |𝑇𝑖𝑡 − 𝑇𝑖,𝑡−1
∗
(𝑚)|,
| ∗
| economic effects are somewhat larger than those obtained in existing
𝑇𝑖𝑡 is the population-weighted average temperature of country 𝑖 in year 𝑡, and 𝑇𝑖,𝑡−1 (𝑚)
is the historical temperature norm of country 𝑖 (based on moving averages of the
studies in the literature and what is generally discussed in policy circles
past 30 years). The coefficients are estimated by the half-panel Jackknife FE (HPJ-FE) (see Fig. 2).
procedure and the standard errors are based on the estimator proposed in Proposition Can adaptation help offset these negative income effects? Repeating
4 of Chudik et al. (2018). Asterisks indicate statistical significance at 1% (***), 5% the counterfactual exercise for different values of 𝑚 highlights the
(**), and 10% (*) levels.
role of adaptation. The shorter the 𝑚, the faster agents treat higher
temperatures as the new norm. Table 6 shows the effects of global
warming over time for various values of 𝑚. The results indicate that
where we base the estimates of 𝑏0𝑇 𝑖 and 𝜎𝑇0 𝑖 on the pre-counterfactual per-capita output losses are lower for 𝑚 = 20 but significantly higher
period 1960–2014 (see Table A.7), and use if it takes longer to adapt to climate change (𝑚 = 40). Overall, we
[ ( ) ( )] ( ) argue that while climate change adaptation could reduce these negative
𝜇𝑇1 𝑖,𝑗 −𝜇𝑇1 𝑖,𝑗 𝜇𝑇1 𝑖,𝑗 economic effects, it is highly unlikely to offset them entirely. More
𝑔𝑇1 𝑖 (𝑚, 𝑏1𝑇 𝑖,𝑗 , 𝜎𝑇0 𝑖 ) = 𝜇𝑇1 𝑖,𝑗 𝛷 −𝛷 + 2𝜔0𝑇 𝑖 𝜙 ,
𝜔0𝑇 𝑖 𝜔0𝑇 𝑖 𝜔0𝑇 𝑖 forceful mitigation policies are needed to limit the damage from climate
change.
(14)
Prior research projects the GDP impact of temperature increases for
some future year, typically 2100, assuming a “no further warming”
[ ( ) ( )] ( ) counterfactual (e.g., Burke et al., 2015; Kalkuhl and Wenz, 2020). Since
𝜇𝑇0 𝑖 −𝜇𝑇0 𝑖 𝜇𝑇0 𝑖
𝑔𝑇0 𝑖 (𝑚, 𝑏0𝑇 𝑖 , 𝜎𝑇0 𝑖 ) = 𝜇𝑇0 𝑖 𝛷 −𝛷 + 2𝜔0𝑇 𝑖 𝜙 , (15) there are no pathways to a scenario in which baseline temperatures
𝜔0𝑇 𝑖 𝜔0𝑇 𝑖 𝜔0𝑇 𝑖 remain constant, we compared the per capita GDP impact of tempera-
ture increases under RCP2.6 and RCP8.5 to a baseline scenario under
( )( ) ( )
𝑚+1 𝑚+1 0 which temperature in each country rises according to its historical trend
𝜇𝑇1 𝑖,𝑗 = 𝑏1𝑇 𝑖,𝑗 , 𝜇𝑇0 𝑖 = 𝑏𝑇 𝑖 , (16)
2 2 of 1960–2014. However, to have a better comparability to previous
( )1∕2 { } papers, we also performed a counterfactual exercise where temperature
and 𝜔0𝑇 𝑖 = 𝜎𝑇0 𝑖 1 + 𝑚1 . To obtain 𝜓̂ 𝑗 , we use the HPJ-FE estimates increases under RCP8.5 are compared to a baseline scenario in which
{ }4 { }4 | ∗ (𝑚)|
of 𝜷 𝓁 𝑙=0 and 𝜑𝑙 𝑙=1 from the ARDL equation with |𝑇𝑖𝑡 − 𝑇𝑖,𝑡−1 | historical temperatures are assumed to remain constant. As expected,
| |
as the climate variable. These estimates and their standard errors are the analysis results in higher per capita income losses in 2100 (4.64
reported in Table 5. Fig. 5 plots the estimates of 𝜓𝑗 for 𝑗 = 0, 1, 2, … , 20, percent for 𝑚 = 20; 7.64 percent for 𝑚 = 30; and 10.67 percent for
∑ ∞
𝑗=1 𝑗 𝜓̂ 𝑗 𝑚 = 40) than those reported in Table 6 under RCP8.5 (4.44, 7.22, and
for which the estimated mean lag is ∑∞ = 3.1943 years.
𝑗=0 𝜓
̂𝑗 9.96 percent, respectively).
We report the real GDP per capita losses from global warming
We showed that economic growth is affected not only by higher
under the RCP 2.6 and RCP 8.5 scenarios, compared to the reference temperatures but also by the degree of climate variability. To study
case, in country heat maps and for the year 2100 only, but make all the role of climate volatility in determining GDP per capita losses,
of the 174 country-specific estimates over various horizons (by year instead of setting 𝜎𝑇1 𝑖 = 𝜎𝑇0 𝑖 , we allow temperature increases to af-
2030, 2050, and 2100) available in Table A.7. Fig. 6 shows that in fect the variability of temperature shocks commensurately. That is,
the absence of climate change policies (under the RCP 8.5 Scenario we keep( the coefficient of variation unchanged, and therefore set
)
with 𝑚 = 30), the percent losses in per-capita incomes by 2100 are 𝜎𝑇1 𝑖 = 𝜇𝑇1 𝑖,𝑗 ∕𝜇𝑇0 𝑖 𝜎𝑇0 𝑖 . The results are reported in Table 7 for the
sizable, regardless of whether a country is rich or poor, and hot or cold. RCP 8.5 scenario and 𝑚 = 30. As expected, the estimated GDP per
Nonetheless, the losses vary significantly across countries depending on capita losses become significantly larger, almost doubling at the global
the country-specific projected paths of temperatures. Fig. 7 shows that level by 2100 to 13.11 percent. For the United States, the losses are
if we managed to limit the increase in average global temperatures likely to be 70 percent higher compared to the baseline counterfactual
to 0.01 ◦ C per annum (the RCP 2.6 scenario), in line with the Paris scenario reported in Table 6. In terms of the channels of impact,
Agreement objective, we would be able to substantially reduce these the increase in the degree of climate variability affects economies by
losses. reducing labour productivity, increasing health problems (e.g., heat-
Table 6 reports the real GDP per capita losses for China, the Eu- related health issues or drought-related water and food shortages),
ropean Union, India, Russia, and the United States, over various time damaging infrastructure (e.g., from flooding in river basins and coasts
horizons. As in Fig. 6, income effects are substantially larger under an and landslides), and disruptions in supply chains—see IPCC (2014) for
unmitigated path (i.e., RCP 8.5). Nonetheless, under both scenarios, details.
the cross-country heterogeneity is significant. Focusing on the RCP 8.5
scenario (with 𝑚 = 30) we observe that the losses vary between 0.50 4. Concluding remarks
and 1.20 percent, 1.53 and 3.77 percent, and 4.35 and 10.52 percent
in 2030, 2050 and 2100, respectively; with a relatively large impact Using data on 174 countries over the period 1960 to 2014, and
estimated for the United States in 2100 (reflecting IPCC’s projections a novel econometric strategy (that differentiates between level and
of a sharp increase in the country’s average temperature in the absence growth effects including over the long term; accounts for bi-directional
of mitigation efforts). feedbacks between economic growth and climate change; considers

10
M.E. Kahn et al. Energy Economics 104 (2021) 105624

( )
Fig. 6. Percent Loss in GDP per capita by 2100 in the Absence of Climate Change Policies (RCP 8.5 Scenario). Notes: The heat map shows 𝛥𝑖ℎ 𝑑𝑖 , see Eq. (13), in year 2100
with 𝑚 = 30, based on the RCP 8.5 scenario. The Mercator projection exaggerates areas far from the equator.

( )
Fig. 7. Percent loss in GDP per capita by 2100 Abiding by the Paris Agreement Objective (RCP 2.6 Scenario). Notes: The heat map shows 𝛥𝑖ℎ 𝑑𝑖 , see Eq. (13), in year 2100
with 𝑚 = 30, based on the RCP 2.6 scenario. The Mercator projection exaggerates areas far from the equator.

asymmetric weather effects; allows for nonlinearity and an implicit in Paris in December 2015, will reduce global income by 1 percent by
model of adaptation; and deals with temperature being trended), we 2100. However, an increase in average global temperatures of 0.04 ◦ C
showed that persistent changes in temperature above time-varying (corresponding to the RCP 8.5 scenario, which assumes higher green-
norms has long-term negative impacts on economic growth. If tem- house gas emissions in absence of climate change mitigation policies)
perature deviates from its historical norm by 0.01 ◦ C annually for an reduces world’s real GDP per capita by 7 percent by 2100, with the
extended period of time, long-term income growth will be lower by size of these income effects varying significantly across countries de-
0.0543 percentage points per year. Furthermore, we illustrated that pending on the pace of temperature increases and variability of climate
these negative long-run growth effects are prevalent in all countries but
conditions in each country. The estimated global per capita GDP losses
to different degrees across climates and income groups. In particular,
under a high-emissions scenario with no policy action (that is RCP 8.5)
our heterogeneous panel data estimates suggested a lower marginal
would almost double if country-specific climate variability were to rise
weather effects in cold and/or rich countries (i.e., slope coefficients
commensurate to temperature increases in each country (with global
were smaller). Nevertheless, we find that income losses are sizable
even in cold climates either because they are warming up much faster income losses amounting to 13 percent by 2100). Overall, abiding
than temperate or hot regions or climate variability is becoming more by the Paris Agreement objective would go a long way in limiting
pronounced in line with faster temperature increases. economic losses from climate change across almost all countries. We
We performed a number of counterfactual exercises where we in- also illustrated that while adaptation to climate change could reduce
vestigated the output effects of annual increases in temperatures under these negative long-run growth effects, it is highly unlikely to offset
mitigated and unmitigated scenarios during 2015–2100. We showed them entirely. Therefore, our findings call for more forceful policy
that keeping the increase in the global average temperature to below responses to the threat of climate change, including more ambitious
2 degrees Celsius above pre-industrial levels as agreed by 190 parties mitigation and adaptation efforts.

11
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Table 6
Percent loss in GDP per capita by 2030, 2050, and 2100 under the RCP 2.6 and RCP 8.5 scenarios.
Year 2030 (ℎ = 16) Year 2050 (ℎ = 36) Year 2100 (ℎ = 86)
𝑚 = 20 𝑚 = 30 𝑚 = 40 𝑚 = 20 𝑚 = 30 𝑚 = 40 𝑚 = 20 𝑚 = 30 𝑚 = 40
World
RCP 2.6 −0.01 −0.01 −0.02 0.06 0.11 0.16 0.58 1.07 1.57
RCP 8.5 0.40 0.80 1.25 1.39 2.51 3.67 4.44 7.22 9.96
China
RCP 2.6 −0.22 −0.45 −0.71 −0.38 −0.80 −1.31 0.24 0.45 0.67
RCP 8.5 0.31 0.58 0.87 0.90 1.62 2.30 2.67 4.35 5.93
European Union
RCP 2.6 −0.04 −0.08 −0.13 −0.06 −0.13 −0.22 0.05 0.09 0.13
RCP 8.5 0.24 0.50 0.80 0.79 1.53 2.35 2.67 4.66 6.69
India
RCP 2.6 0.12 0.26 0.42 0.41 0.81 1.27 1.44 2.57 3.69
RCP 8.5 0.60 1.16 1.78 2.13 3.62 5.08 6.37 9.90 13.39
Russia
RCP 2.6 −0.07 −0.14 −0.23 −0.16 −0.34 −0.56 −0.33 −0.71 −1.19
RCP 8.5 0.51 1.03 1.63 1.62 3.08 4.61 5.28 8.93 12.46
United States
RCP 2.6 0.10 0.20 0.33 0.29 0.60 0.96 0.98 1.88 2.84
RCP 8.5 0.60 1.20 1.86 2.13 3.77 5.39 6.66 10.52 14.32

Notes: We consider persistent increases in temperatures based on the RCP 2.6 and RCP 8.5 scenarios. Numbers are PPP GDP
( )
weighted averages of 𝛥𝑖ℎ 𝑑𝑖 , see Eq. (13), with ℎ = 16, 36 , and 86 (corresponding to the year 2030, 2050, and 2100,
respectively) and 𝑚 = 20, 30, and 40.

Table 7 Chudik, A., Mohaddes, K., Pesaran, M.H., Raissi, M., 2016. Long-run effects in large
Percent loss in GDP per capita by 2030, 2050, and 2100 under the RCP 8.5 scenario: heterogeneous panel data models with cross-sectionally correlated errors. In:
the role of climate variability. Hill, R.C., Gonzalez-Rivera, G., Lee, T.-H. (Eds.), Advances in Econometrics (Volume
Year 2030 Year 2050 Year 2100 36): Essays in Honor of Aman Ullah. Emerald Publishing, pp. 85–135 (Chapter 4).
(ℎ = 16) (ℎ = 36) (ℎ = 86) Chudik, A., Mohaddes, K., Pesaran, M.H., Raissi, M., 2017. Is there a debt-threshold
effect on output growth? Rev. Econ. Stat. 99 (1), 135–150.
World 2.02 5.18 13.11
Chudik, A., Pesaran, M.H., Yang, J.-C., 2018. Half-panel jackknife fixed effects esti-
China 0.78 1.99 5.02
mation of panels with weakly exogenous regressors. J. Appl. Econometrics 33 (6),
European Union 1.45 3.71 9.37
816–836.
India 2.62 6.70 16.92
CIESIN, 2016. Gridded Population of the World, Version 4 (GPWv4): Population
Russia 2.00 5.13 12.94
Density. Columbia University, NASA Socioeconomic Data and Applications Center
United States 2.66 6.81 17.19
(SEDAC), Palisades, NY.
Notes: We consider persistent increases in temperatures based on the RCP 8.5 scenario Dell, M., Jones, B.F., Olken, B.A., 2009. Temperature and income: Reconciling new
( ) ( )
but set 𝜎𝑇1 𝑖 = 𝜇𝑇1 𝑖,𝑗 ∕𝜇𝑇0 𝑖 𝜎𝑇0 𝑖 . Numbers are PPP GDP weighted averages of 𝛥𝑖ℎ 𝑑𝑖 , with cross-sectional and panel estimates. Amer. Econ. Rev. 99 (2), 198–204.
Dell, M., Jones, B.F., Olken, B.A., 2012. Temperature shocks and economic growth:
ℎ = 16, 36, and 86 (corresponding to the year 2030, 2050, and 2100, respectively) and
Evidence from the last half century. Am. Econ. J. Macroecon. 4 (3), 66–95.
𝑚 = 30.
Dell, M., Jones, B.F., Olken, B.A., 2014. What do we learn from the weather? The new
climate-economy literature. J. Econ. Lit. 52 (3), 740–798.
Deryugina, T., Hsiang, S.M., Does the Environment Still Matter? Daily Temperature and
Income in the United States, NBER Working Paper No. 20750.
Data availability
Dhyne, E., Fuss, C., Pesaran, M.H., Sevestre, P., 2011. Lumpy price adjustments: A
microeconometric analysis. J. Bus. Econom. Statist. 29 (4), 529–540.
Mendeley data: http://dx.doi.org/10.17632/hytzz8wftw Gallup, J.L., Sachs, J.D., Mellinger, A.D., 1999. Geography and economic development.
Int. Reg. Sci. Rev. 22 (2), 179–232.
Heal, G., Park, J., 2016. Reflections - Temperature stress and the direct impact of
Appendix A. Theory (A1), relation to literature (A2), temperature climate change: A review of an emerging literature. Rev. Environ. Econ. Policy 10
trends (A3), and individual country results (A4) (2), 347–362.
Henseler, M., Schumacher, I., 2019. The impact of weather on economic growth and
its production factors. Clim. Change 154 (3), 417–433.
Supplementary material related to this article can be found online Hsiang, S.M., 2016. Climate econometrics. Annu. Rev. Resour. Econ. 8 (1).
at https://doi.org/10.1016/j.eneco.2021.105624. International Monetary Fund, 2017. The effects of weather shocks on economic
activity: How can low-income countries cope? In: World Economic Outlook. pp.
117–183 (Chapter 3).
References IPCC, 2014. Climate Change 2014: Impacts, Adaptation and Vulnerability, Contribution
of Working Group II to the Fifth Assessment Report of the Intergovernmental Panel
on Climate Change, Volume 1. Cambridge University Press, Cambridge.
Agrawala, S., Carraro, M., Kingsmill, N., Lanzi, E., Mullan, M., Prudent-Richard, G.,
IPCC, 2021. Climate Change 2021: The Physical Science Basis. Contribution of Working
Private Sector Engagement in Adaptation to Climate Change: Approaches to
Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate
Managing Climate Risks, OECD Environment Working Papers No 39.
Change. Cambridge University Press, Cambridge.
Arguez, A., Durre, I., Applequist, S., Vose, R.S., Squires, M.F., Yin, X., Heim, R.R.,
Jones, B.F., Olken, B.A., 2010. Climate shocks and exports. Amer. Econ. Rev. 100 (2),
Owen, T.W., 2012. Noaa’s 1981-2010 U.S. climate normals: An overview. Bull.
454–459.
Am. Meteorol. Soc. 93 (11), 1687–1697.
Kalkuhl, M., Wenz, L., 2020. The impact of climate conditions on economic production.
Burke, M., Hsiang, S.M., Miguel, E., 2015. Global non-linear effect of temperature on evidence from a global panel of regions. J. Environ. Econ. Manag. 103, 102360.
economic production. Nature 527, 235–239. Letta, M., Tol, R.S.J., 2019. Weather, climate and total factor productivity. Environ.
Cashin, P., Mohaddes, K., Raissi, M., 2017. Fair weather or foul? The macroeconomic Resour. Econ. 73, 283–305.
effects of El Niño. J. Int. Econ. 106, 37–54. Matsuura, K., Willmott, C.J., 2015. Terrestrial air temperature and precipitation:
Chudik, A., Mohaddes, K., Pesaran, M.H., Raissi, M., 2013. Debt, Inflation and Growth: Monthly and annual time series (1900 - 2014), v 4.01.
Robust Estimation of Long-Run Effects in Dynamic Panel Data Models. Federal Mejia, S.A., Mrkaic, M., Novta, N., Pugacheva, E., Topalova, P., 2018. The Ef-
Reserve Bank of Dallas, Globalization and Monetary Policy Institute Working Paper fects of Weather Shocks on Economic Activity: What are the Channels of
No. 162. Impact? International Monetary Fund, IMF Working Paper WP/18/144.

12
M.E. Kahn et al. Energy Economics 104 (2021) 105624

Mendelsohn, R., 2016. Measuring weather impacts using panel data. Sachs, J.D., Warner, A.M., 1997. Sources of slow growth in African economies. J. Afr.
Newell, R.G., Prest, B.C., Sexton, S.E., 2021. The GDP-temperature relationship: Econ. 6 (3), 335–376.
Implications for climate change damages. J. Environ. Econ. Manag. 108, 102445. Schlenker, W., Auffhammer, M., 2018. The cost of a warming climate. Nature 557
Nickell, S., 1981. Biases in dynamic models with fixed effects. Econometrica 49 (6), (News & Views Forum), 498–499.
1417–1426. Stern, N., 2007. The Economics of Climate Change: The Stern Review. Cambridge
Nordhaus, W.D., 1992. The ‘‘DICE’’ Model: Background and Structure of a Dynamic University Press, Cambridge.
Integrated Climate-Economy Model of the Economics of Global Warming. Technical Tol, R.S.J., 2009. The economic effects of climate change. J. Econ. Perspect. 23 (2),
Report 1009, Cowles Foundation Discussion Paper No. 1009. 29–51.
Nordhaus, W.D., 2006. Geography and macroeconomics: New data and new findings. Tol, R.S.J., 2014. Correction and update: The economic effects of climate change. J.
Proc. Natl. Acad. Sci. USA 103 (10), 3510–3517. Econ. Perspect. 28 (2), 221–226.
Obama, B., 2017. The irreversible momentum of clean energy. Science. Tol, R.S.J., 2017. Population and trends in the global mean temperature. Atmósfera 30
Pesaran, M.H., 1997. The role of economic theory in modelling the long run. Econom. (2), 121–135.
J. 107, 178–191. Tol, R.S.J., 2018. The economic impacts of climate change. Rev. Environ. Econ. Policy
Pesaran, M.H., Shin, Y., 1999. An autoregressive distributed lag modelling approach 12 (1), 4–25.
to cointegration analysis. In: Strom, S. (Ed.), Econometrics and Economic Theory Tol, R.S.J., The Economic Impact of Weather and Climate, CESifo Working Paper No.
in 20th Century: The Ragnar Frisch Centennial Symposium. Cambridge University 8946.
Press, Cambridge, pp. 371–413 (Chapter 11). Vose, R.S., Applequist, S., Squires, M., Durre, I., Menne, M.J., Williams, C.N., Fen-
Pesaran, M.H., Smith, R., 1995. Estimating long-run relationships from dynamic imore, C., Gleason, K., Arndt, D., 2014. Improved historical temperature and
heterogeneous panels. J. Econometrics 68 (1), 79–113. precipitation time series for U.S. climate divisions. J. Appl. Meteorol. Climatol.
53 (5), 1232–1251.

13