American Economic Association

The Economic Impacts of Climate Change: Evidence from Agricultural Output and Random
Fluctuations in Weather: Reply
Author(s): Olivier Deschnes and Michael Greenstone
Source: The American Economic Review, Vol. 102, No. 7 (DECEMBER 2012), pp. 3761-3773
Published by: American Economic Association
Stable URL: http://www.jstor.org/stable/41724654
Accessed: 14-11-2015 00:12 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to The American Economic
Review.

http://www.jstor.org

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
 Economic
American Review
2012, 3761-3773
102(7):
http://dx.doi.org/10.1257/aer.102.73761

The Economic Impacts of Climate Change:

Evidence from Agricultural Output and Random
Fluctuations in Weather: Reply

By Olivier Deschnes and Michael Greenstone*

Fisheret al. (2012) (hereafter,FHRS) have uncoveredcodingand data errorsin our

paper,Deschnes and Greenstone(2007) (hereafter, DG). We acknowledgeand are
embarrassedby these mistakes.We are gratefulto FHRS foruncoveringthem.We
hope thatthisReply will also contribute to advancingtheliteratureon thevitalques-
tionof theimpactof climatechange on theUS agriculturalsector.
FHRS' main critiquesof DG are as follows: (i) thereare errorsin theweatherdata
and climatechange projectionsused by DG; (ii) theclimatechange projectionsare
based on the Hadley 2 model and scenarios,ratherthanthe more recentHadley 3
model and scenarios; (iii) standarderrorsare biased due to spatial correlation;
(iv) theinclusionof stateby yearfixedeffectsdoes notleave enoughweathervaria-
tion to obtain meaningfulestimatesof the relationshipbetween agricultureprof-
its and weather;(v) storage and inventoryadjustmentin response to yield shocks
invalidatethe use of annual profitdata; and (vi) FHRS argue thata better-specified
hedonic model produces robustestimates,unliketheresultsreportedin DG.
Four of these critiqueshave littlebasis and we respondto themhere in the intro-
duction.Specifically,withrespectto:

(ii) The more recentdaily climatepredictionswere not available when we wrote

DG. Nevertheless,the most importantissue is providingthe reliable esti-
mates of climate change and in this note we reportestimatesbased on the
climatemodel we used in DG and a morerecentone thatwe gained access to
in themeantime.

(iii) In the primarytable on agriculturalprofits,DG reportstwo sets of standard

errorswiththefirstclusteredat thecountylevel and thesecondbased on a vari-
ance-covariancematrixthataccountsforspatialcorrelation, using themethod
proposedin Conley (1999). Thus, theclaim of FHRS 2012 seems overblown.
to ease comparisonsofpapersin thisliterature,
Nevertheless, thisnotewill adopt
the FHRS conventionof reportingestimatedstandarderrorsclusteredat the
countyand statelevels;we findthatinferenceis largelyunaffected
by thechoice
betweenthesedifferent assumptionsaboutthevariance-covariance matrix.

*Deschnes:
DepartmentofEconomics, ofCalifornia,
University SantaBarbara,
2127NorthHall,Santa
CA93106,IZA,andNBER(e-mail:
Barbara, ; Greenstone:
[email protected]) MITDepartment
ofEconomics,
50Memorial
E52-359, Drive,
Cambridge,MA02142-1347,andNBER(e-mail: Wethank
[email protected]).
Wolfram
Schlenker
forseveral
useful
conversations
andproviding
thedatafiles
andcomputer
codeusedbyFisher,
Hanemann, andSchlenker.
Roberts,
Toviewadditional visit
materials, thearticle
pageathttp://dx.doi.Org/10.1257/aer.102.7.3761.

3761

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
3762 THEAMERICAN
ECONOMIC
REVIEW 2012
DECEMBER

(iv) We demonstrate below thatitis indeedpossible to make meaningfulinference

whenthemodels includestateby yearor regionby yearfixedeffects.Further,
thedata soundlyrejectthehypothesisof zero regionalprice,productivity, and/
or cost shocks.For thesereasons,we continueto believe thatthemostreliable
specificationsare thosethatinclude stateor regionby year fixedeffects.

(vi) The onlineAppendix,available on theAER website,reportson cross-sectional

(hedonic) models thatrelate land values and climate variables to inferthe
impactsof climatechange using a varietyof approachesto model theclimate
variables.As in DG, theseresultsagain demonstratethatpredictedimpactsof
climatechange are heavilydependenton functionalformchoices of the tem-
peraturevariables,thecovariatesused foradjustment,and theparticularyearof
data used to fitthemodel.Thereforewe maintainourconclusionfromDG that
thehedonicapproachis unlikelyto providecredibleestimatesof theimpactof
climatechangeon theagricultural sectordue to problemsof omittedvariables.

The remainderof the paper then assesses the impact of the remainingtwo cri-
tiques on the estimatesof the impact of climate change on US agriculturalprofits.

Using a correctedversionof our data file,a varietyof specificationsthatdo and do
not accountforlocal shocks,and theclimatemodel (i.e., Hadley 2) available when
we wrote DG, we findthatclimate change is projectedto reduce annual agricul-
turalsectorprofitsby about US$(2002) 4.5 billion by the end of the century.1 We
obtain similarresultswhen we apply the same to
specifications a data filegraciously
providedby FHRS. These resultscontrastwithDG's findingof a statisticallyinsig-
nificantincreaseof roughly$1.3 billion. Using a 3 percentdiscountrateand annual
projectionsof climate changes,the presentdiscountedvalue of the change in agri-
culturalprofitsbetween 2010 and 2100 is -$66 billion.2 To put this in context,
historicalannual agriculturalsectorprofitsare about $33 billion.
Notably,more recent climate model projections (i.e., the CommunityClimate
System Model 3 (CCSM 3) and A2 scenario) indicate greaterwanning and the
applicationof these projectionslead to largerdamage estimates.The use of such
climatechange predictionscauses thechange in annual agriculturalsectorprofitsto
increase in magnitudeto about $9.9 billion by the end of the century.The present
discountedvalue of projectedprofitchanges with these projectionsover the next
90 yearsis $164 billion.
The remainingpoint raised by FHRS pertainsto the fact thatthe farmrevenue
measurein thecensus of agricultureincludesproductssold, regardlessof theiryear
of production.Thus, therelationshipbetweenannualprofitsand annualweatherreal-
izationsmaybe confoundedby inventory adjustments.The textbooksolutionto such
issues of dynamicinventoryadjustmentin agriculturaland othersettingsis to use a
distributedlag model and computecumulativeeffects.Thus, the impactof a year's
weatherrealizationis capturedover several years. In this setting,the coefficients
on the lag of temperature tend to have the opposite sign as the contemporaneous
temperature variables.Since the "full" impactof temperature froma distributed lag

1Alldollar in2002dollars.
areexpressed
figures
2Theseresults intheonline
arepresented Appendix.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
VOL102NO.7 ANDGREENSTONE:
DESCHNES ECONOMICS
OFCUMATE REPLY3763
CHANGE:

model is thesumofthecoefficients,theprojectedimpactsof climatechangefromthis

model are morethan50 percentsmallerthanthosedescribedabove. This approachis
moredemandingof thedata and theestimatesare less precisethanis ideal, however.
Finally,it is worthunderscoringa point thatwe make in DG. All of these esti-
matesare derivedundertheunrealisticassumptionof no technologicalprogressand
adaptationover the remainderof the century.It seems reasonable to assume these
economic forceswill contributeto reducingthepredicteddamages.

I. CorrectedImpactsofWeatherFluctuationson AgriculturalProfits

This sectionreportsestimatesof therelationshipbetweenweatherfluctuations on

agriculturalprofitsfrom data filesthat in
correctsour mistakes DG. In particular,we
have correctedthe weatherand climate projectiondata and reconstructedthe main
samples fromtheUS census of agricultureused in DG.
We brieflydescribetheconstructionof theweatherdata samples and climatepro-
jections samples, withmore details available in the online Appendix forthe inter-
estedreader.3The daily temperature data are drawnfromtheNational ClimaticData
CenterSummaryof the Day Data Files. The key variables are the daily maximum
and minimumtemperature, and we definedaily average temperatureas the simple
averageof theminimumand maximumtemperature. We select weatherstationsthat
are less than7,000 feetin elevationand thatwere operational(i.e., had nonmissing
measurements)in all 183 days of a year's growingseason (i.e., Aprilto September).
The station-leveldata is aggregatedat thecountylevel by takingan inverse-distance
weightedaverage of all thevalid measurementsfromstationsthatare located within
a 200 km radius of each county's centroid.The growingseason rainfalldata was
taken from the Parameter-ElevationRegressions on IndependentSlopes Model.
This model generatesmonthlyestimatesof totalprecipitationand average tempera-
turesat 4 X 4 kilometersgridcells fortheentireUnited States.
The correctionsalterthe summarystatistics.The biggestimpactof thedata errors
in DG is thatgrowingseason degree-dayswere too low: thefarmland-weighted aver-
age in theDG data is 2,561 while in thecorrectedsample,thecorrespondingaverage
is 3,821.4 In addition,79 countieswere incorrectly droppeddue to the errorsin the
weatherdata. In contrast,thegrowingseason rainfallvariablein DG was error-free.
We utilize two sets of daily predictedclimate change data. The firstone is from
theHadley 2 model coupled withtheIS92a scenario (which we label forsimplicity
Hadley 2), the same used in DG. The variables containedin thisfile are daily pre-
cipitationand daily minimumand maximumtemperatures.The data is reportedat
gridpointsseparatedverticallyand horizontallyby 0.5 over thecontinentalUnited
States.The second is fromtheNational CenterforAtmosphericResearch's CCSM 3
under the A2 scenario, which togetherpredictslargertemperatureincreases than
Hadley 2. The variables available fromthe CCSM 3 A2 files are the daily mean
temperatures and precipitationlevels foreach day duringtheyears 2000-2099. The

3Thecorrected aswellastheSTATA
data, areposted
ontheAER's website.
programs,
FigureAl intheonlineAppendix thequartiles
displays ofhistorical
distribution
ofgrowing
season
degree-
Itisevident
days. thatthespatial
discontinuities
that theDG(2007)growing
plagued season dataare
degree-days
notpresent
inthecorrected
data.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
3764 THEAMERICAN REVIEW
ECONOMIC DECEMBER
2012

CCSM 3 gridspans theentireglobe; latitudeand longitudepointsare bothseparated

by 1.4. We use the416 gridpointsthatfall on land in thecontiguousUnited States
to develop climatepredictionsforthecontiguousUnited States.
For both sets of climate predictionvariables, we use inverse-distanceweighted
averagingto assign grid point predictionsto counties.All grid points located in a
200 km radiusof a county'scentroidare used to imputetheclimateprediction.This
approach produces observationsthatvary at the countyx day x year level.5 From
these,we definedpredictedend-of-century climate change forany county-yearas
the differencebetween the Hadley 2/CCSM 3 model-predictedaverage growing
season weatherover 2070-2099 and the 1970-2000 average of the same growing
season weathervariable.6The farmland- weightedpredictedchangein averagegrow-
ing season degree-daysover 2070-2099 forHadley 2 forthecountiesin our sample
is 673, whichcorrespondsto an 18 percentincrease over the 1970-2000 average of
3,821.7 The predictedchange fromCCSM 3 A2 is 1,441, which correspondsto a
35 percentincrease over the 1970-2000 baseline.8
The primarydata file is comprisedof a balanced sample of countieswith valid
observationson farmrevenuesand productionexpenditures(thetwo variablesused to
constructfarmprofits), totalacresoffarmland, and acres ofirrigated
farmlandin 1987,
1992, 1997, and 2002.9This sampleis meantto replicatetheone used in DG as closely
as possible while correctingforthe issues outlinedby FHRS. The resultingsample,
whichwe label the"REPLY" sample,has 2,342 countiesfora totalof 9,368 county-
yearobservations,and accountsfor84 percentof US farmland.10 By comparison,the
sampleused in DG had 2,262 countiesfora totalof 9,048 county-year observations.
We used thesedata to fit:

(1) Yct= ac + 7, + X'ctir+ ^ ifi(Wict

) + uct,
i
wherec denotes a countyand t referencesa year.The dependentvariable is annual
agriculturalprofits(defined as the differencebetween revenues and production
expenses) per acre and the equation is weightedby farmlandacres. The equation
includes a full set of countyfixedeffects,ac9 and year indicators,jt. We also con-
sider specificationsthatreplace the year fixed effectswith year effectsthatvary
geographicallyto allow for local shocks to productivity, inputprices, and output
prices.The Xctvectorincludes the same set of soil characteristicsas in DG.

5InDG(2007), theclimatepredictionsfromHadley atthestate*year

2 varied levelonly.
6These measuresofpredicted
climatechange areanalyzedintheonline seeFigure
Appendix; A2.
7FortheHadley 2 model,ouralgorithm predictscoolingbytheendofthecentury for26counties, primarily
located
inColorado.Thisispossiblyduetolackofadjustment Theinclusion
forelevation. orexclusion ofthese
counties
doesnotaltertheresults
meaningfully.
8FHRSreport estimatesbasedontheHadley 3 modelandB2scenario,whichalsopredicts
greater temperature
than
increases Hadley2. Theunweighted increaseingrowing seasondegreedaysis 720Celsius, which when
converted
toFahrenheitiscomparabletotheincrease byCCSM3 A2.WedidnotusetheHadley
predicted 3 B2
because
predictions wedonothaveaccess version
toa daily ofthese for
predictions thefull21stcentury.
9Countieswithzeroacresoffarmland aredropped fromthesample.
10Among these theonethat
variables, ismost missing
frequently intheCensusofAgricultures acres
isirrigated
offarmland.
Abalanced sampleforallcountieswithvalid onfarm
observations sales, expenditures
production (the
twovariables
usedtocalculatefarm andtotal
profits), acresoffarmlandin1987,1992,1997,and2002has2,963
counties
fora total
of11,852 county-year andaccounts
observations, for98percentofUSfarmland. Werefer to
this
sampleasthe"FULL"sample. Weobtain similar
qualitatively ifwefocus
estimates onthebalanced panelof
2,963counties
instead(seeonline
Appendix).

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
VOL102NO.7 DESCHNES
ANDGREENSTONE:
ECONOMICS
OFCUMATE REPLY3765
CHANGE:

The variablesof interestare theweatherones, Wict. As in DG, we model tempera-

turewith a quadratic in growingseason degree-days.11'12 Growing season degree
days are calculated fromthedaily average temperature witha base of 46.4 F and a
ceiling of 89.6 F. Precipitationis modeled witha quadraticin totalgrowingseason
rainfallin countyc in year t.
Table 1 reportson a reanalysis of the annual profitsdata, which is the primary
outcome in DG. The column (a) entriesreportestimatesfroma version of equa-
tion (1) thatrestrictsthe year effectsto be constantacross the country'scounties.
This is the specificationthatFHRS favorin theircomment.Column (b) entriesare
based on specificationsthatallow foryeareffectsspecificto each of thenine USDA
Farm Resource regions.13The column (c) estimatingequation includes year effects
specificto each of the nine US census divisions,and column (d) includes stateby
year fixedeffectsas in DG.
Panel A of Table 1 reportsthemarginaleffectsof thegrowingseason degree-days
and precipitationvariables evaluated at the national sample means. The table also
reportsthestandarderrorassociatedwitheach marginaleffect,estimatedwithcluster-
ing at thecountylevel (in parentheses)and at thestatelevel (in squarebrackets).The
estimatesare allowed to varyby whethera countyis irrigated, whichis definedby hav-
ing more than 10 percentof the farmlandirrigated.Among the eightestimates(four
specifications and twosetsofcounties),thedegree-daymarginaleffectsare onlystatis-
ticallysignificant fornonirrigatedcountiesin thecolumn (la) specificationwithyear
fixedeffects.This specificationindicatesthat100 additionalgrowingseason degree-
days reducesprofitsby $ 1.27 peracre; meanprofitsperacre in thissample are $3 1.30.
It is apparentthatthestrongestevidencein favorof a negativerelationshipbetween
temperature and agriculturalprofitscomes fromthecolumn (a) specification,which
is the one stronglypreferredby FHRS. This specification,however,is the most
susceptibleto bias due to unobservedlocal shocks to prices,costs, or productivity.
The remainingspecificationstryto balance controllingforthese shocks withdiffer-
ent varietiesof year by regionfixedeffects,while at the same timeleaving enough
weathervariationformeaningfulinference.
There are threeimperfectdiagnosticsof the effortto balance these goals. First,
panel B indicates thatthe data decisively reject the null hypothesisof zero local
shocks to agriculturalprofitsin columns (b), (c), and (d). This testwas conducted
jointlyin irrigatedand nonirrigated countiesso theentriesin columns (la)-(ld) are
identical to those in columns (2b)-(2d). Second, the coefficientson the growing

11The resultsarequalitatively
similartothose thatusea measure ofdegree-days
derived
from
subsequent fitting
a sinusoidal
curvebetweenminimum andmaximum temperatures asinSchlenker andRoberts
(2009).Alimitation
ofthesinusoidal
approachisthatitimposesa fixedparametric nonlineardistribution
oftemperatures
within each
dayacrossgeographyandtime ofyear.
12TheReply'sSupplementaryAppendix available athttp://www.econ.ucsb.edu/~olivier/research.html
reports
ona secondapproachtomodeling temperaturethatfollows from thework ofDeschnes andGreenstone (2011);
Deschnes, andGuryan
Greenstone, (2009);Schlenker andRoberts (2009);andBurgessetal. (2011),whoall
theimportance
highlight ofextreme inmodels
temperatures formortality,
infantbirth
weight,andcropyields.
this
Specifically, approachcharacterizes
exposuretogrowing season daily
temperatureswith
a setoftemperature-
or"bins"
daycategories that spanthegrowing season dailytemperature distribution.
Theadvantage ofthebin
approachoverthedegree-daysoneisthattheonly functionalform restriction
isthattheimpactofthedaily mean
onfarm
temperature profitsisconstantwithin
5F degree intervals.
EstimatesofthePDVofthechange inagricul-
tural between
profits 2010and2100from temperature-day binmodels anddegree-daysmodelsarebroadly similar
totheestimates
basedonmodeling with
temperature a quadraticindegreedays.
3Seehttp://www.ers.usda.gov/briefing/arms/resourceregions/resourceregions.htm.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
3766 THEAMERICAN
ECONOMIC
REVIEW DECEMBER
2012

Table1- In-Sample
Estimates
oftheEffect
ofGrowing
SeasonWeather
onFarmProfits
BasedonCorrected
Data

(N= 7,743) Irrigated

counties
Nonirrigated (N= 1,625)
counties
(la) (lb) (lc) (Id) (2a) (2b) (2c) (2d)
Panel
A.Marginal
effects
(atsample
mean)
season
Growing X100 -1.27 -0.39
degree-days 0.10 - 0.18 -1.51 0.87 1.48 1.09
error
Standard clustered
bycounty(0.25) (0.35) (0.37) (0.75) (2.19) (1.87) (2.17) (2.87)
Standard clustered
error bystate [0.39] [0.41] [0.54] [0.56] [1.81] [1.60] [1.82] [2.48]
season
Growing precipitation -0.58 0.61 0.13 0.03 0.06 2.30 1.77 0.91
Standard
error
clustered
bycounty (0.24) (0.24) (0.23) (0.26) (1.62) (1.59) (1.56) (1.53)
Standard
error
clustered
bystate [0.55] [0.39] [0.35] [0.33] [1.48] [1.27] [1.45] [1.69]
Panel
B.Tests onvarious
ofsignificance models
foryear
fixedeffects
oninteracted
F-statistic effects -
year 29.45 29.21 10.66 - 29.45 29.21 10.66
Rvalue] - [0.01] [0.01] [0.01] - [0.01] [0.01] [0.01]
PanelC.Tests ofweather
ofequality variables
across andnonirrigated
irrigated counties
F-statistic
ondegree-days 0.01 0.51 0.46 0.27 0.01 0.51 0.46 0.27
[p-value] [0.91] [0.47] [0.50] [0.60] [0.91] [0.47] [0.50] [0.60]
onprecipitation
F-statistic 0.14 1.11 1.09 0.34 0.14 1.11 1.09 0.34
[p-value] [0.71] [0.29] [0.30] [0.56] [0.71] [0.29] [0.30] [0.56]
Yeareffects Yes No No No Yes No No No
USDAregion x year
effects No Yes No No No Yes No No
Census x year
division effects No No Yes No No No Yes No
x year
State effects No No No Yes No No No Yes
Notes: Alldollarfigures inbillions
of2002constant dollars. counties
"Irrigated" aredefined as thosewhere
10percent ormoreofthetotal farmland Themeans
isirrigated. ofthedependent variable farm
(i.e.,county-level
peracre)innonirrigated
profits andirrigatedcountiesare$31.3and$85.8, There
respectively. are2,342 fora total
of9,368 county-year Standard
observations. inparentheses
errors areclustered
atthecounty level.Standard
errors
inbrackets areclusteredatthestate
level. andp-values
F-statistics inpanelB areidenticalforcolumns (la) and
(2a),(lb) and(2b),(lc) and(2c),and(Id) and(2d)because theyeareffects
arenotinteracted status
byirrigation
inthemodels. F-statistics
andp-valuesinpanel C areidentical
forcolumns (la) and(2a),(lb) and(2b),(lc) and
(2c),and(Id)and(2d).Thisisbecause these tests
arefortheequalityoftheweathervariablesacross and
irrigated
nonirrigatedcounties.
Seethetext formore details.

season degree-daysvariables change across specificationsand even switch signs.

This is consistentwith specification(a) confoundinglocal weatherand economic
shocks and (at least the attenuation)is consistentwith an increased influenceof
measurementerror.Third,the standarderrorson thetemperature variables are gen-
erallylargerfromthe (b)-(d) specificationsbut not uniformlyso and in almost all
cases theincreasesare less than50 percentof thespecification(a) standarderrors.14
Our conclusion fromthistable is not thatthereis a single specificationthatcon-
tainsthetruth,butratherthateach of themhas plusses and minuses.This is probably
our greatestmethodologicaldifferencewithFHRS. In theremainderof theanalysis,
we continueto reportresultsfromall fourof these specifications.We also report
combinedestimates,which are calculated as the weightedaverage of thekey coef-
ficients,wheretheweightis theinverseof thestandarderrors.
Finally,panel C shows the resultsfroma testof equality of the weatherparam-
etersacross irrigatedand nonirrigated counties.We fail to rejectthenull hypothesis
of equalityacross all specificationsand forbothweathervariables.As a result,most

14Thestandard
errors
clustered levelwhich
atthestate implicitly fora higher
control ofspatial
degree correla-
arenotuniformly
tion than
larger thecounty-clustered asinFHRS.
ones,

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
VOL102NO.7 DESCHNES
ANDGREENSTONE:
ECONOMICS
OFCUMATE REPLY3767
CHANGE:

of our remaininganalysis is based on models thatassume thatthe weatherparam-

etersare the same across irrigatedand nonirrigatedcounties while allowing foran
interceptdifference.This testis conductedjointlyacross irrigatedand nonirrigated
countiesso theentriesin (la) and (2a), (lb) and (2b), etc. are identical.

n. CorrectedEstimatesoftheImpactof ClimateChange on Farm Profits

This sectiondevelops correctedpredictedimpactsof climate change on US agri-

culturalprofits.Specifically,we combine theestimatesfromtheestimationof equa-
tion(1) withtheprojecteddifferencesin growingseason weatherfromHadley 2 and
CCSM 3 A2. The predictedimpact on aggregatefarmprofitsforcountyc in year t
is givenby15

(2) IMPACT = ACRESc x

whereA Wict is thepredictedchange in weathervariablei in countyc in yeart.These

changes are specificto a climatechange model and scenario.We "reweight"thecal-
culationssince theregressionmodel is forprofitsper acre and the variableACRESc
representtheaverageacres of farmlandduringthesample periodin countyc. Finally,
to obtaintheimpactforthecountryas a whole, we sum thecounty-specific impacts
(IMPACT^ across all countiesin the sample.
Table 2 reportson the predictedimpact of climate change on annual farmprof-
its at the end of the centuryfromequation (2). Panel A uses the Hadley 2 climate
predictions.There are some minordifferencesbetween the samples, calculation of
weathervariables,and calculation of climate predictionsthatFHRS utilize and the
ones we utilize.16As we discussed above and in the online Appendix,these differ-
ences reflectdifferencesin data resolutionand in themethodsto createcounty-level
weatherdata and climatepredictionsfrommonitorobservationsand gridpointpre-
17
dictions,respectively.
Consequently,thetable beginsby demonstrating theinfluenceof thesedifferences
on the results.Row 1 is obtained using FHRS's 2012 preferredsample, weather

15Inmodelswith quadratics ingrowing seasonweather,eachcounty'spredictedimpactiscalculated

asthedis-
crete inper-acre
difference atthecounty's
profits predicted
degree-daysandprecipitation
after
climatechangeand
itscurrent
climate(i.e.,theaverageover the1970-2000 period).
16With totheweather
respect thecorrelation
variables, between theDGvariables forgrowingseason weather
andtheFHRSonesareallinexcess of0.98.
17One difference isthatFHRSderive their
climate model and
key change predictions
usingmonthly-level data,
thisdataallowsforcorrecting foranymodel errorbycomparing themodel-predictedhistorical
baseline
withthe
actualhistorical
baseline. A limitation
isthatitisnecessarytomake assumptionsaboutthewithin-monthdistribu-
tionofdailytemperatures toderive variables
daily-level frommonthly-levelvariables.
Incontrast,
weusedaily-level model datathatallowsustoderive ondaily
predictions variables
and,inturn,
growing seasondegree-days withoutmaking assumptionsaboutthewithin-month between
relationship monthly
minimum andmaximum temperaturesanddaily temperature.Tothebestofourknowledge, thedaily-level
model
datafrom Hadley2 andCCSM3A2doesnotcontain sufficient
model-based onthehistorical
predictions baseline
tocorrectthedaily-levelclimatechange formodel
predictions error.
Withrespecttotheissueofmodel error,theonlineAppendix thepresent
reports discounted
valueofpredicted
climatechangedamages between 2010and2099.These estimates
includea setbasedontheHadley3AlFImodel
andscenariothatadjust forthedifferenceinthemodel-predictedhistorical
baselineandtheactual
historical
base-
line.Theresultingestimates arequantitatively
similartotheonesfrom theotherclimatemodelswhere baseline
correction
isnotfeasible.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
3768 THEAMERICAN
ECONOMIC
REVIEW 2012
DECEMBER

variables, and climate predictions,which were graciously provided by Wolfram

Schlenker.Everythingis the same in row 2, exceptthattheFHRS weathervariables
are replaced withtheDG ones. Relativeto row 2, row 3 now utilizestheDG sample
(86 additionalcounties) and the DG climate predictions.The preferredestimates
are in row 4, which imposes the restrictionthatthe weathercoefficientsare equal
in irrigatedand nonirrigatedcounties as was supportedby the testin Table 1. Our
interpretation of these resultsis thatthe differencesbetween FHRS' data and our
correcteddata are notconsequential,since withina specificationor columnall of the
estimateshave overlappingconfidenceintervals.
With respect to the substantiveissue of predictedclimate change impacts, the
row 4 Hadley 2 estimatesrangefrompredictedlosses of $7.5 to $1.7 billion,corre-
spondingto -23 percentto -5 percentof currentannual agriculturalsectorprofits.
The weightedaverage of these estimatesis a change in annual agriculturalprofits
of -$4.5 billion or 14 percentat the end of the century,when the weightsare the
inverseof the standarderrors.Furthermore, just as in Table 1, allowing for local
shocks (i.e., the "b," "c," and "d" columns) tends to reduce the magnitudeof the
predictedloss, althoughthisdoes not always come at theexpense of reduced statis-
tical precision.For example, the state-clusteredstandarderrorsin column (lb) are
always smallerthanthe state-clustered standarderrorsin column (la).
Row 5 in panel B reportsthe correspondingend-of-century annual climate dam-
age estimates using the CCSM 3 A2 predictions.These are necessarilylargerin
absolute value since this climate model predictsa significantlylargerincrease in
growingseason temperatures,as well as a reductionin growing season rainfall.
The predictedlosses range from$14.8 to $4.8 billion. The inversestandarderror
weightedaverageof thepredictedimpactsunderCCSM 3 A2 is -$9.9 billion,more
thantwice as large as underthe Hadley 2 model; thisis about 30 percentof current
annual agriculturalprofits.
The online and supplementaryAppendices reportsthe presentdiscountedvalue
(PDV) of the predictedannual impact of climate change on aggregatefarmprof-
its over 2010-2099, based on a discount rate of 3 percent18.The mean of the
PDV estimates from the Hadley 2 projections and the four specificationsindi-
cates that the US agriculturalsector is predicted to sufferlosses of $66 billion
over the remainderof the twenty-first century.The correspondingfigurefromthe
CCSM3 A2 predictionsis a loss of $164 billion. Thus, these estimates,which do
not allow forlong-runadaptationor directedtechnicalchange, implythatclimate
change will cause a loss of about 2 and 5 years of currentprofitsin theagricultural
sector,respectively.

III. Dynamics,Storage,and DistributedLag Models

FHRS make theimportant pointthatin a givenyearfarmersare able to storesome

of theirgrainoutputand sell it in futureyears.19The 2006 US StatisticalAbstract

18Theseestimates
utilizeyearly climate
county-level change foreachyear
predictions 2010-2099.
between
Inpractice, occurs
storage onthefarm, instorage
off-farm spacesrented
byfarmers, incommer-
andoff-farm
cialstorage
facilities thecritique
Therefore
(i.e.,elevators). thatstorage therelationship
confounds farm
between
profitsandweather
depends crucially farmers
onwhether maintain
ownership orwhether
ofthegrain are
grains
purchasedbya secondparty toorduring
prior storage.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
VOL.102NO.7 ANDGREENSTONE:
DESCHNES ECONOMICS
OFCUMATE REPLY3769
CHANGE:

Table2- Comparison
ofEstimates
ofthePredicted
Impact
ofClimate
ChangeonUSAggregate
Farm BasedonFHRSandDGSamples,
Profits, DataandCode,forAverage
Yearover2070-2099,
Billions
of2002Dollars

(la) (lb) (lc) (Id)

PanelA.Predicted inaverage
impact yearover2070-2099under
Hadley2
1. FHRSsample
ofcounties,
FHRSweather FHRSclimate
variables, -11.1 -7.2 -3.0 0.2
predictions
Allowweather
coefficients
tovary
across andnonirrigated
irrigated counties (1.9) (3.1) (2.3) (4.7)
[3.8] [3.7] [5.4] [5.7]
2.FHRSsampleofcounties,
DGweather FHRSclimate
variables, predictions -9.0 -5.6 -2.3 -2.5
Allowweather
coefficients
tovary
across andnonirrigated
irrigated counties (2.1) (3.3) (2.6) (4.6)
[3.4] [2.7] [3.8] [4.4]
3.DGsampleofcounties,
DGweather DGclimate
variables, predictions -7.7 -5.0 -2.2 -2.1
Allowweather
coefficients
tovary
across andnonirrigated
irrigated counties (1.8) (2.9) (2.2) (3.9)
[3.0] [2.4] [3.3] [4.0]
4.DGsampleofcounties,
DGweather DGclimate
variables, predictions -7.5 -4.9 -2.2 -1.7
Restrict
weather
coefficients
tobeequalacross andnonirrigated
irrigated counties(1.9) (2.9) (2.3) (4.0)
[3.0] [2.3] [3.4] [3.8]
PanelB.Predicted inaverage
impact yearover2070-2099
underCCSM3A2
5.DGsample ofcounties,
DGweather DGclimate
variables, predictions - 14.8-11.5 -5.7 -4.8
Restrict
weathercoefficients
tobeequalacross andnonirrigated
irrigated counties(3.8) (6.4) (4.6) (8.3)
[6.4] [4.5] [7.4] [8.9]
Yeareffects Yes No No No
USDAregion x year
effects No Yes No No
Census x year
division effects No No Yes No
x year
State effects No No No Yes
Notes:Table2 reportsestimatesofthepredicted impact ofclimate changeonUS annual aggregatefarm profits
fortheaverage yearover2070-2099. Allestimatesarederivedfrom models with ingrowing
quadratics season
degree-daysandprecipitation.
Allestimates arederivedfrom models withquadraticsingrowing seasondegree-
daysandprecipitation,
andtheeffects oftheweather areallowed
variables tovary across andnonirrigated
irrigated
counties,
except inrows4 and5. Standard errorsinparenthesesareclusteredatthecounty level.Standarderrors
inbrackets
areclusteredatthestate level.
Estimates inrows 1and2 arebasedona sample of9,024observations.
inrows
Estimates 3-5arebased ona sample of9,368 Row1usesthesame
observations. regression model,
sample
(N= 9,024),data, andcodeasFHRS,except thatitreportspredicted
impactsinprofits
($2002 rather
billion) than
percent
impacts. Inaddition,
itestimatestheFHRSmodel byweighting theregression
byannual acresoffarmland
(asopposed) tohistorical
average farmland asinFHRS.Further, itestimates
theFHRSmodel while allowingfor
USDA-region specific
yeareffects(column lb) andUSCensus divisionspecific
yeareffects(column lc).Row2
usesthecorrected DGweather sample, measure,
profit andcode,butusesFHRS'sample (N= 9,024)anddataon
Hadley2 climate change
predictions.Row3 isthecorrected DGestimatesderived underHadley2 andthat allows
theeffects
oftheweather variables tovaryacross andnonirrigated
irrigated counties.Row4 isthecorrected DG
estimates
derived underHadley 2 andthatrestrict
theeffectsoftheweather variables
across andnonirri-
irrigated
gatedcountiestobethesame, assupported bytheevidence inTable1.Row5 isthecorrected DGestimates derived
underCCSM3A2,based onthesamespecification asrow4.

Table 804 reveals that the absolute value of the change in inventoriesaccounts
for only 1.4 percent of "total cash receipts frommarketings"duringthe period
1994-2003. In craftingDG, thisstatisticcaused us to conclude thatstoragewas not
a major factor.FHRS, however,show convincinglythatamong farmswithoutlive-
stock,thevalue of a year's productionexceeds sales in bountifulyears and is below
it in lean years.They argue thatthisinvalidatestheuse of annual profitdata to learn
about theimpactsof weatherrealizations.
In practice,thisdynamicinventoryadjustmentmeans that,forexample, a bushel
of cornthatis storedthisyear will be sold in a subsequentyear.Analogously,draw-
ing down inventoriesthisyearwill reduce thecrops available forsale in futureyears.
The pointis thatthefullimpactof a year's weatherrealizationon profitscan onlybe

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
3770 THEAMERICAN
ECONOMIC
REVIEW DECEMBER
2012

observedoverperiodslongerthana yearbecause inventory decisions allow farmers

to spreadtheimpactsover severalyears.20
The relevantissue forthispaper is thatthiscompensatorybehaviorsuggeststhat
it may be importantto consideran alternativeversionof equation (1) to capturethe
fullimpactof a year's weatherrealization.Specifically,theideal empiricalapproach
in the presence of farmers'inventorymanagementis a distributedlag model that
relatesannual profits,whichincludes revenuesfromsales of productsregardlessof
theirdate of production,to currentand lagged weathervariables.The inclusionof
lagged weathervariablesand calculationof thecumulativeeffectof weathershocks
is necessarybecause thequantitiesthata farmersells in a given yearare affectedby
the fullhistoryof weatherrealizationsthroughher storagedecisions. This solution
is not especially novel; indeed, Stock and Watson (2003) motivatestheirdiscus-
sion of distributedlag models withthe example thata year's weathershock affects
orangejuice marketsover severalperiods.
Table 3 presentsthe estimatesfromsimple distributedlag models thatallow for
a dynamicrelationshipbetweenobservedfarmprofitsand a year's weatherrealiza-
tion. The resultsare reportedin the followingorder: (i) predictedimpacts associ-
ated withthecontemporaneousweathervariables; (ii) thelagged weathervariables;
and (iii) the cumulativeimpact (i.e., the sum). In practice,we include a single lag
because we have a relativelyshortpanel and thisapproachis demandingof thedata.
There are several importantfindingsin Table 3 thatare evidentin both panels A
and B. First,the inclusion of the lagged weathervariables leaves the impact of the
contemporaneousweathervariables largelyunchangedin the contextof theirstan-
darderrors(to see this,comparethe (i) estimateswiththeestimatesin rows 4 and 5
of Table 2). Second, the impact of the lagged weathervariables tends to have the
opposite sign of the impact of the contemporaneousvariables althoughit is usu-
ally of a smaller magnitude.This findingis consistentwith economic theorythat
predictsthatthe full impact of a year's weatherrealizationmay be apparentover
periodslongerthana year due to farmers'use of inventorymanagementto smooth
shocks to income. Third,it is apparentthatincludinglags reduces the precisionof
theestimatedoverall effect.
In thecontextof makingprojectionsabout climatechange,thekey findingis that
models thataccountforlagged weathergenerallypredictsmallercumulativelosses
than the ones that account only for contemporaneousweather.For example, the
weightedaverage (again using the inverseof the standarderrorsas the weight) of
the Hadley 2 baseline model's estimatesin panel A is -$1.3 billion, compared to
-$4.5 billionfromthemodel thatjust includescontemporaneousweathervariables.
In thecase of theCCSM 3 A2 estimates,thecorrespondingestimatesare -$3.4 bil-
lion and -$9.9 billion fromthemodels withand withoutthelag, respectively.

20Thestructural between
relationship decisions
storage realizations
andweather isa complicated that
process
dependsonseveralfactors theweather
including realization's impact
expected oncurrentandfuturecropprices,
thelength
costs,
storage oftimebefore without
a cropcanbestored spoiling corn
(stored andsoybeans canspoil
duetomolds andinsects),andtheinterest
rate. is a fascinating
ofthisbehavior
examination
A careful topicfor
research a long
butrequires paneldataset
with onfarmer
information
detailed Suchananalysis
behavior. isbeyond
thescopeofthis
note.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
VOL.102NO.7 ANDGREENSTONE:
DESCHNES OFCUMATE
ECONOMICS REPLY3771
CHANGE:

Table3- Estimates
ofthePredicted
ImpactofClimate
Change onUSAggregate Farm for
Profits,
Average
Yearover2070-2099, of2002Dollars,inModelsthatAllow
Billions
forLagged ofWeather
Effects
(la) (lb) (lc) (Id)
A.Predicted inaverage
impact yearover
2070-2099
under 2
Hadley
Modelwith weather
lagged (1Lag)
ofcontemporaneous
(i)Impact weather -8.4 -6.5 -4.0 -0.9
(2.1) (3.1) (2.4) (3.6)
oflagged
(ii)Impact weather 7.3 4.8 4.1 -2.8
(2.6) (2.3) (2.4) (3.8)
(iii)Cumulative ofcontemporaneous
impact andlagged
weather - 1.1 - 1.7 0.1 - 3.7
(2.8) (3.6) (3.2) (5.7)
B.Predicted inaverage
impact yearover
2070-2099 CCSM3A2
under
Model with weather
lagged (1Lag)
ofcontemporaneous
(i) Impact weather -16.7 -15.8 -10.2 -4.3
(4.5) (6.9) (5.2) (7.9)
oflagged
(ii)Impact weather 14.1 11.1 9.8 -4.3
(5.8) (5.4) (5.5) (8.4)
(iii)Cumulative ofcontemporaneous
impact andlagged
weather -2.6 -4.7 -0.3 -8.6
(5.9) (7.5) (6.4) (11.7)
Yeareffects Yes No No No
USDAregion x year
effects No Yes No No
Census x year
division effects No No Yes No
x year
State effects No No No Yes
Notes:Figuresareinbillions
of2002constantdollars.
Eachcounty's predicted
impactiscalculated
asthediscrete
difference
inper-acre atthecounty's
profits predicteddegree-days andprecipitation
after
climate
change (aver-
agedoverthe2070-2099 anditscurrent
period) climate(i.e.,theaverage
over
the1970-2000 Theresult-
period).
ingchange inper-acre ismultiplied
profits bythenumber ofacres offarmland
inthecountyandthenthenational
effect
isobtainedbysumming acrossall2,342countiesinthe"REPLY" Thesamecalculation
sample. isapplied
tocontemporaneous andlaggedweathervariables.
Average annual aggregate inthe2,342counties
profits inthe
sampleareUS$(2002) 32.8billion.
Standarderrorsareclusteredatthecounty
level.
Seethetext
formore details.

IV. Conclusions

FHRS (2012) have uncoveredcoding and data errorsin our paper (DG 2007). We
are embarrassedby these mistakesand gratefulto FHRS fordiscoveringthemand
advancingknowledgeon thisimportantissue.
Our reanalysis of agriculturalprofitswithcorrecteddata presentedhere and in
the online Appendix leads to threeprimaryfindings.First,contraryto the results
in DG (2007), the correcteddata suggest thatan immediate shiftto the projected
end-of-the-century climate would reduce agriculturalprofits.This impact is larger
when projectionsfrommore recentclimate models are used and smallerin econo-
metric models that allow for local shocks to input and output prices and pro-
ductivity.Second, the PDV over the remainderof the centuryof the projected
impacts froma recentclimate model is roughly$164 billion, or about 5 years of
currentannual profits.This estimateis likely to overestimatethe loss, because it
fails to allow forany technological advances or adaptation in response to higher
temperatures.Third, the estimated losses are more than 50 percent smaller than
those fromthe standardapproach and generally statisticallyinsignificantwhen
one uses a textbookdistributedlag model and calculates the dynamic cumulative

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
3772 THEAMERICAN REVIEW
ECONOMIC 2012
DECEMBER

effectsthat account for farmers'dynamic inventoryadjustmentsin response to

temperaturerealizations.21
The meaningof theseresultslies in theeyes of thebeholder.For whatit is worth,
we believe thattheseresultsfailto make a convincingcase forlargenegativeimpacts
of climatechange on aggregateprofitsin theUS agriculturalsector.In thisrespect,
conclusionthansome of FHRS's important
theylead to a different workon thelikely
impacts of climate change on the of
yields particularcrops (see also Schlenkerand
Roberts2009). This difference may simply reflect
the differencebetween cropyields
and profitsas outcomes,wherethelatteris more amenable to adaptationeven in the
shortrun,and providesa morecompleteindicatorof productivity in theagricultural
sectorthancrop yields alone. In contrast,recentresearchsuggeststhattheremay be
substantialnegativeimpactson agricultureand healthin poorercountries,especially
thosewithalreadyintemperateclimates(Guiteras2009; Burgess et al. 201 1). Much
uncertainty remains,however,about thelikelyeconomic impactsof climatechange,
and so further researchis necessary.

REFERENCES

Burgess,Robin,Olivier Deschnes,DavidDonaldson, andMichaelGreenstone. 2011."Weatherand

DeathinIndia:Mechanisms andImplications forClimate Change." Unpublished.
Conley,T. G. 1999."GMMEstimation withCrossSectional Dependence. JournalofEconometrics
92(1): 1-45.
Deschnes, Olivier,andMichaelGreenstone. 2007a. TheEconomic ImpactsofClimateChange:Evi-
dencefrom Agricultural
OutputandRandom FluctuationsinWeather."American EconomicReview
97 (1): 354-85.
Deschnes, Olivier,andMichaelGreenstone. 2007b.The Economic ImpactsofClimateChange:Evi-
dencefrom AgriculturalOutput andRandom FluctuationsinWeather:Dataset"American Eco-
nomic Review. http://www.aeaweb.org/articles.php?doi=10. 1.354.
1257/aer.97.
Deschnes, Olivier,andMichaelGreenstone. 2011."Climate Change,Mortality, Evi-
andAdaptation:
dencefrom Annual inWeather
Fluctuations intheUS."American Economic Journal:
AppliedEco-
nomics 3 (4): 152-85.
Deschnes, Olivier,andMichaelGreenstone. 2012."TheEconomic ImpactsofClimateChange:Evi-
dencefrom AgriculturalOutputandRandom inWeather:
Fluctuations Reply:Dataset."
American
Economic Review. http://dx.doi.Org/10.1257/aer.102.7.3761.
Deschnes, Olivier, MichaelGreenstone,andJonathan Guryan. 2009."ClimateChangeandBirth
Weight." American EconomicReview 99 (2): 21 1-17.
Fisher,
Anthony C.,W.MichaelHanemann, MichaelJ.Roberts, andWolfram 2012. The
Schlenker.
Economic Impacts of Climate
Change: Evidence from Output
Agricultural and Random Fluctua-
tionsinWeather: Comment."American Economic Review.102(7): 3749-60.
Gibson,W.P.,C. Daly,T.Kittel,D. Nychka,C. Johns, N.Rosenbloom, A.McNab,andG.Taylor. 2002.
"Development ofa 103-year
high-resolution climatedatasetfortheconterminous UnitedStates."
In Proc.y13thAMS Conf.on AppliedClimatology , Amer.MeteorologicalSoc.,Portland,OR,
May 13-16,181-83.http://www.prism.oregonstate.edu/pub/prism/docs/appclim02-103yr
dataset-gibson.pdf.
Guiteras,Raymond. 2009."TheImpact ofClimate ChangeonIndian Unpublished.
Agriculture."
NationalCenter Research.
forAtmospheric 2007."Community ClimateSystemModel."http://www.
cgd.ucar.edu/csm/.

21Further,
animportant that
setofresults totheReply's
isrelegated Appendix
supplementary account
takes ofa
recent that
literature emphasizesthat
themeaningful
impactsofclimate
change atthe
tobeconcentrated
arelikely
tomodel
andsoitisimportant
temperatures
highest temperatureinways allowfornonlinearities
that (Deschnes
andGreenstone andGuryan
Greenstone,
2007;Deschnes, andRoberts
2009;Schlenker etal.2011).
2009;Burgess
Inourview,
future inthis
research areashould from
depart polynomial for
models degree-days models
andconsider
that
account forhigher
jointly degrees andfortheimpact
ofnonlinearity ofdaily
asopposedtopurelyseasonal
inweather.
variability

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions
VOL.102NO.7 ANDGREENSTONE:
DESCHNES ECONOMICS
OFCLIMATE REPLY3773
CHANGE:

National
AgriculturalStatistical
Service.1987-2002."CensusofAgriculture."
US Departmentof
Agriculture,
http://www.agcensus.usda.gov/index.php.
National
ClimaticDataCenter. 1968-2002."SurfaceLandDailyCooperative
Summary oftheDay."
http://cdo.ncdc.noaa.gov/pls/plclimprod.
Schlenker,
Wolfram, andMichaelJ.Roberts. 2009."Nonlinear Effects
Temperature Indicate
Severe
DamagestoUS CropYieldsunder ClimateChange." oftheNational
Proceedings Academy ofSci-
ences106(37): 15594-98.
James
Stock, H.,andMarkW.Watson. 2003.IntroductiontoEconometrics.
Boston:
Addison-Wesley.

This content downloaded from 132.174.255.116 on Sat, 14 Nov 2015 00:12:28 UTC
All use subject to JSTOR Terms and Conditions