LETTER • OPEN ACCESS You may also like

- Economic growth dominates rising

Persistent effect of temperature on GDP identified potential flood risk in the Yangtze River
and benefits of raising dikes from 1991 to
from lower frequency temperature variability 2015
Wei Qi, Lian Feng, Hong Yang et al.

- Spatiotemporal changes in both asset

To cite this article: B A Bastien-Olvera et al 2022 Environ. Res. Lett. 17 084038 value and GDP associated with seismic
exposure in China in the context of rapid
economic growth from 1990 to 2010
Jidong Wu, Cailin Wang, Xin He et al.

- A systematic review of the evidence on

View the article online for updates and enhancements. decoupling of GDP, resource use and
GHG emissions, part II: synthesizing the
insights
Helmut Haberl, Dominik Wiedenhofer,
Doris Virág et al.

This content was downloaded from IP address 88.230.8.175 on 04/10/2022 at 21:33

 Environ. Res. Lett. 17 (2022) 084038 https://doi.org/10.1088/1748-9326/ac82c2

LETTER

Persistent effect of temperature on GDP identified from lower

OPEN ACCESS
frequency temperature variability
RECEIVED
2 May 2022 B A Bastien-Olvera1,∗, F Granella2,4 and F C Moore3
REVISED 1
8 July 2022
Geography Graduate Group, University of California Davis, Davis, CA, United States of America
2
Bocconi University, Milan, Italy
ACCEPTED FOR PUBLICATION 3
Department of Environmental Science and Policy, University of California Davis, Davis, CA, United States of America
20 July 2022 4
RFF-CMCC European Institute on Economics and the Environment, Milan, Italy
∗
PUBLISHED Author to whom any correspondence should be addressed.
10 August 2022
E-mail: [email protected]
Original content from Keywords: climate change, climate economics, economic growth, temperature variability
this work may be used
under the terms of the Supplementary material for this article is available online
Creative Commons
Attribution 4.0 licence.
Any further distribution
of this work must Abstract
maintain attribution to
the author(s) and the title It is well established that temperature variability affects a range of outcomes relevant to human
of the work, journal
citation and DOI.
welfare, including health, emotion and mood, and productivity across a number of economic
sectors. However, a critical and still unresolved empirical question is whether temperature variation
has a long-lasting effect on economic productivity and, therefore, whether damages compound
over time in response to long-lived changes in temperature expected with climate change. Several
studies have identified a relationship between temperature and gross domestic product (GDP), but
empirical evidence as to the persistence of these effects is still weak. This paper presents a novel
approach to isolate the persistent component of temperature effects on output using lower
frequency temperature variation. The effects are heterogeneous across countries but collectively,
using three different GDP datasets, we find evidence of persistent effects, implying temperature
affects the determinants of economic growth, not just economic productivity. This, in turn, means
that the aggregate effects of climate change on GDP may be far larger and far more uncertain than
currently represented in integrated assessment models used to calculate the social cost of carbon.

1. Introduction when temperature variations affect the productiv-

ity of labor or capital) with a few recent exceptions
A large body of evidence now exists showing a rela- [7–14].
tionship between temperature fluctuations and eco- Despite its importance for determining the
nomic productivity [1–4]. Temperature has been aggregate costs of climate change, evidence on the
shown to influence output at global [1, 2], national persistence of the impacts of temperature shocks is
[5, 6], and regional scales [3], affecting a wide range of sparse and contradictory [15]. Dell et al [2] show that
sectors in both high-income and low-income coun- persistent and non-persistent effects can produce
tries. The persistence of these impacts has first- identical contemporaneous effects on the growth rate
order implications for the magnitude of climate but can be distinguished using lagged temperature
change damages: if temperature fluctuations affect effects. Using global national accounts data, they
the determinants of economic growth (e.g. depreci- fit a reduced-form model with lagged temperature
ation of capital or the total factor productivity growth terms and find evidence that effects of temperat-
rate), then they have a persistent impact on the level ure shocks in poorer countries do not revert within
of economic output. In this case, climate change dam- ten years, implying large negative effects of higher
ages are cumulative and may be orders of magnitude temperatures on economic growth, at least in the
larger than currently represented in models used for medium-term. Burke et al [1] use a similar dataset
the cost–benefit analysis of climate change, which to find robust evidence for a non-linear, hill-shaped
mostly assume non-persistent damages (for example, relationship between contemporaneous temperature

© 2022 The Author(s). Published by IOP Publishing Ltd

Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

and GDP growth. However, evidence for persistent some of these fluctuations, such as El Nino South-
impacts on the economy is weaker since the sum of ern Oscillation with a period of two to seven years,
lagged effects has large standard errors with confid- are well understood [21], spectral analysis of atmo-
ence intervals that include both zero and very large spheric time series reveal fluctuations at all possible
negative effects. In a model-selection exercise based frequencies [22, 23]. Figure 1(a) shows this variability
on cross-validation, Newell et al [16] show that total in the US temperature time series between 1960 and
climate damages are highly sensitive to the question 2017 [24, 25]. We use a low-pass filter to successively
of persistence and to the functional form of empir- remove high-frequency variation and obtain temper-
ical models used to estimate effects, but also find that ature time series that preserve only lower-frequency
out-of-sample cross-validation tests are insufficiently oscillations.
powerful to disambiguate between alternate mod- Temperature variability at different timescales
els of impact persistence. At a smaller spatial scale, will produce distinct economic dynamics depend-
Deryugina and Hsiang [5] found evidence of per- ing on the persistence of economic impacts. This is
sistent but declining effects during the first ten years illustrated in figure 1(b), which shows the change
after a temperature shock in individual U.S. counties. in GDP growth and GDP level expected under tem-
Deryugina and Hsiang [5], and Colacito et al [17] perature shocks of different durations and alternate
found that increases in summer and fall temperat- models of economic impact. Dell et al [2] derive a
ure could have persistent effects on the gross state simple equation for a model that includes both non-
product of U.S. states. persistent level effects (β) and persistent growth effects
A major empirical challenge is that estimating the (γ), given baseline growth rate g:
sum of lagged effects, particularly for a non-linear
function, can produce large standard errors and high gt = g + γTt + β∆Tt (1)
uncertainty. For instance, in the quadratic specifica-
where Tt is the deviation in temperature from some
tion used by Burke et al, identifying cumulative effects
mean value in period t and ∆Tt is the change in tem-
over ten years requires estimating and summing 20
perature between period t and t − 14 . Although it
regression coefficients [1]. The uncertainty around
is likely that some economies experience both levels
this statistic depends on the variance and covariance
and growth effects simultaneously, we use two styl-
of all 20 parameter estimates. More recent empir-
ized cases in figure 1(b) to illustrate how the timescale
ical investigations of climate impacts on economic
of temperature variation interacts with the models of
growth have focused on resolving detail at the sub-
economic impact. In the pure ‘level effects’ model we
national scale [3, 17, 18], or on resolving impacts on
set the growth effect to zero (i.e. γ = 0) so that:
the production process [19]. While they suggest some
persistence in temperature effects, the uncertainty gt = g + β∆Tt . (2)
around this key question relevant to understanding
the aggregate costs of climate change remains largely In the ‘growth effects’ model, we set the level effect
unresolved. to zero (i.e. β = 0) so that:

2. Methods gt = g + γTt . (3)

A one-year temperature shock equally reduces

Here, we propose a statistical test to identify the pres-
GDP in a level effects model and in a growth effects
ence of persistent effects of temperature on output
model (figure 1(b), left column). However, when
using lower-frequency temperature variation. We first
temperature returns to the baseline so does GDP
use a simulation exercise to demonstrate the power
in the level model, but not in the growth model
of the test to discriminate between cases with and
(figure 1(b), bottom-left panel). The two models thus
without persistent effects of temperature. Second,
produce distinct long-term effects on GDP: growth
we implement this test on individual country-level
effects on GDP keep accumulating as the duration of
temperature and economic growth time-series. The
the temperature excursion increases, but level effects
test complements previous approaches that have used
disappear when temperature returns to its baseline. It
either lagged temperatures or out-of-sample tests to
is this effect of past temperature shocks on the future
attempt to resolve the question of impact persistence
level of GDP, occurring because temperature affects
but which, as described above, have mostly produced
the determinants of economic growth, that we refer
ambiguous results.
to in this manuscript as ‘persistent’ impacts.
The essence of the approach is that persistent and
transient impacts on economic output can be distin-
guished using temperature variation occurring at dif- 4 We inherit the taxonomy of ‘levels effects’ and ‘growth effects’

ferent frequencies. Internal variability of the climate from [2]. While we focus on GDP growth, the terms originate in
reference to effects on GDP. A level effect alters the level of GDP,
system gives rise to oscillations at different times-
and when temperature reverts to the baseline so does production.
cales. This is an intrinsic characteristic of non-linear A growth effect alters the growth rate, thus its effects are cumulative
dynamic systems like the Earth’s climate [20]. While and persistent.

2
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

Figure 1. Temperature fluctuations (demeaned and detrended) and their effects on GDP. Panel (A). US population-weighted
temperature fluctuations after detrending and filtering higher-frequency variation [24]. The top blue line shows the US
temperature time series. Lower lines show the filtered time series, progressively expanding the periodicity range of the filter from
2 years up to 3, 5, 10, and 15 years. The time series are spread across the y-axis for visual purposes; all time series oscillate around
zero because they were demeaned and detrended before filtering. Panel (B). Upper panel: temperature shocks at decreasing
frequencies. Mid panel: effects of those shocks on GDP growth under levels and growth models. Lower panel: effects of
temperature shocks on GDP.

Note that the effects illustrated in figure 1(b) do particularly in response to temperature shocks of dif-
not include any variation in the impact of temper- ferent lengths.
ature shocks as a function of the shock duration. The duration of temperature excursions from a
The question of whether longer-period temperature mean value is key to identifying the presence of
excursions, more analogous to the type of perman- growth effects (figure 1(b) top row). The correlation
ent warming expected from climate change, produce between temperature and GDP growth in a growth
either larger (via compounding effects and intensi- effects model does not depend on the duration of
fication) or smaller (via adaptation) impacts com- the temperature anomaly, but breaks down in a level
pared to shorter temperature shocks has been widely effects model as the length of the excursion grows
debated [26–30]. The question of persistence— (figure 1(b), middle row). This happens because there
whether the level of GDP is affected by past tem- are no level effects if temperature is constant but away
perature shocks—is distinct from this issue however. from the baseline.
The distinction between persistent vs non-persistent Therefore it should be possible, in principle, to
impacts arises because of how temperature affects detect the presence of persistent effects in empirical
the economy; non-persistent effects arise through data using different timescales of temperature vari-
temporary effects on productivity (crop yield losses ability. It is a common practice in signal processing
from extreme heat are one example) whereas persist- problems to decompose time series into a sum of
ent effects arise from impacts on factors that have a periodic components with varying frequencies, amp-
long-lived effect on economic production (destruc- litudes and phases [31], widely used in a variety of
tion of capital in extreme events for instance). Adapt- fields like audio processing, electrical engineering,
ation or intensification would somewhat alter the and climate science [32–34]. This approach allows
shape of the responses shown in the right column the time-series to be reconstructed using a specific
of figure 1(b), but the levels and growth models subsets of desired frequencies. A low-pass filter is
would still produce qualitatively different dynamics, a version of the time series that only preserves low

3
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

Figure 2. Simulation exercise demonstrating the divergence of regression results with increasing frequency filters under two
alternate models of temperature impacts on economic production, a non-persistent ‘level only’ model (purple) and a fully
persistent ‘growth only’ model (pink). A third semi-transparent pink line shows a combined model with opposite signs of growth
and level effects.

frequency components. Following studies in the cli- adding an independent and identically distributed
mate literature [35], we use a low-pass filter to remove noise component. We regress the simulated economic
inter-annual variations and obtain temperature time growth data on temperature after filtering out vary-
series that preserve only lower-frequency oscillations. ing ranges of frequencies from the temperature time
If changes in temperature do not influence the under- series, and adjusting the regression estimate to avoid
lying determinants of growth (levels only model), a small bias introduced by the changing amplitude
the estimated effect of low-frequency temperature of temperature variations at lower frequency filters
anomalies on GDP growth should converge towards (Supplementary material 2).
zero from the estimated effect of unfiltered temper- Figure 2 shows the mean value of the estimated
ature data. In contrast, if changes in temperature coefficients and its confidence interval for all the sim-
alter the determinants of growth (presence of persist- ulations. Without any filtering using only contempor-
ent effects), the correlation between temperature and aneous temperatures, growth and level impacts are
GDP growth should be detectable after the temperat- indistinguishable, as originally pointed out by Dell
ure data is filtered. et al [2]. But filtering out high frequencies in the tem-
Figure 2 demonstrates this effect in a simulation perature data produces divergent effects: the estim-
exercise. It shows results from time series regressions ated effect under the growth only model remains
of simulated economic growth on simulated temper- detectable while the coefficients in the level model
ature at different levels of filtering under two stylized attenuates markedly. In other words, the different
cases—one in which there are only non-persistent patterns in figure 2 mean that these two possible
damages (i.e. the level effects-only model, purple line) worlds—one with and one without persistent tem-
and one with only persistent damages (i.e. the growth perature impacts—could potentially be distinguished
effects-only model, pink line), following equations using this method. In essence, a statistical test on the
(2) and (3) respectively. Additionally, to illustrate coefficient for the filtered data is a test for the presence
one of the many possible combinations, another of growth effects, and is independent of the presence,
semi-transparent line shows a simulation with mixed or sign, of level effects.
growth and level effects with opposite signs. While previous literature used lagged temperature
The random temperature time series used in the estimates to test for growth effects, we show through a
simulations preserve the frequency distribution of simulation that using a low-pass filter is more efficient
the Earth’s natural oscillations by matching the spec- in distinguishing between levels and growth effects at
tral decomposition on 1500 years of pre-industrial the medium to long term in a context where data is
global temperatures based on the Last Millennium limited to 70 years. Supplementary figure 1 compares
Reanalysis [36]. Using this decomposition we gener- the coefficients estimated with the filtering approach
ate 10 000 random 350 year temperature time series (left panel) and the sum of the lagged coefficients
that preserve this frequency distribution but with ran- for a full distributed lag model (middle) and a more
dom phase shifts [37] and then simulate economic parsimonious version that reduces the number of
dynamics for each temperature time series under the estimated coefficients by imposing smoothness on the
two alternate impacts models using equations (2) lag structure (right). The distributed lag model is as
and (3), and the combined effect using equation (1), powerful at distinguishing levels from growth effects

4
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

when the number of lags and the length of filtering 3. Results and discussion
are small. However, filtering grows more efficient for
greater number of lags and longer filters, as the dis- The behavior of the estimates θ̂f for each country con-
tributed lag model becomes increasingly noisy. This tains information about the persistence of temperat-
suggests that the low-pass filtering test can be a help- ure effects on the economy. In particular, non-zero
ful complement to existing approaches using lagged low-frequency estimates signal presence of growth
temperature in investigating the persistence of effects effects, as shown in the simulations (figure 2). We find
over the medium to long run in data scarce contexts. that 39 countries have low-frequency estimates that
We use our test to investigate the persistence of are statistically different from zero at the 90% con-
temperature effects on economic production. We use fidence level (of which 18 might be expected as false
GDP data from the World Bank covering 217 coun- positives given the number of comparisons). Further,
tries from 1961 to 2017 [38], merging this dataset looking across all countries there is not strong evid-
with population-weighted temperature and rainfall ence for systematic trends in coefficients towards zero
data from University of Delaware [24, 25]. To identify at lower frequency variation, as would be expected if
whether country-level temperature impacts have per- impacts operated only through non-persistent level
sistent effects we performed the following regression effects.
for each country and length of filter: Figure 3(A) shows the values of θ̂f for all coun-
tries at different levels of filtering, binned into two
gt = θf Tt,f + πf Pt,f + ϵt (4) broad categories: a converging-towards-zero effect
(blues), where the absolute value of θ̂f decreases at
where Tt,f and Pt,f are the population-weighted tem- lower frequencies (as expected by the presence of level
perature and rainfall in year t after demeaning, effects only, or by the combination of a level effect
detrending, and filtering out frequencies higher than and a smaller growth effect), and a not-converging-
f. The filters f are low-pass filters that filter-out any towards-zero effect (oranges), where the absolute
oscillations with periods shorter than 3, 5, 10, and value of θ̂f increases at lower frequencies (explained
15 years, or f = unfiltered when no filter was applied. only by the presence of persistent effects). In addi-
The low-pass filter algorithm requires data that spans tion, there is a third category we describe as ‘unclas-
at least twice the upper bound periodicity, which res- sified’ (grey) where the absolute value of θ̂f increases
ults in some countries not having estimates for all the but changes sign between the unfiltered and the most
levels of filtering due to missing data at earlier time filtered estimates. This behavior could be explained
periods. Country-specific quadratic time trends are by levels and growth effects of opposite signs; yet,
removed from all variables (growth, temperature and these countries are conservatively not classified as
rainfall) prior to analysis to address concerns of non- either converging or not converging. Within the two
stationarity in the weather and economic time-series. groups of converging and not converging countries,
Excluding rainfall from equation (4) would bias the we further identify subsets of countries where the
estimate of θ, since rainfall is known to correlate with filtered estimates are either statistically larger (i.e.
both GDP growth [39] and temperature [40]. How- intensifying; dark orange) or smaller (converging;
ever, we restrict the analysis to temperature and leave dark blue) from the unfiltered estimates.
the discussion of results on precipitation to the sup- Figure 3(B) tracks features of countries’ estim-
plementary material. ates that are key to detect the presence of growth and
Given the lack of strong prior empirical evidence levels effects. The left column divides countries based
for the persistence of temperature effects, or strong on the statistical significance of the unfiltered estim-
theoretical or empirical evidence regarding drivers of ate, the middle column shows the statistical signific-
heterogeneity in the response, the analysis focuses at ance of the countries’ most filtered estimate, and the
the country level to give more flexibility and allow right column shows whether estimates show conver-
estimates to differ across countries. On the other ging, not converging, or intensifying effects. Among
hand, this comes at the cost of larger statistical uncer- the 27 countries whose unfiltered estimate is statist-
tainty. We analyze the evidence for persistence across ically different from zero, the 15 year filtered coef-
all countries at the global scale by separately pooling ficients of 18 countries are not statistically different
the positive and negative estimates of θf and estimat- from zero, meaning only level effects were detected in
ing the following regression model those countries (purple lines in figure 3(B)). Presence
of growth effects (figure 3(B), pink lines) is detected
θ̂f,c = Ff + ϵf,c (5) in the remaining 9 countries and in 30 other countries
whose unfiltered estimate was not statistically differ-
where the value of the temperature coefficient estim- ent from zero.
ate in country c at filtering level f is regressed on a The middle column of figure 3(B) shows that in
vector of indicators of the level of filtering, clustering 18 countries where growth effects have been detected
standard errors at the continent level. the filtered estimates are statistically larger than the

5
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

Figure 3. Panel (A) Country-level estimates of the temperature effect on economic growth. For visualization purposes only, each
line connects the estimated coefficients from regressions at different levels of filtering of the temperature data. Lines are color
coded depending on the trend from the unfiltered to the most filtered estimate: orange when the absolute value of coefficients
increases with filtering (‘Not converging to zero’); dark orange when the difference between unfiltered and most filtered is
significant at 10% (‘Intensifying’); blue when the absolute value of coefficients decreases with filtering (‘Converging to zero’), and
dark blue when the trend is statistically significant at 10% (‘∗ Converging to zero’; not found in this results); grey when the most
filtered estimate is larger than the unfiltered but with opposite sign. The graph only shows countries with estimates below the 99th
percentile for readability. Panel (B) The left-hand side of the chart displays the number of countries for which there is evidence of
growth effects, in pink, and evidence of level effects, in purple. The right-hand side classifies 15 year filtered estimates by the type
of trend using the same color code as panel (A).

unfiltered estimates (i.e. ‘intensifying effect’), a pat- as 1790–2009, developed to examine the persistence
tern that is consistent with level and growth effects of macroeconomic shocks [41, 42]. Secondly, we
of opposite sign. Among the remaining 137 coun- use the Maddison Project database that standardizes
tries that do not attain conventional statistical signi- country-level GDP per capita for 170 countries for
ficance of the most filtered estimate, more countries several centuries [43]. Due to the sparsity of temper-
have non-converging estimates (n = 65, orange lines) ature and rainfall records pre-1900, we use only post-
than converging estimates (n = 27, blue lines). There 1900 data for both datasets. Supplementary figure 4
is no country where the filtered estimate is signific- replicates figure 3 for these two alternate datasets cov-
antly smaller than the unfiltered estimate. ering different subsets of countries and much longer
We performed the same analysis using two altern- time-periods than the World Bank data. We again fail
ative economic growth datasets that span a longer to find strong evidence that estimates systematically
time period but include fewer countries. Firstly we converge towards zero using lower frequency vari-
used the Barro-Ursua dataset, with annual data on ation, as would be expected if impacts to the economy
economic growth of 43 countries starting as early operated only through non-persistent levels effects.

6
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

Figure 4. Pooled estimates of countries with positive and negative unfiltered coefficients across different levels of filtering using
three alternative datasets.

Pooling estimates from all countries, we are able 4. Discussion

to evaluate evidence, at the global level, for conver-
ging estimates at lower frequency filters. We thus The question of the persistence of climate damages is
estimate equation (5). Where the temperature coef- a first-order problem for climate change economics.
ficient estimate θ̂f,c in country c at filtering level f Studies that allow climate change to affect the determ-
is regressed on a vector with the levels of filtering inants of economic growth tend to produce far lar-
F, clustering standard errors at the continent level ger aggregate climate change costs than studies that
to allow for cross-country correlation and weight- impose only level effects on production [11, 13–15].
ing the observations by the inverse standard error. In response to the permanent shifts in temperat-
Patterns such as divergence or convergence towards ure expected with climate change, persistent impacts
zero as filtering increases would cancel out if, as it operate via effects on the growth rate compound
shown in figure 3, upper panel, there are both positive over time, producing far larger aggregate damages
and negative effects. We therefore perform the ana- over the long time frames relevant for assessing cli-
lysis separately for countries with positive and neg- mate change costs. Yet, impacts have been modeled
ative unfiltered estimates. If only non-persistent level as non-persistent by the numerous integrated assess-
effects were present, we would expect to see the negat- ment studies that since the 1990s have calculated
ive (positive) estimates converging towards zero, res- climate damages and evaluated optimal climate
ulting in a positive (negative) coefficient estimate on policy.
the filtering variables F. In contrast with previous literature that mod-
Figure 4 shows the cumulative estimated effect els non-linear effects of temperature on growth, we
for each level of filtering, and shows that, across all analyze the temperature-growth relationship with
countries, we do not see evidence for this attenuating country-level regressions. The smaller temperature
effect. Instead, the regression results show evidence ranges allow us to accurately model the effects
of persistent effect where the average value estimated using a linear approximation (see supplementary
using lower frequency temperature variation is sim- material and supplementary figure 2). In addition,
ilar to the value estimated using unfiltered data (see instead of using high-frequency, year-to-year tem-
supplementary table 1). perature variation to estimate climate impacts on
Finally, supplementary figure 5 examines evid- the economy, here we use lower frequency vari-
ence for heterogeneity in the marginal effect of tem- ation. Our identification strategy focuses on the per-
perature between countries, specifically whether they sistent effect of temperature by adjusting for time
are associated with either per capita GDP or mean trends and country-specific dynamics (via demeaning
temperature. Using only estimates significantly differ- and detrending) but uses lower-frequency temperat-
ent from zero at the unfiltered and 15 year filter levels ure variability instead of lags to distinguish between
(i.e. countries for which evidence of persistent effects growth and levels effects. Using a low-pass filter
is strongest), we find some evidence that impacts are instead of lags avoids adding noise terms together
negatively correlated with countries’ mean temperat- that could prevent identifying medium run persistent
ure as found in previous studies [1], but no systematic effects (see supplementary material and supplement-
differences in the estimated effects between rich and ary figure 1).
poor countries (supplementary figure 8 shows a sim- Applying this test to three different datasets of
ilar pattern resulting from a distributed lag non-linear economic growth, we fail to find strong evidence of
model under a panel analysis). only non-persistent effects. There are two key pieces

7
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

of evidence. First, we found statistically significant raising questions about whether these can be extra-
persistent temperature impacts on economic growth polated to much larger levels of warming expected
in 22% (19%; 8%) of the countries using the World with climate change.
Bank (Maddison Project; Barro-Ursua) dataset. Sig- Finally, we note that our approach is not able
nificant effects in these regressions implies the persist- to distinguish between a levels effect that contin-
ence of temperature impacts at least over the 15 year ues compounding over the 15 year time-frame of
period of our lowest-frequency regressions. Secondly, our lowest-frequency estimates but then subsequently
we examine how regression estimates change using reverses, and a ‘pure’ growth effect in which there
lower frequency temperature variation. The lack of is no subsequent reversal. Differentiating these two
persistent effects, as posited by the vast majority types of effects is a question of what happens in time-
of integrated assessment studies estimating climate frames longer than 15 years, which is an inherently
damages, would imply convergence of these estimates difficult empirical question due to the relatively short
towards zero. But we fail to find evidence of such con- time span of data available. However, either interpret-
vergence. At the individual country level, only 15% ation of the filtered results (i.e. 15 years of continu-
(21%; 34%) of countries have effects that converge ously worsening levels effects followed by reversal or
towards zero. For many more countries, the estim- a fully persistent effect) implies persistence of dam-
ated effects either do not converge towards zero or ages over time periods longer than a decade. Either
intensify over time, an effect that could be due to interpretation would imply larger aggregate climate
adaptation or coping dynamics, competing growth damages than the standard approach to representing
and levels effects with different signs, or a reduction climate change costs in integrated assessment mod-
in attenuation bias with longer filter lengths (though els, which assumes no persistence or compounding
this effect is likely small, as described more fully effects.
in supplementary figure 6). Pooling evidence from While providing evidence of persistent impacts of
across all countries produces stable effect sizes with temperature shocks on growth, our framework does
lower frequency variation for all three datasets, at least not isolate the mechanisms by which they arise. Past
over the 10–15 year period. Therefore, the evidence studies have modeled persistent impacts as resulting
suggests a sensitivity of aggregate economic output from a slow-down in total factor productivity growth
to temperature shocks persisting over at least the 10– [11, 12], changes to the capital depreciation rate [11],
15 year time frame and a conspicuous absence of evid- or impacts to the stock of natural capital [44]. Other
ence for fully non-persistent levels impacts. studies leave the mechanism of growth rate impacts
Like previous work, we find both positive and unspecified [9, 13]. Letta and Tol [19] investigate this
negative effects of temperature on different countries. question and suggest impacts arise through effects
It should be remarked that decade-long temperat- on total factor productivity growth, but more work
ure excursions used to estimate the effects here are is needed to understand exactly how these impacts
very small in amplitude (the median amplitude for manifest.
15 year filtered temperature is 0.11 ◦ C). While figure 3 A consistent and unsurprising finding from past
shows the effect of 1 ◦ C increase in temperature, the work is that allowing for persistent damages, because
actual magnitude of temperature variation over this of their compounding nature, vastly increases the
time-scale is much smaller and it is an open question uncertainty in climate change impact projections. For
whether these effect sizes can be extrapolated to much instance, Newell et al [16] estimate confidence inter-
larger changes in temperature expected with climate vals on damage estimates that allow for growth-rate
change. effects orders of magnitude larger than those that
This highlights a fundamental empirical chal- restrict impacts to only the level of GDP. Similarly,
lenge in estimating the effects of climate change. in a recent modeling study, Kikstra et al [45] show
Climate change will produce large (∼2 ◦ C–4 ◦ C) that the persistence of economic damages is the most
and sustained changes in temperature. The histor- important parameter determining aggregate climate
ical record contains both large but short temperature change costs. Our findings do not show strong evid-
excursions and much smaller but longer temperature ence for the presence of only non-persistent impacts
variation. Previous papers [1, 2] have examined the and instead suggest compounding effects over at
effect of high frequency variation, raising the ques- least a decadal time frame. Therefore, restricting
tion of whether these estimates can be extrapolated to modeling of climate change damages to only non-
longer-lasting temperature changes (e.g. due to effects persistent levels effects likely greatly under-states both
of adaptation, compounding effects, or the dynamics the uncertainty and the downside risk associated with
of persistent vs transient economic impacts). Here we climate change.
instead focus on the opposite—lower-frequency but
much smaller variation (at least in the filtered estim- Data availability statement
ates). This gives more confidence that effects estim-
ated are representative of impacts of sustained tem- The code to replicate the analysis and figures is in:
perature change, at least over the medium run, while https://github.com/BerBastien/TempEffectGDP.

8
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

Acknowledgments [11] Moore F C and Diaz D B 2015 Temperature impacts on

economic growth warrant stringent mitigation policy Nat.
Clim. Change 5 127
The authors thank Marshall Burke, Jeffrey Shrader,
[12] Moyer E J, Woolley M D, Matteson N J, Glotter M J and
and Max Auffhammer for comments on an earlier Weisbach D A 2014 Climate impacts on economic growth as
version of this manuscript. Funding for this work was drivers of uncertainty in the social cost of carbon J. Legal
provided by NSF, Award Number 1924378 (FCM), Stud. 43 401–25
[13] Ricke K, Drouet L, Caldeira K and Tavoni M 2018
and by the EU H2020 research and innovation pro-
Country-level social cost of carbon Nat. Clim. Change
gramme under the Marie Skłodowska-Curie Grant 8 895
Agreement No. 681228, GEMCLIME (FG). [14] Estrada F, Tol R S J and Gay-García C 2015 The persistence of
shocks in GDP and the estimation of the potential economic
costs of climate change Environ. Model. Softw. 1 155–65
Author contributions [15] Piontek F, Drouet L, Emmerling J, Kompas T, Méjean A,
Otto C, Rising J, Soergel B, Taconet N and Tavoni M 2021
Conceptualization and supervision: F C M Integrated perspective on translating biophysical to
Methodology: B A B O, F C M, F G economic impacts of climate change Nat. Clim. Change
11 563–72
Coding: B A B O
[16] Newell R G, Prest B C and Sexton S E 2021 The
Investigation: B A B O, F C M, F G GDP-temperature relationship: implications for climate
Writing: B A B O, F C M, F G change damages J. Environ. Econ. Manage. 1 102445
[17] Colacito R, Hoffmann B and Phan T 2019 Temperature and
growth: a panel analysis of the United States J. Money Credit
Conflict of interest Bank 51 313–68
[18] Damania R, Desbureaux S and Zaveri E 2020 Does rainfall
matter for economic growth? Evidence from global
Authors declare that they have no competing sub-national data (1990–2014) J. Environ. Econ. Manage.
interests. 102 102335
[19] Letta M and Tol R S J 2019 Weather, climate and total factor
productivity Environ. Resour. Econ. 73 283–305
ORCID iDs [20] Lorenz E N 1963 Deterministic nonperiodic flow J. Atmos.
Sci. 20 130–41
[21] Imbers J, Lopez A, Huntingford C and Allen M R 2013
B A Bastien-Olvera  https://orcid.org/0000-0002-
Testing the robustness of the anthropogenic climate change
3734-4267 detection statements using different empirical models J.
F Granella  https://orcid.org/0000-0002-2349- Geophys. Res. 118 3192–9
0132 [22] Hasselmann K 1976 Stochastic climate models Part I. Theory
Tellus 28 473–85
F C Moore  https://orcid.org/0000-0003-3866-
[23] Mann M E, Steinman B A and Miller S K 2020 Absence of
9642 internal multidecadal and interdecadal oscillations in climate
model simulations Nat. Commun. 11 49
[24] Matsuura K and Willmott C J 2018 Terrestrial air
References temperature: 1900–2017 gridded monthly time series
(Newark, DE: University of Delaware) (available at: http://
[1] Burke M, Hsiang S M and Miguel E 2015 Global non-linear climategeogudeledu/∼climate/html_pages/Global2017/READ
effect of temperature on economic production Nature MEGlobalTsT2017.html)
527 235 [25] Matsuura K and Willmott C J 2018 Terrestrial Precipitation:
[2] Dell M, Jones B F and Olken B A 2012 Temperature shocks 1900–2017 Gridded Monthly Time Series (Newark, DE:
and economic growth: evidence from the last half century Electronic Department of Geography, University of
Am. Econ. J. 4 66–95 Delaware) p 19716
[3] Kalkuhl M and Wenz L 2020 The impact of climate [26] Moore F C and Lobell D B 2014 Adaptation potential of
conditions on economic production. Evidence from a global European agriculture in response to climate change Nat.
panel of regions J. Environ. Econ. Manage. 1 102360 Clim. Change 4 610–4
[4] Carleton T A and Hsiang S M 2016 Social and economic [27] Burke M and Emerick K 2016 Adaptation to climate change:
impacts of climate Science 353 6304 evidence from US agriculture Am. Econ. J. 8 106–40
[5] Deryugina T and Hsiang S 2017 The marginal product of [28] Mérel P and Gammans M 2018 Climate econometrics: can
climate Report No.: w24072 (Cambridge, MA: National the panel approach account for long-run adaptation?
Bureau of Economic Research) p w24072 (available at: American Journal of Agricultural Economics 103 1207–38
www.nber.org/papers/w24072.pdf) (Accessed 10 December [29] Taraz V 2017 Adaptation to climate change: historical
2020) evidence from the Indian monsoon Environ. Dev. Econ.
[6] Schlenker W and Roberts M J 2009 Nonlinear temperature 22 517–45
effects indicate severe damages to US crop yields under [30] Kolstad C D and Moore F C 2020 Estimating the economic
climate change Proc. Natl Acad. Sci. 106 15594–8 impacts of climate change using weather observations Rev.
[7] Dietz S and Stern N 2015 Endogenous growth, convexity of Environ. Econ. Policy 14 1–24
damage and climate risk: how Nordhaus’ framework [31] Duhamel P and Vetterli M 1990 Fast Fourier transforms: a
supports deep cuts in carbon emissions Econ. J. 125 574–620 tutorial review and a state of the art Signal Process. 19 259–99
[8] Gazzotti P et al 2021 Persistent inequality in economically [32] Bergland G 1969 Fast Fourier transform hardware
optimal climate policies Nat. Commun. 12 3421 implementations—an overview IEEE Trans. Audio
[9] Glanemann N, Willner S N and Levermann A 2020 Paris Electroacoust. 17 104–8
Climate Agreement passes the cost-benefit test Nat. [33] Ghil M 2002 Advanced spectral methods for climatic time
Commun. 11 110 series Rev. Geophys. 40 1003
[10] Hänsel M C et al 2020 Climate economics support for the [34] Smith J O 2007 Mathematics of the Discrete Fourier Transform
UN climate targets Nat. Clim. Change 10 781–9 (DFT): With Audio Applications (CA: W3k Publishing) p 323

9
Environ. Res. Lett. 17 (2022) 084038 B A Bastien-Olvera et al

[35] Mann M E, Steinman B A and Miller S K 2014 On forced [41] Barro R J and Ursúa J F 2008 Macroeconomic crises since
temperature changes, internal variability, and the AMO 1870 National Bureau of Economic Research wp 13940
Geophys. Res. Lett. 41 3211–9 (available at: https://doi.org/10.3386/w13940) (Accessed 25
[36] Tardif R, Hakim G J, Perkins W A, Horlick K A, Erb M P, July 2022)
Emile-Geay J, Anderson D M, Steig E J and Noone D 2019 [42] Barro R and Ursua J 2010 Barro-Ursua macroeconomic data
Last millennium reanalysis with an expanded proxy database (available at: https://scholar.harvard.edu/barro/publications/
and seasonal proxy modeling Clim. Past 15 4 barro-ursua-macroeconomic-data)
[37] Figwer J 1997 A new method of random time-series [43] Bolt J and van Zanden J L 2020 Maddison style estimates of
simulation Simul. Pract. Theory 5 217–34 the evolution of the world economy. A new 2020 update p 44
[38] World Bank 2021 World Bank open data (available at: [44] Bastien-Olvera B A and Moore F C 2021 Use and non-use
https://data.worldbank.org/) value of nature and the social cost of carbon Nat. Sustain.
[39] Kotz M, Levermann A and Wenz L 2022 The effect of rainfall 4 101–8
changes on economic production Nature 601 223–7 [45] Kikstra J S, Waidelich P, Rising J, Yumashev D, Hope C and
[40] Fischer E M and Knutti R 2016 Observed heavy precipitation Brierley C M 2021 The social cost of carbon dioxide under
increase confirms theory and early models Nat. Clim. climate-economy feedbacks and temperature variability
Change 6 986–91 Environ. Res. Lett. 16 094037

10