Weather Fluctuations, Expectation Formation, and Short-Run Behavioral Responses To Climate Change - Ji, 2021

Environmental and Resource Economics (2021) 78:77–119
https://doi.org/10.1007/s10640-020-00525-x
Weather Fluctuations, Expectation Formation, and Short‑Run

Behavioral Responses to Climate Change
Xinde Ji1 · Kelly M. Cobourn2
Accepted: 5 November 2020 / Published online: 23 November 2020

© Springer Nature B.V. 2020
Abstract
One premise adopted in most previous studies is that weather fluctuations affect economic
outcomes contemporaneously. Yet under certain circumstances, the impact of weather fluc-
tuations in the current year can be carried over into the future. Using agricultural produc-
tion as an example, we empirically investigate how past weather fluctuations affect eco-
nomic decision-making by shifting agents’ subjective expectations over future climate. We
find that agricultural producers do not form expectations on future climate using only long-
run normals, and instead engage in a combination of heuristics, including the availability
heuristic and the reinforcement strategy. Adopting these learning mechanisms causes farm-
ers to significantly over-react to more recent fluctuations in weather and water availability
when making ex ante acreage and crop allocation decisions.
Keywords Weather fluctuation · Climate change adaptation · Expectation formation ·

Agricultural production
JEL Classification Q54 · Q15 · Q25
There has been widespread interest in the economic literature on understanding the impacts
of climate change on economic outcomes and corresponding adaptation measures. A popu-
lar approach to do so is to use random weather fluctuations to identify the effects of climate
change on economic outcomes, such as agricultural production (Deschênes and Greenstone
2007; Schlenker and Roberts 2009), gross domestic product (Dell et al. 2012), and long-
run adaptation (Deschênes and Greenstone 2011). One premise adopted in most studies
is that weather fluctuations affect economic outcomes contemporaneously. For example,
extreme heat events decrease crop yields (e.g. Schlenker and Roberts 2009; Lobell et al.
2013), but this negative effect applies only to the particular year in which the extreme heat
event occurs. However, under certain circumstances, the impact of weather fluctuations
* Xinde Ji
xji@brandeis.edu
Kelly M. Cobourn
kellyc13@vt.edu
1
Brandeis University, 415 South St MS 021, Waltham, MA 02453, USA
2
Virginia Tech, 310 West Campus Dr, Blacksburg, VA 24061, USA
13
Vol.:(0123456789)
78 X. Ji, K. M. Cobourn
in the current year can be carried over into the future. Studies have shown that extreme
weather realizations can affect long-term individual economic well-being (Ebenstein et al.
2016; Deryugina et al. 2018), future macroeconomic output (Dell et al. 2012), and induce
and intensify political instability, human conflicts, and wars in the future (Zhang et al.
2008; Hsiang et al. 2013; Zheng et al. 2014; Iyigun et al. 2017).
One particular pathway through which past weather fluctuations can affect future eco-
nomic outcomes is the process of expectation formation. Not only do weather fluctuations
affect contemporaneous outcomes, they also shape what people expect in terms of future
climate. Both economists and psychologists have established that recent realizations of an
event can disproportionately affect human perceptions on the underlying probability of that
event happening (e.g. Tversky and Kahneman 1974; Kahneman et al. 1991; Camerer and
Loewenstein 2011). Studies have found that subjective expectations to climate are incon-
sistent with predictions from rational expectations (Cameron 2005), such that people sig-
nificantly over-adjust their expectations to climate in response to recent, local, and extreme
weather events (Marx et al. 2007; Deryugina 2013; Konisky et al. 2016; Lee et al. 2018).
Through expectation formation, economic decisions that require forward-looking inference
on climate may be subject to a disproportionately large influence from past weather fluc-
tuations. In that sense, weather fluctuations can cause short-run economic losses due to
sub-optimal decision-making based on over-adjusted climate expectations.
Using agricultural production as an example, we empirically investigate how eco-
nomic agents adjust and adapt present decision-making in response to past shocks, and
from which we shed light on how economic agents learn from and act upon those shocks.
Agricultural production is uniquely affected by weather fluctuations and the correspond-
ing expectation formation process. Many agricultural production decisions are made
before the realization of weather. Examples include decisions on acreage, crop allocation,
planting time, seed variety, and tillage method. These ex ante decisions are made based
on subjective expectations over climate rather than weather realizations during the grow-
ing season. Previous studies have found that climate change induces long-run adaptation
in acreage (Cui 2020) and crop-allocation decisions (Seo and Mendelsohn 2008; Ji et al.
2018; Hagerty 2020). This literature usually focuses on farmers’ response to long-run cli-
mate normals, which ignores the fact that recent realizations can have a disproportionately
large role in driving subjective expectations over climate. Indeed, other studies have found
that agricultural production decisions can also be affected by various types of short-run
expectation shifters, including monsoon patterns (Miller 2016), spring snowpack forecasts
(Manning et al. 2017), and previous realizations of temperature (Miao et al. 2015; Cohn
et al. 2016), precipitation (Ding et al. 2009; Kala 2017), and water availability (Buck et al.
2014; Hagerty 2020). However, there is still scant literature that comments on the specific
mechanisms by which previous realizations enter the expectation formation process, how
they may affect multiple ex ante decisions, and how expectation-induced adaptation can be
affected by multiple factor inputs simultaneously.
Our empirical analysis focuses on two ex ante agricultural production decisions: (1)
acreage decisions, i.e., how much land to plant, and (2) crop-allocation decisions, i.e.,
what type(s) of crop to grow on land in production. We estimate fixed-effect models on
acreage, crop allocation, and their aggregate effect using farm-level data for an irrigated,
multi-crop agricultural system. We use multiple lags of growing degree days (GDD),
extreme degree days (EDD), precipitation, and surface water availability (as governed
by prior appropriation water rights) as independent variables to identify the effects of
past weather realizations on agricultural decision making. A unique feature of our data-
set is that irrigation water is distributed by collective water providers, whose water is
13
Weather Fluctuations, Expectation Formation, and Short-Run… 79
acquired under prior appropriation water rights. Under appropriative water rights, cross-
sectional variation in surface water availability comes from differences in water provid-
ers’ water rights attributes. In contrast, time-series variation in surface water availabil-
ity is governed by the stochastic natural inflows that determine the total water supply
available to be allocated across all water users in the region. Our identification strategy
relies on time-series variation for individual farmers by differencing out cross-sectional
variation using fixed effects. This approach allows us to control for average climate and
water right attributes, while simultaneously controlling for other unobserved individual-
specific heterogeneity, such as risk attitudes that may drive insurance adoption. This
allows us to capture the effect of shifting expectations due to the fluctuations in weather
and surface water availability for individual farmers and their subsequent behavioral
responses.
Using this identification strategy, we test for three types of expectation formation mech-
anisms (heuristics). Our starting hypothesis is that farmers form expectations using long-
run climate normals, a strategy consistent with the rational Bayesian learning framework. If
a farmer follows the Bayesian updating rule and uses long-run normals, then the effect of a
recent shock on his/her posterior expectation, and thus ex ante production decisions in sub-
sequent years, will be small, time-independent in magnitude, and long-lasting in duration.
We test this long-run-normal strategy against two alternative heuristics identified in the
behavioral economics literature. These alternatives are: (1) the availability heuristic, which
predicts that the effects on expectations and decisions are larger for more recent shocks,
and smaller, or even trending towards zero, for shocks farther away; and (2) the reinforce-
ment strategy, which predicts the effects on expectations and decisions to be cyclical.
Our empirical application is to Water District 1 (WD1) in the Upper Snake River Basin
of Idaho, for which we assemble a farm-level, spatially explicit panel dataset spanning
2007–2016. We identify the geographical boundary for each farm, and spatially match the
farm with a series of high-resolution geospatial datasets documenting field-level cropland
acreage, crop allocation, temperature, and precipitation. We also acquire data on water
rights characteristics, water curtailments by day of the growing season, and daily water
deliveries for collective water providers in WD1, and match them to the farms to which
they supply water.
Our results indicate that previous shocks in weather and surface water availability sig-
nificantly affect both acreage and crop-allocation decisions. On average, an increase in
GDD in past years increases cropland acreage, an increase in EDD in past years reduces
cropland acreage, and past precipitation shocks do not have a significant effect on acreage.
We do not find significant effects of weather shocks on crop allocation decisions. With
respect to surface water availability, we find evidence of cyclical responses to water avail-
ability shocks for crop-allocation decisions: an increase in curtailment days, i.e., a negative
shock in water availability, decreases expected profitability on planted lands in the first
and third years, and increases it in the second and fourth years following the shock. Past
increases in curtailment days also decrease cropland acreage, though the cyclical pattern is
less prevalent than in the case of crop allocation decisions.
These results provide evidence to reject the hypothesis that farmers only use long-run
normals to form expectations on climate and surface water availability. Instead, our results
are consistent, to varying degrees, with the alternative expectation–formation heuristics
considered. We find evidence to support the recency heuristic, as our results suggest that
the effect of endowment shocks dissipates in 4–5 years. We find limited evidence to sup-
port the availability heuristic, as there is no clear pattern of gradually decreasing effects
over time. Finally, we present evidence consistent with the reinforcement strategy, as
13
responses to previous weather shocks exhibit cyclical behavior in the model on crop alloca-
tion, a key characteristic consistent with reinforcement learning.
This study makes several contributions to the economic literature. First, we show that ex
ante economic decision-making may be disproportionately affected by past weather fluc-
tuations through cognitive heuristics other than Bayesian updating. This complements the
existing literature assessing the economic impact of climate change by demonstrating a
mechanism through which past weather fluctuations may affect future economic decisions
and outcomes. Second, our study emphasizes the need to incorporate short-run behavio-
ral responses when assessing climate change impacts. We show that weather fluctuations
create the potential for an over-adjustment cost as farmers overreact to short-run weather
signals. The existing approaches to modeling climate change impacts, such as Ricardian
regression, contemporaneous fixed-effect, or long-differences, are not able to identify this
short-run response. Our results are particularly relevant given that climate change mod-
els project a change in long-run averages as well as increased variability, particularly in
temperature. As weather shocks become more frequent and larger in magnitude, this over-
adjustment cost may also increase. Third, this study links two important threads in the eco-
nomic literature, one from the behavioral economics literature on learning mechanisms, the
other on agricultural adaptation to climate change. Previous studies in these areas either
investigate subjective perceptions over climate change without linking them to behavioral
responses or demonstrate adaptive responses to past weather fluctuations without tracing
them back to their underlying learning mechanisms.
1 Expectation Formation
In this section, we use a stylized framework to motivate learning mechanisms by which

expectations are formed and updated using previous weather and water availability realiza-
tions. We start by motivating the case for using the long-run normal as the expectation for
weather and water availability, which serves as our baseline learning mechanism. We then
discuss how producers may deviate from the long-run normal and use alternative mecha-
nisms to learn from the past.
Imagine a representative farmer who holds a single natural endowment governed by a
random variable W, which is independently and identically distributed (i.i.d) with mean
𝜇 and variance 𝜎 2 . The expectation formation process can be generalized as a process of
assigning weights to previous realizations, i.e.,
T
∑
Et (W) = 𝛽s wt−s (1)
s=1
where the expectation of the endowment at time t, Et (W) , is learned from T previous reali-
zations, wt−T to wt−1 , with a set of weights 𝛽T , … , 𝛽1 . We present three different types of
learning mechanisms in this section: long-run normals, the availability heuristic, and the
reinforcement strategy. All thrre of them are essentially different sets of weights for previ-
ous realizations.
We begin with inference based on long-run normals, i.e., an expectation based on the
average of a time series of historical realizations. Long-run normals are widely produced
and used by meteorologists and hydrologists as a baseline estimate for long-run climate
13
and water flow conditions.1 Normals have also been frequently used in the economics liter-
ature to approximate farmers’ expectations over climate (Mendelsohn et al. 1994; Schlen-
ker et al. 2005; Burke and Emerick 2016).
The use of long-run normals is appealing because it is consistent with the weights gen-
erated from several rational learning frameworks. Not only does it correspond to an unbi-
ased maximum likelihood estimate (MLE) of the underlying climate, it also corresponds
to a rational Bayesian learning framework with a non-informative prior.2 The Bayesian
learning framework builds on the Bayesian updating rule, a standard tool used by econo-
mists to model rational probability updating behaviors (Camerer and Loewenstein 2011).
While empirical evidence on the validity of the Bayesian updating rule is mixed (e.g. Bar-
beris et al. 1998; Charness and Levin 2005; Rabin 2002; Anwar and Loughran 2011; Chi-
ang et al. 2011), the Bayesian updating framework serves as a common starting point for
numerous studies on rational learning and decision-making under risk and uncertainty.
With its links to the Bayesian learning and MLE framework, we argue that if long-run
normals provide a rational expectation for the underlying long-run climate (as suggested
by their use in climate science), they are thus the best prediction for the realizations in
the upcoming season. Following that logic, economic decisions made using expectations
formed on Bayesian learning/long-run normals will be optimal. On the other hand, if
recent realizations are disproportionately weighted, subjective expectations over climate
will deviate from the long-run normal. In that case, we expect that the use of alternative
learning mechanisms is likely to generate economic losses by driving sub-optimal decision
making.
Two key hypotheses can be drawn from the assumption that the farmer uses long-run
normals to define expectations over weather. First, given a lengthy time series of previ-
ous realizations, the weather realization for any single year has a relatively small effect
on the expectation. This is because each single realization is weighted by T1 , which results
in a small influence on the posterior expectation if T is sufficiently large. Second, no mat-
ter how distant (lagged) the realization, each enters the expectation with the same weight.
Thus the effect of each realization on the expectation is independent of its timing.
The literature on cognitive psychology and behavioral economics has pointed out mul-
tiple alternative heuristics, i.e., learning mechanisms, that deviate from long-run normals.
One such heuristic is the availability heuristic, proposed by Tversky and Kahneman (1974).
The availability heuristic describes the phenomenon in which individuals place higher
probabilities on events that are foremost in their memory. Often known as the recency
bias, the availability heuristic has been documented in multiple empirical settings involv-
ing learning and expectation formation, and especially on events with small probabilities,
for example, flood risks (Gallagher 2014), climate change (Deryugina 2013), and financial
markets (Malmendier and Nagel 2011). In the case of agricultural production, studies have
also suggested that farmers often turn to familiar patterns from the past two agricultural
seasons, or to those happening at the time of other memorable events (Marx et al. 2007).
1
According to the World Meteorological Organization (WMO), normals are defined as “period averages
computed for a uniform and relatively long period comprising at least three consecutive 10-year periods.”
Other organizations and agencies use the same standard, for example, USGS, National Weather Service,
and USDA-NRCS, to produce their weather and surface water flow normals.
2
The assumption of a non-informative prior is not needed for our empirical analysis because any individ-
ual-specific prior will be subsumed by the inclusion of individual level fixed-effects.
13
In our application, we characterize the availability as a set of unequal weights between dif-
ferent past realizations. This implies that:
T
∑
Et (W) = 𝛽s wt−s = 𝛽1 wt−1 + 𝛽2 wt−2 + ⋯ + 𝛽T wt−T (2)
s=1
where unlike in the case of long-run normals, the weights 𝛽s for each realization wt−s under
the availability heuristic are unequal, and dependent on how the salience of previous reali-
zations. The qualitative pattern of these weights can take several forms. One possibility
that corresponds to the availability heuristic is that producers give higher weights to reali-
zations that are more recent, i.e.,
𝛽1 ≥ 𝛽2 ≥ ⋯ ≥ 𝛽T (3)
, because more recent realizations are more salient in one’s memory. Another possibility
is that producers draw inference only from experiences in the recent past and ignores the
more recent past, i.e.,
𝛽s > 0 when s ≤ S
(4)
𝛽s = 0 when S < s < T
such that any realization more distant than S years receives zero weight. In both cases, pro-
ducers put larger weight on the more recent shocks and over-react to weather fluctuations,
relative to the reactions using long-run normals.
A second learning mechanism we consider here is the reinforcement strategy, sometimes
referred to as the “win-stay, lose-shift” heuristic. Under the reinforcement strategy, an indi-
vidual expects a shock they have personally experienced to recur in the future (Roth and Erev
1995; Camerer and Ho 1999; Chiang et al. 2011). When a previous realization deviates from
the individual’s expectation, he/she will expect that positive or negative shock to be carried
over to future realizations. Although Bayesian learning exhibits some positive feedback, rein-
forcement learning amplifies those feedback mechanisms to the degree that is inconsistent
with behavior predicted by rational expectation.
We formally express the reinforcement strategy as a case in which the farmer adjusts his/
her expectation by reinforcing the most recent shock, i.e., the difference between the most
recent realization, wt−1 , and the expectation prior to that realization, Et−1 (W):
T
∑ [ ]
Et (W) = 𝛽s wt−s + 𝛾 wt−1 − Et−1 (W) (5)
s=1
where 𝛾 > 0 measures the magnitude of the reinforcement behavior. Simplifying Eq. 5 by
expanding the previous expectation term Et−1 (W) yields:
T
∑
Et (W) = 𝛽̃s wt−s (6)
s=1
where 𝛽̃s takes the following form:
𝛽̃1 =𝛽1 + 𝛾
(7)
𝛽̃s =𝛽s − 𝛾 𝛽̃s−1 for any s ≥ 2
13
Equation 7 suggests that the reinforcement strategy generates cyclical weights. The key
reason for cyclical weights is that the farmer adjusts his/her expectations based on the
shock wt−1 − Et−1 (W) . As such, the farmer relies too heavily on the realization in the year
prior t-1, wt−1 , but puts too little weight on the shock at t-2, wt−2 , which was the key com-
ponent in forming the year t-1 expectation, Et−1 (W) . Expanding the iterated expectations
leads us to conclude that the farmer will also over-utilize wt−3 , under-utilize wt−4 , etc. If 𝛾
is sufficiently large, it is even possible that the weights for even years, i.e., 𝛽̃2 , 𝛽̃4 , etc., can
become negative.
The mechanism for reinforcement learning can also be understood by linking it with ex
ante production decisions through a win-stay-lose-shift strategy. Say a farmer experiences
a positive shock in precipitation in year t-1, relative to what he/she expected. He/she could
have planted more acres if he/she knew precipitation would exceed the expectation. Using
the reinforcement strategy, he/she will over-adjust by planting more acres in year t, in the
expectation that this year’s precipitation is going to be more like last year’s. This will lead
to a larger-than-optimal planted acreage because, on average, precipitation will not be as
good as it was last year. Following the realization in year t, which is a negative precipita-
tion shock relative to year t-1, the farmer over-adjusts to a smaller-than-optimal acreage in
year t+1. This pattern implies that a positive shock in year t-1 can lead to a negative adjust-
ment two years later, in year t+1.
We formally write the hypotheses for the four learning mechanisms as follows:
Hypothesis 1 (Long-run Normal) Given a lengthy time series of previous realizations,

the realization from any single year has a relatively small impact on the expectation.
Hypothesis 2 (Long-run Normal) The timing of a realization is irrelevant to expectation
formation.
Hypothesis 3 (Availability Heuristic) Each realization will have different weights in the
expectation, with more recent realizations having larger weights and more distant ones
having smaller weights.
Hypothesis 4 (Availability Heuristic) More recent realizations will have positive
weights, whereas more distant realizations will have zero weight.
Hypothesis 5 (Reinforcement Strategy) Past realizations enter into expectations with
cyclical weights.
Figure 1 shows a stylized illustration of learning mechanisms mentioned above, including

long-run normals, different representations of the availability heuristic, and the reinforce-
ment strategy with various magnitudes.
For the purpose of illustrating relevant learning mechanisms, we have treated the under-
lying long-run distributions for climate and surface water flows as stationary over time.
It is worth noting that these long-run distributions are potentially trending as a result of
global or regional scale environmental change. However, we do not explicitly consider cli-
mate trends in this section for two reasons. First, long-run climate trends in temperature
and precipitation are much smaller in magnitude than year-to-year fluctuations. Long-run
climate trends imply a change of about 0.02◦ Cannually (Intergovernmental Panel on Cli-
mate Change 2014), whereas annual temperature realizations have a standard deviation of
around 0.7 ◦ C per year in our sample. Over a relatively short period of time (10 years for
our empirical analysis), climate trends are not likely to cause large shifts in the long-run
distribution of weather realizations. Second, our empirical analysis is sufficiently flexible
to control for trending means of climate or water availability in the data. Trends in mean
13
Fig. 1 Heuristics and weights on past shocks
temperature, precipitation, and surface water availability will be absorbed in the year fixed-
effects, such that the weather shocks included in our empirical models are on average sta-
tionary over time.
Our analysis does not preclude the possibility that producers interpret weather shocks
as signals of a trending climate. In fact, both learning mechanisms we listed above are
consistent with some type of over-interpreting weather fluctuations as climate trends. Rein-
forcement learning is a direct result of interpreting perceived trends in weather fluctuations
as actual climate trends. More interestingly, the availability heuristic is consistent with the
Bayesian learning framework with a time-variant climate governed by a random walk pro-
cess. Kala (2017) showed that under such a case, Bayesian learning leads to over-weighing
recent shocks relative to more distant ones, and the over-weight is proportional to the vari-
ability of the climate trend. However, because the actual climate trend is much smaller in
magnitude than year-to-year weather fluctuations, it is not likely for farmers using a Bayes-
ian learning framework to exhibit significant availability bias if the variability of the cli-
mate trend is correctly estimated. In fact, one of the possible alternative explanations for
the availability heuristic is that farmers use Bayesian learning, but over-estimate the vari-
ability of climate trends and underestimate the variability of weather. In such a case, we
would expect to see farmers over-weighting more recent weather shocks and under-weight-
ing more distant ones.
2 Data
Our empirical analysis focuses on Water District # 1 (WD1) in the State of Idaho, which
governs water allocation for the upper Snake River from Minidoka Dam to Milner Dam,
as well as tributaries along that segment of the Snake River. WD1 is the largest water dis-
trict in Idaho, accounting for water deliveries to nine major reservoirs and approximately
350 water users (Olenichak 2015). Among the 350 water users, approximately 200 are
13
Table 1 Summary statistics of variables

Variables Mean Median Min Max SD
Ex ante decision
Expected profit from crop-allocation ($/Acre) 277.86 255.85 73.27 686.28 129.51
Total expected profit ($/Acre) 199.3 184.42 0.19 686.27 121.07
% Planted Land 72.05 78.69 21.65 100 23.19
Land uses
% Alfalfa 55.36 59.1 0 100 36.84
% Barley 14.32 1.35 0 100 25.01
% Corn 10.88 0 0 100 22.08
% Potato 3.92 0 0 100 14.1
% Sugarbeet 2.52 0 0 100 10.91
% Wheat 10.05 0.73 0 100 20.01
% Fallow 2.94 0 0 96.97 8.92
Weather
Growing degree days 1549.48 1537.91 1017.66 2123.75 254.46
Extreme degree days 61.77 49.62 1.43 312.83 49.95
Summer precipitation (mm) 117.69 109.52 31.38 264.42 49.23
Water curtailment
No. days, 50% curtailment, summer 43.82 49 0 92 30.57
Others
Winter precipitation (mm) 150.17 146.65 63.37 266.38 35.8
Surface water supply index-reservoir 1.1 1.08 − 3.14 3.47 2.12
Farm area (acre) 112.8 67.15 4.58 1572.89 150.84
individual irrigators, and the rest are collective water management institutions, including
public irrigation districts and private irrigation and canal companies. Water is used pre-
dominantly for agricultural irrigation in WD1.
We chose WD1 for our analysis because the district operates and keeps the longest and
the most accurate information regarding water availability and curtailment decisions in the
state. Records on daily water calls for each river segment are available dating back to 1988,
including detailed information on how water calls are determined (see Olenichak 2015).3
Using these records as well as the adjudicated water rights database from the Idaho Depart-
ment of Water Resources (IDWR), we are able to reconstruct daily water availability for
every water user within WD1.
We collect data for 1056 unique farms for 2007–2016, with information on farm bound-
aries, attributes for water rights held by respective water providers, and daily water calls.
We spatially match these farms with geo-referenced databases on land use and crop alloca-
tion, temperature, and precipitation. Our data assembling strategy is similar to those used
in previous studies, including Leonard and Libecap (2016); Browne (2017) and Ji and
Cobourn (2018). Table 1 provides summary statistics.
3
Only Water District #11, the governing body of the Bear River, has similar temporal coverage regarding
water calls in Idaho. WD11 is not an ideal study site because the river is collectively managed by Idaho,
Utah, and Wyoming, which leads to complex water right structures.
13
Focusing our analysis on one particular sub-region allows us to address the challenge of
measuring irrigation water availability, an essential input and the “most important omitted
variable” (Schlenker et al. 2007) in irrigated agriculture in the western U.S. There exist
significant hydrological and legal differences in water allocation institutions between and
within states, as do efforts and abilities to measure, document, and enforce those institu-
tions. By focusing on a sub-region of which water management institutions are consist-
ent, well-managed, and well-documented, our study addresses the challenges presented by
omitted or poorly measure water availability, while simultaneously shedding light on the
importance of irrigation water to other semi-arid regions.
2.1 Surface Water Availability
Irrigation water deliveries in WD1 are administered under the doctrine of prior appropria-
tion, based on the principle of “first in time, first in right.” Available water in the district
is prioritized to fulfill senior water rights, or those that are established earliest in time.
Junior water rights are curtailed if there is not enough water to fulfill senior water rights on
any given day. An individual farmer either possesses his/her own water rights, or receive
water deliveries from an irrigation water provider, such as an irrigation district, irrigation
company, or canal company. In the latter case, the water provider serves as the ”middle-
man” between WD1 and the farmer. The water provider acquires water from WD1 through
water rights owned in common by all farmers within that provider’s service area and dis-
tributes water to its members (shareholders). Service areas of most surface water providers
are determined long before our study period, and remain so since then.4 Additionally, in
our analysis, we focus on farmers that can only acquire water from the water provider that
services the specific plots of land, thus eliminating the possibility that farmers switching
water providers/sources during our study period.5
We obtain daily curtailment records for each farm in our dataset from 1988 to 2017.
Curtailment decisions in WD1 are made using a computerized water accounting model
that calculates water availability in each of 36 unique river segments in the district (hereby
referred to as “reaches”).6 Each day, the district calculates water availability over the entire
sub-basin. It then issues a cutoff priority date for each reach following the prior appropria-
tion rule, with the goal of allocating all available water to the most senior water right hold-
ers throughout the entire district. For each reach, the cutoff priority date corresponds to the
priority date of the last water right fulfilled. All rights with a priority date later than the
cutoff priority date are not fulfilled. Depending on the spatial and temporal distribution of
water flows in the basin, the cutoff priority date can differ across different reaches and/or
on different days.7
4
Ji and Cobourn (2018) documents that most irrigation districts are formed between 1890 and 1920, with
the purpose of securing fundings for large irrigation projects. Thus boundaries of water providers are less
likely to be correlated with production conditions a century later.
5
In rare cases, farms may choose to acquire their own water rights apart from getting water from water
providers. We exclude those farmers in our analysis.
6
Figure 5 illustrates the reaches in WD1.
7
For example, on May 20th, 2014, reaches on most of the main stem of the Snake River, as well as reaches
on upstream tributaries of the Grey River, Salt River, Henry’s Fork, and Falls River, share a cutoff priority
date of December 14th, 1891. Reaches on the Teton River, Willow Creek, as well as the lower main stem of
the Snake River have a cutoff priority date that differs from the main stem. Table 6 shows the output from
the water accounting model on May 20th, 2014.
13
We also construct a database of water rights portfolios using administrative records

from WD1 and the Idaho Department of Water Resources (IDWR). A water portfolio is a
combination of one or more water rights owned by the same farm. Water rights in a water
portfolio share the same geographical place of use, with each water right having (poten-
tially) different priority dates, diversion points, and diversion rates.8 Almost all collective
irrigation institutions hold a water portfolio that contains multiple water rights. We then
combine water curtailment records with the water portfolio database, from which we are
able to determine daily curtailments for each water portfolio. Although curtailment deci-
sions for a single water right are binary on any given day, i.e., either curtailed or filled,
water deliveries for a farm with a water portfolio may be partially filled, i.e., some water
rights in the portfolio are curtailed, and others are filled. We characterize daily curtailment
decisions for a water portfolio as the percentage of diversions curtailed.
Since acreage and crop allocation decisions are made over the entire growing season,
we need to aggregate daily curtailments into annual realizations. We count the number of
curtailment days annually and create 12 different measures of curtailment days by varying
three aggregation methods: (1) total number of days (TND) curtailed versus longest cur-
tailment streak (Streak); (2) Growing Season (April to October) versus Summer Months
(June to August); and (3) 100% versus more than 80% versus more than 50% percent of
permitted diversions curtailed. Our main empirical models present estimates using the total
number of days in summer months where more than 50% percent of the permitted diver-
sion is curtailed (TND.summer.50%), one of the more responsive measures. We conduct
robustness checks with other constructed measures, presented in the results section.
Our dataset on water curtailments improves upon previous studies on prior appropria-
tion water rights in both spatial and temporal coverage. In terms of spatial coverage, previ-
ous studies usually proxy the underlying distribution of curtailments through direct com-
parison of priority dates (Lee et al. 2017) or some order-preserving transformation over
priority date (Brent 2017; Ji and Cobourn 2018). However, priority date only preserves the
order of water seniority within each reach, but can differ widely in the probability distribu-
tions of curtailment across reaches and sub-basins. By directly measuring water availabil-
ity using curtailment records, we eliminate the concern that spatial heterogeneity in water
availability for the same priority date across reaches may be correlated with some location-
specific unobservables, which can potentially bias statistical inference.
In terms of temporal coverage, most previous studies rely on time-invariant attributes
of water rights (e.g. Mukherjee and Schwabe 2015; Brent 2017; Lee et al. 2017; Ji and
Cobourn 2018) or some long-run expectations over water delivery (e.g. Schlenker et al.
2007), which cannot capture dynamically evolving expectations over water deliveries. Two
exceptions are Buck et al. (2014) and Manning et al. (2017), which measure the amount of
water delivered and curtailed, respectively, over a growing season at the county level. Our
dataset goes a step farther to quantify variation at the level of the water right and create
various measures of water curtailment attributes.
8
Geographical boundaries are obtained from the water rights place of use (WRPOU) database from
IDWR. Diversion point, diversion volume, and priority date are from Olenichak (2015).
13
Fig. 2 Map of farms in water district 1 of Idaho. Each green dot denotes one farm. Size of the dot indicates
relative size of the farm. Blue line denotes the main stem of the Snake River. Lower-right panel denotes the
relative location of Water District 1 (Red line denotes the boundary of the district.)
2.2 Farm Boundaries
Farm boundary data were purchased from FarmMarketID, a private firm that collects data
on farm characteristics for marketing purposes. We obtain geo-referenced farmland owner-
ship for the year 2016, where each farm in our sample owns one or more common land
units (CLU).9 For this study, we use farms that are serviced by only one collective water
provider, either an irrigation district or an irrigation or canal company.10 By doing so, we
exclude farms that have access to any individual water rights, or those within the service
boundaries of more than one water provider. These measures ensure that water availability
and farm boundaries are clearly defined for the farms in the analysis. Figure 2 shows the
locations of the farms included in our analysis.
9
According to the definition from USDA-Farm Service Agency, a CLU is an individual contiguous farm-
ing parcel, which is the smallest unit of land that has four characteristics: (1) a permanent, contiguous
boundary; (2) common land cover and land management; (3) a common owner, and/or (4) a common pro-
ducer association.
10
We define this using the criteria that over 90% of the total land mass of the farm is covered by that pro-
vider.
13
We make several assumptions regarding water rights and farm boundaries. First, we
assume that farm boundaries remain constant from 2007 to 2016. Due to data availability
constraints, we are unable to obtain farm boundaries earlier than 2016. We do not expect
this to jeopardize inference if farmland sales are accompanied by the sale of corresponding
right(s) to be serviced by water providers, which is usually the case. Second, we assume
that within each irrigation district, water is allocated proportionally to each farm. This
assumption is widely adopted by previous studies, including Schlenker et al. (2007) and
Buck et al. (2014). Our main inference will not be jeopardized as long as water is not allo-
cated using a seniority-based system within each irrigation district. As Ji and Cobourn
(2018) documented, water within irrigation districts is usually allocated to members in pro-
portion to the share(s) owned by each. Similar share systems are widely adopted by irriga-
tion and canal companies in our sample.11
2.3 Acreage and Crop Allocation
Acreage and crop-allocation data are obtained from the USDA Cropland Data Layer
(CDL), a field-scale remote sensing product developed for the continental U.S. using
satellite imagery and calibrated classification algorithms. For each farm, we identify the
percentage of land allocated to six major crops in the region: alfalfa, barley, corn, potato,
sugarbeets, and wheat, as well as land in fallow. These six crops and fallow account for
over 98% of the total cropland area in WD1 (National Agricultural Statistics Service
2007–2016).12 Therefore, we do not expect the exclusion of land in other crops to affect our
results. Our farm level fixed-effect approach offers additional safeguards against omitted
crop types as well as general misclassification in the CDL, as time-invariant misclassifica-
tions will be eliminated by the transformation.
2.4 Climate and weather
We obtain daily temperature and precipitation data from the PRISM climate dataset, a
medium-scale geo-referenced weather dataset with a spatial resolution of 4km. PRISM
has been used extensively in previous studies on climate change, such as Schlenker et al.
(2007). Growing degree-days (GDD), extreme degree-days (EDD), and growing season
precipitation are derived from PRISM data by accumulating daily temperature and pre-
cipitation over the growing season. Following Schlenker et al. (2007), we evaluate GDD
by accumulating daily degree days over the growing season, using cutoffs of 8 ◦ C and 32 ◦
C.13 We calculate EDD by accumulating daily degree days‘ exceeding the upper threshold,
11
Personal communication with Brian Olmstead, General Manager of the Twin Falls Canal Company,
6/16/2016.
12
This is calculated based on the number of pixels in the Cropland Data Layer in each land use within each
farm. There are 60 other categories of crops reported in the CDL. The majority of land in other crops is in
dry beans (1.1% on average across years). All other crops occupy less than 1% of the land base in agricul-
ture (e.g., oats, canola, peas, onions, rye, mint, hops, grapes, tree crops).
13
We construct GDD by accumulating degree days using the following equation:
⎧b −b if tmean ≥ b2
⎪ 2 1
GDDt =⎨ tmean − b1 if b2 > tmean > b1
⎪0
⎩
if tmean ≤ b1 (8)
�
GDD = GDDt
t
13
using a cutoff temperature of 32 ◦C.14 We calculate growing season precipitation as the

cumulative amount precipitation (in mm) between April and September, which reflects the
standard duration of the growing season in the region.
3 Empirical Strategy
Our empirical question examines how previous realizations in weather and water avail-
ability, transmitted through the expectation formation process, affect agricultural produc-
ers’ ex ante decision making. One inherent problem for many agricultural decisions, such
as acreage and crop allocation, is the timing mismatch between decision making earlier
in the growing season and the realization of key endowments later during the season. In
the planting stage of year t, producer i maximizes expected profit by making a production
decision, yit , based on his/her subjective expectations over a vector of J natural endow-
ments Eit (𝐖) , where 𝐖 ≡ W1 , … , WJ . At harvest, realizations of those endowments, 𝐖it ,
are observed, and the producer earns a realized profit. Assuming that the marginal effects
of endowments are separable from each other, previous studies have shown theoretically
that under certain conditions, the producer’s profit maximization problem yields a series
of linear marginal effects between optimal production decisions and the expected level of
endowments (Moore and Negri 1992; Ji et al. 2018; Cui 2020).15 Our conceptual model
can be written as:
∑
yit = 𝛼j Eit (Wj )
j
T
∑ (10)
Eit (Wj ) = 𝛽js wi,j,t−s
s=1
where the producer’s decision, yit , is a linear function of expectations over each natural
endowment, Eit (Wj ) ; and the expectation over each endowment is in turn a linear combina-
tion of its historical realization, wi,j,t−s . The expectation formation process follows Eq. 6,
where each of the previous T shocks, wi,j,t−s , contributes to the expectation with a weight of
𝛽js . These unspecified weights nest the three learning mechanisms presented in the previ-
ous section.
Footnote 13 (continued)
where b1 = 8 and b2 = 32 is the lower and upper threshold of GDD, and tmean is the daily mean tempera-
ture.
14
We calculate EDD by accumulating daily degree days‘ using a cutoff temperature of 32 ◦ C, i.e.,
{
tmax − b2 if tmax > b2
EDDt =
0 if tmax ≤ b2
∑ (9)
EDD = EDDt
t
where b2 = 32 is the threshold for EDD, and tmax is the daily maximum temperature.
15
While the non-separability between endowments can potentially bias estimates when modeling farm
profit (Fezzi and Bateman 2015; Hendricks 2018), in a land allocation model the impact of separability is
negligible (Ji et al. 2018).
13
Combining these two steps and adding in different types of natural endowments, our
final empirical model becomes:
T T T
∑ ∑ ∑
yit = 𝛾1s GDDi,t−s + 𝛾2s EDDi,t−s + 𝛾3s Precipitationi,t−s
s=1 s=1 s=1
T
(11)
∑
+ 𝛾4s WaterCurtailmenti,t−s + 𝜇i + vt + 𝜀it
s=1
where yit is the outcome of interest for producer i at year t; GDDi,t−s is the s years’ lag of
growing degree days; EDDi,t−s is the s years’ lag of extreme degree days; Precipitationi,t−s
is the s years’ lag of growing season precipitation, and WaterCurtailmenti,t−s is the s
years’ lag of irrigation water availability measured by days of curtailment. The model also
includes farm-level fixed effect, 𝜇i , and year fixed-effect, vt . Here the coefficient of lag s for
endowment j, 𝛾js , identifies the product of the marginal effect of an endowment on the deci-
sion, 𝛼j , and the lag-specific weight, 𝛽js , both of which appear in Eq. 9. We are not able to
separately identify 𝛼j and 𝛽js from our empirical model in Eq. 11, but since 𝛼j remains con-
stant for each endowment over time, we are able to identify the relative weights between
different lags.
Equation 11 can be used to test for all three learning mechanisms laid out in the pre-
vious section. Specifically, if a farmer uses long-run normals to infer the expectation
for endowment j, we expect all lagged coefficients on that endowment to be equal, i.e.,
𝛾j1 = 𝛾j2 = ⋯ = 𝛾jT . If a farmer engages in availability heuristic, we expect two patterns:
(1) the coefficients are larger when the shocks happen closer to the current period, i.e.
𝛾j1 > 𝛾j2 > ⋯ > 𝛾jT ; and (2) the coefficients are different from zero for lags closer to the
current period, but to be non-different from zero for lags farther away from the current
period, i.e., |𝛾js | > 0 for some s ≤ S and 𝛾s ≃ 0 for s > S . If a farmer engages in reinforce-
ment heuristic, we expect coefficients to be cyclical, as documented in Eqs. 6 and 7 .
We model three outcomes of interest in our empirical models: (1) the crop-allocation
effect, summarized as an expected profit from planted land; (2) the acreage effect, i.e.,
the fraction of planted land; and (3) total expected profit, which summarizes acreage
and crop-allocation decisions.
Crop-allocation adjustment is one of the main mechanisms through which farmers
adapt to climate conditions and water constraints (Seo and Mendelsohn 2008; Kala et al.
2012; Cobourn et al. 2017; Hagerty 2020). We aggregate crop mixes into a single meas-
ure of long-run expected profit, constructed as a function of field-scale crop allocation
decisions, state-level crop prices, average state-level crop yields, and region-specific
costs of production for the six major crops in the region. This reflects the differences in
expected profitability of growing different types of crops: some crops (e.g., corn) have
higher expected profits but are sensitive to drought and extreme heat, other crops (e.g.,
wheat) have lower expected profits but are tolerant to drought.
Acreage adjustment is another mechanism used by farmers to adapt to climate change
(Miao et al. 2015; Cohn et al. 2016; Cui 2020). Farmers may decide to adjust crop acre-
age by fallowing land or by expanding or shrinking production on marginal lands when
facing differing climate conditions and water constraints. We define acreage as the frac-
tion of a farm’s land base that is used to produce the six major crops mentioned above.
Finally, we aggregate crop-allocation and acreage margins to construct a measure-
ment of total expected profit, which is the product of the two. This gives us the total
13
extensive margin adjustment when facing climate shocks. All three outcome variables
reflect farmers’ ex ante production decisions undertaken early in the growing season.
We employ a panel fixed-effect approach to identify our main effect of interest by
exploiting time-series variation in previous weather and water availability shocks for each
individual farm. This approach offers two advantages over a cross-sectional or pooled
approach. First, by comparing the same individual at different time periods, we are able
to simultaneously control for the effect of location-specific average temperature and pre-
cipitation, as well as average water curtailments specific to each water right. Although
realizations of water availability vary from year to year, the underlying water portfolios for
the same farm remain constant, and thus the long-run mean water availability should also
remain constant.16 Second, since weather and water availability shocks are plausibly ran-
dom, we eliminate concerns about time-invariant omitted variable biases and spatially-cor-
related measurement errors, which plague empirical studies using cross-sectional regres-
sions (Auffhammer and Schlenker 2014; Blanc and Schlenker 2017).
The effect we capture here differs from those recovered in popular empirical approaches,
including the Ricardian approach (e.g. Mendelsohn et al. 1994; Schlenker et al. 2005), the
long-difference approach (e.g. Burke and Emerick 2016), and the contemporaneous fixed-
effect approach (e.g. Deschênes and Greenstone 2007; Schlenker and Roberts 2009). The
Ricardian approach draws inference from cross-sectional variation in climate, which recov-
ers the long-run difference in economic outcomes for different individuals as well as any
corresponding adaptation responses. Similarly, the long-difference approach also recovers
the long-run impact of climate, but through time-series variations in long-run climate nor-
mals for the same individual. Our approach is most similar to the contemporaneous fixed-
effect approach, yet there are still subtle differences between the two. The contempora-
neous fixed-effect approach recovers the impact of weather shocks on contemporaneous
economic outcomes, as well as any short-run adaptation responses that are contemporane-
ous or ex post to weather realizations. In contrast, our approach captures short-run ex ante
adaption responses undertaken prior to weather realizations by using fixed-effect models
with lagged dependent variables. In addition, our approach directly tests for the effect of
weather fluctuations on resource-allocation decisions, while most contemporaneous fixed-
effect studies focus on the combined effect of these decisions and actual weather and water
availability, such as crop yield or revenue.
4 Results
4.1 Estimates for Ex Ante Production Decisions
Results from our main model are presented in Table 2. First, we discuss results for acre-
age decisions, presented in Column (1). We find statistically significant effects of GDD
and EDD on cropland acreage. A one degree-day increase in GDD significantly increases
planted acreage by 0.034% and 0.055% in the 2nd and 3rd-year lags. The 1st, 4th, and
5th-year lags are not significant. A one degree-day increase in EDD decreases acreage by
16
Water portfolios were established from the late 1800s through the mid-1900s and rarely include new
rights with priority dates after the year 2000. Even if new rights are added to the portfolio, these newer
water rights have close to zero probability of receiving water during the growing season. This means that
the fundamentals of water portfolios remain mostly unchanged during our study period.
13
0.084%, 0.063%, and 0.153% in the 1st, 2nd and 3rd-year lags. It increases acreage by
0.095% for the 4th-year lag, and is statistically insignificant for the 5th year. We do not
find a significant impact of past precipitation on planted acreage. For water availability
variables, we find that shocks in water curtailments significantly decreases planted acre-
age in the 2nd year, but not significant for other lags. The direction of these effects are
all as expected: greater GDD increases crop yields, which encourages expansion onto
marginal lands for crop production and discourages fallow; greater EDD decreases crop
yields, which discourages the use of marginal lands for crop production and encourages
fallow; longer curtailments reduce expected supplemental irrigation water supply, leading
to decreasing crop yields, and thus agricultural land uses.
To evaluate the magnitude of producer responses, we calculate the discrete effects of
a one-standard-deviation (one-standard-deviation) increase in the independent variables,
which suggests that responses to past shocks are economically sizeable. A one-standard-
deviation increase in GDD increases acreage by 2.3%, 4.3%, and 7.0%, and a one-stand-
ard-deviation increase in EDD decreases acreage by 2.5%, 1.9%, and 4.6% in the 1st, 2nd,
and 3rd year. The sizes of the effect are even larger when aggregated over multiple years.
Columns (2) and (3) of Table 2 present two models using 3-year and 5-year moving aver-
ages as independent variables. A one-standard-deviation increase in GDD for three and five
consecutive years will increase planted acreage by 14.6% and 14.0%.17 A one-standard-
deviation increase in EDD for three and five consecutive years will decrease acreage by
8.8% and 9.0%.18 This suggests that farmers respond to past weather shocks by significantly
changing acreage decisions in the short run. The magnitude of the responses is on par with
estimates of the long-run acreage responses to temperature in the Eastern United States
(Cui 2020). We find a one-standard-deviation increase in curtailment days for the water
availability variables decreases planted acreage by 1.0% for the 2nd-year lag, and 1.2%
over three years. The 5-year moving average estimate for curtailment is not significant.
Next, we discuss results on crop-allocation decisions. Our main model, presented in
Column 4 of Table 2, suggests that most estimates for the three weather variables are sta-
tistically insignificant except for the 5th-year lag of EDD and the 2nd-year lag of precipita-
tion. In contrast, we find strongly significant water availability effects on crop-allocation
decisions, and the effects exhibit cyclical patterns. Producers react to a one-standard-devi-
ation increase in curtailment days by planting less profitable crop mixes in the first and
third year, which decreases their expected profit by $5.4/acre and $5.7/acre. In contrast, in
the 2nd and 4th year, producers plant more profitable crop mixes when seeing an increase
in curtailment days, of which a one-standard-deviation increase in curtailment leads to
$8.8/acre and $7.5/acre. Estimates for the first four years are highly significant statistically,
while the 5th year lag is insignificant.
However, when evaluated over multiple years, the cyclical responses to water availabil-
ity shocks disappear. Columns (5) and (6) in Table 2 presents models where we use 3-year
and 5-year moving averages (MA3 and MA5, respectively) of water shocks instead of year-
specific shocks in the fixed-effect model. We find that both MA3 and MA5 of water curtail-
ment do not have a statistically significant impact on crop allocation. This indicates that
17
A back-of-the-envelope calculation suggests that a one-standard-deviation increase in GDD corresponds
to about 0.7 ◦ C of warming, assuming that warming is evenly spread out over the growing season.
18
A back-of-the-envelope calculation suggests that one-standard-deviation increase in EDD corresponds
to about 1.1 ◦ C of warming, assuming that warming is evenly spread over a 30-day period where daily tem-
peratures exceed the EDD threshold.
13
the short-run cyclical responses to past curtailment shocks would be concealed if either
Ricardian regression or the long-difference approaches were adopted, as both approaches
smooth yearly shocks into medium to long-run normals.
Finally, we turn to models on total expected profits, presented in Column (7) of Table 2.
EDD has the most robust impact on total expected profit among the three variables: the
1st, 2nd, 3rd, and 5th-year lags of EDD are negative and statistically significant. GDD
and precipitation both have positive effects on total expected profit. However, most of the
estimates are statistically imprecise: only the 2nd lag of GDD and the 2nd and 3rd lags
of precipitation are statistically significant, while other lags of GDD and precipitation are
statistically insignificant. We do find statistically significant responses to all weather vari-
ables when evaluated using multi-year moving averages (presented in Column (8) and (9)
of Table 2), and the direction of the responses are as expected. Increases in GDD and pre-
cipitation have positive effects on total expected profit, while increases in EDD have nega-
tive effects on total expected profit. The magnitudes of effects of moving average estimates
are larger than estimates for individual-year estimates.
For the water curtailment variables, we again find evidence of cyclical responses in total
expected profit to water availability shocks. Water curtailment days have negative and sig-
nificant effects on total profit for the 2nd and 4th-year lags, and have positive and signifi-
cant effects for the 1st and 3rd-year lags. A one-standard-deviation increase in curtailment
days significantly decreases total expected profit by $3.76/acre in the first year, increases it
by $5.00/acre in the second year, decreases it by $4.63/acre in the third year, and increases
it by $10.70/acre in the fourth year. The fifth lag is not statistically significant. We again
generate two aggregate measures for the multi-year impacts on total profit, and find that
water availability has no significant impact over either the 3-year or the 5-year averages.
Overall, we find that agricultural producers significantly respond to recent shocks
in natural endowments by changing ex ante production decisions in the short run. These
responses generally differ between different types of endowments. A GDD shock has strong
positive effect on cropland acreage, no significant effect on crop allocation, and weak posi-
tive effect on total expected profit. An EDD shock has a strong negative impact on cropland
acreage, almost no effect on crop allocation, and a strong negative effect on total expected
profit. Precipitation has a weak positive effect on crop allocation, and no effect on planted
acreages. A water availability shock exhibits cyclical effects on crop-allocation decisions,
but this effect is canceled out when evaluated over 3 or 5-years’ time. It also negatively
affects planted acreage, though the effect is not as strong as the effects on crop allocation.
4.2 Expectation Formation Mechanisms
Our results suggest that agricultural producers deviate from using long-run normals in sev-
eral ways. First, we find that ex ante adaptation behaviors significantly respond to previous
shocks in endowments. Responses to previous shocks are persistent across different types
of endowments, as well as across different adaptation mechanisms. This contradicts the
hypothesis that farmers’ expectations are formed through Bayesian updating using a long
history of realizations, and the hypothesis that the expectations are drawn from long-run
normals. Second, we find that the magnitudes of adaptation responses vary widely between
different lags. This further contradicts the long-run normals hypothesis, which predicts that
each lagged realization enters the posterior expectation with equal weight.
Instead, we find evidence that, to varying degrees, support alternative heuristics for
expectation formation. We find evidence for the availability heuristic, which predicts that
13
more distant lags will not be considered as part of the expectation formation process. We
find that a number of endowments have significant effects on ex ante decisions in the first
three shocks, for example, water curtailment on crop allocation, GDD, EDD, and water
curtailment on planted acreage, and GDD, EDD, precipitation, and water curtailment on
total expected profit. Most of these effects die off in 5 years, with the only exception being
the effect of EDD on crop allocation and total expected profits. To provide further evidence
on potential recency biases, we estimate additional models by including one additional lag
(lag 6) into our framework. The extended models suggest that none (0 out of 20) of the
coefficient estimates for year 6 are statistically significant at the 10% level (see Appendix
Table 10). This lends additional support to our claim on the availability heuristic, such that
producers over-utilize more recent realizations when making production decisions.
We find limited evidence to support the hypothesis that that more recent shocks carry
greater weight than more distant shocks, another prediction of the availability heuristic.
Most effects do not die off gradually, and instead exhibit either cyclical patterns (e.g., the
effect of water curtailment variables on all dependent variables) or die off suddenly at the
fourth or fifth lag (e.g., the effect of GDD and EDD on acreage).
We find evidence consistent with reinforcement learning, which predicts that previous
shocks enter expectations in a cyclical fashion. The cyclical pattern appears strongly in
the water curtailment variables the crop allocation responses. In the first year, the farmer
reacts to drought by adjusting the crop allocation towards less water-intensive (and lower
profit) crops. This is likely an over-reaction since the long-run distribution of water cur-
tailment has not changed significantly. In the second year, the farmer realizes that his/her
strategy under-performs by planting too little to water-intensive crops, and over-correct his/
her belief and plants more water-intensive crops. In the third year, the farmer under-per-
forms again by planting too much of the water-intensive crop and again over-corrects in the
opposite direction. As such, the reinforcement strategy causes the farmer to move back and
forth in his/her expectations and the corresponding adjustments in crop mix. The cyclical
pattern is also present for the acreage response to water curtailments, and the crop alloca-
tion response to GDD, though in both cases, the magnitudes of the response are estimated
imprecisely.
Both of the alternative heuristics examined here capture more pronounced reactions to
weather and water shocks over the short run, relative to the Bayesian expectation forma-
tion process. We argue that using medium or long-run differences to estimate the impact of
climate change, such as is common in Ricardian regression or the long-difference method
(Burke and Emerick 2016), may fail to capture this effect. This is problematic, especially
if reinforcement learning dominates the learning process, as we have found for shocks
in water curtailment: a cyclical pattern of water curtailment on crop allocation and over-
all profit exists, but these effects are entirely masked in the MA3 and MA5 models. This
empirical observation is also documented in Buck et al. (2014), which suggests that farm-
land prices are only significantly affected by water delivery shocks by the first two orders
of moving averages, but not by MA3 or higher orders.
4.3 Robustness Checks
4.3.1 Crop Share Models
To further shed light on the mechanism through which agricultural producers react to
weather and water shocks, we present regression results explaining the shares of land
13
allocated by crop categories. Specifically, we group our crops into three categories:
water-intensive crops, drought-tolerant crops, and the perennial crop, following the
agronomic literature and field knowledge of our study region. Water-intensive crops,
including potatoes, sugarbeets, and corn, generally have higher expected profit (around
$600/acre), but are more sensitive to heat and/or water stress (Yonts et al. 2003; King
et al. 2006; FAO 2020). Drought-tolerant crops, including barley and wheat, have
lower expected profit (around $100/acre), but are less sensitive to heat and/or water
stress. All three crops in the water-intensive category must be produced with irrigation,
while field crops like barley and wheat can potentially be produced without irrigation
in our study region. We use the only perennial crop, alfalfa, as the baseline category
because it lies in the middle of the two above categories in terms of both profitability
and drought sensitivity. Alfalfa has a moderate expected profit ($250/acre), and is gen-
erally less sensitive to shorter periods of under-irrigation comparing to some of our
drought-sensitive crops (Shewmaker et al. 2011; Obidiegwu et al. 2015). In addition,
producers can adapt to water stress by targeted irrigation or harvesting at an earlier
time (Liu et al. 2018). All these factors make alfalfa the default choice in our study
region, with over half of the agricultural lands devoted to it. Overall, these categories
capture broad differences between types of crops, which are of greatest importance,
whereas the differences within each category, for example, between potato and sugar-
beets, are less important because they are both of higher profit and drought-sensitive.
Results from these crop-share models are presented in Figs. 3 and 4, and are generally
in line with our main models. In the case of water-intensive crops, we find evidence in sup-
port of the availability heuristic, such that all estimates for the last (5th) lag are statistically
insignificant. Furthermore, producers’ responses to EDD and precipitation decrease over
time, which correspond to one of the key features of the recency heuristic. The signs of the
responses are all intuitive: increases in GDD and precipitation support planting of more
water-intensive, higher profit crops, while an increase in EDD leads to less water-intensive
planting choices.
In the case of drought-tolerant crops, we find mixed evidence regarding the availabil-
ity heuristic. Producers decrease their shares of drought-tolerant crops when they see an
increase in GDD. Combining with the model on water-intensive crops, this indicates that
producers substitute drought-tolerant crops for the perennial crop when GDD increases.
The magnitude of response towards GDD generally decreases over time, which corre-
sponds to the recency hypothesis. On the other hand, crop share responses to EDD and
precipitation do not support the availability heuristic, as in both cases, the magnitudes
of the coefficients do not decrease or trend over zero over time. Increases in EDD leads
to smaller shares of drought-tolerant crops in the 3rd and 4th year, and larger shares of
those in the 5th year, which suggest that producers primarily adapt to EDD by substitut-
ing between drought-tolerant crops and alfalfa, while keeping the share of water-intensive
crops unchanged. An increase in precipitation decrease shares of drought-tolerant crops in
the 2nd and 5th year. This is also within our expectation as an increase in water availability
shifts the production towards more water-intensive crops.
More interestingly, we find consistent evidence in support of the reinforcement heuris-
tic. In both crop share models, producers exhibit cyclical adjustment patterns when fac-
ing a shock in water curtailment. Specifically, when encountering a water curtailment
shock, producers increase the share of drought-tolerant crops and decrease the share of
water-intensive crops in the 1st and 3rd years. They adjust in the opposite direction in
the 2nd and 4th year by decreasing the share of drought-tolerant crops and increasing the
share of water-intensive crops. The estimated coefficients are statistically significant and
13
Fig. 3 Marginal effects on shares of water intensive crops. Black dot denotes the estimated marginal effect,
and grey ribbon denotes 90% confidence interval
economically sizeable: a one-standard-deviation change (20.3 curtailment days) in water

curtailment will lead to cyclical adjustments of up to 2% change in the share of water-
intensive and drought-tolerant crops.
13
Fig. 4 Marginal effects on shares of drought tolerant crops. Black dot denotes the estimated marginal effect,
and grey ribbon denotes 90% confidence interval
4.3.2 Placebo Tests
Since both crop allocation and acreage decisions are made prior to the actual realization,
it is expected that the ex post information, i.e., shocks happening after the decisions, will
13
Table 2 Parameter estimates of all lags, MA3, and MA5 models
Variables Planted acreage Crop allocation Total expected profit
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Growing degree days

lag1(GDD) 0.000184 − 0.0257 0.0370
(0.000165) (0.0765) (0.0774)
lag2(GDD) 0.000341*** 0.0423 0.161***
(0.000119) (0.0528) (0.0323)
lag3(GDD) 0.000551*** − 0.0822 0.0567
(0.000152) (0.0664) (0.0608)
lag4(GDD) − 0.000110 − 0.000102 0.0975 − 0.0482 0.0675 − 0.0655
(0.000145) (8.28e−05) (0.0839) (0.0567) (0.0665) (0.0501)
lag5(GDD) 0.000148 0.000215 − 0.0485 − 0.0200 0.00293 0.0248
(0.000218) (0.000131) (0.119) (0.124) (0.110) (0.0942)
MA3(GDD) 0.00115*** 0.0147 0.336**
Weather Fluctuations, Expectation Formation, and Short-Run…
(0.000291) (0.143) (0.141)

MA5(GDD) 0.00110*** 0.00762 0.324**
(0.000275) (0.183) (0.159)
Extreme degree days
lag1(EDD) − 0.000842*** − 0.186 − 0.383**
(0.000183) (0.208) (0.184)
lag2(EDD) − 0.000632** − 0.228 − 0.403**
(0.000243) (0.168) (0.158)
lag3(EDD) − 0.00153*** − 0.0963 − 0.415***
(0.000530) (0.234) (0.151)
lag4(EDD) 0.000948** 0.000789*** 0.116 0.195 0.195 0.299*
(0.000388) (0.000148) (0.175) (0.188) (0.219) (0.160)
99
13
Table 2 (continued)
100

(1) (2) (3) (4) (5) (6) (7) (8) (9)
13
lag5(EDD) − 0.000242 − 0.000186 − 0.272* − 0.236 − 0.350*** − 0.313***
(0.000416) (0.000349) (0.146) (0.155) (0.110) (0.114)
MA3(EDD) − 0.00295*** − 0.378 − 1.089***
(0.000488) (0.458) (0.306)
MA5(EDD) − 0.00301*** − 0.509 − 1.312***
(0.000748) (0.486) (0.262)
Growing season precipitation
lag1(prec) − 1.90e−05 − 0.106 − 0.0592
(0.000144) (0.187) (0.145)
lag2(prec) − 4.25e−05 0.378*** 0.245**
(0.000234) (0.122) (0.0941)
lag3(prec) 0.000253 0.0973 0.146*
(0.000172) (0.0993) (0.0816)
lag4(prec) − 7.62e−05 − 2.36e−05 0.115 0.142 0.0947 0.112
(0.000287) (0.000309) (0.203) (0.217) (0.180) (0.174)
lag5(prec) − 1.31e−06 6.51e−05 0.162 0.193* 0.126 0.186
(0.000308) (0.000276) (0.106) (0.111) (0.126) (0.129)
MA3(prec) − 1.21e−05 0.464** 0.385**
(0.000207) (0.224) (0.168)
MA5(prec) 0.000486 0.725*** 0.656***
(0.000485) (0.232) (0.200)
Water curtailment
lag1(Curtailment) 7.42e−05 − 0.266*** − 0.185*
(0.000175) (0.0939) (0.0981)
X. Ji, K. M. Cobourn
(1) (2) (3) (4) (5) (6) (7) (8) (9)
lag2(Curtailment) − 0.000486** 0.432*** 0.246*

(0.000180) (0.147) (0.123)
lag3(Curtailment) − 0.000128 − 0.282** − 0.228**
(0.000185) (0.107) (0.103)
lag4(Curtailment) − 0.000178 − 0.000108 0.368*** 0.302*** 0.245*** 0.216***
(0.000181) (0.000196) (0.0982) (0.104) (0.0786) (0.0768)
lag5(Curtailment) 0.000192 0.000136 − 0.120 − 0.00519 − 0.0714 − 0.00106
(0.000165) (0.000166) (0.118) (0.113) (0.0770) (0.0839)
MA3(Curtailment) − 0.000614* − 0.0607 − 0.126
(0.000357) (0.190) (0.193)
MA5(Curtailment) − 0.000550 0.351 0.181
(0.000422) (0.299) (0.248)
Columns 1–3 present models on planted acreage; Columns 4–6 present models on expected profit from crop-allocation. Columns 7–9 presents models on total expected profit;
Winter precipitation, individual fixed-effects, and year fixed effects are added to all models, and are suppressed from table. Robust standard errors in parenthesis, clustered at
the level of water providers; “lag#” denotes number of years that the variable in question took place prior to the current growing season. “MA3” denotes the 3 year moving
average of a variable, i.e. the average of the last 3 lags
A triple asterisk indicates p < 0.01 ; a double asterisk indicates p < 0.05; a single asterisk indicates p < 0.1
101
13
Table 3 Parameter estimates for placebo tests

Variables Planted Acreage Crop Allocation Total expected profit
(1) (2) (3) (4) (5) (6)
GDD − 0.000504** − 0.000636*** 0.0597 0.111 − 0.0848 − 0.107

(0.000201) (0.000177) (0.0946) (0.0907) (0.105) (0.0855)
lead1(GDD) 3.35e−05 0.0531 0.0309
(8.05e−05) (0.0506) (0.0570)
EDD 8.91e−05 − 0.000172 0.0797 0.00233 − 0.0169 − 0.112
(0.000229) (0.000330) (0.180) (0.155) (0.140) (0.156)
lead1(EDD) 0.000140 − 0.235* − 0.113
(0.000206) (0.138) (0.137)
prec − 0.000204 0.000224 0.152 0.236 0.0687 0.194
(0.000230) (0.000333) (0.126) (0.150) (0.108) (0.159)
lead1(prec) 8.71e−05 0.0524 0.0586
(6.20e−05) (0.0815) (0.0703)
Curtailment 0.000229 4.51e−05 0.0361 0.0852 0.0718 0.0630
(0.000244) (0.000253) (0.0759) (0.0772) (0.0667) (0.0742)
lead1(Curtailment) − 0.000421 0.276 0.0905
(0.000326) (0.184) (0.144)
Columns (1) and (2) presents the model on planted acreage; Columns (3) and (4) present models on
expected profit from crop-allocation; Column (5) and (6) presents the model on total expected profit; All
models include five lags of GDD, EDD, prec, and TND.Summer.50%, winter precipitation of the current
year, as well as individual and time fixed-effects. These variables are suppressed from reporting; Lagged
water curtailment variables used here are total number of days of when at least 50% of water supply are
curtailed during the entire growing season (TND.Summer.50%), and are suppressed from reporting; Robust
standard errors in parenthesis, clustered at the level of water providers
A triple asterisk indicates p < 0.01 ; a double asterisk indicates p < 0.05 ; a single asterisk indicates p < 0.1
have no effect on ex ante decisions. We check for this hypothesis by augmenting the main
models with a series of ex post shocks, which serve as placebos. This includes realizations
of GDD, EDD, precipitation, and water curtailment for the current and the immediate next
years.
Results for the placebo tests are presented in Table 3. We find that only 3 out of 36 pla-
cebo variables are statistically significant: the effect of current year precipitation on crop
allocation and on total expected profit, and the effect of GDD on planted acreage. All other
placebo variables have statistically insignificant effects on ex ante production decisions.
This suggests that our estimate on {textitex ante decisions are not likely to be driven by ex
post realizations of weather and water availability.
4.3.3 Additional Ex Ante Information
We also check if any additional ex ante information, observed by the econometrician and
potentially also by farmers, influences our results. To introduce a source of bias, the ex
ante information must have both cross-sectional and time-series variation, which rules out
13
common climate and weather prediction models at the regional scale.19 We run robustness
tests using one additional piece of information: reservoir surface water supply index (SWSI
Reservoir), which measures the water storage level at the beginning of the growing sea-
son.20 We have already included an additional variable on ex ante information, cumulative
winter precipitation between November and February, in the main model. Though Both
variables are documented to affect acreage decisions by the previous literature (Manning
et al. 2017), our empirical results (presented in Table 7) suggest that our main models are
robust to the addition of ex ante information, as the parameter estimates in the extended
models are very close to the main models.
4.3.4 Stationarity and Predictability of Weather Shocks
Until now, we have assumed that weather and water availability shocks are largely unpre-
dictable events. If the current shocks can actually be predicted using past shocks, then it
could be beneficial for farmers to rely more heavily on past realizations when forming
expectations. Here we check for two possibilities that may lead to a rational expectation for
over-reliance on recent events.
First, we check for whether included weather variables are stationary over time. If there
exists time-dependence (which may or may not be caused by climate change), then it could
be rational for producers to rely on some sort of alternative learning mechanisms other than
the long-run normals. We check this possibility by performing the Im-Pesaran-Shin (IPS)
test for unit roots in panel data models (Im et al. 2003), recommended for cases with fixed
T and large N (Baltagi Badi 2013). To mimic our main fixed-effect models, we perform the
test with demeaned variables but allow for time trends, and present these results in Table 8.
We find that for all weather variables (GDD, EDD, and prec) as well as different forms of
water curtailment variables, the IPS test strongly rejects the null hypothesis that the panel
has unit-roots. This provides suggestive evidence that the weather variables are stationary
over our study period.
Second, we specifically check for the predictability of water curtailment shocks, noting
that unlike climate variables, water availability is not only governed by stochastic natu-
ral processes, but additionally by inter-year storage management that could cause water
availability to be serially correlated from year to year. We check this possibility using a
leave-one-out validation strategy. For each year from 2007 to 2016, we train a prediction
model using data for the past 15 years, leaving out the year in question. The prediction
model is constructed using an elastic net algorithm with 10-fold cross-validation (Hastie
et al. 2009). We then use the post-elastic-net prediction model to generate predictions for
the water curtailment shock, and compare that with the actual shock for the year we have
left out. If the predictive model is able to explain some variations for the current shock,
then the unpredictability assumption is violated.
19
For example, long-run climate forecasts from IPCC and short-run weather patterns drawn from El Niño-
southern oscillation (ENSO) forecasts only have time-series variations at the regional scale, and will be
assimilated in the time fixed effects.
20
We acquire reservoir level at the beginning of the growing season (Mar.31) from the monthly water out-
look report published by NRCS. Reservoir levels are then rank-standardized based on 30-year historical
observations. We do not use the NRCS forecasts on surface water flows because (1) empirically, the forecast
does not predict the actual surface water flows well ( R2 < 0.1 ); and (2) the prediction is mainly based on El
Niño-southern oscillation (ENSO) forecasts and winter precipitation, both of which are controlled in our
model.
13
Table 4 Dynamic panel estimates

Dependent variable Crop allocation
FE AB 1-lag AB 2-lag
(1) (2) (3)
Lagged dependent variables

lag1(y) − 0.203*** − 0.116***
(0.0345) (0.0429)
lag2(y) 0.104**
(0.0408)
Growing degree days
lag1(GDD) − 0.0257 0.106 − 0.150
(0.0765) (0.0771) (0.117)
lag2(GDD) 0.0423 − 0.0343 0.0381
(0.0528) (0.0841) (0.0960)
lag3(GDD) − 0.0822 0.153 0.0307
(0.0664) (0.107) (0.147)
lag4(GDD) 0.0975 − 0.0524 0.0303
(0.0839) (0.0739) (0.112)
lag5(GDD) − 0.0485 0.160** 0.267***
(0.119) (0.0801) (0.0783)
Extreme degree days
lag1(EDD) − 0.186 − 0.0252 0.0545
(0.208) (0.174) (0.204)
lag2(EDD) − 0.228 − 0.168 − 0.325
(0.168) (0.217) (0.230)
lag3(EDD) − 0.0963 − 0.272 0.0355
(0.234) (0.231) (0.273)
lag4(EDD) 0.116 0.529** 0.0548
(0.175) (0.248) (0.296)
lag5(EDD) − 0.272* − 0.502** − 0.698***
(0.146) (0.202) (0.260)
lag1(prec) − 0.106 − 0.0269 − 0.00138
(0.187) (0.147) (0.162)
lag2(prec) 0.378*** 0.0847 0.305
(0.122) (0.155) (0.193)
lag3(prec) 0.0973 0.0171 0.271
(0.0993) (0.142) (0.165)
lag4(prec) 0.115 0.205 0.199
(0.203) (0.143) (0.160)
lag5(prec) 0.162 0.426*** 0.604***
(0.106) (0.147) (0.176)
Water curtailment
lag1(Curtailment) − 0.266*** − 0.175 − 0.124
(0.0939) (0.124) (0.181)
13
Dependent variable Crop allocation
FE AB 1-lag AB 2-lag
(1) (2) (3)
lag2(Curtailment) 0.432*** 0.377*** 0.380***

(0.147) (0.0989) (0.102)
lag3(Curtailment) − 0.282** − 0.0328 − 0.0789
(0.107) (0.125) (0.137)
lag4(Curtailment) 0.368*** 0.396*** 0.411***
(0.0982) (0.106) (0.120)
lag5(Curtailment) − 0.120 0.297** 0.223
(0.118) (0.137) (0.168)
Arellano–Bond’s z-statistics − 16.475*** − 15.004***
H0 : no autocorrelation 2.069** − 0.602
Columns 1 presents two-way fixed effects models, identical to the main specification. Columns 2 and 3 are
Arellano–Bond estimators with one and two period of lags for the dependent variable; Winter precipitation,
individual and year fixed effects are added to the fixed effect models, and are suppressed from reporting.
Winter precipitation and annual crop prices are added to the Arellano–Bond models, and are suppressed
from reporting. Robust standard errors in parenthesis, clustered at the level of water providers; “lag#”
denotes number of years that the variable in question took place prior to the current growing season
We present results from the predictive models in Table 4. Though the models perform
well on in-sample predictions, they perform poorly for the omitted year. All prediction
models increase the variations for the test sample rather than decreasing it, indicating that
these models have no predictive power over future water curtailments. Thus, we rule out
the possibility that previous shocks in water curtailment can be used to predict the shock in
the upcoming growing season.
4.3.5 Whether Crop Rotation Drives Cyclical Coefficients
Our main estimation framework suggests that there exist cyclical patterns in the responses
to past shocks, which is consistent with reinforcement learning. An alternative explanation
for this pattern might be that farmers adjust crop allocation cyclically because of crop rota-
tion rather than responding to weather shocks. Econometrically, this means there is likely
negative serial correlation in crop allocation if that is true.
We check for this possibility by estimating an Arellano and Bond (1991)’s dynamic
panel estimator. The Arellano–Bond estimator directly includes lagged dependent varia-
bles on the right-hand side, and uses a generalized method of moments (GMM) framework
to estimate a first-differenced model. Arellano–Bond’s z-test suggests that including two
lags of the dependent variable is just sufficient to wipe out serial correlation in the first-
differenced error term, and thus we focus our interpretation on the model with two lags,
presented in Column (3) of Table 5.
We find that the first lag of crop allocation has a negative and significant effect of
−0.116 , while the second lag has a positive and significant effect of 0.104. This suggests
13
Table 5 Predictive model of Year R2 within sample R2 out of sample

TND.Season.100%
2007 0.280 − 0.151
2008 0.239 − 0.262
2009 0.203 − 1.176
2010 0.396 − 0.860
2011 0.303 − 0.424
2012 0.238 − 0.812
2013 0.329 − 0.165
2014 0.170 − 0.395
2015 0.132 − 1.319
2016 0.136 − 1.194
Training samples are constructed using observations within 15 years

to the year of interest. Testing samples are observations of the year of
interest. Elastic net mixing parameter 𝛼 is fixed at 0.5. Lasso penalty
parameter 𝜆 is selected as 1-standard-deviation larger than the optimal
𝜆 that minimizes prediction error. Both within sample and out of sam-
ple r2 are calculated as 1 − Total Sum of Squares
Residual Sum of Squares
that, on average, about 10% of land each year is constrained in crop types because of crop
rotations. However, we also find that although the parameter estimates for water curtail-
ment variables are slightly different, the cyclical pattern largely remains. Estimates of water
curtailment variables are positive and significant for the 2nd and 4th-year lags. Estimates
for the 1st and 3rd-year lags are still negative, but quantitatively smaller and statistically
insignificant. Nevertheless, the dynamic panel model presented here still exhibits strong
cyclical patterns in the water curtailment variables. As such, we conclude that although
there exist crop rotation effects, this explanation is not likely the main driver of the cyclical
responses to water shocks.
4.3.6 Choice of Aggregation Methods for the Water Curtailment Variable
In the main specification, we choose to present one specific way to aggregate daily curtail-
ment records to the growing season: the total number of days during summer months that
at least 50% of the water is curtailed (TND.Summer.50%).21 However, unlike the well-doc-
umented non-linear effects of temperature on agricultural production, there is little guid-
ance in the previous literature, theoretically or empirically, documenting which aggregation
should be used in the case of irrigation water availability. As such, we conduct robustness
checks by running the empirical analysis with different aggregation methods. Results of
these alternative specifications are presented in Table 9.
We find that using different aggregation methods leads to similar results qualita-
tively, although there are subtle differences between different aggregation methods.
In the planted acreage model, all models indicate negative effects of curtailment on
planted acreage, but the timing of those effects are different. Two other variables,
21
We compound our aggregation methods with three components: TND versus Streak denotes whether
we aggregate the total number or the longest streak of curtailment days. Season versus Summer denotes
whether the aggregation is over the entire season or only in summer months. 100% versus 50% denotes the
qualification for a curtailment day using all, or half, of the water in a portfolio is curtailed.
13
TND.Season.100% and Streak.season.100%, have negative and statistically significant

impacts on planted acreage in the farther lags. This is different from the response to the
main variable, TND.Summer.50%, which shows a significant and negative effect in the
2nd lag. In the crop allocation model, we find strong cyclical effects in all water cur-
tailment variables. Similar to the main model using TND.Summer.50%, the other three
models also show negative estimates for the 1st and 3rd lags, and positive estimates for
the 2nd and 4th lag, albeit some estimates are not statistically significant. In the total
expected profit model, patterns are similar to those in the crop allocation models, such
that all four aggregation methods show cyclical patterns, with the 1st and the 3rd lags
negative, and the 2nd and the 4th lags positive. In the TND.season.100% model, the
5th lag is also negative and statistically significant, indicating that the cyclical pattern
might span further.
Overall, these results indicate that agricultural producers are more responsive to the
total days of curtailments rather than curtailment streaks, and to summer months rather
than the entire growing season. On the other hand, while by using the chosen aggrega-
tion method we are able to capture most of the producer responses to water curtail-
ment, our chosen model will inevitably omit some of the responsive mechanisms not
incorporated in our aggregation method. As such, our main specification represents a
lower bound of the actual responses to water shortages.
5 Conclusion
A weather shock is the realization of climate at a certain point in time, but the economic
impacts of weather fluctuations have the potential to affect agricultural production for mul-
tiple growing seasons in the future. In this study, we show that past weather fluctuations
can affect future agricultural decision making through the mechanism of expectation for-
mation. Using a fixed-effect approach, we find significant behavioral responses in acreage
and crop-allocation decisions to past shocks in natural endowments, including temperature,
precipitation, and water availability. This suggests that instead of using the long-run nor-
mals, agricultural producers use alternative learning mechanisms that rely more heavily on
recent weather realizations. This over-weighting of recent realizations can result in eco-
nomic losses if farmers over-react to recent signals by making acreage and crop-allocation
decisions that are sub-optimal relative to those made under expectations based on a longer
and more representative history of weather observations.
There are several policy implications that follow from this study. First, our results stress
the need to incorporate short-run behavioral responses to the assessment of climate change
impacts. Economic decisions are made by agents who are prone to various cognitive biases
when evaluating future uncertainties. Assuming textbook rational expectations from eco-
nomic agents overlooks an important behavioral response to weather fluctuations. Second,
our study shows that the impact of climate change is caused not only by shifts in the mean,
but also by shifts in climate variability. This is in similar essence with Schlenker (2006),
who postulated economic losses stemming from the same ex ante decision problem facing
weather variability, such that even under the correct set of climate expectations, ex ante
crop choice decisions will still be sub-optimal when evaluated ex post because of imperfect
foresight. The economic losses we document here are parallel to that in Schlenker (2006)’s,
as in our case, the losses come not from the weather uncertainty itself, but from the learn-
ing process that agents use to infer weather expectations. As the magnitude of weather
13
fluctuations becomes larger, so do shifts in the subjective expectations over future climate,
and thus the resulting economic losses from short-run behavioral responses. Third, our
study provides an example of the economic value of information regarding future uncer-
tainty. Reliable forecasts of future climate and water availability, which may in the form
of long-run climate normals, spring snowpack forecasts, monsoon or El Niño patterns, as
well as extension programs that focus on the proper use of these sets of information, reduce
biases when predicting future weather and water realizations. As a result, access to these
types of information can alleviate the economic losses induced by cognitive biases from
the expectation formation process.
Our study has the potential to provide decision-support to agricultural producers and
policymakers. Our results will be directly applicable to places under similar institutional
and climatic settings, for example, the Yakima River Basin, the Central Valley of Cali-
fornia, or the Murray-Darling Basin. More broadly, our insights on producers’ behavio-
ral response to weather fluctuation can apply to places around the world where heat and
water stresses threaten agricultural production, especially in places where irrigation serves
as a crucial channel for agriculture to adapt to climate change. Going beyond agriculture,
our study can also be used as a foundation to consider other types of ex ante adaptation
decisions that require forward-looking inference regarding climate or other types of uncer-
tainty. For example, one could consider the effect of a recent heatwave on the adoption
of irrigation, farmland value, or air-conditioner purchases. All of these behaviors require
agents to form expectations over future climate, which are subject to the influence of recent
shocks. In a broader sense, this study fits into the larger literature on learning and decision-
making under uncertainty. Although we study a particular set of uncertainties arising from
stochastic climatic and hydrological processes, our study could shed light on other types
of uncertainty, for example, flood risk (Atreya et al. 2013), financial markets (Malmendier
and Nagel 2011), and energy markets (Fell and Kaffine 2018).
Our study is limited in the sense that we only show behavioral changes due to weather
fluctuation, but we do not estimate the associated economic losses. Evaluating the latter
question requires field level crop-specific yield data, which we do not have at this time.
Future research may quantify these welfare effects through the use of new remote sens-
ing methods to generate data on field-level crop yields (Donaldson and Storeygard 2016;
Chance et al. 2017).
Acknowledgements The authors gratefully acknowledge support from the NASA Land Cover/Land Use
Change (LCLUC) Program (Award NNX14AH15G), the Virginia Tech Institute for Critical Technology and
Applied Science, and the National Science Foundation’s Dynamics of Coupled Natural and Human Systems
(CNH) program (Award 1517823). All errors are our own.
Appendix: Additional Figures and Tables
See Fig. 5 and Tables 6, 7, 8, 9 and 10.
13
Fig. 5 Schematic illustration of relative positions of Snake River and tributary reaches in the Upper Snake
River Basin (from Olenichak 2015)
13
Table 6 Sample water accouting record in WD1
110
Reach Actual date Actual reach flow Reach gain Total natural Reach diversion Natural flow Remaining Last right priority
reach flow diversion natural flow
13
Snake River
S1 May 15 1690 685 685 0 0 685 1891-12-14
S2 May 15 3810 2120 2805 0 0 2805 1891-12-14
Greys River
G1 May 15 488 488 488 0 0 488 1891-12-14
Salt River
L1 May 15 483 483 483 0 0 483 1891-12-14
Snake River
53 May 16 9380 773 4549 0 0 4549 1891-12-14
S4 May 16 10,200 842 5391 22 20 5371 1891-12-14
S5 May 16 5999 0 5391 4201 3340 2031 1891-12-14
56 May 16 5570 85 5476 514 430 1686 1891-12-14
Henrys Fork
H1 May 13 40 10 10 0 0 10 1891-12-14
H2 May 14 1410 512 522 0 0 522 1891-12-14
H3 May 15 2220 818 1340 8 8 1332 1891-12-14
H4 May 15 2216 0 1340 4 0 1332 1891-12-14
Falls River
F1 May 15 20 10 10 0 0 10 1891-12-14
F2 May 15 902 882 892 0 0 892 1891-12-14
F3 May 15 474 69 961 497 490 471 1891-12-14
Henrys Fork
H5 May 15 1560 − 100 2201 1030 550 1153 1891-12-14
H6 May 15 901 0 2201 659 490 663 1891-12-14
Teton River
T1 May 15 347 354 354 7 7 347 1885-06-01

T2 May 15 0 163 517 510 510 0 1885-06-01

T3 May 15 0 34 551 34 34 0 1885-10-17
Henrys Fork
H7 May 16 1480 579 3331 0 0 1242 1891-12-14
Snake River
S7 May 16 7010 − 40 8767 0 0 2888 1891-12-14
S8 May 16 6180 581 9348 1411 1400 2069 1891-12-14
S9 May 16 5970 0 9348 210 203 1866 1891-12-14
Willow Cr
W1 May 16 31 41 41 10 10 31 1883-04-01
W2 May 16 64 − 4 37 0 0 27 1883-04-01
W3 May 16 28 − 16 21 20 11 0 1883-04-01
W4 May 16 28 0 21 0 0 0 1883-04-01
Snake River
S10 May 17 5640 86 9455 444 330 1622 1891-12-14
S11 May 17 2920 − 332 9123 2388 1173 117 1891-12-14
S12 May 17 2788 0 9123 132 117 0 1891-12-14
S13 May 18 2795 7 9130 0 0 7 1900-10-11
Portneuf R.
P1 May 18 55 55 55 0 0 55 1900-10-11
Snake River
S14 May 19 10,868 2440 11,625 188 0 2502 1900-10-11
S15 May 19 8850 -78 11,547 1838 10 2414 1900-10-11
111
13
112
13
S16 May 20 0 149 11,696 9100 2563 0 1900-10-11
Table shows the curtailment record for May 20th 2015 (Time at Milner, ID). The water right accounting system displays the ”actual date” for each reach of when the block of
water arriving at Milner on May 20th passed through the reach and how it was distributed (diverted) according to the reach priority date occurring on the actual date displayed
for each reach (Olenichak 2015)
Table 7 Parameter estimates with ex ante information

Variables (1) (2) (3)
Planted acreage Crop allocation Total expected profit
Growing degree days

lag1(GDD) 0.000188 − 0.0315 0.0346
(0.000167) (0.0770) (0.0786)
lag2(GDD) 0.000344*** 0.0458 0.165***
(0.000118) (0.0537) (0.0335)
lag3(GDD) 0.000559*** − 0.0803 0.0575
(0.000153) (0.0652) (0.0593)
lag4(GDD) − 0.000100 0.0992 0.0738
(0.000140) (0.0842) (0.0666)
lag5(GDD) 0.000134 − 0.0821 − 0.0262
(0.000223) (0.111) (0.105)
Extreme degree days
lag1(EDD) − 0.000835*** − 0.169 − 0.367*
(0.000186) (0.205) (0.182)
lag2(EDD) − 0.000643** − 0.234 − 0.411**
(0.000240) (0.169) (0.160)
lag3(EDD) − 0.00154*** − 0.0719 − 0.401**
(0.000531) (0.235) (0.151)
lag4(EDD) 0.000973** 0.135 0.220
(0.000393) (0.173) (0.214)
lag5(EDD) − 0.000248 − 0.287* − 0.360***
(0.000423) (0.149) (0.112)
lag1(prec) − 5.82e−05 − 0.117 − 0.0690
(0.000147) (0.192) (0.148)
lag2(prec) − 4.28e−05 0.379*** 0.246**
(0.000237) (0.124) (0.0950)
lag3(prec) 0.000262 0.0867 0.145*
(0.000194) (0.100) (0.0828)
lag4(prec) − 5.34e−05 0.137 0.124
(0.000274) (0.204) (0.179)
lag5(prec) − 6.90e−06 0.158 0.125
(0.000307) (0.105) (0.122)
Water curtailment
lag1(Curtailment) 0.000103 − 0.231** − 0.151*
(0.000174) (0.0897) (0.0886)
lag2(Curtailment) − 0.000439** 0.463*** 0.288**
(0.000190) (0.147) (0.115)
lag3(Curtailment) − 0.000156 − 0.323*** − 0.268**
(0.000187) (0.109) (0.107)
lag4(Curtailment) − 0.000222 0.388*** 0.234**
(0.000204) (0.116) (0.0925)
13
Variables (1) (2) (3)
Planted acreage Crop allocation Total expected profit
lag5(Curtailment) 0.000194 − 0.101 − 0.0598

(0.000174) (0.118) (0.0801)
Ex ante information
winter prec 0.000610*** − 0.0406 0.185
(0.000220) (0.209) (0.163)
SWSI reservoir − 0.00473 − 1.483 − 3.378
(0.00698) (3.064) (2.132)
Columns (1)–(3) present models with different dependent variables: dependent variables are planted acre-
age, crop allocation, total expected profit, respectively; All models mimic the main setup, except that they
are augmented with the spring reservoir levels index (SWSI Reservoir). Individual and time effects are
included in all models, and suppressed from reporting; Robust standard errors in parenthesis, clustered at
the level of water providers
Table 8 Im-Pesaran-Shin test for Variables Test statistic

unit roots in panel data
gdd − 21.38***
edd − 29.18***
prec − 37.28***
TND.season.100% − 44.47***
Streak.season.50% − 44.67***
TND.summer.50% − 44.28***
Streak.summer.0% − 44.28***
Null hypothesis is that unit roots exist in all time series. Alternative
hypothesis is that at least one time series are stationary
A triple asterisk indicates p < 0.001 ; a double asterisk indicates
p < 0.01 ; a single asterisk indicates p < 0.05
13
Table 9 Parameter estimates with different aggregation methods for water curtailment variables
Water curtailment TND.summer.50% TND.season.100% Streak.summer.50% Streak.season.100%
Variable (1) (2) (3) (4)
Panel A
Dependent variable: planted acreage
lag1(Curtailment) 7.42e−05 − 2.40e−07 9.37e−05 0.0000458
(0.000175) (0.000201) (0.000227) (0.000268)
lag2(Curtailment) − 0.000486** − 0.000277 − 0.000386 − 0.000274
(0.000180) (0.000222) (0.000271) (0.000169)
lag3(Curtailment) − 0.000128 − 0.000268** − 0.000129 − 0.000196
(0.000185) (0.000116) (0.000195) (0.000166)
lag4(Curtailment) − 0.000178 − 0.000432** − 0.000264 − 0.000474**
(0.000181) (0.000188) (0.000248) (0.000196)
lag5(Curtailment) 0.000192 − 0.000236* 0.000225 − 0.000315**
(0.000165) (0.000119) (0.000219) (0.000132)
Panel B
Dependent variable: crop allocation
lag1(Curtailment) − 0.266*** − 0.195** − 0.211** − 0.130
(0.0939) (0.0819) (0.0967) (0.103)
lag2(Curtailment) 0.432*** 0.258 0.491*** 0.248
(0.147) (0.159) (0.146) (0.149)
lag3(Curtailment) − 0.282** − 0.0711 − 0.378*** − 0.103
(0.107) (0.129) (0.136) (0.102)
lag4(Curtailment) 0.368*** 0.318** 0.460** 0.303*
(0.0982) (0.146) (0.172) (0.169)
lag5(Curtailment) − 0.120 − 0.0663 0.0185 0.00272
(0.118) (0.115) (0.109) (0.124)
Panel C
Dependent variable: total expected profit
lag1(Curtailment) − 0.185* − 0.166* − 0.142 − 0.0713
(0.0981) (0.0839) (0.101) (0.0990)
lag2(Curtailment) 0.246* 0.132 0.318** 0.110
(0.123) (0.102) (0.119) (0.104)
lag3(Curtailment) − 0.228** − 0.109 − 0.297** − 0.107
(0.103) (0.106) (0.118) (0.0739)
lag4(Curtailment) 0.245*** 0.137* 0.280* 0.0974
(0.0786) (0.0805) (0.139) (0.112)
lag5(Curtailment) − 0.0714 − 0.134* 0.0124 − 0.0924
(0.0770) (0.0715) (0.0835) (0.0799)
Columns (1)–(4) present models using different aggregation method for all five lags of water curtailment as
explanatory variables: Column (1) uses total number of days during summer months where at least 50% of
the water are curtailed, which is identical to the main specification in Table 2; Column (2) uses total number
of days during the growing season where 100% of the water are curtailed; Column (3) use days of the longest
streak of curtailment during summer months where at least 50% of the water are curtailed; and Column (4)
uses days of the longest streak of curtailment during the growing season where 100% of the water are curtailed;
Panel A presents models where the dependent variable is planted acreage. Panel B presents models where the
dependent variable is expected profit from crop allocation. Panel C presents models where the dependent vari-
able is total expected profit; All models include five lags of GDD, EDD, prec, winter precipitation of the current
year, as well as individual and time fixed-effects in addition to the curtailment variables. These variables are
suppressed from reporting; Robust standard errors in parenthesis, clustered at the level of water providers
13
116
13
Table 10 Parameter estimates with extended lags
(1) (2) (3) (4) (5)
Planted acreage Crop allocation Total expected profit Water-intensive crops Drought-tolerant crops
lag6(GDD) − 5.00e−05 − 0.0457 − 0.0768 − 8.07e−05 0.000116

(0.000203) (0.0617) (0.0599) (0.000129) (0.000171)
lag6(EDD) 0.000115 0.205 0.202 0.000161 − 0.000660
(0.000550) (0.279) (0.335) (0.000566) (0.000718)
lag6(prec) 0.000373 − 0.0820 0.00302 − 0.000448 − 0.000346
(0.000250) (0.128) (0.110) (0.000275) (0.000301)
lag6(Curtailment) − 6.50e−05 0.0260 − 0.0352 3.09e−05 − 2.57e−05
(0.000184) (0.134) (0.0846) (0.000259) (0.000316)
Columns (1)–(5) present models with different dependent variables: dependent variables are planted acreage, crop allocation, total expected profit, water-intensive crops, and
drought-tolerant crops, respectively; All models mimic the main setup, except that they are augmented with one additional lag (lag 6) of weather variables. Five lags of GDD,
EDD, prec, winter precipitation of the current year, as well as individual and time effects are included, and suppressed from reporting; Robust standard errors in parenthesis,
clustered at the level of water providers
References
Anwar S, Loughran TA (2011) Testing a Bayesian learning theory of deterrence among serious juvenile
offenders. Criminology 49(3):667–698
Arellano M, Bond S (1991) Some tests of specification for panel data: Monte carlo evidence and an applica-
tion to employment equations. Rev Econ Stud 58(2):277–297
Atreya A, Ferreira S, Kriesel W (2013) Forgetting the flood? An analysis of the flood risk discount over
time. Land Econ 89(4):577–596
Auffhammer M, Schlenker W (2014) Empirical studies on agricultural impacts and adaptation. Energy Econ
46:555–561. https://doi.org/10.1016/j.eneco.2014.09.010
Baltagi Badi H (2013) Econometric analysis of panel data. Wiley
Barberis N, Shleifer A, Vishny R (1998) A model of investor sentiment. J Financ Econ 49(3):307–343
Blanc E, Schlenker W (2017) The use of panel models in assessments of climate impacts on agriculture.
Rev Environ Econ Policy 11(2):258–279. https://doi.org/10.1093/reep/rex016
Brent DW (2017) The value of heterogeneous property rights and the costs of water volatility. Am J Agric
Econ 99(1):72–102. https://doi.org/10.1093/ajae/aaw057
Browne OR (2017) Do secure property rights affect resource allocation and firm production? Evidence
from water right adjudications in Idaho. In: 2017 Annual meeting, July 30–August 1, 2017, Chicago,
Illinois, Agricultural and Applied Economics Association, Agricultural and Applied Economics
Association
Buck S, Auffhammer M, Sunding D (2014) Land markets and the value of water: hedonic analysis using
repeat sales of farmland. Am J Agric Econ 96(4):953–969. https://doi.org/10.1093/ajae/aau013
Burke M, Emerick K (2016) Adaptation to climate change: evidence from us agriculture. Am Econ J Econ
Policy 8(3):106–40. https://doi.org/10.1257/pol.20130025
Camerer C, Ho HT (1999) Experience-weighted attraction learning in normal form games. Econometrica
67(4):827–874
Camerer CF, Loewenstein G (2011) Behavioral economics: past, present, future. In: Camerer CF, Loewen-
stein G, Rabin M (eds) Advances in behavioral economics. Chap. 1. Princeton University Press,
Princeton
Cameron TA (2005) Updating subjective risks in the presence of conflicting information: an application to
climate change. J Risk Uncertain 30(1):63–97
Chance EW, Cobourn KM, Thomas VA, Dawson BC, Flores AN (2017) Identifying irrigated areas in the
snake river plain, Idaho: evaluating performance across compositing algorithms, spectral indices, and
sensors. Remote Sens 9(6):546. https://doi.org/10.3390/rs9060546
Charness G, Levin D (2005) When optimal choices feel wrong: a laboratory study of bayesian updating,
complexity, and affect. Am Econ Rev 95(4):1300–1309
Chiang Y-M, Hirshleifer D, Qian Y, Sherman AE (2011) Do investors learn from experience? Evidence
from frequent ipo investors. Rev Financ Stud 24(5):1560–1589
Cobourn KM, Ji X, Mooney S, Crescenti NF (2017) The effect of prior appropriation and water right portfo-
lios on agricultural land-allocation decisions. Working paper
Cohn AS, VanWey LK, Spera SA, Mustard JF (2016) Cropping frequency and area response to climate vari-
ability can exceed yield response. Nat Clim Change 6(6):601
Cui X (2020) Climate change and adaptation in agriculture: evidence from US cropping patterns. J Environ
Econ Manag 101:102306
Dell M, Jones BF, Olken BA (2012) Temperature shocks and economic growth: evidence from the last half
century. Am Econ J Macroecon 4(3):66–95
Deryugina T (2013) How do people update? The effects of local weather fluctuations on beliefs about global
warming. Clim Change 118(2):397–416
Deryugina T, Kawano L, Levitt S (2018) The economic impact of hurricane katrina on its victims: evidence
from individual tax returns. Am Econ J Appl Econ 10(2):202–33
Deschênes O, Greenstone M (2007) The economic impacts of climate change: evidence from agricultural output
and random fluctuations in weather. Am Econ Rev 97(1):354–385. https://doi.org/10.1257/aer.97.1.354
Deschênes O, Greenstone M (2011) Climate change, mortality, and adaptation: evidence from annual fluctua-
tions in weather in the us. Am Econ J Appl Econ 3(4):152–185
Ding Y, Schoengold K, Tadesse T (2009) The impact of weather extremes on agricultural production methods:
does drought increase adoption of conservation tillage practices? J Agric Resour Econ 395–411
Donaldson D, Storeygard A (2016) The view from above: applications of satellite data in economics. J Econ
Perspect 30(4):171–198. https://doi.org/10.1257/jep.30.4.171
Ebenstein A, Lavy V, Roth S (2016) The long-run economic consequences of high-stakes examinations: evi-
dence from transitory variation in pollution. Am Econ J Appl Econ 8(4):36–65
13
FAO (2020) Land and water: crop information. http://www.fao.org/land-water/databases-and-software/crop-

information/en/. Accessed 29 June 2020
Fell H, Kaffine DT (2018) The fall of coal: joint impacts of fuel prices and renewables on generation and emis-
sions. Am Econ J Econ Policy 10(2):90–116
Fezzi C, Bateman I (2015) The impact of climate change on agriculture: nonlinear effects and aggrega-
tion bias in Ricardian models of farmland values. J Assoc Environ Resour Econ 2(1):57–92. https://doi.
org/10.1086/680257
Gallagher J (2014) Learning about an infrequent event: evidence from flood insurance take-up in the united
states. Am Econ J Appl Econ 6(3):206–233
Hagerty N (2020) The scope for climate adaptation: evidence from water scarcity in irrigated agriculture. Work-
ing paper
Hastie T, Tibshirani R, Friedman J (2009) The elements of statistical learning: data mining, inference, and pre-
diction. Springer
Hendricks NP (2018) Potential benefits from innovations to reduce heat and water stress in agriculture. J Assoc
Environ Resour Econ 5(3):545–576. https://doi.org/10.1086/697305
Hsiang SM, Burke M, Miguel E (2013) Quantifying the influence of climate on human conflict. Science
341(6151):1235367
Im KS, Pesaran MH, Shin Y (2003) Testing for unit roots in heterogeneous panels. J Econom 115(1):53–74
Intergovernmental Panel on Climate Change (2014) Climate change 2013: the physical science basis: Working
Group I contribution to the Fifth assessment report of the Intergovernmental Panel on Climate Change
Iyigun M, Nunn N, Qian N (2017) Winter is coming: the long-run effects of climate change on conflict, 1400–
1900, Working paper, NBER working paper no. 23033
Ji X, Cobourn KM (2018) The economic benefits of irrigation districts under prior appropriation doctrine: an
econometric analysis of agricultural land-allocation decisions. Can J Agric Econ 66(3):441–467. https://
doi.org/10.1111/cjag.12165
Ji X, Cobourn KM, Weng W (2018) The effect of climate change on irrigated agriculture: water–temperature
interactions and adaptation in the Western U.S., working paper
Kahneman D, Knetsch JL, Thaler RH (1991) Anomalies: the endowment effect, loss aversion, and status quo
bias. J Econ Perspect 5(1):193–206
Kala N (2017) Learning, adaptation, and climate uncertainty: evidence from Indian agriculture. Working paper
Kala N, Kurukulasuriya P, Mendelsohn R (2012) The impact of climate change on agro-ecological zones: evi-
dence from Africa. Environ Dev Econ 17(6):663–687
King BA, Stark JC, Wall RW (2006) Comparison of site-specific and conventional uniform irrigation manage-
ment for potatoes. Appl Eng Agric 22(5):677–688
Konisky DM, Hughes L, Kaylor CH (2016) Extreme weather events and climate change concern. Clim Change
134(4):533–547
Lee G-E, Rollins K, Singletary L (2017) An empirical analysis of water allocation efficiency through the prior
appropriation doctrine: a case study in the Carson River Valley, Neveda. Selected paper prepared for pres-
entation at the 2017 agricultural and applied economics association annual meeting, Illinois, Chicago
Lee G-E, Loveridge S, Winkler JA (2018) The influence of an extreme warm spell on public support for govern-
ment involvement in climate change adaptation. Ann Am Assoc Geogr 108(3):718–738
Leonard B, Libecap GD (2016) Collective action by contract: prior appropriation and the development of irriga-
tion in the Western United States. Working paper, National Bureau of Economic Research Working Paper
No. 22185
Liu Y, Wu Q, Ge G, Han G, Jia Y (2018) Influence of drought stress on afalfa yields and nutritional composi-
tion. BMC Plant Biol 18(1):13
Lobell DB, Hammer GL, McLean G, Messina C, Roberts MJ, Schlenker W (2013) The critical role of extreme
heat for maize production in the united states. Nat Clim Change 3(5):497–501. https://doi.org/10.1038/
nclimate1832
Malmendier U, Nagel S (2011) Depression babies: do macroeconomic experiences affect risk taking? Q J Econ
126(1):373–416
Manning DT, Goemans C, Maas A (2017) Producer responses to surface water availability and implications for
climate change adaptation. Land Econ 93(4):631–653
Marx SM, Weber EU, Orlove BS, Leiserowitz A, Krantz DH, Roncoli C, Phillips J (2007) Communication
and mental processes: experiential and analytic processing of uncertain climate information. Glob Environ
Change 17(1):47–58
Mendelsohn R, Nordhaus WD, Shaw D (1994) The impact of global warming on agriculture: a Ricardian analy-
sis. Am Econ Rev 84(4):753–771
Miao R, Khanna M, Huang H (2015) Responsiveness of crop yield and acreage to prices and climate. Am J
Agric Econ 98(1):191–211
13
Miller BM (2016) Does validity fall from the sky? Observant farmers and the endogeneity of rainfall. In: AGU
fall meeting abstracts PA11D-03
Moore MR, Negri DH (1992) A multicrop production model of irrigated agriculture, applied to water allocation
policy of the bureau of reclamation. J Agric Resour Econ 17(1):29–43
Mukherjee M, Schwabe K (2015) Irrigated agricultural adaptation to water and climate variability: the eco-
nomic value of a water portfolio. Am J Agric Econ 97(3):809–832
National Agricultural Statistics Service (2007–2016) USDA national agricultural statistics service cropland data
layer. Published crop-specific data layer. https://nassgeodata.gmu.edu/CropScape/
Obidiegwu JE, Bryan GJ, Jones HG, Prashar A (2015) Coping with drought: stress and adaptive responses in
potato and perspectives for improvement. Front Plant Sci 6:542
Olenichak T (2015) Concepts, practices and procedures used to distribute water within water district no. 1.
Water District No. 1 of Idaho. http://www.waterdistrict1.com/water%20accounting%20manual.pdf
Rabin M (2002) Inference by believers in the law of small numbers. Q J Econ 117(3):775–816
Roth AE, Erev I (1995) Learning in extensive-form games: experimental data and simple dynamic models in
the intermediate term. Games Econ Behav 8(1):164–212
Schlenker W (2006) Inter-annual weather variation and crop yields. Working paper
Schlenker W, Roberts MJ (2009) Nonlinear temperature effects indicate severe damages to us crop yields under
climate change. Proc Natl Acad Sci 106(37):15594–15598. https://doi.org/10.1073/pnas.0906865106
Schlenker W, Hanemann WM, Fisher AC (2005) Will US agriculture really benefit from global warming?
Accounting for irrigation in the hedonic approach. Am Econ Rev 95(1):395–406
Schlenker W, Hanemann WM, Fisher AC (2007) Water availability, degree days, and the potential impact of
climate change on irrigated agriculture in California. Clim Change 81(1):19–38. https://doi.org/10.1007/
s10584-005-9008-z
Seo SN, Mendelsohn R (2008) An analysis of crop choice: adapting to climate change in South American
farms. Ecol Econ 67(1):109–116
Shewmaker GE, Allen RG, Neibling WH (2011) Alfalfa irrigation and drought. University of Idaho, College of
Agricultural and Life Sciences, Moscow
Tversky A, Kahneman D (1974) Judgment under uncertainty: heuristics and biases. Science
185(4157):1124–1131
Yonts CD, Palm KL, Reichert DL (2003) Late season irrigation management for optimum sugarbeet produc-
tion. J Sugar Beet Res 40(1/2):11–28
Zhang P, Cheng H, Edwards RL, Chen F, Wang Y, Yang X, Liu J, Tan M, Wang X, Liu J et al (2008) A test of
climate, sun, and culture relationships from an 1810-year chinese cave record. Science 322(5903):940–942
Zheng J, Xiao L, Fang X, Hao Z, Ge Q, Li B (2014) How climate change impacted the collapse of the ming
dynasty. Clim Change 127(2):169–182
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
13

Weather Fluctuations, Expectation Formation, and Short-Run Behavioral Responses To Climate Change - Ji, 2021

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Weather Fluctuations, Expectation Formation, and Short-Run Behavioral Responses To Climate Change - Ji, 2021

Uploaded by

Copyright:

Available Formats

Environmental and Resource Economics (2021) 78:77–119

Weather Fluctuations, Expectation Formation, and Short‑Run

Accepted: 5 November 2020 / Published online: 23 November 2020

Keywords Weather fluctuation · Climate change adaptation · Expectation formation ·

JEL Classification Q54 · Q15 · Q25

In this section, we use a stylized framework to motivate learning mechanisms by which

where 𝛽̃s takes the following form:

Hypothesis 1 (Long-run Normal) Given a lengthy time series of previous realizations,

Figure 1 shows a stylized illustration of learning mechanisms mentioned above, including

Fig. 1 Heuristics and weights on past shocks

Table 1 Summary statistics of variables

2.1 Surface Water Availability

We also construct a database of water rights portfolios using administrative records

2.3 Acreage and Crop Allocation

using a cutoff temperature of 32 ◦C.14 We calculate growing season precipitation as the

4.1 Estimates for Ex Ante Production Decisions

4.2 Expectation Formation Mechanisms

4.3.1 Crop Share Models

economically sizeable: a one-standard-deviation change (20.3 curtailment days) in water

Growing degree days

(0.000291) (0.143) (0.141)

Variables Planted acreage Crop allocation Total expected profit

lag2(Curtailment) − 0.000486** 0.432*** 0.246*

Table 3 Parameter estimates for placebo tests

GDD − 0.000504** − 0.000636*** 0.0597 0.111 − 0.0848 − 0.107

4.3.3 Additional Ex Ante Information

4.3.4 Stationarity and Predictability of Weather Shocks

Table 4 Dynamic panel estimates

Lagged dependent variables

lag2(Curtailment) 0.432*** 0.377*** 0.380***

4.3.5 Whether Crop Rotation Drives Cyclical Coefficients

Table 5 Predictive model of Year R2 within sample R2 out of sample

Training samples are constructed using observations within 15 years

4.3.6 Choice of Aggregation Methods for the Water Curtailment Variable

TND.Season.100% and Streak.season.100%, have negative and statistically significant

Appendix: Additional Figures and Tables

See Fig. 5 and Tables 6, 7, 8, 9 and 10.

T1 May 15 347 354 354 7 7 347 1885-06-01

T2 May 15 0 163 517 510 510 0 1885-06-01

Table 7 Parameter estimates with ex ante information

Growing degree days

lag5(Curtailment) 0.000194 − 0.101 − 0.0598

Table 8 Im-Pesaran-Shin test for Variables Test statistic

lag6(GDD) − 5.00e−05 − 0.0457 − 0.0768 − 8.07e−05 0.000116

FAO (2020) Land and water: crop information. http://www.fao.org/land-water​/datab​ases-and-softw​are/crop-

You might also like

Fig. 1 Heuristics and weights on past shocks

Table 1 Summary statistics of variables

2.1 Surface Water Availability

2.3 Acreage and Crop Allocation

4.1 Estimates for Ex Ante Production Decisions

4.2 Expectation Formation Mechanisms

4.3.1 Crop Share Models

lag2(Curtailment) − 0.000486 0.432* 0.246*

Table 3 Parameter estimates for placebo tests

GDD − 0.000504 − 0.000636* 0.0597 0.111 − 0.0848 − 0.107

4.3.3 Additional Ex Ante Information

4.3.4 Stationarity and Predictability of Weather Shocks

Table 4 Dynamic panel estimates

lag2(Curtailment) 0.432* 0.377* 0.380***

4.3.5 Whether Crop Rotation Drives Cyclical Coefficients

Table 5 Predictive model of Year R2 within sample R2 out of sample

4.3.6 Choice of Aggregation Methods for the Water Curtailment Variable

Table 7 Parameter estimates with ex ante information

Table 8 Im-Pesaran-Shin test for Variables Test statistic

FAO (2020) Land and water: crop information. http://www.fao.org/land-water/databases-and-software/crop-