P ROBABILITY D ISTRIBUTIONS OF C ROP Y IELDS :

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

A B AYESIAN S PATIAL Q UANTILE
R EGRESSION A PPROACH
A. F ORD R AMSEY

Probability distributions of crop yields are important for understanding technological change, the
effects of weather on crop production, and production risk. It can be difficult to model these distribu-
tions because they are time-varying and do not follow a particular parametric form. To overcome
some of the empirical challenges inherent in yield modeling, we implement a Bayesian spatial quan-
tile regression model for the conditional distribution of yields. The statistical model is semiparamet-
ric, borrows information across space and quantile level, and models the complete quantile process.
We use the model in two empirical applications where flexible modeling of the yield distribution is
essential. First, we evaluate the effects of weather across quantiles and conduct a Bayesian test of the
difference in the rate of technological change at opposite ends of the yield distribution. We then de-
rive crop insurance premium rates and compare the predictive performance of the Bayesian spatial
quantile regression to other existing approaches for modeling time-varying yield distributions.

Key words: Bayesian spatial quantile regression, crop insurance, crop yields, probability distribu-
tions, technological change, weather.
JEL codes: C11, Q10, Q16, Q54.

Quantities and prices are the fundamental yield distribution as it relates to technological
units of economic analysis. In agriculture, change and crop modification (Chavas and
crop quantities are usually reported as crop Shi 2015; Tolhurst and Ker 2015; Lusk, Tack,
yield, that is, quantity per unit of land. and Hendricks 2017), production technology
Observed yields are random variables drawn and agricultural profitability, (Griliches 1957;
from a probability distribution. Randomness Dixon 1980), the effects of climate and
in yields results from weather over the grow- weather (Roberts, Schlenker, and Eyer 2013;
ing period, invasions of pests, interactions be- Tack and Ubilava 2015; Tack, Coble, and
tween weather and pests, soil quality, and a Barnett 2018), and crop insurance programs
number of other factors. The distribution of (Ker and Goodwin 2000; Zhu, Goodwin, and
yields also changes with the introduction of Ghosh 2011; Sherrick et al. 2014). However,
new technologies and management practices. limited data on yields in any one location
Researchers pursue inference on the prob- complicates empirical modeling of the yield
ability distribution of yields as a methodolog- distribution. Yields result from a biological
ical problem or the first step in economic process of plant growth and are usually mea-
analysis. Previous work has examined the sured only once a year. In addition, yield dis-
tributions do not follow a specific parametric
form, vary over time with technological
change, and often exhibit skewness and other
A. Ford Ramsey is an assistant professor in the Department of
Agricultural and Applied Economics at Virginia Tech. The au- non-normal features. Flexibly modeling a
thor wishes to thank the editor, three anonymous referees, and time-varying distribution with potentially
seminar participants at Meiji University, Peking University, many moments changing across time is a fun-
Beijing Institute of Technology, and the 30th Triennial
Conference of the International Association of Agricultural damentally difficult problem.
Economists for their helpful comments and suggestions. This To overcome these empirical difficulties,
work was supported by the USDA National Institute of Food
and Agriculture, Hatch project 1016069. Correspondence may be
we model the probability distribution of crop
sent to: aframsey@vt.edu. yields using Bayesian spatial quantile

Amer. J. Agr. Econ. 0(0): 1–20; doi: 10.1093/ajae/aaz029

C The Author(s) 2019. Published by Oxford University Press on behalf of the Agricultural and Applied Economics
V
Association. All rights reserved. For permissions, please email: journals.permissions@oup.com
2 June 2019 Amer. J. Agr. Econ.

regression. We apply the model in two empir- normal distribution in Botts and Boles
ical situations where the economic analysis (1958), the beta distribution in Nelson and

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

depends crucially on having a flexible, but Preckel (1989), and the gamma distribution
easily interpreted, statistical model. In the in Gallagher (1987). Goodwin and Ker (1998)
first application, we examine how the effects and Ker and Goodwin (2000) used nonpara-
of weather vary across quantiles of the yield metric techniques. As a result of short sam-
distribution. We also conduct a Bayesian test ples, available yield data may not provide
of the hypothesis that technological change conclusive results with respect to the choice
primarily impacts higher quantiles of the dis- of an optimal distribution. Short samples are
tribution. This test is conditional on weather, particularly problematic for nonparametric
for example, the quantile regressions control estimators with slow convergence rates. Ker
for weather covariates. The second applica- and Coble (2003) proposed a semiparametric
tion involves rating area-yield crop insurance method as a compromise between parametric
policies. We derive the time-conditional efficiency and nonparametric flexibility.
probability distribution of yields from the Outside of the two-stage methodology for
Bayesian spatial quantile regression model. estimation of the yield density, several
We then implement a rating game to com- authors have examined the use of distribu-
pare our proposed approach with classical tions with embedded trend functions. Zhu,
quantile regression and the two-component Goodwin, and Ghosh (2011) embedded trend
mixture model of Tolhurst and Ker (2015). functions within three and four parameter
The Bayesian spatial quantile regression is beta distributions. Tolhurst and Ker (2015)
a comprehensive approach that provides a used a mixture of two normal distributions
powerful modeling solution for measuring a with components corresponding to worst case
time-varying distribution. The model identi- and average growing conditions. These mod-
fies changes in the distribution whether they els naturally allow the distribution to change
manifest in the mean, variance, or higher with time but may require ad hoc constraints
moments. No parametric assumptions are in estimation. The interpretation of two nor-
made on the distribution itself. The model mal components as representing catastrophic
also provides results for the entire quantile and average conditions breaks down when
process. The implication is that, instead of the component means cross or when the in-
having point estimates to describe relation- tercept terms do not have proper ordering.
ships between variables, a smooth function Interpretation of the parameters of highly pa-
characterizes the relationship across all possi- rameterized beta distributions can be chal-
ble quantiles. The model prohibits quantile lenging when measures of central tendency
crossing because such behavior would invali- and dispersion are functionals of these
date the resulting probability distribution. parameters. Time-trending distributions have
Lastly, the model borrows information across much in common with quantile regression in
quantile levels and locations. This is a useful terms of their flexibility, but unlike the con-
feature because limited yield data is available straints for a valid quantile process, the con-
in most locations and even less data is avail- straints for parameters of the distributions do
able for estimation at extreme quantiles. not always come from statistical theory.
A robust body of literature has dealt with A somewhat distinct line of research has
yield distributions conditional on time. From dealt with the effects of weather, soil type,
an empirical perspective, estimation of yield and other covariates on crop yields. Much of
distributions is usually divided into two com- the motivation for these studies has come
ponents: detrending of yields and estimation from increased scientific understanding of cli-
of the conditional yield density. Models for mate change and the impacts that changing
deterministic trends have included linear temperatures and extreme weather could
functions, polynomials, and splines. have on American agriculture. Schlenker and
Normalized yields are constructed using Roberts (2006) considered the nonlinear
residuals from the deterministic trend and effects of weather in determining yields;
treated as observed yields for the purposes of Roberts, Schlenker, and Eyer (2013) ex-
density estimation. It is common in two-stage tended their results to additional covariates.
applications to correct yields for heteroske- Tack and Ubilava (2013) and Tack and
dasticity; this normalization is detailed in Ubilava (2015) examined the ways that spe-
Harri et al. (2011). Parametric examples of cific weather phenomena—the El Ni~ no
conditional yield distributions include the Southern Oscillation—can affect yields and
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 3

drive losses in the federal crop insurance pro- quantile process implied by the regressions is
gram. Antle (2010) used partial-moment coherent.

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

functions as a means of capturing asymme- Sanglestsawai, Rejesus, and Yorobe (2014)
tries in the effects of inputs on crop produc- used instrumental variables quantile regres-
tion. His results, along with those of Tack, sions to measure the impact of Bt corn on
Harri, and Coble (2012), highlight the neces- crop yields in the Philippines. The method is
sity of accounting for the effects of weather similar to classical quantile regression in that
and other covariates on higher moments of it only provides point estimates of the coeffi-
the yield distribution. Higher moments are cients. These authors found differences in the
inputs for the economic study of decision- impact of Bt technology over different quan-
making under uncertainty. tiles of the yield distribution. As far as tech-
Recognizing that biological constraints re- nological change is the sum of the effects of
sult in scarce yield data, several authors have individual technologies and changes in man-
explicitly modeled correlation of yields across agement practices, we might then expect to
space and time. Ozaki et al. (2008) used a find differential impacts across time when ag-
Bayesian conditional autoregressive model to gregated across individual technologies.
capture spatial dependence across regions of Sanglestsawai et al. (2017) built on the work
Brazil. Ker, Tolhurst, and Liu (2016) devel- of Sanglestsawai, Rejesus, and Yorobe (2014)
oped a method for combining similar densi- by showing Bt corn to have yield-increasing,
ties using Bayesian model averaging and variance-increasing, and skewness-increasing
applied the model to a number of crops and effects in a stochastic production function
counties in the United States. These authors framework.
found that private insurers could extract rents Similar to Sanglestsawai, Rejesus, and
from government insurance programs using Yorobe (2014), Chavas and Shi (2015) con-
averages of independent county-level models. sidered the impact of genetically modified
Park, Brorsen, and Harri (2019) compared seed technology on crop yields. Measuring
Bayesian model averaging to Bayesian krig- the impact using conditional quantile regres-
ing. Bayesian model averaging does not use sion, these authors found that the benefits of
distance information to smooth insurance genetic modification varied across both space
rates across space and contrasts with and management practices. A novel aspect of
Bayesian kriging where spatial smoothing is their study was the use of plot-level experi-
explicit. This makes it possible to construct mental data from Wisconsin, which allowed
densities in locations without yield data. Tack for a careful examination of production prac-
and Holt (2016) showed that spatial correla- tices. These authors noted that changes in
tion in corn yields tends to be stronger during management practices have had large effects
cases of extreme weather; this phenomenon on productivity. Chavas and Shi (2015) sug-
is otherwise known as state dependence. gested quantile regression as a promising
More fundamentally, spatial correlation method for analyzing agricultural risk,
affects the relevance of the central limit theo- weather, and technological change.
rem to aggregate yields as well as the degree Using a quantile autoregression model,
of systemic risk across insured units in crop Chavas et al. (2019) modeled the impacts of
insurance programs (Ker and Goodwin weather on crop yield distributions in Italy.
2000). Their results indicated that the effects of hot
Several recent studies have examined and dry weather are larger in the lower tail of
aspects of crop yields using frequentist quan- the yield distribution; the effects are much
tile regression. Barnwhal and Kotani (2013) smaller in the upper tail of the distribution.
focused on the heterogeneous impacts of These types of asymmetries in the effects of
weather across the yield distribution using a temperature are similar to those noted by
relatively short (34 years) time series of Barnwhal and Kotani (2013). We observe
yields. Developing countries have limited similar asymmetries in the empirical applica-
yield data and this exacerbates the effects of tion to American crop yields that follows.
data scarcity. Providing support for the Lastly, the appendix to Tolhurst and Ker
quantile regression approach, Barnwhal and (2015) presented some comparisons of the
Kotani (2013) found significant effects of two-component normal mixture versus fre-
weather and trend variables across quantiles. quentist quantile regression. The estimated
Regressions across quantiles were independent; quantiles were set equal to the mixture
in such cases, there is no guarantee that the weights for the upper and lower components
4 June 2019 Amer. J. Agr. Econ.

of the mixture. There was a relatively close one form or another; a common assumption
agreement between the two methods. is that the parametric form of the distribution

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

However, no comparison was made at any is the same across all locations.
other quantiles. While the two models may Fully nonparametric approaches are an alter-
be similar at certain quantiles, this does not native to parametric methods. Nonparametric
imply that they are similar at all quantiles. In methods can capture features of a distribution
other words, the conditional distribution of that parametric analyses would hide. Although
yields is unlikely to be similar. nonparametric density estimators are
In this article, we propose a Bayesian spa- completely flexible, they converge to the true
tial quantile regression approach for model- distribution at a slow rate (Li and Racine
ing the yield distribution. We measure the 2007). While nonparametric strategies are
effects of weather across quantiles, test largely suitable for estimating a stationary yield
whether technological change differs across density, we are interested in modeling a time-
quantiles, and rate crop insurance policies. varying distribution.
With respect to the last application, by bor- Given the basic methodological problem
rowing information across space and quantile (modeling a time-varying distribution of un-
level, this model may provide improved rat- known form), some models using nonpara-
ing in the lower tail of the yield distribution. metric Bayes are available. Most of these
The model is more comprehensive in the models are classified as density regression.
sense that it deals with several aspects of the Examples include Griffin and Steel (2006)
yield distribution simultaneously. For in- and Dunson, Pillai, and Park (2007).
stance, the models of Tolhurst and Ker Dirichlet processes are the building blocks of
(2015) and Zhu, Goodwin, and Ghosh (2011) such models; the resulting distribution is an
do not include other covariates beyond the infinite mixture. While density regression is
time trend. The model of Park, Brorsen, and attractive for its flexibility, allowing the data
Harri (2019), while providing an attractive to speak for itself, it is difficult to interpret
approach for spatial smoothing, assumes a the effects of covariates. The incorporation of
generalized Pareto conditional distribution.1 covariates is important for studies involving
We build on these earlier studies by incorpo- crop yields because the effects of technology
rating additional features into the economet-
and weather are easily confounded (Lusk,
ric model while retaining model flexibility.
Tack, and Hendricks 2017). There can also be
important interactions among weather varia-
bles (Roberts, Schlenker, and Eyer 2013), not
Bayesian Spatial Quantile Regression to mention spatial dependence between
yields at different locations.
The time-conditional distribution of crop As a compromise between parametrics and
yields is usually skewed or characterized by nonparametrics, Koenker and Bassett (1978)
non-normal features: an empirical observa- proposed quantile regression. Quantile re-
tion with significant evidence (Ker and gression has the advantage of linearity in
Goodwin 2000; Harri et al. 2009; Antle 2010; parameters providing easily interpreted co-
Tack, Harri, and Coble 2012). These aspects variate effects with a flexible conditional dis-
of the distribution may persist even after ac- tribution. Suppose there is a simple random
counting for weather. Based on this evidence, sample of size N and that we wish to model
many studies use alternative parametric dis- the response yi as a function of K covariates
tributions to model the conditional distribu- Xi . The standard quantile regression model is
tion of yields. These distributions can
incorporate aspects of non-normality such as ð1Þ qðsjXi Þ ¼ Xi 0 bðsÞ
skewness or excess kurtosis. However, para-
metric models can still involve restrictions of where qðÞ is the quantile function defined as
Pðyi < qðsjXi ÞÞ ¼ s 2 ½0; 1. The coefficients
b vary with the quantile s resulting in differ-
ent covariate effects at different quantiles of
1
There is no theoretical reason why the method in Park, the conditional distribution. Quantile regres-
Brorsen, and Harri (2018) cannot be applied to other distribu- sion thus provides a distribution-free ap-
tions. The problem is computational; it is computationally de-
manding to smooth multiple parameters. In any event, a proach to the estimation of the conditional
parametric assumption is still required. density.
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 5

We can obtain an estimate of equation 1 single quantile curve, is that information is

for several quantiles by maximum likelihood shared across quantile levels. With proper

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

techniques. The problem is one of optimiza- constraints, the quantile regressions cannot
tion with the regression quantile given by the cross. The constraint embeds information
solution to about the regressions at other quantiles.
" Formulating such a model is clearly more
X complicated than classical quantile regression
ð2Þ min sjyi  Xi 0 bj because it requires specification of parameter
b2RK 0
ijyi Xi b
# effects over a theoretically infinite number of
X
þ ð1  sÞjyi  Xi 0 bj : quantile levels. To construct the process, let
ijyi < Xi 0 b M be a number of basis functions with Bm ðsÞ
a basis function of s. Then the coefficients of
Classical quantile regression has since been the quantile process can be modeled as
applied in many economic studies where the
conditional distribution of the response is of X
M

primary interest (Buchinsky 1994; Jalan and ð4Þ bj ðs; sÞ ¼ Bm ðsÞajm ðsÞ
m¼1
Ravallion 2001; Sanglestsawai, Rejesus, and
Yorobe 2014). The quantile curves are usu-
ally estimated independently so estimates at where j ¼ 1; . . . ; K and the ajm ðsÞ are basis
extreme quantiles do not borrow information coefficients that determine the shape of the
from estimates at quantiles near the mass of quantile process. This semiparametric ap-
the distribution. Because of their indepen- proach was first proposed by Reich, Fuentes,
dent estimation, quantile curves can cross, and Dunson (2011), who used quantile re-
thus rendering the distribution of the depen- gression to study trends in summer ozone lev-
dent variable invalid. Several authors have els. Several extensions of the model have
proposed methods for the simultaneous esti- since been implemented. These include
mation of quantile curves, and empirical Reich (2012), who accounted for residual cor-
applications indicate that more precise coeffi- relation in predictors, Reich and Smith
cient estimates result (Bondell, Reich, and (2013), who considered censored data, and
Wang 2010; Tokdar and Kadane 2012). Smith et al. (2015), who developed a hierar-
Bayesian methods for quantile regression chical variant of the model.
extend to a number of domains including lim- To obtain estimates of the quantile process,
ited independent variables, instrumental vari- it is necessary to identify a suitable basis that
ables, spatial data, and regularization (Hallin, permits the monotonicity constraint on the
Lu, and Yu 2009; Lancaster and Jun 2010; Li, quantile process. Selection can be made
Xi, and Lin 2010; Benoit and Poel 2012). among a number of different basis functions,
Bayesian models are able to incorporate both including parametric quantile functions
temporal and spatial information into estima- (Reich 2012), bernstein polynomials (Reich,
tion of the quantiles while providing benefits Fuentes, and Dunson 2011), or splines (Smith
in terms of inference and the construction of et al. 2015). The R package BSquare imple-
conditional densities. Consider an extension ments non-spatial versions of Bayesian simul-
of the hypothetical random sample of equa- taneous quantile regression (Reich 2013). On
tion (1) where the response yi is observed at selecting the form of the basis, the analyst
space/time location ðs; tÞi . The time and must specify the number of basis functions. If
location of the ith observation are given by ti Bernstein polynomials are used, then
and si. The model of equation (1) can be !
modified as M
ð5Þ Bm ðsÞ ¼ sm ð1  sÞMm
m
ð3Þ qðsjXi ; si Þ ¼ Xi 0 bðs; si Þ
which is the Bernstein basis polynomial.
where the effects of the covariates now vary Equation (4) is specifically a Bernstein polyno-
by both quantile level and location. A valid mial of degree M. Bernstein polynomials have
quantile process will be non-decreasing in the proven to be popular for statistical and economic
quantile for all values of the predictors. The problems where shape constraints must be im-
advantage of a model and estimation proce- posed (Ryu and Slottje 1996; Chak, Madras, and
dure for the quantile process, as opposed to a Smith 2005; Wang and Ghosh 2014). As the
6 June 2019 Amer. J. Agr. Econ.

degree of the Bernstein polynomial goes to parameters and can include interactions and
infinity, it converges uniformly to any arbi- other transformations of variables. The spa-

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

trary continuous function on the interval [0, tial Gaussian process implies a distribution at
1]. all points in space; it is possible to generate a
The relevant constraint is that the process distribution for a location without data.
in equation (4) must be monotonic increasing Although this spatial dependence means that
in s. Using the Bernstein polynomial basis, estimates are somewhat more conservative
Reich, Fuentes, and Dunson (2011) show that than if estimated for each location indepen-
this reduces to cjm ¼ ajm  ajm1  0 for dently of all others, they are likely to be more
m ¼ 2; . . . ; M. The constraints are ensured via accurate given the relatively short samples
a latent
P unconstrained variable, where ajm ðsÞ routinely dealt with in yield modeling.
¼ M l¼1 jl ðsÞ and
c
8
>
> X
K
< c ðsÞ c ðsÞ þ
jm 1m Iðcjm ðsÞ < 0Þcjm ðsÞ  0 Technological Change and the Effects of
ð6Þ cjm ðsÞ¼
> j¼2 Weather on Crop Yields
>
:
0 otherwise
To determine if the effects of weather vary
Specifying the cjm ðsÞ as a spatial Gaussian across quantiles of the yield distribution, and
process results in smoothing across spatial to conduct a Bayesian test of the hypothesis
locations. As explained in Reich, Fuentes, that technological change primarily affects
and Dunson (2011), the processes have yields at higher quantiles, we apply the ap-
spatial covariance Covðcjm ðsÞ; cjm ðs0 ÞÞ ¼ r2j proximate Bayesian spatial quantile regres-
expðjjs  s0 jj=qj Þ and mean cjm ðKÞ. The spa- sion model of Reich, Fuentes, and Dunson
tial covariance depends on r2j , which is the (2011) to all-practice corn and soybean yields
variance of rjm ðsÞ and qj, which is a parame- from Illinois.2 The National Agricultural
ter controlling the range of spatial Statistics Service provided data from 1950 to
correlation; K is the hyperparameter set for 2015. In terms of acres of corn harvested,
the centering distribution. Illinois is often the second-largest producer
Two approaches are available for estima- in the United States, and typically the largest
tion of the model. The full model is estimable producer of soybeans.
using MCMC for small and moderate sam- We mapped the yield in each county to
ples. Reich, Fuentes, and Dunson (2011) de- the centroid of the corresponding areal unit.
veloped an approximate method that first fits We retained counties with a complete yield
separate quantile regressions over a number history and available weather data; this
of quantiles. The Bayesian spatial model then resulted in 70 spatial locations (out of 102
smooths over these initial estimates. In terms counties in Illinois) for corn and soybeans.
of mean squared error, coverage, and power, In terms of space-time observations, the
the approximate and full Bayesian spatial datasets for corn and soybeans both contain
models performed better than classical quan- 4,620 observations. Figure 1a plots the tem-
tile regression (Reich, Fuentes, and Dunson poral evolution of soybean yields by year,
2011). The MCMC algorithms used in estima- while figure 1b plots the unconditional den-
tion can be found in appendices to Reich, sity of yields.
Fuentes, and Dunson (2011) and Reich We use the weather variables of
(2012); we do not restate them here. Schlenker and Roberts (2009); similar vari-
The Bayesian spatial quantile regression ables are found in Tack, Harri, and Coble
model ultimately provides a means of captur- (2012), Tolhurst and Ker (2015), and
ing changes in a conditional distribution
while allowing a high degree of flexibility in
2
terms of covariate effects. Therefore, the Irrigation status is an omitted variable in the analysis. This
causes a problem when we wish to draw inference on the effects
model lies somewhere between density re- of weather. For instance, if irrigation increases yields and is nega-
gression (a fully flexible nonparametric ap- tively correlated with precipitation, we might expect the coeffi-
cient on precipitation to be biased downward. Because we are
proach) and parametric methods. Covariates dealing with aggregate yields, we could think of the percentage
affect higher moments and more generally of irrigated acres as being captured by the time trend. However,
the shape of the distribution. These effects acreage of irrigated corn and soybeans in Illinois is relatively
low. To avoid bias from the omitted variable, we drop any coun-
vary across locations and across quantile lev- ties where the amount of irrigated harvested acreage is greater
els. The quantile regressions are linear in the than 10% as reported in the 2012 Agricultural Census.
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 7

(a) (b)

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

(c) (d)

Figure 1. Illinois all-practice soybean yields

Note: (a) Illinois soybean yields over time, (b) unconditional distribution of Illinois soybean yields, (c) mean soybean yield over sample, and (d) minimum
soybean yield over sample

Rejesus et al. (2015). These authors mea- marginal impact of additional precipitation
sure temperature over the growing season or an ameliorating effect of precipitation on
(April to September) as the aggregate num- high temperatures.
ber of degree days in three bins: 0 to 9 Changes in technology and weather are
centigrade, 10 to 29 centigrade, and temporal phenomena. Identifying the effect
greater than 29 centigrade. Precipitation is of technological change requires controlling
measured cumulatively over the same pe- for weather and, conversely, identifying the
riod. The weather variables are at the effect of weather requires controlling for
county level and provide spatial variation technological change. Both Lobell,
necessary to identify weather impacts on Schlenker, and Costa-Roberts (2011) and
county-level yields. The binning of degree Tack, Coble, and Barnett (2018) made similar
days admits a piecewise linear relationship observations. Considering yields over short
between temperature and the response. periods of time where the sample of weather
Interaction terms can be added to allow for may not be representative exacerbates the
additional effects, such as a decreasing problem. The manner in which the weather
8 June 2019 Amer. J. Agr. Econ.

variables enter the model is one difference function of time and weather, the quantile
between this approach and the normal mix- process is modeled. In contrast to the mo-

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

ture of Tolhurst and Ker (2015). We measure ment matching method, the conditional dis-
the rate of technological change after control- tribution is not required to have a parametric
ling for weather. Tolhurst and Ker (2015) form and linearity in quantiles does not imply
measure the rate of technological change and linearity in moments.
then view weather as affecting the probability Moreover, the model allows for differences
that an observation will be drawn from a spe- in the quantile process across counties. This
cific component of the mixture distribution.3 is in contrast to other studies where the
As noted, both weather and technological parameters of the regressions have been
change should be directly modeled if we wish allowed to vary across states (Tack, Harri,
to ascertain their individual effects on the and Coble 2012) or are intended to measure
yield distribution. the average effect of weather across all coun-
Due to the constraints necessary to pro- ties (Tack, Coble, and Barnett 2018). While
duce a valid quantile function, we map the models with varying intercept terms allow for
variables to the unit interval before estima- some flexibility, they imply that the effect of
tion. Inference on the parameters is robust to weather is the same for all counties. Because
this normalization, which is common in management practices, varieties, soil quality
Bayesian analysis, although it is typically and type, and other variables vary across
implemented to speed Markov chain Monte counties, this may not be a reasonable as-
Carlo simulations. For time, the transformed sumption. The Bayesian spatial model admits
variable we use in the estimation is given by county-varying coefficients that smooth
year minus fyear divided by nyears. The con- across space.
stant fyear is the first year in the sample We estimated the model using the approxi-
(1950) and nyears is the number of years in mate method of Reich, Fuentes, and Dunson
the time series for the crop. The temperature (2011).4 The prior for the spatial range was
variables and precipitation variable were nor- Gamma(0.5, 0.5) and the prior for the varian-
malized using the normal distribution cumu- ces was Inverse Gamma(0.1, 0.1). A skew-t
lative density function. The estimated distribution was used for the centering distri-
quantile function is then bution of the model. For the MCMC algo-
rithm to effectively search the parameter
ð7Þ qðsjX i ; si Þ ¼ b0 ðs; si Þ þ b1 ðs; si Þt space, the centering distribution should be
þ b2 ðs; si Þlowi flexible enough to match the unconditional
þ b3 ðs; si Þmedi distribution of the response. The priors for
the parameters of the skew-t centering distri-
þ b4 ðs; si Þhighi bution were location Normal(0, 100), scale
þ b5 ðs; si Þpcpn þ ei Uniform(0, 100), skew Normal(0, 20), and
degrees of freedom Uniform(30, 100). The
where low, medium, high, and pcpn refer to Bernstein polynomials were of degree ten,
the normalized degree day variables and nor- allowing for considerable flexibility in model-
malized precipitation variables defined ing the covariate effects. We took a total of
previously. 5,000 draws from the Markov chain Monte
The quantile function is indexed by loca- Carlo algorithm with 2,500 discarded as burn-
tion si. As advanced by Antle (2010) and in. The model was evaluated at all quantiles
Tack and Ubilava (2013), yield observations from 0.1 to 0.9 by 0.05, and for reasons which
come from a distribution that varies across will become clear, the octiles 0.125, 0.376,
both space and time. The Bayesian spatial 0.625, and 0.875.
quantile regression shares some similarities Results for soybeans in a single county are
with the moment matching method suggested in figure 2. Figure 2a shows actual soybean
by Antle (2010). The key difference is that in- yields in Christian County, Illinois, as well as
stead of modeling specific moments as a

4
The algorithm was implemented in R on an Intel Xeon
3
This may be because incorporating weather directly in the CPU E5-1650 at 3.60 GHz. Total time to estimate the model with
normal mixture model would require more complicated con- 5,000 draws for one crop and one state was just under 30 minutes.
straints on the components of the mixture. For instance, one We conducted sensitivity analysis with alternative priors and
would need to keep the mixture components from crossing across found that there were no changes in the substantive results of the
time and at all possible values of weather covariates. model. The MCMC algorithm appeared to converge.
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 9

(a) (b)

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

(c) (d)

Figure 2. Model results for Christian County, Illinois soybeans

Note: Weather variables are fixed at sample means to generate the fitted quantile curves. (a) Fitted quantile regressions, (b) posterior distribution of b1 (0.75)
 b1 (0.25), (c) temperature quantile processes, and (d) precipitation quantile process.

posterior mean quantile curves with all cova- effectively borrows information from the
riates except year fixed at their sample bulk of the distribution and partially deals
means. The quantile model provides a close with the problem of outliers.
fit to the bulk of the conditional distribution In Christian County, Illinois, the number of
and weather covariates have the effect of low- and medium-degree days increases
shifting the quantiles up or down. The slopes yields at all quantiles. Consistent with other
of the quantile curves at higher quantiles studies, figure 2 shows that high-degree days
tend to be larger than the slopes at lower have a severe, negative impact on crop yields.
quantiles. Slightly more than 5% of the Precipitation over the growing season has a
observations lie above the first and last ven- small, positive impact, although we cannot
tiles. As the other predictors have been fixed, make the statement with a large amount of
this does not imply an egregious model fit. credibility given the width of the credible
Were extremely high heat included, for in- intervals. While the effect of low- and me-
stance, the fitted quantile curves would shift dium-degree days is essentially constant
down and better capture catastrophically low across quantiles, the effect of high-degree
yields. The infrequency with which cata- days is most negative at the low end of the
strophic yields occur makes the estimation at yield distribution. In contrast, precipitation
low quantiles particularly difficult. By pre- has a greater effect at low quantiles. The
venting crossing quantiles, the model impacts of the weather variables are
10 June 2019 Amer. J. Agr. Econ.

(a) (b)

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

(c) (d)

Figure 3. Year and high-heating degree day effects by quantile and county
Note: Each curve represents the quantile process for a single county in the sample. (a) Year effect for soybeans, (b) high HDD effect for soybeans, (c) year ef-
fect for corn, and (d) high HDD effect for corn.

generally consistent with past literature in Figures 3b and 3d also support the earlier
terms of their signs. results on the effect of heating degree days
The plots of greatest interest, which summa- on yields in Christian County, Illinois. Across
rize differences in the time coefficient across most counties, high heating degree days have
quantiles, are figures 3a and 3c. Plotting the ef- a larger negative effect at lower quantiles of
fect of time for all counties in the sample the yield distribution. However, while most
reveals that the effect of technological change, counties have quantile processes of similar
as embodied in the coefficient associated with shape, the magnitudes vary. This suggests
the trend, is greater at higher quantiles for that researchers should allow models of
most counties in the sample. These results pro- yields to vary across counties when practical.
vide some support for the hypothesis of Lobell Flexibility is difficult to achieve in many sit-
et al. (2014) but do not constitute a statistical uations because of short time series of yields.
test. Weather has been at least partially con- Nonetheless, by bringing spatial information
trolled, so the observation of this behavior in into a model, it is possible to measure county-
spite of the weather variables lends credence specific effects.
to the idea that there are tradeoffs to be had We conduct a Bayesian test of the hypothe-
in breeding new crop varieties. The growth at sis that technological change primarily
upper quantiles is most marked for corn and impacts higher quantiles. As an example,
less evident in soybeans. consider a single county where we are
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 11

Table 1. Summary Trend Coefficients across Counties

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

Raw value Yield scale (Bu./Ac.)
Crop Mean Min. Max. Mean Min. Max.
b1 ð:50Þ Corn 118.60 109.1 128.8 1.80 1.65 1.95
Soybeans 26.84 20.14 33.82 0.40 0.31 0.51
b1 ð0:75Þ  b1 ð0:25Þ Corn 8.01 4.65 12.64 0.12 0.07 0.19
Soybeans 1.91 0.53 3.67 0.03 0.01 0.06

interested in whether the coefficient on time the heart of the hypothesis that technological
is greater at the 75th percentile compared to change is making yields more sensitive to, for
the 25th percentile. Within each Markov instance, extreme heat. The question is
chain Monte Carlo draw, we obtain the dif- whether, conditional on given weather condi-
ference between b1 ð:75Þ and b1 ð:25Þ, and con- tions, yield distributions show increased vari-
struct a posterior distribution of the ance or negative skew.
difference. This posterior distribution is We argue that the question cannot be an-
shown in figure 2b for Christian County, swered by solely examining yields over time.
Illinois. If the posterior mean is greater than If weather is not evenly distributed over the
zero, and zero is not within the given credible sample period and a once-in-100-year (or
set (two standard deviations), then we view worse) weather event occurs near the end of
the effect of technological change as being the sample, then the difference in the rate of
different across the two quantiles. change at different quantiles could be dis-
Based on the posterior distributions of the torted. The problem is the same as if there is
coefficients, we found that 44 of 70 counties correlation between weather and time, but is
for soybeans had greater technological exaggerated because we are considering ex-
change at the 75th percentile than at the 25th treme quantiles and thus giving additional
percentile of the conditional distribution. The weight to a small portion of the data. If we
test showed this behavior in 47 out of 70 wished to only regress on time, we would
counties for corn. Table 1 shows a numerical need a long enough sample to characterize
summary of the trend coefficients at the me- the distribution of weather and the distribu-
dian and the difference in the 75th and 25th tion of yields. It is difficult to argue that 50 or
percentiles. At least in Illinois, there is some even 100 years of yield and weather data is
limited evidence that the upper mass of the adequate for this purpose.
yield distribution has grown more rapidly Figure 4 shows conditional yield distribu-
than the lower tail. In this framework, it is tions over time in two counties for soybeans
very easy to test differences between other and corn, respectively. The use of quantile re-
quantiles as well. A general finding across all gression does not exclude more formal meas-
crops was that the difference in the trend ures of the skewness and kurtosis of the
coefficients was small at quantiles above the conditional distributions. While conventional
median but large when considering quantiles skewness and kurtosis are estimated by sam-
below the median. ple averages, several quantile-based alterna-
For comparison, we fit the model with only tives exist. These alternatives replace, more
a time trend to the same sample of Illinois or less, the mean with the median for location
corn and soybean yields. Conducting the and the variance with an interquartile range
same Bayesian test, we found that 70 out of for dispersion. Based on the estimated quan-
70 counties had greater growth at higher tiles, we calculated the coefficient of quantile
quantiles. Controlling for weather has major variation, the skewness coefficient of Hinkley
impacts on the results of the test and the em- (1975), and the kurtosis coefficient of Crow
pirical implications. Our conclusion, if we did and Siddiqui (1967). Kim and White (2004)
not control for weather, would be that tech- found that the skewness and kurtosis meas-
nological change has a major impact on the ures compare favorably to classical skewness
yield distribution and is significantly increas- and kurtosis coefficients, and are robust to
ing risk in all counties. On controlling for outliers.
weather, the conclusion is radically different Table 2 shows summary statistics for the
and much weaker. This difference strikes at estimated coefficient of quantile variation,
12 June 2019 Amer. J. Agr. Econ.

(a) (b)

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

(c) (d)

Figure 4. Conditional yield distributions for selected counties

Note: Weather variables are fixed at sample means to generate the conditional distributions. (a) Christian County, Illinois Soybeans, (b) Adams County,
Illinois Soybeans, (c) Christian County, Illinois Corn, and (d) Adams County, Illinois Corn.

Table 2. Quantile-based Moments across County Distributions

Moments in 2015 Counties with
decreasing moments
Measure Definition Crop Mean Std. dev. Min. Max. Number Percent
Q75 Q25
Dispersion 100  Q75 þQ25 Corn 5.58 0.88 4.01 8.07 70 100
(CQV) Soybeans 4.67 1.00 3.05 7.72 64 91
Skewness ðQ75 QQ5075ÞðQ
Q25
50 Q25 Þ
Corn 0.13 0.05 0.29 0.02 5 7
Soybeans 0.04 0.04 0.02 0.13 15 21
ðQ87:5 Q62:5 ÞþðQ37:5 Q12:5 Þ
Kurtosis Q75 Q25  1:23 Corn 0.11 0.06 0.02 0.23 30 43
Soybeans 0.05 0.11 0.16 0.27 69 99

skewness, and kurtosis, using quantile-based distribution. In all counties, the median yield
measures for the last year in the sample. The increased over time and the variance of the
kurtosis measure is normalized so that a data has increased. Mean measurements of
value of zero corresponds to a normal skewness and kurtosis across the counties
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 13

indicate negative skew and excess kurtosis covariates in the model or including a longer
for corn. The coefficient of quantile variation sample of yields so that the distribution of

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

has decreased over time in almost all coun- weather is more complete. It would also be
ties. The moment results accord with those of useful to see continued investigation of the
figure 4 and provide quantitative evidence of impacts of specific technologies alongside ag-
the graphical results. gregate studies of technological change.
We do find, however, that conditional on
weather being fixed at sample means, yield
distributions are not by-and-large becoming
more negatively skewed over time. Kurtosis Crop Insurance Ratemaking
is decreasing in most soybean counties but
only 43% of corn counties. This result may Crop insurance ratemaking makes extensive
arise from our use of a robust estimator of use of probability distributions of crop yields.
skewness and kurtosis. We also tested for de- Insurers derive prices for crop insurance poli-
creasing skewness without the weather varia- cies using a historical loss distribution or the
bles. As expected, the distributions were distribution of the random variable generat-
heavily skewed, with negative skewness in- ing the loss. For insurance that pays out on
creasing over time. The key point is that the lost yields, this random variable is the yield it-
comparison is essentially over two distribu- self. While the Bayesian spatial quantile
tions. One is conditional on time only. The regressions do not directly produce a proba-
other is conditional on time and weather. As bility distribution, estimates of the quantile
we argue above, the latter is the more rele- process can be combined with adaptive ker-
vant distribution for testing a hypothesis nel density estimation to provide a time-
about technological change conditional on conditional distribution of crop yields. In
weather. This is especially so when the sam- what follows, we compare the Bayesian spa-
ple may not be representative of the distribu- tial quantile regression model to frequentist
tion of weather. quantile regression and the two-component
These results for the higher moments of normal mixture model developed in Tolhurst
the distributions have implications for pro- and Ker (2015). In all cases, the models in-
duction risk in general. Any interpretation clude a linear trend across time. Based on the
will depend on the specific risk measure used. estimated trend coefficients (and in the case
Dispersion of the yield distribution is not in- of the normal mixture model, the mixing
creasing over time, suggesting less overall weight), we project the yield distributions for-
risk under normal weather conditions. We ward to generate the actuarial quantities used
also find that few counties have decreasing to construct premium rates.
skewness under the same weather conditions. We may suspect a priori that the Bayesian
Skewness and kurtosis have been used as spatial quantile regression will provide more
measures of downside risk. Because we gen- accurate rates when compared with non-
erally find decreasing skewness and kurtosis, spatial frequentist quantile regression. This is
technological change has also reduced down- because the model borrows information
side risk controlling for weather variables. across space and quantile level. Intuitively, if
The quantile-based measures do not consider a model prohibits quantile crossing, it must
the deep tails of the distribution and there- adjust one or both of the crossing quantile
fore do not provide ready information about curves using information on both quantile
extreme risk. curves. The Bayesian spatial quantile regres-
The Bayesian spatial quantile regression sion makes this adjustment based on informa-
allows the researcher to bring covariates into tion at all other quantiles. While the
the model for a conditional distribution. theoretical implication of crossing quantiles
While results for Illinois on the effects of is an invalid distribution, whether quantiles
weather covariates accord with the findings cross in practice is an empirical question.
of previous literature, the incorporation of Figure 5 shows a conventional estimate of the
covariates has a major impact on inferences quantile process for soybean yields in Brown
drawn about technological change. We find County, Illinois. Even if we ignore the patho-
mixed evidence that technological change logical behavior at extremely low quantiles,
makes yields more sensitive to weather for the quantile curves cross inside the range of
given weather conditions. Future work the data. Because the presence of crossing
should consider including additional weather quantiles relates to the data and the quantile
14 June 2019 Amer. J. Agr. Econ.

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

50

40
Yield (Bu./Ac.)

30

20

1960 1980 2000

Year

Figure 5. Conventional estimate of the quantile process for Brown County, Illinois soybeans

levels at which we conduct the regressions, it ð10Þ EðindemnityÞi ¼Pðyi < k^yi Þ
is not difficult to find crossing quantiles when
estimating the quantile process.  ðk^yi Eðyi jyi < k^yi ÞÞ
Federal crop insurance policies are avail-
able at farm and aggregate levels. Area yield where yi is the realized yield. The expected
insurance is based on county-level yields, and indemnity is the product of the probability of
although farm-level insurance is more popu- a loss occurring multiplied by the expected
lar, area-based policies protect against moral loss given that a loss occurs. In practice, the
hazard and adverse selection (Miranda 1991). expected indemnity under an area-yield pol-
For an area yield insurance policy, the pre- icy is scaled by a specified price. Because the
mium rate is given by price is not stochastic, and therefore not ger-
mane to this rating exercise, we do not con-
ð8Þ ratei ¼ EðindemnityÞi =liabilityi sider it here. The loss ratio for a portfolio of
policies is given by
where both the indemnity and liability de-
pend on the coverage level. The subscript i P
N
maxðk^yi  yi ; 0ÞÞ
refers to the county over which the policy is
written. With a given coverage level ð11Þ loss ratio ¼ i¼1
P
N
k 2 ð0; 1Þ, the liability under the policy is ratei k^yi
i¼1

ð9Þ liabilityi ¼ k^yi

with indexing determining the portfolio of
policies for which the loss ratio is defined
where ^yi is the mean or predicted county
(e.g., we could index over counties to obtain
yield under a given model.5 The expected in-
a loss ratio in a single year or index over
demnity is
years and obtain the loss ratio for a single
county).
We consider all-practice corn yields from
Illinois, Indiana, Iowa, Minnesota, and
5
Nebraska and all-practice soybean yields
For the purposes of the rating exercises considered in this from Illinois, Indiana, Iowa, and Nebraska.
section, ^yi is assumed to be the mean of the projected yield
distribution. The analysis includes only counties with a
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 15

Table 3. Estimated Loss Ratios

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

Crop State/Coverage 70% coverage 90% coverage
Model NM QR BSQR NM QR BSQR
Corn Illinois 2.93 1.19 1.96 1.20 1.01 1.23
Indiana 2.88 1.02 1.58 1.25 1.08 1.35
Iowa 0.22 0.09 0.45 0.60 0.49 0.81
Minnesota 0.00 0.00 0.00 0.21 0.16 0.30
Nebraska 1.57 0.76 9.25 1.33 1.08 1.80
Soybeans Illinois 0.66 0.26 3.38 0.85 0.76 1.38
Indiana 2.25 0.27 0.98 1.00 0.95 1.18
Iowa 0.87 0.43 1.52 1.02 0.98 1.43
Minnesota 0.31 0.15 0.28 0.687 0.68 0.88

complete yield series from 1950 to 2015.6 favorable loss ratios over the period; they are
Because recent changes to federal crop insur- close to one in most cases. The frequentist
ance (such as the supplemental coverage op- quantile regressions tend to produce loss ratios
tion) now allow for higher coverage levels, that are lower than those produced by the
we calculate premium rates for 70% and 90% Bayesian spatial model. We did not check for
coverage. The estimated Bayesian spatial crossing of the frequentist regressions, so it
quantile regression is could be that some of the estimated distribu-
tions are invalid. The average loss ratios of the
ð12Þ qðsjX i ; si Þ ¼ b0 ðs; si Þ þ b1 ðs; si Þt: normal mixture model are close to those esti-
mated using the Bayesian quantile regressions.
We use the same prior distributions as in However, examination of the loss ratios does
the first application and take the same not constitute a statistical test of the accuracy
amount of draws when running the MCMC of the models; it only suggests that either of the
algorithm. The Bayesian spatial quantile re- models could be a reasonable approach in
gression model is estimated in each state sep- terms of controlling outlays under the program.
arately.7 For the frequentist quantile Similar to Park, Brorsen, and Harri (2019),
regressions, we estimate the quantile process Ker, Tolhurst, and Liu (2016), and as origi-
at all possible quantiles that the data permits. nally developed by Ker and McGowan (2000),
The densities are constructed using adaptive we compare methods using a rating game. The
kernel density estimation fitted to projected rating game is an out-of-sample exercise
values from the quantile regressions. We esti- intended to assess the predictive performance
mate the normal mixture model as in Tolhurst of the methods. In the game, a private insurer
and Ker (2015). The premium rate is calcu- cedes a certain number of policies in its port-
lated in each county for the years between folio to the government. This mimics the
2001 and 2015. For example, data from 1950 Standard Reinsurance Agreement that gov-
to 2014 are used to estimate the premium rate erns the relationship between private insurers
for 2015. The estimation of the premium rate and the federal government in the federal
for 2014 uses data from 1950 to 2013. crop insurance program. The private insurer
Table 3 shows the average loss ratio across has an internal rating system that may be
all years for each crop/state/coverage/model more accurate than the rating system
combination. The rates generated from the employed by the government in pricing crop
Bayesian spatial quantile regressions have insurance policies. This leads to a decision
rule for the insurer in deciding whether to
cede or retain a policy. If the insurer’s rate is
6
For corn, the sample includes 76 counties in Illinois, 65 in higher than the government’s rate, the policy
Indiana, 95 in Iowa, 58 in Minnesota, and 71 counties in is ceded as the government is underestimating
Nebraska. For soybeans, the sample includes 85 counties in
Illinois, 64 in Indiana, 96 in Iowa, and 56 in Minnesota.
the risk inherent in the policy. Likewise, if the
7
We might be able to obtain better estimates by varying the insurer’s rate is lower than the government’s
priors by state. There might also be gains in running the model rate, then the policy is retained.
over several states and incorporating more spatial information.
However, the 48 hours of computational time required to pro- The results of a randomization test sug-
duce the estimates in this section was already high. gested in Ker, Tolhurst, and Liu (2016) are
16 June 2019 Amer. J. Agr. Econ.

Table 4. Randomization Test of BSQR vs. Normal Mixture

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

Crop Coverage State % Payouts % Insurer retains LR Gov. LR Ins. P-value
Corn 70 Illinois 6 40 8.47 0.50 0.00
Indiana 7 30 6.37 0.59 0.00
Iowa 3 59 0.28 0.20 0.46
Minnesota 0 53 0.00 0.00 –
Nebraska 5 36 8.24 1.44 0.20
90 Illinois 20 56 2.33 0.65 0.00
Indiana 22 61 2.50 0.81 0.00
Iowa 14 69 1.22 0.45 0.00
Minnesota 9 58 0.47 0.10 0.00
Nebraska 28 51 1.88 1.43 0.16
Soybeans 70 Illinois 3 56 6.11 0.29 0.01
Indiana 2 24 3.63 0.32 0.09
Iowa 4 55 3.05 0.30 0.02
Minnesota 3 38 1.30 0.02 0.00
90 Illinois 16 66 1.76 0.64 0.00
Indiana 19 62 1.88 0.73 0.00
Iowa 20 73 1.56 0.80 0.01
Minnesota 26 64 1.36 0.48 0.00

shown in table 4. For all state/crop/coverage The difference is statistically significant at the
combinations, the table gives the percentage 5% level in six out of ten states. The insurer
of policies retained by the insurer using can obtain rents at all state/coverage combina-
Bayesian spatial quantile regression com- tions for soybeans, of which seven out of eight
pared to the government using the normal are statistically significant at the 5% level.
mixture model. The case where the govern- Good performance in the lower tail does not
ment operates using frequentist quantile re- imply worse performance in the bulk of the
gression can be seen in table 5. Using the distribution; the Bayesian spatial model is
decision rule and the government rating found to be statistically significant in more
method, we calculate loss ratios for the insurer cases at the 90% coverage level.
and the government. To construct the p-val- The Bayesian spatial quantile regression
ues, we select a random sample of policies also performs well against the frequentist
with the size of the sample equal to the num- quantile regressions. Insurers extract rents in
ber of policies retained by the insurer. The loss eight out of ten state/coverage combinations
ratio is calculated from the random sample for corn. The results are statistically signifi-
and repeating this procedure 5,000 times yields cant in three out of ten state/coverage combi-
5,000 loss ratios. These loss ratios are used to nations. Similar results hold for soybeans
construct the distribution of loss ratios under with all state/coverage combinations generat-
the null hypothesis that the decision rule for ing rents and five out of eight being statisti-
the insurer has no effect. The p-value is the cally significant. The differences between the
area under this distribution of loss ratios and two forms of quantile regression are gener-
to the left of the estimated insurer loss ratio. ally smaller than those between the quantile
Compared to the normal mixture model regression and normal mixture. We expect
and the frequentist quantile regression, the this to be the case because we are using the
Bayesian spatial quantile regression performs approximate approach of Reich, Fuentes, and
much better in the lower tail of the distribu- Dunson (2011) to estimate the Bayesian spa-
tion. Against the normal mixture, insurers can tial model. This approach uses frequentist
extract economic rents in all states and at all estimates to inform the Bayesian MCMC
coverage levels for corn, with the exception of algorithm.
Minnesota, which requires 70% coverage.8 The results of the insurance pricing exer-
cise and rating game show that Bayesian spa-
tial quantile regression can be a potent model
8
Over the 15 years, there were actually zero losses at the for constructing conditional distributions.
70% coverage level in all counties in Minnesota. Therefore, we This has to do with the spatial aspects of the
cannot calculate a p-value for tables 4 and 5.
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 17

Table 5. Randomization Test of BSQR vs. Quantile Regression

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

Crop Coverage State % Payouts % Insurer retains LR Gov. LR Ins. P-value
Corn 70 Illinois 6 71 4.11 0.67 0.00
Indiana 7 67 2.14 0.88 0.06
Iowa 3 88 0.03 0.17 0.54
Minnesota 0 90 0.00 0.00 –
Nebraska 5 49 7.76 0.97 0.17
90 Illinois 20 69 1.75 0.74 0.00
Indiana 22 75 1.34 1.01 0.14
Iowa 14 84 0.74 0.49 0.11
Minnesota 9 91 0.37 0.13 0.02
Nebraska 28 34 1.50 0.91 0.13
Soybeans 70 Illinois 3 87 0.82 0.37 0.36
Indiana 2 79 0.88 0.28 0.28
Iowa 4 80 1.99 0.37 0.05
Minnesota 3 67 0.77 0.07 0.00
90 Illinois 16 79 1.49 0.68 0.01
Indiana 19 73 1.44 0.79 0.01
Iowa 19 77 0.98 0.93 0.42
Minnesota 25 70 1.18 0.53 0.00

model and the ability of the model to more distribution of yields is central to the eco-
accurately capture behavior at extreme quan- nomic analysis.
tiles. The fact that the Bayesian spatial model In the first application, we use Bayesian
offers competitive performance when com- spatial quantile regression to determine if the
pared with the normal mixture suggests that effects of weather are different across quan-
borrowing information across both space and tiles of the yield distribution, and to test the
quantile level is important. The spatial com- hypothesis that technological change occurs
ponent is not the only feature of the model faster at upper quantiles. We find that the
that results in these favorable results. But as impacts of some weather covariates vary
we have not separated the spatial modeling across quantiles. In particular, the negative
and modeling of the individual yield distribu- effects of extreme heat (as measured by heat-
tions, we cannot say which of the two facets ing degree days) are largest in the lower tail
of the Bayesian spatial model is more of the conditional yield distribution. We also
important. find significant variation in the effects of
weather across counties, implying that yield
modelers should consider spatially-varying
effects in their analyses. After controlling for
Conclusion weather in the quantile regressions, a
Bayesian test of the trend coefficients yields
Limited data in most locations, spatial corre- some evidence that technological change has
lation among yields, and the need for a statis- a greater impact at upper quantiles of the
tical model that allows for flexibility confront yield distribution. We argue that this test con-
empirical applications using a probability trolling for weather is more appropriate be-
distribution of crop yields. The principal cause it tests for differences in technological
problem is that the distribution of yields is change conditional on given weather condi-
time-varying and has an unknown form. For tions. The technology trend is much less pro-
economic analysis, where parameters of the nounced when compared with the results of
model facilitate inference, the model must Tolhurst and Ker (2015). As a result, the
also yield easily interpreted results. The ten- yield distributions are less skewed when con-
sion is between model flexibility on the one ditioned on weather.
hand and usability on the other. We suggest We then price area-yield crop insurance
Bayesian spatial quantile regression as an ap- policies using the Bayesian spatial quantile
propriate model for the yield distribution and regression. Data are limited in each county
apply it in two empirical settings where the and limited at extreme quantiles of the yield
18 June 2019 Amer. J. Agr. Econ.

distribution; borrowing information across American Journal of Agricultural

space and quantile level results in improved Economics 92 (5): 1294–309.

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

accuracy in rating. The rates and loss ratios Barnwhal, P., and K. Kotani. 2013. Climatic
derived from policies using the proposed Impacts Across Agricultural Crop Yield
Bayesian spatial model are in line with rates Distributions: An Application of
from other approaches. Following Ker and Quantile Regression on Rice Crops in
McGowan (2000) and Ker, Tolhurst, and Liu Andhra Pradesh, India. Ecological
(2016), we conduct a rating game where we Economics 87: 95–109.
assume that a private insurer operates using Benoit, D.F., and D.V.D. Poel. 2012. Binary
the Bayesian spatial quantile regression and Quantile Regression: A Bayesian
the government prices insurance using either Approach Based on the Asymmetric
the two-component normal mixture of Laplace Distribution. Journal of Applied
Tolhurst and Ker (2015) or classical quantile Econometrics 27: 1175–88.
regression. The insurer is able to extract eco- Bondell, H., B. Reich, and H. Wang. 2010.
nomically significant rents when using the Non-Crossing Quantile Regression
Bayesian spatial quantile regression as its Curve Estimation. Biometrika 97 (4):
preferred model. 825–38.
Bayesian estimation can be time consuming Botts, R.R., and J.N. Boles. 1958. Use of
when compared with frequentist quantile re- Normal-Curve Theory in Crop Insurance
gression. Bayesian models also require a host Ratemaking. Journal of Farm Economics
of prior assumptions, and researchers might 40 (3): 733–40.
view these points as drawbacks of the proposed Buchinsky, M. 1994. Changes in the U.S.
approach. In choosing priors, there is no objec- Wage Structure 1963–1987: Application
tive criteria. But as noted by Gelman (2008), of Quantile Regression. Econometrica 62
there is no objective criteria for choosing a like- (2): 405–58.
lihood. Future work could consider several Chak, P., N. Madras, and B. Smith. 2005.
additions and improvements. Additional Semi-Nonparametric Estimation with
weather variables would improve model pre- Bernstein Polynomials. Economics
dictions. Another useful development would Letters 89 (2): 153–6.
be to focus on the spatial aspects of the model Chavas, J.-P., S. Di Falco, F. Adinolfi, and F.
and how it could be used to construct distribu- Capitanio. 2019. Weather Effects and
tions in counties without yield data. We did not Their Long-Term Impact on the
compare the Bayesian model to other spatial Distribution of Agricultural Yields:
models in our rating game. Such a comparison, Evidence from Italy. European Review of
for instance, to the models of Ker, Tolhurst, Agricultural Economics 46 (1): 29–51.
and Liu (2016) or Park, Brorsen, and Harri Chavas, J.-P., and G. Shi. 2015. An Economic
(2019), would be a commendable extension. Analysis of Risk, Management, and
Empirical modeling of crop yields is likely Agricultural Technology. Journal of
to continue to grow in practical importance. Agricultural and Resource Economics 40:
As climate projections improve in accuracy 63–79.
and include additional variables, improved Crow, E., and M. Siddiqui. 1967. Robust
yield projections will follow suit. Yield distri- Estimation of Location. Journal of the
butions in crop insurance contexts should also American Statistical Association 62 (318):
see continued focus from agricultural econo- 353–89.
mists. Crop insurance remains the most impor- Dixon, R. 1980. Hybrid Corn Revisited.
tant instrument of United States agricultural Econometrica 48 (6): 1451–61.
policy under the 2018 Farm Bill, and has Dunson, D.B., N. Pillai, and J.-H. Park. 2007.
grown in size in developing countries (Mahul Bayesian Density Regression. Journal of
and Stutley 2010). In both cases, appropriate the Royal Statistical Society: Series B
statistical models for the probability distribu- (Statistical Methodology) 69 (2): 163–83.
tion of crop yields will be imperative. Gallagher, P. 1987. U.S. Soybean Yields:
Estimation and Forecasting with
Nonsymmetric Disturbances. American
References Journal of Agricultural Economics 69 (4):
796–803.
Antle, J.M. 2010. Asymmetry, Partial Gelman, A. 2008. Rejoinder. Bayesian
Moments, and Production Risk. Analysis 3 (3): 467–78.
Ramsey Probability Distributions of Crop Yields: A Bayesian Spatial Quantile Regression Approach 19

Goodwin, B.K., and A.P. Ker. 1998. Kurtosis. Finance Research Letters 1 (1):
Nonparametric Estimation of Crop Yield 56–73.

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

Distributions: Implications for Rating Koenker, R., and G. Bassett. 1978.
Group–Risk Crop Insurance Contracts. Regression Quantiles. Econometrica 46
American Journal of Agricultural (1): 33–50.
Economics 80 (1): 139–53. Lancaster, T., and S.J. Jun. 2010. Bayesian
Griffin, J.E., and M.J. Steel. 2006. Order- Quantile Regression Methods. Journal of
Based Dependent Dirichlet Processes. Applied Econometrics 25 (2): 287–307.
Journal of the American Statistical Li, Q., and J.S. Racine. 2007. Nonparametric
Association 101 (473): 179–94. Econometrics: Theory and Practice.
Griliches, Z. 1957. Hybrid Corn: An Princeton, NJ: Princeton University
Exploration in the Economics of Press.
Technological Change. Econometrica 25 Li, Q., R. Xi, and N. Lin. 2010. Bayesian
(4): 501–22. Regularized Quantile Regression.
Hallin, M., Z. Lu, and K. Yu. 2009. Local Bayesian Analysis 5 (3): 533–56.
Linear Spatial Quantile Regression. Lobell, D.B., M.J. Roberts, W. Schlenker, N.
Bernoulli 15 (3): 659–86. Braun, B.B. Little, R.M. Rejesus, and
Harri, A., K.H. Coble, A.P. Ker, and B.J. G.L. Hammer. 2014. Greater Sensitivity
Goodwin. 2011. Relaxing to Drought Accompanies Maize Yield
Heteroscedasticity Assumptions in Area- Increase in the U.S. Midwest. Science 344
Yield Crop Insurance Rating. American (6183): 516–9.
Journal of Agricultural Economics 93 (3): Lobell, D.B., W. Schlenker, and J. Costa-
707–17. Roberts. 2011. Climate Trends and
Harri, A., C. Erdem, K.H. Coble, and T.O. Global Crop Production since 1980.
Knight. 2009. Crop Yield Distributions: A Science 333 (6042): 616–20.
Reconciliation of Previous Research and Lusk, J., J. Tack, and N.P. Hendricks. 2017.
Statistical Tests for Normality. Review of Heterogeneous Yield Impacts from
Agricultural Economics 31 (1): 163–82. Adoption of Genetically Engineered
Hinkley, D. 1975. On Power Transformations Corn and the Importance of Controlling
to Symmetry. Biometrika 62 (1): 101–11. for Weather. In Understanding
Jalan, J., and M. Ravallion. 2001. Behavioral Productivity Growth in Agriculture, ed.
Responses to Risk in Rural China. W. Schlenker. Chicago, IL: University of
Journal of Development Economics 66 Chicago Press.
(1): 23–49. Mahul, O., and C.J. Stutley. 2010.
Ker, A.P., and K. Coble. 2003. Modeling Government Support to Agricultural
Conditional Yield Densities. American Insurance: Challenges and Options for
Journal of Agricultural Economics 85 (2): Developing Countries. Washington, DC:
291–304. The World Bank.
Ker, A.P., and B.K. Goodwin. 2000. Miranda, M.J. 1991. Area-Yield Crop
Nonparametric Estimation of Crop Insurance Reconsidered. American
Insurance Rates Revisited. American Journal of Agricultural Economics 73 (2):
Journal of Agricultural Economics 82 (2): 233–42.
463–78. Nelson, C.H., and P.V. Preckel. 1989. The
Ker, A.P., and P. McGowan. 2000. Weather- Conditional Beta Distribution as a
Based Adverse Selection and the U.S. Stochastic Production Function.
Crop Insurance Program: The Private American Journal of Agricultural
Insurance Company Perspective. Journal Economics 71 (2): 370–8.
of Agricultural and Resource Economics Ozaki, V.A., S.K. Ghosh, B.K. Goodwin, and
25: 386–410. R. Shirota. 2008. Spatio-Temporal
Ker, A.P., T.N. Tolhurst, and Y. Liu. 2016. Modeling of Agricultural Yield Data
Bayesian Estimation of Possibly Similar with an Application to Pricing Crop
Yield Densities: Implications for Rating Insurance Contracts. American Journal
Crop Insurance Contracts. American of Agricultural Economics 90 (4):
Journal of Agricultural Economics 98 (2): 951–61.
360–82. Park, E., B.W. Brorsen, and A. Harri. 2019.
Kim, T.-H., and H. White. 2004. On More Using Bayesian Kriging for Spatial
Robust Estimation of Skewness and Smoothing in Crop Insurance Rating.
20 June 2019 Amer. J. Agr. Econ.

American Journal of Agricultural Sherrick, B.J., C.A. Lanoue, J.D. Woodard,

Economics 101 (1): 330–51. G.D. Schnitkey, and N.D. Paulson. 2014.

Downloaded from https://academic.oup.com/ajae/advance-article-abstract/doi/10.1093/ajae/aaz029/5548582 by Nottingham Trent University user on 18 August 2019

Reich, B.J. 2012. Spatiotemporal Quantile Crop Yield Distributions: Fit, Efficiency,
Regression for Detecting Distributional and Performance. Agricultural Finance
Changes in Environmental Processes. Review 74 (3): 348–63.
Journal of the Royal Statistical Society: Smith, L.B., B.J. Reich, A.H. Herring, P.H.
Series C (Applied Statistics) 61 (4): Langlois, and M. Fuentes. 2015.
535–53. Multilevel Quantile Function Modeling
Reich, B.J., M. Fuentes, and D.B. Dunson. with Application to Birth Outcomes.
2011. Bayesian Spatial Quantile Biometrics 71 (2): 508–19.
Regression. Journal of the American Tack, J., K. Coble, and B. Barnett. 2018.
Statistical Association 106 (493): 6–20. Warming Temperatures Will Likely
Reich, B.J., and L.B. Smith. 2013. Bayesian Induce Higher Premium Rates and
Quantile Regression for Censored Data. Government Outlays for the U.S. Crop
Biometrics 69 (3): 651–60. Insurance Program. Agricultural
Reich, L.S.B. 2013. BSquare: Bayesian Economics 49 (5): 635–47.
Simultaneous Quantile Regression. R Tack, J., A. Harri, and K. Coble. 2012. More
package version 1.1. Than Mean Effects: Modeling the Effect
Rejesus, R.M., K.H. Coble, M.F. Miller, R. of Climate on the Higher Order
Boyles, B.K. Goodwin, and T.O. Knight. Moments of Crop Yields. American
2015. Accounting for Weather Journal of Agricultural Economics 94 (5):
Probabilities in Crop Insurance Rating. 1037–54.
Journal of Agricultural and Resource Tack, J., and M. Holt. 2016. The Influence of
Economics 40 (2): 306–24. Weather Extremes on the Spatial
Roberts, M.J., W. Schlenker, and J. Eyer. Correlation of Corn Yields. Climatic
2013. Agronomic Weather Measures in Change 134 (1–2): 299–309.
Econometric Models of Crop Yield with Tack, J., and D. Ubilava. 2013. The Effect of
Implications for Climate Change. El Nino Southern Oscillation on U.S.
American Journal of Agricultural Corn Production and Downside Risk.
Economics 95 (2): 236–43. Climatic Change 121 (4): 689–700.
Ryu, H.K., and D.J. Slottje. 1996. Two ———. 2015. Climate, Risk, and Insurance:
Flexible Functional Form Approaches Measuring the Effects of Enso on U.S.
for Approximating the Lorenz Curve. Crop Insurance. Agricultural Economics
Journal of Econometrics 72 (1–2): 46 (2): 245–57.
251–74. Tokdar, S., and J. Kadane. 2012.
Sanglestsawai, S., R.M. Rejesus, and J.M. Simultaneous Linear Quantile
Yorobe. 2014. Do Lower Yielding Regression: A Semiparametric Bayesian
Farmers Benefit from Bt Corn? Evidence Approach. Bayesian Analysis 7 (1):
from Instrumental Variables Quantile 51–72.
Regressions. Food Policy 44: 285–96. Tolhurst, T.N., and A.P. Ker. 2015. On
Sanglestsawai, S., D.G.P. Rodriguez, R.M. Technological Change in Crop Yields.
Rejesus, and J.M. Yorobe. 2017. American Journal of Agricultural
Production Risk, Farmer Welfare, and Bt Economics 97 (1): 137–58.
Corn in the Philippines. Agricultural and Wang, J., and S. Ghosh. 2014. Shape Restricted
Resource Economics Review 46 (3): Nonparametric Regression with Bernstein
507–28. Polynomials. Computational Statistics &
Schlenker, W., and M. Roberts. 2006. Data Analysis 56: 2729–41.
Nonlinear Effects of Weather on Corn Zhu, Y., B.K. Goodwin, and S.K. Ghosh.
Yields. Review of Agricultural 2011. Modeling Yield Risk Under
Economics 28 (3): 391–8. Technological Change: Dynamic Yield
Schlenker, W., and M.J. Roberts. 2009. Distributions and the U.S. Crop
Nonlinear Temperature Effects Indicate Insurance Program. Journal of
Severe Damages to U.S. Crop Yields un- Agricultural and Resource Economics
der Climate Change. Proceedings of the 36 (1): 192–210.
National Academy of Sciences 106 (37):
15594–8.