Nonlinear Temperature Effects Indicate Severe Damages to

U.S. Crop Yields under Climate Change

Wolfram Schlenker and Michael J. Roberts

Supporting Information Appendix

Contents
1 Data Appendix 1

2 Regression Models 6

3 Regression Results and Significance Levels 9

4 Out-of-Sample Prediction Accuracy 10

5 Nonlinearities and Climate Change Impacts 14

6 Cross-section versus Time Series 15

7 Geographic Subregions 17
7.1 North versus South . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
7.2 East versus West . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

8 Temporal Subsets 21

9 Temperature - Precipitation Interaction 23

10 Growing Season and Time Separability 24

11 Planted versus Harvested Area 31

12 Year Fixed Effects 33

13 Linear Within-Day Temperature Interpolation 35

14 Uniform Climate Change Scenarios 37

1 Data Appendix
Dependent Variable
Yields for corn, soybeans, and cotton for the years 1950-2005 are reported by the U.S.
Department of Agriculture’s National Agricultural Statistical Service (USDA-NASS). We
include all reported yields, even though some appear spuriously low. These outliers are
few. If we drop outliers, results do not change, but the cutoff point for omission becomes
somewhat arbitrary, so we include them. The descriptive statistics are:

Table A1: Descriptive Statistics: Crop Yields

Crop Mean Minimum Maximum Std. Dev.

Corn (Bushels/Acre) 75.60 0.04 203.49 35.52
Soybeans (Bushels/Acre) 27.00 1.80 249.39 9.35
Cotton (Pounds/Acre) 451.48 8.00 3433.00 234.86

Yields equal total county-level production divided by harvested acres. For about 80%
of the observations, NASS reports planted acres in addition to harvested acres. Because
planted acres may not be harvested in a low-yielding year, yield per harvested acre may
cause bias. As a sensitivity check, we derived an alternative yield measure by taking total
production over total acres planted, with results reported in Section 11 below. The results
are very similar to the results using harvested acres. Since the area planted is not reported
in all areas and years, our analysis focuses on the larger sample of output per acre harvested
that is the standard USDA definition of yield.
In the main paper we report results for corn and soybean yields in counties east of the 100
degree meridian. We do this because counties west of the 100 degree meridian are supported
by large and heavily subsidized irrigation systems. For cotton the main paper reports results
for all cotton yields, since most cotton (58%) is irrigated and the sample of counties is much
smaller. This gives us 105981 observations with corn yields, 82385 observations with soybeans
yields, and 31540 observations with cotton yields.
Given the potential importance of irrigation and particularly how irrigation might dif-
ferentially affect yield outcomes in eastern and wester regions, Section 7 reports separate
results for counties east and west of the 100th meridian for all three crops. Results differ
substantially for corn and soybeans, but to a lesser degree for cotton, and we hence pool
observations from the East and West for the latter.

1
Past Weather
Earlier statistical studies use average temperature over an entire season or months, which
can hide extreme temperatures that occur within a month or even during a fraction of a day.
Our fine-scale weather aids identification of nonlinear effects which are diluted when weather
outcomes are averaged over time or space. Construction of these data is briefly described
here and in more detail in [1].
The basic steps are as follows. We first develop daily predictions of minimum and max-
imum temperature on a 2.5x2.5mile grid for the entire United States. We then derive the
time a crop is exposed to each 1 degree Celsius interval in each grid cell. These predictions
are merged with a satellite scan that allows us to select only those grid cells with cropland.
We then aggregate the whole distribution of outcomes for all days in the growing season in
each county. Since our study emphasizes nonlinearities, it is important to derive the time
each grid cell is exposed to each 1 degree Celsius interval before aggregating to obtain the
county-level distribution in each growing season. This preserves within-county variation in
temperatures in each year.1
We obtain geographic specificity of weather outcomes using the Parameter-elevation Re-
gressions on Independent Slopes Model (PRISM), widely regarded as one of the best ge-
ographic interpolation procedures ([Link] PRISM accounts for
elevation and prevailing winds to predict weather outcomes on 2.5x2.5 mile grid across the
contiguous United States. The PRISM data, however, are reported on a monthly time scale.
We therefore combine the advantages of the PRISM model (good spatial interpolation) with
better temporal coverage of individual weather stations (daily instead of monthly values).
We do this by pairing each of the 259,287 PRISM grid cells that cover agricultural area in
a LandSat satellite scan with the closest seven weather stations having a continuous record
of daily observations.2 We then estimate a separate regression for each grid cell, where the
dependent variable is the monthly PRISM grid cell estimate and the explanatory variables
1
An alternative method to approximate the distribution of daily temperatures from the distribution of
average monthly temperatures is developed by [2]. This method appears appropriate for predicting the
average frequency that a certain weather outcome will be realized, but less appropriate in predicting a
specific frequency of a weather outcome in a particular year. As a result, these methods work well in a
forward-looking cross-sectional analysis where the dependent variable is tied to weather expected outcomes
rather than realized outcomes (for example, the link between land values and climate), but less well in our
analysis where the dependent variable (yield) linked to specific weather outcomes. Obtaining daily values
on a small scale requires a spatial interpolation procedure to approximate daily weather outcomes between
individual weather stations.
2
Stations have a continuous record if at least 90 percent of all months have at most 3 missing observations.
Missing values are filled in by regressing the daily observations at a station on the seven closest weather
stations including half-month fixed effects.

2
are the monthly averages at each of the seven closest weather stations, plus fixed effects
for each month. The R-squares are usually in excess of 0.999. The estimated relationship
between monthly PRISM grid cell averages and monthly averages at each of the seven closest
stations is then used to predict daily records at each PRISM grid cell from the daily records
at the seven closest weather stations.
A cross-validation exercise is used to test the accuracy of the daily weather predictions.
For this exercise we construct a daily weather record at each PRISM cell that harbors a
weather station without using that weather station in the interpolation procedure. We then
compare predicted daily outcomes to actual outcomes within all PRISM cells that harbor
weather stations. The mean absolute error is 1.36◦ C for minimum temperature and 1.49◦ C
for maximum temperature. Due to the law of large numbers, our county-level distribution
estimates contain much less error because they average errors over all grid cells in each
county and all days of the growing season.
The distribution of temperatures within each day is approximated using a sinusoidal
curve between predicted minimum and predicted maximum temperatures [3]. In a sensitivity
check, reported in Section 13, we instead use a linear interpolation between minimum and
maximum temperature. Both methods give similar results. Using the sinusoidal or linear
interpolation we then estimate time in each 1◦ C-degree temperature interval between −5◦ C
and +50◦ C. Finally, we construct the area-weighted average time at each degree over all
PRISM grid cells in a county. The agricultural area in each cell was obtained from LandSat
satellite images.3 The weather variables are summed over the six-month period from March
through August for corn and soybeans, and the seven-month period April through October
for cotton.
Boxplots in Figure A1 show historical temperature distributions and how they vary over
time and across counties. Whiskers indicate the minimum and maximum exposure to a each
temperature range across all counties. The box marks the 25%-75% range, and the middle
line within each box is the median across all counties and years.

Climate Change Scenarios

Climate change predictions are drawn from the Hadley III model.4 It is one of the models that
form the basis for the report by the Intergovernmental Panel on Climate Change (IPCC).
3
Vince Breneman and Shawn Bucholtz at the Economic Research Service provided us with the agricultural
area in each PRISM grid cell. Since we use the LandSat scan of a given year, we are not able to pick up
shifts in growing regions.
4
[Link]

3
 Figure A1: Descriptive Weather Statistics
Corn − Average Climate Soybeans − Average Climate Cotton − Average Climate
20 20 20
Exposure During Growing Season (Days)

Exposure During Growing Season (Days)

Exposure During Growing Season (Days)

18 18 18

16 16 16

14 14 14

12 12 12

10 10 10

8 8 8

6 6 6

4 4 4

2 2 2

0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs show distributions of time each crop was exposed to each 1◦ C interval during the 184-
day growing season March-August for corn and soybeans and the 214-day growing season April-October
for cotton. The lowest interval has no lower bound and includes the time temperatures fall below 0◦ C.
The highest interval has no upper bound and includes the time temperatures are above 39◦ C. For each
temperature interval, the range between minimum and maximum time across counties and years is shown by
whiskers, the 25%-75% percentile range is outlined by a box, and the median is added as a solid bold line.

We obtain monthly model output for both minimum and maximum temperatures under four
major emissions scenarios (A1FI, A2, B1, and B2) for the years 1960-2099. Each emission
scenario rests on different assumptions about population growth and use of alternative fuels,
among other factors [4]. The B1 scenario assumes CO2 emissions soon decline and therefore
results in the slowest rate of warming over the next century. The A1FI assumes continued
use of fossil fuels and the largest increase in CO2 -concentrations and temperatures. The
other two scenarios are between these extremes.
Predicted weather under climate change is derived as follows. At each of 216 Hadley
grid nodes covering the United States we find the predicted difference in monthly mean
temperature for 2020-2049 (medium-term), 2070-2099 (long-term), and historic averages
(1960-1989). Next, predicted changes in monthly minimum and maximum temperature at
each 2.5x2.5 mile PRISM grid are calculated as the weighted average of the monthly mean
change in the four surrounding Hadley grid points, where the weights are proportional to
the inverse squared distance and forced to sum to one. In a final step, we add the predicted
absolute changes in monthly minimum and maximum temperatures at each PRISM grid to
observed daily time series from 1960 to 1989. In other words, we shift the entire historical
distribution of each month in each grid cell by the predicted mean monthly change of each
climate scenario. An analogous approach was used for precipitation, except that we use the
relative ratio of future predicted rainfall to historic rainfall instead of absolute changes. Each

4
 Figure A2: Climate Change Predictions under the Hadley HCM3 Model
Corn: Hadley III − B1 Scenario (2020−2049) Corn: Hadley III − A1FI Scenario (2020−2049) Corn: Hadley III − B1 Scenario (2070−2099) Corn: Hadley III − A1FI Scenario (2070−2099)

20 20 20 20

15 15 15 15
Change in Exposure (Days)

Change in Exposure (Days)

Change in Exposure (Days)

Change in Exposure (Days)

10 10 10 10

5 5 5 5

0 0 0 0

−5 −5 −5 −5

−10 −10 −10 −10

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Soybeans: Hadley III − B1 Scenario (2020−2049) Soybeans: Hadley III − A1FI Scenario (2020−2049) Soybeans: Hadley III − B1 Scenario (2070−2099) Soybeans: Hadley III − A1FI Scenario (2070−2099)

20 20 20 20

15 15 15 15
Change in Exposure (Days)

Change in Exposure (Days)

Change in Exposure (Days)

Change in Exposure (Days)

10 10 10 10

5 5 5 5

0 0 0 0

−5 −5 −5 −5

−10 −10 −10 −10

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Cotton: Hadley III − B1 Scenario (2020−2049) Cotton: Hadley III − A1FI Scenario (2020−2049) Cotton: Hadley III − B1 Scenario (2070−2099) Cotton: Hadley III − A1FI Scenario (2070−2099)

20 20 20 20

15 15 15 15
Change in Exposure (Days)

Change in Exposure (Days)

Change in Exposure (Days)

Change in Exposure (Days)

10 10 10 10

5 5 5 5

0 0 0 0

−5 −5 −5 −5

−10 −10 −10 −10

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs show the distribution of predicted temperature changes for each crop and each 1◦ C interval
during the March-through-August 184-day growing season for corn and soybeans and the April-through-
October 214-day growing season for cotton. The lowest interval has no lower bound and includes the change
in time that temperatures fall below 0◦ C. The highest temperature interval has no upper bound and includes
the change in time that temperatures are above 39◦ C. For each temperature interval, whiskers show the range
of duration changes across counties, the box outlines the 25%-75% percentile of changes, and the solid bold
line inside each box shows the median change. The top row displays predicted changes for corn-growing
counties, the middle row shows soybeans, and the bottom row shows cotton. The first two columns show
slow-warming (B1) and fast-warming (A1FI) predictions for the medium term (2020-2049) and last two
columns show slow- and fast-warming predictions for the long-term (2070-2099).

county’s weather outcomes are the area-weighted averages of all PRISM grids.
Figure A2 shows the shift in the temperature distribution under the B1 and A1FI scenar-
ios in the medium-term (2020-2049) and long-term (2070-2099) for each of the three crops
in our data. Each figure shows a series of box plots, one for each degree Celsius. Each box-
plot summarizes the predicted change in the frequency of that specific temperature across
all counties growing that crop. Temperatures below 22◦ C generally become less frequent in
corn and soybeans counties and temperatures below 25◦ C become less frequent in cotton

5
counties. Temperatures above these levels generally become more frequent.

2 Regression Models
The regression models assume temperature effects on yields are cumulative over time and
that yield is proportional to total exposure. This implies the temperature effect on yield
growth rates are the same and additively substitutable over time. The empirical validity of
this assumption is considered in more detail below in Section 10. Specifically, yield growth
g(h) depends nonlinearly on heat h such that log yield, yit , in county i and year t is
 h
yit = g(h)φit (h)dh + zit δ + ci + it (A1)
h

where φit (h) is the time distribution of heat over the growing season in county i and year t.
We fix the growing season to months March through August for corn and soybeans and the
months April through October for cotton. Observed temperatures during this time period
range between the lower bound h and the upper bound h. Other control factors are denoted
zit and include a quadratic in total precipitation as well as a quadratic time trend for each
state to capture technological change. A time-invariant county fixed effect ci is to control
for heterogeneity, such as soil type and quality. We allow the error terms it to be spatially
correlated using the non-parametric routine by [5].
Although time separability is partially rooted in agronomy, we implicitly validate this
assumption by showing a statistically significant relationship between the cumulative dis-
tribution of temperatures and yields. We would not observe this if time separability were
not appropriate, because random pairing of various temperatures over a season and between
years would not provide clear identification. In Section 10 we consider the assumption of
time-additivity of temperature effects in more detail. These alternative models estimate a
separate growth function g(h) for various sub-periods of the growing season. In one alterna-
tive specification we split the six-month growing season into three two-month intervals and
estimate a growth function for each one of them by omitting the remaining four months. In
a second specification we jointly estimate a separate growth function for each month of the
growing season. In a third specification we test for whether the temperature response func-
tion is different July than it is in other months. These considerably more flexible alternative
models do little to improve the overall fit of the model and give similar predictions under
climate-change scenarios.

6
 A special case of time-separable growth is the concept of growing degree days, typically
defined as the sum of truncated degrees between two bounds. For example, bounds of 8◦ C
and 32◦ C for ”beneficial heat” have been suggested by [6]. A day of 9◦ C hence contributes
1 degree day, a day of 10◦ C contributes 2 degree days, up to a temperature of 32◦ C, which
contributes 24 degree days. All temperatures above 32◦ C also contribute 24 degree days.
Degree days are then summed over the entire season. These particular bounds have been
implemented in a cross-sectional analysis by [7]. Thus, growing degree days are the special
case of our model where (using the above bounds as an example)
⎧
⎪
⎪ if h ≤ 8
⎪
⎨
0
g(h) = h − 8 if 8 < h < 32
⎪
⎪
⎪
⎩24 if 32 ≤ h

The appropriate bounds for growing degree days are still debated, partly because earlier
studies use a limited number of observations from field experiments to identify them. There
is also uncertainty about temperature effects above the upper bound. Degree days above
34◦ C are sometimes included as a separate variable and speculated to be harmful.
Using data on exposure to each 1-degree Celsius temperature interval, we approximate
the above integral with


49
yit = g(h + 0.5)[Φit (h + 1) − Φit (h)] + zit δ + ci + it (A2)
h=−5

where Φit (h) is the cumulative distribution function of heat in county i and year t.
We consider three specifications for g(h). First, we approximate g(h) using dummy vari-
ables for each three-degree temperature interval. We obtain similar results when estimating
more flexible models with dummy variables for each one-degree interval. We report results
for three-degree intervals in order to make figures easier to interpret. This step function ef-
fectively regresses yield on season-total time within each temperature interval. Temperatures
above 39◦ C (102 degrees Fahrenheit) occur less frequently and we therefore lump all time a
plant is exposed to a temperature above 39◦C into one category. Similarly, we lump all time
where temperature is below freezing into the interval [−1; 0]. The historical temperature

7
distribution is displayed in Figure A1. The model becomes5


39
yit = γj [Φit (h + 3) − Φit (h)] +zit δ + ci + it . (A3)

j=0,3,6,9,... xit,j

The second specification assumes g(h) is an m-th order Chebychev polynomial of the
form g(h) = m j=1 γj Tj (h), where Tj () is the j − th order Chebyshev polynomial. Chebyshev
polynomials are a relatively parsimonious approximation for the function g(h), assuming it
is smooth. By interchanging the sum we obtain


39 
m
yit = γj Tj (h + 0.5) [Φit (h + 1) − Φit (h)] + zit δ + ci + it
h=−1 j=1


m 
39
= γj Tj (h + 0.5) [Φit (h + 1) − Φit (h)] +zit δ + ci + it (A4)
j=1 h=−1

xit,j

where xij,t is the exogenous variable obtained by summing the j − th Chebyshev polyno-
mial evaluated at each temperature interval midpoint, multiplied by the time spent in each
temperature interval. Successively higher-order polynomials were estimated until the rela-
tionship appeared stable.
The third specification models g(h) as piecewise linear function, which is similar to the
concept of growing degree days. In this case growth increases linearly up to an endoge-
nous threshold and then decreases linearly above the threshold. We loop over all possible
thresholds, estimate the least-squares segment slopes for each one, and pick the threshold
and segment slopes with the best fit.
The predicted change in production when the average weather variables change from
the 1960-1989 average w̄i0 to the new values w̄i1 is using the regression coefficients on the
weather variables β w :6
N
i=1 ai ew̄i1 βw +c¯i
impact = N
−1
i=1 ai ew̄i0 βw +c¯i
where ai is the average growing area for counties i = 1...N and ci is the county fixed effect.
5
The omitted category, captured by the model’s intercept, is the time temperature is below 0◦ C.
6
For some sensitivity checks, we use regression coefficients estimated from one subset of data and use
them to make climate-change predictions for another subset of data. For example, in one sensitivity check
we use southern counties to estimate β w and then use these coefficient estimates to predict climate-change
impacts for all counties east of the 100 degree meridian.

8
It should be noted that the c̄i act like another scale factor besides the area ai . If we omit
them, the results differ by at most 1%, so the reweighing has a minor impact on the results.

3 Regression Results and Significance Levels

Regression results are displayed in Figure 1 in the main paper. Here we replicate the analysis
and present each of the three model specifications as a separate row in Figure A3. We also

Figure A3: Nonlinear Relation Between Temperature and Yields With Confidence Bands
Corn Soybeans Cotton
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Point Estimate Point Estimate Point Estimate
95% Confidence Band 95% Confidence Band 95% Confidence Band
99% Confidence Band 99% Confidence Band 99% Confidence Band
99.5% Confidence Band 99.5% Confidence Band 99.5% Confidence Band
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn Soybeans Cotton
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Point Estimate Point Estimate Point Estimate
95% Confidence Band 95% Confidence Band 95% Confidence Band
99% Confidence Band 99% Confidence Band 99% Confidence Band
99.5% Confidence Band 99.5% Confidence Band 99.5% Confidence Band
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn Soybeans Cotton
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Point Estimate Point Estimate Point Estimate
95% Confidence Band 95% Confidence Band 95% Confidence Band
99% Confidence Band 99% Confidence Band 99% Confidence Band
99.5% Confidence Band 99.5% Confidence Band 99.5% Confidence Band
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs display changes in log yield if the crop is exposed for one day to a particular 1◦ C temperature
interval where we sum the fraction of a day temperatures fall within each interval. The 95%, 99%, and 99.5%
confidence bands, after adjusting for spatial correlation, are added. Curves are centered so the exposure-
weighted impact is zero.

9
include three confidence bands given that our data set is rather large: 95%, 99% and 99.5%.
Note that even for the latter, the piecewise linear function shows a statistically significant
increase in temperature followed by a statistically significant decrease.

4 Out-of-Sample Prediction Accuracy

Tables A2-A4 report encompassing tests that compare our model’s predictions to others in
the literature. Comparisons are based on out-of-sample forecast accuracy. Each model is
estimated 1000 times, where each replication randomly selects 48 of the 56 years in our full
sample. Relative performance is measured according to the accuracy of each model’s predic-
tion for the omitted 8 years of the sample (about 14 percent). We sample years instead of
observations as year-to-year weather fluctuations are random, but there is considerable spa-
tial correlations across counties within a year.7 Models include our own three specifications
of temperature effects (step function, polynomial (8th order), and piecewise linear), a model
with average temperatures for each of four months [8], an approximation of growing-degree
days based on monthly average temperatures (Thom’s formula) used in [7], and a measure
of growing degree days that is calculated using daily mean temperatures used by [9].8 As a
baseline, we also report a model with county fixed effects and time trends but no weather
variables. Comparisons are made using the root-mean squared prediction error (RMS) and
Welch t-tests against the null hypothesis that the RMS in our 1000 replications is the same
for any comparison of two models.
Models that average temperatures over time or space have significantly inferior out-of-
sample predictions relative to our new three models. There is little difference in forecasting
ability between the step function, polynomial, and piecewise linear model, but large and
statistically significant differences between these three models and the other models. For
all three crops – corn, soybeans, and cotton – our three new specifications do not give
significantly different RMS, yet when paired with any of the other models, the reduction in
RMS is statistically significant at the 0.1% level.

7
We would like to thank a referee for pointing this out.
8
We use degree days bounds of each cited study, but the ranking of models does not change if instead we
were to use bounds suggested by this study.

10
 Table A2: Model Comparison Test for Out-Of-Sample Prediction Accuracy for Corn

RMS Welch Test for Equal Forecasting Accuracy (RMS)

Step Piecewise Monthly Degree Days Degree Days
Function Linear Averages (Thom) (Daily Mean)
Polynomial (8th-order) 0.2259 0.10 0.68 13.94 23.58 25.76
Step Function 0.2260 0.58 13.90 23.57 25.75
Piecewise Linear 0.2265 13.49 23.24 25.44
Monthly Averages 0.2387 9.22 12.02
Degree Days 8-32◦ C, >34◦ C (Thom) 0.2481 3.18

11
Degree Days 8-32◦ C (Daily Mean) 0.2516
County-Fixed Effects (No Weather) 0.2688

Notes: Table compares alternative temperature specifications for corn according to out-of sample prediction accuracy. The first column reports
the average root mean squared out-of sample prediction error (RMS) from 1000 replications. Rows are sorted from best forecast performance
(lowest average RMS) to worst. The last five columns present pair-wise Welch t-tests against the null hypothesis of equal RMS.
Step Function, Polynomial (8th order), and Piecewise Linear are the models developed in this paper; Monthly Averages uses a
quadratic specification in both average temperature and total precipitation for the months January, April, July, and October [8]; Degree Days
8-32◦ C, >34◦ C (Thom) uses Thom’s formula to estimate degree days (which are based on daily data) from monthly average temperature
data [7]; Degree Days 8-32◦ C (Daily Mean) first derive the average temperature for each day from daily temperature readings and then
construct degree days from this average [9];County fixed Effects (No weather) is a baseline model with county fixed effects and state-level
trends but no weather measures.
 Table A3: Model Comparison Test for Out-Of-Sample Prediction Accuracy for Soybeans

RMS Welch Test for Equal Forecasting Accuracy (RMS)

Step Piecewise Monthly Degree Days Degree Days
Function Linear Averages (Thom) (Daily Mean)
Polynomial (8th-order) 0.1962 0.78 0.89 22.62 23.86 30.75
Step Function 0.1966 0.10 21.86 23.07 30.04
Piecewise Linear 0.1967 21.96 23.19 30.17
Monthly Averages 0.2096 0.66 9.32
Degree Days 8-32◦ C, >34◦ C (Thom) 0.2100 8.86

12
Degree Days 8-32◦ C (Daily Mean) 0.2161
County-Fixed Effects (No Weather) 0.2300

Notes: Table compares alternative temperature specifications for soybeans according to out-of sample prediction accuracy. The first column
reports the average root mean squared out-of sample prediction error (RMS) from 1000 replications. Rows are sorted from best forecast
performance (lowest average RMS) to worst. The last five columns present pair-wise Welch t-tests against the null hypothesis of equal RMS.
Step Function, Polynomial (8th order), and Piecewise Linear are the models developed in this paper; Monthly Averages uses a
quadratic specification in both average temperature and total precipitation for the months January, April, July, and October [8]; Degree Days
8-32◦ C, >34◦ C (Thom) uses Thom’s formula to estimate degree days (which are based on daily data) from monthly average temperature
data [7]; Degree Days 8-32◦ C (Daily Mean) first derive the average temperature for each day from daily temperature readings and then
construct degree days from this average [9]; County fixed Effects (No weather) is a baseline model with county fixed effects and state-level
trends but no weather measures.
 Table A4: Model Comparison Test for Out-Of-Sample Prediction Accuracy for Cotton

RMS Welch Test for Equal Forecasting Accuracy (RMS)

Polynomial Step Degree Days Monthly Degree Days
(8th-order) Function (Thom) Averages (Daily Mean)
Piecewise Linear 0.3350 1.41 1.55 3.85 4.41 5.89
Polynomial (8th-order) 0.3379 0.14 2.35 2.87 4.33
Step Function 0.3382 2.20 2.72 4.17
Degree Days 8-32◦ C, >34◦ C (Thom) 0.3425 0.50 2.04
Monthly Averages 0.3434 1.57

13
Degree Days 8-32◦ C (Daily Mean) 0.3462
County-Fixed Effects (No Weather) 0.3484

Notes: Table compares alternative temperature specifications for cotton according to out-of sample prediction accuracy. The first column reports
the average root mean squared out-of sample prediction error (RMS) from 1000 replications. Rows are sorted from best forecast performance
(lowest average RMS) to worst. The last five columns present pair-wise Welch t-tests against the null hypothesis of equal RMS.
Step Function, Polynomial (8th order), and Piecewise Linear are the models developed in this paper; Monthly Averages uses a
quadratic specification in both average temperature and total precipitation for the months January, April, July, and October [8]; Degree Days
8-32◦ C, >34◦ C (Thom) uses Thom’s formula to estimate degree days (which are based on daily data) from monthly average temperature
data [7]; Degree Days 8-32◦ C (Daily Mean) first derive the average temperature for each day from daily temperature readings and then
construct degree days from this average [9]; County fixed Effects (No weather) is a baseline model with county fixed effects and state-level
trends but no weather measures.
5 Nonlinearities and Climate Change Impacts
The last section showed how a model that accounts for nonlinear temperature effects improves
model accuracy as measured by the out-of-sample prediction error. This section shows how
an account of nonlinear temperature effects influences predictions under projected climate
change scenarios.
Figure A4 replicates Figure 2 of the main paper except that it adds the results for a
linear model in average growing season temperature. All other controls (a quadratic in
precipitation, quadratic time trends by state, and county fixed effects) are the same. The
linear temperature term therefore should pick up the marginal effect at the sample mean.
Since predicted climate change can lead to very significant non-marginal changes, the linear
model can be misleading. The predicted climate change impacts in Figure A4 are much lower
and statistically significantly different. Accounting for the nonlinear effect of temperatures
has a big effect on predicted climate change impacts.

Figure A4: Predicted Climate Change Impacts on Crop Yields under the Hadley III Climate
Model
Corn Soybeans Cotton Corn Soybeans Cotton
40 40
B1 B2 A2 A1 B1 B2 A2 A1 B1 B2 A2 A1 B1 B2 A2 A1 B1 B2 A2 A1 B1 B2 A2 A1

20 20
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

0 0

−20 −20

−40 −40

−60 −60
Step Function
−80 Polynomial (8th−order) −80
Piecewise Linear
Average Growing Temperature
−100 −100

Notes: Graphs show predicted percentage changes in crop yields under four emissions scenarios. The left
panel shows predicted impacts in the medium-term (2020-2049) and the right panel shows the long-term
(2070-2099). A star indicates the point estimates and whiskers show the 95% confidence interval after
adjusting for spatial correlation. The color corresponds to the regression models in Figures 1 and 2 of the
main paper. A model using average temperature over the growing season is added in green.

14
6 Cross-section versus Time Series
A panel data analysis with county fixed effects is identified from time-series variation in
weather, which cannot account for grower adaptation to changes in climate. A clear ad-
vantage of the panel-data analysis is that year-to-year variations in weather are plausibly
random to farmers, so this approach embodies a viable natural experiment. In contrast, a
cross-sectional analysis compares yields and grower management choices across areas with
typically different weather (i.e., different climates). While this approach can accounts for
a wide array of grower adaptations, much like a hedonic model of land values, a poten-
tial problem with cross sectional analysis is confounding from omitted variables associated
with location and climate. Unlike year-to-year weather variations, climate variations are less
plausibly random. It is informative to consider both methods of identification and compare
them.
This comparison, replicated for each of the three crops, is presented in Figure A5. The
cross-sectional regression fits each county’s average deviation from the nationwide average
yield as a function of each county’s average temperature distribution. Since the cross-
sectional regression is most susceptible to omitted variables biases, we also consider a spec-
ification with soil controls as used in [7]. The time-series regression fits nationwide average
yield as a function of each year’s acre-weighed average temperature distribution. Another

Figure A5: Nonlinear Relation Between Temperature and Yields Using Various Sources of
Identification
Corn Soybeans Cotton
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Panel Panel Panel
Cross−section Cross−section Cross−section
Cross−section (Soil Controls) Cross−section (Soil Controls) Cross−section (Soil Controls)
Time Series Time Series Time Series
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs show changes in log yield if the crop is exposed for one day to a particular 1◦ C temperature
interval where within-day temperature variations are counted. Each graph shows results for the full panel,
pure cross-section, and pure time-series as the sources of identification. The fitted curves are centered so
the exposure-weighted impact is zero. Precipitation and other controls are included in estimation but not
reported.

15
regression uses the full panel. Since there are only 56 observations in the time series, com-
parisons are made using the piecewise linear model, which has just two degrees of freedom
for temperature.
The similarity of the estimated temperature growth functions is notable given the very
different variations used to identify them. It seems unlikely that omitted variables could
arise in a manner to indicate the same nonlinear relationship in the cross-section as the time
series. Holding the locations where crops are grown fixed, this gives a strong indication that
growers have historically been unable to easily adapt to warmer temperatures.
The cross-section and time-series models also make similar predictions under climate

Figure A6: Predicted Climate Change Impacts under the Hadley III Climate Model Using
Various Sources of Identification
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40

−60 −60
Panel
−80 Cross−section
−80
Cross−section (Soil Controls)
Time Series
−100 −100
Soybeans Soybeans
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40

−60 −60
Panel
−80 Cross−section
−80
Cross−section (Soil Controls)
Time Series
−100 −100
Cotton Cotton
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40

−60 −60
Panel
−80 Cross−section
−80
Cross−section (Soil Controls)
Time Series
−100 −100

Notes: Graphs show predicted percent changes in crop yields under four emissions scenarios. Left panels
display predicted impacts in the medium-term (2020-2049), right panels in the long-term (2070-2099). A
star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The line labeled “panel” replicates Figure 2 in the main paper. Lighter colors show predictions
from the pure cross-section and pure time-series coefficients.

16
change. These comparisons are shown in Figure A6, except that standard errors become
larger in the case of cotton. Again, this suggests adaptation possibilities have either been
too costly or have not been available. Farmers might, of course, change the crops they grow.
And future adaptation possibilities may differ from those in the past.

7 Geographic Subregions
7.1 North versus South
To test the sensitivity of the temperature-growth function to location, we divided the sample
into three mutually exclusive geographical regions: the most northern (and coolest) states,
the most southern (and warmest) states, and those in the middle.
If farmers can adapt to permanently higher temperatures by growing different varieties,
one would expect yields in the South would be less sensitive to extreme heat. We replicate

Figure A7: Geographic Subsets

17
Figure 1 from the main paper for each subregion. Each plot also includes the empirical
distribution of temperatures within each subregion as green histogram. These show how
much warmer southern counties are in comparison to the northern counties.
When moving to hotter counties (columns on the right), there are two countervailing
factors. First, the threshold where temperatures become harmful becomes slightly lower
in warmer areas (the threshold in the piecewise linear regression is 30◦ C in the northern
subsample, 29◦ C in the middle, and 28◦ C in the southern subsample. Second, the slope of
the decline seems to be less steep in the warmer south. It is an empirical question which of
the two effects will dominate. We hence conduct a policy experiment where we assume that
all farmers are growing the corn varieties from the South and face their response function.
We replicate the climate change impacts of Figure 2 in the main paper by evaluating the
impacts for the same counties, but alternatively using only counties in the South to identify
the regression coefficients. Predicted damages are approximately the same as shown in
Figure A9.

Figure A8: Nonlinear Relation Between Temperature and Yields For Various Geographic
Subregions
Corn − North Corn − Middle Corn − South
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel display changes in log yield if the crop is exposed for one day to
a particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each
interval. The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the
Polynomial regression. The three panels from left to right, respectively, only use northern, middle, and
southern counties in the estimation of the coefficients. Curves are centered so the exposure-weighted impact
is zero. Histograms at the bottom of each panel display the average temperature exposure among all counties
in the data.

18
 Figure A9: Predicted Climate Change Impacts under the Hadley III Climate Model
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (all obs)
−60 Step Function (South) −60
Polynomial (all obs)
−80 Polynomial (South)
−80
Piecewise Linear (all obs)
Piecewise Linear (South)
−100 −100

Notes: Graphs show predicted percent changes in crop yields under four emissions scenarios. The left panel
displays predicted impacts in the medium-term (2020-2049) and the right panel in the long-term (2070-2099).
A star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The darker colors replicate Figure 2 in the main paper. Corresponding lighter colors only use
the southern counties in the estimation of the regression coefficients but then evaluate the climate change
impacts for the same counties as in Figure 2 of the main paper.

7.2 East versus West

Our baseline model uses corn and soybean yields in counties east of the 100 degree meridian.
While some eastern states rely on irrigation (for example, 79 percent of the corn area in
Arkansas was irrigated in 2007)9, there is an essential difference between irrigation in the
East and West. The West relies on large subsidized irrigation infrastructures to capture and
carry large amounts of surface water, while the East relies predominantly on groundwater
irrigation systems. Because groundwater irrigation is a possible mitigation strategy, we want
to include such counties in the analysis, but not counties with irrigation systems that would
seem unlikely to be built in the future (due to both physical and political constraints).
Figure A10 shows the estimated relationship for counties east of the 100 degree meridian
(right column) and west of the 100 degree meridian (left column). While the effect of extreme
temperature (difference in height of the downward-sloping portion) is different in the East
and West for corn and soybeans, it is less so for cotton. We therefore pool all counties for
cotton, which is highly irrigated even in the East. The percent of soybeans, corn, and cotton
irrigated in the United States is 9%, 18%, and 58%, respectively.

9
Census of Agriculture, Volume 1, Chapter 2, Table 26.

19
Figure A10: Nonlinear Relation Between Temperature and Yields For Eastern and Western
United States
Corn − Western United States Corn − Eastern United States
0.025 0.025

0 0
Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025

−0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear

Exposure (Days)

Exposure (Days)
−0.075 −0.075
10 10
5 5
0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius)
Soybeans − Western United States Soybeans − Eastern United States
0.025 0.025

0 0
Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025

−0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)
−0.075 −0.075
10 10
5 5
0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius)
Cotton − Western United States Cotton − Eastern United States
0.025 0.025

0 0
Log Yield (Bales)

Log Yield (Bales)

−0.025 −0.025

−0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

−0.075 −0.075
10 10
5 5
0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to
a particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each
interval. The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the
Polynomial regression. The three rows use corn, soybeans, and cotton yields. The left column uses counties
west of the 100 degree meridian, while the right column uses counties east of the 100 degree meridian. Curves
are centered so the exposure-weighted impact is zero. Histograms at the bottom of each panel display the
average temperature exposure among all counties in the data.

20
8 Temporal Subsets
In this section we report regression results when the sample is split into two time periods
of equal length, 1950-1977 and 1978-2005. The purpose here is to explore whether relative
heat tolerance has changed over time as technological change boosted average yields by a
factor of two to three over the 56 year time period, depending on the crop. Although average
yields in the more recent panel are substantially greater than those in the earlier period, the
relative relationships between temperature and yield growth are similar.
Since plants have not become more heat tolerant over time in our sample, we find very
little difference in our predicted climate change impacts if the regression coefficients are
identified using only the years 1978-2005 rather than the full sample. This comparison is
shown in Figure A12. While we do not observe an increase in heat tolerance towards the

Figure A11: Nonlinear Relation Between Temperature and Yields for Various Temporal
Subsets
Corn (Years 1950−1977) Soybeans (Years 1950−1977) Cotton (Years 1950−1977)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn (Years 1978−2005) Soybeans (Years 1978−2005) Cotton (Years 1978−2005)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)

−0.075 −0.075 −0.075

10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to
a particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each
interval. The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the
Polynomial regression. The top row uses only the first half of our panel in the estimation of the coefficients
and the bottom row uses only the second half. Curves are centered so the exposure-weighted impact is zero.
Histograms at the bottom of each panel display the average temperature exposure among all counties in the
data.

21
end of our sample, a NSF funded study just completed a draft sequence of the corn genome,
which might make it easier to develop new corn varieties with greater heat tolerance [10].

Figure A12: Predicted Climate Change Impacts under the Hadley III Climate Model for
Various Temporal Subsets
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (all obs)
−60 Step Function (1978−2005) −60
Polynomial (all obs)
−80 Polynomial (1978−2005)
−80
Piecewise Linear (all obs)
Piecewise Linear (1978−2005)
−100 −100
Soybeans Soybeans
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (all obs)
−60 Step Function (1978−2005) −60
Polynomial (all obs)
−80 Polynomial (1978−2005)
−80
Piecewise Linear (all obs)
Piecewise Linear (1978−2005)
−100 −100
Cotton Cotton
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (all obs)
−60 Step Function (1978−2005) −60
Polynomial (all obs)
−80 Polynomial (1978−2005)
−80
Piecewise Linear (all obs)
Piecewise Linear (1978−2005)
−100 −100

Notes: Graphs show predicted percentage changes in crop yields under four emissions scenarios. Left panels
display predicted impacts in the medium-term (2020-2049), right panels in the long-term (2070-2099). A
star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The darker colors replicate Figure 2 in the main paper. Corresponding lighter colors only use
the years 1978-2005 counties in the estimation of the regression coefficients but then evaluate the climate
change impacts for the same counties as in Figure 2 of the main paper.

22
9 Temperature - Precipitation Interaction
All models we have considered so far include a quadratic specification of total growing-season
rainfall, but do not consider interactions between temperature and precipitation. Here we
report estimates from four regression models that divide the sample by quartiles of total
precipitation in June and July. The estimates for corn have a similar shape to that of the

Figure A13: Nonlinear Relation Between Temperature and Yields and Precipitation Inter-
actions
Corn − Step Function Soybeans − Step Function Cotton − Step Function
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Precipitation Quartile 1 Precipitation Quartile 1 Precipitation Quartile 1
Precipitation Quartile 2 Precipitation Quartile 2 Precipitation Quartile 2
Precipitation Quartile 3 Precipitation Quartile 3 Precipitation Quartile 3
Precipitation Quartile 4 Precipitation Quartile 4 Precipitation Quartile 4
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn − Polynomial (8th order) Soybeans − Polynomial (8th order) Cotton − Polynomial (8th order)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Precipitation Quartile 1 Precipitation Quartile 1 Precipitation Quartile 1
Precipitation Quartile 2 Precipitation Quartile 2 Precipitation Quartile 2
Precipitation Quartile 3 Precipitation Quartile 3 Precipitation Quartile 3
Precipitation Quartile 4 Precipitation Quartile 4 Precipitation Quartile 4
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn − Piecewise Linear Soybeans − Piecewise Linear Cotton − Piecewise Linear
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 −0.05 −0.05

Precipitation Quartile 1 Precipitation Quartile 1 Precipitation Quartile 1
Precipitation Quartile 2 Precipitation Quartile 2 Precipitation Quartile 2
Precipitation Quartile 3 Precipitation Quartile 3 Precipitation Quartile 3
Precipitation Quartile 4 Precipitation Quartile 4 Precipitation Quartile 4
−0.075 −0.075 −0.075
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs display changes in log yield if the crop is exposed for one day to a particular 1◦ C temperature
interval where we sum the fraction of a day temperatures fall within each interval. The sample is split into
four quartiles based on total precipitation in June and July. The 95% confidence band, after adjusting for
spatial correlation, are added as dotted lines for the step function as well as the polynomial. Curves are
centered so the exposure-weighted impact is zero.

23
pooled sample up to the critical temperature of 29◦C. The decline above the threshold,
however, is less steep for subsamples with greater precipitation. Note, however, how the
confidence band widens for higher temperatures.
The same interaction effect does not hold for soybeans.
There is some evidence that precipitation partly mitigates damages from extreme tem-
peratures, especially for corn. We did estimate models with richer interactions between
temperature and rainfall, but these models do not predict yields out-of-sample significantly
better than additively separable model reported above. It is possible that the relatively poor
predictive power of precipitation in comparison to temperature stems from greater measure-
ment error in the precipitation variable as spatial interpolation is more difficult. Since we
do not find a significant correlation between temperatures and rainfall in the raw daily data,
omitting temperature-rainfall interactions should not bias our predictions. That is, they
should give the unbiased estimates of the average effects of temperature and rainfall.

10 Growing Season and Time Separability

An important assumption of the empirical model is the additive separability of temperature
effects over the growing season. We fix the growing season to the months March through
August for corn and soybeans in the main paper, even though northern regions tend to plant
later than southern regions, and planting dates may vary from year to year depending on
weather conditions. We explore the sensitivity of the results to various definitions of the
growing season in Figure A14-Figure A16.
Each figure shows nine alternative specifications of the growing season together with
the baseline (top left for corn and soybeans, and center middle for cotton). The estimated
temperature effects appear similar regardless of how we shift the growing season. Columns
in the first two rows vary the start date (March in the first column, April in the second, and
May in the third) while rows vary the end date (for corn and soybeans: August in the top
row and September in the second row; for cotton: September in the top row and October
in the second row). The third row breaks the growing season into three two-month periods
and still obtains similar results by only including temperature readings from a two-month
period in the regression. This lends support to the assumption of additive separability. Note,
however, how the confidence band becomes much larger if we limit the sample to only the
first two months or the last two months. In case of the former, the temperature distribution
included very few hot temperatures as shown in the green exposure histogram at the bottom.

24
In case of the latter, the exposure to temperatures below 10◦ C is limited.
There is some discussion in the agronomic literature that the effect of temperatures can
vary over the season, especially during the corn flowering period [11, 12]. While the flowering
period varies by region, it mostly falls within the month of July. Figure A17 therefore
replicates our baseline models by including separate temperature variables for the months
of July, while all other months are pooled.10 The estimated relationship changes somewhat,
but in different directions for corn and soybeans. The F-statistic against the null hypothesis
that the 10 additional temperature variables for July are the same as the other months is
5.16 for corn, 6.54 for soybeans, and 4.10 for cotton. The statistics reject the null hypothesis
that temperature effects are the same. Despite statistical significance, however, the model fit
improves very little for both crops and the overall shape is still comparable. The R-square of
the corn model without any weather variables (just fixed effects and time trends) is 0.6624.
Adding the 14 temperature variables and two precipitation variables in our baseline model
(step function) increases the R-square to 0.7654. When we add 10 temperature variables for
July (omitting the bottom four categories due to lacking data), the R-square increases to just
0.7698.11 If we include a separate set of temperature variables for each of the six months,
adding an additional 70 temperature variables over our baseline model, the R-square again
only increases slightly to 0.7774. More importantly though, the predicted climate change
impacts are not statistically significantly different as shown in Figure A18.

10
We would like to thank a referee for suggesting this sensitivity check. Since temperatures hardly ever
drop below 10◦ C in July, we exclude these categories for July.
11
The analogous numbers for soybeans are 0.4792 (no weather), 0.6253 (baseline model), 0.6322 (separate
July variables added); and for cotton they are 0.2954 (no weather), 0.3728 (baseline model), 0.3866 (separate
July variables added).

25
Figure A14: Nonlinear Relation Between Temperature and Yields for Various Growing Sea-
sons for Corn
Corn (March − August) Corn (April − August) Corn (May − August)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn (March − September) Corn (April − September) Corn (May − September)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn (March − April) Corn (May − June) Corn (July − August)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to a
particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each interval.
The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the Polynomial
regression. Each plot uses a different definition of the growing season. Curves are centered so the exposure-
weighted impact is zero. Histograms at the bottom of each panel display the average temperature exposure
among all counties in the data.

26
Figure A15: Nonlinear Relation Between Temperature and Yields for Various Growing Sea-
sons for Soybeans
Soybeans (March − August) Soybeans (April − August) Soybeans (May − August)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Soybeans (March − September) Soybeans (April − September) Soybeans (May − September)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Soybeans (March − April) Soybeans (May − June) Soybeans (July − August)
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bushels)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to a
particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each interval.
The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the Polynomial
regression. Each plot uses a different definition of the growing season. Curves are centered so the exposure-
weighted impact is zero. Histograms at the bottom of each panel display the average temperature exposure
among all counties in the data.

27
Figure A16: Nonlinear Relation Between Temperature and Yields for Various Growing Sea-
sons for Cotton
Cotton (March − September) Cotton (April − September) Cotton (May − September)
0.025 0.025 0.025

0 0 0
Log Yield (Bales)

Log Yield (Bales)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Cotton (March − October) Cotton (April − October) Cotton (May − October)
0.025 0.025 0.025

0 0 0
Log Yield (Bales)

Log Yield (Bales)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Cotton (April − May) Cotton (June − July) Cotton (August − September)
0.025 0.025 0.025

0 0 0
Log Yield (Bales)

Log Yield (Bales)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to a
particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each interval.
The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the Polynomial
regression. Each plot uses a different definition of the growing season. Curves are centered so the exposure-
weighted impact is zero. Histograms at the bottom of each panel display the average temperature exposure
among all counties in the data.

28
Figure A17: Nonlinear Relation Between Temperature and Yields in July Compared to
Remaining Months
Corn − Step Function Soybeans − Step Function Cotton − Step Function
0.1 0.1 0.1

0.05 0.05 0.05

Log Yield (Bushels)

Log Yield (Bushels)

0 0 0

Log Yield (Bales)

−0.05 −0.05 −0.05

−0.1 −0.1 −0.1

−0.15 −0.15 −0.15

July July July
Remaining Months Remaining Months Remaining Months
−0.2 −0.2 −0.2
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn − Polynomial (8th order) Soybeans − Polynomial (8th order) Cotton − Polynomial (8th order)
0.1 0.1 0.1

0.05 0.05 0.05

Log Yield (Bushels)

Log Yield (Bushels)

0 0 0

Log Yield (Bales)

−0.05 −0.05 −0.05

−0.1 −0.1 −0.1

−0.15 −0.15 −0.15

July July July
Remaining Months Remaining Months Remaining Months
−0.2 −0.2 −0.2
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)
Corn − Piecewise Linear Soybeans − Piecewise Linear Cotton − Piecewise Linear
0.1 0.1 0.1

0.05 0.05 0.05

Log Yield (Bushels)

Log Yield (Bushels)

0 0 0
Log Yield (Bales)

−0.05 −0.05 −0.05

−0.1 −0.1 −0.1

−0.15 −0.15 −0.15

July July July
Remaining Months Remaining Months Remaining Months
−0.2 −0.2 −0.2
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs show changes in log yield if the crop is exposed for one day to a particular 1◦ C temperature
interval where we sum the fraction of a day temperatures fall within each interval. The growing season is
split into July and the sum of the remaining months. The 95% confidence band, after adjusting for spatial
correlation are added as dotted lines for the step function as well as the polynomial. Curves are centered so
the exposure-weighted impact is zero.

29
Figure A18: Predicted Climate Change Impacts under the Hadley III Climate Model using
Separate Temperature Variables for July
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Pooled Months)
−60 Step Function (July Separate) −60
Polynomial (Pooled Months)
−80 Polynomial (July Separate)
−80
Piecewise Linear (Pooled Months)
Piecewise Linear (July Separate)
−100 −100
Soybeans Soybeans
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Pooled Months)
−60 Step Function (July Separate) −60
Polynomial (Pooled Months)
−80 Polynomial (July Separate)
−80
Piecewise Linear (Pooled Months)
Piecewise Linear (July Separate)
−100 −100
Cotton Cotton
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Pooled Months)
−60 Step Function (July Separate) −60
Polynomial (Pooled Months)
−80 Polynomial (July Separate)
−80
Piecewise Linear (Pooled Months)
Piecewise Linear (July Separate)
−100 −100

Notes: Graphs show predicted percentage changes in crop yields under four emissions scenarios. Left panels
display predicted impacts in the medium-term (2020-2049), right panels in the long-term (2070-2099). A
star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The dark line replicate Figure 2 in the main paper. Lighter colors include a separate set of
temperature variables for the month of July, while all remaining months are pooled.

30
11 Planted versus Harvested Area
The USDA yield data reports harvested bushels divided by harvested acres. This presents
a potential selection problem during extremely bad years when yields are close to zero and
farmers choose not to harvest. The yield measure may therefore underestimate the harmful
effects of heat waves. In Figure A19 we consider an alternative yield measure that divides
the total production quantity for corn by acres planted instead of acres harvested. If we use
the alternative yield measure we obtain similar results. The corresponding climate change
impacts are insensitive to the chosen yield definition as shown in Figure A20. Because the
data on planted acres is missing for 20% of our data for corn, we use production per harvested
acre in the main paper.

Figure A19: Nonlinear Relation Between Temperature and Yields using Production per Area
Planted
Corn − Area Planted Soybeans − Area Planted Cotton − Area Planted
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to
a particular 1◦ C temperature interval where we sum the fraction of a day temperatures fall within each
interval. The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the
Polynomial regression. Yield is total production divided by the area planted. Curves are centered so the
exposure-weighted impact is zero. Histograms at the bottom of each panel display the average temperature
exposure among all counties in the data.

31
Figure A20: Predicted Climate Change Impacts under the Hadley III Climate Model using
Production per Area Planted
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Harvested Area)
−60 Step Function (Planted Area) −60
Polynomial (Harvested Area)
−80 Polynomial (Planted Area)
−80
Piecewise Linear (Harvested Area)
Piecewise Linear (Planted Area)
−100 −100
Soybeans Soybeans
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Harvested Area)
−60 Step Function (Planted Area) −60
Polynomial (Harvested Area)
−80 Polynomial (Planted Area)
−80
Piecewise Linear (Harvested Area)
Piecewise Linear (Planted Area)
−100 −100
Cotton Cotton
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Harvested Area)
−60 Step Function (Planted Area) −60
Polynomial (Harvested Area)
−80 Polynomial (Planted Area)
−80
Piecewise Linear (Harvested Area)
Piecewise Linear (Planted Area)
−100 −100

Notes: Graphs show predicted percentage changes in crop yields under four emissions scenarios. Left panels
display predicted impacts in the medium-term (2020-2049), right panels in the long-term (2070-2099). A
star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The dark line replicates Figure 2 in the main paper where yield is total production divided by
the area harvested. Lighter colors instead use total production divided by area planted.

32
12 Year Fixed Effects
All regression so far included quadratic time trends by state. A smooth trend will not be
able to pick up sudden discrete jumps, i.e., after the introduction of a new crop variety with
a significant yield boost or other temporal shocks. We therefore replicate the analysis with
year fixed effects instead. Results are shown in Figure A21, which are similar to our initial
estimates. The corresponding climate change impacts are generally insensitive to the chosen
interpolation method as shown in Figure A22, but they do diverge somewhat for smaller
temperature increases for soybeans.

Figure A21: Nonlinear Relation Between Temperature and Yields using Year Fixed Effects
Corn − Linear Interpolation Soybeans − Linear Interpolation Cotton − Linear Interpolation
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day to
a particular temperature. Regressions include year fixed effects instead of quadratic time trends by state.
The 95% confidence band, after adjusting for spatial correlation, is added as grey area for the Polynomial
Regression. Estimated growth curves are centered so the exposure-weighted impact is zero. Histograms at
the bottom of each panel display the average temperature exposure among all counties in the data.

33
Figure A22: Predicted Climate Change Impacts under the Hadley III Climate Model using
Year Fixed Effects
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Quadratic Time Trend)
−60 Step Function (Year Fixed Effects) −60
Polynomial (Quadratic Time Trend)
−80 Polynomial (Year Fixed Effects)
−80
Piecewise Linear (Quadratic Time Trend)
Piecewise Linear (Year Fixed Effects)
−100 −100
Soybeans Soybeans
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Quadratic Time Trend)
−60 Step Function (Year Fixed Effects) −60
Polynomial (Quadratic Time Trend)
−80 Polynomial (Year Fixed Effects)
−80
Piecewise Linear (Quadratic Time Trend)
Piecewise Linear (Year Fixed Effects)
−100 −100
Cotton Cotton
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Quadratic Time Trend)
−60 Step Function (Year Fixed Effects) −60
Polynomial (Quadratic Time Trend)
−80 Polynomial (Year Fixed Effects)
−80
Piecewise Linear (Quadratic Time Trend)
Piecewise Linear (Year Fixed Effects)
−100 −100

Notes: Graphs show predicted percentage changes in crop yields under four emissions scenarios. Left panels
display predicted impacts in the medium-term (2020-2049), right panels in the long-term (2070-2099). A
star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The dark line replicate Figure 2 in the main paper where we use quadratic time trends by state.
Lighter colors instead use year fixed effects in the regression equation.

34
13 Linear Within-Day Temperature Interpolation
In this section we test the sensitivity of results to the chosen sinusoidal interpolation between
minimum and maximum temperature within each day. We replicate the analysis using a
linear interpolation instead. Results are shown in Figure A23, which are similar to our
initial estimates. The corresponding climate change impacts are insensitive to the chosen
interpolation method as shown in Figure A24.

Figure A23: Nonlinear Relation Between Temperature and Yields using a Linear Interpola-
tion between Minimum and Maximum Temperature within Each Day
Corn − Linear Interpolation Soybeans − Linear Interpolation Cotton − Linear Interpolation
0.025 0.025 0.025

0 0 0
Log Yield (Bushels)

Log Yield (Bushels)

Log Yield (Bales)

−0.025 −0.025 −0.025

−0.05 Step Function −0.05 Step Function −0.05 Step Function

Polynomial (8th−order) Polynomial (8th−order) Polynomial (8th−order)
Piecewise Linear Piecewise Linear Piecewise Linear
Exposure (Days)

Exposure (Days)

Exposure (Days)
−0.075 −0.075 −0.075
10 10 10
5 5 5
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Temperature (Celsius) Temperature (Celsius) Temperature (Celsius)

Notes: Graphs at the top of each panel show changes in log yield if the crop is exposed for one day
to a particular temperature. The distribution of temperatures within a day is derived by using a linear
interpolation between minimum and maximum temperature rather than a sinusoidal interpolation. The 95%
confidence band, after adjusting for spatial correlation, is added as grey area for the Polynomial Regression.
Estimated growth curves are centered so the exposure-weighted impact is zero. Histograms at the bottom
of each panel display the average temperature exposure among all counties in the data.

35
Figure A24: Predicted Climate Change Impacts under the Hadley III Climate Model using
a Linear Interpolation between Minimum and Maximum Temperature within Each Day
Corn Corn
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Sinusoidal Int.)
−60 Step Function (Linear Int.) −60
Polynomial (Sinusoidal Int.)
−80 Polynomial (Linear Int.)
−80
Piecewise Linear (Sinusoidal Int.)
Piecewise Linear (Linear Int.)
−100 −100
Soybeans Soybeans
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Sinusoidal Int.)
−60 Step Function (Linear Int.) −60
Polynomial (Sinusoidal Int.)
−80 Polynomial (Linear Int.)
−80
Piecewise Linear (Sinusoidal Int.)
Piecewise Linear (Linear Int.)
−100 −100
Cotton Cotton
0 0
Impacts by 2020−2049 (Percent)

Impacts by 2070−2099 (Percent)

B1 B2 A2 A1 B1 B2 A2 A1

−20 −20

−40 −40
Step Function (Sinusoidal Int.)
−60 Step Function (Linear Int.) −60
Polynomial (Sinusoidal Int.)
−80 Polynomial (Linear Int.)
−80
Piecewise Linear (Sinusoidal Int.)
Piecewise Linear (Linear Int.)
−100 −100

Notes: Graphs show predicted percentage changes in crop yields under four emissions scenarios. Left panels
display predicted impacts in the medium-term (2020-2049), right panels in the long-term (2070-2099). A
star indicates the point estimates and whiskers show the 95% confidence interval after adjusting for spatial
correlation. The dark line replicate Figure 2 in the main paper where we used a sinusoidal approximation
between minimum and maximum temperature. Lighter colors instead use a linear interpolation between
minimum and maximum temperature.

36
14 Uniform Climate Change Scenarios
Table A5 reports predicted yield impacts under a range of uniform temperature changes
rather than climate change scenarios from the Hadley III climate model. Impacts on total
production are based on the most flexible functional form (step-function) and correspond to
the blue line of the non-uniform climate change scenarios in Figure 2 of the main paper.

Table A5: Predicted Climate Change Impacts under Uniform Climate Change

Corn Soybeans Cotton

Impact s.e. Impact s.e. Impact s.e.
◦
Temperature +1 C -6.38 (0.53) -3.17 (0.46) -2.59 (1.57)
Temperature +2◦ C -14.87 (1.02) -8.89 (0.90) -8.34 (3.08)
Temperature +3◦ C -24.76 (1.47) -16.98 (1.25) -16.30 (4.54)
Temperature +4◦ C -35.30 (1.84) -26.88 (1.51) -25.48 (6.01)
◦
Temperature +5 C -45.75 (2.21) -37.66 (1.73) -35.61 (7.36)
◦
Temperature +6 C -55.88 (2.66) -48.82 (1.97) -46.72 (8.39)
Temperature +7◦ C -65.19 (3.12) -59.16 (2.24) -58.28 (8.72)
◦
Temperature +8 C -73.42 (3.56) -68.18 (2.53) -69.08 (8.33)
Temperature +9◦ C -80.26 (3.87) -75.66 (2.80) -78.08 (7.52)
◦
Temperature +10 C -85.86 (4.07) -81.69 (2.99) -85.06 (6.43)
Precipitation -50% -10.44 (1.47) -11.80 (1.29) +6.38 (2.23)
Precipitation -40% -7.43 (1.12) -8.52 (0.98) +5.13 (1.64)
Precipitation -30% -4.85 (0.79) -5.66 (0.69) +3.86 (1.13)
Precipitation -20% -2.73 (0.49) -3.26 (0.43) +2.58 (0.69)
Precipitation -10% -1.10 (0.23) -1.36 (0.20) +1.30 (0.32)
Precipitation +10% +0.57 (0.20) +0.81 (0.18) -1.30 (0.28)
Precipitation +20% +0.60 (0.38) +1.07 (0.35) -2.61 (0.53)
Precipitation +30% +0.09 (0.55) +0.76 (0.53) -3.93 (0.78)
Precipitation +40% -0.95 (0.73) -0.10 (0.71) -5.24 (1.02)
Precipitation +50% -2.50 (0.92) -1.49 (0.92) -6.57 (1.28)

Notes: Table reports predicted yield changes and standard errors (in parentheses) under unform temperature
changes and precipitation change. Predictions are based on our baseline (step function) model.

37
References
[1] Schlenker W, Roberts MJ (2006) Nonlinear Effects of Weather on Corn Yields. Rev Agr
Econ 28:391-398.

[2] Thom HCS (1966) Normal Degree Days Above Any Base by the Universal Truncation
Coefficient. Mon Weather Rev 94:461-465.

[3] Snyder RL (1985) Hand Calculating Degree Days. Agr Forest Meterol 35:353-358.

[4] Nakicenovic N, Swart R (2000) Special Report on Emissions Scenarios (Cambridge

University Press, Cambridge).

[5] Conley TG (1999) GMM Estimation with Cross Sectional Dependence. J Econometrics
92:1-45.

[6] Ritchie JT, NeSmith DS (1991) in Modeling Plant and Soil Systems. Agronomy 31, eds
Hanks J, Richie JT (American Society of Agronomy, Madison), pp 5-29.

[7] Schlenker W, Hanemann WM, Fisher AC (2006) The Impact of Global Warming on
U.S. Agriculture: An Econometric Analysis of Optimal Growing Conditions. Rev Econ
Stat 88:113-125.

[8] Mendelsohn R, Nordhaus WD, Shaw D (1994) The Impact of Global Warming on Agri-
culture: A Ricardian Analysis. Am Econ Rev 84:753-771.

[9] Deschênes O, Greenstone M (2007) The Economic Impacts of Climate Change: Evidence
from Agricultural Output and Random Fluctuations in Weather. Am Econ Rev, 97:354-
385.

[10] National Corn Growers Association, February 27 2008,

[Link]

[11] Hallauer AR (2000) Speciality Corns (CRC Press, Boca Raton).

[12] Thomison P (2002) Drought and Heat Stress Effects On Corn Yield Potential (Crop
Observation and Recommendation Network, Ohio State University).

38