Académique Documents
Professionnel Documents
Culture Documents
Title
Estimating economic damage from climate change in the United States
Permalink
https://escholarship.org/uc/item/8db26620
Journal
Science, 356(6345)
ISSN
0036-8075
Authors
Hsiang, S
Kopp, R
Jina, A
et al.
Publication Date
2017-06-30
DOI
10.1126/science.aal4369
Peer reviewed
climate change in the United States and 8.5] (20), a probability distribution for GMST
change is constructed based on an estimated
distribution of equilibrium climate sensitivity,
Solomon Hsiang,1,2*† Robert Kopp,3*† Amir Jina,4† James Rising,1,5† historical observations, and a simple climate model
Michael Delgado,6 Shashank Mohan,6 D. J. Rasmussen,7 Robert Muir-Wood,8 (SCM) (19). (ii) The joint spatiotemporal distribu-
Paul Wilson,8 Michael Oppenheimer,7,9 Kate Larsen,6 Trevor Houser6 tion of monthly temperature and precipitation is
constructed from a broad range of global climate
Estimates of climate change damage are central to the design of climate policies. Here, models (GCMs), statistically downscaled from the
we develop a flexible architecture for computing damages that integrates climate science, Coupled Model Intercomparison Project 5 (CMIP5)
econometric analyses, and process models. We use this approach to construct spatially archive (21) and assigned a probability of realiza-
explicit, probabilistic, and empirically derived estimates of economic damage in the United tion such that the distribution of 21st-century
States from climate change. The combined value of market and nonmarket damage GMST change mirrors the distribution from the
across analyzed sectors—agriculture, crime, coastal storms, energy, human mortality, and SCM. Tails of the distribution beyond the range
labor—increases quadratically in global mean temperature, costing roughly 1.2% of gross present in the CMIP5 archive are represented by
domestic product per +1°C on average. Importantly, risk is distributed unequally across “model surrogates” constructed by scaling pat-
locations, generating a large transfer of value northward and westward that increases terns from CMIP5 models using the GMST pro-
E
conomically rational management of the rapid growth in a spatially resolved, empirical un- used when we compute damage probability dis-
global climate requires that the costs of re- derstanding of these relationships (14). Yet inte- tributions for specific RCP scenarios. (iii) We then
ducing greenhouse gas emissions be weighed grated assessments of climate change and their construct a set of 10 daily projections for each
against the benefits of doing so (or, converse- calculation of the social cost of carbon do not re- climate realization by superimposing daily weather
ly, the costs of not doing so). A vast liter- flect these advances (15–17). residuals relative to monthly climatologies that are
ature has considered this problem, developing, Here, we develop an integrated architecture to resampled in yearly blocks from the period 1981 to
among other insights, our understanding of the compute potential economic damages from cli- 2010 (Fig. 1B).
optimal timing of investments (1), the role of un- mate change based on empirical evidence, which A distribution of empirically grounded econom-
certainty (2), the importance of future adaptation we apply to the United States. Our risk-based ic impacts is computed for each joint realization
(3), the role of trade (4), and the potentially large approach is grounded in empirical longitudinal of county-level daily temperature and precipita-
impact of unanticipated tipping points (5, 6). analyses of nonlinear, sector-specific impacts, sup- tion: (iv) Econometrically derived dose-response
Integrated assessment models that value the bene- plemented with detailed energy system, inunda- functions (13) estimating the nonlinear effects
fits of greenhouse gas abatement are used by gov- tion, and cyclone models. Built upon a calibrated of temperature, rainfall, and CO2 on agriculture
ernments to estimate the social cost of climate distribution of downscaled climate models, this (22, 23), mortality (24, 25), crime (26, 27), labor
change (7, 8), which in turn informs the design approach is probabilistic and highly resolved (28), and energy demand (24) are constructed via
of greenhouse gas policies. However, the estima- across geographic space while taking into account Bayesian meta-analysis (29) (e.g., Fig. 1, C to H,
ted benefits of greenhouse gas abatement—or the spatial and sectoral covariance of impacts in and SM sections B and C). Following the approach
conversely, the “damages” from climate change— each possible future. Our framework is designed and criteria laid out in (30), we only employ studies
are conceptually and computationally challeng- to continuously integrate new empirical find- that are nationally representative, spatially dis-
ing to construct. Because of this difficulty, previous ings and new climate model projections as the aggregated, and account for temporal displacement
analyses have relied on rough estimates, theorized supporting subfields of research advance in the and unobserved heterogeneity across locations,
effects, or limited process modeling at continen- future. When applied to the U.S. economy, this along with the additional criterion that studies
tal scales or larger (9–11), with no systematic cali- approach provides a probabilistic and empirically statistically identify marginal distortions in the
bration to observed human-climate linkages (12). derived “damage function,” linking global mean distribution of experienced daily temperatures
Since the original development of these models, surface temperature (GMST) to market and non- (13, 14). (v) Econometric uncertainty is accounted
methodological innovations (13) coupled with data market costs in the United States, built up from for by resampling from the 26 posterior functions
availability and computing power have fueled empirical analyses using micro-level data. in (iv) (fig. S4). (vi) County-level daily projections
from (iii) are mapped onto the distribution of pos-
1
Global Policy Laboratory, Goldman School of Public Policy, System architecture sible responses from (v) to construct 3143 county-
University of California, Berkeley, CA, USA. 2National Bureau We developed the Spatial Empirical Adaptive Global- level joint distributions for 15 impacts across 29,000
of Economic Research, Cambridge, MA, USA. 3Department of
Earth and Planetary Sciences and Institute of Earth, Ocean,
to-Local Assessment System (SEAGLAS) to dynam- possible states of the world during 2000 to 2099
and Atmospheric Sciences, Rutgers University, New ically integrate and synthesize research outputs (SM sections D and E), although for display pur-
Brunswick, NJ, USA. 4Department of Economics and Harris across multiple fields in near-real time. We use poses we primarily summarize 2080 to 2099 im-
School of Public Policy, University of Chicago, Chicago, IL, SEAGLAS to construct probabilistic, county-level pacts here.
USA. 5Energy Resource Group, University of California,
Berkeley, CA, USA. 6Rhodium Group, New York, NY, USA.
impact estimates that are benchmarked to GMST A parallel approach is necessary to estimate
7
Woodrow Wilson School of Public and International Affairs, changes. [See section A of the supplementary energy demand changes and coastal impacts: (vii)
Princeton University, Princeton, NJ, USA. 8Risk Management materials (SM) for additional details (18).] Energy demand estimated in (iv) is used as a
Solutions, Newark, CA, USA. 9Department of Geosciences, County-level projections of daily temperature partial calibration for the National Energy Mod-
Princeton University, Princeton, NJ, USA.
*Corresponding author. Email: shsiang@berkeley.edu (S.H.);
and precipitation are constructed and sampled eling System (NEMS) (31) (SM section G). NEMS
robert.kopp@rutgers.edu (R.K.) †These authors contributed following a three-step process that simultaneously is then run with different weather realizations
equally to this work. captures the probability distribution of climate to estimate energy supply costs. (viii) Cyclone
exposure is simulated via analytical wind field Fig. 1. Recombining previous research results
models (32) that force a storm surge model (33), as composite inputs to SEAGLAS. (A) Forty-
with cyclogenesis and storm tracks generated via four climate models (outlined maps) and model
either (i) semiparametrically resampling histori- surrogates (dimmed maps) are weighted so that
cal activity (34) or (ii) resampling from projected the distribution of the 2080 to 2099 GMST
storm tracks and intensities (35) (SM section H). anomaly exhibited by weighted models matches
(ix) Inundation from localized probabilistic sea the probability distribution of estimated GMST
level rise projections (36) interacting with storm responses (blue-gray line) under RCP8.5.
surge and wind exposure in (viii) are mapped onto Analogous display for precipitation in fig. S1.
a database of all coastal properties maintained (B) Example of 10 months of daily residuals in
by Risk Management Solutions, where engineer- New York City, block resampled from historical
ing models predict damage (SM section H). observations at the same location and
Finally, economic impacts are aggregated and superimposed on monthly mean projections for
indexed against the GMST in their corresponding a single model (GFDL-CM3) and scenario
climate realization to construct multidimensional (RCP8.5) drawn from (A). (C to H) Examples of
probabilistic damage functions suitable for ap- composite (posterior) county- level dose-
plication in integrated assessment modeling: (x) response functions derived from nonlinear
Direct impacts from (vi), (vii), and (ix) are ag- Bayesian meta-analysis of empirical studies
gregated across space or time within each sector. based on selection criteria in (30). Median
Monetizing the value of nonmarket impacts estimate is black, central 95% credible interval is
(deaths and crime) using willingness-to-pay or blue-gray. To construct probabilistic impact pro-
to reduce yields ∼12.1 (±0.7) % per °C (see also mining, agriculture, and manufacturing) (Fig. 3, expected direct annual economic damage 0.6
figs. S11 and S12 and tables S10 and S11). E and F, and table S14). to 1.3% of state gross domestic product (GDP) for
Rising mortality in hot locations more than Property crime increases as the number of cold South Carolina, Louisiana, and Florida in the
offsets reductions in cool regions, so annual na- days—which suppress property crime rates (fig. median case, and 0.7 to 2.3% for the 95th per-
tional mortality rates rise ∼5.4 (±0.5) deaths per S4)—falls but then flattens for higher levels of centile of MSL rise (Fig. 4E) (RCP8.5). Nation-
100,000 per °C (Fig. 3C). For lower GMST changes, warming because hot days do not affect property ally, MSL rise would increase annual expected
this is driven by mortality between ages 1 and 44 crime rates. Violent crime rates increase linearly storm damages roughly 0.0014% GDP per cm if
and by infant mortality and ages ≥45 for larger at a relatively precise 0.88 (±0.04) % per °C in capital and storm frequency remain fixed (Fig.
GMST increases (fig. S13 and table S12). GMST (Fig. 3, G and H, and table S15). 4F). Accounting for the projected alteration of
Electricity demand rises on net for all GMST Coastal impacts are driven by the amplifica- the TC distribution roughly doubles the damage
changes, roughly 5.3 (± 0.14) % per °C, because tion of tropical cyclone and extratropical cyclone from MSL rise, the two combined costing an es-
rising demand from hot days more than offsets storm tides by local MSL rise and by the alter- timated additional 0.5 (±0.2) % of GDP annually
falling demand on cool days (Fig. 3D and table ation of the frequency, distribution, and intensity in 2100 when aggregated nationally (Fig. 4G).
S13). Because total costs in the energy sector are of these cyclones (SM section H). Rising MSL
computed using NEMS, demand is not statistical- increases the storm tide height and floodplain Uncertainty
ly resampled as other sectors are (SM section G). during cyclones: Fig. 4, A to D, illustrates how 1- At the county level, conditional upon RCP, un-
Total hours of labor supplied declines ∼0.11 in-100-year floodplains evolve over time due to certainty in direct damages is driven by climate
(±0.004) % per °C in GMST for low-risk workers, MSL rise (RCP8.5) with and without projected uncertainty (both in GMST and in the expected
who are predominantly not exposed to outdoor changes in cyclones for two major coastal cities. spatiotemporal distribution of changes conditional
temperatures, and 0.53 (±0.01) % per °C for high- Coastal impacts are distributed highly unequally, on GMST), by within-month weather exposure,
risk workers who are exposed (∼23% of all em- with acute impacts for eastern coastal states with and by statistical assumptions and sampling used
ployed workers, in sectors such as construction, topographically low cities; MSL rise alone raises to derive dose-response functions, as well as by
form suggests losses of ∼1.2% GDP per 1°C on empirically, and may evolve in the future, its in- differences are more extreme for the richest 5%
average in our sample of scenarios (table S16). fluence on damages is an important area for fu- and poorest 5% of counties, with average inter-
The greatest direct cost for GMST changes larger ture investigation. vals for damage of −1.1 to 4.2% and 5.5 to 27.8%,
than 2.5°C is the burden of excess mortality, with respectively.
sizable but smaller contributions from changes Risk and inequality of total We note that it is possible to adjust the aggre-
in labor supply, energy demand, and agricultural local damages gate damage function in Fig. 5A to capture socie-
production (Fig. 5B). Coastal storm impacts are Climate change increases the unpredictability and tal aversion to both the risk and inequality in Fig.
also sizable but do not scale strongly with GMST between-county inequality of future economic out- 5C. In SM section K, we demonstrate one approach
because projections of global MSL are dependent comes, effects that may alter the valuation of cli- to constructing such inequality-neutral, certainty-
on RCP but are not explicitly calculated as func- mate damages beyond their nationally averaged equivalent damage functions. Depending on the
tions of GMST (36), causing the coastal storm expected costs (45). Figure 5C displays the prob- parameters used to value risk and inequality,
contribution to the slope of the damage function ability distribution of damage under RCP8.5 as a accounting for these factors may dramatically
to be relatively muted. It is possible to use alter- fraction of county income, ordering counties by influence society’s valuation of damages in a man-
native approaches to valuing mortality in which their current income per capita. Median damages ner similar to the large influence of discount rates
the loss of lives for older and/or low-income indi- are systematically larger in low-income counties, on the valuation of future damages (46). This find-
viduals are assigned lower value than those of increasing by 0.93% of county income (95% confi- ing highlights risk and inequality valuation as cri-
younger and/or high-income individuals (44), an dence interval = 0.85 to 1.01%) on average for tical areas for future research.
adjustment that would alter damages differently each reduction in current income decile. In the
for different levels of warming based on the age richest third of counties, the average very likely Discussion
and income profile of affected individuals (e.g., range (90% credible interval, determined as the Our results provide a probabilistic, national dam-
fig. S13). Here, we focus on the approach legally average of 5th and 95th percentile values across age function based on spatially disaggregated,
adopted by the U.S. government for environmental counties) for damages is −1.2 to 6.8% of county empirical, longitudinal analyses of climate im-
cost-benefit analysis, in which the lives of all indi- income (negative damages are benefits), whereas pacts and available global climate models, but
viduals are valued equally (37). Because the VSL for the poorest third of counties, the average it will not be the last estimate. Because we use
parameter is influential, challenging to measure range is 2.0 to 19.6% of county income. These stringent selection criteria for empirical studies,
simultaneously forced by direct damages in all 3. K. C. de Bruin, R. B. Dellink, R. S. Tol, Clim. Change 95, 63–81 45. C. Gollier, Pricing the Planet’s Future: The Economics of
sectors, net market losses in general equilibrium (2009). Discounting in an Uncertain World (Princeton Univ. Press,
4. A. Costinot, D. Donaldson, C. Smith, J. Polit. Econ. 124, 2013).
tend to be larger than direct damages by ∼50% 205–248 (2016). 46. M. L. Weitzman, J. Econ. Lit. 45, 703–724 (2007).
(mortality is excluded from both). These simu- 5. D. Lemoine, C. Traeger, Am. Econ. J. Econ. Policy 6, 137–166 47. J. A. Patz, D. Campbell-Lendrum, T. Holloway, J. A. Foley,
lations are relatively coarse approximations of (2014). Nature 438, 310–317 (2005).
the complex national economy and do not cap- 6. Y. Cai, K. L. Judd, T. M. Lenton, T. S. Lontzek, D. Narita, Proc. 48. S. M. Hsiang, Proc. Natl. Acad. Sci. U.S.A. 107, 15367–15372
Natl. Acad. Sci. U.S.A. 112, 4606–4611 (2015). (2010).
ture international trade effects, but they suggest 7. N. Stern, Stern Review: The Economics of Climate Change 49. G.-R. Walther et al., Nature 416, 389–395 (2002).
that the spatial reallocation of economic activity (Cambridge Univ. Press, 2006). 50. O. Deschênes, M. Greenstone, Am. Econ. Rev. 97, 354–385
within the United States may not easily mitigate 8. Interagency Working Group on Social Cost of Carbon, (2007).
the economic damage from climate change. Technical support document: Social cost of carbon for 51. M. Burke, K. Emerick, Am. Econ. J. Econ. Policy 8, 106–140
regulatory impact analysis under Executive Order 12866, Tech.
Our results are “bottom-up” micro-founded rep., United States Government (2010).
(2016).
estimates of U.S. damages, although parallel analy- 52. T. Deryugina, S. M. Hsiang, NBER Working Paper 20750
9. R. L. Revesz et al., Nature 508, 173–175 (2014).
(NBER, 2014); www.nber.org/papers/w20750.
ses have employed “top-down” macro-level ap- 10. N. Stern, J. Econ. Lit. 51, 838–859 (2013).
53. S. M. Hsiang, A. Jina, NBER Working Paper 20352 (NBER,
proaches that estimate how overall productivity 11. R. S. Tol, J. Econ. Perspect. 23, 29–51 (2009).
2014); www.nber.org/papers/w20352.
12. R. S. Pindyck, J. Econ. Lit. 51, 860–872 (2013).
measures (such as GDP) directly respond to tem- 13. S. M. Hsiang, Annu. Rev. Resour. Econ. 8, 43–75 (2016).
perature or cyclone changes without knowledge 14. T. A. Carleton, S. M. Hsiang, Science 353, aad9837 AC KNOWL ED GME NTS
of the underlying mechanisms generating those (2016). This research was funded by grants from the National Science
losses. This alternative approach can be compared 15. R. E. Kopp, S. M. Hsiang, M. Oppenheimer, Impacts World 2013 Foundation, the U.S. Department of Energy, Skoll Global Threats
Conference Proceedings (Potsdam Institute for Climate Impact
to our estimates of market losses only, as they will Research, Potsdam, Germany, 2013), pp. 834–843.
Fund, and by a nonpartisan grant awarded jointly by Bloomberg
not account for nonmarket valuations. Our mar- Philanthropies, the Office of Hank Paulson, and Next Generation.
16. W. Pizer et al., Science 346, 1189–1190 (2014). The methodology and results presented represent the views
ket estimates are for a 1.0 to 3.0% loss of annual 17. M. Burke et al., Science 352, 292–293 (2016). of the authors and are fully independent of the granting