Review of Agricultural Economics—Volume 31, Number 2—Pages 303–329 Estimating the Value of an Early-Warning System Michael J. Roberts, David Schimmelpfennig, Michael J. Livingston, and Elizabeth Ashley An early-warning system generates economic value to the extent that it improves decision making. The value of the information hinges on the degree to which a timely response, aided by warnings, facilitates successful damage mitigation. USDA’s Coordinated Frame- work for Soybean Rust includes a network of sentinel soybean plots and wild kudzu stands monitored by extension agents for the presence of soybean rust, a potentially recur- ring threat to the U.S. soybean crop since 2005. The linchpin in this early-warning system is a website that provides near real-time, county-level information on the location of the disease. We consider factors that may influence information value. I nformation is valuable when it allows decision makers to adjust their actions to better suit the situation at hand. When information allows many individuals to improve their decision making, a governmental role in its provision may be justi- fiable, because individuals may be unable to coordinate and finance its collection and dissemination. Of course, just because information can embody the public good attributes of nonrivalry and nonexcludability does not necessarily imply a role for public policy. One must still consider social costs and benefits. While cost calculation might be relatively straightforward, valuing information can be challenging. In this paper, we consider an illustrative example provided by a recent USDA-led effort to provide real-time information for a new threat to the U.S. soybean crop, Phakopsora pachyrhizi, a fungus that causes soybean rust (SBR). SBR, a recurrent problem for soybean producers in much of the southern hemisphere, was first detected in the United States in 2004, late enough in the growing season that it posed no threat to that year’s soybean crop. After overwin- tering in southern states along the Gulf of Mexico, SBR posed a new, uncertain, Michael J. Roberts, David Schimmelpfennig, and Michael J. Livingston are with USDA Economic Research Service. Elizabeth Ashley is at the Department of Economics, University of California at Santa Barbara. Views expressed are the authors’ and not necessarily those of the USDA. DOI: 10.1111/j.1467-9353.2009.01439.x
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Review of Agricultural Economics—Volume 31, Number 2—Pages 303–329
Estimating the Value of anEarly-Warning System
Michael J. Roberts, David Schimmelpfennig,Michael J. Livingston, and Elizabeth Ashley
An early-warning system generates economic value to the extent that it improves decisionmaking. The value of the information hinges on the degree to which a timely response,aided by warnings, facilitates successful damage mitigation. USDA’s Coordinated Frame-work for Soybean Rust includes a network of sentinel soybean plots and wild kudzustands monitored by extension agents for the presence of soybean rust, a potentially recur-ring threat to the U.S. soybean crop since 2005. The linchpin in this early-warning systemis a website that provides near real-time, county-level information on the location of thedisease. We consider factors that may influence information value.
Information is valuable when it allows decision makers to adjust their actions tobetter suit the situation at hand. When information allows many individuals to
improve their decision making, a governmental role in its provision may be justi-fiable, because individuals may be unable to coordinate and finance its collectionand dissemination. Of course, just because information can embody the publicgood attributes of nonrivalry and nonexcludability does not necessarily imply arole for public policy. One must still consider social costs and benefits.
While cost calculation might be relatively straightforward, valuing informationcan be challenging. In this paper, we consider an illustrative example providedby a recent USDA-led effort to provide real-time information for a new threat tothe U.S. soybean crop, Phakopsora pachyrhizi, a fungus that causes soybean rust(SBR). SBR, a recurrent problem for soybean producers in much of the southernhemisphere, was first detected in the United States in 2004, late enough in thegrowing season that it posed no threat to that year’s soybean crop. After overwin-tering in southern states along the Gulf of Mexico, SBR posed a new, uncertain,
� Michael J. Roberts, David Schimmelpfennig, and Michael J. Livingston are with USDAEconomic Research Service.� Elizabeth Ashley is at the Department of Economics, University of California at SantaBarbara.Views expressed are the authors’ and not necessarily those of the USDA.
DOI: 10.1111/j.1467-9353.2009.01439.x
304 Review of Agricultural Economics
and potentially severe threat to the U.S. soybean crop at the beginning of the 2005growing season (Skokstad).
The USDA-led framework provides real-time SBR information via a website(http://www.sbrusa.net) that reports findings from sentinel plots where expertsregularly monitor for soybean rust. Findings are pooled together with weatherforecasts and aerobiological analyses to forecast the likely future spread of thefungus. The overarching purpose of the framework is to provide farmers withsufficient notice so they can make appropriate decisions as to the use of preventiveand curative fungicides on their soybean fields.
In the three full years since its first detection in the United States, SBR hasposed little threat to the U.S. soybean crop. Given the expense of developingthe website and its underlying infrastructure, some have questioned whether theframework was a worthwhile endeavor. After all, if farmers had simply managedtheir crops as if there were no SBR threat, they may have fared as well or bet-ter than they did in the presence of the Coordinated Framework. However, thisview overlooks a key point: although weather conditions have not yet facilitateddispersion of SBR spores to key soybean-producing regions, this could not havebeen known in advance. A potential SBR threat existed at the beginning of the2005 season, but how farmers might have prepared for that threat in the absenceof the USDA framework is not clear. Indeed, without the framework, individ-ual farmers may have incurred even greater expense by monitoring their ownfields, perhaps spraying fungicides for a threat that did not exist in their area, orforgoing planting entirely.1 More generally, this view overlooks the fundamentalnotion that information value should be assessed from an ex ante perspective.Quantifying the ex ante value involves determining the expected value of actionswith and without the benefit of information and subtracting the latter from theformer. This can be challenging because it involves determining decisions thatwould have been made without the information, and what the consequences ofthose decisions would have been. Perhaps more elusively, it also hinges on whatfarmers’ expectations would have been without the USDA framework.
We develop estimates of the ex ante value of information provided by the USDAframework from the vantage point of the beginning of the 2005 growing season(Roberts et al.). We show how various factors influence the size of this value, in-cluding the costs and efficacy of available fungicides, farmers’ prior beliefs aboutthe likelihood of infection, the perceived accuracy of the framework’s SBR fore-casts, and farmers’ risk preferences. The value may also depend on how soybeanprices would be affected by SBR-induced production shocks. Our analysis buildson standard Bayesian decision theory.
Most empirical value-of-information studies consider only the value of perfectinformation. In this study, we develop a set of assumptions that allows us toconsider the value of information over a continuum of information qualities.This is useful in gauging the value of more realistic information systems. It alsofacilitates straightforward valuation of marginal improvements in informationquality.
We find the value of information depends on many factors, but most impor-tantly farmers’ prior beliefs about SBR risk at the beginning of the growing sea-son and the accuracy of the system’s forecast. These factors cannot be quantifiedprecisely, so we consider information values over a range of assumptions aboutprior beliefs, forecast accuracy, and other factors. Even if forecasts are imprecise,
Estimating the Value of an Early-Warning System 305
resolving only 20% of SBR infection uncertainty for all fields planted with soy-beans, the system’s value in 2005 was an estimated $11 million. If forecasts re-solved 80% of infection uncertainty, the estimated value was $395 million. Ouranalysis suggests that the value of the information in 2005 likely exceeded costsof developing the information, reported to be between $2.5 and almost $5 million.
Three additional factors affect estimated information values: anticipated priceshocks in the event of a large rust outbreak, soybean farmers’ aversion to risk,and heterogeneity of farmers’ prior beliefs of an infestation. We find that all ofthese factors tend to reduce the largest estimated values and increase the smallestestimated values, but the effects are relatively small in magnitude. The potentialbenefits of the framework suggest that similar programs for other crop pests canbe cost effective if, as in the case of soybean rust, preventive action can stronglymitigate damages in the event of an outbreak.
Bayesian Updating with Information AccuracyIn this section, we review Bayesian decision theory (Hirschleifer and Riley;
Schimmelpfennig and Norton) and use it to develop a simple model to valueinformation provided by the USDA-led framework about impending SBR infec-tions. We also develop a set of simplifying assumptions to derive a scalar indexof information accuracy (Lawrence).
We begin by characterizing the problem in terms of a payoff matrix that mapsa finite set of possible farmer actions x ∈ {x1, x2, . . . , xX} and a finite set of mutu-ally exclusive states s ∈ {s1, s2, . . . , sS} to an X × S matrix of possible outcomes,the elements of which are denoted by px,s. The unconditional probability thatany given state s will occur is �s. These probabilities are subjective: they per-tain to a farmer’s beliefs about the chances that each state will occur, and weassume these beliefs are consistent with the laws of probability (all �s ≥ 0 and�s�s = 1).
An information signal is modeled as a random variable M, which might realizeoutcomes such as mL, signaling “low risk of infestation,” or mH, signaling a “highrisk of infestation.” In general, there may be any number of possible messages. Themessage is valuable if it arrives before the farmer chooses an action and causes thefarmer to change beliefs about the probability that states s will occur. The posteriorprobability, the probability of s given M (�sM), is linked to the prior probability(�s) using Bayes rule
�s, M = Pr[s | M] = �sPr[M | s]
Pr[M].(1)
The quality of the information signal depends on the magnitude of the differ-ence between �sM and �s. A perfect information signal would cause a completeresolution of uncertainty, so that if the state s∗ ultimately arises, �s∗M = 1 and�−s∗M = 0. Thus, for perfect information, the number of possible outcomes fromthe message M must equal the number of possible states S, and the distributionof M must be identical to the distribution defined by the probabilities �s. That is,Pr[M = s] = �s and Pr[s = M | M] = 1. For example, if there are two states of theworld, “infestation” and “no infestation,” then analogous messages that forecast“impending infestation” or “no impending infestation” would occur with the
306 Review of Agricultural Economics
same frequency as infestations themselves (Pr[M = s] = �s), and the messageswould always be correct (Pr[s = M | M] = 1).
The more realistic and interesting case is when information is imperfect. Ingeneral, the number of possible messages may be greater than, less than or equalto the number of possible states, and both the conditional and unconditionalprobability distributions of messages can take many forms. One way to simplifythe issue is to assume the message is a forecast that predicts which state willoccur, and that the unconditional probability distribution of messages equals theprobability distribution of states. Hence, as in the case of perfect information,Pr[M = s] = �s. We therefore define the unconditional probability density functionof M using the same notation as the states, e.g., �M. Unlike the case of perfectinformation, however, Pr[s = M | M] < 1; that is, it is anticipated that the forecastmay be inaccurate. From Bayes rule, we can see that in this special case
�s M = Pr[s | M] = Pr[M | s] = Pr[M ∩ s]�s
.(2)
By simplifying the problem in this way, we can model information quality usinga single index of message accuracy � ∈ (0, 1), where information quality tends tozero as � tends to 0 and information quality tends to perfect accuracy as � tendsto 1. We do this by setting
Pr[M ∩ s] = �s�M + �(1 − �s�M).(3)
This expression implies that M and s are independent when � = 0 (the mes-sage contains no information) and perfectly correlated (Pr[M ∩ s] = �s) when� = 1 (perfect information). The expression in (3) is simply a linear interpola-tion between these two extremes. Under this set of assumptions, the posteriordistribution is linked to the prior distribution by
�s M = �s�M + �(1 − �s�M)�s
.(4)
The value of information depends on the effect the message has on the decisionsfarmers make. Without a message, farmers maximize expected profit given theirprior beliefs, �s. We denote these actions by x∗. With information, farmers choosetheir actions to maximize expected profits conditional on the message M. Wedenote these actions by x∗ | M. Payoffs are given by the combination of the statethat occurs and action chosen and denoted pxs. Taking expectations, the ex antevalue of information given by message M is therefore
VOI =∑
M
∑
s
�s M psx∗ | M −∑
s
�s psx∗ .(5)
Estimating the Value of the SBR Monitoring FrameworkTo estimate the value of the SBR monitoring framework we must delineate farm-
ers’ possible management strategies (actions) and possible payoffs. We assume
Estimating the Value of an Early-Warning System 307
Figure 1. Decision tree without information about soybean rustinfection
Infest(prob=π)
SBR
PREVENTDECISION
NoInfest
Monitor
NoApply
Payoff(5.)
Payoff(1.)
Payoff(2.)
Apply
SBR SBR
Apply Curative
CUREDECISION
CUREDECISION
Payoff(3.)
Payoff(4.)
NoCurative
Infest(prob= π)
NoInfest
Infest(prob= π)
NoInfest
Payoff(6.)
Notes: Square boxes indicate farmers’ decisions and the circles represent nature’s random decisionwhether or not to infest. The payoffs (1–6) are described in table 1.
Table 1. Possible outcomes stemming from P. pachyrhizi threat
Management Strategy Infection No Infection
Apply preventive treatment 1 21% yield loss, cost of $25.63/acre Cost of $25.63/acre
Monitor fields and apply 3 4curative treatment if SBR 7% yield loss, cost of $20.52/acre Cost of $6.71/acre
No rust management 5 625% yield loss Base return
Source: Johansson et al.
three management strategies: (a) apply a preventive fungicide before soybean rustoccurs; (b) intensively monitor fields and apply a curative fungicide if soybeanrust is detected; or (c) do nothing. The payoffs and profit-maximizing strategiesdepend on the costs of preventive and curative fungicides, monitoring costs,expected yield losses in the event of an infection, soybean prices, and farmers’perceptions of the probability that infection will occur. The decision tree in figure 1shows how the three strategies, crossed with two possible states (infestation andno infestation), lead to six possible payoffs, which are summarized in table 1.2
The costs in this table are treatment cost, not the cost of the yield loss.
308 Review of Agricultural Economics
Figure 2. Decision tree with partial information about soybean rustinfection
Infest(prob= H)
SBR
PREVENTDECISION
NoInfest
Monitor
No Apply
Payoff(5.)
Payoff(1.)
Payoff(6.)
Payoff(2.)
Apply
SBR SBR
CUREDECISION
CUREDECISION
‘‘High Risk’’signal
(prob=π)
‘‘Low Risk’’signal
(prob=1-π)
INFO
PREVENTDECISION
Monitor
No Apply
Apply
Infest(prob=
SBR
NoInfest
Payoff(5.)
Payoff(1.)
Payoff(6.)
Payoff(2.)
SBR SBR
CUREDECISION
CUREDECISION
Apply Curative
Payoff(3.)
Payoff(4.)
NoCurative
Apply Curative
Payoff(3.)
Payoff(4.)
NoCurative
Infest(prob=
NoInfest
Infest(prob=
NoInfest
Infest(prob=
NoInfest
Infest
NoInfest
Notes: Square boxes indicate farmers’ decisions and the circles represent nature’s random decisionsabout the information signal and whether or not to infest. The payoffs (1–6) are described in table 1.
Figure 2 illustrates the decision tree for an environment with an imperfectinformation signal. In this environment, management decisions are made afterreceiving an information signal, M, which changes farmers’ beliefs about infec-tion from the prior, � to posterior �H or �L, depending on whether it is a “high-”or “low-risk” signal. The accuracy of the signal, �, ranges continuously between0 and 1: the higher is �, the more the signal changes farmers’ beliefs. If infor-mation quality were perfect, we would expect only two signals, one perfectlyforecasting an impending arrival of SBR and one perfectly forecasting the nonar-rival of SBR—that is, �H would equal 1 and �L would equal zero. To approx-imate a continuum of information qualities, we suppose there remain just twosignals, but that the signal itself may have different levels of accuracy. If neitherof the two signals contained informational content, they would not alter farmers’priors (� = �H = �L), and farmers would choose the same management strat-egy in the partial information environment as they would in the no-informationenvironment.
Because we do not have objective estimates for information accuracy, we eval-uate farmers’ optimal conditional strategies and expected profits over a range ofaccuracies: � = 0.2 (low), � = 0.5 (medium), and � = 0.8 (high). One may think ofthese information qualities as the proportion of uncertainty resolved by the partialinformation. We then calculate farmers’ overall expected profits by multiplyingconditional expected profits by the probability of each signal and summing them,as per equation (5).
Estimating the Value of an Early-Warning System 309
Quantifying Payoffs and Probabilities
Soybean Yield ImpactsFungicide efficacy trials from Brazil and Paraguay in 2001–2003, aggregate
yield data for 10 Brazilian states during 1993–2002, and data on the introduc-tion of P. pachyrhizi into those states were used to estimate rust-free yields andtreated and untreated yield impacts (Livingston et al.). Rust-free yields aver-aged 2.604 (±0.422) metric tons per hectare, and treated and untreated yieldsaveraged 2.578 (±0.201) and 2.025 (±0.363) metric tons per hectare. Estimatedtreated and untreated yields were therefore lower by an average 4.3% (±5.2%) and25.0% (±11.9%), respectively, than estimated rust-free yields.
We use the untreated yield impacts to estimate payoffs when rust occurs butno fungicide is applied. Because the treated yield impacts were estimated withyield data reported from soybean plots sprayed with curative, preventive, orcurative plus preventive fungicides, we must separate impacts of the differentkinds of treatments. Replicating the methods in Johansson et al., we find that theaverage yield impact for the preventive class of fungicides is −0.97%. The averageyield impact for the curative class of fungicides is −6.95% with a mean of 1.39applications evaluated (see table 2 and Livingston et al.).3
Prior Infection ProbabilitiesWe develop regional proxies for prior probabilities of SBR infestations using
data on wheat stem rust, a disease that spreads through the air much like SBR.Stem rust epidemics of wheat for 1921–62 (Hamilton and Stakman) are used toestimate how often P. pachyrhizi spores may be present in most states where soy-beans are produced (U.S. Department of Agriculture, 2005). We also use dataon daily temperature extremes, rainfall, and humidity for 1992–2001 to estimatethe proportion of years in which conditions may favor the development of soy-bean rust in each state (Livingston et al.). P. pachyrhizi may be able to overwinteralong the coastlines of Alabama, Florida, Georgia, Louisiana, Mississippi, andTexas (Pivonia and Yang); therefore, we assume that climatic conditions will fa-vor introduction of soybean rust in all years for these states. In addition, becauseP. pachyrhizi is an obligate parasite that can not live without a host plant, we usedata on the most likely soybean planting and harvesting dates for each state (U.S.Department of Agriculture, 1997) to estimate how often climatic conditions andhost availability may favor rust epidemics.
To estimate state-level prior probabilities for rust infections, we use the productof the proportion of years that stem rust epidemics actually occurred and the pro-portion of years that climates favored the spread of rust. This assumes climaticconditions affecting the dispersal of spores and those affecting the broader estab-lishment of rust in an area are independent (Hamilton and Stakman). To convertstate-level prior probabilities to regional probabilities, we weight states by aver-age 1995–2004 soybean production (U.S. Department of Agriculture, 1998–2005).Over all U.S. soybean acres, these calculations imply an average prior probabilityof rust infection equal to 0.53. Across regions, the priors are 0.67, 0.55, 0.55, 0.49,0.62, 0.43, 0.76, and 0.51 for Appalachia, Corn Belt, Delta, Lake States, North-east, Northern Plains, Southeast, and the Southern Plains, respectively.4 These
310 Review of Agricultural Economics
Table 2. Yields with preventive and curative fungicides
Source: (a) Bayer (2003a) (Trials 1 and 2). Lower bound of rust-free yield estimate for Mato Grassodo Sul 2001–02. (b) Bayer (2003b) (Trial 14). The estimate for rust-free yield in Minas Gerais 2002–03.(c) Bayer (2003b) (Trial 15). The estimate for rust-free yield in Minas Gerais 2002–03. (d) Bayer(2003b) (Trial 16). The estimate for rust-free yield in Minas Gerais 2002–03. (e) BASF (2003) (Jesus,Paraguay). The upper bound of the rust-free yield estimate for Parana 2002–03. (f) BASF (2003)(Pirapo, Paraguay). The estimate for rust-free yield in Mato Grasso do Sul 2002–03.Note: Blank fields indicate no data: Each study considers either preventive or curative fungicidetreatments.
probabilities are used to estimate information values in the base case (table 3) andother scenarios.
Our rather high estimated priors may appear inconsistent with the actual U.S.SBR experience after 2005 since SBR has not yet posed a serious threat to theoverall U.S. soybean crop. It is possible, however, that our estimated infestation
Estimating the Value of an Early-Warning System 311
Tab
le3.
Bas
eca
sein
form
atio
nva
lues
Exp
ecte
dV
alu
eV
alu
eP
rior
Yie
ldH
igh
Hig
hL
owL
owE
Vof
Info
ofIn
foB
elie
fof
No
Info
Wit
hou
tIn
form
atio
nR
isk
Ris
kR
isk
Ris
kW
ith
Per
Per
Infe
ctio
nD
ecis
ion
SB
RQ
ual
ity
(�)
Dec
isio
nE
VD
ecis
ion
EV
Info
Farm
Acr
eR
egio
n(P
rob
abil
ity)
(P,M
,or
N)
(Bu
shel
s/A
cre)
(Sca
le0
to1)
(P,M
,or
N)
(Dol
lars
)(P
,M,o
rN
)(D
olla
rs)
(Dol
lars
)(D
olla
rs)
(Dol
lars
)
App
alac
hia
0.67
M35
.80
0.2
P77
,530
M83
,015
79,3
5871
40.
640.
67M
35.8
00.
5P
77,2
82M
89,5
7281
,379
2,73
42.
450.
67M
35.8
00.
8P
77,0
34N
99,7
4384
,604
5,95
95.
33C
orn
Bel
t0.
55M
44.6
00.
2P
91,7
76M
96,3
9093
,873
166
0.22
0.55
M44
.60
0.5
P91
,497
M10
0,41
395
,549
1,84
22.
480.
55M
44.6
00.
8P
91,2
17N
106,
510
98,1
694,
461
6.01
Del
ta0.
55M
31.8
00.
2M
93,0
57M
103,
850
97,9
630
00.
55M
31.8
00.
5P
87,4
15N
114,
271
99,6
231,
659
0.85
0.55
M31
.80
0.8
P86
,890
N13
0,02
210
6,49
68,
533
4.36
Lak
eSt
ates
0.49
M41
.50
0.2
M56
,831
M60
,227
58,5
730
00.
49M
41.5
00.
5P
55,7
20M
62,7
0859
,304
731
1.37
0.49
M41
.50
0.8
P55
,509
N67
,085
61,4
462,
873
5.38
Nor
thea
st0.
62M
38.7
00.
2P
41,2
76M
43,8
8942
,281
172
0.36
0.62
M38
.70
0.5
P41
,145
M46
,559
43,2
281,
119
2.37
0.62
M38
.70
0.8
P41
,015
N50
,697
44,7
392,
630
5.56
Con
tinu
ed
312 Review of Agricultural Economics
Tab
le3.
Con
tin
ued
Exp
ecte
dV
alu
eV
alu
eP
rior
Yie
ldH
igh
Hig
hL
owL
owE
Vof
Info
ofIn
foB
elie
fof
No
Info
Wit
hou
tIn
form
atio
nR
isk
Ris
kR
isk
Ris
kW
ith
Per
Per
Infe
ctio
nD
ecis
ion
SB
RQ
ual
ity
(�)
Dec
isio
nE
VD
ecis
ion
EV
Info
Farm
Acr
eR
egio
n(P
rob
abil
ity)
(P,M
,or
N)
(Bu
shel
s/A
cre)
(Sca
le0
to1)
(P,M
,or
N)
(Dol
lars
)(P
,M,o
rN
)(D
olla
rs)
(Dol
lars
)(D
olla
rs)
(Dol
lars
)
Nor
ther
n0.
43M
36.3
00.
2M
67,7
17M
72,9
1670
,688
00
Plai
ns0.
43M
36.3
00.
5P
63,7
67N
77,1
4171
,409
721
0.82
0.43
M36
.30
0.8
P63
,428
N83
,496
74,8
954,
207
4.78
Sout
heas
t0.
76M
25.2
00.
2M
1,88
7M
4,07
82,
408
00
0.76
M25
.20
0.5
P1,
765
N7,
146
3,04
663
81.
440.
76M
25.2
00.
8P
1,71
5N
11,0
953,
949
1,54
03.
48So
uthe
rn0.
51M
26.0
00.
2M
21,2
44N
29,6
4725
,343
371
0.24
Plai
ns0.
51M
26.0
00.
5M
15,6
52N
39,0
6927
,075
2,10
21.
380.
51M
26.0
00.
8P
13,4
99N
48,4
9130
,568
5,59
63.
67O
ther
0.53
M33
.90
0.2
M66
,216
M72
,469
69,1
590
00.
53M
33.9
00.
5P
63,1
94N
77,8
3770
,085
926
0.84
0.53
M33
.90
0.8
P62
,869
N86
,977
74,2
145,
056
4.61
Not
es:I
nd
ecis
ion
colu
mns
(b),
(e),
and
(g),
Mis
mon
itor
/cu
re,P
ispr
even
t,an
dN
isd
ono
thin
g.In
(f),
(h),
and
(j),E
Vis
expe
cted
valu
e.Z
ero
valu
esof
info
rmat
ion
are
due
toro
und
ing
ofd
iscr
ete
dat
a.
Estimating the Value of an Early-Warning System 313
rates continue to be reasonable, even after 2005 when U.S. farmers had not anyfirsthand experience with SBR. Some support for this view stems from the factthat, even following a benign year in 2005, survey data indicate many farmersremained concerned about possible SBR infection in 2006. For example, whileover 50% thought it was “very unlikely” that their fields would be infested, and22% thought it “somewhat unlikely;” there were still 6% and 3% of farmers whorespectively thought infestation was “somewhat” and “very likely” in their fields.Seventeen percent were uncertain whether their fields would be infested or not.This means over a quarter of surveyed soybean farmers thought that infestationwas more likely than not, or were uncertain. While we cannot discern preciseprobabilistic priors from qualitative survey responses, these data do indicate thatfarmers perceived significant infestation risk in 2006. Thus, it is plausible thatbeliefs were just somewhat more pessimistic a year earlier in 2005.
In 2006–2008, the actual incidence of SBR steadily increased. Plant pathologistsnow believe it will take a number of years for P. pachyrhizi spores to build upto natural equilibrium levels.5 Thus, individual-season prior infection beliefs arelikely rising again after an initial fall after the 2005 season.
Given the difficulty of determining prior beliefs of infection and the funda-mentally subjective nature of those beliefs, we supplement our base-case priorprobabilities with a sensitivity analysis that varies prior infection probabilitiesfrom 10% to 120% of our base-case estimates. We also look more closely at thesoybean-rich Corn Belt, considering for that region a continuum of prior beliefsthat ranges from zero to one.
Base Case ResultsThe value of information is the difference in expected profits between the
partial-information environment (figure 2) and the no-information environment(figure 1). Results reported in table 3 indicate that the value of information variesbetween zero and $6.01 per U.S. soybean acre, depending on information qualityand assuming prior infestation beliefs equal our estimated infestation probabili-ties. In table 4, we report information values for a whole range of priors, scaled asa percent of our estimated priors. A striking result of this sensitivity analysis is theextreme nonlinearity of the information value with respect to the prior infestationprobabilities.
To further characterize the nonlinear relationship between information valueand prior infestation beliefs, we focus on the Corn Belt region, which has thenation’s largest share of soybean production. For this region, we examine thefull range of possible priors (figure 3). Values for all three information qualitiespeak at prior infection probabilities of � = 0.19 and 0.63. These probabilities markswitching points in farmers’ optimal strategies. Below a 19% chance of infection,the best strategy is to do nothing; between a 19% and 63% chance of infection,the best strategy is monitoring and application of curative fungicides if infected;above a 63% chance of infection, the best strategy is to use preventive fungicides.For a given quality of information, values are highest near these switching points,because information has the greatest scope for altering farmers’ decisions and re-ducing the chance of ex post errors. Of course, the better the quality of information,the more ex post errors are reduced, creating greater value. When information is
314 Review of Agricultural EconomicsT
able
4.S
ensi
tivi
tyan
alys
is
Pri
or(P
erce
nta
geof
Bas
eC
ase
Reg
ion
alA
ssu
mp
tion
s)In
foR
egio
nQ
ual
ity
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
110%
120%
App
alac
hia
Low
077
3,79
12,
414,
674
3,79
0,56
10
00
00
2,87
8,57
23,
627,
729
1,41
7,39
0M
ediu
m2,
163,
377
4,84
9,24
88,
057,
614
10,6
87,4
887,
103,
426
4,13
9,74
33,
517,
786
4,47
5,20
37,
478,
491
11,0
24,8
6510
,796
,467
7,28
2,72
1H
igh
5,29
7,41
411
,021
,060
17,1
70,9
3922
,646
,064
21,8
10,9
9021
,402
,149
21,4
19,5
4021
,863
,163
22,7
33,0
1924
,029
,107
21,4
33,5
6815
,435
,828
Cor
nB
elt
Low
3,84
1,39
013
,601
,571
29,2
80,5
4212
,101
,905
00
00
07,
733,
674
36,0
62,7
3336
,087
,619
Med
ium
23,0
77,4
5949
,854
,163
80,3
30,1
1175
,728
,903
49,6
50,7
0133
,572
,259
27,3
32,8
7243
,762
,883
63,3
24,2
2286
,016
,890
111,
840,
886
107,
481,
912
Hig
h53
,107
,148
108,
946,
524
167,
518,
130
190,
045,
566
186,
262,
155
185,
210,
973
186,
892,
021
191,
305,
297
198,
450,
803
208,
328,
537
220,
938,
501
202,
966,
395
Del
taL
ow0
017
6,85
773
6,02
91,
545,
310
1,71
4,96
415
8,62
50
00
00
Med
ium
621,
699
1,39
9,71
72,
334,
053
3,42
4,70
74,
671,
680
5,18
5,23
53,
879,
005
2,72
9,09
51,
948,
953
2,06
1,54
62,
515,
688
3,61
8,74
1H
igh
1,68
0,54
83,
497,
716
5,45
1,50
27,
541,
909
9,76
8,93
411
,242
,843
10,8
77,2
6810
,648
,312
10,5
55,9
7610
,600
,259
10,7
81,1
6211
,098
,684
Lak
eSt
ates
Low
517,
193
2,29
1,75
05,
323,
673
9,61
2,96
03,
015,
380
00
00
00
3,53
3,83
6M
ediu
m5,
120,
637
11,0
27,1
2717
,719
,470
25,1
97,6
6521
,317
,482
14,2
36,7
1710
,290
,200
7,94
9,25
79,
392,
784
13,9
77,2
1519
,269
,805
25,2
70,5
52H
igh
12,0
83,4
0624
,764
,417
38,0
43,0
3151
,919
,250
54,2
48,8
4253
,189
,603
52,7
27,9
6852
,863
,938
53,5
97,5
1254
,928
,690
56,8
57,4
7359
,383
,860
Nor
thea
stL
ow50
,552
346,
358
887,
418
976,
802
00
00
050
2,62
61,
631,
964
819,
977
Med
ium
751,
825
1,65
6,93
42,
715,
326
3,23
0,07
22,
104,
172
1,32
2,58
71,
050,
968
1,43
3,85
22,
276,
353
3,26
6,24
64,
185,
548
3,07
5,09
1H
igh
1,79
0,77
73,
701,
825
5,73
3,14
37,
187,
802
6,96
8,80
26,
870,
072
6,89
1,61
47,
033,
425
7,29
5,50
87,
677,
861
7,96
2,50
16,
190,
259
Nor
ther
nL
ow0
568,
344
2,60
7,84
75,
817,
571
10,1
97,5
1513
,667
,862
3,32
2,52
90
00
00
Plai
nsM
ediu
m4,
599,
361
9,93
0,11
015
,992
,247
22,7
85,7
7230
,310
,685
36,4
87,1
6728
,409
,137
21,0
62,4
9414
,447
,239
12,2
42,0
7911
,579
,232
13,4
07,3
81H
igh
11,5
16,2
3623
,624
,791
36,3
25,6
6649
,618
,860
63,5
04,3
7475
,902
,390
73,9
06,8
2472
,503
,577
71,6
92,6
5071
,474
,043
71,8
47,7
5572
,813
,787
Sout
heas
tL
ow0
00
014
2,40
647
3,03
122
0,74
00
00
416,
584
128,
548
Med
ium
56,0
2017
4,98
735
6,90
160
1,76
290
9,56
91,
280,
323
1,03
0,39
370
0,86
475
1,34
21,
094,
027
1,22
3,34
755
4,74
9H
igh
284,
989
632,
394
1,04
7,68
91,
529,
051
2,07
6,47
92,
689,
972
2,68
5,90
02,
605,
350
2,59
0,86
52,
642,
446
2,38
1,56
51,
223,
667
Sout
hern
Low
00
00
08,
171
130,
908
288,
325
328,
898
132,
508
00
Plai
nsM
ediu
m35
,112
91,8
9817
0,35
827
0,49
239
2,30
153
5,78
470
0,94
188
7,77
294
4,75
675
1,77
258
0,46
243
9,14
3H
igh
158,
169
338,
428
540,
777
765,
217
1,01
1,74
71,
280,
367
1,57
1,07
81,
883,
879
2,06
7,24
92,
001,
068
1,95
6,97
61,
934,
975
Oth
erL
ow0
34,9
6752
7,93
01,
337,
880
2,46
4,81
71,
250,
017
00
00
00
Med
ium
922,
216
2,04
2,54
93,
360,
998
4,87
7,56
46,
592,
247
5,84
6,32
44,
215,
036
2,78
1,86
52,
347,
647
2,49
1,10
83,
379,
917
4,99
6,88
2H
igh
2,37
2,74
54,
912,
049
7,61
7,91
410
,490
,339
13,5
29,3
2314
,076
,145
13,7
06,0
4613
,502
,507
13,4
65,5
2813
,595
,110
13,8
91,2
5114
,353
,952
U.S
.Tot
alL
ow4,
409,
135
17,6
16,7
8141
,218
,941
34,3
73,7
0817
,365
,427
17,1
14,0
453,
832,
802
288,
325
328,
898
11,2
47,3
8141
,739
,010
41,9
87,3
71M
ediu
m37
,347
,706
81,0
26,7
3213
1,03
7,07
714
6,80
4,42
612
3,05
2,26
310
2,60
6,14
080
,426
,338
85,7
83,2
8510
2,91
1,78
613
2,92
5,74
716
5,37
1,35
216
6,12
7,17
2H
igh
88,2
91,4
3118
1,43
9,20
327
9,44
8,79
234
1,74
4,05
835
9,18
1,64
737
1,86
4,51
537
0,67
8,25
937
4,20
9,45
038
2,44
9,11
039
5,27
7,12
140
8,05
0,75
238
5,40
1,40
6
Estimating the Value of an Early-Warning System 315
Figure 3. Possible information values in the Corn Belt
Table 5. Aggregate information values for the United States
Information Quality
Low Medium High(� = 0.2) (� = 0.5) (� = 0.8)
Scenario (in Dollars) (in Dollars) (in Dollars)
Base caseU.S. total 11,247,380 132,925,747 395,277,121Average per acre 0.16 1.84 5.46
Risk aversionU.S. total 16,880,465 136,418,926 391,344,966Average per acre 0.23 1.88 5.41
Price feedbackU.S. total 28,773,280 130,259,030 376,391,350Average per acre 0.40 1.80 5.20
Heterogeneous beliefsU.S. total 16,777,090 102,300,370 275,889,857Average per acre 0.23 1.41 3.81
of low quality (� = 0.2), there are some priors for which the information providedwill not cause farmers to change their management behavior, thus producing novalue at all.
Using estimated probabilities of infection for prior beliefs in each region andaggregating on a per-acre basis across all regions, we estimate in table 5 thatthe aggregate value of information ranges between $11 million and $395 mil-lion, depending on information quality. In the next section, we show how these
316 Review of Agricultural Economics
information values were determined under the assumptions that farmers arestrongly risk averse, that soybean prices are influenced by soybean-rust-inducedyield losses, and farmers’ prior beliefs are heterogeneous within each region. Ascan be seen in table 5, these scenarios all increase the lowest information val-ues and reduce the highest values in comparison to the base case. Despite thebroad range of information values estimated, all far exceed the program’s esti-mated cost in the first year, which ranged between $2.5 million and $5.0 million(U.S. Department of Agriculture, 2005).6 Because the value increases between$2 million and $3 million for each percentage point of uncertainty resolved by theframework, efforts to improve timeliness and accuracy of infection forecasts maywell be cost-effective.
Alternative ScenariosIn the base case scenario described in the preceding section, we evaluate farm-
ers’ expected-profit-maximizing management decisions with and without infor-mation and estimate the value of information as the difference in expected profitsbetween the two environments. In the following scenarios, the concept is similarbut some assumptions are changed or relaxed in order to explore the sensitivityof the base-case estimates to these assumptions.
Alternative Scenario 1: Information Values of Risk-Averse FarmersWe first consider how our results change if farmers are strongly risk averse.
More specifically, we assume farmers’ preferences can be characterized by con-stant relative risk aversion (CRRA), with a coefficient of relative risk aversionequal to 4. This may be expressed with the utility function: u(W) = −AW−3/3,where W indicates wealth, and A is an arbitrary constant. We made this assump-tion to throw into stark relief the effect of risk aversion on information values.
Using data from USDA’s 2003 Agricultural Resource Management Survey, weestimate base wealth for each region by weighting farm households’ net worthby their number of soybean acres. We find that wealth varies considerably acrossfarm sizes and across the country, with the average soybean acre being associ-ated with a household net worth of $1,649,807 in Appalachia and $1,348,667 inthe Corn Belt, but only $918,870 in Delta states. Average net worth is $1,430,615in Lake States, $1,030,815 in the Northeast, $1,389,427 in the Northern Plains,$1,300,438 in the Southeast, and $1,572,391 in the Southern Plains. Net worth forthe average soybean acre of farms outside these regions is $915,964. Calculatinginformation values for risk-averse farmers’ proceeds similarly to the base casedescribed earlier, except that farmers are assumed to maximize expected utilityrather than expected profits.
In only a few cases does the extreme level of risk aversion cause farmers’ deci-sions to differ from the base case. This assumption changes information values,however, mainly because different information environments may lead to largedifferences in profit variability. For example, consider a farmer who would haveapplied the preventive strategy without information. Suppose that if armed witha high-quality forecast, the farmer chooses prevention in response to a “high-risk” signal and does nothing in response to “low risk.” The information wouldcause her average profits to increase but would also cause her profit variability to
Estimating the Value of an Early-Warning System 317
increase, so the information would be valued less by this farmer than by anotherwho is less risk-averse. When the information signal is of poorer quality, informa-tion may increase or decrease the amount of risk, so the information value maybe relatively greater or smaller than in the risk-neutral base case.
Table 6 reports results from the analysis of risk-averse farmers. Differences fromthe base-case scenario are generally modest and vary somewhat across regionsand information qualities. In the Corn Belt, for example, a strongly risk-aversefarmer values low-quality information at $0.37 per acre versus the base-case valueof $0.22. Alternatively, the same risk-averse farmer values high-quality informa-tion at $5.96 per acre, which is slightly less than the value of $6.01 in the base case.More realistic assumptions about the level of risk aversion would imply smallerdifferences from the base case.7 U.S. aggregate information values (see table 5)range from almost $17 million (low quality) to over $391 million (high quality).
Alternative Scenario 2: Price Feedback EffectsThe base case scenario assumed soybean prices were fixed at the May 2, 2005,
futures price. However, both economic theory and historical evidence indicatethat soybean prices vary with yield, and because each decision (prevent, moni-tor/cure, or no management) and each outcome (rust infection or no rust infection)leads to different yields, we consider the additional possibility that postharvestsoybean prices are endogenous. Table 7 presents information values similar tothose in table 3, except infestations and farm management decisions are assumedto influence market prices for soybeans.
Equilibrium soybean price is determined as follows: individual farmers, takingexpected postharvest price as given, maximize their own profits, while the indus-try as a whole, which is made up of these individual profit-maximizing farmers,satisfies one of the equations below. In these equilibrium equations, the May 2futures price equals the average of all potential end-of-season soybean prices,weighted by the market-perceived probabilities that these prices will be realized.In the case where no information is available, this means:
The above conditions are only correct if the geography of the soybean marketcoincides with that of the SBR infection and message probabilities. Given the widevariation in climate conditions across the United States and the global nature of
318 Review of Agricultural Economics
Tab
le6.
Info
rmat
ion
valu
esw
ith
risk
aver
sion
Pri
orC
Eof
Hig
hL
owC
Eof
Val
ue
ofV
alu
eof
Bel
ief
ofB
ase
No
Info
EU
No
Info
rmat
ion
Ris
kR
isk
EU
Wit
hIn
foP
erIn
foP
erIn
fect
ion
Wea
lth
Dec
isio
nE
UN
oIn
fo.
Qu
alit
y(�
)D
ecis
ion
Dec
isio
nIn
foFa
rmA
cre
Reg
ion
(Pro
bab
ilit
y)(D
olla
rs)
(P,M
,or
N)
Info
.(D
olla
rs)
(Sca
le0
to1)
(P,M
,or
N)
(P,M
,or
N)
(Dol
lars
)(D
olla
rs)
(Dol
lars
)
App
alac
hia
0.67
1,64
9,80
7M
1.35
478
,371
0.2
PM
79,2
4887
70.
780.
671,
649,
807
M1.
354
78,3
710.
5P
M81
,249
2,87
82.
570.
671,
649,
807
M1.
354
78,3
710.
8P
N84
,302
5,93
15.
30C
orn
Bel
t0.
551,
348,
667
M0.
889
93,5
000.
2P
M93
,772
272
0.37
0.55
1,34
8,66
7M
0.88
993
,500
0.5
PM
95,4
461,
946
2.62
0.55
1,34
8,66
7M
0.88
993
,500
0.8
PN
97,9
254,
425
5.96
Del
ta0.
5591
8,87
0M
(1.1
84)
96,5
560.
2M
M96
,556
00
0.55
918,
870
M(1
.184
)96
,556
0.5
PM
98,0
851,
529
0.78
0.55
918,
870
M(1
.184
)96
,556
0.8
PN
104,
795
8,23
94.
21L
ake
Stat
es0.
491,
430,
615
M0.
990
58,4
760.
2M
M58
,476
00
0.49
1,43
0,61
5M
0.99
058
,476
0.5
PM
59,2
5177
51.
450.
491,
430,
615
M0.
990
58,4
760.
8P
N61
,329
2,85
35.
34N
orth
east
0.62
1,03
0,81
5M
(0.7
00)
42,0
170.
2P
M42
,241
223
0.47
0.62
1,03
0,81
5M
(0.7
00)
42,0
170.
5P
M43
,183
1,16
62.
460.
621,
030,
815
M(0
.700
)42
,017
0.8
PN
44,6
352,
618
5.53
Con
tinu
ed
Estimating the Value of an Early-Warning System 319
Tab
le6.
Con
tin
ued
Pri
orC
Eof
Hig
hL
owC
Eof
Val
ue
ofV
alu
eof
Bel
ief
ofB
ase
No
Info
EU
No
Info
rmat
ion
Ris
kR
isk
EU
Wit
hIn
foP
erIn
foP
erIn
fect
ion
Wea
lth
Dec
isio
nE
UN
oIn
fo.
Qu
alit
y(�
)D
ecis
ion
Dec
isio
nIn
foFa
rmA
cre
Reg
ion
(Pro
bab
ilit
y)(D
olla
rs)
(P,M
,or
N)
Info
.(D
olla
rs)
(Sca
le0
to1)
(P,M
,or
N)
(P,M
,or
N)
(Dol
lars
)(D
olla
rs)
(Dol
lars
)
Nor
ther
n0.
431,
389,
427
M0.
929
70,4
610.
2M
M70
,461
00
Plai
ns0.
431,
389,
427
M0.
929
70,4
610.
5P
N71
,022
562
0.64
0.43
1,38
9,42
7M
0.92
970
,461
0.8
PN
74,6
084,
147
4.71
Sout
heas
t0.
761,
300,
438
M0.
493
2,37
50.
2M
M2,
375
00
0.76
1,30
0,43
8M
0.49
32,
375
0.5
PN
3,01
363
71.
440.
761,
300,
438
M0.
493
2,37
50.
8P
N3,
910
1,53
53.
47So
uthe
rn0.
511,
572,
391
M1.
181
24,5
160.
2M
N24
,543
270.
02Pl
ains
0.51
1,57
2,39
1M
1.18
124
,516
0.5
MN
26,2
951,
778
1.17
0.51
1,57
2,39
1M
1.18
124
,516
0.8
PN
29,9
775,
461
3.58
Oth
er0.
5391
5,96
4M
(1.4
92)
68,6
660.
2M
M68
,666
00
0.53
915,
964
M(1
.492
)68
,666
0.5
PM
69,6
1394
70.
860.
5391
5,96
4M
(1.4
92)
68,6
660.
8P
N73
,620
4,95
44.
52
Not
es:I
nth
ed
ecis
ion
colu
mns
,Mis
mon
itor
/cu
re,P
ispr
even
t,an
dN
isd
ono
thin
g.E
Uin
dic
ates
“exp
ecte
dut
ility
”an
dC
Ein
dic
ates
“cer
tain
tyeq
uiva
lenc
e,”
whi
chis
the
cert
ain
leve
lofp
rofit
sw
ith
sam
eut
ility
asth
eac
tual
,unc
erta
inle
velo
fpro
fits.
320 Review of Agricultural Economics
Tab
le7.
Info
rmat
ion
valu
esw
ith
pri
ceef
fect
s
Pri
orH
igh
Low
Low
Val
ue
ofV
alu
eof
Bel
ief
ofN
oIn
foE
VN
oIn
form
atio
nR
isk
Ris
kR
isk
EV
Wit
hIn
foP
erIn
foP
erIn
fect
ion
Dec
isio
nIn
foQ
ual
ity
(�)
Dec
isio
nD
ecis
ion
EV
Info
Farm
Acr
eR
egio
n(P
rob
abil
ity)
(P,M
,or
N)
(Dol
lars
)(S
cale
0to
1)(P
,M,o
rN
)(P
,M,o
rN
)(D
olla
rs)
(Dol
lars
)(D
olla
rs)
(Dol
lars
)
App
alac
hia
0.67
M78
,519
0.2
PM
84,5
3379
,442
923
0.83
0.67
M78
,519
0.5
PM
90,0
7081
,174
2,65
52.
370.
67M
78,5
190.
8P
N10
0,50
784
,540
6,02
15.
39C
orn
Bel
t0.
55M
93,3
020.
2P
M99
,471
93,8
1951
70.
700.
55M
93,3
020.
5P
M10
1,72
295
,424
2,12
12.
860.
55M
93,3
020.
8P
N10
7,54
797
,568
4,26
55.
75D
elta
0.55
M98
,203
0.2
MM
104,
540
98,2
030
00.
55M
98,2
030.
5P
N11
2,91
499
,763
1,56
00.
800.
55M
98,2
030.
8P
N12
9,99
610
6,85
98,
657
4.43
Lak
eSt
ates
0.49
M58
,608
0.2
MM
59,9
4558
,608
00
0.49
M58
,608
0.5
PM
62,8
2659
,217
610
1.14
0.49
M58
,608
0.8
PN
67,1
0861
,239
2,63
24.
93N
orth
east
0.62
M41
,936
0.2
PM
45,3
0742
,216
279
0.59
0.62
M41
,936
0.5
PM
47,1
9243
,152
1,21
62.
570.
62M
41,9
360.
8P
N51
,267
44,4
812,
545
5.38
Con
tinu
ed
Estimating the Value of an Early-Warning System 321
Tab
le7.
Con
tin
ued P
rior
Hig
hL
owL
owV
alu
eof
Val
ue
ofB
elie
fof
No
Info
EV
No
Info
rmat
ion
Ris
kR
isk
Ris
kE
VW
ith
Info
Per
Info
Per
Infe
ctio
nD
ecis
ion
Info
Qu
alit
y(�
)D
ecis
ion
Dec
isio
nE
VIn
foFa
rmA
cre
Reg
ion
(Pro
bab
ilit
y)(P
,M,o
rN
)(D
olla
rs)
(Sca
le0
to1)
(P,M
,or
N)
(P,M
,or
N)
(Dol
lars
)(D
olla
rs)
(Dol
lars
)(D
olla
rs)
Nor
ther
nPl
ains
0.43
M70
,621
0.2
MM
72,0
3270
,621
00
0.43
M70
,621
0.5
PM
/N
77,8
0470
,588
(33)
(0.0
4)0.
43M
70,6
210.
8P
N83
,612
74,5
513,
930
4.47
Sout
heas
t0.
76M
2,41
90.
2M
M3,
997
2,41
90
00.
76M
2,41
90.
5P
N7,
570
3,03
761
81.
390.
76M
2,41
90.
8P
N11
,239
3,96
21,
543
3.48
Sout
hern
Plai
ns0.
51M
24,8
070.
2M
M/
N30
,006
24,7
03(1
03)
(0.0
7)0.
51M
24,8
070.
5M
N38
,578
26,6
051,
798
1.18
0.51
M24
,807
0.8
PN
48,7
8730
,456
5,64
93.
71O
ther
0.53
M69
,085
0.2
MM
72,4
8169
,267
00
0.53
M69
,085
0.5
PN
78,1
6270
,097
830
0.76
0.53
M69
,085
0.8
PN
86,8
8474
,039
4,77
24.
35
Not
es:I
nth
ed
ecis
ion
colu
mns
,Mis
mon
itor
/cu
re,P
ispr
even
t,an
dN
isd
ono
thin
g.In
colu
mn
head
ings
,EV
ind
icat
esex
pect
edva
lue.
322 Review of Agricultural Economics
the soybean market, this is unrealistic. Allowing probabilities to vary by location,however, quickly renders the price equilibrium condition intractable. Even withthe simplifying assumption that infection and message probabilities are homoge-neous within each of the eight major soybean-producing regions,8 the number ofterms on the left-hand side of the partial information equilibrium equation climbsto 6,561. After all, there are three yield-determining outcomes (“high risk” mes-sage with infection, “low risk” message with infection, and no infection) possiblein each region, producing 38 characterizations for the nation as a whole.
Though we do not explicitly model the regional interconnectedness that pro-duces this host of outcomes, we allow for such effects by not including alternatelocation yield shocks as explanatory variables in our regressions of postharvestsoybean price on regional yields (there are eight regressions—one for each region).Thus, our coefficient estimates will capture not only the effect of a specific region’syield on postharvest soybean price but also the effects of other yield shocks thatare correlated with the region’s own yield shocks (presumably those occurringin locations close in distance and weather). While these estimates provide someinsight into how regional soybean prices and yields have been spatially correlatedhistorically, there is no guarantee soybean rust will exhibit similar spatial effectsas the weather and production shocks of the past decades. If, for example, soy-bean rust were to spread quickly over the entire soybean-producing part of NorthAmerica, a rust-induced regional price increase would likely be greater than theincrease resulting from yield loss caused by a more localized drought.
Our first step is to aggregate, using production-weighted averages, state-leveldata (U.S. Department of Agriculture 1950–2004) to the regional level. Next, inorder to abstract from yearly variations in output while still accounting for pro-ductivity increases over time, we fit a smooth trend curve for yields in all ninesoybean production regions. Examples of these trends for the Corn Belt and South-east are given in figure 4.
Figure 4. Yield trends and shocks in the Corn Belt and Southeast
1950 1960 1970 1980 1990 2000
2530
3540
45
Corn Belt Yields and Yield Trend, 1950-2004
Year
Yie
ld (
bush
els/
acre
)
1950 1960 1970 1980 1990 2000
1618
2022
24
Southeast Yields and Yield Trend, 1950-2004
Year
Yie
ld (
bush
els/
acre
)
Estimating the Value of an Early-Warning System 323
This fitting process allows us to calculate, for each region in each year, a percent-age residual yield (i.e., the difference between actual yield and yield predictedby the trend, divided by the yield predicted by the trend). We approximate thepercentage change in regional soybean prices by calculating the year-to-year dif-ference in the natural logarithm of the real price. By regressing this value on thepercentage residual yield, we obtain an estimate of the percentage change in pricethat would result from a percentage deviation from the yield trend.9, 10
In computing the results in table 7, we first find the equilibria11 in which allfarmers within a region make identical management decisions. In just two cases,such equilibria do not exist: the Northern Plains with an information quality of0.5 and the Southern Plains with an information quality of 0.2. For these regions,we consider equilibria where farmers apply one strategy to a share of acreagewithin a region and apply another strategy to the remainder. In these two cases,in equilibrium, farmers are indifferent between monitoring and no-managementstrategies when they receive the low-risk signal. An equilibrium results in theNorthern Plains scenario when, in response to a low-risk signal, about 35% ofacreage is monitored, the remainder is unmanaged, and the postharvest price,when the signal indicates low risk but infection occurs anyway, is $6.91. Similarly,the Southern Plains is in equilibrium when, in the face of low risk, about 27% ofacreage is monitored; 73% is unmanaged; and the postharvest price, when thesignal indicates a low risk-signal but infection occurs anyway, is $6.45.
For most regions, accounting for price effects has a small influence on informa-tion values. The effect is somewhat greater in the Corn Belt and Northern Plains
Figure 5. Density of farmers’ heterogeneous prior beliefs in the CornBelt
324 Review of Agricultural Economics
Tab
le8.
Info
rmat
ion
valu
esw
ith
het
erog
eneo
us
pri
orb
elie
fs
Ave
rage
Ave
rage
Ave
rage
Pri
or“a
lph
a”“b
eta”
Val
ue
ofV
alu
eof
Bel
ief
ofP
aram
eter
Par
amet
erN
oIn
foE
VN
oIn
form
atio
nIn
foP
erIn
foP
erIn
fect
ion
for
Bet
afo
rB
eta
Dec
isio
nIn
foQ
ual
ity
(�)
Farm
Acr
eR
egio
n(P
rob
abil
ity)
Pri
orP
rior
(P,M
,or
N)
(Dol
lars
)(S
cale
0to
1)(D
olla
rs)
(Dol
lars
)
App
alac
hia
0.67
2.00
1.00
M78
,644
0.2
239
0.21
0.67
2.00
1.00
M78
,644
0.5
1,46
51.
310.
672.
001.
00M
78,6
440.
83,
957
3.54
Cor
nB
elt
0.55
1.20
1.00
M93
,707
0.2
182
0.25
0.55
1.20
1.00
M93
,707
0.5
1,13
21.
530.
551.
201.
00M
93,7
070.
83,
001
4.04
Del
ta0.
551.
201.
00M
97,9
630.
239
10.
200.
551.
201.
00M
97,9
630.
52,
360
1.21
0.55
1.20
1.00
M97
,963
0.8
6,50
43.
33L
ake
Stat
es0.
490.
951.
00M
58,5
730.
213
70.
260.
490.
951.
00M
58,5
730.
580
11.
500.
490.
951.
00M
58,5
730.
82,
185
4.09
Nor
thea
st0.
621.
601.
00M
42,1
090.
210
70.
230.
621.
601.
00M
42,1
090.
566
01.
390.
621.
601.
00M
42,1
090.
81,
793
3.79
Con
tinu
ed
Estimating the Value of an Early-Warning System 325
Tab
le8.
Con
tin
ued
Ave
rage
Ave
rage
Ave
rage
Pri
or“a
lph
a”“b
eta”
Val
ue
ofV
alu
eof
Bel
ief
ofP
aram
eter
Par
amet
erN
oIn
foE
VN
oIn
form
atio
nIn
foP
erIn
foP
erIn
fect
ion
for
Bet
afo
rB
eta
Dec
isio
nIn
foQ
ual
ity
(�)
Farm
Acr
eR
egio
n(P
rob
abil
ity)
Pri
orP
rior
(P,M
,or
N)
(Dol
lars
)(S
cale
0to
1)(D
olla
rs)
(Dol
lars
)
Nor
ther
nPl
ains
0.43
0.75
1.00
M70
,688
0.2
179
0.20
0.43
0.75
1.00
M70
,688
0.5
1,07
71.
220.
430.
751.
00M
70,6
880.
83,
009
3.42
Sout
heas
t0.
763.
201.
00M
2,40
80.
281
0.18
0.76
3.20
1.00
M2,
408
0.5
450
1.01
0.76
3.20
1.00
M2,
408
0.8
1,11
12.
51So
uthe
rnPl
ains
0.51
1.05
1.00
M24
,973
0.2
243
0.16
0.51
1.05
1.00
M24
,973
0.5
1,46
50.
960.
511.
051.
00M
24,9
730.
83,
734
2.45
Oth
er0.
531.
131.
00M
69,1
590.
223
40.
210.
531.
131.
00M
69,1
590.
51,
396
1.27
0.53
1.13
1.00
M69
,159
0.8
3,84
63.
51
Not
es:
Inth
ed
ecis
ion
colu
mn,
Mis
mon
itor
/cu
re.
Inot
her
colu
mn
head
ings
,E
Vis
expe
cted
valu
e,an
d“a
lpha
”an
d“b
eta”
para
met
ers
are
thos
ein
the
beta
prob
abili
tyd
istr
ibut
ion.
326 Review of Agricultural Economics
because yield shocks in these soybean-intensive regions have larger estimatedprice effects. In the Corn Belt, for example, the value of information increasesfrom the table 3 amount of $0.22 per acre to $0.70 per acre when it is of low qualityand declines from $6.01 to $5.75 when it is of high quality. In the Southeast, wherethe estimated price effects are far smaller, the values of both low- and high-qualityinformation are unchanged at zero and $3.48 per acre, respectively.
When we account for price feedback effects, small changes in expected yieldlead to small changes in expected price (i.e., the futures price). If informationcauses an increase in expected yield, expected prices tend to decline. If the ex-pected price decline is large enough, farmers’ total expected profits might de-cline because of the information, even though individual farmers find the infor-mation valuable (because individual farmers take prices as given). For soybeanconsumers, however, this price decline is a gain—it represents a transfer fromproducers to consumers. Of course, the opposite is true if the information causesa decline in expected yield: prices increase, and consumers experience a welfareloss because of the information’s existence
Alternative Scenario 3: Average Information Values for Farmswith Heterogeneous Beliefs
A third departure from the base case results comes from changing our assump-tions about farmers’ prior beliefs. Whereas in the base case we assume all farmerswithin a region hold the same prior beliefs about the probability of infection, inthis scenario we assume farmers have heterogeneous prior beliefs. Specifically,we assume beliefs within a region are distributed according to a beta distributionwith the beta parameter equal to one and the alpha parameter set so that the meanequals the prior probability of infection in the base case (table 3). This assumptionimplies that farmers’ beliefs vary widely within each region.
The assumed distribution for the Corn Belt is plotted in figure 5. The height ofthe density curve shows the relative proportion of farmers assumed to have theprior belief plotted along the horizontal axis. We estimate average informationvalues for each region and information quality by taking 1,000 random draws fromthe assumed beta distribution, using each draw as the value for P, calculating theassociated information values from each draw, and then taking the average of theinformation values across all 1,000 draws (table 8).
In general, heterogeneous prior beliefs tend to reduce the highest informationvalues and increase the lowest ones. The highest values decline because theyare associated with the highest-value prior beliefs—those near the critical prob-abilities that mark the switching points between strategies. With heterogeneousbeliefs, these high-value prior beliefs are averaged with lower value prior beliefs,bringing down the overall average. Conversely, the lowest value prior beliefs areaveraged with higher value prior beliefs, which bring those information valuesup.
ConclusionThe value of disease information to farmers depends on many factors, but par-
ticularly their perceived risk of being infected with SBR at the beginning of theseason and the accuracy of forecasts provided by the program. Over a broad range
Estimating the Value of an Early-Warning System 327
of plausible parameters, the value of information provided by the SBR Coordi-nated Framework could have been substantial even though the disease did notspread to the major soybean growing regions in the Midwest. If the forecasts werepoor, resolving only 20% of SBR infection uncertainty for all fields planted withsoybeans, the estimated value of the program was $11 million during the first year.If the forecasts resolved 80% of infestation uncertainty, the estimated value was$395 million. The estimated value is also sensitive to prior beliefs about the likeli-hood of infestation. Unlike forecast accuracy, however, the relationship betweeninformation value and prior beliefs is highly nonlinear. While some alternativeprior beliefs would lead to lower information values as compared to our baseline,others give much higher values.
The sensitivity of information value to the extent of resolved uncertainty sug-gests that the potential value of information will be greatest for pest problemsthat can be forecasted accurately, that farmers have little experience with, andthat have large potential impacts on crop production that can be mitigated usingpreventive management activities (Carlson). Other key drivers of the total valueof information are the size and value of the crop in question. For instance, the neartripling of soybean commodity prices since 2005 would imply a near tripling ofour estimated value of SBR forecasts, holding all else the same.
Two other more subtle features affect estimated information values: anticipatedprice shocks in the event of large SBR outbreaks and the risk aversion of soybeanfarmers. We find that both of these effects reduce the largest estimated values andincrease the smallest ones, but the magnitudes of these effects are modest.
By examining the value of information across a range of forecast accuracies, wehave illustrated the value of marginally improving information quality. Since themarginal information value appears large in the case of SBR, in future work it maybe worthwhile to consider models that explicitly incorporate features of the infor-mation collection system. In the SBR framework, information quality is linked tothe number of sentinel plots used for monitoring, where the plots are located spa-tially, and how frequently the sentinel plots are monitored. It would be interestingto consider the optimal mix of these policy choices. While such a model would beextremely complex, the potential social gains from its construction would seemto warrant the effort.
Such a model would need to capture the spatial and dynamic patterns of infes-tations. Our discrete two-period timing of information flows and decision makingand the way we deal with spatial dependencies are crude relative to the continu-ous and spatially dependent processes that exist in reality. In a more sophisticatedmodel, prior beliefs would be determined at different points in time for differ-ent farmers, since the timing of the growing season varies geographically. Evenwithout the monitoring framework, many farmers would have had more infor-mation thantheir beginning-of-the-season prior beliefs, say from weather reportsand news about infections further south. Thus, the relevant decision-time prior isitself uncertain at the beginning of the season. Because we found information tobe valuable across many priors, integrating over these priors would likely lead toinformation values not drastically different from the ones we have estimated. Amore sophisticated model, however, would provide insight into the appropriatestructure of information systems—where and how frequently monitoring wouldbe most valuable.
328 Review of Agricultural Economics
A more sophisticated model may also be able to account for the externalitiesassociated with farmers’ pest management choices. Each farmer’s monitoringand fungicide application decisions affect the likelihood that his or her neighborswill be infected. This physical externality embodies another market failure thatsits alongside the public information problem. While the optimal policy solutionwould likely involve separate tools for each problem, in practice, solving oneproblem may either mitigate or exacerbate the other.
AcknowledgmentsThe authors would like to thank two anonymous reviewers for valuable input on an earlier draft,
and the GAO for providing us data from their report (GAO-05-668R) ‘‘USDA’s Preparation for AsianSoybean Rust,’’ which was presented as a briefing for Senator Tom Harkin (Ranking DemocraticMember Committee on Agriculture, Nutrition, and Forestry) United States Senate on 28 April 2005.The authors also thank NASS for allowing us to access unpublished 2005 soybean planting intentionsdata, David Nulph and Vince Breneman for terrific GIS support, and Gabriella LaRocca for assistancewith the wheat stem rust data. The views expressed are those of the authors and should not beattributed to the Economic Research Service or the U.S. Department of Agriculture.
Endnotes1In this case study, we do not consider farmers’ planting decisions, only their fungicide application
decisions, provided they do plant. By ignoring this decision, we underestimate the value of informationprovided by the framework.
2Costs and yield losses reported in table 1 are estimates. Yield data, before and after the arrival of P.pachyrhizi are not available for the United States, nor are efficacy trial data for U.S. fungicides. Efficacydata also were not available at the time of this study for climatic regions similar to the United States.Thus, to estimate treated and untreated yield impacts of soybean rust epidemics relative to rust-freeyields, we evaluated impacts of rust on soybean yields in South America. Details are provided in anonline supplement to Roberts et al.
3Poorly timed sprays could lead to the need for additional applications (Dorrance, Draper, andHershman). We do not attempt to quantify the value of improved timing, which would likely increasethe presented estimates for the value of information.
4The regions contain the following states: Appalachia (KY, NC, TN, VA, and WV), Corn Belt (IA,IL, IN, MO, and OH), Delta (AR, LA, and MS), Lake States (MI, MN, and WI), Northeast (CT, DE, MA,MD, ME, NH, NJ, NY, PN, RI, and VT), Northern Plains (KS, ND, NE, and SD), Southeast (AL, FL,GA, and SC), and the Southern Plains (OK and TX).
5From personal communication with Douglas G. Luster, Research Leader of USDA AgriculturalResearch Service, Foreign Disease Weed Science Research Unit.
6Totals depend on which fixed, start-up costs are included. Extension agents and land-grant pro-fessors volunteered their time, which is not counted in the cost estimate. The testing labs in Beltsville,Maryland, used equipment that had already been purchased, and USDA scientists involved in severalother research projects carried out the tests.
7A farmer with this utility function and wealth of $200,000, values an additional dollar 16 times asmuch as a farmer with $400,000 and 625 times as much as a farmer with $1 million.
8Defined in “Prior Infection Probabilities” section.9Only the 20 most recent observations (1984–2004) are included in this regression.10Percentage change in price from year to year will depend not only on this year’s yield shocks, but
also on yield shocks that may have affected the previous year’s price. However, including previousyear yield residuals as an explanatory variable does not lead to significant changes in estimates of thecoefficients on current-year price-shock effects.
11The relevant equilibrium condition is expressed by equation (7).
ReferencesCarlson, G.A. “A Decision Theoretic Approach to Crop Disease Prediction and Control.” Am. J. Agric.
Econ. 52, (May 1970):216–23.Dorrance, A.E., M.A. Draper, and D.E. Hershman. Using Foliar Fungicides to Manage Soybean Rust. 2005.
Available at http://www.oznet.ksu.edu/library/plant2/MF2680.pdf.
Estimating the Value of an Early-Warning System 329
Hamilton, L.M., and E.C. Stakman. “Time of Stem Rust Appearance on Wheat in the Western Mis-sissippi Basin in Relation to the Development of Epidemics from 1921 to 1962.” Phytopathology57(1967):609–14.
Hirshleifer, J., and J.G. Riley, The Analytics of Uncertainty and Information, Cambridge: CambridgeUniversity Press, 1992.
Johansson, R.C., M.J. Livingston, J. Westra, and R. Guidry. “Simulating the U.S. Impacts of AlternativeAsian Soybean Rust Treatment Regimes.” Agric. Resour. Econ. Rev. 35(2006):116–27.
Lawrence, D.B. The Economic Value of Information. New York: Springer-Verlag Inc., 1999.Livingston, M.J., R.C. Johansson, S. Daberkow, M. Roberts, M. Ash, and V. Breneman. Economic and
Policy Implications of Wind-Borne Entry of Asian Soybean Rust into the United States. Washington DC:Economic Research Service, U.S. Department of Agriculture, Outlook Report No. OCS04D02,April 2004.
Pivonia, S., and X.B. Yang. “Assessment of the Potential Year-Round Establishment of Soybean Rustthroughout the World.” Plant Disease 88(2004):523–29.
Roberts, M.J., D. Schimmelpfennig, E. Ashley, and M.J. Livingston. The Value of Plant DiseaseEarly-Warning Systems: A Case Study of USDA’s Soybean Rust Coordinated Framework. Wash-ington, DC: US Department of Agriculture, Econ. Res. Rep. ERR-18, April 2006. Available athttp://www.ers.usda.gov/Publications/ERR18/.
Schimmelpfennig D.E., and G.W. Norton. “What Is the Value of Agricultural Economics Research?”Am. J. Agric. Econ., 85 (February 2003):81–94.
Skokstad, E. “Plant Pathologists Gear up for Battle with Dread Fungus.” Science 306(2004):1672–73.U.S. Department of Agriculture. A Coordinated Framework for Soybean Rust Surveillance, Reporting, Pre-
diction, Management and Outreach. Riverdale, MD, May 2005. Available at http://www.aphis.usda.gov/plant health/plant pest info/soybean rust/downloads/coordfram041405.pdf.