The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Temperature and Growth: A Panel
Analysis of the United States
WP 18-09
Riccardo Colacito
University of North Carolina at Chapel
Hill
Bridget Hoffmann
Inter-American Development Bank
Toan Phan
Federal Reserve Bank of Richmond
Temperature and Growth:
A Panel Analysis of the United States
Riccardo Colacito
Bridget Hoffmann
Toan Phan∗
Working Paper No. 18-09
————————————————————————————————————–
Abstract
We document that seasonal temperatures have significant and systematic effects
on the U.S. economy, both at the aggregate level and across a wide cross-section of
economic sectors. This effect is particularly strong for the summer: a 1o F increase in
the average summer temperature is associated with a reduction in the annual growth
rate of state-level output of 0.15 to 0.25 percentage points. We combine our estimates
with projected increases in seasonal temperatures and find that rising temperatures
could reduce U.S. economic growth by up to one-third over the next century.
JEL classification: O44; Q51; Q59; R11.
This draft: March 30, 2018.
————————————————————————————————————–
∗
Colacito is affiliated with the Kenan-Flagler Business School and the Department of Economics at
the University of North Carolina at Chapel Hill. Hoffmann is affiliated with the Research Department
at Inter-American Development Bank. Phan is affiliated with the Federal Reserve Bank of Richmond.
Emails: ric@unc.edu, bridgeth@iadb.org, and toanvphan@gmail.com. The authors acknowledge
helpful comments from the editor and two anonymous referees, as well as helpful discussions with Ravi
Bansal, Marshall Burke, Tatyana Deryugina, Don Fullerton, Duane Griffin, Solomon Hsiang, Benjamin
Jones, Ju Hyun Kim, and Mike Roberts. We also thank conference/seminar participants at the American
Economic Association, Econometric Society World Congress, Association of Environmental and Resource
Economists, European Economic Association Annual Congress, Triangle Resource and Environmental
Economics, University of Illinois at Chicago, UNC–Chapel Hill, and University of Hawaii. The views
expressed herein are those of the authors and not those of the Federal Reserve Bank of Richmond or
the Federal Reserve System, or the Inter-American Development Bank, its Board of Directors, or the
countries they represent. All errors are ours.
1
Introduction
We analyze the effect of average seasonal temperatures on the growth rate of U.S. output. We find that seasonal temperatures, particularly summer temperatures, have significant and systematic effects on the U.S. economy, both at the aggregate level and
across a wide cross-section of economic sectors. A 1o F increase in the average summer
temperature is associated with a reduction in the annual growth rate of state-level output of 0.15 to 0.25 percentage points.
As global average temperatures are predicted to continue rising over this century, many
scholars and policymakers have raised warnings of the potential for dramatic damages
to the global economy (e.g., Stern (2007), Field et al. (2014)). The economics literature
has documented substantial negative effects of global warming on economic growth in
developing economies (e.g., Gallup, Sachs and Mellinger (1999), Nordhaus (2006), Burke
et al. (2009), Dell, Jones and Olken (2012)). For the U.S., however, it has been challenging to provide systematic evidence that rising temperatures affect the growth rate of
economic activities beyond sectors that are naturally exposed to outdoor weather conditions (see Mendelsohn and Neumann (1999), Schlenker and Roberts (2006; 2009), and
Burke and Emerick (2015) for an analysis of an agricultural industry). We contribute to
this literature by providing comprehensive evidence that rising temperatures do affect
U.S. economic activities, at both the aggregate and industry levels.
We overcome existing challenges by exploiting random fluctuations in seasonal temperatures across years and states. Using a panel regression framework with the growth
rate of state GDP, or gross state product (henceforth GSP), and average seasonal temperatures of each U.S. state, we find that summer and fall temperatures have opposite
effects on economic growth. An increase in the average summer temperature negatively
affects the growth rate of GSP, while an increase in the fall temperature positively affects this growth rate, although to a lesser extent. The different signs of the two effects
1
suggest that previous studies’ aggregation of temperature data into annual temperature averages (e.g., Dell et al. (2012)) may mask the heterogeneous effects of different
seasons.
The summer effect dominates the fall effect in our recent sample (post-1990), leading to
a negative net economic effect of rising temperatures. This implies that the U.S. economy is still sensitive to temperature increases, despite the progressive adoption of adaptive technologies such as air conditioning (Barreca et al. (2015)). We also document that
the temperature effects are particularly strong in states with relatively higher summer
temperatures, most of which are located in the South. However, we do not find any evidence that the effect of temperature on GDP in the South is driven by the relatively less
developed states. This implies that the channel through which temperature affects GDP
in this part of the country must be distinct from the one documented in the literature
for developing economies.
We revisit the conjecture that only a small fraction of the sectors of the economy are
sensitive to rising temperatures in developed economies, implying that the aggregate
economic impact of warming on the U.S. will be limited (Schelling (1992), Mendelsohn
(2010), Nordhaus (2014)). Our results show that rising summer temperatures have a
pervasive effect in the entire cross-section of industries, above and beyond the sectors
that are traditionally deemed as vulnerable to changing climatic conditions. Figure 1
documents that, in the most recent part of our sample, an increase in the average summer temperature negatively affects the growth rate of output of many industries, including finance, services, retail, wholesale, and construction, which in total account for more
than a third of national gross domestic product (GDP). Only a limited number of sectors, such as utilities (1.8% of national GDP), which includes providers of energy, benefit
from an increase in the average summer temperature.1 To the best of our knowledge,
our paper is the first in the literature to systematically document the pervasive effect of
1
Section 4.2 provides a comprehensive break-down of these results across different samples and subindustries.
2
0.1
0.05
0
-0.05
-0.1
-0.15
-0.2
Mining
Utilities
Manufacturing
Transportation
Governement
Communication
Retail
Construction
Wholesale
Agriculture
Services
-0.25
All Industries = -0.256 (0.060)
Finance,
Real Estate,
and Insurance
Effect of Summer Temperature on GDP times Industry Share
0.15
Figure 1: Decomposition of the Summer temperature effect in the cross-section of industries.
For each industry, the horizontal line represents the point estimate of the impact of Summer
temperature on the growth rate of industry GDP times the industry share of GDP. The bottom
and top portions of each rectangle represent 90% confidence intervals, while the outer limits
of each boxplot represent the 95% confidence interval of each estimated coefficient. Standard
errors are clustered at the year level. The number denoted as “All Industries” is the sum of all
the industry coefficients multiplied by the corresponding industry share. All estimates refer to
the post-1997 sample as documented in Table 4A of Section 4.2.
summer temperatures on the cross-section of industries in the U.S..
We document that temperature may affect economic activities through its impact on
labor productivity. In our empirical analysis, an increase in the average summer temperature decreases the annual growth rate of labor productivity, while an increase in
the average fall temperature has the opposite effect. While our finding sheds light on
the effects of temperature on labor productivity at the macroeconomic level, it is also
3
consistent with existing studies of this relationship at the microeconomic level. For
example, Zivin and Neidell (2014) have found that warmer temperatures reduce labor
supply in the U.S., and Cachon, Gallino and Olivares (2012) have documented that high
temperatures decrease productivity and performance.
Our paper also contributes to the growing debate on the long-term economic consequences of rising global temperatures (e.g., Mendelsohn and Neumann (1999) and Tol
(2010)). We combine our estimates of the effects of seasonal temperatures on the growth
rate of U.S. output with several projections of the expected U.S. temperature change
over the next century. We conduct our analysis under a “business as usual” benchmark,
in which there is no additional mitigation and the estimated effects of temperature on
economic growth remain unchanged over the long horizon. We document that the projected increases in summer and fall temperatures could reduce the growth rate of annual
nominal GDP by up to 1.2 percentage points, which is roughly a third of the historical
average nominal growth rate of about 4% per year.
Our analysis highlights the complex ways in which temperatures affect economic activities, and it reveals the need to disaggregate the data into seasons and industries to
uncover the full extent of this impact. By providing specific estimates of the effect of
temperature on economic activities in the U.S., our empirical analysis informs a growing body of literature focused on general equilibrium models of climate change, including
integrated assessment models. These models constitute the basis of many policy recommendations regarding the regulation of greenhouse gas emissions (e.g., Golosov et al.
(2014), Acemoglu et al. (2012), Bansal and Ochoa (2011), and Bansal, Ochoa and Kiku
(2014)). All of these models critically rely on empirical estimates of the impacts of rising
temperatures on aggregate economic activities. In the absence of specific estimates for
the U.S., the parameters of these “climate damage functions” are generally calibrated to
match cross-country estimates (e.g., Nordhaus and Sztorc (2013)). In this respect, our
analysis helps bridge the gap between the theoretical and empirical literatures and will
4
enable researchers to sharpen the policy recommendations based on this class of models,
especially for the U.S.
Our focus on quarterly temperature fluctuations allows us to combine our estimates
with existing climatological projections, which are typically only available at lower frequencies. Our analysis differs from a large literature that uses daily temperature fluctuations to study outcomes as diverse as agriculture, local income, birth weight, mortality,
and time allocation (Schlenker and Roberts (2006), Deschênes and Greenstone (2007),
Deryugina and Hsiang (2014), Deschênes, Greenstone and Guryan (2009), Deschênes
and Greenstone (2011), and Zivin and Neidell (2014)). Whereas studies using high frequency data typically focus on the effect of a change in the observed distribution of daily
temperature, we focus on the effect of a change in the mean seasonal temperature.
In a similar vein, Bloesch and Gourio (2015) analyze the impact of temperature and
snowfall during the coldest months of the year (November through March) on the growth
rate of quarterly economic activities. They find that snowfall has a negative effect on
some routinely employed economic indicators such as nonfarm payroll, housing permits,
and housing starts. Our analysis is broadly consistent with their results, which suggests that a drop in temperatures in cold weather seems to have a negative impact on
economic activities. While they focus on the quarter-to-quarter effect of temperature
on economic activity to measure a potential bounce-back effect following adverse winter
weather conditions, we assess the cumulative effect on the annual growth rate of output
and emphasize the effect of summer temperatures.
The rest of the paper is organized as follows. Section 2 provides a description of the
main datasets that we employ in our analysis. Section 3 describes our main results and
documents the stability of the estimated effects over time. Section 4 documents several
economic mechanisms driving the main results, including the effect of temperature on
labor productivity, on the growth rate of output in the cross-section of industries, and in
the cross-section of U.S. regions. Section 5 analyzes the long-term consequences of global
5
warming for the aggregate U.S. economy, in addition to providing robustness checks of
our main results. Section 6 concludes. The Online Appendix provides a comprehensive
set of robustness exercises.
2
Data
This section describes our data sources and the procedures we use to aggregate weatherrelated data. We refer the reader to the appendix for additional details.
2.1
Weather data
We use daily station-level weather data from the National Oceanic and Atmospheric
Administration (NOAA) Northeast Regional Climate Center. This dataset contains daily
observations on average temperature, precipitation, and snowfall across U.S. weather
stations. Throughout the paper, the unit of temperature is degrees Fahrenheit. The
longest common sample across all weather stations starts in 1869 and ends in 2012.
In this study, we focus on the 1957–2012 sample, which coincides with the period for
which we have data on GSP (see below). For each weather station, we deseasonalize
the raw data by regressing daily observations on 12 dummies representing the months
of the year and subtract the corresponding estimated monthly component from each
observation (see appendix A.2.1 for details on deseasonalization).2
We aggregate daily weather observations to quarterly averages by taking the average of
the daily observations in each season. We define the winter as January through March,
the spring as April through June, the summer as July through September, and the fall
2
We have followed the common practice in the macroeconomics literature of not correcting the standard
errors of our analysis to account for the deseasonalization. However, we provide a robustness check of our
results using a data set which does not seasonally adjust temperatures.
6
as October through December. Our definition of seasons coincides with the definition of
quarters commonly encountered in the macroeconomics literature, and thus will allow
our analysis to contribute to future developments of macroeconomic models that include
climate-related variables. We analyze average seasonal temperatures in order to establish a connection between long-term temperature changes and economic activities.
This connection can be more accurately assessed using a lower-frequency temperature
measure. We also consider alternative definitions of seasons in the robustness checks in
section 5.3.
To aggregate weather data from the station level to the county and state level, we employ ArcGIS, a geographic information system, to obtain the coordinates for the centroid
of each of the 3,144 counties and county equivalents, as well as each weather station.
The country, state, and county borders used in ArcGIS are from 2013 topographically
integrated geographic encoding and referencing (TIGER) shape files. These shape files,
along with the area and population of each county are obtained from the U.S. Census
Bureau. We then follow a standard aggregation method (e.g., Deschênes and Greenstone
(2012)). For each county we weight the daily temperature, precipitation, and snowfall
of each weather station in a 500 km radius of the county’s centroid by the inverse of
the straight-line distance between the station and the county centroid. In this way,
the closest weather stations are assigned a larger weight in determining each county’s
weather.
Finally, to aggregate to the state level, we weight the weather observations of each
county in a state in proportion to either the corresponding county’s area or population.
Weighting by area assigns larger weights to larger counties, while weighting by population assigns larger weights to more densely populated areas. We use area weights in
the main analysis in the text, but our results are very similar across different weighting
schemes (see section 5.3 and appendix A.6.2). We aggregate state-level weather data to
the country level by following the same procedure.
7
In section 5.3, we document that our results are robust to using non-deseasonalized
gridded temperature data. We use the NOAA U.S. Climate Divisions’ nClimDiv dataset,
which provides absolute monthly temperature averages for each state, derived from
area-weighted averages of 5km × 5km grid-point temperature estimates interpolated
from station data.3 In appendix A.6.1, we replicate all of our results by using gridded
temperature data from the NOAA nClimDiv data set. We show that our results are
robust to the use of this alternative dataset.
2.2
State-level economic data
We use data on nominal GSP between 1957 and 2012 for all 50 states and the District
of Columbia. GSP is defined as the value added in production by the labor and capital
of all industries located in that state. Data for 1957–1962 come from the U.S. Census
Bureau Bicentennial Edition, and data for the 1963–2012 sample come from the U.S.
Department of Commerce’s Bureau of Economic Analysis (BEA). The data frequency
is annual. From the BEA, we also collected data for national GDP, nominal GSP per
capita, real GSP, and industry output data for the 1963–2011 sample. Industry data
for 1963–1997 are categorized using the Standard Industrial Classification (SIC) codes,
while data for 1997–2011 follow the North American Industry Classification System
(NAICS). Finally, annual employment data at the state level (measured in thousands of
employees) are collected from the Bureau of Labor and Statistics for the sample 1990–
2012.
3
See ftp://ftp.ncdc.noaa.gov/pub/data/cirs/climdiv/climdiv-inv-readme.txt
more details.
8
for
3
Main results
In this section we report our main empirical results. First, as a benchmark, we show
that the relationship between temperatures and growth is not statistically significant in
time-series regressions at the whole-country level, consistent with findings in the existing literature. Then, we improve the analysis by using panel regressions with weather
and economic data from all 50 states plus the District of Columbia (henceforth “the
cross-section of states”).
For the baseline specifications that we consider, we always include the lagged dependent
variable and the average seasonal temperatures. We motivate the inclusion of lagged
GDP growth rates with the strong empirical evidence supporting the claim of first order
auto-correlation of this variable in the cross-section of U.S. states (see table A5 in the
appendix for a complete set of estimates).4 Furthermore, we document in table A6 in
the appendix that the correlation of seasonal temperatures is typically very low. This
supports the claim that our results are unlikely to be affected by multicollinearity.
Our main findings are as follows: (1) an increase in the average summer temperature
negatively affects the growth rate of GSP, and (2) an increase in the average fall temperature positively affects growth, although to a lesser extent. Both effects are statistically
and economically significant. In section 5.3, we perform a comprehensive set of robustness checks and show that the summer effect is generally very robust, while the fall
effect is less so.
Our finding on the compositional effect of seasonal temperatures on GDP is relevant
because it implies that temperature fluctuations may also affect the volatility of GDP
growth. Indeed, given the modest degree of correlation of seasonal temperatures (see table A6 in the appendix), our results indicate that temperature’s volatility will contribute
4
A common concern about including lagged dependent variables in our regressions is that this could
give rise to the Nickell (1981) bias. We show in section 5.3 that our results are robust to excluding the
lagged dependent variable.
9
Table 1: Main results:
Effects of annual and seasonal temperatures on GSP growth
Time Series
Whole Year
−0.396
Winter
−0.071
Spring
−0.027
Summer
−0.414
Fall
0.042
(0.382)
(0.179)
(0.334)
(0.385)
(0.287)
−0.154
Panel Analysis
0.006
0.001
0.003
(0.111)
(0.069)
(0.105)
(0.049)
(0.025)
(0.044)
(0.065)
(0.032)
(0.051)
(0.072)∗∗
(0.047)∗∗∗
(0.065)∗∗
0.102
(0.055)∗
(0.040)∗∗∗
(0.054)∗
Notes: The first column reports the estimated coefficients on average annual temperature from
a regression of the economic growth rate on its lag and the average annual temperature (regressions (1) and (3)). The four columns on the right report the estimated coefficients for each of
the four seasonal temperature averages (regressions (2) and (4)). The top panel (“Time Series”)
reports the estimated coefficients using GDP and weather data aggregated to the national level
(regressions (1) and (2)). The bottom panel (“Panel Analysis”) reports estimated coefficients using state-level GSP and weather data (regressions (3) and (4)). In the panel regressions, all 50
states and the District of Columbia are included and each state is weighted by the proportion,
averaged over the whole sample, of its GSP relative to the national GDP. All specifications include the lagged dependent variable, and the panel specifications include state and year fixed
effects. Temperatures are in degrees Fahrenheit. The sample is 1957–2012. Standard errors are
in parentheses. In the bottom panel, the standard errors are clustered by year, by state, and by
both dimensions. ***, **, and * denote significance at the 1%, 5%, and 10% levels.
to GDP volatility. This is an important result in light of an abundant macro literature
on the welfare costs of economic fluctuations. This literature, which was started by Lucas (1987), is interested in the question of how much would economic agents be willing
to pay in order to eliminate all sources of fluctuations in business cycles. Equivalently,
if temperature does contribute to business cycle fluctuations, then our analysis is relevant because it points out that there are welfare consequences associated with large
temperature variations.
10
3.1
Benchmark: Time-series regressions with country-level data
We consider two time-series regressions. The first is a regression of the aggregate
growth rate of national GDP on the average annual temperature:
∆yt = βTt + ρ∆yt−1 + α + εt ,
(1)
where ∆yt denotes the growth rate of national GDP between years t − 1 and t; Tt denotes
the annual average temperature in year t in degrees Fahrenheit; and the lagged growth
rate ∆yt−1 controls for autocorrelation.
The second is a regression of the growth rate of aggregate GDP on the average temperatures of the four seasons:
∆yt =
X
βs Ts,t + ρ∆yt−1 + α + εt ,
(2)
s∈S
where Ts,t denotes the average temperature in season s ∈ S = {winter, spring, summer, f all}
in year t.
The first row of table 1 reports the results of these regressions. The column “Whole
Year” reports the estimate for the coefficient β in equation (1). The remaining columns
report the estimations for coefficients βs in equation (2). As the table shows, none of the
estimated coefficients are statistically significant. These results confirm the difficulty of
identifying the effect of temperature on economic growth in the U.S. documented in the
extant literature.
3.2
Panel regressions with state-level data
We explore the impact of temperature on the growth rates of GSP in the cross-section of
states using two panel specifications that mirror our time-series analysis. The first is a
11
regression of the growth rate of GSP on the state-level annual average temperature:
∆yi,t = βTi,t + ρ∆yi,t−1 + αi +αt +εi,t ,
(3)
where ∆yi,t and Ti,t denote GSP growth and the annual average temperature in state i in
year t, respectively, while αi and αt denote state and year fixed effects. We again include
the lagged GSP growth rate as a control to capture the degree of autocorrelation of the
dependent variable.
In the second specification, which is the main specification of the paper, we disaggregate
the annual average temperature into four average seasonal temperatures:
∆yi,t =
X
βs Ti,s,t + ρ∆yi,t−1 +αi +αt +εi,t ,
(4)
s∈S
where the variables are defined as above and Ti,s,t denotes the temperature in degrees
Fahrenheit in state i, year t, and season s. We show in section A.6.4 of the appendix that
our results are robust to using the Arellano and Bond (1991) estimation methodology.
Since some states have larger GSPs and thus contribute more to national GDP than
others, in both panel specifications we weight each state by the proportion of its GSP
relative to the entire country’s GDP over the whole sample (see appendix A.2.2 for details). In section 5.3 we conduct robustness checks with alternative weighting schemes.
We report the results of these panel regressions in table 1. The column “Whole Year”
refers to the specification in equation (3). We report the estimated coefficient for β
and standard errors. The results indicate that the effect of average temperature at the
annual level is again not statistically significant, confirming the findings in the timeseries specifications.
However, when we break down annual temperatures into the four seasonal temperatures, the results change substantially. The rightmost four columns of table 1 report
12
the estimates for the βs seasonal coefficients with associated standard errors, clustered
by year, by state, and in both dimensions. The table shows the relationship between
average summer and fall temperatures and economic growth rates. These effects are
both statistically and economically significant: a 1o F increase in the average summer
temperature is associated with a reduction in the annual GSP growth rate by 0.154 percentage points, while a 1o F increase in the average fall temperature is associated with
an increase in the annual GSP growth rate by 0.102 percentage points.5
The opposite signs of the effects of summer and fall temperatures on GSP growth rates
may partially explain the difficulty of obtaining statistically significant estimates using annual temperatures. Even though the magnitudes of the summer and fall effects
are comparable, we document through robustness checks (section 5.3) and the exercise
below that the summer effect is much more robust than the fall effect.
3.3
Growth vs. level effects
We test whether the response of GSP to temperature in the previous section is an effect
on the level of output or on its growth rate. It is important to distinguish between these
two hypotheses, because the effects on the growth rate compound over time and thus are
more quantitatively important than effects on the level of output (Pindyck (2011; 2013),
Dell et al. (2012)).
To illustrate the quantitative significance of growth effects compared to level effects, and
to set the stage for our empirical methodology, consider the following simple example.
Assume that the aggregate output of a certain state follows the process:
yt = α + yt−1 + βTt + βlag Tt−1 + εt ,
5
(5)
We note that the time series point estimates for the summer are larger than the corresponding panel
coefficients, but they are statistically insignificant. This means that by using state level data we obtain
an increase in precision due to the increased ability to connect temperatures to local economic variables.
13
where β and βlag denote the impacts of current and lagged average temperatures T (of,
for instance, the current and last summer) on output growth. For simplicity, we assume
that εt = 0, ∀t. Consider a shock in year t = 1 that permanently increases the average
temperature T by one degree Fahrenheit, from T0 = 0 to Tt = 1, ∀t ≥ 1 (illustrated in the
left panel of figure 2). This temperature path is motivated by climatologists’ predictions
that temperatures will rise permanently by the middle and end of this century (see
section 5.2 for details). Along this hypothetical temperature path, the level and the
growth rate of output would be
yt = (y0 + β) + (t − 1) [α + (β + βlag )] , and
∆y1 = α + β,
and
∆yt = α + (β + βlag ) ,
(6)
∀t ≥ 2,
(7)
respectively. We consider three cases. If β = βlag = 0, then temperatures have no
economic effect. We refer to this situation as the No Effect case. If β + βlag = 0, then an
increase in temperature has a permanent impact on the level of output (see equation (6)),
but it affects the growth rate of output for only one period (see equation (7)). We refer to
this situation as the Level Effect case. If β + βlag 6= 0, temperature permanently affects
both the level and the growth rate of output. We refer to this situation as the Growth
Effect case.
We illustrate these three scenarios in the right panel of figure 2. Over the span of 50
years, if temperatures permanently affect the growth rate of output (the Growth Effect
case), then the level of output would be substantially lower than what it would be in
the No Effect case (dashed-dot vs. dashed line). This is in sharp contrast to the case in
which temperature has a permanent effect only on the level of output (the Level Effect
case). In this scenario, after 50 years output is only marginally lower compared to the
baseline case (solid vs. dashed line).
We follow the logic of this example to test whether average seasonal temperatures affect
14
100
1.2
90
80
70
0.8
60
0.6
50
40
0.4
0.2
No Effect
30
Level Effect
20
OUTPUT
TEMPERATURE
1
10
Growth Effect
0
0
0
10
20
30
40
50
0
TIME
10
20
30
40
50
TIME
Figure 2: Growth vs. level effect. The left panel depicts a permanent increase of 1 degree
Fahrenheit in the level of temperature that takes place at year 1. The right panel shows the
levels of output associated with the temperature path reported in the left panel. The three lines
are constructed according to equation (5), by setting β = βlag = 0 (dashed line), β = −βlag =
−0.170 (solid line), and β = −0.170, βlag = −0.153 (dash-dot line). In all three cases µ = 2 and
εt = 0, ∀t.
the growth rate of GSP in the data. Specifically, we estimate the following equation:
∆yi,t =
X
s∈S
βs Ti,s,t +
X
βlag,s Ti,s,t−1 +ρ∆yi,t−1 + αi + αt + εi,t .
(8)
s∈S
|
{z
lagged terms
}
Then we test whether we can reject the null hypothesis that the sum of the contemporaneous and lagged coefficients for each season is equal to zero, that is, H0 : βs + βlag,s = 0,
for each season s.6
Our results, reported in table 2, indicate that lagged temperatures are generally statistically significant, with the exception of the winter season. The signs of the effects for
summer and fall do not change when considering the lagged temperatures. Most importantly, from the p-values of the Wald tests reported on the last row of the table, we can
strongly reject the null hypothesis that the sums of the contemporaneous and lagged
temperature coefficients are equal to zero (βs + βlag,s = 0) for summer and fall. This
6
Note that by setting βs ≡ β/4, βlag,s ≡ βlag /4, for each season s, we obtain the specification for average
annual temperature in (3) augmented with lagged temperature. We omit this case from our investigation,
since we did not find any statistically significant effect associated with annual temperatures in table 1.
15
Table 2: Growth vs. level effects
Contemporaneous temp.
One-year lagged temp.
Sum of coefficients
Wald test’s p-value
Winter
−0.008
Spring
−0.012
(0.051)
(0.029)
(0.048)
(0.059)
(0.032)
(0.046)
0.004
0.121
(0.053)
(0.023)
(0.046)
(0.063)∗
(0.039)∗∗∗
(0.054)∗∗
−0.004
0.109
Summer
−0.170
(0.076)∗∗
(0.045)∗∗∗
(0.067)∗∗
−0.153
(0.079)∗
(0.053)∗∗∗
(0.075)∗∗
−0.323
Fall
0.108
(0.050)∗∗
(0.038)∗∗∗
(0.048)∗∗
0.066
(0.060)
(0.029)∗∗
(0.049)
0.174
(0.084)
(0.031)
(0.075)
(0.086)
(0.045)∗∗
(0.068)
(0.115)∗∗∗
(0.077)∗∗∗
(0.108)∗∗∗
(0.077)∗∗
(0.052)∗∗∗
(0.067)∗∗∗
[0.961]
[0.893]
[0.956]
[0.208]
[0.018]
[0.110]
[0.007]
[0.000]
[0.003]
[0.027]
[0.002]
[0.009]
Notes: This table reports results of the growth vs. level regression (8). The first row (“Contemporaneous temp.”) reports estimates for coefficient β of the effect of contemporaneous temperature
on economic growth, while the second row (“One-year lagged temp.”) reports estimates for coefficient βlag of the effects of one-year lagged temperature. The third row (“Sum of coefficients”)
reports the sum of β and βlag . The last row (“Wald test p-value”) reports the p-values for the
Wald test of whether β + βlag is significantly different from zero. Temperatures are in degrees
Fahrenheit. The sample is 1957–2012. The regressions are weighted by constant GSP shares.
The standard errors, clustered by year, by state, and by both dimensions, are in parentheses.
***, **, and * denote significance at the 1%, 5%, and 10% levels.
evidence supports the hypothesis that increases in summer and fall temperatures have
lasting effects on output growth. Table A4 in the appendix provides several robustness
checks for the result reported in table 2.
In what follows, we will focus primarily on an econometric specification that omits
lagged temperatures for parsimony. While this results in a bias in our estimated coefficients, we note that such bias affects primarily the estimate of the autoregressive
coefficient, and only marginally the estimated temperature coefficients (as can be noted
by comparing tables 1 and 2). For completeness the robustness of all the results of our
subsequent analysis including lagged temperatures are reported in section A.6.3 of the
appendix.
16
3.4
Stability of the effects through time
We explore how the estimated coefficients in the main panel regression (4) evolve through
time. This exploration is relevant because it could be the case that the negative economic effects of summer temperatures are diminished in the more recent part of the
sample due to adaptation (for example, due to widespread adoption of air conditioning
technologies as documented by Barreca et al. (2015)).
We re-run the regression specified in equation (4) but delay the beginning of the sample
by one year at a time. We repeat this exercise until the sample starts in 1990; past
this year, the sample size becomes very small, thus compromising the power of our
estimation. The results, reported in figure 3 show that the summer coefficient remains
negative and statistically significant at the 10% level as the sample shrinks; the pointestimate for the summer effect is −0.154 in the full sample and −0.246 in the post-1990
sample. However, the fall coefficient is no longer statistically significant in the post-1990
sample; the point-estimate for the fall effect is 0.102 in the full sample and 0.031 (and
indistinguishable from zero) in the post-1990 sample. This finding is consistent with the
results of our robustness checks (section 5.3): the summer effect is very robust, but the
fall effect is not.
4
Economic mechanisms
In this section, we explore potential mechanisms through which temperatures affect
the growth rate of GSP. First, we show that summer and fall temperatures affect the
growth of labor productivity. Second, we disaggregate GSP into industry groups and
show that, in the post-1997 sample, an increase in the average summer temperature
negatively affects output growth in various industry groups (including food services and
drinking places; insurance; wholesale; retail; and agriculture, forestry, and fishing) and
17
Winter
Estimated Coefficients
0.3
0.2
0.2
0.1
0.1
0
0
-0.1
-0.1
-0.2
-0.2
-0.3
-0.3
-0.4
-0.4
-0.5
-0.5
1960
1965
1970
1975
1980
1985
1990
1960
Summer
0.3
Estimated Coefficients
Spring
0.3
1965
1970
0.2
0.1
0.1
0
0
-0.1
-0.1
-0.2
-0.2
-0.3
-0.3
-0.4
-0.4
-0.5
1980
1985
1990
1980
1985
1990
Fall
0.3
0.2
1975
-0.5
1960
1965
1970
1975
1980
1985
1990
1960
Year
1965
1970
1975
Year
Figure 3: Stability across time of the effect of average seasonal temperatures on GSP growth.
Each panel reports the estimated coefficients of average temperature for the corresponding season. Dots correspond to the coefficients estimated over the sample starting with the year reported on the horizontal axis and ending in 2012. The panel regressions are for the entire crosssection of the U.S. Each state is weighted by its relative GSP. Regressions include state and year
fixed effects. The grey areas represent 90% confidence intervals. Standard errors are clustered
at the year level. The solid lines are linear fits of the dots in each panel.
positively affects growth in the utilities and mining sectors. Third, we show that the
effect of temperature on GSP is particularly strong in Southern states.
4.1
Effect on labor productivity
We study the possibility that temperature affects economic growth through labor productivity. Following Bernard and Jones (1996), we define labor productivity for each
state as the ratio between total private industry output and employment. The decision
18
to restrict our focus to private industries is dictated by the fact that the Bureau of Labor
Statistics reports data on state-level employment only for private industries. We verify
in our robustness checks (see section 5.3) that the main results reported in table 1 are
still valid for this specific subset of industries. Similarly, our choice to analyze labor
productivity as opposed to total factor productivity is based on data availability.7
In the top panel of table 3 we report the results of our analysis of the growth rate of
annual labor productivity. Specifically, we estimate the coefficients of the following specification:
∆ai,t =
X
βs Ti,s,t + ρ∆ai,t−1 +αi +αt +εi,t ,
(9)
s∈S
where ∆ai,t denotes the growth rate of productivity in state i at year t, and all other
variables are defined as in the previous sections. Specification (9) corresponds to our
baseline specification in (4) but replaces the growth rate of GSP with the growth rate
of productivity. The last two columns of table 3 document that summer and fall temperatures again have significant effects on the growth rate of labor productivity. These
results confirm our findings in table 1 and also provide a possible pathway through
which seasonal temperatures may affect economic growth. Specifically, an increase in
the average summer temperature negatively affects productivity growth, which in turn
results in a reduction in output. A drop in the average fall temperature seems to be
detrimental for productivity, thus resulting in a lower growth rate of GSP.
The bottom panel of table 3 reports the results of our analysis of the growth rate of
employment for private industries. The estimates in this panel correspond to the specification in equation (4), except that the dependent variable is the growth rate of employment rather than the growth rate of GSP. The results indicate that the association
between average summer and fall temperatures and the growth rate of employment is
not statistically significant. Taken together, the results in the top and bottom panels of
7
Garofalo and Yamarik (2002) built a dataset for state-level real capital stock. However, the sample
over which real GSP and real capital stock are deflated using the same method is limited, thus impairing
the construction of a panel of total factor productivity series.
19
Table 3: Effects of temperatures on productivity growth and employment growth
Productivity
Employment
Winter
−0.033
Spring
−0.020
(0.067)
(0.041)
(0.055)
(0.065)
(0.031)
(0.028)
0.013
−0.086
(0.032)
(0.015)
(0.024)
(0.051)∗
(0.051)∗
(0.055)
Summer
−0.152
(0.087)∗
(0.050)∗∗∗
(0.063)∗∗
Fall
0.132
(0.048)∗∗∗
(0.054)∗∗
(0.039)∗∗∗
0.008
−0.021
(0.059)
(0.037)
(0.049)
(0.042)
(0.019)
(0.032)
Notes: This table reports results for panel regressions of state productivity growth rate on temperatures, using the entire cross-section of 50 states and the District of Columbia. Productivity
is defined as output over employment in the private sector. All specifications include the lagged
dependent variable, state and year fixed effects. States are weighted in the panel regression
by the proportion, averaged over the whole sample, of their GSP relative to that of the whole
country. The columns refer to the analysis conducted by regressing jointly on the four seasonal
averages. Winter is defined as Jan.–Mar., spring as Apr.–Jun., summer as Jul.–Sep., fall as Oct.–
Dec. Temperatures are in degrees Fahrenheit. The sample is 1990–2011. The standard errors,
clustered by year, by state, and by both dimensions, are in parentheses. ***, **, and * denote
significance at the 1%, 5%, and 10% levels.
table 3 suggest that a main mechanism through which summer and fall temperatures
affect GSP growth is productivity growth, rather than employment growth.
Our results are in line with other findings in the literature. For example, Cachon et al.
(2012) document that heat and snow significantly affect output and productivity in automobile plants. The occurrence of six or more days with temperatures above 90 degrees
Farenheit reduces the weekly production of U.S. automobile manufacturing plants by an
average of 8 percent. Given that automobile manufacturing largely takes place indoors,
the authors argue that this finding suggests there are limitations of air conditioning;
it is possible that there are important areas in the production process, such as loading
and unloading areas, that are difficult to cool or warm. Bloesch and Gourio (2015) also
document that cold weather negatively affects production in various industries. We will
return to this discussion in the industry analysis below.
Several other studies also document effects of temperature on productivity and performance. In a survey of workplace and laboratory studies with objective measures of per-
20
formance, Seppänen et al. (2006) document that performance at office tasks decreases at
high temperatures. Similarly, Adhvaryu et al. (2014) find that productivity in garment
factories in Bangalore, India, decreases at high temperatures. Using repeated cognitive assessments from the National Survey of Youth, Zivin, Hsiang and Neidell (2015)
study the effect of short-run weather shocks on cognitive performance and find that an
increase in outdoor temperature decreases math performance.
4.2
Industry analysis
A common perception is that the effects of global warming are limited to agriculturerelated sectors, which constitute a relatively small fraction of GDP in developed countries. In this section, we revisit this idea and explore the composition of the effects
documented in the panel regressions of table 1. Our analysis is guided by existing microlevel evidence of possible mechanisms through which temperature may affect economic
activities.
First, high temperatures negatively affect the productivity of workers, especially in what
Cachon et al. (2012) call interface areas, such as loading and unloading areas, which are
difficult to cool with air conditioning. This constitutes a potential pathway through
which high temperatures may affect sectors such as retail, wholesale, and construction.
Second, the fact that high temperatures exert a negative impact on health is well established in the literature. For instance, Isaksen et al. (2015a; 2015b) have documented
that hot, humid days increase the risk of hospitalization and death in the state of Washington. Choudhary and Vaidyanathan (2014) provide evidence that increases in summer
temperatures are associated with an increase in heat stress illness hospitalizations.8 Focusing on community hospitals, Merrill et al. (2008) report that the hospitalization costs
8
Additionally, Chan et al. (2013) have shown that during the hot season in Hong Kong, hospital admissions increased by 4.5% for every increase of 1 degree Celsius above the seasonal average temperature.
21
from exposure to heat are in the order of $40 million per year, billed roughly equally to
government payers (Medicare and Medicaid) and private insurance companies. Since
the increase in hospitalization costs can lead to an increase in insurance payouts, we
therefore hypothesize that an increase in summer temperatures can negatively affect
the insurance sector.
Third, several cognitive biases may be at work to explain the negative impact of rising
summer temperatures on several sectors of the economy. High summer temperatures
may affect household demand in the retail sector. For instance, high temperatures may
adversely impact what Starr-McCluer (2000) call “households’ shopping productivity”
and in addition may negatively affect customers’ perception of wait time (Baker and
Cameron (1996)) and social interactions with strangers (Griffit and Veitch (1971)), inducing them to spend less time shopping. High summer temperatures may also affect
the real estate sector, as the real estate market is characterized by a “search-and-match”
mechanism, a large part of which takes place outdoors, and many prospective homebuyers search for houses in the summer (Ngai and Tenreyro (2014)). There may also be
spillovers between different sectors. For example, if agricultural income falls, agricultural workers may reduce their demand for other goods, affecting sectors such as retail
or real estate.
To investigate these conjectured industry-level effects, we break down the total GSP of
each state into 12 large industry groups according to the BEA classifications. These
groups are listed in descending order of national GDP share in the first column of table 4A (for a detailed list of industries in each group, see appendix table A2). The last
column of table 4A provides the post-1997 average share of national GDP for each group.
These groups are non-overlapping and together account for 100% of gross product.
One caveat of this exercise is a data limitation due to the change in the classification
of industries, from the Standard Industrial Classification system (SIC) to the North
American Industry Classification System (NAICS), that took place in 1997. In several
22
instances this substantially affects the composition of specific industries (see, for example, the breakdowns of the “Services” and “Communication/Information” categories in
appendix table A2). Furthermore, the BEA website warns that “users of GDP by state
are strongly cautioned against appending the two data series in an attempt to construct
a single time series”.9 In order to prevent our results from picking up effects that may
be due to these changes, we report the results of our analysis over two separate subsamples (pre- and post-1997). The split at 1997 significantly reduces the sample size and,
hence, the power of our statistical analysis. For this reason, we estimate only the effect
of summer temperature—the season whose effects on economic growth is strongest in
our analysis.10
Specifically, for each group of industries j, we estimate the following equation:
j
j
j
= βsummer
Ti,summer,t + ρ∆yi,t−1
+αi +αt +εi,t ,
∆yi,t
(10)
j
where ∆yi,t
denotes the output growth of industry group j in state i at year t.
As a benchmark, we also regress equation (10) where j is total GSP (i.e., we repeat
regression (4) for the pre- and post-1997 samples separately and dropping all seasonal
temperature variables except for the summer). We report the results in the first row of
table 4A. Consistent with our previous finding, the table shows that the estimated effect
of the average summer temperature on GSP growth appears to be larger in the most
recent portion of the sample.
9
The full cautionary note is available at https://www.bea.gov/regional/docs/product/.
In figure A2 of the appendix, we document that adding additional seasons to the specification does not
alter our main conclusion and only marginally affects the statistical significance of the results. Furthermore, we show that by extending the sample to include the entire pre-1997 sample, it appears that the
results are primarily driven by “Agriculture”, i.e. the sector that has traditionally been the most exposed
to high temperature. However, as noted, these results need to be interpreted with caution due to the
BEA’s recommendation against combining pre- and post-1997 series.
10
23
Gross state product
Table 4A: Industry analysis
Pre-1997
Post-1997
−0.188 ∗∗
−0.250
(0.095)
(0.062)∗∗∗
(0.076)∗∗
Services†
0.019
(0.070)
(0.050)
(0.062)
Finance, insurance, real estate
−0.209
(0.241)
(0.228)
(0.271)
Manufacturing
Government
Retail
Wholesale
Utilities
−0.051
(0.071)
(0.063)
(0.070)
(0.164)
(0.086)
(0.128)
−0.052
−0.241
−0.158
12.9
12.2
6.6
(0.189)
(0.083)∗∗∗
(0.146)∗
−0.284 ∗
5.9
−0.235 ∗∗∗
−0.294
4.5
−0.224
−0.379
(0.171)
(0.163)∗
(0.164)∗
(0.732)
(0.405)
(0.662)
0.150
0.189
(0.221)
(0.138)
(0.187)
0.338
−0.152
−2.489 ∗∗
(0.995)
(0.443)∗∗∗
(0.952)∗∗∗
Continues on next page.
24
4.4
(0.446)
(0.194)∗
(0.372)
(0.125)
(0.196)
(0.187)
(0.539)
(0.572)
(0.515)
Agriculture, forestry, fishing
20.5
(0.384)
(0.158)∗∗∗
(0.329)
−0.068
(0.248)
(0.202)∗
(0.220)
Mining
−0.437
0.067
(0.236)
(0.199)
(0.232)
Transportation
25.7
(0.075)
(0.076)∗∗∗
(0.064)∗∗∗
(0.623)
(0.420)
(0.513)
(0.088)
(0.092)∗∗
(0.068)∗∗∗
Construction
−0.206 ∗∗∗
(0.215)
(0.102)
(0.160)
(0.104)
(0.062)∗∗
(0.084)∗
Communication/Information†
(0.197)
(0.067)∗∗∗
(0.156)
−0.058
(0.073)
(0.060)
(0.070)
Avg. GDP share (%)
100
3.0
0.621
1.8
0.954
1.4
(0.377)∗
(0.230)∗∗∗
(0.264)∗∗
(1.524)
(0.300)∗∗∗
(1.251)
−2.203 ∗∗
(0.969)
(0.502)∗∗∗
(0.751)∗∗∗
1.1
Table 4B: Industry group analysis: Services and Finance, Insurance, Real Estate
Post-1997 Ave GDP share (%)
Professional and business services
−0.219
11.6
(0.127)∗
(0.098)∗∗
(0.076)∗∗∗
Educational services, health care, social assistance
−0.004
7.7
(0.047)
(0.064)
(0.043)
Other services, except government
−0.253
2.6
Food services and drinking places
−0.387
2.0
0.417
1.0
Arts, entertainment, and recreation
(0.136)∗
(0.103)∗∗
(0.099)∗∗
(0.155)∗∗
(0.148)∗∗∗
(0.127)∗∗∗
(0.274)
(0.203)∗∗
(0.229)∗
Accommodation
0.025
0.9
(0.270)
(0.359)
(0.335)
Finance, insurance, real estate
Real estate
−0.435
11.4
(0.399)
(0.125)∗∗∗
(0.333)
Federal Reserve banks, credit intermediation,
and related services
−0.254
Insurance, carriers and related activities
−1.299
2.6
Securities, commodity contracts, and investments
−0.287
1.3
3.6
(0.463)
(0.354)
(0.407)
(0.630)∗∗
(0.548)∗∗
(0.632)∗∗
(0.531)
(0.337)
(0.375)
Rental and leasing services, lessors of intangible assets
−0.030
1.3
(0.244)
(0.290)
(0.169)
Funds, trusts, and other financial vehicles
1.027
0.2
(1.142)
(1.068)
(0.970)
Notes: This table reports results for panel regressions of industry output growth, using the entire
cross-section of 50 states and DC. Industries are classified according to the BEA (see appendix
table A2). All specifications include the lagged dependent variable, and state and year fixed
effects; the independent variable is the average summer temperature. States are weighted in
the panel regression by the proportion, averaged over the whole sample, of their industry output
relative to the whole country’s. The sample is 1997–2011. The last column reports the share of
national GDP that each industry accounts for. Standard errors, clustered by year, by state, and
by both dimensions, are in parentheses. ***, **, and * denote significance at the 1%, 5%, and
25
10% levels.
j
Our results for the estimate of βsummer
for each industry group are reported in table 4A.
The columns labeled “Pre-1997” and “Post-1997” correspond to the estimates for the
1963-1997 and 1997-2011 samples, respectively. As before, for each coefficient estimate,
we report three standard errors, the first clustered by year, the second clustered by state,
and the third clustered in both dimensions.
The two largest sectors of the U.S. economy, “Services” and “Finance, insurance, and
real estate”, account together for almost a half of national GDP and thus warrant further decomposition. Table 4B decomposes these two sectors into an exhaustive list of
subcomponents, using the post-1997 sample and the post-1997 NAICS classifications.
(We have to omit the pre-1997 sample as the pre-1997 SIC classifications do not offer a
decomposition of these two sectors.)
Several important findings emerge from tables 4A and 4B, especially in the post-1997
sample. Table 4A provides evidence for the conjecture that an increase in the average
summer temperature negatively affects the retail and wholesale sectors, which account
for 6.6% and 5.9% of national GDP, respectively. In the post-1997 sample, a 1o F increase
in the average summer temperature is associated with a 0.24 to 0.28 percentage points
reduction in the output growth of these sectors. Table 4A also provides some evidence
for the conjectured effect of summer temperatures on the construction sector, which
accounts for 4.4% of national GDP, however with lesser statistical significance, possibly
because of the short sample.
Table 4B also provides evidence for the conjecture that an increase in the average summer temperature negatively affects the real estate sector, which accounts for 11.4% of
national GDP. In the post-1997 sample, an increase of 1o F in the average summer temperature is associated with a 0.44 percentage points reduction in the output growth of
this sector. While this estimated impact may seem excessively large at first glance, we
note that the average volatility of the growth rate of output of the real estate sector is
3.5%, and the average volatility of summer temperature is equal to 1.5. By re-scaling
26
the estimated coefficient by the the ratio of these volatilities, we obtain a value of −0.19,
which can be interpreted as the number of standard deviations that the growth rate
of output of this industry moves in response to a one standard deviation movement in
summer temperature.11
Furthermore, as we previously hypothesized, table 4B shows that an increase in the
average summer temperature negatively affects insurance, insurance carriers and related activities, which account for 2.6% of national GDP. The effect is substantial: in the
post-1997 sample, a 1o F increase in the average summer temperature is associated with
a 1.30 percentage points reduction in the output growth of this sector.12
Table 4A shows that an increase in the average summer temperature has a substantial
negative effect on agriculture, forestry, and fishing. While the effect on this sector is
intuitive and well studied in the literature (see, inter alia, Mendelsohn and Neumann
(1999), Schlenker and Roberts (2006; 2009), and Burke and Emerick (2015)), we note
that this sector only accounts for about 1% of national GDP. However, its effects also
propagate to other sectors. For example, rising summer temperatures are associated
with a decline in “Food services and drinking places”, which account for about 2% of
national GDP (see table 4B). An effect on households’ shopping productivity, similar to
the one mentioned above for the retail sector, may lead to a decline in the demand for
food services and drinking places due to hot summer temperatures.
Not all industry groups are negatively affected by an increase in summer temperatures.
The utilities and mining sectors, accounting for about 1.8% and 1.4% of national GDP,
respectively, appear to benefit from an increase in the average summer temperature
(see table 4A). This could be due to the higher consumption of energy during warmer
11
In figure A1 of the appendix, we present all of the the regression coefficients for the analysis of tables 4A and 4B using variables standardized by their volatilities and show that the standardized effect is
very similar in the cross-section of industries.
12
The growth rate of output in the insurance sector is very large (about 10% per year). We show in
figure A1 in the appendix that after taking into account this volatility, the effect of temperature in this
sector is comparable to effect on real estate.
27
summers, which translates into larger revenues for these industries.
Overall, our results suggest that the effects of summer temperatures on aggregate economic activity are not due to the isolated impact of rising temperatures on just a few
sectors of the economy. Rather, higher temperatures systematically affect a large crosssection of industries, which in total account for more than a third of national GDP.
4.3
Regional analysis
To determine whether certain broad geographic areas are primarily responsible for the
effects of seasonal temperatures on GSP, we divide the U.S. into four regions: North,
South, Midwest, and West. These regions are identified according to the classification of
the U.S. Census Bureau (see appendix A.4 for the list of states in each region). We then
estimate the effects of seasonal temperatures for each region using a panel regression
of the growth rate of state-level GSP on temperatures.
The results of this regional analysis (reported in table 5) document that the effects
of summer and fall temperatures are statistically and economically significant in the
South. The estimated coefficients for the South are substantially larger than their
country-level counterparts identified in table 1. This indicates that the growth rates
of GSP in Southern states are particularly sensitive to summer and fall temperatures,
while other regions do not appear to be systematically affected.
We argue that the significance of the estimated coefficients for the South region can be
attributed to the relatively higher average temperatures that characterize the states in
this area. To provide evidence in support of this claim, we sort states in descending
order, according to their average summer temperature. As expected, the states in the
South region occupy the highest positions (see appendix table A3). We then estimate
the regression coefficients of equation (4) for the ten states with the highest summer
28
temperatures, and successively re-estimate these coefficients, each time adding the next
temperature-sorted state. The results of this exercise are reported in figure 4.
The bottom left panel of figure 4 documents that the estimated summer coefficient for
the ten warmest states is about three times as large as their whole-country counterpart.
The absolute value of the summer coefficient declines sharply past the first 15 states,
thus highlighting a nonlinearity in the impact of rising temperatures for this season.
Furthermore, a comparison of the bottom two panels of figure 4 reveals that the dramatic rise of the impact coefficient for the warmest states is precisely identified for the
average summer temperature, whereas the coefficient of the average fall temperature is
characterized by a higher degree of uncertainty. Winter and spring temperatures do not
seem to play a major role in this part of our analysis.13
We conclude this section by establishing a potential connection with the industry analysis of section 4.2. As shown in the right panel of figure 5, the relative contribution of
the South to the overall GSP/GDP of the U.S. has substantially increased during our
time period (1957-2012). Furthermore, it seems to be the case that this increase in the
South’s share of GDP has been a widespread phenomenon, involving the entire crosssection of industries (see the left panel of figure 5). Interestingly, the agricultural sector
has displayed the smallest percentage increase across the two sub-samples. Combined
with our regional results in table 5 this seems to suggest that the increased effect that
we estimate in the post-1997 sample is driven by a larger overall contribution of the
South to the country’s economic activity.
13
Our analysis does not preclude the possibility that the average temperature of other seasons may have
additional economic effects. For example, an increase in the average Winter temperature may plausibly
have a positive economic effect in the North, and our empirical evidence cannot reject this null hypothesis.
However, we believe that more evidence is needed to fully establish this channel and leave this as an open
task for future research.
29
Table 5: Effects of seasonal temperatures on GSP growth in different regions
North
South
Winter
0.329
Spring
0.065
Summer
0.240
(0.173)∗
(0.238)
(0.216)
(0.296)
(0.176)
(0.233)
(0.257)
(0.232)
(0.235)
0.152
−0.326
−0.087
(0.167)
(0.077)
(0.142)
Midwest
West
(0.159)
(0.087)∗
(0.130)
(0.163)∗∗
(0.085)∗∗∗
(0.129)∗∗
0.010
−0.158
0.043
(0.089)
(0.055)
(0.074)
(0.144)
(0.104)
(0.125)
(0.162)
(0.075)
(0.130)
−0.000
−0.155
(0.096)
(0.060)
(0.056)
(0.143)
(0.077)∗∗
(0.097)
Fall
−0.255
(0.233)
(0.184)
(0.186)
0.571
(0.194)∗∗∗
(0.063)∗∗∗
(0.157)∗∗∗
−0.116
(0.128)
(0.068)∗
(0.112)
0.028
−0.006
(0.154)
(0.145)
(0.153)
(0.167)
(0.162)
(0.174)
Notes: This table reports results for panel regressions of state GSP growth rate on temperatures,
using the cross-section of U.S. states in each region. Regions are classified according to the Census Bureau. All specifications include the lagged dependent variable, and state and year fixed
effects. States are weighted in the panel regression by the proportion, averaged over the whole
sample, of their GSP relative to the region’s GDP. The columns refer to the analysis conducted
by regressing jointly on the four seasonal averages. Winter is defined as Jan.–Mar., spring as
Apr.–Jun., summer is Jul.–Sep., fall is Oct.–Dec. Temperatures are in degrees Fahrenheit. The
sample is 1957–2012. Standard errors, clustered by year, by state, and by both dimensions, are
in parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% levels.
5
Additional results
In this section we conduct a series of additional exercises to confirm and extend the
analysis above. In section 5.1 we demonstrate that the effect of temperature on GDP
growth is not driven by less developed states. Although this result differs from those
reported in the literature, even U.S. states that are considered less developed are still
highly developed according to common measures of global poverty. In section 5.2 we
combine the estimated impact coefficients from section 3 with various projections of
temperature changes over the next 100 years to provide an assessment of the long-term
impact of rising temperature on U.S. economic growth. In section 5.3 we conclude our
30
Estimated Coefficients
Winter
Spring
0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2
-0.2
-0.4
-0.4
-0.6
10
-0.6
15
20
25
30
35
40
45
10
15
20
25
Estimated Coefficients
Summer
35
40
45
35
40
45
Fall
0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2
-0.2
-0.4
-0.4
-0.6
10
30
-0.6
15
20
25
30
35
40
45
10
Number of States
15
20
25
30
Number of States
Figure 4: Effects of seasonal temperatures in temperature-sorted states. Each panel reports the
estimated coefficients of the average temperature for the corresponding season. Dots correspond
to the coefficients estimated for the number of states reported in the horizontal axis. The grey
areas represent 95% confidence intervals. States are sorted in descending order according to
their average summer temperature. Each state is weighted by its relative GSP in the panel
regressions. State and year fixed effects are included. Standard errors are clustered at the year
level.
investigation by showing that our main finding on the effect of summer temperatures is
robust to alternative specifications.
5.1
Development and Temperature
In this section we explore the interaction between the level of development of a state
and seasonal temperatures. Specifically, we define development levels according to the
Human Development Index (HDI) for American states developed by Lewis and Burd-
31
Share of each industry output produced in the South
Share of US GDP produced in the South
70.00
0.400
60.00
1963-1997
0.350
1997-2011
50.00
0.300
40.00
0.250
30.00
0.200
20.00
0.150
10.00
0.100
0.00
0.050
0.000
1957-1960 1961-1970 1971-1980 1981-1990 1991-2000 2001-2010 2011-2012
Figure 5: Percent of total U.S. GDP produced in the South region. The left panel reports the
break-down by industry. The right panel reports the share of total GDP produced in the South
by decade (or fraction of it) from 1957 to 2012.
Sharps (2014). We introduce an HDI indicator variable equal to 1 for states with values
of the Human Development Index less than or equal to 4.5, which corresponds to the bottom 20% of the distribution. With the exception of Idaho (which is located in the West
region), all of the less developed states are in the South region. Specifically, the following
states in the South are classified as states with “low development”: Alabama, Arkansas,
Kentucky, Louisiana, Mississippi, Oklahoma, South Carolina, Tennessee, and West Virginia. The following states in the South are instead classified as states with “high development”: Delaware, D.C., Florida, Georgia, Maryland, North Carolina, Texas, and
Virginia.
We use the following regression specification to investigate the hypothesis that less developed states are driving the strong negative effect of summer temperature on GSP
32
Table 6: Effects of seasonal temperature by development level
βs
Winter
Spring
Summer
−0.111
Panel A: South
0.139
−0.375
(0.188)
(0.069)
δs
βs
(0.179)∗∗
(0.090) ∗∗∗
0.053
0.068
0.131
(0.072)
(0.046)
(0.136)
(0.111)
(0.120)
(0.064) ∗∗
0.005
(0.049)
(0.026)
δs
(0.185)
(0.116)
Panel B: U.S.
−0.008
(0.065)
(0.032)
−0.155
(0.075) ∗∗
(0.050) ∗∗∗
Fall
0.585
(0.203) ∗∗∗
(0.065) ∗∗∗
−0.042
(0.124)
(0.060)
0.099
(0.052) ∗
(0.040) ∗∗
−0.031
0.121
0.003
0.036
(0.055)
(0.043)
(0.106)
(0.075)
(0.113)
(0.063)
(0.111)
(0.059)
Notes: Results of estimating equation 11. HDI is an indicator equal to 1 for states with HDI
value less than or equal to 4.5. The sample is 1957–2012. Standard errors are in parentheses.
The first set of standard errors is clustered by year and the second set of standard errors is
clustered by state. ***, **, and * denote significance at the 1%, 5%, and 10% levels.
growth in the South:
∆yi,t =
X
(βs Ti,s,t + δs Ti,s,t · I[HDIi ]) + ρ · ∆yi,t−1 +αi +αt +εi,t ,
(11)
s∈S
where ∆yi,t denotes GSP growth in state i in year t, and Ti,s,t denotes the temperature in
degrees Fahrenheit in state i, year t, and season s. We include the interaction between
I[HDIi ], an indicator for low development, and seasonal temperatures. We also include
the lagged GSP growth rate as a control to capture the degree of autocorrelation of the
dependent variable, in addition to αi and αt , which denote state and year fixed effects.
Panel A of table 6 reports the results of estimating the specification above in the South
region and Panel B reports the results for the whole U.S.. Two sets of standard errors are
reported in the table, the first clustered by year and the second clustered by state.14 Both
14
We report only the standard errors clustered in one dimension, because the standard errors clustered
33
in the South region and in the whole country, the coefficient on summer temperature is
negative and significant and the coefficient on the interaction of the indicator for lower
HDI and summer temperature is positive and marginally significant. This indicates
that the negative effect of summer temperature documented in table 5 is not driven by
less developed states.15
The results reported in table 6 are likely driven by the different industrial composition
of more and less developed states in the South, with GSP in states with higher development more highly concentrated in industries exposed to summer temperature. For
example, in the South, the finance sector constitutes a greater percent of GSP in high development states (22% for 1997-2011) than in less developed states (14% for 1997-2011).
Although less developed states have a greater percent of GSP from the agricultural sector, agriculture accounts for a very small fraction of total GSP (see figure 5).
These results likely differ from those reported in the literature because even U.S. states
that are considered less developed are still well above the poverty line, according to the
most common measures of poverty. For example, Arkansas, which is ranked second from
the bottom in terms of the American Human Development Index, has a 2016 GDP per
capita of about $36,000. According to the IMF, this is comparable to the GDP per capita
of developed countries such as Israel, Italy, and Spain. Although it is certainly true that
there is a marked socioeconomic North-South divide in the US, Southern states are still
highly developed according to most international metrics. This implies that the channel
through which temperature affects GDP in this part of the country must be distinct
from that typically established for developing economies. We believe that our evidence
on the widespread effect of temperature on sectors other than agriculture is the key to
explaining this result.
in two dimensions become unreliable when the cluster size shrinks.
15
We have replicated the analysis in equation (11) by focusing on the 10 hottest states in the US. The
results of our estimation for this subset of states indicate that the coefficient βsummer is equal to −0.485
with standard errors of 0.239 and 0.153 (depending on clustering), while the coefficient δsummer is equal to
0.180 with standard errors of 0.154 and 0.110 (depending on clustering). This supports the idea that the
negative effect of temperature on GDP growth is primarily driven by the higher development states.
34
5.2
Combining our results with climate projections
In this section we provide a quantification of the magnitudes of the effects of summer
and fall temperatures estimated in panel regression (4) over a longer horizon. This exercise needs to be interpreted with caution, since it assumes that the impact coefficients
estimated in our main analysis do not change over the time period under consideration,
and it ignores the uncertainty about the point-estimates of the coefficients. Equivalently, one can interpret this case as a “business as usual” benchmark, in which there
is no adaptation or mitigation, and the effects of temperatures on economic growth in
section 3 remain unchanged over the long horizon.
To quantify the potential long-term relevance of the coefficients estimated in section 3,
we employ temperature projections obtained from the Climate Wizard tool (http://
ClimateWizard.org) developed by Girvetz et al. (2009). We use this tool to obtain
projected monthly average temperatures for the U.S. for the period 2070-2099 from 16
general circulation models (GCMs) under three different IPCC greenhouse gas emissions scenarios: A2 (high emissions), A1B (medium emissions), and B1 (low emissions).
For each model and scenario, we consider both the minimum and maximum projected
temperature change in our analysis.
We combine each set of temperature projections with the impact coefficients that we
estimated in section 3. Throughout our analysis, we focus on the coefficients for only the
summer and the fall seasons, given the lack of statistical significance of the coefficients
for winter and spring. We compute the projected impact on the growth rate of GDP as:
E [∆GDP ] =
X
E [∆Ts ] × β̂s ,
s∈{summer,f all}
where E [∆Ts ] and β̂s denote the expected change in the average temperature of season
s and the impact coefficient of season s, respectively. Throughout our analysis, we use
35
Projected reduction in the growth rate of GDP
0
-0.2
-0.4
-0.6
-0.8
-1
-1.2
A2 (High)
A1B (Medium)
B1 (Low)
Figure 6: Projected reduction in the growth rate of GDP for the period 2070-2099 under three
emission scenarios. For each scenario, the bottom and top lines denote the minimum and maximum projected impact, the bottom and top of the rectangle are the first and third quartile of the
distribution of projected impacts, while the horizontal line within the rectangle is the median
projected impact.
β̂summer = −0.154 and β̂f all = 0.102, as reported in Table 1.
We report the results of our analysis in figure 6. Under the most conservative emission scenario (B1), the projected trend in rising temperatures is expected to reduce the
growth rate of U.S. output by 0.2 to 0.4 percentage points over the next 100 years, depending on the specific GCM employed. These figures are not negligible: given a historical average growth rate of nominal U.S. GDP of about 4% per year, our first set of
estimates implies a reduction of the growth rate by up to 10%.
The results are more dramatic when we use the projections obtained under the more
aggressive emission scenarios. For instance, under the High emission scenario (A2), the
estimated reductions in output growth due to rising temperatures could be as large as
1.2 percentage point. Thus, assuming no change in the way in which seasonal tempera-
36
tures affect economic growth, the projected increases in summer and fall temperatures
could potentially reduce economic growth by roughly a third of the historical average
nominal U.S. GDP growth rate.
5.3
Robustness checks
In this section we check the robustness of our results to different specifications of main
regression (4). The results are reported in table 7. Throughout the table (except the
row “Spatial correlation”), we report three standard errors, one clustered by year, one
clustered by state, and one in both dimensions, with the corresponding significance levels. Overall, the table shows that the negative relationship between average summer
temperature and GSP growth is very robust. We also document that the positive relationship between average fall temperature and growth is not supported in several
robustness checks.
The panel labeled “Alternative panel weights” reports the results obtained from using
different weighting schemes for the states in the panel regression. Specifically, we
weight states by population, area, and time-varying GSP. The last weighting scheme
takes into account possible changes over time in the relative distribution of output
across states (see appendix A.2.2). The results indicate that the signs of the estimated
coefficients are generally aligned with the main findings in section 3.
In the panel labeled “Alternative GSP measures” we report the results obtained by replacing the dependent variable of our regression with per-capita GSP, real GSP, or private industries’ GSP. The results of the regressions using these alternative measures
of GSP demonstrate that our earlier results are not driven by the growth rate of population, inflation, or the public sector. The alternative measure results also confirm our
main finding that an increase in average summer temperature has a strong negative effect on economic growth rates. In some cases, the magnitudes of the estimated summer
37
coefficients are even larger than those obtained in our baseline specification. The effect
of the fall season, however, is less robust: an increase in the average fall temperature
does not appear to have a significant effect on real GSP growth.
We also check the robustness of our findings to various definitions of seasons (see the
panel “Alternative definitions of seasons” in table 7. Specifically, in the row labeled “Meteorological,” all seasons are shifted backwards by one month. This means that winter
is defined as including December, January, and February; spring is defined as March,
April, and May; summer is defined as June, July, and August; and fall is defined as
September, October, and November. In the row labeled “Core Seasonal Months,” we
focus only on the subset of months that fall within both the astronomical and meteorological definitions of a given season. Here, winter is defined as January and February,
spring as April and May, summer as July and August, and fall as October and November.
The results indicate that the summer effect is generally robust to the various definitions
of seasons. When we adopt the meteorological definition, the coefficient on summer temperature is negative, although only statistically significant at the 5% level if standard
errors are clustered by state. This may be due to the inclusion of the transitional month
of June, during which temperatures have not yet fully adjusted to the seasonal summer
average. Indeed, when we only focus on the subset of months that are associated with
both the astronomical and meteorological definitions of each season, we get consistently
strong results for the summer (see the row “Core Seasonal Months”). This suggests that
the economic effect of summer temperatures is mainly driven by the months of July
and August. The fall effect, in contrast, is not significant under any of the alternative
seasonal definitions.
In the panel labeled “Alternative temperature data” in table 7, we check the robustness
of our results to different aggregation methods, deseasonalization methods, and sources
of temperature data. As the panel shows, both the summer and fall effects are robust
to aggregating station-level weather data to the state level using county population in-
38
stead of county area (see the row “Temp. weighted by pop.”) and to deseasonalizing temperatures using pre-1950 monthly dummies (see the row “Pre-1950 deseasonalization”).
Appendix A.2.1 describes the method that we used to deseasonalize the temperature
data. Furthermore, in the row “Non-deseasonalized gridded temp.,” we employ gridded
temperature data that is not deseasonalized from the NOAA nClimDiv dataset to show
that the deseasonalization of weather data does not drive our results.
In the panel labeled “Other” in table 7 we check the robustness of our results to several
additional variations of our main specification. In the row labeled “Spatial correlation,”
we adjust standard errors to take into account the possible dependence induced by the
geographical proximity of the states. Specifically, we employ the correction proposed by
Conley (1999) and adapted by Hsiang (2010) to the study of climate-related variables
with spatial correlation. We used a radius of 300 km around the center of each state,
with a uniform spatial weighting kernel. The ordinary least squares regression is an
unweighted state-level panel regression. Our results again show that the summer effect
is statistically significant, at the 10% level, but the fall effect is not.
We also include average precipitation (the row “Controlling for precipitation”) and temperature volatility (the row “Controlling for temp. vol.”) in our main specification. The
temperature volatility of season s in year t is calculated as the standard deviation of the
deseasonalized temperature observations in that season (see appendix A.2.1 for details
on deseasonalization). We find that controlling for these two additional sets of control
variables does not alter our main conclusions regarding the effect of summer and fall
temperatures on GSP growth.
Our results are robust to the exclusion of the lagged growth rate of GSP. This finding
is important in light of the so-called Nickell (1981) bias, which arises in the context of
dynamic panel models with fixed effects in a short sample. The results shown in row
“Excluding AR(1)” of table 7 are from panel regressions that do not included lagged
GSP. As shown, the negative effect of summer temperature is still economically and
39
statistically significant. We also note that the magnitudes of the estimated coefficients
are very close to the ones obtained in table 1, which can be interpreted as evidence
of a small overall impact of the bias on our results. In related studies, Judson and
Owen (1999), Acemoglu et al. (2014), and Deryugina and Hsiang (2014) reach similar
conclusions regarding the extent of the bias.
Finally, the row labeled “Excluding Alaska and Hawaii” shows that our results are robust to excluding the two non-contiguous states of Alaska and Hawaii.
In summary, the battery of tests using various alternative specifications has shown that
the effect of summer temperatures is generally very robust, but that of fall temperatures
generally less so.16
6
Conclusion
In this paper, we analyze the effects of increases in average seasonal temperatures on
economic growth across U.S. states. We find that an increase in the average summer
temperature has a significant and robust negative effect on GSP growth. We also find a
positive, albeit weaker and less robust, effect of an increase in the average fall temperature. In net, the summer effect dominates, and the total impact of increases in seasonal
temperatures is substantial: under the business-as-usual scenario, the projected trends
in rising temperatures could depress U.S. economic growth by up to a third.
Our results are informative for the calibration of the climate damage functions in general equilibrium models, and they also should be helpful in advancing the analysis of
the long-term effects of climate change (e.g., Stern (2007), Nordhaus and Sztorc (2013),
Bansal and Ochoa (2011), Giglio et al. (2015), and Donadelli, Jüppner, Riedel and Schlag
(2017)). These results highlight the importance of improving the next generation of equi16
In appendix section A.6, we provide a series of further robustness checks.
40
librium models for the environment along two dimensions. First, these models should
account for the heterogeneous effects that rising temperatures have on the cross-section
of industries. Second, these models should explicitly model the effects of seasonal temperatures on labor productivity and other economic variables.
Finally, the finding that the effect of summer temperatures is stronger in the states that
are on average warmer than the rest of the country is related to the nonlinear effects
of rising temperatures in the studies of Schlenker and Roberts (2006; 2009). Future research should employ methodologies from these studies to further investigate potential
nonlinearities in the effects of seasonal temperatures.
41
Table 7: Robustness checks
Winter
Alternative panel weights
Time-varying GSP
State population
State area
Alternative GSP measures
Per-capita GSP
Real GSP
Alternative definitions of seasons
Meteorological
Core seasonal months
Summer
−0.148
0.008
−0.008
(0.051)
(0.026)
(0.047)
(0.067)
(0.030)
(0.051)
0.028
−0.024
(0.053)
(0.025)
(0.048)
(0.069)
(0.039)
(0.060)
0.018
0.012
(0.062)
(0.033)
(0.058)
(0.074)
(0.045)
(0.068)
−0.007
0.018
(0.047)
(0.025)
(0.043)
(0.068)
(0.033)
(0.055)
−0.070
−0.016
(0.043)
(0.040)∗
(0.042)∗
Private industries only
Spring
(0.081)
(0.037)
(0.054)
0.013
0.010
(0.063)
(0.029)
(0.055)
(0.083)
(0.041)
(0.065)
0.025
−0.040
(0.043)
(0.016)
(0.033)
(0.053)
(0.038)
(0.044)
0.015
−0.026
(0.041)
(0.016)
(0.035)
(0.050)
(0.023)
(0.035)
,→ Continued on Next Page
42
(0.076)∗
(0.043)∗∗∗
(0.066)∗∗
−0.132
(0.071)∗
(0.039)∗∗∗
(0.060)∗∗
−0.098
(0.066)
(0.054)∗
(0.059)∗
−0.119
(0.071)∗
(0.048)∗∗
(0.064)∗
−0.194
(0.110)∗
(0.087)∗∗
(0.109)∗
−0.207
(0.087)∗∗
(0.060)∗∗∗
(0.076)∗∗∗
−0.083
(0.074)
(0.038)∗∗
(0.059)
−0.145
(0.066)∗∗
(0.033)∗∗∗
(0.055)∗∗∗
Fall
0.105
(0.058)∗
(0.042)∗∗
(0.057)∗
0.131
(0.061)∗∗
(0.043)∗∗∗
(0.062)∗∗
0.079
(0.063)
(0.063)
(0.075)
0.098
(0.053)∗
(0.040)∗∗
(0.053)∗
−0.006
(0.068)
(0.053)
(0.061)
0.114
(0.069)∗
(0.049)∗∗
(0.067)∗
0.025
(0.055)
(0.033)
(0.049)
0.036
(0.050)
(0.027)
(0.045)
Table 7: Robustness checks (continued)
Winter
Alternative temperature data
Temp. weighted by pop.
Pre-1950 deseasonalization
Non-deseasonalized gridded temp.
Other
Spatial correlation
Controlling for precipitation
Controlling for temp. vol.
Excluding AR(1)
Excluding Alaska and Hawaii
Spring
Summer
0.012
−0.004
−0.129
(0.048)
(0.023)
(0.043)
(0.066)
(0.028)
(0.052)
0.001
0.003
(0.049)
(0.025)
(0.044)
(0.065)
(0.032)
(0.051)
0.001
−0.005
(0.042)
(0.023)
(0.038)
(0.057)
(0.028)
(0.044)
0.011
−0.020
(0.046)
(0.061)
0.003
0.008
(0.047)
(0.025)
(0.043)
(0.069)
(0.039)
(0.057)
−0.009
−0.013
(0.050)
(0.024)
(0.045)
(0.062)
(0.030)
(0.046)
0.023
0.014
(0.052)
(0.029)
(0.049)
(0.073)
(0.039)
(0.058)
−0.001
−0.000
(0.048)
(0.026)
(0.044)
(0.065)
(0.032)
(0.051)
(0.074)∗
(0.041)∗∗∗
(0.061)∗∗
−0.154
(0.072)∗∗
(0.047)∗∗∗
(0.065)∗∗
−0.167
(0.064)∗∗∗
(0.047)∗∗∗
(0.058)∗∗∗
−0.109
(0.066)∗
−0.169
(0.077)∗∗
(0.048)∗∗∗
(0.069)∗∗
−0.138
(0.071)∗
(0.042)∗∗∗
(0.061)∗∗
−0.156
(0.080)∗
(0.054)∗∗∗
(0.073)∗∗
−0.153
(0.071)∗∗
(0.048)∗∗∗
(0.074)∗∗
Fall
0.094
(0.057)∗
(0.034)∗∗∗
(0.051)∗
0.102
(0.055)∗
(0.040)∗∗∗
(0.054)∗
0.100
(0.047)∗∗
(0.035)∗∗∗
(0.044)∗∗
0.024
(0.058)
0.093
(0.056)∗
(0.037)∗∗
(0.052)∗
0.106
(0.055)∗
(0.040)∗∗∗
(0.052)∗∗
0.086
(0.059)
(0.036)∗∗
(0.053)
0.118
(0.056)∗∗
(0.040)∗∗∗
(0.054)∗∗
Notes: This table reports robustness checks for main regression (4). Temperatures are in degrees
Fahrenheit. The sample is 1957–2012, except for the row with private industries only, in which
the sample is 1963–2011, and the row with real GSP, in which the sample is 1987–2012. In all
regressions except those in “Alternative panel weights” and “Spatial correlation,” each state is
weighted by the proportion, averaged over the whole sample, of its GSP relative to the whole
country’s GDP. In “Time-varying GSP,” each state in each year is weighted by the proportion of
its GSP relative to the whole country’s GDP in that year. In “State population” and “State area,”
each state is weighted by the proportion, averaged over the whole sample, of its population
or area, respectively. In the row “Core seasonal months,” winter is Jan.–Feb., spring is Apr.–
May., summer is Jul.–Aug., fall is Nov.–Dec. In the row “Spatial correlation”, all states are
equally weighted. Standard errors, clustered by year, by state, and by both dimensions, are in
parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% levels.
43
References
Acemoglu, Daron, Philippe Aghion, Leonardo Bursztyn, and David Hemous, “The Environment
and Directed Technical Change,” The American Economic Review, 2012, 102 (1), 131–166.
, Suresh Naidu, Pascual Restrepo, and James A. Robinson, “Democracy Does Cause Growth,”
National Bureau of Economic Research working paper No. 20004, March 2014.
Adhvaryu, Achyuta, Namrata Kala, and Anant Nyshadham, “The light and the heat: Productivity
co-benefits of energy-saving technology,” working paper, 2014.
Alvarez, Javier and Manuel Arellano, “The Time Series and Cross-Section Asymptotics of Dynamic
Panel Data Estimators,” Econometrica, 2003, 71 (4), 1121–1159.
Arellano, Manuel and Stephen Bond, “Some tests of specification for panel data: Monte Carlo evidence
and an application to employment equations,” The Review of Economic Studies, 1991, 58 (2), 277–
297.
Baker, Julie and Michaelle Cameron, “The effects of the service environment on affect and consumer
perception of waiting time: An integrative review and research propositions,” Journal of the Academy
of Marketing Science, 1996, 24 (4), 338–349.
Bansal, Ravi and Marcelo Ochoa, “Welfare costs of long-run temperature shifts,” National Bureau of
Economic Research working paper No. 17574, 2011.
,
, and Dana Kiku, “Climate Change and Growth Risk,” working paper, 2014.
Barreca, Alan, Olivier Deschenes, and Melanie Guldi, “Maybe next month? Temperature shocks,
Climate Change, and dynamic adjustments in birth rates,” National Bureau of Economic Research
working paper No. 21681, 2015.
Bernard, Andrew B and Charles I Jones, “Productivity and convergence across U.S. States and industries,” Empirical Economics, 1996, 21 (1), 113–135.
Bloesch, Justin and Francois Gourio, “The effect of winter weather on US economic activity,” Economic Perspectives, 2015, 39 (1).
Burke, Marshall and Kyle Emerick, “Adaptation to Climate Change: Evidence from US Agriculture,”
American Economic Journal: Economic Policy, 2015, fortchoming.
Burke, Marshall B, Edward Miguel, Shanker Satyanath, John A Dykema, and David B Lobell,
“Warming increases the risk of civil war in Africa,” Proceedings of the National Academy of Sciences,
2009, 106 (49), 20670–20674.
Cachon, Gerard, Santiago Gallino, and Marcelo Olivares, “Severe Weather and Automobile Assembly Productivity,” working paper, 2012.
Chan, EY, WB Goggins, JS Yue, and P. Lee, “Hospital admissions as a function of temperature, other
weather phenomena and pollution levels in an urban setting in China,” Bulletin of the World Health
Organization, 2013, 91(8), 576–584.
Choudhary, Ekta and Ambarish Vaidyanathan, “Heat Stress Illness Hospitalizations Environmental Public Health Tracking Program, 20 States, 2001–2010,” Morbidity and Mortality Weekly Report,
Surveillance Summaries, 2014, 63 (13), 1–16.
Conley, T.G., “GMM estimation with cross-sectional dependence,” Journal of Econometrics, 1999, 92,
1–45.
44
Dell, Melissa, Benjamin F Jones, and Benjamin A Olken, “Temperature shocks and economic
growth: Evidence from the last half century,” American Economic Journal: Macroeconomics, 2012, 4
(3), 66–95.
Deryugina, Tatyana and Solomon Hsiang, “Does the Environment Still Matter? Daily Temperature
and Income in the United States,” National Bureau of Economic Research working paper No. 20750,
2014.
Deschênes, Olivier and Michael Greenstone, “The economic impacts of climate change: Evidence
from agricultural output and random fluctuations in weather,” The American Economic Review, 2007,
97 (1), 354–385.
and
, “Climate change, mortality, and adaptation: Evidence from annual fluctuations in
weather in the US,” American Economic Journal: Applied Economics, 2011, 3 (4), 152–185.
and
, “The economic impacts of climate change: Evidence from agricultural output and random
fluctuations in weather: Reply,” The American Economic Review, 2012, 102 (7), 3761–3773.
,
, and John Guryan, “Climate change and birth weight,” American Economic Review Papers
and Proceedings, 2009, 99 (2), 211–217.
Donadelli, Michael, Marcus Jüppner, Max Riedel, and Christian Schlag, “Temperature shocks
and welfare costs,” Journal of Economic Dynamics and Control, 2017, 82, 331–355.
Field, Christopher B, Vicente R Barros, Michael D Mastrandrea, Katharine J Mach, Mohamed
A-K Abdrabo, W Neil Adger, Yury A Anokhin, Oleg A Anisimov, Douglas J Arent, Jonathon
Barnett et al., “Climate Change 2014: Impacts, Adaptation, and Vulnerability. Part A: Global and
Sectoral Aspects. Contribution of Working Group II to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change,” 2014.
Gallup, John Luke, Jeffrey D Sachs, and Andrew D Mellinger, “Geography and economic development,” International regional science review, 1999, 22 (2), 179–232.
Garofalo, Gasper A and Steven Yamarik, “Regional convergence: Evidence from a new state-by-state
capital stock series,” Review of Economics and Statistics, 2002, 84 (2), 316–323.
Giglio, Stefano, Matteo Maggiori, Johannes Stroebl, and Andreas Weber, “Climate Change and
Long-Run Discount Rates: Evidence from Real Estate,” Chicago, Harvard, and NYU Working Paper,
2015.
Girvetz, Evan H, Chris Zganjar, George T Raber, Edwin P Maurer, Peter Kareiva, and Joshua J
Lawler, “Applied climate-change analysis: the climate wizard tool,” PLoS One, 2009, 4 (12), e8320.
Golosov, Mikhail, John Hassler, Per Krusell, and Aleh Tsyvinski, “Optimal Taxes on Fossil Fuel in
General Equilibrium,” Econometrica, 2014, 82 (1), 41–88.
Griffit, William and Russell Veitch, “Hot and crowded: Influence of population density and temperature on interpersonal affective behavior,” Journal of Personality and Social Psychology, 1971, 17 (1),
92.
Hsiang, Solomon M, “Temperatures and cyclones strongly associated with economic production in the
Caribbean and Central America,” Proceedings of the National Academy of Sciences, 2010, 107 (35),
15367–15372.
Isaksen, Tania Busch, Michael G. Yost, Elizabeth K. Hom, You Ren, Hilary Lyons, Michael G.
Yost, and Richard A. Fenske, “Increased hospital admissions associated with extreme-heat exposure in King County, Washington, 1990–2010,” Reviews on Environmental Health, 2015, 30(1),
51–64.
45
, Richard A. Fenske, Elizabeth K. Hom, You Ren, Hilary Lyons, and Michael G. Yost, “Increased mortality associated with extreme-heat exposure in King County, Washington, 1980–2010,”
International Journal of Biometeorology, 2015, pp. 1–14.
Judson, Ruth A. and Ann L. Owen, “Estimating Dynamic Panel Data Models: A Guide for Macroeconomists,” Economic Letters, 1999, 65(1).
Lewis, Kristen and Sarah Burd-Sharps, “The American Human Development Index: The Measure of
America 2013-2014,” 2014.
Lucas, Robert E. Jr., Models of business cycles, New York: Basil Blackwell, 1987.
Mendelsohn, Robert, “Climate change and economic growth,” in “Globalization and growth: Implications for a Post-Crisis World,” The International Bank for Reconstruction and Development, 2010,
pp. 285–295.
and James Neumann, The Impact of Climate Change on the United States Economy, Cambridge,
UK: Cambridge University Press, 1999.
Merrill, Chaya T., Mackenzie Miller, and Claudia Steiner, “Hospital Stays Resulting from Excessive
Heat and Cold Exposure Due to Weather Conditions in U.S. Community Hospitals, 200,” HCUP
Statistical Brief 55, Agency for Healthcare Research and Quality, Rockville, MD, 2008.
Newey, Whitney K and Frank Windmeijer, “Generalized method of moments with many weak moment conditions,” Econometrica, 2009, 77 (3), 687–719.
Ngai, L Rachel and Silvana Tenreyro, “Hot and Cold Seasons in the Housing Market,” American
Economic Review, 2014, 104 (12), 3991–4026.
Nickell, Stephen, “Biases in dynamic models with fixed effects,” Econometrica: Journal of the Econometric Society, 1981, pp. 1417–1426.
Nordhaus, W and Paul Sztorc, “DICE 2013R: Introduction and User’s Manual, 2nd Edition,”
http://dicemodel.net 2013.
Nordhaus, William D, “Geography and macroeconomics: New data and new findings,” Proceedings of
the National Academy of Sciences of the United States of America, 2006, 103 (10), 3510–3517.
, A question of balance: Weighing the options on global warming policies, Yale University Press, 2014.
Pindyck, Robert S, “Modeling the Impact of Warming in Climate Change Economics,” in “The Economics
of Climate Change: Adaptations Past and Present,” University of Chicago Press, 2011, pp. 47–71.
, “Climate Change Policy: What Do the Models Tell Us?,” Journal of Economic Literature, 2013, 51
(3), 860–872.
Schelling, Thomas C, “Some economics of global warming,” The American Economic Review, 1992,
pp. 1–14.
Schlenker, Wolfram and Michael J Roberts, “Nonlinear effects of weather on corn yields,” Applied
Economic Perspectives and Policy, 2006, 28 (3), 391–398.
and
, “Nonlinear temperature effects indicate severe damages to U.S. crop yields under climate
change,” Proceedings of the National Academy of sciences, 2009, 106 (37), 15594–15598.
Seppänen, Olli, William J. Fisk, and QH Lei, “Effect of Temperature on Task Performance in Office
Environment,” Earnest Orlando Lawrence Berkeley National Laboratory working paper, 2006.
Starr-McCluer, M., “The Effects of Weather on Retail Sales,” Federal Reserve Board of Governors working paper, 2000.
Stern, Nicholas, The economics of climate change: the Stern review, Cambridge University Press, 2007.
46
Tol, Richard S. J., “The Economic Impact of Climate Change,” Perspektiven der Wirtschaftspolitik, 2010,
11, 13–37.
Zivin, Joshua Graff and Matthew Neidell, “Temperature and the allocation of time: Implications for
climate change,” Journal of Labor Economics, 2014, 32 (1), 1–26.
, Solomon Hsiang, and Matthew Neidell, “Temperature and Human Capital in the Short- and
Long-Run,” National Bureau of Economic Research working paper No. 21157, 2015.
47
A
Appendix (for online publication)
A.1
Weather Stations
We use weather data from 129 weather areas featuring a total of 10,128 individual
weather stations, with the number of weather stations per area ranging from 2 to 295.
The data for a weather area are created by collecting the earliest available data from a
currently active weather station in that area. The data series is then extended further
by using another weather station in the area.1 For example, the weather data series for
the Nashville, TN area is compiled from three individual weather stations over the time
period 1871–2014.2
Using data for weather areas, as opposed to individual weather stations, avoids the
problem of missing daily data without sacrificing a significant amount of temperature
information because the correlation of average temperature reported across stations in
a given area is very high. For example, table A1 shows the correlation between daily
average temperatures reported by individual stations in the Nashville area and in the
Las Vegas area. Individual stations are included in the table if they report at least
60 daily observations per season for each season in the sample 1957–2012. There are
54 stations in the Nashville area and 8 meet the inclusion criteria, and there are 83
stations in the Las Vegas area and 7 meet the inclusion criteria. The correlations in
daily temperature reported across stations are greater than 0.99.
Next, we calculate the correlation between individual stations in each area over the
20-year period with the greatest number of individual stations meeting the inclusion
criteria of 60 daily observations per season. Twenty-year periods beginning in 1959–
1962 have the greatest number of stations meeting the inclusion criteria for Nashville
1
2
http://threadex.rcc-acis.org/
http://threadex.rcc-acis.org/threadex/process_records
Appendix - 1
and the 20-year period beginning in 1969 has the greatest number of stations meeting
the inclusion criteria for Las Vegas. This increases the number of individual stations
to 21 for Nashville and to 16 for Las Vegas. For the 20-year period beginning in 1959,
the minimum correlation between any two stations in Nashville is 0.9882, and for the
20-year period beginning in 1969, the minimum correlation between any two stations in
Las Vegas is 0.9785.
Finally, in order to consider all stations in each area, we impute missing seasonal data
for individual stations that report any daily data in 1957–2012. This includes 53 of 54
weather stations in Nashville and 78 of 83 weather stations in Las Vegas. Specifically,
we consider the seasonal average for a station to be missing if the station does not report
at least 60 daily observations in that season. We replace missing seasonal data with the
mean of the seasonal average of all stations. The mean correlation between stations in
Nashville is 0.9959 and the mean correlation between stations in Las Vegas is 0.9463.
A.2
A.2.1
Additional details of the empirical analysis
Deseasonalization
We regress each raw temperature observation Tj,τ at station j and day τ using the following specification:
Tj,τ =
12
X
γm Ij,m + αj + ετ ,
m=1
where Ij,m is a dummy for month m at station j, αj is a station fixed effect, and τ is an
error term. Then the deseasonalized station observation is
T̃j,τ ≡ Tj,τ − (
12
X
γ̂m Ij,m + α̂j ).
m=1
Appendix - 2
Appendix - 3
1.0000
0.9971
0.9984
0.9977
0.9968
0.9977
0.9977
0.9978
1.0000
0.9973
0.9982
0.9937
0.9980
0.9980
0.9969
Station 2
1.0000
0.9983
1.0000
0.9976
0.9955
1.0000
0.9983
0.9983
0.9957
0.9983
0.9983
0.9957
0.9979
0.9976
0.9965
Panel B: VEF: Las Vegas, NV
Station 3
Station 4
Station 5
Panel A: OHX: Nashville, TN
1.0000
1.0000
0.9979
Station 6
1.0000
0.9979
Station 7
1.0000
Station 8
Station 1
1.0000
Station 2
0.9948
1.0000
Station 3
0.9948
1.0000
1.0000
Station 4
0.9935
0.9928
0.9928
1.0000
Station 5
0.9935
0.9928
0.9928
1.0000
1.0000
Station 6
0.9952
0.9944
0.9944
0.9979
0.9979
1.0000
Station 7
0.9956
0.9968
0.9968
0.9956
0.9956
0.9964
1.0000
Notes: Individual stations are included in the table if they have at least 60 daily observations per season for each season
in the sample (1957–2012).
Station 1
Station 2
Station 3
Station 4
Station 5
Station 6
Station 7
Station 8
Station 1
Table A1: Correlation of individual stations in an area
In the row labeled “Pre-1950 deseasonalization” in table 7, we estimate γm and αj using
weather data up to only 1950.
A.2.2
GSP weights
In panel regressions using constant GSP weights, state i’s weight is calculated as the
proportion of state i’s total GSP over the sample 1957–2012 relative to national GDP (the
total of all states’ GSP) over the sample 1957–2012. Specifically, let gi,1 , . . . , gi,T denote
state i’s GSP in year t = 1, . . . , T ; then the weight of state i in the main specification in
the panel regression (section 3) is
PT
gi,t
PT t=1
P51
.
t=1
i=1 gi,t
In this way, the weight of each state in
the regression is time invariant.
In the “Time-varying GSP” row of table 7, we use time-varying GSP weights instead of
constant GSP weights. In this panel regression, each state i in year t is weighted by the
proportion of state i’s GSP in year t relative to national GDP in year t. Specifically, the
weight of state i in year t is
A.3
gi,t
P51
.
i=1 gi,t
Industry group classification
Table A2 provides the classifications of the industry groups used in the industry analysis in section 4.2, with industry output data and classifications from the Bureau of
Economic Analysis. The column “Pre-1997 classification” uses the industry group categories of the Standard Industrial Classification (SIC). The column “Post-1997 classification” uses the industry group categories of the North American Industry Classification
System (NAICS).
Appendix - 4
Appendix - 5
Manufacturing
Government
Retail trade
Wholesale trade
Communications
Printing and publishing
Motion pictures
Construction
Transportation
Electric, gas, sanitary services
Mining
Agriculture, forestry, fishing
Manufacturing
Government
Retail
Wholesale
Communication/Information
Construction
Transportation
Utilities
Mining
Agriculture, forestry, fishing
Notes: Definitions from the Bureau of Economic Analysis.
Finance, insurance, real estate
Pre-1997 classification
(SIC)
Services
Finance, insurance, real estate
Services
Industry group
Post-1997 classification
(NAICS)
Professional, scientific, technical services
Management of companies and enterprises
Administrative, waste management services
Educational services
Health care and social assistance
Arts, entertainment, and recreation
Accommodation and food services
Other services, except government
Finance and insurance
Real estate and rental and leasing
Manufacturing
Government
Retail trade
Wholesale trade
Publishing industries, except Internet
Motion picture, sound recording industries
Broadcasting and telecommunications
Information and data processing services
Construction
Transportation and warehousing
Utilities
Mining
Agriculture, forestry, fishing, and hunting
Table A2: Industry classifications
Table A3: State ranking by average summer temperature
Rank
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
State
Florida
Louisiana
Texas
Mississippi
Oklahoma
Alabama
Georgia
South Carolina
Arkansas
Arizona
Kansas
North Carolina
Tennessee
Missouri
California
Kentucky
Delaware
Maryland
Virginia
Illinois
New Jersey
Indiana
Nebraska
New Mexico
Ohio
Avg. Summer Temp
80.78
80.18
79.87
78.44
78.21
77.67
77.64
77.47
77.20
77.06
74.70
74.30
74.21
73.65
73.07
72.92
72.89
72.32
71.84
71.58
70.87
70.64
69.80
69.63
69.45
Rank
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
State
Iowa
West Virginia
Nevada
South Dakota
Rhode Island
Utah
Connecticut
Pennsylvania
Massachusetts
New York
Wisconsin
Michigan
North Dakota
Minnesota
Colorado
New Hampshire
Oregon
Vermont
Washington
Montana
Maine
Idaho
Wyoming
Alaska
Avg. Summer Temp
69.13
68.88
68.61
68.02
67.92
67.85
67.61
67.03
66.59
64.70
64.64
64.54
64.44
64.32
63.67
62.95
62.77
62.25
62.07
61.72
61.66
61.62
61.25
47.97
Notes: Hawaii and the District of Columbia are not included. Summer is defined as July, August, and September. Average summer temperature is calculated over the sample 1957–2012.
Monthly temperature data are from NOAA.
A.4
Definitions of U.S. regions and Ranking of States
We follow the U.S. Census Bureau and identify four geographic regions:
1. North: Connecticut, Maine, Massachusetts, New Hampshire, New Jersey, New
York, Pennsylvania, Rhode Island, Vermont;
2. Midwest: Illinois, Indiana, Iowa, Kansas, Michigan, Minnesota, Missouri, Nebraska, North Dakota, Ohio, South Dakota, Wisconsin;
3. South: Alabama, Arkansas, Delaware, Florida, Georgia, Kentucky, Louisiana, Maryland, Mississippi, North Carolina, Oklahoma, South Carolina, Tennessee, Texas,
Virginia, Washington D.C., and West Virginia;
4. West: Alaska, Arizona, California, Colorado, Hawaii, Idaho, Montana, Nevada,
Appendix - 6
Table A4: P -values for Wald tests of growth vs. level effects
2 lags
5 lags
1 lag, no LDV
Winter
[0.547]
[0.148]
[0.373]
[0.039]
[0.835]
[0.663]
Spring
[0.148]
[0.013]
[0.043]
[0.011]
[0.189]
[0.030]
Summer
[0.004]
[0.000]
[0.008]
[0.002]
[0.004]
[0.000]
Fall
[0.101]
[0.008]
[0.410]
[0.124]
[0.022]
[0.004]
Notes: This table reports results of robustness checks of growth vs. level regression (8). Each row
reports the p-values,Pthe first clustered by year and the second clustered by state, for the Wald
test of whether β + βlag is significantly different from zero. The first two rows includes 2 and
5 lags of temperatures, respectively. The last row includes 1 lag of temperature and excludes the
lagged dependent variable (GSP growth). Temperatures are in degrees Fahrenheit. The sample
is 1957–2012. The regressions are weighted by constant GSP shares.
New Mexico, Oregon, Utah,Washington, and Wyoming.
Table A3 displays each state’s ranking by average summer temperature and the average
summer temperature used to determine this rank. This ranking is used to determine
the samples for the results presented in figure 4.
A.5
Additional Results and Robustness Checks
Robustness for growth vs. level test. Table A4 provides several robustness checks
for the test of growth vs. level effects reported in table 2. The table reports the p-value
P
of the Wald test for the hypothesis that βs + βlag,s = 0 for each season s, where the
sum is over all the lags of seasonal temperature. In the table, we include two lags (the
first row), five lags (the second row), and 1 lag while excluding the lagged dependent
variable, i.e., lagged GSP growth (the last row). The results in this table are broadly
consistent with those reported in table 2, especially for the summer.
Autocorrelations of states’ GDP growth rates. Table A5 reports the first order au-
Appendix - 7
Table A5: Autocorrelation of States’ GDP growth rates
State
Alabama
Alaska
Arizona
Arkansas
California
Colorado
Connecticut
Delaware
District of Columbia
Florida
Georgia
Hawaii
Idaho
Illinois
Indiana
Iowa
Kansas
Kentucky
Louisiana
Maine
Maryland
Massachusetts
Michigan
Minnesota
Mississippi
Missouri
AC(1)
0.492 ∗∗∗
0.270 ∗∗
0.623 ∗∗∗
0.444 ∗∗∗
0.658 ∗∗∗
0.668 ∗∗∗
0.612 ∗∗∗
0.181
0.413 ∗∗∗
0.755 ∗∗∗
0.614 ∗∗∗
0.663 ∗∗∗
0.415 ∗∗∗
0.367 ∗∗∗
0.139
0.141
0.405 ∗∗∗
0.334 ∗∗∗
0.474 ∗∗∗
0.439 ∗∗∗
0.604 ∗∗∗
0.671 ∗∗∗
0.197
0.312 ∗∗
0.399 ∗∗∗
0.345 ∗∗∗
(S.E.)
( 0.120 )
( 0.133 )
( 0.108 )
( 0.125 )
( 0.101 )
( 0.102 )
( 0.109 )
( 0.133 )
( 0.127 )
( 0.087 )
( 0.106 )
( 0.100 )
( 0.125 )
( 0.126 )
( 0.134 )
( 0.136 )
( 0.123 )
( 0.129 )
( 0.122 )
( 0.122 )
( 0.105 )
( 0.099 )
( 0.135 )
( 0.128 )
( 0.123 )
( 0.127 )
State
Montana
Nebraska
Nevada
New Hampshire
New Jersey
New Mexico
New York
North Carolina
North Dakota
Ohio
Oklahoma
Oregon
Pennsylvania
Rhode Island
South Carolina
South Dakota
Tennessee
Texas
Utah
Vermont
Virginia
Washington
West Virginia
Wisconsin
Wyoming
AC(1)
0.286 ∗∗
-0.019
0.673 ∗∗∗
0.420 ∗∗∗
0.632 ∗∗∗
0.523 ∗∗∗
0.498 ∗∗∗
0.398 ∗∗∗
0.017
0.319 ∗∗
0.543 ∗∗∗
0.208
0.425 ∗∗∗
0.523 ∗∗∗
0.520 ∗∗∗
-0.128
0.413 ∗∗∗
0.542 ∗∗∗
0.640 ∗∗∗
0.212
0.621 ∗∗∗
0.518 ∗∗∗
0.425 ∗∗∗
0.454 ∗∗∗
0.462 ∗∗∗
(S.E.)
( 0.130 )
( 0.135 )
( 0.099 )
( 0.124 )
( 0.104 )
( 0.117 )
( 0.117 )
( 0.124 )
( 0.139 )
( 0.130 )
( 0.116 )
( 0.131 )
( 0.121 )
( 0.114 )
( 0.113 )
( 0.136 )
( 0.121 )
( 0.114 )
( 0.106 )
( 0.133 )
( 0.105 )
( 0.115 )
( 0.125 )
( 0.120 )
( 0.123 )
Notes: This table reports the first order autocorrelations of nominal GDP growth rates for each
U.S. state. The numbers in parenthesis denote standard errors. ***, **, and * denote significance
at the 1%, 5%, and 10% levels.
tocorrelations of GDP growth rates for the entire cross-section of US states. Our results
document that an overwhelming majority of states display a positive and statistically
significant first order autocorrelation. We use this finding to motivate the inclusion of
the lagged dependent variable in our baseline empirical specification.
Correlations of seasonal temperatures. Table A6 reports the correlations of seasonal temperatures in the cross-section of US states. The median values of correlations
provided in the table are typically positive and very low. The only correlation for which
we cannot reject the null hypothesis of null correlation at the 90% confidence level is the
one between spring and summer average temperatures. The correlations between fall
and summer temperatures, which are the seasons that are the main focus of attention,
are very modest and indistinguishable from zero in the cross-section of states. This finding supports the claim that the results presented are not affected by multicollinearity.
Appendix - 8
Table A6: Correlations of seasonal temperatures
Median
90% CI
Win/Spr
0.208
[ -0.103 , 0.432 ]
Win/Sum
0.144
[ -0.114 , 0.327 ]
Win/Fall
0.044
[ -0.189 , 0.166 ]
Spr/Sum
0.293
[ 0.093 , 0.448 ]
Spr/Fall
0.185
[ -0.048 , 0.325 ]
Sum/Fall
0.133
[ -0.018 , 0.376 ]
Notes: This table reports the median seasonal temperature correlations along with the 90%
confidence intervals for the cross-section of U.S. states.
Standardized industry regressions. In figure A1 we report the results obtained by
standardizing the growth rate of GDP of each industry and the summer temperatures by
their respective average volatilities. The specification is identical to the one used in the
main text for this part of the analysis. The figure shows that a one standard deviation
increase in summer temperature results in a change in the growth of GDP between −0.2
and +0.2 standard deviations, depending on the specific industry under consideration.
Inclusion of additional seasons in the industry regressions. In this section we
consider four additional specifications for the industry analysis presented in the main
text. The first specification also includes the Fall (for which the coefficient is sometimes
significant in our total GSP regressions). The second specification includes the Fall in
addition to estimating a pooled coefficient for Spring and Winter (which are robustly
insignificant in our analysis). The third specification includes all the season with separate coefficients and it is focused on the post-1990 sample. The fourth specification
includes a separate coefficient for each season, estimated using the entire sample. The
results presented in figure A2 document that, despite some marginal loss of power, the
estimated Summer coefficients appear very much in line with those in our benchmark
specification.
We note that the results for Case 3 and Case 4 in figure A2 combine industry state-level
data over the two sub-samples which coincide with the adoption of NAICS codes (i.e.
pre- and post-1997). The results should therefore be interpreted with caution in light
with the following note reported on the BEA website:3
3
Available at https://www.bea.gov/regional/docs/product/. Additional details on the industry changes that took place when NAICS codes were introduced are available at https://www.naics.
com/history-naics-code/.
Appendix - 9
“Cautionary note:
There is a discontinuity in the GDP-by-state time series at 1997, where the
data change from SIC industry definitions to NAICS industry definitions. [...]
This data discontinuity may affect both the levels and the growth rates of
GDP by state. Users of GDP by state are strongly cautioned against appending the two data series in an attempt to construct a single time series for 1963
to 2017.”
A.6
A.6.1
Further Robustness Checks
Gridded temperature data set
In this section, we show that our results are robust to using temperature data that is
not deseasonalized. We re-estimate our results in tables 1-5 using gridded temperature
data from the NOAA nClimDiv data set.4 The results are reported in tables A7-A12.
Throughout for brevity, we only report two standard errors, clustered by year and clustered by state.
4
This data set excludes Hawaii and the District of Columbia.
Appendix - 10
Figure A1: Standardized regression coefficients for the industry analysis reported in tables 4
and 5 of the main text.
Appendix - 11
Case 1: 1997-2011, Fall, Summer and pooled Spring/Winter
0.1
0
12
11
10
9
8
7
6
5
4
3
2
-0.2
1
-0.1
Baseline
case
(as in Figure
1 inSummer
the paper)
Case 2:
1997-2011,
Fall and
0.1
0
10
10
11
11
12
12
10
10
11
11
12
12
10
11
12
99
88
77
66
55
44
33
22
-0.2
-0.2
11
-0.1
Case 3: post-1990, All Seasons
Case 3: 1997-2011, Spring and Summer
0.1
0.1
0
0
99
88
77
66
55
44
33
22
-0.2
-0.2
11
-0.1
-0.1
Case 4: whole sample, All Seasons
0.1
0
9
8
7
6
5
4
3
2
-0.2
1
-0.1
Figure A2: Inclusion of additional seasons in the industry regressions reported in figure 1 in
the main text.
Appendix - 12
Table A7: Robustness of table 1 (using gridded data)
Panel Analysis
Winter
0.001
Spring
−0.005
(0.042)
(0.023)
(0.057)
(0.028)
Summer
−0.167
Fall
0.100
(0.064)∗∗∗
(0.047)∗∗∗
(0.047)∗∗
(0.035)∗∗∗
Notes: See notes to table 1 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Table A8: Robustness of table 2 (using gridded data)
Contemporaneous temp.
One-year lagged temp.
Sum of coefficients
Winter
−0.004
Spring
−0.015
(0.044)
(0.024)
(0.052)
(0.028)
0.006
0.094
Wald test’s p-value
(0.066)∗∗∗
(0.048)∗∗∗
−0.151
∗
(0.046)
(0.022)
(0.051)
(0.032)∗∗∗
0.002
0.079
(0.074)
(0.029)
(0.978)
(0.944)
Summer
−0.181
(0.075)
(0.041)∗
(0.297)
(0.058)
Fall
0.104
(0.044)∗∗
(0.035)∗∗∗
0.042
∗∗
(0.069)
(0.050)∗∗∗
−0.332
(0.051)
(0.025)∗
0.146
∗∗∗
(0.096)
(0.080)∗∗∗
(0.001)
(0.000)
(0.065)∗∗
(0.049)∗∗∗
(0.030)
(0.004)
Notes: See notes to table 2 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Table A9: Robustness of table 3 (using gridded data)
Productivity
Employment
Winter
−0.033
Spring
−0.037
(0.055)
(0.034)
(0.062)
(0.028)
0.014
−0.073
(0.027)
(0.013)
∗
(0.042)
(0.042)∗
Summer
−0.145
(0.086)∗
(0.049)∗∗∗
Fall
0.120
(0.051)∗∗
(0.046)∗∗∗
0.022
−0.008
(0.057)
(0.032)
(0.036)
(0.015)
Notes: See notes to table 3 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 13
Table A10: Robustness of table 4A4 (using gridded data)
Pre-1997
Post-1997
Avg. GDP share (%)
Gross State Product
−0.212
−0.222
100
(0.091)∗∗
(0.064)∗∗∗
Services
0.020
(0.058)
(0.054)
Finance, insurance,real estate
−0.255
(0.233)
(0.233)
Manufacturing
Government
Retail
Communication/Information
Construction
Utilities
20.5
(0.340)
(0.140)∗∗
0.042
−0.045
−0.036
(0.065)
(0.051)
(0.142)
(0.085)
−0.084
−0.285
−0.176
(0.103)∗
(0.060)∗∗∗
−0.224
(0.091)∗∗
(0.109)∗∗
−0.186
−0.267
−0.262
−0.415
(0.940)∗∗∗
(0.445)∗∗∗
5.9
4.5
4.4
(0.400)
(0.154)∗∗∗
0.188
−2.634
6.6
(0.681)
(0.406)
(0.225)
(0.159)
−0.521
12.2
(0.171)
(0.151)∗
0.091
0.205
12.9
(0.174)
(0.093)∗∗∗
(0.109)
(0.167)
(0.508)
(0.413)
Agriculture, forestry, fishing
−0.356
(0.624)
(0.377)
(0.187)
(0.197)
Mining
25.7
(0.066)∗∗∗
(0.076)∗∗
(0.194)
(0.083)
(0.205)
(0.208)
Transportation
−0.184
−0.082
(0.060)
(0.063)
Wholesale
(0.183)
(0.060)∗∗∗
3.0
0.645
1.8
0.662
1.4
(0.386)∗
(0.223)∗∗∗
(1.295)
(0.303)∗∗
−2.022
(0.892)∗∗
(0.468)∗∗∗
1.1
Notes: See notes to table 4A in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 14
Table A11: Robustness of table 4B (using gridded data)
Post-1997 Ave GDP share (%)
Services
Professional and business services
−0.192
11.6
(0.122)
(0.103)∗
0.024
Educational services, health care,
social assistance
7.7
(0.050)
(0.060)
Other services, except government
−0.217
2.6
(0.132)∗
(0.106)∗∗
Food services and drinking places
−0.415
2.0
(0.144)∗∗∗
(0.127)∗∗∗
Arts, entertainment, and recreation
0.375
1.0
(0.242)
(0.202)∗
0.013
Accommodation
0.9
(0.215)
(0.326)
Finance, insurance, real estate
−0.441
Real estate
11.4
(0.358)
(0.112)∗∗∗
Federal Reserve banks, credit
intermediation, and related services
−0.160
Insurance, carriers and related activities
−0.995
3.6
(0.400)
(0.311)
2.6
(0.590)∗
(0.458)∗∗
Securities, commodity contracts, and investments
−0.218
1.3
(0.508)
(0.296)
Rental and leasing services, lessors
of intangible assets
Funds, trusts, and other financial vehicles
−0.012
1.3
(0.227)
(0.277)
0.702
0.2
(1.171)
(1.041)
Notes: See notes to table 4B in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 15
Table A12: Robustness of table 5 (using gridded data)
North
South
Winter
0.195
Spring
0.077
Summer
0.089
Fall
0.084
(0.130)
(0.120)
(0.196)
(0.159)
(0.191)
(0.114)
(0.170)
(0.167)
−0.108
0.162
−0.281
0.570
(0.163)
(0.084)
Midwest
West
(0.157)
(0.064)∗∗
∗
(0.152)
(0.114)∗∗
(0.184)∗∗∗
(0.077)∗∗∗
0.006
−0.117
−0.032
(0.068)
(0.041)
(0.108)
(0.085)
(0.116)
(0.073)
0.042
−0.145
−0.007
0.050
(0.156)
(0.195)
(0.152)
(0.191)
(0.088)
(0.067)
(0.118)
(0.060)∗∗
−0.100
(0.104)
(0.060)∗
Notes: See notes to table 5 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
A.6.2
Temperature aggregated by population data
In this section, we show that our results are robust to using a temperature data set
in which weather station data is aggregated to the state level using county population
instead of county area. We re-estimate our results in tables 1-5 using temperature data
aggregated by population. The results are reported in tables A13-A18.
Appendix - 16
Table A13: Robustness of table 1 (using temperature data aggregated by population)
Panel Analysis
Whole Year
0.017
Winter
0.012
Spring
−0.004
(0.109)
(0.057)
(0.048)
(0.023)
(0.066)
(0.028)
Summer
−0.129
(0.074)∗
(0.041)∗∗∗
Fall
0.094
(0.057)∗
(0.034)∗∗∗
Notes: See notes to table 1 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Table A14: Robustness of table 2 (using temperature data aggregated by population)
Contemporaneous temp.
One-year lagged temp.
Sum of coefficients
Wald test’s p-value
Winter
0.009
Spring
−0.023
(0.051)
(0.025)
(0.062)
(0.030)
0.011
0.132
Summer
−0.142
(0.077)∗
(0.040)∗∗∗
−0.145
∗∗
Fall
0.100
(0.052)∗
(0.033)∗∗∗
0.044
∗
(0.049)
(0.026)
(0.061)
(0.040)∗∗∗
(0.078)
(0.056)∗∗∗
(0.062)
(0.028)
0.020
0.110
-0.287
0.144
∗∗∗
(0.079)
(0.029)
(0.087)
(0.042)∗∗∗
(0.110)
(0.075)∗∗∗
(0.082)∗
(0.044)∗∗∗
[0.804]
[0.507]
[0.211]
[0.012]
[0.011]
[0.000]
[0.085]
[0.002]
Notes: See notes to table 2 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Table A15: Robustness of table 3 (using temperature data aggregated by population)
Productivity
Employment
Winter
−0.029
Spring
−0.045
(0.069)
(0.040)
(0.073)
(0.034)
0.015
−0.082
(0.032)
(0.015)
(0.050)∗
(0.047)∗
Summer
−0.113
Fall
0.148
(0.091)
(0.048)∗∗
(0.051)∗∗∗
(0.053)∗∗∗
0.008
-0.023
(0.058)
(0.035)
(0.041)
(0.016)
Notes: See notes to table 3 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 17
Table A16: Robustness of table 4A (using temperature data aggregated by population)
Pre-1997
Post-1997
Avg. GDP share (%)
Gross State Product
−0.164
−0.187
100
(0.104)
(0.054)∗∗∗
Services†
Finance, insurance,real estate
0.016
Government
Retail
(0.087)
(0.070)∗
−0.249
-0.418
Communication/Information†
0.333
(0.217)
(0.098)
(0.606)
(0.451)
−0.047
−0.038
(0.077)
(0.051)
(0.158)
(0.081)
−0.041
−0.181
Construction
−0.217
(0.105)
(0.062)
(0.175)
(0.154)
−0.182
−0.269
−0.159
(0.234)
(0.185)
Transportation
Utilities
Mining
Agriculture, forestry, fishing
−0.390
0.249
(0.195)
(0.119)∗∗
6.6
5.9
4.5
4.4
3.0
0.600
1.8
1.4
(0.366)∗
(0.206)∗∗∗
0.048
0.596
(0.543)
(0.522)
(1.400)
(0.442)
−2.500
−2.118
(0.941)∗∗∗
(0.390)∗∗∗
12.2
(0.424)
(0.190)∗∗
0.168
0.357
12.9
(0.673)
(0.355)
(0.131)
(0.191)
(0.229)
(0.174)∗∗
20.5
(0.185)
(0.090)∗∗
−0.090
(0.102)∗
(0.080)∗∗
25.7
(0.373)
(0.144)∗∗∗
−0.019
(0.071)
(0.056)
Wholesale
−0.122
(0.072)
(0.049)
(0.221)
(0.193)
Manufacturing
(0.205)
(0.072)∗∗∗
(0.981)∗∗
(0.478)∗∗∗
1.1
Notes: See notes to table 4A in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 18
Table A17: Robustness of table 4B (using temperature data aggregated by population)
Post-1997
Ave GDP share (%)
Services
Professional and business services
−0.083
11.6
(0.145)
(0.085)
Educational services, health care,
social assistance
−0.011
Other services, except government
−0.136
7.7
(0.045)
(0.064)
2.6
(0.122)
(0.092)
Food services and drinking places
−0.284
2.0
(0.146)∗
(0.162)∗
Arts, entertainment, and recreation
0.354
1.0
(0.268)
(0.168)∗∗
0.091
Accommodation
0.9
(0.246)
(0.330)
Finance, insurance, real estate
−0.407
Real estate
11.4
(0.387)
(0.113)∗∗∗
Federal Reserve banks, credit
intermediation, and related services
−0.184
Insurance, carriers and related activities
−1.419
3.6
(0.415)
(0.280)
2.6
(0.612)∗∗
(0.539)∗∗∗
−0.207
Securities, commodity contracts,
and investments
1.3
(0.512)
(0.297)
Rental and leasing services, lessors
of intangible assets
Funds, trusts, and other financial vehicles
0.008
1.3
(0.272)
(0.271)
1.243
0.2
(1.201)
(0.897)
Notes: See notes to table 4B in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 19
Table A18: Robustness of table 5 (using temperature data aggregated by population)
North
South
Midwest
West
Winter
0.345
Spring
0.071
Summer
0.219
(0.219)
(0.261)
(0.285)
(0.191)
(0.271)
(0.240)
−0.056
0.058
−0.263
(0.153)
(0.080)
(0.144)
(0.082)
(0.161)
(0.069)∗∗∗
Fall
−0.355
(0.266)
(0.144)∗∗
0.515
(0.194)∗∗∗
(0.089)∗∗∗
0.001
−0.174
0.093
−0.089
(0.092)
(0.071)
(0.156)
(0.118)
(0.185)
(0.090)
(0.139)
(0.082)
0.052
−0.145
0.102
−0.030
(0.138)
(0.144)
(0.151)
(0.147)
(0.094)
(0.050)
(0.143)
(0.067)∗∗
Notes: See notes to table 5 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 20
A.6.3
Including one-year lagged temperatures
In this section, we show that our results are robust to including the one-year lag of
temperature in all specifications. We re-estimate our results in tables 3-5 including the
one-year lag of seasonal temperature variables. The results are reported in tables A19A23.
Table A19: Robustness of table 3 (including one-year lag of temperature)
Productivity contemporaneous temp.
Productivity one-year lag
Sum of coefficients
Wald test’s p-value
Employment contemporaneous temp.
Winter
−0.050
Spring
−0.040
(0.066)
(0.035)
(0.064)
(0.040)
−0.019
−0.023
−0.083
0.062
(0.067)
(0.055)
(0.065)
(0.026)
(0.119)
(0.058)
(0.071)
(0.119)
−0.069
−0.063
−0.250
(0.114)
(0.067)
(0.099)
(0.046)
[0.553]
[0.309]
0.016
[0.531]
[0.176]
−0.083
(0.026)
(0.016)
Employment one-year lag temp.
Sum of coefficients
Wald test’s p-value
(0.040)∗∗
(0.050)∗
0.074
0.027
(0.035)∗∗
(0.019)∗∗∗
(0.042)
(0.024)
0.090
−0.056
Summer
−0.167
(0.090)∗
(0.055)∗∗∗
(0.141)∗
(0.068)∗∗∗
Fall
0.163
(0.054)∗∗∗
(0.076)∗∗
0.225
(0.112)∗∗
(0.182)
[0.092]
[0.001]
0.002
[0.058]
[0.223]
−0.018
(0.064)
(0.039)
(0.047)
(0.028)
−0.136
−0.007
(0.052)∗∗∗
(0.050)∗∗∗
−0.134
(0.048)
(0.022)
−0.025
(0.044)∗∗
(0.031)∗∗∗
(0.054)
(0.034)
(0.098)
(0.065)∗∗
(0.082)
(0.035)
[0.053]
[0.005]
[0.317]
[0.108]
[0.187]
[0.045]
[0.762]
[0.474]
Notes: See notes to table 3 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 21
Table A20: Robustness of table 4A column 1 (including one-year lag of temperature)
Contemp. One-year lag Sum of Wald test’s
temp.
temp.
coeff.
p-value
Gross state product
−0.186
−0.116
−0.301
(0.095)∗
(0.113)
(0.158)∗
[0.065]
∗∗∗
∗∗
(0.060)
(0.098)
(0.123)
[0.018]
†
Services
0.022
−0.187
−0.164
(0.067)
(0.061)∗∗∗
(0.092)∗
[0.084]
∗∗∗
∗∗
(0.049)
(0.049)
(0.079)
[0.042]
Finance, insurance, real estate −0.205
−0.207
−0.412
(0.243)
(0.177)
(0.331)
[0.222]
∗
(0.235)
(0.143)
(0.219)
[0.066]
Manufacturing
−0.052
−0.221
−0.274
(0.218)
(0.192)
(0.254)
[0.289]
(0.104)
(0.171)
(0.153)∗
[0.079]
Government
−0.068
0.019
−0.049
(0.071)
(0.062)
(0.095)
[0.608]
(0.063)
(0.070)
(0.107)
[0.647]
Retail
−0.050
−0.075
−0.125
(0.075)
(0.081)
(0.102)
[0.227]
(0.060)
(0.059)
(0.079)
[0.119]
Wholesale
−0.153
−0.161
−0.314
(0.104)
(0.107)
(0.166)∗
[0.068]
∗∗
∗
∗∗∗
(0.062)
(0.087)
(0.104)
[0.004]
†
Communication/Information
−0.238
0.073
−0.164
(0.088)∗∗∗
(0.101)
(0.148)
[0.275]
∗∗∗
(0.091)
(0.095)
(0.122)
[0.185]
Construction
−0.216
−0.471
−0.688
(0.238)
(0.256)∗
(0.383)∗
[0.082]
∗∗
∗
(0.184)
(0.230)
(0.360)
[0.062]
Transportation
0.151
−0.037
0.114
(0.124)
(0.149)
(0.213)
[0.595]
(0.195)
(0.097)
(0.254)
[0.654]
Utilities
0.337
0.028
0.364
(0.249)
(0.184)
(0.283)
[0.207]
(0.202)∗
(0.160)
(0.258)
[0.164]
Mining
−0.162
0.138
−0.024
(0.538)
(0.706)
(0.817)
[0.977]
(0.539)
(0.730)
(1.189)
[0.984]
Agriculture, forestry, fishing
−2.556
1.338
−1.218
(0.966)∗∗∗
(0.700)∗
(1.292)
[0.353]
∗∗∗
∗∗∗
∗∗∗
(0.444)
(0.316)
(0.369)
[0.002]
Notes: See notes to table 4A in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 22
Table A21: Robustness of table 4A column 2 (including one-year lag of temperature)
Contemp. One-year lag Sum of Wald test’s
temp.
temp.
coeff.
p-value
Gross state product
−0.269
−0.223
−0.492
(0.183)
(0.151)
(0.212)∗∗
[0.039]
∗∗∗
∗∗
∗∗∗
(0.069)
(0.090)
(0.121)
[0.000]
†
Services
−0.230
−0.242
−0.472
(0.068)∗∗∗
(0.132)∗
(0.162)∗∗∗
[0.013]
∗∗∗
∗
∗∗∗
(0.080)
(0.127)
(0.176)
[0.010]
Finance, insurance, real estate −0.454
−0.184
−0.638
(0.385)
(0.278)
(0.561)
[0.278]
∗∗∗
∗∗∗
(0.155)
(0.128)
(0.205)
[0.003]
Manufacturing
0.014
−0.576
−0.562
(0.622)
(0.515)
(0.821)
[0.507]
(0.410)
(0.416)
(0.492)
[0.259]
Government
−0.071
−0.267
−0.338
∗
(0.159)
(0.148)
(0.238)
[0.180]
(0.085)
(0.141)∗
(0.163)∗∗
[0.043]
Retail
−0.251
−0.093
−0.343
(0.200)
(0.202)
(0.333)
[0.323]
∗∗∗
∗∗∗
(0.083)
(0.070)
(0.103)
[0.002]
Wholesale
−0.294
−0.116
−0.410
(0.173)∗
(0.124)
(0.231)∗
[0.101]
∗
∗∗∗
(0.160)
(0.081)
(0.157)
[0.012]
†
Communication/Information
−0.298
−0.045
−0.343
(0.739)
(0.291)
(0.870)
[0.701]
(0.388)
(0.323)
(0.321)
[0.290]
Construction
−0.410
−0.390
−0.800
(0.442)
(0.338)
(0.612)
[0.215]
∗∗
∗∗
∗∗∗
(0.189)
(0.186)
(0.267)
[0.004]
Transportation
0.201
0.167
0.368
(0.216)
(0.256)
(0.367)
[0.337]
(0.140)
(0.145)
(0.254)
[0.153]
Utilities
0.605
−0.209
0.396
∗
(0.341)
(0.509)
(0.488)
[0.433]
(0.252)∗∗
(0.387)
(0.580)
[0.498]
Mining
0.844
1.673
2.516
(1.593)
(1.361)
(1.848)
[0.198]
∗∗
∗∗
∗∗∗
(0.356)
(0.718)
(0.645)
[0.000]
Agriculture, forestry, fishing
−2.091
1.817
−0.274
(0.999)∗∗
(1.106)∗
(1.764)
[0.879]
∗∗∗
∗∗
(0.463)
(0.758)
(0.852)
[0.749]
Notes: See notes to table 4A in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 23
Table A22: Robustness of table 4B (including one-year lag of temperature)
Contemp. One-year Sum of Wald test’s
temp.
lag temp. coeff.
p-value
Services
Professional and business services
−0.258
−0.391
−0.649
(0.103)∗∗
(0.223)∗
(0.234)∗∗∗
[0.017]
∗∗
∗
(0.113)
(0.225)
(0.307)∗∗
[0.039]
Educational services,
−0.007
−0.022
−0.028
health care, social assistance
(0.048)
(0.061)
(0.085)
[0.746]
(0.065)
(0.047)
(0.087)
[0.745]
Other services, except government
−0.274
−0.257
−0.531
(0.147)∗
(0.178)
(0.277)∗
[0.079]
∗∗∗
∗
(0.104)
(0.154)
(0.214)∗∗
[0.016]
Food services and drinking places
−0.392
−0.063
−0.455
(0.158)∗∗
(0.243)
(0.321)
[0.181]
(0.148)∗∗∗
(0.067)
(0.180)∗∗
[0.015]
Arts, entertainment, and recreation
0.371
−0.439
−0.068
(0.258)
(0.247)∗
(0.419)
[0.873]
(0.202)∗
(0.160)∗∗∗
(0.281)
[0.809]
Accommodation
0.015
−0.182
−0.167
(0.275)
(0.240)
(0.460)
[0.723]
(0.365)
(0.177)
(0.489)
[0.734]
Finance, insurance, real estate
Real estate
−0.460
−0.268
−0.728
(0.402)
(0.234)
(0.572)
[0.227]
(0.121)∗∗∗
(0.112)∗∗
(0.181)∗∗∗
[0.000]
Federal Reserve banks, credit
−0.271
−0.185
−0.456
intermediation, and related services
(0.455)
(0.246)
(0.443)
[0.324]
(0.364)
(0.370)
(0.611)
[0.459]
Insurance, carriers
−1.310
-0.124
−1.435
and related activities
(0.624)∗∗
(0.595)
(0.837)∗
[0.112]
∗∗
(0.539)
(0.513)
(0.684)∗∗
[0.041]
Securities, commodity contracts,
−0.310
−0.167
−0.477
and investments
(0.574)
(0.611)
(1.012)
[0.646]
(0.329)
(0.412)
(0.526)
[0.368]
Rental and leasing services,
−0.035
−0.081
−0.116
lessors of intangible assets
(0.250)
(0.419)
(0.558)
[0.839]
(0.292)
(0.334)
(0.487)
[0.813]
Funds, trusts, and
0.984
−0.442
0.542
other financial vehicles
(1.067)
(1.156)
(1.175)
[0.653]
(1.143)
(1.591)
(2.344)
[0.818]
Notes: See notes to table 4B in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 24
Table A23: Robustness of table 5 (including one-year lag of temperature)
North contemp.
North one-year lag
Sum of coefficients
Wald test’s p-value
South contemp.
South one-year lag
Sum of coefficients
Winter
0.276
Spring
0.063
Summer
0.115
(0.165)∗
(0.225)
(0.301)
(0.170)
(0.268)
(0.220)
0.263
0.300
−0.022
0.114
(0.205)
(0.162)
(0.196)
(0.197)
(0.306)
(0.224)
(0.225)
(0.242)
Midwest contemp.
0.539
0.363
0.093
−0.111
(0.339)
(0.244)
(0.333)
(0.438)
(0.257)
(0.319)
[0.045]
[0.147]
−0.102
[0.290]
[0.175]
0.049
[0.781]
[0.837]
−0.325
[0.668]
[0.738]
0.549
(0.167)
(0.092)
(0.138)
(0.051)
0.009
0.152
(0.120)
(0.082)
(0.182)
(0.086)∗
−0.092
[0.638]
[0.197]
−0.017
(0.084)
(0.054)
Midwest one-year lag
−0.126
(0.090)
(0.060) ∗∗
Sum of coefficients
Wald test’s p-value
West contemp.
West one-year lag
Sum of coefficients
Wald test’s p-value
(0.236)
(0.163)
(0.262)∗∗
(0.335)
(0.195)
(0.069)
Wald test’s p-value
Fall
−0.224
0.201
(0.211)
(0.082)∗∗
[0.343]
[0.026]
−0.223
(0.140)
(0.098)∗∗
0.243
(0.142) ∗
(0.105) ∗∗
(0.173)∗
(0.087)∗∗∗
−0.390
(0.207)∗
(0.141)∗∗∗
−0.715
(0.244)∗∗∗
(0.182)∗∗∗
[0.005]
[0.001]
−0.003
(0.159)
(0.069)
−0.154
(0.161)
(0.131)
(0.179)∗∗∗
(0.054)∗∗∗
0.326
(0.187)∗
(0.111)∗∗∗
0.876
(0.285)∗∗∗
(0.147)∗∗∗
[0.003]
[0.000]
−0.162
(0.115)
(0.078) ∗∗
0.176
(0.115)
(0.076) ∗∗
−0.143
0.020
−0.157
0.015
(0.119)
(0.091)
(0.20)3
(0.113)
(0.255)
(0.116)
(0.157)
(0.083)
[0.235]
[0.143]
0.018
[0.921]
[0.861]
−0.121
[0.540]
[0.204]
0.068
[0.927]
[0.864]
−0.014
(0.096)
(0.069)
(0.149)
(0.069)∗
(0.147)
(0.149)
(0.168)
(0.165)
0.057
0.047
−0.152
-0.069
(0.124)
(0.072)
(0.162)
(0.169)
(0.149)
(0.116)
(0.150)
(0.082)
0.075
−0.075
−0.083
−0.083
(0.165)
(0.094)
(0.253)
(0.213)
(0.221)
(0.212)
(0.249)
(0.115)
[0.650]
[0.440]
[0.769]
[0.732]
[0.708]
[0.702]
[0.739]
[0.482]
Notes: See notes to table 5 in the main text. Standard errors, first clustered by year and second
clustered by state, are in parentheses.
Appendix - 25
A.6.4
Arellano-Bond estimator
In this section, we show that our results in tables 1-5 are robust to using GMM estimators developed by (Arellano and Bond, 1991) that produce consistent estimates of a
dynamic panel for finite T . We use first differences with respect to time. Because T
is fairly large, using all possible instruments could lead to a bias of “too many instruments” (Newey and Windmeijer (2009)), so we restrict the number of intruments and
use one step GMM estimators with a naive weighting matrix. These estimators remain
consistent when T (the number of time periods) and N (the number of states) and the
number of instruments is large (Alvarez and Arellano (2003)). We use lags 2-10 as instruments, use small sample adjustments, and estimate robust standard errors. The
results are reported in tables A24-A29.
Table A24: Robustness of table 1 (using Arellano-Bond)
Panel Analysis
Whole Year
−0.026
Winter
−0.002
Spring
−0.049
(0.082)
(0.032)
(0.034)
Summer
−0.134
Fall
0.115
(0.042)∗∗∗
(0.056)∗∗
Notes: See notes to table 1 in the main text. Robust standard errors are in parentheses.
Table A25: Robustness of table 2 (using Arellano-Bond)
Contemporaneous temp.
Winter
−0.020
(0.034)
One-year lagged temp.
Sum of coefficients
Wald test’s p-value
Spring
−0.059
Summer
−0.191
(0.030)∗
(0.043)∗∗∗
0.012
0.083
(0.027)
(0.048)∗
Fall
0.143
(0.054)∗∗∗
−0.137
0.126
(0.055)∗∗
(0.040)∗∗∗
−0.008
0.024
(0.043)
(0.052)
−0.328
(0.075)∗∗∗
((0.083)∗∗∗
[0.860]
[0.651]
[0.000]
[0.002]
0.270
Notes: See notes to table 2 in the main text. Robust standard errors are in parentheses.
Appendix - 26
Table A26: Robustness of table 3 (using Arellano-Bond)
Productivity
Employment
Winter
−0.011
Spring
0.002
(0.050)
(0.052)
−0.004
−0.068
(0.016)
Summer
−0.152
Fall
0.142
(0.065)∗∗
(0.032)∗∗
(0.059)∗∗
0.048
−0.012
(0.036)
(0.021)
Notes: See notes to table 3 in the main text. Robust standard errors are in parentheses.
Table A27: Robustness of table 4A (using Arellano-Bond)
Pre-1997
Post-1997
Avg. GDP share (%)
Gross state product
−0.164
−0.236
100
(0.061)∗∗∗
Services†
Finance, insurance, real estate
0.110
−0.165
25.7
−0.137
−0.505
20.5
−0.052
0.344
12.9
(0.148)
(0.512)
−0.101
−0.011
(0.046)∗∗
(0.227)
Manufacturing
Government
Retail
Wholesale
Communication/Information†
Construction
Transportation
Utilities
Mining
Agriculture, forestry, fishing
(0.067)∗∗∗
(0.057)∗
(0.076)∗∗
(0.179)∗∗∗
−0.011
−0.122
(0.075)
(0.108)
−0.136
−0.267
(0.081)∗
−0.066
(0.135)*
(0.423)
−0.038
−0.218
(0.148)
(0.226)
0.243
0.131
(0.287)
(0.114)
0.454
5.9
0.129
4.5
4.4
3.0
1.8
(0.340)
−0.110
0.272
(0.526)
(0.324)
−3.054
−2.456
(0.486)∗∗∗
6.6
(0.165)
−0.223
(0.202)∗∗
12.2
(0.104)
(0.690)∗∗∗
1.4
1.1
Notes: See notes to table 4A in the main text. Robust standard errors are in parentheses.
Appendix - 27
Table A28: Robustness of table 4B (using Arellano-Bond)
Post-1997
Ave GDP share (%)
Services
−0.102
Professional and business services
11.6
(0.095)
−0.044
Educational services, health care,
social assistance
Other services, except government
7.7
(0.067)
−0.238
2.6
(0.108)∗∗
−0.339
Food services and drinking places
2.0
(0.152)∗∗
0.513
Arts, entertainment, and recreation
1.0
(0.194)∗∗∗
0.219
Accommodation
0.9
(0.273)
Finance, insurance, real estate
−0.640
Real estate
11.4
(0.134)∗∗∗
Federal Reserve banks, credit
intermediation, and related services
Insurance, carriers and related activities
−0.181
3.6
(0.304)
−1.637
2.6
(0.673)∗∗
Securities, commodity contracts,
and investments
Rental and leasing services, lessors
of intangible assets
Funds, trusts, and other financial vehicles
0.309
1.3
(0.376)
0.149
1.3
(0.320)
1.176
0.2
(0.847)
Notes: See notes to table 4B in the main text. Robust standard errors are in parentheses.
Table A29: Robustness of table 5 (using Arellano-Bond)
North
South
Midwest
West
Winter
0.322
Spring
0.079
Summer
0.205
Fall
−0.236
(0.248)
(0.194)
(0.225)
−0.115
0.148
−0.257
(0.096)
(0.136)
−0.010
−0.156
0.028
−0.108
(0.059)
(0.103)
(0.076)
(0.084)
−0.029
−0.087
0.053
0.002
(0.069)
(0.070)
(0.178)
(0.174)
(0.168)
0.672
(0.096)∗∗∗
(0.069)∗∗∗
Notes: See notes to table 5 in the main text. Robust standard errors are in parentheses.
Appendix - 28
@FedFRASER