Essays on U.S. Agricultural Policy: Subsidies, Crop Insurance, and
Environmental Auctions
by
Barrett Edward Kirwan
Submitted to the Department of Economics
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Economics
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2005
© Barrett Edward Kirwan, MMV. All rights reserved.
The author hereby grants to MIT permission to reproduce and to distribute publicly paper and electronic
copies of this thesis document in whole or in part.
-
I1/_,
, I/
Signatureof Author.........................................................
.1...
...... . ...
Department of Economics
August 15, 2005
_s
Certifiedby........................................................
.....................
David H. Autor
Associate Profesor of Economics
A , r ofThesis Supervisor
Certified by .............................................................
,
....
T
Michael Greenstone
M Associate Professor of Economics
'?
-- 7 Thesis Supervisor
Accepted by ...............................................................................................................
Peter Temin
Elisha Gray II Professor of Economics
Chairperson, Department Committee on Graduate Students
MASSACHUSETTSINST'TE I
OF TECHNOLOGY
ARCHIVES
IFT T T 7, I
Z'
SE 3 U 20U5
LIBRARIES
I
Essays on U.S. Agricultural Policy: Subsidies, Crop Insurance, and
Environmental Auctions
by
Barrett Edward Kirwan
Submitted to the Department of Economics
on August 15, 2005 in partial fulfillment of the requirements
for the degree of Doctor of Philosophy in Economics
ABSTRACT
This thesis investigates the unintended consequences of government policy, specifically policy
meant to benefit agricultural producers. The first chapter asks how agricultural subsidies affect
farmland rental rates. Chapter two examines the relative effects of ad hoc disaster payments and
crop insurance, focusing on the potential to smooth income shocks, and the final chapter studies
the targeting efficiency of the Conservation Reserve Program.
Chapter one exploits exogenous variation in agricultural subsidies to measure the effects
on farmland rental rates. The primary finding is that landlords capture one-fifth of the marginal
subsidy dollar per acre. This finding stands in contrast to the standard assumption that landlords
immediately capture the entire subsidy. There is some evidence that the share of the subsidy
captured by landlords increases as the farmland rental market becomes more competitive. The
analysis also indicates that the lower effective price of land induces tenants to rent more land
such that they gain roughly $1 per dollar of subsidy. Taken together, the results suggest that
agricultural subsidies benefit farmers, as well as individuals that own agricultural land.
Chapter two exploits random weather variation and exogenous variation in crop
insurance take-up in order to investigate the relative effect of two forms of government-provided
disaster relief on farm failure rates. I find that disaster relief in the form of ad hoc disaster
payments slightly reduces the average farm failure rate, while average farm failure rates increase
under a crop insurance regime. The relative effect suggests that farm failure rates increase by 1.7
percentage points (about 30-percent) under the crop insurance regime. Excessively generous ad
hoc disaster payments and moral hazard provide possible explanations for these findings. These
findings suggest that government-provided
crop insurance plays an important role in farmer risk
management.
Chapter three estimates the premiums received by Conservation Reserve Program
participants above their reservation rents. Findings suggest that participants receive substantial
premiums, and the premiums have increased over time as uncertainty about the program has
decreased.
Thesis Supervisor: David H. Autor
Title: Associate Professor of Economics
Thesis Supervisor: Michael Greenstone
Title: 3M Associate Professor of Economics
3
4
Acknowledgements
The research presented in this thesis has benefited greatly from the guidance and mentoring of
my advisors. I am deeply grateful to David Autor. He has challenged me to think carefully with
his insightful suggestions and critiques. His kindness and encouragement have been equally
valuable. I also am extremely thankful to Michael Greenstone, whose clear vision and
encouragement motivated me from the beginning. His rigorous intellect helped focus the
research, and I hope the analyses reflect the wisdom he imparted. Jonathan Gruber provided the
foundation for this research through his teaching, his encouragement and his passionate example.
This research owes a great debt to his intellectual legacy.
Several others deserve special mention for their contributions and suggestions. Most
notably, conversations and collaboration with Michael Roberts of the U.S. Department of
Agriculture-Economic Research Service provided the impetus of this research. Jim Poterba
offered many helpful comments.
Access to the data used in Chapter 1 was generously facilitated by James Breuggen and
Jameson Burt of the U.S. Department of Agriculture-National Agricultural Statistics Service.
The George Schultz fund at MIT generously funded several trips to Washington, D.C. to
access the confidential data used in Chapter 1. I also wish to acknowledge financial support
from the National Science Foundation during my first three years of graduate school.
I am extremely grateful to my classmates at MIT from whom I learned far more than just
economics. Their camaraderie and friendship will be with me always. In particular, Mara,
Melissa, Manuel, and Cindy continue to be a source of inspiration. Gaining their friendship
made graduate school worthwhile.
I am truly thankful for the love and encouragement of my family. My parents have been
a wellspring of advice, encouragement, and prayers. Brock, Jennifer, Todd, and Whitney have
revealed the strength of the human spirit to me. Above all, I am profoundly grateful to my wife,
Jennifer, and to my children, Ella and Samuel. Their love, devotion, and strength are at the heart
of this thesis.
5
6
Contents
The Incidence of U.S. Agricultural Subsidies on Farmland Rental Rates
1
Introduction .......................................................
2
Literature Review................................................................................................
12
3
Institutional
15
3.1
3.2
4
Background
9
......................................................
Overview of the Farmland Rental Market
..
..................... 15
Agricultural Subsidy Policy.....................................................
16
3.2.1
The Federal Agricultural Improvement and Reform Act.......................... 18
Data .......................................................................................................................................
19
4.1
4.2
5
6
Sample Selection.......................................................................................................20
Variable Creation...............................
21
4.2.1
Dependent Variable................................................................................... 21
4.2.2
Independent Variables............................................................................... 22
Empirical Strategy-Identification ............................
23
5.1
Econometric Ideal .....................................................................................................
5.2
Unobserved Heterogeneity........................................................................................25
5.3
5.4
Simultaneity Bias ......................................................................................................
Expectation Error ......................................................................................................
26
27
Estimation and Results......................................................................................................... 29
6.1
Ordinary Least Squares........................................................
29
6.1.1
6.1.2
7
8
9
23
A Cross-Sectional Approach ........................................................
A New Approach .......................................................
29
30
6.2
Instrumental Variables.......................................................
6.3
The Heterogeneity of Rental Rate Incidence across Region and Farm Size............ 34
32
6.3.1
Resource Regions.......................................................
34
6.3.2
Sales Class .......................................................
35
Interpretation...........................................................................................................
Variable Factors of Production........................
Tenant's Incidence.........................................
10
Conclusion .........................................
11
Appendix...............................................................
36
37................
37
40
41
........................................
45
Adversity and the Propensity to Fail: The Impact of Disaster Payments and Multiple Peril
Crop Insurance on U.S. Farm Failure Rates
I
2
3
4
Introduction..............
........................................
Background......................................................................
63
.......67
2.1
Two Disaster-Risk Mitigation Programs .........................................
70
2.2
The Federal Crop Insurance Reform Act of 1994.
71
Empirical Strategy .........................................
Data ..........................................
4.1
...................................
72
75
Farm Failure Rates.................................................................................................... 75
7
4.2
D isaster Classification ........................ ....
4.2.1
Precipitation Indices ..................................................................................
4.2.2
5
Em pirical Results ..................................................................................................................
5.1
5.2
79
81
Overview................................
81
Randomized Experiments......................................................................................... 82
5.2.1
Randomization........................................................................................... 82
Disaster Payments ...................................................................................... 84
5.2.2
5.2.3
5.3
6
Disaster Classification............................
76
77
Crop Insurance ...........................................................................................
85
Relative Effectiveness.............................................................................................. 86
Interpretation and Conclusion ..............................................................................................
88
How Cost-Effective are Land Retirement Auctions? Estimating the Difference between
Payments and Willingness to Accept in the Conservation Reserve Program
1
Introduction .........................................................................................................................
107
2
Institutional Background..................................................................................................... 108
3
4
The CRP Bidding Mechanism ...............................
2.1
Sum m ary Statistics ..............................................................................................................
A M odel of Bidding Behavior ............................................................................................
4.1
5
6
Empirical Model ...................................
Results .........................................
Conclusion ..............................
109
110
112
115
117
118
8
Chapter
1
The Incidence of U.S. Agricultural Subsidies
on Farmland Rental Rates
1
Introduction
The primary goal of U.S. agricultural policy over the past century has been to increase farmers'
income. Since 1973, direct payments to agricultural producers have been a vital instrument for
supporting that goal'.
open question.
Whether agricultural subsidies actually benefit farmers, however, is an
Nearly half of all farmland in the United States is rented, almost all of it from
non-farmers. Agricultural subsidies may not benefit farmers if non-farmer landlords are able to
adjust rental rates to capture the subsidies paid to agricultural producers.
In the United States, agricultural subsidies are a significant transfer payment to farmers.
Among income support policies, subsidies rank among the highest in expenditures per recipient.
In 1999, the average subsidy was $17,561 per recipient household.
Compare this to $2,052
annually per recipient household in food stamps; an average total unemployment compensation
claim of $3,118; or $4,460 per recipient in annual benefits from SSI, an income support for the
needy aged, blind, and disabled.
The total amount spent on farm payments in 1999, $22.7
billion, rivals the spending of one of the largest transfer programs, the earned income tax credit,
which allocated $30.5 billion in credits to low-income tax filers. The size of the farm subsidy
program alone emphasizes the importance of understanding whether the stated policy goals are
being met.
See Orden, Paarlberg, and Roe (1997) for a history of agricultural policy.
9
Recent increased scrutiny of U.S. agricultural policy highlights the importance and
urgency of this issue. Media stories about Congressmen and sports and movie stars receiving
agricultural
subsidies
have
focused
the domestic
debate
on agricultural
policy 2 .
The
international debate has also focused on U.S. domestic agricultural policy due to record subsidy
payments from 1999 through 2001. This topic is playing a key role in the current Doha round of
World Trade Organization negotiations.
Despite the prevalence of rental agreements, and the importance of subsidies as a policy
measure, there is little empirical work examining the relationship between subsidies and
farmland rental rates. Without a clear understanding of the connection between subsidies and
rental rates farm policy might be severely misguided. Rented farmland is almost entirely owned
by non-farmers. In 1999 alone, 94 percent of landlords were non-farmers. Policy aimed at
benefiting farmers fails if non-farmer landlords extract the entire subsidy dollar.
This paper
estimates the extent to which the marginal subsidy dollar is reflected in rental rates. In an effort
to inform both the public debate and the academic literature, I quantify the effectiveness of
government subsidies as a policy instrument used to benefit farmers, and I utilize an incidence
measure to inform various assumptions in the literature surrounding farmland value
determination.
This paper investigates the relationship between farmland rental rates and agricultural
subsidies by analyzing a panel of farm-level production data. The data are from confidential
United States Department of Agriculture (USDA) - National Agricultural Statistical Service
(NASS) Census of Agriculture micro files from 1992 and 1997. The nature of these data
provides two identifying sources of variation.
2
The first comes from the differential acreage
For example, Elizabeth Becker. "Some Who Vote on Farm Subsidies Get Them as Well." The New York Times
Sept. 1, 2001, and Billy Heller. "$24M Baseballer Reaps Farm Aid." The New York Post Mar. 27, 2002.
10
enrollment in the subsidy program across farms. To illustrate, consider two otherwise identical
fields, one is completely enrolled in the subsidy program, while the other is not.
Under this
scenario any difference in rental rates will be due to differences in subsidies. The second source
of variation is due to a policy change in 1996, the Federal Agricultural Improvement and Reform
(FAIR) act, which exogenously changed subsidy rates. This provides a quasi-experimental
design that allows me to examine a change in subsidy rates that is plausibly uncorrelated with
other behavioral changes.
I employ several techniques to exploit these sources of variation and identify the effect of
subsidies on farmland rental rates. I use farm fixed effects and time-varying county fixed effects
to control for unobserved farm heterogeneity and regional shocks, and I exploit the exogeneity
imposed on subsidy payments by the 1996 legislation to overcome bias due to simultaneity.
I
use instrumental variables techniques to overcome possible expectation error that could bias the
estimated effect of subsidies on farmland rental rates by using the legislated, pre-determined
1997 subsidy payments as an instrumental variable to address expectation error in the 1992
subsidies.
The analysis finds that between 18 and 20 percent of the marginal subsidy dollar is
reflected in increased rental rates. This implies that of the $5.5 billion in farm subsidies in 1997,
at most $1.1 billion may have gone to non-farmers. Although the literature typically assumes
that every subsidy dollar is captured by landowners, i.e., full incidence (e.g., Chambers 1995),
these findings suggest that the standard assumptions may need to be re-evaluated to more
accurately reflect the factor market for land.
Based on these findings, economic theory suggests that subsidies may have effects
beyond the farmland rental market. Because farm subsidies are factor-specific subsidies to land,
11
the marginal subsidy dollar effectively lowers the rental rate by 80 cents. The lower rental price
of land induces substitution away from other variable inputs toward rental land, while the lower
marginal cost might result in greater output and increased variable input use. The ultimate effect
on other variable factors of production is an empirical question. The analysis finds a significant,
positive response of expenditure on nearly all variable inputs. Overall, the marginal per-acre
subsidy dollar increases per-acre expenditure by 35 cents.
The increased output ultimately results in a gain to tenant farmers. The analysis shows
that the net returns of tenant farmers increase one-for-one with the marginal subsidy dollar,
supporting the notion that ultimately the farm operator benefits from the marginal subsidy dollar.
The paper proceeds as follows.
Section 2 reviews the existing literature, noting the lack
of empirical work on subsidy incidence.
Section 3 details the institutional facts about the
farmland rental market and subsidy policy. Section 4 describes the data. Section 5 lays out the
empirical strategy employed in this investigation, emphasizing the identifying assumptions.
Section 6 presents the incidence on landlords and provides evidence of the robustness of the
estimated incidence.
Section 7 provides a plausible explanation of the results.
Section 8
examines the subsidy's effect on non-land factors of production. Section 9 provides evidence
that the incidence on tenants is complementary to the incidence on landlords. Finally, section 10
interprets the findings in light of existing assumptions and policy objectives and suggests
directions for future research.
2
Literature Review
This paper is one of the first to examine the incidence of subsidies on farmland rental rates, but it
has a foundation in a much broader literature on the determinants of farmland values. Alston
12
(1986) and Robison,et al. (1985), among others, examined the asset value of land as the present
discounted value of the stream of returns to land, which they represented with the cash rental
rate. Melichar (1979) noted that the discount rate in such models is comprised of a discount rate
and the expected growth rate of the stream of returns.
More recently Weersink, et al. (1999)
disaggregated the stream of returns to farmland into the return from farming and the subsidies to
land.
They estimated separate discount rates for each stream, finding that subsidies are
discounted less heavily than the returns to production.
Virtually all of the research on subsidy incidence has examined the relationship between
subsidies and land values.
This investigation has taken many approaches.
Morehart, et al.
(2001) and Weersink, et al. (1999) are characteristic of those who use a present discounted value
of the stream of subsidies in their analysis.
Shoemaker, et al. (1989) approach the problem by
using a computable general equilibrium model in order to ascertain the extent to which subsidies
are capitalized into land values.
Both of these approaches suffer from assuming perfect
incidence and essentially examine only the process of capitalization.
represent a separate approach using pooled cross-sections
Barnard, et al. (1996)
in order to assess the degree of
capitalization. Although this approach does not assume perfect incidence, the findings cannot be
separated into an incidence estimate and discount and expected growth rate estimates.
The approach taken in this paper contributes to the discussion by explicitly separating the
incidence estimate from the discount rate. By focusing on farmland rental rates, rather than land
value, I disentangle the incidence question from the question of expectations over future policy.
This not only contributes to the current knowledge about who benefits from current subsidies and
to what degree, but it also lays the groundwork
to more carefully address the questions
concerning the expected growth rate of government subsidies.
13
Researchers have recently begun examining the effects of subsidies on rental rates (Lence
and Mishra 2003; Goodwin, et al. 2004).
A prominent feature of this research is the implicit
assumption that once one controls for observable characteristics, no unobservables remain that
are correlated with government payments. This assumption is the 'selection on observables'
assumption. To illustrate the hazard of relying on this assumption, consider the confounding
effect of the inherent productivity of land due to complex interactions among soil characteristics
and climate. This characteristic is an unobserved determinant of land values that is correlated
with subsidies because subsidies are a function of historic yield.
In the presence of this
unobservable characteristic, the coefficient on government payments will capture both the effect
of subsidies and the effect of unobserved productivity. Since both effects are positively related
to land values, the parameter of interest will be larger than it would be if it identified the true
effect of subsidies. This paper overcomes the problem by using a farm fixed effect to purge
permanent, unobserved characteristics that determine rental rates.3
In contrast, this study uses disaggregated panel data to control for unobserved
heterogeneity and county-specific
secular trends.
The timing of the data allows me to exploit
exogenous variation in subsidies from the mid-1990s policy change. The policy change acts as a
'quasi-experiment' by changing the subsidy in a way that was unanticipated and uncorrelated
with the farmer's behavior, thereby reinforcing the identification of the true effect of subsidies
on farmland rental rates.
3 Interestingly, Hoch (1958 and 1962) and Mundlak (1961) developed the fixed effects model to account for
unobserved heterogeneity in estimating agricultural production functions.
14
3
Institutional Background
3.1
Overview of the farmland rental market
Renting farmland is a common practice in U.S. agriculture, where more than 45 percent of the
917-million farmland acres are rented (USDA 2001b). A typical tenant rents 65 percent of the
land he farms, paying either in cash or in shares of production. Sixty percent of rented farmland
is paid for with cash, 24 percent with shares of production, and 11 percent with a cash/share
combination.
Those who cash lease pay $60 per acre on average. The average rental rate among
those who grow subsidized crops is slightly below the national average at $50 per acre.
The rental market is characterized by short-term contracts in long-term relationships.
Rental contracts commonly take the form of year-to-year "handshake" agreements.
formal, written rental contracts exist they also generally last one year.
When
In spite of frequent
renewal, these arrangements are typically found in long-term landlord-tenant relationships.
Allen and Lueck (1992) report an average landlord-tenant relationship length of 11.5 years in
Nebraska and South Dakota, and Sotomeyer, et al. (2000) report 11.3-year landlord-tenant
relationships in Illinois.
by March 1St.
The parties enter into the rental agreements early in the year, typically
A rental rate negotiated in the spring is based on the expected returns from
farming during the upcoming year. Among the expected returns are government payments,
which typically are made after the harvest. The mistiming between the setting of the rental rate
and the realization of government payments results in forecast error. The forecast error involved
in the agreed upon rental rate may serve to confound the relationship between subsidies and
rental rates if not adequately addressed. Below I detail the instrumental variables (IV) strategy
used to address the issue.
15
The incidence measures the division of subsidy benefits between the tenant and the
landowner. Whenever a land-owning farmer rents out land, the incidence question asks whether
the tenant or land-owning farmer benefits from the policy.
To the extent that rented land is
owned by a non-farmer, the incidence denotes how much of each dollar leaves agriculture.
In
the United States, non-farmers own 94 percent of the rented land, or 340 million acres-twice
the size of Texas (USDA, NASS 2001b).
This makes the issue of subsidy incidence a
particularly poignant one if agricultural subsidies are meant to benefit farmers. In effect, one can
view the incidence as describing the portion of each subsidy dollar not going to a farmer.
3.2
Agricultural Subsidy Policy
A key component of domestic agricultural policy is the support of farmers' income. The U.S.
farmers' income is supported by a two-tier system: 1) price supports and 2) income supports, i.e.,
subsidies. Price supports have existed since the farm program began in the 1930s. For most of
the century, they were the chief mechanism for transferring money to the agricultural sector.
Prices are supported by the government's promise to purchase the commodity at a predetermined
price.
Although the government once took direct possession of these purchases, today it
generally augments the price the farmer receives from sales to a third party.
Income supports became a distinct objective with the introduction of production subsidies
in 1973. Since their inception, subsidies have been calculated in a consistent way. An acre of
land used to grow one of seven crops-wheat, corn, sorghum, barley, oats, rice, and cottoncould qualify to be subsidized.
A qualifying acre receives a subsidy equal to a national subsidy
rate times an output yield assigned to that acre. Between 1985 and 2002 the assigned yield,
16
called a program yield, approximately equaled the farm-average yield between 1980 and 19854.
An acre that qualifies for a subsidy is called a base acre. A farm's total crop-specific subsidy is
the product of the subsidy rate, the program yield, and the total number of base acres for that
crop. That is,
subsidy = subsidy rate * program yield * base acres.
Prior to 1996, the subsidy rate was designed in a way that made subsidy payments
counter-cyclical. This was done by calculating the crop-specific deficiency rate as the difference
between a legislated target price and the national average price received. As the national
average price fell, the difference between it and the target price increased, causing the subsidy to
increase.
It is important to note that program yield and base acreage are specifically tied to a
qualifying acre of land. As an acre changes ownership, these parameters are transferred with the
land. As such, agricultural subsidies are factor-specific subsidies to land. However, not all land
planted to a subsidized crop qualified for the subsidy because subsidies were closely associated
with supply control measures through the base acreage allotment. Prior to 1996, total subsidized
acres could not be greater than an average of the number of acres planted to that crop over the
previous five years.
If a farmer's planted acreage exceeded his base, he was disqualified from
receiving subsidies for that year, although his increased planting would enter into a five-year
moving average when calculating the farm's total base acres in subsequent years.
This feature
provided farmers with some means to affect current and future subsidies through current
behavior. (See Duffy, et al. (1995) for an analysis of the effect of base acres on land values.)
" Prior to 1985 the program yield was the 5-year Olympic average of the farm's annual yield. The current program
yield is the average of the 1980 - 1985 program yields.
17
As an additional supply control measure, farmers were required to remove a designated
proportion of their base acres from production each year and leave it fallow. This program was
called the Acreage Reduction Program (ARP), but it was commonly referred to as the "set-aside"
program. The Department of Agriculture annually established the set-aside proportion, which
could be as high as 25 percent.
In 1992, producers were required to leave 5 percent of their
wheat, corn, sorghum, and barley base acres fallow; 10 percent of their cotton base acres fallow;
and none of their rice or oats base acres fallow.
These planting constraints and supply control measures imposed costs that kept many
producers from participating in the subsidy program. Before 1996, participation rates usually
ranged from 60 to 80 percent of qualified acres5 . After the 1996 policy change, which removed
the costly constraints, nearly all qualified acres were enrolled in the farm program. 6
The
identification strategy relies, in part, on the cross-sectional variation in acres participating in the
subsidy program.
3.2.1
The Federal Agricultural Improvement and Reform Act
The 1990s saw an end to the supply control measures and the completion of the process to
decouple subsidies from contemporaneous production.
In a dramatic move, the Federal
Agricultural Improvement and Reform (FAIR) Act froze base acres at their 1995 level and
removed all planting restrictions, including set-aside requirements. With the exception of certain
fruits and vegetables, producers were given complete planting flexibility, while they still
received subsidies based on their 1985 program yield and their 1995 acreage base. At the same
time, the subsidy rate was no longer conditioned on the commodity's price, as it was before
5
Land qualified to receive subsidies for oats is the single exception, where the participation has been between 20
and 40 percent of qualified acres
6
This variation will be used in future work by the author to measure the costs of program participation.
18
1996, but it was exogenously determined by the policy. I exploit this innovation to identify the
effect of subsidies on farmland rental rates.
Two aspects of this reform are significant to the analysis below. First, by freezing the
farm-specific program parameters (program yield and base acres), the policy divorced farmer
behavior from the subsidy. As described below, this unique feature of the legislation allows me
to control for simultaneity that would otherwise serve to confound the incidence estimates.
Second, the 1996 legislation removed the uncertainty of subsidy payments by establishing a
schedule of annual payments from 1996 to 2002.
I exploit the post-1995 known subsidy
payments as an IV for pre-1995 ex ante uncertain subsidy payments. Failure to address this issue
would result in attenuation bias of the incidence estimate.
4
Data
The primary source of data used in the analysis described below is the U.S. Census of
Agriculture, a quinquennial census of those who produce at least $1,000 of agricultural goods.
These data are confidential micro files accessed under an agreement with the USDA Economic
Research Service and the USDA National Agricultural Statistics Service (NASS). The data are
available at NASS in Washington, DC 7.
The Census of Agriculture contains farm-level information such as total acres farmed,
total acres rented, and total government payments received. It also contains detailed information
on the number of acres harvested and the total production of each crop. The value of production
is reported for 13 crop groups and for all livestock. The census also collects information on the
corporate structure of the operation as well as demographic information on the primary operator.
7 Any
interpretations
and conclusions derived from the data represent the author's views and not necessarily those of
NASS.
19
Additionally, approximately one in three farms receives the census's long form which
requests further information on the production and financial structure of the operation8
Recipients are asked to report production expenses in 15 cost groups, including cash rent paid on
land and buildings. Recipients are further asked to report the estimated value of the farm's land
and buildings.
The long form also collects information on the type and amount of equipment
used, the number of workers hired, and the use of agricultural chemicals and fertilizers.
4.1
Sample Selection
Two years of the Census of Agriculture, 1992 and 1997, are used to create a balanced panel of
farming operations. Each year has roughly 1.6 million respondents. Although the total number
of respondents remains roughly equal over the two years, this masks a great deal of turnover (see
Hoppe & Korb, 2001). There are 1,040,305 farm operations observed in both years of the data 9 .
The population of interest is all farms that could potentially receive subsidies in the base year. I
approximate that population by focusing on the 437,979 farms that grew any one of the seven
program crops in the base year. The first column of Table 1 contains the summary statistics for
this population.
Because I am concerned with the financial structure of farming, I limit the
sample to those who returned the long-form in both years, yielding 100,936 observations in each
year.
Since rental rates are the outcome of primary interest, I include only those who report
paying cash rent in both years' ° .
I remove farms that are primarily renting structures by
The Census attempts to collect financial information from all farms with sales over $500,000, and randomly
8
samples from the remaining farms.
9 This is not indicative of the failure rate of farms, but rather the incomplete response rate to the census. About 85%
of farm operations respond.
10
Renting in 1997 is a potentially endogenous decision since the propensity to rent farmland in 1997 might be
influenced by the 1996 policy change. This source of endogeneity could cause downward bias if farms experiencing
a high incidence stopped renting, but those experiencing low incidence continued to rent. I investigate this using a
two-step Heckman sample selection correction. A well-known test of sample selection bias is a simple significance
20
dropping those with imputed rental rates greater than two standard deviations above the mean"
The cutoff is $470, which is conservative considering the highest response to the 1997 June
Agricultural Survey, a nationwide survey that specifically asked for cropland cash rental rates,
was $448.
The final analysis sample consists of 58,302 farms observed over two years.
Table 1 contains the summary statistics of the estimation sample and the population of
farms growing subsidized crops. The median farmer in this sample has operated a farm for 23
years, compared to 22 years in the population.
The average net returns per acre to a farmer in the
analysis are $92.50, a bit less than the average in the full sample of those reporting net returns.
Of those receiving subsidies in the sample the average total government payment is $16,614.
The median is $10,000.
This suggests that a few farms receive large subsidies. The sample
statistics of key variables can be found in Table 1.
4.2
Variable Creation
4.2.1 Dependent Variable
The Census of Agriculture does not report the per-acre rental rate, however respondents do
report the total amount paid in cash rent. The total acres rented on a cash, share, or free basis
also are reported.
From these two variables, I create the per-acre rental rate by dividing total
cash rent by total acres rented. Admittedly, the resulting rental rate will be too small for farms
that cash rent part of the land and share rent another part. The resulting measurement error is not
classical measurement error. Since the calculated rental rate is less than or equal to the true cash
rental rate, the expected value of the measurement error must be greater than zero. As long as
test of the coefficient on the inverse Mills ratio. This coefficient is statistically insignificant, and the null hypothesis
of no sample selection bias cannot be rejected. See Appendix table 1 for the coefficient estimates.
'' This drops less than 1% of the sample.
21
the non-classical measurement error is uncorrelated with the regressors, only the intercept will be
confounded. However, if the measurement error is positively correlated with the magnitude of
government payments, the estimated effect of government payments on rental rates will be
biased downward'2 .
I investigate this source of bias using the microfiles of the 1999 Agricultural Economics
and Land Ownership Survey (AELOS), a follow-up survey to the 1997 Census of Agriculture
which asked for detailed information about the type of lease arrangement, the total cash rent
paid, and the value of the share of production paid as rent.
estimate of the non-classical measurement
Using these data I construct an
error, and I estimate the relationship between the
measurement error and government payments'l3. When controlling for county-level unobserved
heterogeneity, I find no relationship between government payments and the measurement error,
implying no bias in the primary analysis.
4.2.2
Independent Variables
The subsidy variable is constructed from the reported government payments variables. Every
agricultural producer is asked to report non-price support payments received from the
government. Producers are asked to report both total payments received and payments received
from the Conservation
Reserve Program (CRP).
By subtracting
CRP payments from the
reported total payments, I construct an approximate measure of subsidy receipts. In 1992,
subsidies accounted for 68 percent of direct government payments net of Conservation Reserve
Program payments. The remaining 32 percent are from disaster relief (17 percent) and other
programs (USDA, NASS, 1996). In 1997, even fewer total payments are attributable to non-
12
13
Appendix 1 formally lays out the potential for bias and details the analysis that is summarized below.
The estimates are reported in Appendix Table 1.
22
subsidy sources, with only 4 percent of direct government payments going to neither subsidies
nor CRP payments (USDA, NASS, 2001).
The remaining covariates are constructed directly from variables contained in the Census
of Agriculture. All regressors are measured on a per-acre basis, using total farmland acres in the
denominator.
5
Empirical Strategy - Identification
Here I lay out the obstacles that must be overcome in order to identify the effect of government
subsidies on farmland rental rates. I begin by specifying a simple rental rate equation that, under
ideal circumstances, yields the incidence measure. In the face of a less than ideal experiment, I
lay out the modifications necessary to identify the parameters of the conditional expectation
function. After setting out a fixed-effect estimation equation, I detail the instrumental variables
procedure necessary to overcome attenuation bias in the econometric model. The resulting IV
model overcomes the obstacles separating the real-world situation from the econometric ideal.
5.1
Econometric Ideal
If subsidies were randomly assigned, then the parameter identified by a regression of the rental
rate on subsidy per acre would be the proportion of each extra subsidy dollar per acre reflected in
higher rental rates. Denote the rent on acre i at time t by ri, and let gi, be the amount of subsidy
payments associated with acre i. Then we may write
(1)
where
i, is the residual.
ri = a + gi Y
+i,
Random assignment identifies y* as the incidence of agricultural
subsidies on farmland rental rates.
23
However, subsidies are not randomly assigned, git is most likely correlated with
it, and
the resulting OLS estimate of yin equation (1) will be biased. The subsidy is a function of yield
and crop choice; hence, it is an endogenous variable reflecting the characteristics of the land and
the producer's behavior. This endogeneity problem can be overcome by addressing three issues:
unobserved heterogeneity, simultaneity, and farmer's expectation error due to the mistiming of
rental contracts and subsidy payments 14 . The innovation of my analysis is to address all three
problems and identify the parameter of interest in the conditional expectation function. First, I
use farm fixed effects to control for unobserved
heterogeneity,
such as different
land
characteristics and entrepreneurial skill. Second, I control for simultaneity by exploiting a
unique aspect of the policy change that divorced producer behavior from subsidy payments.
Finally, I am able to overcome the expectation error by using an IV strategy.
The best instrumental variables for agricultural subsidies are the program parameters
(program yield and base acres) underlying the subsidies.
These parameters fulfill the
requirements of instrumental variables because, as detailed below, within the fixed effects model
they are highly correlated with the subsidy and plausibly uncorrelated with the error term. Data
on the program parameters are unavailable at the farm level, but two farm level variables closely
reflect the program parameters.
They are the 1992 set-aside acres and the 1997 subsidy level.
The 1992 acres that were set aside as part of the ARP are a linear function of the base acres, and
the 1997 subsidy is a known, deterministic function of the underlying program parameters.
These features allow me to use the 1992 set-aside acres and the 1997 level of government
payments as instrumental variables for the 1992-1997 change in government payments.
14 Measurement
error will also be a problem as only 68% of government payments in 1992 were subsidies. This is
much less of a problem in 1997 when 98% of government payments were subsidies. Instrumental variables
techniques will address this issue.
24
5.2
Unobserved Heterogeneity
Many farm characteristics cannot be observed by the econometrician, yet they are influential to
both subsidies and farmland rental rates. Among these are farm-level soil properties and farmer
human capital and entrepreneurial skill. Transient shocks, such as drought or pests, also may
affect rental rates and government subsidies. Typical analyses are performed at the county or
regional level, under the assumption of farm homogeneity within the geographic unit of
observation. However, differences in farm size, structure, and productivity within a county serve
to confound the conventional analysis'5
The unique nature of the data allows me to control for permanent farm-level
characteristics that cause yto be inconsistent.
One source of bias comes through the unobserved
characteristics, such as farm productivity, that positively influence both subsidies and rental
rates. This positive correlation between government payments and the unobserved factors that
influence productivity will result in an upward bias to incidence estimates and confound y as a
measure of the effect of subsidies on rental rates. Including farm and time-varying county fixed
effects allows me to overcome this source of bias. Rewriting equation (1) using ft as the fixed
effect for farm i in year t yields:
(2)
r,=a +gi, + i, +f +C,+Ci,
The parameter C, is the time-varying county effect which allows for shocks, such as weather or
pests, that impact everyone within a localized region. Xi, is a vector of observable covariates
such as yield, selection and production of crops, occurrence of irrigation, farm size, sales, and
costs.
"5Current work by the author examines the information lost due to aggregation.
25
Controlling for unobserved heterogeneity in this way has the advantage of avoiding the
inherently nonlinear relationships between soil characteristics and productivity. Because of this
nonlinear relationship, even explicitly using soil characteristics as controls cannot overcome the
omitted variable bias-a
problem other researchers have faced (e.g. Moss, et al. 2002). This
method of conditioning on unobserved farm-level characteristics also overcomes the bias
inherent in studies on more aggregate units of observation (Lence and Mishra 2003).
The estimating equation used in this study is obtained from equation (2) by first
differencing the data to absorb the farm effect, resulting in
(3)
Ar = C + Agiy + Xi92 + AEi .
The first difference of the control variables is not included because the 1997 level of these
variables are potential outcomes influenced by the exogenous subsidy change. Instead, the 1992
values of these variables are included in the estimating equation.
In a panel with t=2, the
coefficients estimated from first difference data will be identical to those obtained by including
the individual fixed effects.
5.3
Simultaneity Bias
Simultaneity bias arises when at least one of the explanatory variables is determined
simultaneously along with the response variable. Prior to the 1996 FAIR Act, output prices
played a role in determining both subsidies and rental rates. When expected prices were high,
rental rates were high and expected subsidies were low. Thus, simultaneity caused a negative
relationship between the subsidy and the rental rate.
A unique feature of the 1996 FAIR Act allows me to exploit the policy change itself to
overcome the simultaneity problem. The 1996 legislation was titled the "Freedom to Farm Act"
26
because it lifted planting restrictions and divorced the subsidy from prices and producers'
behavior.
Such a divorce meant that prices no longer determined
the subsidy rate, and
simultaneity bias ceased to be a problem for the incidence estimates. Hence, the policy provides
an exogenous change in subsidy rates, and its structure eliminates the obstacle to identification
caused by simultaneity bias.
5.4
Expectation Error
Without the obstacle of simultaneity bias, the remaining problem to be addressed is expectation
error, which causes attenuation bias. As detailed earlier (see Section 3.1), rental rates are set
according to expected receipts, including expected subsidy payments. Prior to the 1996 FAIR
Act, subsidy payments were conditioned on the market price and thus were unknown until after
the harvest, while rental rates were agreed upon before planting in the spring. To see the effects
of this mistiming on the incidence parameter, rewrite equation (2) using the expected government
payments, gi*,
r,=a +gy +f +C,+e,.
(4)
Actual government payments will equal the expected government payment and an expectation
error,
g =g
(5)
+
Substituting for expected subsidy receipts in equation (4) yields
(6)
'
= a + g,
+
f
+
C, + , - ci .
The expectation error becomes part of the error term in the estimating equation. Assuming the
expected subsidy and the expectation error are uncorrelated, i.e. Cov(g, cE/)=OVt,s, implies
that realized government payments, git, are correlated with the error term in equation (6). The
27
effect on the coefficient of interest is the same as classical errors in variables, namely attenuation
bias.
The 1996 FAIR Act reduces the complexity of the problem by eliminating expectation
error in 1997. Recall that in 1996 the subsidy rates were exogenously predetermined for the next
seven years. Because of this feature of the legislation, there was no expectation error in 1997.
The expected government payments for 1997 and 1992, respectively, are:
(7)
gi97 = gi97
(8)
g92 = gi92 -92 i .
*gg
Substituting (7) and (8) into equation (2) and first differencing results in
Ar = C + Agiy + Ae i - Ei92.
(9)
An adequate instrument is correlated with the change in government subsidies and
uncorrelated with the composite error term in equation (9).
requirements,
the 1992 set-aside acres, denoted as
sai92,
Two variables meet these
and the 1997 subsidy level.
Both
variables are assumed to be strictly exogenous. That is, conditional on the fixed effects, sai92and
gi97 are uncorrelated with both i92 and i97. Thus, they are uncorrelated with zAi. Furthermore,
both variables are uncorrelated with the second error term,
ig9. The 1992 set-aside acres are
proportional to the base acres, and are known when rental rates are set. Thus, under rational
expectations, the 1992 set-aside acreage is uncorrelated with the expectation error. The 1997
subsidy level is uncorrelated with the second error term due to the absence of expectation error in
1997 which allows one to write the orthogonality condition asE(i
92gi97)=
0. This condition
holds if the subsidy shock in 1992 contained no information for the expected subsidy in 1997, a
reasonable assumption.
28
Since the 1997 level of government payments is uncorrelated with the composite error, it
is therefore a good instrumental variable insofar as it is correlated with the 1992-1997 change in
government payments. The top panel of Table 4 contains the results of the following first stage
equation of a two-stage least squares estimation strategy:
(10)
Ag
i = C + gi97 5 + sai92 ; + u i
The coefficient of determination and the F-statistics are very high for both instruments, satisfying
the requirement that the instruments are correlated with the endogenous variable.
6
Estimation and Results
6.1
Ordinary Least Squares
6.1.1
A Cross-Sectional Approach
The approach taken by this paper has the advantage of controlling for unobserved characteristics
of the farm, such as the operators entrepreneurial
skill and the productivity
of the land.
Estimates that fail to account for these characteristics likely suffer an upward bias as productivity
and skill are probably positively correlated with the subsidy. Subsidies are a direct function of
the productivity
of the land, and more skillful farm operators may better understand the
complexities involved in maximizing subsidy payments.
Estimation methods that fail to
adequately control for these variables will attribute rental rate variation to the subsidy that should
be attributed to productivity and skill.
I explore the consequences of unobserved heterogeneity by ignoring the panel nature of
the data and estimating the incidence in each year separately and in the pooled cross-section.
Table 2 contains the results of this exercise.
Panel A reports the estimates for 1992, panel B
29
reports the estimates for 1997, and panel C reports the estimates obtained by pooling the two
cross-sections. Column one contains the results of a bivariate regression of the rental rate on the
per-acre subsidy, while column two includes as covariates the proportion of the farm planted to
13 different crop groups, the output yield of 8 crops, sales, variable factor expenditures, farm
size, and the proportion irrigated.
It is noteworthy that in 1992, the incidence estimate found
from the simple bivariate relationship is far below perfect incidence.
In fact, the estimate in
column one suggests that only 46 cents of the marginal subsidy dollar is reflected in rental rates.
As controls are added, the incidence estimates change dramatically, falling nearly fifty percent in
panels B and C, with a 30 percent drop in panel A.
In such a setting, county fixed effects may adequately control for unobserved
heterogeneity if farms within a county are sufficiently homogeneous. Columns three and four
present the incidence estimates when a county fixed effect is included. The specification
reported in column three includes no covariates, and it yields incidence estimates very similar to
those in column 2. Column four includes covariates, and the incidence estimate declines even
more to 0.23 in the 1992 cross-section, 0.39 in the 1997 cross-section, and 0.26 in the pooled
cross-sections.
A notable feature of Table 2 is the extensive change in the incidence estimate as more
controls are added. This parameter instability causes one to wonder how much bias continues to
exist because of other unobservable and excluded covariates.
6.1.2
A New Approach
Table 3 contains the incidence estimates when I control for unobserved heterogeneity, as
delineated in equation (3) above. Columns one and two report the coefficient from a random and
30
fixed effects regression respectively.
The random effects regression will be consistent and
efficient if the unobserved qualities of the farm are uncorrelated with the regressor (here per-acre
government payments).
However, the random effects model will be inconsistent if the farm
effects are correlated with the regressors. Performing a Hausman test on the appropriateness of
the random effects model effectively tests whether the unobserved farm characteristics are in fact
biasing the estimate. The Hausman test statistic is reported at the bottom of columns one and
two in Table 3. It soundly rejects the random effects model, suggesting that one should be
concerned with unobserved heterogeneity when seeking a consistent estimate of the incidence
parameter.
Columns two and three of Table 3 report the results of the fixed effects model including
farm and time-varying county effects. Column two contains no covariates, while column three
includes the 1992 level of the covariates listed above. As noted earlier, the 1997 values of these
covariates are potential outcomes affected by the changing subsidy and their inclusion in this
specification would serve to confound the incidence estimate. Compared to the estimates in
Table 2, the incidence estimate is lower once one accounts for unobserved farm characteristics,
signifying a decrease in the upward bias. Another important feature of columns two and three is
the stability of the incidence estimates.
Once farm fixed effects are included, additional
covariates do little to change the estimate, suggesting that farm fixed effects account well for
potential sources of bias.
The most notable characteristic of Table 3 is the low incidence estimate; the estimated
incidence is only 0.18. In other words, 18 cents of the marginal subsidy dollar is reflected in
higher rental rates. As discussed earlier, conventional wisdom and economic theory suggest that
the land owner should capture all of the Ricardian rents, including the subsidy. The evidence
31
presented here suggests that the truth lies far from perfect incidence; the landlord is only able to
capture 18 cents of the marginal subsidy dollar.
One may be concerned that this estimate suffers from attenuation bias caused by
measurement error or expectation error. The next section details the instrumental variables
strategy used to account for these sources of bias and demonstrates that the conclusions reached
above are essentially unchanged.
6.2
Instrumental Variables
As discussed earlier, the measurement error and expectation error concerns apply to the 1992
government payments and can be addressed with instrumental variables. The ideal instruments
for 1992 government payments would be the farm-specific subsidy parameters: program yield
and base acres.
These parameters are known in advance, are highly correlated with actual
subsidy payments, and are uncorrelated with the idiosyncratic shocks to prices that ultimately
determine subsidy payments. Thus, program yield and base acres are good instruments because
they are correlated with the realized subsidy payment and uncorrelated with shocks that
contribute to the expectation error.
Unfortunately, data on program yields and base acres are unavailable.
In order to closely
approximate the ideal instruments and stay within the constraints on data availability one must
look to linear functions of these variables, such as set-aside acres and 1997 government
payments.
Since set-aside acres are an exogenously
determined proportion of base acres they
provide a good instrument for the 1992 subsidy level. As noted earlier, base acres are known
when the rental contracts are agreed upon, thus they are highly correlated with the expected
32
subsidy.
The 1997 subsidy payments are a potentially good instrument because in 1997 the
subsidy on an acre of qualified land equaled the 1992 program yield multiplied by the 1997
subsidy rate. The 1997 subsidy should thus be highly correlated with the ideal instrument, the
1992 program parameters.
The first panel of Table 4 reports the first stage of a two-stage least squares estimation
strategy. Column one reports the coefficients and the F-statistic when no covariates are included
in the fixed effects specification.
Both instruments are significant predictors of the change in
subsidies. The F-statistic is 20,011.
Stock and Staiger (1997) have suggested that strong
instruments have an F-statistic greater than five. These instruments easily meet the criteria.
The second panel of Table 4 reports the IV estimates. The IV estimates, 0.186 and 0.214,
are slightly higher than the OLS estimates, revealing some attenuation.
However, the small
difference between the OLS and IV estimates suggests that expectation error is not a large
concern in this investigation.
By accounting for the potential sources of bias, namely unobserved heterogeneity,
simultaneity bias, and expectation error, I have found that the landlord captures only 20 cents of
the marginal subsidy dollar. Economic theory predicts that if land is the only specific factor of
production and if markets are perfectly competitive then the landlord will capture the entire
marginal subsidy dollar. These findings call into question the validity of these assumptions in
the short run.
33
6.3
The Heterogeneity of Rental Rate Incidence across Region and Farm
Size
One might be concerned that the results presented mask variation in response across region or
the size of the farm operation.
Regions within the U.S. differ substantially in the crops grown
and the predominant lease contract type.
Noting that each crop is subsidized separately, one
might worry that the incidence differs according to crop and subsidy regime.
also influence the size of the incidence.
Farm size might
For instance, perhaps large farms are better able to
negotiate for lower rental rates, and they are driving the relatively low incidence.
I explore the
robustness of the results by estimating the incidence separately for different regions and different
farm sizes.
6.3.1 Resource Regions
Table 5 reports the regional'
6
mean and median rental rates and per-acre subsidies for 1992 and
1997. This Table highlights the heterogeneity in both the levels and in the 1992-1997 changes of
these variables.
Five of the nine regions experienced a decline in mean and median rental rates.
While the average per-acre subsidy almost uniformly declined, five of the regions experience an
increase in the median per-acre subsidy.
Table 6 presents the OLS and IV incidence estimates by resource region. A prominent
feature across resource regions is the stability of the incidence estimate, which generally ranges
from 0.17 to 0.30 with a couple of zeros.
The incidence is somewhat higher in the Heartland
region, where subsidized crops are a significant share of output. Regions with fewer farms and
less subsidized crop production generally have lower incidence, sometimes not significantly
16
The USDA has established 9 resource regions corresponding to predominant crop mix and farming practices.
34
different than zero, but nearly always significantly different than one.
In spite of regional
differences highlighted in Table 5, the incidence estimates are all relatively similar; no one
region seems to be driving the results.
6.3.2 Sales Class
Another concern could be that the incidence differs by farm size. Large farms might be able to
negotiate lower rental rates and thereby keep a larger share of the subsidy. Alternatively, small
farms might be better acquainted with the landlord and hence receive a more favorable rental
rate. In the analysis below I define farm size to be gross sales, and I adopt the classification
system espoused by the USDA. 17
Table 7 reports the summary statistics for rental rates and per-acre subsidies by sales
class.
Interestingly, both rental rates and subsidies monotonically
increase with sales class,
presumably reflecting the higher productivity of the land farmed by larger farms.
Columns one and two of Table 7 report the estimated subsidy incidence when the data are
treated as pooled cross-sections and the panel nature of the data is ignored. These columns reveal
a significant degree of variation in the estimated incidence across farm size when one fails to
control for the unobserved characteristics of the farm. Notably, the incidence estimates are
smallest for the smallest and largest farms, suggesting that the unobserved characteristics are more
highly correlated with size for these two sales classes than for the others.
estimate once controls are added is also noteworthy.
The change in the
The relative instability of these estimates
supports the position that the unobserved farm characteristics
may play a role in biasing the
incidence estimate.
17For example, see http://www.ers.usda.gov/Briefing/FarmStructure/Gallery/farmsbyconstantdollars.htm
35
Columns
three
through
six report the OLS and IV estimates
once unobserved
heterogeneity is accounted for. The incidence estimates do not significantly differ across farm
sizes, ultimately settling around 0.2. These results give further confidence in the twenty percent
incidence reported above.
7
Interpretation
Economic theory predicts, and economists have long held, that under competitive markets,
incidence is perfect and landlords are able to extract the entire marginal subsidy dollar.
Intuitively, a subsidized parcel of farmland should result in competition among renters to secure
the subsidy.
Such competition will result in an increased rental rate as potential tenants bid
against each other until ultimately the price of that parcel of land fully reflects the return from
the subsidy.
The evidence presented here suggests that this does not always happen.
In the
short-run, competition only serves to increase the rental rate by 20 cents. A possible explanation
of this fact is imperfect rental markets.
In a market with many landlords and few renters, the
landlords may implicitly share the subsidy dollar in an attempt to attract tenants.
In order to examine this hypothesis I create five measures of rental market concentration.
The first measure is the proportion of farmers in a county who rent some land, and the second is
the proportion of farmland in a county that is rented. The next measure is an approximation of
the tenant-landlord ratio within a county. This measure is an approximation because each farmer
reports the number of landlords from whom they rent land, but there is no way to tell whether a
single landlord rents to multiple tenants. This approximate tenant-landlord ratio will be lower
than the actual ratio.
Finally, I calculate two Herfindahl indexes to measure rental market
concentration. One Herfindahl defines market share over total county rental expenditure, the
36
other defines market share over the total number of rented acres. I interact these measures with
the change in government payments in order to determine whether the marginal effect of
subsidies changes as rental markets become less concentrated.
The results from this exercise are found in Table 9. If rental market imperfections do not
play a role, I should see no relationship between the incidence and rental market concentration.
These results suggest that I can safely reject the null hypothesis that the subsidy incidence is not
related to rental market concentration. Nearly every measure of concentration returns a sign and
magnitude that is consistent with the alternative hypothesis that subsidy incidence decreases with
rental market concentration. For instance, the first row of Table 9 indicates that if all farms in a
county rented land the incidence would increase to about 25 percent. These findings suggest that
rental market imperfection is a plausible explanation for the incidence finding reported above.
8
Variable Factors of Production
Agricultural subsidies are defacto specific subsidies to land. As such, they serve to reduce the
rental price of land. A 20 percent subsidy incidence on landlords, as found above, implies that
the marginal per-acre subsidy dollar lowers the rental rate by 80 cents.
By lowering the rental rate of land, the subsidy changes the relative prices of the factors
of production.
The altered relative prices will have an indeterminate impact on the use of non-
land inputs. Because the relative price of non-land inputs has increased, producers will substitute
away from them toward the relatively cheaper land.
At the same time, since total costs are
lower, the farmer may expand output and demand more non-land variable inputs. The net result
on the non-land factors of production is theoretically ambiguous.
37
The subsidy will directly affect the demand for land. Facing a lower rental rate, farmers
will rent more acres. The first row of Table 10 supports this implication.
Since the number of
rented acres will have a mechanical, negative relationship with the subsidy per acre, the rented
acres are regressed on the total subsidy. One should thus interpret the coefficient as the change
in rented acres due to the marginal subsidy dollar rather than the marginal subsidy dollar per
acre. The results of this estimation suggest that increasing the subsidy by one dollar per acre (a
total increase of $1,040 for the median farm) leads to between one and three more acres rented.
This result supports the theoretical prediction about the direct effect of subsidies.
I measure the indirect effect of subsidies on non-land variable factors of production by
applying the empirical strategy developed above to expenditures on 13 variable factors of
production.
The data on variable input expenditures are reported by those returning the census
long form. Using the same sample of cash renters as above, I treat the expenditure on each input
as the dependent variable and estimate the effect of subsidies using equation (9).
The appropriate specification in this setting differs slightly from that used above in order
to adequately account for the trend in expenditures.
The time-varying county effect not only
captures local shocks, but also accounts for the county-specific
trend in expenditure.
If the
expenditure trend is proportional to the farm size then the per-acre transformation is sufficient
for the county effect to capture the change.
However, if the trend is proportional to total
expenditures and not to farm size then a log-log specification would be preferred.
reason that factor expenditure
proportional to the acres farmed.
growth for livestock
It stands to
and feed, for instance, will not be
Therefore, a more appropriate specification is the log-log
model.
38
Table 10 reports the OLS coefficient on government payments for each of the 13 factor
expenditure regressions and for a total expenditure regression. All of the regressions include a
complete set of covariates as specified earlier. For comparison purposes, column one reports the
results from the per-acre specification. The largest response comes from expenditures that are
least likely to grow in proportion to farm size, such as livestock and feed. Column two of Table
10 reports the elasticity estimates from a log-log specification. Columns three and four report
the incidence which is calculated by evaluating each elasticity at the mean and median of the
dependent and independent variables. Using OLS, I estimate that the marginal subsidy dollar
result in a 35 cent increase of total variable factor expenditures.
Expenditures might also suffer from expectation error in much the same way as rental
rates. To account for this, I instrument for the change in government payments using the 1992
set-aside acres and the 1997 government payments.
Table 11 reports the IV estimates.
The
estimate is essentially unchanged from the OLS.
The indirect effect of the subsidy increases expenditure on variable factors by about 35
cents, revealing a positive production response to the factor-specific subsidy.
This response is
consistent with the cost of farming an additional acre of land. As noted above, increasing the
subsidy payment by one dollar per acre for the median farm would increase the total subsidy
payment by $1,040 and result in about one more acre rented.
The average cost of farming an
acre of land, according to the summary statistics in Table 1, is about $385.
The expenditure
results imply that an extra $1,040 subsidy results in $365 more dollars spent on other variable
factors of production, an amount consistent with the cost of farming one additional acre. This
result is noteworthy since under WTO rules subsidies are allowed if they are nondistortionary.
The evidence presented here suggests that these subsidies might not meet that criterion.
39
9
Tenant's Incidence
The above analysis found landlords capturing 20 cents of the marginal subsidy dollar, and
tenants increasing expenditures on variable inputs by 35 cents. Yet the question of how the
subsidy ultimately affects the tenant remains. If variable inputs are perfectly elastically supplied,
and the farmer is a risk-neutral profit maximizer, then one would expect the farmer's net returns
to increase by at least 80 cents.
incidence on the tenant.
The production response should not dissipate the subsidy
Using the same empirical strategy as the one detailed in section 5, I
measure the tenant's incidence by regressing net returns on government payments. Here the
instrumental variable strategy is slightly different. There is no expectation error associated with
the dependent variable in this specification; net returns are calculated as the total revenue
(including government payments) less total variable costs. However, measurement error in the
1992 subsidy measure remains.
Fortunately, for the same economic reasons as above, the 1997
subsidy level is an appropriate instrument for the change. Namely, the cross-sectional variation
in 1997 subsidy payments accurately reflects the variation in program yield and base acres, and
these parameters are independent of the component of 1992 subsidies that suffers from
measurement error.
Table 12 reports the ultimate effect of the marginal subsidy dollar on the farmer's net
returns. As with the other estimates, the IV in columns three and four results are slightly larger
the OLS results in columns one and two, but the result is consistent.
Tenants ultimately benefit
one for one from the marginal subsidy dollar. The production response outlined above seems to
be sufficient to make up for the 20 cents extracted by the landlord.
40
10
Conclusion
This paper has investigated the direct effect of agricultural subsidies on farmland rental rates, the
indirect effect on other factors of production, and the ultimate effect on the farm operator. In the
investigation, I have overcome three significant obstacles: unobserved heterogeneity,
simultaneity, and expectation errors. Using a nationally representative dataset of individual
farms, I have controlled for unobserved heterogeneity with fixed effects. I exploited the 1996
FAIR Act to account for any concern about simultaneity bias, and I used IV techniques to
overcome measurement and expectation error, thereby identifying the effect of subsidies on
rental rates.
The analyses are based on individual-level data from 1992 and 1997, years that
bracket the 1996 policy change.
The evidence on the incidence of agricultural subsidies demonstrates that some, but not
all, of the subsidy is passed to landlords. The point estimates suggest that, on average, about 20
cents of the marginal subsidy dollar per acre are passed to landowners in the form of higher
rents. Considering that 94 percent of landlords are not farmers, this incidence implies that if all
subsidy recipients were tenant farmers, then non-farmer landlords would have received about
$1.1 billion of the $5.5 billion in agricultural subsidies paid to farmers in 1997, signifying a
"leakage" of 20 percent. Of course, not all subsidy recipients are tenants, and the distribution of
subsidy payments between tenants and owners will result in a lower leakage. Overall, a vast
majority of the aid to farmers stays in the farm sector. Evidence was presented indicating that
tenants ultimately benefit one-for-one from the marginal subsidy dollar. Farm policy appears to
accomplish its stated purpose to increase farmers' income.
The evidence presented in this paper also demonstrates the production effects caused by
agricultural subsidies.
Because incidence is incomplete and the subsidy is factor-specific, the
41
marginal subsidy dollar effectively reduces the farmland rental rate by 80 cents. In response,
farmers rent more land and purchase more variable inputs in order to produce a crop on the extra
acreage. Overall, the marginal subsidy dollar increases variable factor expenditures by 35 cents.
As a result of the production response, farmers ultimately benefit one-for-one from the marginal
subsidy dollar.
These results provide a first step in accurately characterizing the farmland rental market.
The body of literature that has investigated the effect of government payments on land values has
typically assumed full incidence because land has a zero elasticity of supply while all other
factor inputs are supplied with infinite elasticity. However, this paper's results show that in the
short-run such an assumption is untenable.
Future work should investigate the role played by market imperfections or sticky prices
in the farmland rental market. Imperfect rental markets provide a plausible explanation; the
evidence presented suggests that the incidence may increase to twenty-five percent if all farmers
within a county were tenants. A large presence of long-term tenant-landlord relationships might
cause rental rates to adjust slowly. Incidence may be higher in the long run as rents adjust when
new tenant-landlord relationships are formed, but the time frame of this analysis could be too
short to capture this occurrence.
42
References
Allen, D.W. and D. Lueck.
1992. "The 'Back Forty'
on a Handshake:
Specific Assets,
Reputation, and the Structure of Farmland Contracts." The Journal of Law, Economics,
and Organization 8(2), 366-376.
Alston, J.M. 1986. "An Analysis of Growth of U.S. Farmland Prices, 1963-82." American
Journal ofAgricultural Economics 68(1), 1-9.
Barnard,
C.H.,
G. Whittaker,
D. Westenbarger,
and
M. Ahearn.
1997. "Evidence
of
Capitalization of Direct Government Payments into U.S. Cropland Values." American
Journal of AgriculturalEconomics79(5), 1642-1650.
Chambers, R.G. 1995. "The Incidence of Agricultural Policies." Journal of Public Economics
57, 317-35.
Duffy, P.A., C.R. Taylor, D.L. Cain, and G.J. Young. 1994. "The Economic Value of Farm
Program Base." Land Economics 70(3), 318-29.
Goodwin, B.K. and F. Ortalo-Magne. 1992. "The Capitalization of Wheat Subsidies into
Agricultural Land Values." Canadian Journal of Agricultural Economics 40, 37-54.
Goodwin, B.K., A. Mishra, and F. Ortalo-Magne, 2004. "Landowners'
of
Agricultural
Subsidies."
Working
Paper
http://research.bus.wisc.edu/fom/documents/lr-latest.pdf
Riches: The Distribution
downloaded
from
Hoch, I. 1958. "Simultaneous Equations Bias in the Context of the Cobb-Douglass Production
Function." Econometrica 30, 566-578.
Hoch, I. 1962. "Estimation of Production Function Parameters Combining Time Series and
Cross-Section Data." Econometrica 30, 34-53.
Hoppe, R.A. and P. Korb. 2001. "Farm Operations Facing Development: Results from the
Census Longitudinal File." Presented at the Symposium: "Urbanization and Agriculture:
Measuring Urban Influence and Its Impact on Farmland Values and Costs of
Production. " AAEA 2001 Annual Meeting.
Lence, S.H. and A.K. Mishra. 2003. "The Impact of Different Farm Programs on Cash Rents."
American Journal ofAgricultural Economics 85, 753-61.
Melichar, E. 1979. "Capital Gains versus Current Income in the Farming Sector." American
Journal ofAgricultural Economics 61(Dec.), 1086-1106.
Morehart, M., J. Ryan, and R. Green. 2001. "Farm Income and Finance: The Importance of
Government Payments." Presented at the Agricultural Outlook Forum, 2001.
43
Moss, C.B., G. Livanis, V. Breneman, and R.F. Nehring. 2002. "Productivity
versus Urban
Sprawl: Spatial Variations in Land Values." In Agricultural Productivity: Measurement
and Sources of Growth, V.E. Ball and G.W. Norton, eds. Boston, MA: Kluwer Academic
Publishers, pp. 1 1 7-33.
Mundlak, Y. 1961. "Empirical Production Function Free of Management Bias." Journal of
Farm Economics 43, 44-56.
Robison, L.J., D.A. Lins, and R. VenKataraman. 1985. "Cash Rents and Land Values in U.S.
Agriculture." American Journal of Agricultural Economics 67(4), 794-805.
Sotomayer, N.L., P.N. Ellinger, and P.J. Barry. 2000. "Choice among Leasing Contracts in Farm
Real Estate." Agricultural Finance Review 60, 71-84.
Staiger, D. and J.H. Stock. 1997. "Instrumental Variables Regression with Weak Instruments."
Econometrica 65(3), 557-86.
Orden, D., R. Paarlberg, and T. Roe. 1999. Policy Reform in American Agriculture. Chicago, IL:
The University of Chicago Press.
U.S. Department of Agriculture, National Agricultural Statistics Service. 1996. Agricultural
Statistics 1996 Washington, DC: U.S. Government Printing Office.
U.S. Department of Agriculture, National Agricultural Statistics Service. 2001a. Agricultural
Statistics 2001 Washington, DC: U.S. Government Printing Office.
U.S. Department of Agriculture, National Agricultural Statistics Service. 2001b. "1997 Census
of Agriculture: Agricultural Economics and Land Ownership Survey (1999)." Vol. 3,
Special Studies, Part IV.
Weersink, A., S. Clark, C.G. Turvey, and R. Sarker. 1999. "The Effect of Agricultural Policy on
Farmland Values." Land Economics 75(3), 425-439.
44
11
Appendix
1
The census of agriculture reports the total cash rent paid and the total number of acres rented on
a cash, share, or free basis. I use the ratio of these variables to construct the rental rates variable,
which is the dependent variable in the primary analysis. Because the number of acres rented for
cash are not reported separately
constructed
from the number rented for a share of production,
cash rental rate may be smaller than the true cash rental rate.
the
The induced
measurement error is not classical measurement error because it has a positive expected value.
Formally, the true cash rental rate, r,*, equals the calculated cash rental rate, ri,, plus
measurement error, pit:
ri, = ri+
,
Since the calculated cash rental rate is less than or equal to the true cash rental rate, r, 2>r,, the
measurement error must be greater than or equal to zero, E(u, ) 2 0, with equality only when all
farms cash rent only, i.e., when there is no measurement error.
The potential inconsistency due to the non-classical measurement error can be seen using
the estimating equation:
, f, + C,+J, .
rJ=r, + ,Ui,
=:a + gr + ,Xi+
To see the source of inconsistency, write the linear projection of the measurement error, pit, onto
government payments, git, observed farm characteristics, Xi,, and a county fixed effect, z,:
Uit= o0+ l g ,, + 2X, +r, +vit.
0
Substituting into the estimating equation and rearranging yields:
ri, = (a -
)+ ( -
)gi, + ( -
45
2 )Xi, + f, + (C, - , )+ 1c , + Vo,.
This equation illustrates the possible inconsistency. When the measurement error is ignored the
plim of the coefficient on government payments is: plim
= 7-
I.
If government payments
are positively correlated with the non-classical measurement error, i.e. lI>0, the estimate will be
biased downward.
I investigate this source of bias using the 1999 Agricultural Economics and Land
Ownership Survey, which sampled over 26,000 farms nationwide and asked detailed questions
about the farm operation, including the number of acres rented for cash and the total rent paid on
those acres, and the number of acres rented for a share of production and the value of the share of
production paid on those acres. Using the data from this survey I construct the 'true' cash rental
rate by dividing the total cash rent paid by the acres rented for cash. I reproduce the rental rate
variable used in the main analysis by dividing the total cash rent paid by the acres rented for cash
summed with the acres rented for a share of production.
The difference between the true cash
rental rate and the rental rate variable used in the main analysis is an estimate of the non-classical
measurement error.
By regressing the calculated non-classical measurement error on the per-acre government
payments, observable farm characteristics, and a county fixed effect I estimate the magnitude
and direction of the potential bias. Appendix table 1 contains the results of that analysis. The
population of interest consists of farms that either cash and share rent or cash rent only. These
criteria resulted in a sample size of 11,924. It should be noted that on average there are about 5
farms per county in this analysis, resulting in less power for the county fixed effects. Columns 1
and 2 report the results when there are no fixed effects and when there are state-level fixed
effects, respectively.
In each of these specifications the coefficient on government payments in
statistically significant, equaling 0.06 without fixed effects and 0.03 with state-level fixed
46
effects. These results imply that if one assumes that the measurement error is independent of
unobserved county fixed effects, such as local institutional factors, the primary results in the
paper should be increased by 0.06, resulting in an incidence estimate of 0.24. Column 3 of
appendix table 1 reports the most general specification controlling for local unobserved
heterogeneity with county fixed effects. The result is a statistically insignificant coefficient of
.00329 on the government payment variable.
Based on this estimate one cannot reject the null
hypothesis that government payments and the non-classical measurement error are uncorrelated,
hence the primary estimates reported in the paper are consistent.
47
Table
1
Summary Statistics: U.S. Census of Agriculture Micro Files
Program Crop Producers
Sampled Renters
N=437,979
N=58,302
Mean
Variable
Size (Acres)
Sales ($/Acre)
Expenditures ($/Acre)
Irrigated Acres
Proportion in Pasture
Proportion Irrigated
Proportion in Cropland
Subsidies ($)
(including zeros)
Subsidies ($)
(exluding zeros)
Subsidies ($/ Acre)
(including zeros)
Subsidies ($/ Acre)
(exluding zeros)
Rental Rate ($/ Acre)
Median
Year
(1)
(2)
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
1992
1997
712.988
737.155
282.923
329.000
632.886
227.470
494.116
47.803
54.290
0.182
0.195
0.058
0.060
0.802
0.769
5,621.83
5,148.58
9,497.64
7,987.03
8.116
14.338
13.712
22.242
n/a
n/a
Mean
(3)
325.000
163.288
182.471
127.567
137.000
0.000
0.000
0.063
0.057
0.000
0.000
0.900
0.880
1,124.00
1,345.00
4,800.00
4,128.00
3.391
4.418
10.188
9.767
n/a
n/a
Median
(4)
1,638.330
1,743.750
1,040.000
1,116.000
465.250
483.460
369.717
385.997
299.546
303.898
233.362
230.172
138.924
161.452
0.135
0.133
0.100
0.000
0.000
0.015
0.008
0.000
0.106
0.000
0.821
0.824
17,234.20
14,287.55
22,643.12
17,918.44
12.947
11.085
17.011
13.902
0.912
0.921
10,393.93
9,919.00
15,447.56
13,900.00
9.514
9.000
13.541
11.687
47.605
34.468
50.123
36.543
Notes: The sample consists of all farms that returned the Census of Agriculture's long form in both 1992 &
1997, payed cash rent in both years, and reported growing program crops in 1992. There are 58,302 farms in
each year of the sample. All monetary values have been adjusted to 1997 dollars.
48
Table 2 - Annual Cross Sections
Effect of Subsidies on Farmland Rental Rates
Dep Var: Farm Average Rental Rate ($/acre)
Farm Level Analysis
No Controls
0.463
**
(0.013)
0.309
(0.015)
0.033
(0.009)
-0.029
(0.010)
-0.452
(0.095)
Costs
Farm Size (Acres)
Proportion
Irrigated
Proportion
Pasture
**
0.261
County Fixed
Effects and
Covariates
(4)
(3)
(2)
A. 1992 Cross Section
(1)
Government
Payments
Sales
Covariat es
County Fixed
Effects
**
(0.012)
0.199
(0.013)
0.021
(0.001)
-0.019
(0.001)
-0.200
(0.064)
**
**
**
**
**
22.282 **
29.985 **
(1.073)
-33.222
(1.168)
(1.377)
-14.006
(1.463)
**
0.320
**
**
B. 1997 Cross Section
Government
Payments
Sales
0.946
**
(0.022)
0.019
(0.004)
-0.013
(0.003)
-0.295
(0.052)
Farm Size (Acres)
Proportion
Irrigated
Proportion
Pasture
0.645
(0.000)
**
**
**
**
**
**
(0.017)
0.011
(0.001)
-0.008
(0.001)
-0.086
(0.064)
**
**
23.441 **
27.648 **
(1.072)
-33.341
(1.343)
(1.422)
-16.732
(1.553)
**
C. Pooled Cross Sections
0.371
**
0.303
**
(0.010)
(0.012)
0.022
**
(0.004)
Costs
-0.017 **
Farm Size (Acres)
(0.004)
-0.372
(0.050)
Proportion
Irrigated
Proportion
Pasture
0.424
(0.017)
(0.022)
Costs
Government
Payments
Sales
0.493
**
23.284 **
(0.762)
-33.097
(0.907)
**
0.230
(0.010)
0.014
(0.000)
-0.011
(0.000)
-0.141
(0.045)
29.251
(0.978)
-14.888
(1.052)
**
**
**
**
**
**
**
long formin both 1992& 1997,
Notes: The sampleconsistsof all farmsthat returnedtheCensusof Agriculture's
payedcashrent in both years,and reportedgrowingprogramcropsin 1992. Thereare 58,302farmsin eachyearof
standarderrors areshowninparentheses. The dependentvariableiscash
the sample. Heteroskedasticity-robust
rentalrate. Whenincluded,thecovariatesare the 1992values;the 1997valuesare not includedbecausetheyare
potentialoutcomes. The controlsarethe 1992valuesof per-acresales,per-acrevariablecostsexcludingrent, farm
size ( 1,000sof acres),proportionof farmlandthat is irrigated,proportionof farmlandin pastureor range,proportion
of total fannacres plantedto corn, wheat,oats, barley,sorghum,cotton,rice,soybeans,other grains& beans,hayand
seeds,vegetables,fruits& nuts, and other crops. Allmonetaryvalueshavebeenadjustedto 1997dollars. **
indicatessignificanceat 99thpercentile.
49
Table 3 - Individual Farm Effects Estimates
Dependent Variable: 1992 - 1997 Change in the Cash Rental Rate
Random
Effects
Farm Fixed
Effects
(1)
Government
Payments
Sales
0.408
(2)
**
(0.009)
0.183
(0.061)
Costs
Farm Size (Acres)
Proportion
Irrigated
Proportion
Pasture
Hausman Test
p -value
County x
Time Effects
(3)
**
0.183
Time Effect
(4)
**
0.172
(0.018)
-0.008
(0.018)
-0.007
(0.005)
0.007
(0.006)
0.006
(0.006)
(0.007)
0.099
0.109
(0.055)
-5.269
(1.958)
-4.678
(0.051)
1.038
(1.321)
-1.790
(1.845)
**
**
**
(1.564)
1033.100
(0.000)
Notes: The sample consists of all farms that returned the Census of Agriculture's long form in both 1992 &
1997, payed cash rent in both years, and reported growing program crops in 1992. There are 58,302 farms in
each year of the sample. Heteroskedasticity-robust standard errors are shown in parentheses. The dependent
variable is the change in the cash rental rate, except in the random effects model where the dependent variable
is the cash rental rate. When included, the covariates are the 1992 values; the 1997 values are not included
because they are potential outcomes. The controls are the 1992 values of per-acre sales, per-acre variable costs
excluding rent, farm size (1,000s of acres), proportion of farmland that is irrigated, proportion of farmland in
pasture or range, proportion of total farm acres planted to corn, wheat, oats, barley, sorghum, cotton, rice,
soybeans, other grains & beans, hay and seeds, vegetables, fruits & nuts, and other crops. All dollars have been
inflated to 1997 dollars. ** indicates significance at 99th percentile.
50
Table 4 - Farm Effects IV Estimates
1992 Set-aside and 1997 Subsidy as Instruments
No
Covariates
Covariates
(1)
1997 Government
Payments
1992 Set-Aside
Acres
Sales
0.830
(0.005)
-327.517
(2.771)
Costs
Farm Size
Proportion Irrigated
Proportion Pasture
First Stage F-stat
R2
Government
Payments
Sales
20,011
0.31
0.186
(0.020)
Costs
Farm Size
Proportion Irrigated
Proportion Pasture
Overidentification Test
p -value
20.77
0.00
(2)
A. First Stage
**
0.902
(0.004)
**
-323.647
(2.636)
0.0005
(0.0002)
-0.0007
(0.0003)
0.1527
(0.0183)
-3.112
(0.3941)
3.644
(0.4185)
2,057
**
**
**
**
**
**
**
0.51
B. IV Estimates
**
0.214
(0.019)
-0.0075
(0.0011)
0.0065
(0.0013)
0.098
(0.0808)
-4.953
(1.7449)
-4.778
(1.8513)
21.25
0.38
**
**
**
**
**
Notes: The sample consists of all farms that returned the Census of Agriculture's long
form in both 1992 & 1997, payed cash rent in both years, and reported growing
program crops in 1992. There are 58,302 farms in each year of the sample. Standard
errors are shown in parentheses. The dependent variable is the change in the cash
rental rate. All specifications include a farm fixed effect and a time-varying county
fixed effect. The instruments are the proportion of the farm enrolled in set-aside in
1992 and the per-acre subsidies in 1997. When included, the covariates are the 1992
values of the variable; the 1997 values are not included because they are potential
outcomes. The controls are the 1992 values of per-acre sales, per-acre variablecosts
excluding rent, farm size (1,000s of acres), proportion of farmland that is irrigated,
proportion of farmland in pasture or range, proportion of total farm acres planted to
corn, wheat, oats, barley, sorghum, cotton, rice, soybeans, other grains & beans, hay
and seeds, vegetables, fruits & nuts, and other crops. All dollars have been inflated to
1997 dollars. In panel A, the F-statistic is for the test of whether the coefficients
on the instruments are jointly equal to zero. In panel B, the F-statistic is for the
test of overidentifying restrictions. **indicates significance at the 99th percentile.
51
Table 5 - Region Summary Statistics
Rental Rate
Subsidy per Acre
Mean
Median
Mean
Median
Region
Heartland
N=18,920
Northern
Crescent
1992
1997
(1)
55.28
60.28
(2)
48.84
53.92
(3)
13.62
13.79
(4)
12.51
12.77
1992
1997
49.75
50.06
39.43
38.87
9.11
8.83
5.04
6.87
Northern
Great Plains
N=6, 198
1992
1997
25.85
25.08
16.83
18.07
9.01
6.53
7.66
5.54
Prarie
Gateway
1992
1997
19.50
19.68
10.79
10.81
10.84
7.94
7.87
5.98
1992
1997
28.48
25.40
19.45
17.05
6.85
5.49
0.00
2.11
1992
1997
53.24
51.55
38.99
37.50
8.69
8.40
2.00
5.19
Fruitful Rim
N=3,358
1992
1997
70.75
73.45
53.45
55.56
22.15
14.01
7.77
7.18
Basin and
Range
1992
1997
42.80
42.62
24.00
20.71
6.90
5.43
1.59
2.00
1992
1997
35.88
34.75
28.48
26.43
27.29
16.74
20.46
13.68
N=10,280
N=7,049
Eastern
Uplands
N=1,921
Southern
Seaboard
N=6,250
N=1,263
Mississippi
Portal
N=3,063
Notes: Data are from the confidential microfiles of the Census of Agriculture.
Regions are as defined by the USDA's Economic Research Service as Farm
Resource Regions.
52
Table 6 - Regional Incidence Estimates
OLS
Region
Heartland
No
Covariates
(1)
0.274
**
(0.051)
Northern
Crescent
0.137
(0.056)
Northern
Great Plains
0.063
(0.074)
Prarie
Gateway
0.222
(0.044)
Eastern
Uplands
0.033
(0.066)
Southern
Seaboard
0.194
Fruitful Rim
Basin and
**
**
(0.134)
Mississippi
Portal
0.137
(0.028)
Covariates
(2)
(3)
(4)
0.288
**
0.148
0.270
**
(0.030)
**
0.150
0.294
**
(0.030)
**
0.173
**
(0.056)
(0.053)
(0.053)
0.065
(0.073)
0.063
(0.068)
0.146
(0.068)
**
0.171
(0.047)
**
0.190
(0.042)
0.041
(0.072)
0.003
(0.113)
**
0.216 **
(0.055)
0.255 ** 0.315 **
(0.064)
(0.064)
**
0.235 **
(0.055)
(0.082)
0.006
(0.116)
0.082
(0.142)
**
0.107
(0.028)
**
**
0.121
(0.049)
0.040
Range
Covariates
No
Covariates
(0.050)
(0.051)
0.221
(0.055)
IV
**
0.240
0.164
(0.043)
0.074
(0.114)
**
0.259 **
(0.064)
0.087
(0.155)
**
0.166
**
(0.039)
Notes:. The sample consists of all farms that returned the Census of Agriculture's long form in
both 1992 & 1997, payed cash rent in both years, and reported growing program crops in
1992. There are 58,302 farms in each year of the sample. Heteroskedasticity-robust standard
errors are shown in parentheses.
The dependent variable is cash rental rate. When included,
the covariates are the 1992 values; the 1997 values are not included because they are
potential outcomes. The controls are the 1992 values of per-acre sales, per-acre variable
costs excluding rent, farm size (1,000s of acres), proportion of farmland that is irrigated,
proportion of farmland in pasture or range, proportion of total farm acres planted to corn,
wheat, oats, barley, sorghum, cotton, rice, soybeans, other grains & beans, hay and seeds,
vegetables, fruits & nuts, and other crops. All monetary values have been adjusted to 1997
dollars. ** indicates significance at 99th percentile.
53
Table 7 - Summary Statistics by Sales Class
Rental Rate
Subsidy per Acre
Mean
Median
Mean
Median
1992 Sales
(1)
(2)
(3)
(4)
0.00
0.00
< $10,000
N= 2,597
1992
1997
24.56
16.41
27.65
17.05
3.01
4.94
$10,000 - $100,000
35.45
N= 31,587
1992
1997
32.97
23.71
20.63
8.87
7.53
4.15
4.37
$100,000 - $250,000
N= 5 7,091
1992
1997
41.86
44.48
30.09
32.01
12.17
10.57
8.59
8.54
$250,000 - $500,000
1992
1997
49.03
50.01
37.61
10.83
38.76
13.38
12.16
1992
1997
60.43
64.35
47.74
50.46
15.06
12.45
11.29
10.69
N= 49,222
$500,000 +
N= 26,562
10.51
Notes: Data are from confidentialU.S. Census of Agriculture microfiles. Sale class taxonomy
as used by the USDA Economic Research Service. Statistics weighted by sample weights
developed by the USDA National Agricultural Statistical Service. Observations per sales class
represent weighted totals.
54
Table 8 - Sales Class Incidence Estimates
OLS
No Fixed Effects
1992 Sales
< $10,000
$10,000-$100,000
No
Covariates
Covariates
(1)
(2)
0.411
** 0.195
(0.110)
(0.098)
0.615
**
$250,000-$500,000
0.671
(0.018)
**
0.541
**
0.425
(0.026)
**
0.221
0.463
(0.021)
**
0.429
**
**
0.294
**
**
0.190
(0.024)
**
0.171
**
0.184
(0.031)
0.227
**
(0.055)
0.173
(0.024)
**
0.180
**
**
0.208
(0.032)
No
Covariates
(3)
-0.207
(1.707)
**
Covariates
(4)
-4.644
(11.708)
0.100
0.192
(0.083)
(0.079)
0.152
(0.041)
**
0.177
**
0.222
(0.046)
**
0.211
(0.038)
**
0.211
**
(0.031)
(0.032)
(0.023)
(0.022)
(0.027)
IV
Fixed Effects
Covariates
(4)
-1.027
(1.826)
(0.055)
(0.022)
(0.020)
$500,000+
0.334
No
Covariates
(3)
-0.360
(0.852)
(0.021)
(0.021)
$100,000- $250,000
OLS
Fixed Effects
**
0.244
**
(0.045)
Notes: Dependentvariableis therental rate. Columnsthreethroughsixare estimatedin firstdifferences;therethe dependentvariableis the change
in farmlandrentalrates. Controls,whenincluded,are the 1992values;the 1997valuesare not includedbecausetheyarepotentialoutcomes. The
controlsarethe 1992valuesof per-acresales;per-acrevariablecosts, excludingrent; farmsize ( ,000sof acres);proportionof farmlandthat is
irrigated;proportionof farmlandinpasture or range;proportionof total farmacresplantedto corn, wheat,oats, barley,sorghum,cotton,rice,
soybeans,othergrains & beans,hay and seeds,vegetables,fruits& nuts, and other crops;and 1992yieldof corn, wheat,oats, barley,sorghum,
cotton,rice, and soybeans.All dollarshavebeen inflatedto 1997dollars. ** indicatessignificanceat 99thpercentile. Salesclasstaxonomydefined
by theUSDAEconomicResearchService.
55
Table 9 - The Effect of Rental Market Concentration on the Subsidy Incidence
Coefficient on Government Payments & the Interaction Term Reported
OLS
Main Effect
Interaction Term
IV
Interaction
(1)
(2)
Herfindahl
Market Share = Rental Expenditures
-0.090
(0097
(0.097)
0.213
(0.044)
-0.042
-0.042
(0.065)
0.147
(0.034)
0.400
(0.40)
(0.140)
-0.242
(0.122)
0.368
0.368
(0.114)
-0.077
(0.112)
Herfindahl
Market Share = Acres Rented
0.220
(0.026)
Proportion of Farmers who are Tenants
Tenant-Landlord Ratio
Proportion of Farm Land that is Rented
**
**
**
-0.162
(0.038)
**
**
**
Main Effect
Interaction
(3)
(4)
0.009
(0.020)
(0.020)
0.048
(0.021)
0.019
0.019
(0.019)
0.135
(0.016)
0.246
(0.026)
(0.026)
0.338
(0.052)
0.275
0.275
(0.027)
0.148
(0.075)
0.170
(0.021)
**
**
**
-0.050
(0.025)
**
**
**
*
*
Notes: The sample consists of all farms that returned the Census of Agriculture's long form in both 1992 & 1997, payed cash rent in
both years, and reported growing program crops in 1992. There are 58,302 farms in each year of the sample. Heteroskedasticityrobust standard errors are shown in parentheses. The dependent variable is change in the cash rental rate. The interaction terms are
county level measures of rental market concentration. The Herfindahl indexes treat market share as a farm's proportion of total rental
expenditure in the county, and a farms proportion of the total acres rented in the county. The instruments are the per-acre subsidies in
1997 and the per-acre subsidies in 1997 interacted with the measure of rental market concentration . All monetary values have been
adjusted to 1997 dollars. ** indicates significance at 99th percentile.
56
Table 10
Effect of Subsidies on Variable Factor Expenditures - OLS Regressions
Coefficient on Government Payments Reported
. Log-log Specification
Incidence Incidence
at the
at the
Mean
Median
OLS
(3)
(4)
Acres
0.003 **
(0.000)
0.011 **
(0.000)
0.001
0.001
Livestock
0.218
(0.099)
0.571
(0.131)
0.035
(0.005)
0.075
(0.007)
0.078
(0.006)
0.043
(0.005)
0.010
(0.004)
0.027
(0.003)
0.020
(0.002)
0.024
(0.002)
0.044
(0.003)
0.014
(0.001)
0.026
0.000
0.077
0.009
**
0.022
0.023
**
0.051
0.053
**
0.079
0.068
**
0.016
0.018
0.020 **
0.008
0.006
**
0.072
0.042
**
0.031
0.033
**
0.022
0.008
**
0.066
0.066
**
0.006
0.005
0.016 **
0.033
0.027
0.339
0.349
Dependent
Variable
Feed
Seed
Fertilizer
Chemicals
Fuel
Electricity
Labor
Repair
Machine
Rental
Interest
Property Taxes
(2)
(1)
**
**
**
**
**
**
0.014 **
(0.004)
0.074 **
(0.022)
0.061 **
(0.009)
0.016
(0.012)
0.092 **
(0.015)
0.004
(0.004)
0.098 **
(0.002)
0.030
(0.004)
0.021
(0.002)
0.041
(0.005)
0.042
(0.003)
0.015
(0.002)
Other
Expenditures
(0.040)
(0.001)
Total Variable
Factors
1.289 **
(0.251)
0.016
(0.001)
**
Notes: The sample consists of all farms that returned the Census of Agriculture's long form in
both 1992& 1997, payed cash rent in both years, and reported growing program crops in 1992.
There are 58,302 farms in each year of the sample. Standard errors are shown in parentheses.
All specificationsinclude covariates from 1992; the 1997 values are not included because they are
potential outcomes. The controls are the 1992 values of per-acre sales, per-acre variablecosts
excluding the dependent variable; farm size (l,000s of acres), proportion of farmlandthat is
irrigated, proportion of farmland in pasture or range, proportion of total farm acres planted to
corn, wheat, oats, barley, sorghum, cotton, rice, soybeans, other grains & beans, hay and seeds,
vegetables, fruits & nuts, and other crops. Due to computationalconstratints, the robust
regressions do not include the time-varyingcounty effect. The log-log specificationreports the
elasticityestimates, which are evaluated at the mean and median of the subsidyand the dependent
variableto obtain the incidence estimates. All monetary values have been adjusted to
1997 dollars. ** indicates significanceat 99th percentile.
57
Table 11 - Effect of Subsidies on Variable Factor Expenditures
IV Results - 1992 Log Set-aside and 1997 Log Subsidies Instruments
Coefficient on Log Government Payments Reported
Log-log Specification
No
Covariates
Dependent
Variable
Acres
Livestock
Feed
Seed
Fertilizer
Chemicals
Fuel
Electricity
Labor
(1)
Incidence at Incidence at
the Mean
th,e Median
(2)
(3)
0.002
0.002
0.025
0.025
**
0.095
0.095
**
0.030
0.030
**
0.069
0.069
**
0.083
0.083
**
0.021
0.021
**
0.009
0.009
0.039 **
0.095
0.095
0.037
0.037
**
0.025
0.025
**
0.070
0.070
**
0.008
0.008
0.015
(0.001)
0.009
(0.005)
0.033
(0.004)
0.027
(0.002)
0.032
(0.002)
0.046
(0.004)
0.019
(0.001)
0.023
(0.003)
(0.004)
Repair
0.026 **
Other
Expenditures
(0.002)
0.047
(0.006)
0.044
(0.004)
0.021
(0.003)
0.021
(0.002)
**
0.044
0.044
Total Variable
Factors
0.019 **
(0.001)
0.399
0.399
Machine
Rental
Interest
Property Taxes
Covariates
Incidence at Incidence at
the Mean
the Median
(5)
(6)
0.002
0.002
0.020
0.000
**
0.088
0.010
**
0.029
0.029
**
0.063
0.064
**
0.085
0.074
**
0.018
0.021
**
0.008
0.005
**
0.081
0.047
**
0.032
0.034
**
0.024
0.009
**
0.061
0.062
**
0.007
0.006
**
0.038
0.031
0.017 **
(0.001)
0.351
0.362
(4)
0.020
(0.001)
0.008
(0.005)
0.030
(0.004)
0.025
(0.002)
0.029
(0.002)
0.047
(0.004)
0.016
(0.001)
0.019
(0.003)
0.033
(0.004)
0.022
(0.002)
0.045
(0.006)
0.039
(0.004)
0.017
(0.003)
0.018
(0.002)
Notes: The sample consists of all farms that returned the Census of Agriculture's long form in both 1992 & 1997, payed cash
rent in both years, and reported growing program crops in 1992. There are 58,302 farms in each year of the sample. Standard
errors are shown in parentheses. Farm and time-varying county effects are included in all specifications. The instruments are
the log of acres in the farm enrolled in set-aside in 1992 and the log of subsidies in 1997. When included, the covariates are
from 1992; the 1997 values are not included because they are potential outcomes. The controls are the 1992 values of peracre sales, per-acre variable costs excluding the dependent variable; farm size (1,000s of acres), proportion of farmland that is
irrigated, proportion of farmland in pasture or range, proportion of total farm acres planted to corn, wheat, oats, barley,
sorghum, cotton, rice, soybeans, other grains & beans, hay and seeds, vegetables, fruits & nuts, and other crops. All monetary
values have been adjusted to 1997 dollars. ** indicates significance at 99th percentile.
58
Table 12
Effect of Subsidies on Net Returns
Dependent Variable: Net Returns ($/Acre)
OLS
IV
(1)
Government
Payments
Farm Size
0.959
(0.086)
0.976
Pasture
No
(3)
**
(0.089)
0.656
(0.576)
27.500
Irrigated
Acreage Controls
(2)
**
1.076 **
(0.128)
(4)
1.080
**
(0.124)
0.660
(0.576)
27.647 **
**
(11.494)
(11.494)
-22.202 **
(11.240)
-22.462 **
(11.242)
Yes
No
Yes
Notes. Net returns here are total sales plus subsidies minus total variable costs. The
instruments are the 1992 proportion of the farm enrolled in set-aside and the 1997 peracre subsidy. The sample consists of all farms that returned the Census of Agriculture's
long form in both 1992 & 1997 and reported growing program crops and paying cash
rent in both years. There are 58,302 farms in the sample. Acreage controls are the 1992
total farm acres planted to corn, wheat, oats, barley, sorghum, cotton, rice, soybeans,
other grains & beans, hay and seeds, vegetables, and fruits & nuts. A farm fixed effect,
year effect, and time-varying county fixed effect are included. Standard errors are
shown in parentheses. ** indicates significance at 99th percentile.
59
Appendix Table 1 - Sample Selection Correction
Dependent Variable:
1992 - 1997 Change in the Cash Rental Rate
(1)
Government Payments
0.181
Inverse Mills Ratio
Include Crop-Mix
Covariates In Second
(2)
**
0.295
(0.017)
(0.075)
5.174
64.587
(5.864)
(38.525)
N
Y
**
Stage?
Notes: The sample consists of all farms that returned the Census of
Agriculture's long form, reported growing program crops, and cash
rented in 1992. There are 151,534 farms in the sample. The coefficients
reported are from the second step in a Heckman two-step sample
selection correction model. The dependent variable is the change in the
cash rental rate. When included, the covariates are the 1992 values; the
1997 values are not included because they are potential outcomes. The
controls are the 1992 values of per-acre sales, per-acre variable costs
excluding rent, farm size (1,000s of acres), proportion of farmland that
is irrigated, proportion of farmland in pasture or range, proportion of
total farm acres planted to corn, wheat, oats, barley, sorghum, cotton,
rice, soybeans, other grains & beans, hay and seeds, vegetables, fruits &
nuts, and other crops. Model 1 identifies off of the exclusion restriction;
model 2 identifies off of the non-linearity of the inverse Mills ratio. All
dollars have been inflated to 1997 dollars. ** indicates significance at
99th percentile.
60
AppendixTable 2 - Non-Classical Measurement Error
Dependent Variable: Calculated Measurement Error from 1999 AELOS
(1)
(2)
**
0.033
(3)
Government
0.066
**
0.003
Payments
(0.008)
(0.006)
(0.006)
Fixed Effects
N
None
11,924
State
11,924
County
11,924
Notes. The data are from the 1999 Agricultural Economics and Land Ownership Survey.
The sample consists of all farms that reported cash renting or both cash and share
renting. Heteroskedasticity-robust standard errors are shown in parentheses. The
dependent variable is constructed as the difference between the cash rental rate, and the
total cash rent paid divided by the sum of the acres cash rented and the acres share
rented. The covariates are the amount of sales per acre, the total cost per acres
excluding rent, the value of assets per acre, the value of debt per acre, and the value of
new machinery per acre. All dollars have been deflated to 1997 dollars. ** indicates
significance at 99th percentile.
61
62
Chapter 2
Adversity and the Propensity to Fail:
The Impact of Disaster Payments and Multiple Peril Crop Insurance on
U.S. Farm Failure Rates
1
Introduction
"Crop insuranceand a systemof storagereserveswouldhelp to
protect the incomeof individualfarmers againstthe hazardsof
cropfailure; it wouldhelp to protect consumersagainstshortages
offood supply and againstextremesofprices; andfinally, it would
assist in providing a morenearly evenflow offarm supplies,thus
stabilizingfarm buyingpower and contributingto the securityof
businessand employment."
- Franklin D. Roosevelt, February 18, 1937
A key purpose of government-provided insurance is to help recipients smooth income or
consumption during adversity.
Thus, the marginal contribution of government-provided
insurance may be small when there are many other means of risk management.
Farmers in
particular have a myriad of risk management devices such as self-insurance through savings,
crop diversification,
and many financial instruments
and contract types, to name a few.
Although much research has examined adverse selection and moral hazard in crop insurance
markets, few have addressed the income smoothing and business security benefits of this
program. As policy makers seek the optimal size, scope, and program design, they require
information on both the costs and the benefits of crop insurance, as well as the effectiveness of
various crop insurance designs. Recent legislation focused on crop insurance demonstrates the
salience of this information.
Weather events have always played a significant role in the decline in the number of
farms in the U.S. Spurred by the 1862 Homestead Act, the number of farms in the U.S. grew
63
rapidly from 2 million in 1862 to 6.5 million in 1917. From 1917-1921 farmers in the newly
settled Midwest and Great Plains experienced their first major drought, which slowed the growth
in the number of farms. As demonstrated in figure 2, the number of farms in the U.S. peaked in
1935 and declined precipitously until the last decade when numbers have leveled off at about 2
million farms.'8 As vividly captured in Steinbeck's Grapes of Wrath, the decade long drought
that caused the Dust Bowl in the 1930s contributed greatly to the initial decline in farms in the
late 1930s. More recently, major weather events such as the 1988 drought in the Midwest and
the 1993 flooding in the Midwest have captured media attention and galvanized support for the
family farmer.
The popular perspective on the declining number of farms overlooks the economic
efficiency brought about through farm failure.
Economists since Hayek and Schumpeter have
recognized the economic benefits of reallocating productive resources from inefficient firms to
efficient firms. The near century-long decline in farm numbers is due in no small part to rapid
technological advancement and shifting factor prices (Barkley 1990; Huffman and Evenson
2001). Thus, from an efficiency standpoint the declining number of farms is not particularly
disconcerting, but the relative increase in the decline due to temporary shocks, such as weather
shocks, may be a source of concern. Credit market imperfections may prevent viable firms from
smoothing temporary shocks, resulting in inefficient failures due to adverse shocks.
An important element of perfect credit markets is the existence of state-contingent
contracts, e.g. insurance. For instance, in the presence of liquidity constraints efficient firms
benefit from the ability to insure against temporary adverse shocks (Kiyotaki and Moore 1999;
Krishnamurthy 2003). The absence of insurance contracts, for instance due to asymmetric
18 Since
1974, the definition of a farm has been based on $1,000 worth of agricultural production, a nominal amount
whose real value has eroded thereby liberalizing the definition of a farm and contributing to the apparent leveling off
of farm numbers.
64
information, may be a market imperfection leading to inefficient farm turnover. In this case,
government-provided insurance increases efficiency by overcoming the market failure and
allowing viable farms to smooth temporary shocks.
Insurance
market
failures
due to adverse
selection provide
the key motive for
government intervention. The government addresses these insurance market failures in a number
of different ways.
Typically, the government mandates that all individuals or all firms
participate in an insurance market and pay premiums through specific taxes (e.g. Social Security
for individuals and unemployment insurance for firms) or through premiums to private insurers
(e.g. workers compensation).
However, in the case of multiple peril crop insurance (MPCI), the
government tries to entice farmers by generously subsidizing the insurance premiums paid by
individuals to private insurers that are subsequently reinsured by the government.
participation
Due to low
(about 25-percent of eligible acres in the 80s and early 90s) adverse selection
continued to be a significant problem (Just et al., 1999; see Goodwin and Smith, 1995, for a
comprehensive review of early studies).
Moral hazard comprises another cost that influences the design of government-provided
insurance programs.
In the crop insurance literature investigators have focused on two moral
hazard-induced types of behavior: input use and cropland choice. Several researchers examined
input use, such as fertilizers and chemicals, and found that input use is negatively correlated with
crop insurance (Quiggen et al., 1993; Babcock and Hennessy, 1996; Goodwin and Smith, 2003).
Others have investigated the influence of crop insurance on the type and amount of land in
production (Claasen et al., 2005; Wu, 1999; Goodwin et al., 2004).
This body of work has
generally found that crop insurance is associated with more environmentally sensitive land being
brought into production.
65
In contrast to the voluminous work examining the costs associated with crop insurance,
little has been done to investigate the benefits.
This paper fills that gap. By exploiting random
variation in weather and a major policy change in 1994, I exploit two natural experiments that
identify the causal effect of crop insurance on farm failure rates.
My empirical design also
allows me to address crop insurance program design issues by comparing the effectiveness of
crop insurance versus ad hoc, ex post disaster payments. Making this comparison also gives
insight into the crop insurance induced crowd-out of other risk management behaviors.
In the analysis I focus on two regimes, one primarily consisting of ex post disaster
payments in 1994 and the other a purely crop insurance regime in 1995. I find that during the
disaster payments regime farms experiencing a disaster are slightly less likely to fail than farms
that are not hit by disaster, while under the crop insurance regime I find that disaster-stricken
farms are about 1.3 percentage points more likely to fail.
The lower failure rate under the
disaster payment regime does not necessarily indicate that it is more efficient.
In light of the
'efficient failure' argument above, the lower failure rate more likely speaks to the inefficient
allocation of funds under the politically driven disaster payments regime. Garrett et al. (2003)
provide evidence that key political figures allocate relatively more disaster payments to their
constituents.
Hence the lower failure rate may indicate that inefficient farms are being
artificially propped up.
Although the absolute effects of crop insurance are immeasurable without a base period
free from government provided assistance, I compare the two regimes to find that farms are
about 2 percentage points (30-percent) more likely to fail under mandatory crop insurance than
under a program primarily consisting of ex post disaster payments. Moral hazard associated with
66
ex ante crop insurance provides a plausible explanation of the increased failure rate. Future
research with more appropriate data will test this hypothesis.
Although somewhat surprising, the relative effect supports the idea that government
provided crop-loss assistance does not completely crowd out other means of risk management.
If farmers were able to fully insure through some other means, variation in crop insurance
regimes would not alter failure rates. The 30-percent relative effect suggests that government
intervention can potentially play a significant role in this market.
The rest of the paper is structured as follows. Section 2 lays out the policy and program
background.
Section 3 introduces the empirical strategy.
The data are discussed in section 4.
Section 5 presents the empirical results. The interpretation and conclusion are put forward in
section 6.
2
Background
Passed as part of New Deal legislation, the Agricultural Adjustment Act of 1938 introduced the
Federal Crop Insurance Program as a form of social insurance meant to smooth the consumption
of the 32 million people living on farms in the United States. Spurred on by the decade-long
drought that produced the Dust Bowl, policy makers established a national crop insurance
product for wheat that insured against losses due to, "drought, flood, hail, wind, winterkill,
lightning, tornado, insect infestation, plant disease, and ... other unavoidable causes..." (52 Stat.
74). Private firms had previously tried and failed to provide multiple peril crop insurance
(MPCI). Owing to the geographic concentration of their policyholders, private insurance firms
succumbed to geographically correlated weather events, such as drought and flooding. Private
67
insurance companies succeeded only in providing insurance for hail and fire, events that are
typically spatially uncorrelated.
Due to poor farm-level yield data, as well as problems with the program design, the
Federal Crop Insurance Corporation (FCIC), a U.S. Department of Agriculture agency
established to manage the Federal Crop Insurance Program, failed to take in more in premiums
than it paid out in indemnities for the first 8 years of operation. In spite of this, while the FCIC
made adjustments to the program, MPCI was extended to cotton in 1942 and flax in 1945. In
1947 crop insurance premiums finally exceeded indemnity payments. However, in that same
year Congress severely reduced the program size, essentially making it an experimental program.
From 1947 until 1981 the Federal Crop Insurance Program remained a relatively small
experimental program; by 1980 only about half of the counties in the U.S. and just 26 crops were
eligible for insurance coverage.
While crop insurance was in an experimental phase, Congress pursued another method to
address the crop insurance market failure.
The Agriculture and Consumer Protection Act of
1973 and the Rice Production Act of 1975 established a disaster payments program, which, for
some crops, overlapped with the protection provided by federal crop insurance. Essentially,
disaster payments amounted to free universal coverage for a select group of crops. 19 Low-yield
payments were made to farmers who participated in income- and price-support programs when
yields fell below two-thirds of normal. This program was popular with farmers because it
provided catastrophic coverage with no premium. However, detractors claimed that the program
resulted in heavily distorted behavior. In particular, they argued that the moral hazard behavior
of farmers planting on high-risk, environmentally sensitive land imposed a great cost to society.
As a result, the program ended in 1981, having paid out $3.4 billion in total.
19barley, oats, corn, rice, sorghum, upland cotton, and wheat
68
The modem era of federal crop insurance began with passage of the Federal Crop
Insurance Act of 1980. The Act authorized crop insurance expansion to every county in the
United States with significant agriculture and to every crop with sufficient actuarial data to
establish premiums.
In 1981, 252 additional counties received crop insurance, resulting in 1,340
additional county programs (a county program is a particular crop-county combination).
Crop
insurance extended to 1,050 additional counties in 1982, adding 8,278 county programs.
Crop insurance policies were consistently structured from 1980 - 1994. Producers
selected from three guaranteed yield levels (50-, 65-, or 75-percent of their insurable yield) and
from three guaranteed price levels. Price election levels were determined from FCIC forecasts of
expected prices. The top price election level was set at 90-100 percent of the expected market
price.
If the producer's
yield fell below the elected coverage level, the producer received an
indemnity payment equal to the product of the elected price coverage and the yield shortfall.
This yield shortfall was determined by the amount that actual yields fell short of the farm's
insured yield. Determination of the farm's insured yield initially was based on area average
yields.
This method
increased the adverse selection problem as the risk pool became
increasingly filled with farms that had loss-risks above the area average (Skees and Reed 1986).
Consequently, after 1985 a farm's insured yield was based on the preceding ten years of the
farm's actual production data.
Among government-provided
insurance, crop insurance is unique in that it has not
typically overcome adverse selection by mandating participation. Instead the Federal Crop
Insurance Program has sought to entice producers to join by subsidizing premiums. From 1980 1994 the subsidy reduced premiums by 30-percent for the 50- and 65-percent coverage levels,
and by about 17-percent for the 75-percent level.
69
2.1
Two Disaster-Risk Mitigation Programs
The 1980 Act sought to create an insurance program that would replace disaster relief measures.
However, Congress quickly established a precedent2 0 of providing ad hoc, ex post disaster relief,
thereby sustaining the pattern established in the 1970s of having two parallel mechanisms for
dealing with crop-loss risk: crop insurance and disaster payments.
In spite of premium subsidies, enrollment in the Federal Crop Insurance Program was
low and grew little through the 1980s. Table 1 contains measures of program size from 1981 -
2001. From 1981 - 1988 participation remained around 20-percent, increasing to 40-percent in
1989 only as a consequence of mandatory insurance for those receiving disaster payments in
1988.
The implicit guarantee of ex post disaster payments is a key reason for the historically
low participation rates. The U.S. General Accounting Office (1989) reported that, "other federal
disaster assistance programs provide farmers with direct cash payments at no cost to the farmers,
resulting in the perception [among farmers] that crop insurance is unnecessary."
As many others
have pointed out (e.g., Just et al. 1999; Glauber 2004), those who did participate were adversely
selected and more likely to collect indemnity payments.
Table 1 contains the annual loss ratio
(indemnities paid out divided by premiums collected), which averaged 1.47 in 1981 - 1994 and
dropped below 1 only in 1994. The consistent payout of more indemnities that premiums
collected from 1981 - 1994 supports the idea that those who chose to purchase crop insurance
20
In 1986, a severe drought in the Southeast prompted Congress to pass supplemental disaster legislation.
Then, in
1988 an extreme drought hit the Midwest, prompting congress to pass the Disaster Assistance Act of 1988 allocating
$3.5 billion to producers who suffered crop loss. Congress subsequently passed 16 ad hoc, ex post facto disaster
bills between 1989 - 2005.
70
expected a positive return, while those who did not chose to rely on the costless disaster
payments.
2.2
The Federal Crop Insurance Reform Act of 1994
The Federal Crop Insurance Reform Act of 1994 (FCIRA) constituted the most radical change in
the Federal Crop Insurance Program since coverage was expanded to all counties in 1980. The
1994 Act established a mandatory minimum level of crop insurance coverage in order for
producers to qualify for any other government-sponsored program, including subsidy and credit
programs.
For the first time, adverse selection in the crop insurance industry was addressed
through mandatory participation.
All insurable crops had to have at least catastrophic coverage,
which insured production losses falling below 50% of expected yield, indemnified at 60% of the
expected price of the insured crop.
The Act facilitated compliance by authorizing a 100%
subsidy on the catastrophic insurance premiums, requiring producers to simply pay a $50
administrative fee per crop insured. 2 '
The Act also greatly expanded the options and subsidies for coverage beyond the
catastrophic level, termed "buy-up" coverage. The maximum buy-up coverage level increased
from 75- to 85-percent, and several tiers were added to the three coverage levels allowed earlier.
The average subsidy also increased significantly. For example, the subsidy for the 65-percent
yield coverage increased from 30-percent to 41.7-percent of the premium. The subsidy also
varied by the coverage level under FCIRA, with the highest buy-up coverage level (85-percent)
receiving a 13-percent subsidy.
As with the Federal Crop Insurance Act of 1980, the stated goal of FCIRA was to
eliminate ad hoc, ex post disaster payments.
21
In the subsequent three years this goal was
The administrative fee was capped at $200 per producer per county with an absolute cap of $600 per producer.
71
generally accomplished with only $14 million in disaster payments made in 1995, $2 million in
1996, and none in 1997 (see table 1). However, when market conditions collapsed in 1998,
Congress quickly reverted to providing disaster payments ($1.9 billion) in spite of 70-percent
participation in the crop insurance program.
The FCIRA was responsible for a large increase in both the number of acres insured and
the level of coverage.
Figure 1 illustrates the total number of acres insured from 1981 - 2003.
As figure 1 illustrates, the number of insured acres increased dramatically due to the 1994 Act.
The total number of acres insured increased from 99.6 million in 1994 to 220.5 million in 1995.
Of these 220.5 million acres, over half, 115 million acres, were covered by the newly mandated
catastrophic coverage. Coverage levels also increased as farmers took advantage of increased
subsidies.
Mandatory
participation
lasted only one year.
Through
the Federal
Agriculture
Improvement and Reform (FAIR) Act of 1996, Congress modified the crop insurance rules to
allow farmers to opt out of catastrophic coverage.
Table 1 and figure 1 both show a steady
decline of insured acres after 1995 as farmers opted out until 1998 when 40 million fewer acres
were insured.
3
Empirical Strategy
The objective of this paper is to investigate the role of government provided crop insurance in
risk management by comparing the efficacy of disaster payments and crop insurance in reducing
the failure rates of farms that receive government.
I accomplish this by exploiting variation in
weather and the exogenous change in the take-up of crop insurance induced by the Federal Crop
Insurance Act of 1994.
72
Agricultural subsidy recipients were the target of the 1994 crop insurance mandate, and
thus are the focus of my investigation. They also are both politically important and economically
important. Farms that receive agricultural subsidies produce a substantial majority of the value
of all agricultural output. Subsidized farms produce the major export crops, providing 39% of
the world's corn, 19% of the world's cotton, and 28% of the world's wheat exports.
These
farms also are the objects of much policymaker attention. Over the past 20 years Congress has
passed five major agricultural bills primarily aimed at this group of farmers, directing over $305
billion to their support.
The 1994 crop insurance expansion, which expanded coverage by 120%, provides a
source of exogenous variation in crop insurance coverage. By exploiting this exogenous
variation I overcome the endogeneity problem. Farmers who choose to purchase insurance
coverage are likely to be different than their uninsured counterparts in many ways.
Some
differences will be unobservable to the econometrician, thereby resulting in an inconsistent
estimate of the relationship between crop insurance and farm failure rates.
Mandating that all
farms participate in the crop insurance program overcomes this source of bias.
The 1994-1995 time period also provides a stark regime shift between politically-driven
disaster payments and privately-administered
crop insurance.
In the decade prior to 1995
Congress had doled out over $10.3 billion in disaster payments. In 1993, one of the worst
flooding years this century, farmers received nearly $2.5 billion in disaster payments.
The
disaster payments regime was particularly relevant in 1994. It was a relatively normal weather
year in which indemnity payments were the lowest in the decade, yet over half a billion dollars in
disaster payments were made.
The regime shifted drastically in 1995 when farmers were
required to have crop insurance, and disaster payments sharply decreased to $14 million.
73
In
1996 disaster payments fell to $2 million, and in 1997 no disaster payments were made for the
first time in over a decade 2 2 . This dramatic regime shift sets the stage to directly compare the
efficacy of the two primary methods used to address losses caused by extreme weather events.
In spite of the dramatic regime shift, uniform application of the federal policy change
makes it challenging to identifying the relative effect of crop insurance with treatment-control
comparisons. I overcome this obstacle in a novel way by using weather shocks to form the
treatment and control groups. Although specific weather events are generally unpredictable,
long-run weather patterns are not geographically random. For example, the odds of having snow
and freezing temperatures in January in Ithaca, NY are pretty high. I address this issue by using
conditional climate indices to come a close as possible to random assignment. These climate
indices use the county-month specific weather distribution to calculate the conditional cumulative
probability of an observed weather event. In other words, conditional on observing Tompkins
County, NY in January, the climate index allows one to determine the likelihood of observing 48
inches of snow or more. Because the climate indices standardize the measure of an extreme
event, all counties have an equal likelihood of being assigned to the treatment group, i.e. the
treatment group is randomly selected.
Using this empirical approach, I estimate the effects of disaster relief.
Under the
assumption of random assignment, I estimate the effect of disaster assistance under each regime
by simply comparing the average farm failure rate in the control group with that of the treatment
group.
I augment the simple differences with regression analysis in the event that the
'randomization' results in non-equivalent treatment and control groups.
This will reduce
sampling error and provide a check on the random assignment assertion. To further check the
22 This
dramatic shift was primarily caused by moving disaster payments from off-budget to on-budget, thereby
requiring proponents of disaster payments to find cuts elsewhere in the budget to compensate for the disaster
expenditures.
74
random assignment assertion I design a difference-in-differences estimator to provide an
alternate estimate of the effects of disaster relief under the crop insurance regime.
Finally, these two exogenous events, weather and mandatory crop insurance, allow me to
create a simple difference-in-differences
estimator to calculate the relative effectiveness of crop
insurance versus disaster payments at smoothing exogenous shocks by reducing farm failure
rates.
The identifying assumption of this estimation strategy is that there are no shocks in
disaster stricken counties that are associated with the regime change but not caused by the
regime change. If extreme weather events are truly randomly assigned over time, then the
identifying assumption is a fairly weak one.23
4
Data
4.1
Farm Failure Rates
I use USDA administrative data obtained under the Freedom of Information Act in order to
construct annual, county-level farm failure rates. The data consist of every payment made by the
Farm Services Agency (FSA) from 1990 until 2001.
Agriculture's
(USDA)
administrative
arm, which
programs, as well as disaster relief programs.
The FSA is the U.S. Department of
administers
price
and income support
Although the data are not explicitly an
enumeration of every farm in the United States it does contain information on the population of
interest, every farm that received agricultural subsidies.
If a farm is observed at two points in
A means of addressing concerns about this assumption is to compare the outcomes of subsidized crop farms to
farms in the same county that were unaffected by the crop insurance reforms. Two groups that qualify as withintreatment-controls are farms that primarily raise livestock and farms that grow non-insured crops, i.e. crops for
which insurance is unavailable through the Federal Crop Insurance Program, typically fruits and vegetables.
Livestock farms did not experience the crop insurance regime shift; they continued to qualify for the Livestock
E mergency Assistance, a disaster payment-type program. Non-insured crops had a slight change to their disaster
payments mechanism, but it was essentially the same. (See Lee et al. (1996) for discussion of the differences.)
Performing this difference-in-differences-in-differences
estimation requires data on livestock or non-insured crop
farm failure rates, which are unavailable.
23
75
time, I extrapolate the farm's existence during the intervening period. If a farm drops out of the
dataset and is never observed again, I label that farm as having failed in the last year observed.
A key feature of these data is that in 1996 virtually all farms that qualified for subsidies received
subsidy payments 2 4 . Hence, farms that may have chosen to opt-out of the subsidy program in
1994 or 1995 (when the participation rate was about 85%) are not counted as failures. Figure 3
illustrates the annual farm failure rate for this population.
4.2
Disaster Classification
The exact definition of an agricultural disaster is nebulous at best. Ex post measures such as crop
yield loss presumably demonstrate the effects of weather shocks, but these measures are also
determined by farmer behavior, making them endogenous to the question at hand. A much more
satisfying classification is to quantify weather shocks as they occur. However, the heterogeneity
in climate makes it impossible to use temperature and precipitation directly. Ten inches of rain
in July in Florida will have a much different effect than ten inches in Montana.
Therefore, in
classifying drought and excess precipitation I employ the Palmer Drought Severity Index (PDSI)
and the Standardized Precipitation Index (SPI). These indices are widely used by climatologists.
For example, the U.S. Drought Monitor and state drought monitors use these indices to monitor
drought conditions, the National Weather Services Climate Prediction Center uses the PDSI as
their primary measure of drought, and disaster legislation often uses the PDSI to trigger disaster
declarations.
The indices also allow one to easily and directly compare weather shocks across
different climates, e.g. Florida and Montana. The indices also provide a standardized measure of
24
This feature is important. Farms may temporarily drop out of the dataset without having actually failed, but a
failed farm cannot temporarily show up in the data. This asymmetry potentially induces one-sided measurement
error in the dependent variable. Since virtually all farms are observed in 1996 the measurement error caused by
some farms temporarily opting out is dramatically reduced.
76
weather extremes.
By these measures a farm in drought-prone West Texas is no more likely to
experience a 'disaster' than a corn farm in Iowa.
4.2.1
Precipitation Indices
The Standard Precipitation Index provides a standardized measure of precipitation relative to the
county-month
specific time-series distribution
of precipitation.
The SPI is essentially a z-
statistic constructed by transforming the location-specific time-series precipitation distribution
into a standard normal distribution.
One can then confidently speak about the likelihood of an
event at least as extreme as the one observed and make geographic comparisons of precipitation-
related weather events. The units of the precipitation distribution can be varied. If one were
interested in long run precipitation anomalies one could look at the distribution of one- or five-
year total precipitation. Because agriculture is very sensitive to short run weather shocks, I
chose to examine the two-month total precipitation distribution for this analysis. The first panel
of table 2 contains the event scale commonly associated with the SPI.
The Palmer Drought Severity Index incorporates temperature, precipitation, soil
characteristics, and location to develop a supply and demand model for water. A drought is
classified by the extent of excess demand for water, while excess wetness classification depends
on the excess supply of water. The PDSI is a widely used measure of drought that considers the
rainfall history, making it less sensitive to transitory events than the SPI. However, since
agriculture responds to short-term transitory weather shocks, I also incorporate an index used to
create the PDSI, the moisture anomaly index, generally know as the "Z" index. The Z index is
the difference between the actual precipitation that fell in a specific month and the amount of
precipitation needed to maintain normal soil moisture, adjusted for the climatic characteristics of
77
the location. Essentially the Z index can be used to show how wet or dry it was during a single
month without regard to recent precipitation trends. The commonly used classification system
for the Z-Index and the PDSI are shown in the panel B and C respectively of table 2.
Using Unix programs available from the National Agricultural Decision Support System
and the National Drought Monitor, I calculate the PDSI, its components, and the SPI for each
county in the United States. The publicly available datasets typically report the PDSI and the
SPI at the climate division, a level of geography than encompasses multiple counties. Climate
division data are best used "to assess large-scale climatic features or anomalies with respect to a
long period or century-scale perspective" (Guttman 1995). Although agriculture certainly suffers
from such large-scale events, even smaller scale, relatively local weather events can have
adverse effects. In order to exploit the smaller scale variation in weather, I calculate the PDSI
and the SPI on the county level by using county-level weather data derived from the Parameterelevation Regressions on Independent Slopes Model (PRISM) and county-level soil
characteristics from the State Soil Survey Database (STATSGO) 2 5 .
The PRISM modeling system generates estimates of precipitation and temperature at 4x4
kilometer grid cells for the entire United States. These estimates are based upon data collected
by the National Oceanic & Atmospheric Administration (NOAA) Cooperative Weather Stations
at over 3,000 locations nationwide from 1970 - 2001. In order to focus on the weather occurring
on cropland these data are used in conjunction with data on the land use within each 4x4 km grid
to calculate the average temperature and total precipitation for cropland.
The PDSI also requires a measure of the available water capacity (AWC) of the soil in
the location where the temperature and precipitation are collected.
25
The AWC is the amount of
1 am grateful to Olivier Deschenes and Michael Greenstone for generously providing the PRISM data used in this
analysis.
78
water that a soil can store that is available for use by plants (USDA-NRCS 1998). I calculate the
average AWC for all cropland in each county using STATSGO data. STATSGO is a dataset
collected by the US Department of Agriculture - National Soil Conservation Service containing
soil characteristics, up to 6 layers deep, across the entire United States. In order to determine the
characteristics of soil associated with cropland, I combine STATSGO with the National
Resource Inventory (NRI), which collects data every five years on the land use of 844,000 points
across the entire United States. I employ the land use data from 1992, the most recent year prior
to the policy reform.
The final input into the PDSI calculation is the latitude corresponding to each data point.
To obtain these data I use geographical information system (GIS) software to calculate the
latitude of each county centroid.
4.2.2
Disaster Classification
Establishing objective disaster criteria is vital to the analysis that follows.
myriad of factors contribute to disaster, of which weather is just one.
Unfortunately, a
Consequently, the
climatology literature contains a set of possible criteria, but not a clearly defined standard. I
draw upon that literature and rely upon the three indices outlined above in order to develop a
'disaster/non-disaster'
classification.
The PDSI, the Z-Index, and the SPI capture the short- to
medium-term moisture events that are most applicable to agriculture. The National Climate
Center uses these three indices as the core of its 'Short-Term Blend Weather Monitor' used to
monitor weather events that could have significant effects on agriculture. 2 6 As noted earlier,
table 2 contains the classification system of each index.
26 http://www.cpc.ncep.noaa.gov/products/predictions/experimental/edb/droughtblend-access-page.html
79
I ultimately use all three indices in order to establish the disaster classification.
I
calculate a county specific index for each month of the year. Because of my focus on extreme
events I calculate the most extreme positive value of a county's index across all twelve months
within a year, and I calculate the most extreme negative value of a county's index across all
twelve months. Disasters occurring during the growing season have a significantly greater
impact on agriculture than disasters during the non-growing season. Therefore I limit my focus
to extreme index values occurring between May and October.
Once I calculate the relevant
extremes, I classify a county as having experienced a disaster if the PDSI and the Z-Index both
exceed the 'severe' threshold on their respective scales, and the SPI value has a less than 5-
percent likelihood. These criteria result in 795, 355, and 476 counties classified as disaster
counties in 1993, 1994, and 1995 respectively.
The cross-sectional distributions of the moisture indices are shown in table 3. For the
question at hand the interesting part of the distribution is not the center, as is usually reported,
but the tails, i.e. the extreme events. In order to take a closer look at the tails, the upper panel of
table 3 presents the
5th,
h
15 t ,
average of each county's
h
8 5t ,
and
95 th
percentiles.
annual monthly-maximum
county's annual monthly-minimum
The lower panel of table 3 shows the
index value and the average of each
index value for 1993, 1994, and 1995.
Each of the indices
characteristics can be seen in table 3. The Z-Index is most responsive to monthly precipitation
and registers the greatest number of extreme values. The PDSI is a medium-term measure that
adjusts slowly, as demonstrated by the continual downward trend in the monthly maximum
PDSI, even though both the Z-Index and the SPI increase in 1995. The SPI has responsiveness
between the Z-Index and PDSI, being very responsive to significant weather events.
80
The three panels of figures 4, 5, and 6 each show the geographic distribution of the three
weather indices in 1993, 1994, and 1995 respectively.
from table 3:
Climatologically,
These figures underscore the information
1993 was an extremely wet year, whereas 1994 was a relatively normal year.
1993 and 1995 were fairly similar, wet years.
In the analysis that follows I
will demonstrate the comparison between 1993 and 1995 as well as 1994 and 1995.
5
Empirical Results
5.1
Overview
In this section I exploit two natural experiments in order to examine the 'survival smoothing
'2 7
effects of disaster payments and crop insurance, the two most heavily used means of crop-loss
insurance. I create randomized experiments by taking advantage of exogenous extreme weather
events. In both natural experiments examined below I rely on random weather events to classify
counties into disaster (treatment) and non-disaster (control) groups.
Random assignment
allows me to avoid the endogeneity
concerns associated with
directly analyzing the impact of disaster or indemnity payments on farm failure rates. Farmer
behavior associated with adverse selection and moral hazard affect the loss-likelihood and
thereby determine the size of the disaster or indemnity payment.
This same behavior also is
plausibly correlated with the propensity to fail, thereby resulting in a spurious, non-causal
relationship between payments and failure rates. By relying on the random component of croploss payments I avoid these concerns.
27
1 adapt this term from the consumption smoothing effects that are examined in the social insurance literature.
Since I am focused on failure rates, I deem a constant failure rate across states of nature 'complete survival
smoothing.'
81
In addition to calculating the survival smoothing effects of disaster payments and crop
insurance, one would like to examine the absolute effects of these two programs on farm failure
rates. For example, policy makers would benefit from knowing the decrease in farm failure rates
when a farmer moves from an uninsured state to being covered by catastrophic crop insurance.
Unfortunately
for this investigation,
there has not been a time when farmers have not been
potentially covered by disaster payments, crop insurance, or both.2 8 Nevertheless, because of the
virtual elimination of disaster payments and the increased insurance coverage in 1995 I can
compare the relative effectiveness of these two programs. This is a fundamentally relevant
policy question as Congress has operated these two programs in parallel intermittently for the
past 35 years with frequent debate about the relative benefits.
In the analysis that follows I first examine each of these programs separately, and then I
examine their relative effectiveness and test the survival smoothing hypothesis.
5.2
Randomized Experiments
5.2.1
Randomization
Random weather shocks play an intricate role in agricultural yield reduction. The type, severity,
and timing of weather events interact in complicated, non-linear ways to influence output
(Roberts and Schlenker, 2005). For example, a wet spell can have dramatically different effects
on wheat yield if it occurs in early May instead of later June. Consequently, it is very difficult to
develop a single metric to quantify a disaster. In order to ensure that the classification used in
this paper corresponds to actual yield losses, I present the average per-acre disaster payment and
the average proportion of crop liabilities indemnified for each treatment and control group.
28
In future work I will exploit yet to be obtained data and the dramatic expansion in crop insurance availability due
to the 1980 Federal Crop Insurance Act to answer this question.
82
Panel A of table 5 reports the average per-acre disaster payments in 1993 and 1994 for
the disaster and non-disaster counties. The standard errors are in parenthesis, and the group size
is in brackets.
The disaster/non-disaster
difference is also presented, along with its standard
error. The average per-acre disaster payment for disaster counties in 1994 was $5.11. In 1993
the amount was nearly double that at $10.34, due to the worst flooding in the Midwestern United
States this century.
For the non-disaster counties the average per-acre disaster payment was
$2.90 and $8.54 for 1994 and 1993 respectively.
In both cases the disaster counties receive
significantly greater disaster payments, validating the classification scheme.
The presence of substantial disaster payments in non-disaster counties underscores the
importance of using exogenous weather events in this investigation. Although some of the
imperfect correlation between disaster payments and weather is surely due the complicated
relationship between yield and weather, moral hazard effects and fraud provide an additional,
more insidious explanation for the magnitude disaster payments to non-disaster counties. A
somewhat less nefarious explanation is also viable. Disaster payments require Congressional
action, and the political economy may result in large "disaster" payments directed to the
constituents of key politicians. Garrett et al. (2003) demonstrate that between 1992 and 1999
about $7 billion of disaster aid was the result of political influence rather than due to disaster
losses. This political economy may interact with moral hazard to magnify the effects. Others
also have noted that counties that do not objectively appear to experience a disaster receive
considerable disaster payments (Lee et al. 1996).
Panel B of table 5 demonstrate that disaster counties received greater indemnities as a
proportion of total liabilities than did non-disaster counties.
In 1995 11.2-percent of crop
insurance liabilities were indemnified in disaster counties, compared with 8.2 percent in non-
83
disaster counties. This again shows high, but imperfect, correlation between actual weather
events and compensated losses.
In a randomized experiment, the treatment groups should be statistically similar. Table 4
the disaster and non-disaster
county characteristics
as measured
in the 1992 Census of
Agriculture. With a few rare exceptions, the treatment and control groups look statistically
identical, further supporting the quasi-random experimental design.
5.2.2
Disaster Payments
Using the quasi-random experiment described above I evaluate the effect of ad hoc disaster
payments on farm failure rates by simply comparing the average farm failure rate of the
randomly assigned disaster counties to the average farm failure rate of the non-disaster counties.
I make this comparison in two ways. First, table 5 panel C shows the averages failure rate of
each group and the failure rate difference between the two groups.
displayed in parentheses, and the group size is in brackets.
The standard errors are
The average failure rate in non-
disaster counties is 6.5-percent in 1994 and 5.4-percent in 1993. The average failure rate in
disaster counties is 5.9-percent and 5.0-percent in 1994 and 1993 respectively.
Both the 1994
failure rate difference between groups and the 1993 difference show that under the disaster
payments regime farms were slightly less likely to fail when hit by a disaster. The difference is
small, but statistically significant, suggesting that disaster payments help sustain farms that
otherwise would fail, regardless
of whether they faced a disaster.
In terms of survival
smoothing, these findings suggest that disaster payments are not simply a substitute for other
means of self-insurance. Clearly, government intervention aids in smoothing temporary shocks.
84
An alternative method for analyzing this experiment is with an ordinary least squares
regression.
This approach provides a check on the contention that treatment is quasi-random.
Under true randomization other observed county characteristics should not influence the
estimate, but the sampling variability should decrease when conditioning on covariates. Panel A
in table 6 lays out the regression results using the county-crop mix as controls. The regression
results are consistent with the simple differences outlined above, although the estimate changes
somewhat with the inclusion of controls, suggesting some caution should be used when
interpreting this in a strictly experimental sense.
Although disaster payments provided the vast majority of assistance in 1993 and 1994, it
should be noted that the disaster payments regime included some farms that also held crop
insurance.
Disaster payments were more than twice the net indemnities (indemnities minus
premiums) paid by crop insurance in 1993, and net indemnities were zero in 1994. Hence, the
conclusion that disaster payments provide for survival smoothing is a statement about the entire
regime, in which ex post disaster payments play a dominant role.
5.2.3
Crop Insurance
I also examine the influence of crop insurance on farm failure rates through a quasi-randomized
natural experiment.
As with the experiment investigating the disaster payments regime, I
categorize counties into disaster and non-disaster groups using exogenous, random weather
events. The lower half of panel C in table 5 contains the results of this experiment.
The results
from this experiment are generally consistent with an increased failure rate when faced by a
disaster under the crop insurance regime. Farms with crop insurance have a 1.3 percentage point
85
increased failure rate when hit by a disaster. This raises the average failure rate to 8.3-percent
from a 7-percent failure rate when not hit by a disaster.
These findings suggest that crop insurance fails to provide complete survival smoothing.
High deductibles provide a plausible explanation. Deductibles ranged from 35-percent to 50percent of production.
The most widely held coverage level, the mandated catastrophic
coverage, had a 50-percent of production deductible. Such a large deductible possibly provided
an insufficient buffer for some farmers, resulting in permanent effects of large, temporary shocks
(Basker, 2003).
As with the disaster payments regime, I use ordinary least squares regressions to augment
the experimental approach by including covariates in the analysis. The third row of panel A in
table 6 contains the regression estimates.
Once covariates are included in the analysis the
coefficient decreases to 0.7, becoming statistically insignificant in spite of the slightly smaller
standard error. This estimate signifies nearly no change in failure rates due to disaster under the
crop insurance regime, although the original, simple-difference
estimate is still within the
confidence interval.
5.3
Relative Effectiveness
One would like to use the 1994-1995 policy change that more than doubled the rate of insurance
coverage to examine the absolute effect of crop insurance coverage. As noted earlier, this is
unfortunately not possible because ad hoc, ex post disaster payments were the norm prior to the
1994 Act. Instead, it is informative to compare the effectiveness of the disaster payments regime
to that of the crop insurance regime. Knowing their relative effectiveness at decreasing farm
86
failures in the face of disaster is the first step to understanding the incentive effects of each
regime and to designing an optimal government-provided crop-loss insurance scheme.
I evaluate the relative effectiveness of the two regimes by comparing the outcomes of the
previous
two quasi-random
experiments.
Panel C in table 5 contains the results of the
experiments in 1993, 1994, and 1995. I report the relative effects of the two regimes in the last
two rows of panel C by taking the difference between each of the disaster payment regime results
and the 1995 crop insurance regime results.
These difference-in-differences estimates
demonstrate the relative effectiveness of these two programs.
The results of these comparisons are remarkably similar. Relative to the 1993 regime,
farms suffering a disaster in 1995 are 1.7 percentage points more likely to fail. That likelihood is
1.9 percentage points when 1994 is the base period.
Both of these results are statistically
significant.
I extend my analysis of the relative effects controlling for observable county
characteristics in a regression framework. As with the quasi-random experiments, any failure of
the randomization resulting in treatment group heterogeneity could be problematic. Through
regression analysis I also address the concern that the composition of the treatment group is
allowed to differ between the two periods by controlling for other observables that affect failure
rates. The regression equation has the following form:
Fi, =a + , X, + 82r, +
3disasteri,+
4 (r,x disasteri,).
In this equation i indexes counties and t indexes years (1 if after the policy change, 0 if before).
F;, is the farm failure rate, X is a vector of observable county characteristics,
effect, and disaster is a dummy variable for disaster counties in either regime.
87
is a fixed time
The form of the regression resembles the typical difference-in-differences regression.
However, the interpretation of the coefficients is slightly different because the treatment group is
not the same in both periods. The time fixed effect controls for time-series changes in the failure
rate. The disaster variable captures the effect of a disaster in the first period while the second
order interaction term captures the relative effect between the two regimes. The set of covariates
includes the proportion of cropland planted to 12 different crops.
Panel B of table 6 reports the coefficients of the interaction term for the various criteria.
Controlling for covariates, I cannot reject the hypothesis that there is no change in farm failure
rates when moving from the 1993 disaster payments regime to the 1995 crop insurance regime.
However, relative to the 1994 disaster payment regime, a farmer is 1.2 percentage points more
likely to fail with crop insurance in 1995 after controlling for crop mix. This estimate is similar
to the point estimate of 1.9 percentage points from the simple difference-in-differences
6
estimate.
Interpretation and Conclusion
The analysis above found a decreased failure rate for the disaster payments regime, and less than
complete smoothing for the crop insurance regime.
I estimate that, relative to the disaster
payments regime, farm failure rates increase by between 1.7 - 1.9 percentage points under the
crop insurance regime. From an average failure rate of 7.2-percent this change represents a 30-
percent increase in the likelihood of failure when faced with a natural disaster under the crop
insurance regime.
This is a surprising finding for several reasons. First, the ultimate structure of ex post
disaster payments and mandatory catastrophic crop insurance is very similar. Ex post disaster
payments provided coverage of loss greater than 65-percent of yield at 60-percent of the
88
expected price. The catastrophic coverage was less generous, covering only 50-percent of crop
lass at 60-percent of the expected price. This reduced generosity might play a factor in the
findings, but one would expect the overall coverage increase due to substantially increased
premium subsidies to have had an off-setting effect.
An alternative explanation could be in the moral hazard incentives that are likely more
substantial with an ex ante insurance plan. Ex ante insurance provides incentives for the farmer
to shirk early on in the growing season, when reduced tillage, chemical treatment, or fertilizer
will have the greatest impact. Coverage uncertainty under the ex post disaster payments regime
reduces the incentives for moral hazard behavior. Even if disaster payments were implicitly
anticipated, the Congressional action necessary to activate coverage is likely somewhat
capricious. Under this sort of coverage, the moral hazard incentives might only set in late in the
season when it became certain that Congress would act.
Delaying moral hazard may be a factor contributing to the greater effectiveness of the
disaster payments regime. More data are required in order to investigate this explanation.
Obtaining data on fertilizer and chemical input will help in evaluating these claims. 2 9
Alternative explanations are also viable. Fraud might play a role if it is better addressed
under the crop insurance regime than under ex post disaster payments.
The Federal Crop
Insurance Program is operated primarily through private insurers that are reinsured by the U.S.
Department of Agriculture. The private insurers might be better at policing and weeding out
fraudulent claims. If this is the case and if fraudulent claims are made by farms that are more
likely to fail, then fraud could be a factor in the finding.
In other words, the greater likelihood of farm survival under the disaster payments regime
is not necessarily a good thing. Perhaps the greatest cost of an ad hoc, ex post disaster payments
29
The author is currently pursuing use of confidential survey data in order to address this issue.
89
program is in terms of the disutility of farmers forced to bear the ex ante risk of failure due to
random weather events. Public policy will greatly benefit from research more closely examining
the potential costs associated with these program structures, in addition to the moral hazard and
adverse selection costs studied by others and the survival smoothing benefits outlined in this
paper.
Although the absolute effects of crop insurance and disaster payments remain unknown,
finding a positive relative effect allows me to conclude that disaster payments do provide
survival smoothing and do not perfectly crowd out other means of risk mitigation. This finding
suggests that there is a significant role for government-provision of crop-loss insurance.
90
References
Babcock, B.A., and D.A. Hennessy. 1996. "Input Demand under Yield and Revenue Insurance."
American Journal ofAgricultural Economics 78(2), 416-427.
Barkley, E. P. 1990. "The Determinants of the Migration of Labor out of Agriculture in the
United States, 1940 - 1985." American Journal of Agricultural Economics 72(3), 427-435.
Basker, E. 2002. "Do Temporary Shocks have Permanent Effects?
Agricultural Sector." Unpublished PhD Dissertation, MIT.
Evidence from the
Claasen, R., R.N. Lubowski, and M.J. Roberts. 2005. "Extent, Location, and Characteristics of
Land Cropped Due to Insurance Subsidies." Selected paper prepared for the American
Agricultural Economics Association Annual Meeting.
Garrett, T.A., T.L. Marsh, and M.I. Marshall. 2003. "Political Allocation of Agricultural
Disaster Payments in the 1990s." Federal Reserve Bank of St. Louis Working Paper 2003005A.
Glauber, J.W. 2004. "Crop Insurance Reconsidered."
American Journal of Agricultural
Economics 86(5), 1179-1195.
Goodwin, B.K., and V.H. Smith. 1995. The Economics of Crop Insurance and Disaster Aid.
Washington, DC: The AEI Press.
Goodwin, B.K., and V.H. Smith. 2003. "An Ex Post Evaluation of the Conservation Reserve,
Federal Crop Insurance and Other Government Programs: Program Participation and Soil
Erosion." Journal ofAgricultural and Resource Economics 28(, 201-216.
Goodwin, B.K., M.L. Vandeveer, and J. Deal. 2004. "An Empirical Analysis of Acreage Effects
of Participation in the Federal Crop Insurance Program." American Journal of Agricultural
Economics 86(4), 1058-1077.
Huffman, W. E. and R. E. Evenson. 2001. "Structural
and Productivity
Change in US
Agriculture, 1950 - 1982." Agricultural Economics 24(2), 127-147.
Just, R.E., L. Calvin, and J. Quiggin. 1999. "Adverse Selection in Crop Insurance: Actuarial and
Asymmetric Information Incentives." American Journal of Agricultural Economics 81(4),
834-849.
Kiyotaki, N. and J. Moore. 1999. "Indexation and Insurance: A Stochastic Model of Credit
Cycles." Working Paper, London School of Economics.
Krishnamurthy, A. 2003. "Collateral Constrains and the Amplification Mechanism." Journal of
Economic Theory 111()277-292.
91
Lee, H., J. Harwood, A. Somwaru. 1997. " Implications of Disaster Assistance Reform for NonInsured Crops." American Journal of Agricultural Economics 79(2), 419-429.
Quiggen, J., G. Karagiannis, and J. Stanton. 1993. "Crop Insurance and Crop Production:
Empirical Study of Moral Hazard and Adverse Selection." Australian Journal of Agricultural
Economics 37(, 95-113.
Skees, J.R., and M.R. Reed.
1986. "Rate-Making for Farm-Level Crop Insurance: Implications
for Adverse Selection." American Journal of Agricultural Economics 68(3), 653-659.
U.S. General Accounting Office. 1989. "Disaster Assistance: Crop Insurance Can Provide
Assistance More Effectively than Other Programs." GAO/RCED-89-211.
Wu, J. 1999. "Crop Insurance, Acreage Decisions, and Nonpoint-Source Pollution." American
Journal of AgriculturalEconomics81(2), 305-320.
92
Table 1. Crop Insurance and Disaster Payment Details
Crop
Total
Year
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
Acres
Policies
(thousands) -- (mil)
45.0
416.8
42.7
386.0
27.9
310.0
42.7
389.8
414.6
48.6
48.7
406.9
433.9
49.1
461.0
55.6
949.7
101.7
893.7
1991
706.2
1992
1993
1994
663.1
678.8
1995
1996
1997
1998
1999
2000
2001
2002
800.4
2,039.2
1,623.3
1,320.1
1,242.4
1,287.9
1,323.2
1,297.9
1,259.4
101.3
82.3
83.1
83.7
99.6
220.6
205.0
181.9
181.7
196.3
206.5
211.8
215.5
Participation Premium Indemnity
Rate
($ mil)
($ mil)
16
15
12
16
18
20
22
25
40
40
33
31
32
38
41/85
44/76
43/67
46/69
53/73
59/76
63/78
66/80
Total
Loss
Ratio
376.8
396.1
285.8
433.9
439.8
379.7
365.1
436.4
819.4
835.5
736.4
758.7
755.6
407.3
529.1
583.7
638.4
683.1
615.7
369.8
1,067.6
1,215.3
1,033.6
958.5
922.5
1,653.4
1.08
1.34
2.04
1.47
1.55
1.62
948.9
600.9
1,543.0
1,838.4
1,773.8
1,874.2
2,304.4
2,540.1
1,566.5
1,491.0
990.1
1,659.3
2,416.1
0.63
1.02
0.81
0.56
0.89
1.05
1.02
1.00
1.39
2,961.9
2,915.9
2,591.0
2,959.4
4,054.0
Disaster
Payments
($ mil)
306
115
1
0
0
556
1.01
10
2.45
1.48
1.24
1.30
1.22
3,386
2.19
1,500
6
960
872
2,461
577
14
2
0
1,913
1,452
2,348
0
2,145
Notes: Source: U.S. Department of Agriculture - Risk Management Agency.
Participation rate calculated as acres insured as percent of eligible acres. For 1995 - 2002, the first number is the
participation in "buy-up" and the second number is total parti
93
Table 2. Climate Indices
Index Value
Classification
A. Standardized Precipitation Index
2.0+
Extremely Wet
1.5 to 1.99
Very Wet
1.0 to 1.49
Moderately Wet
-0.99 to 0.99
Near Normal
-1.0 to -1.49
-1.5 to -1.99
Moderately Dry
Severely Dry
-2.0 and less
Extremely Dry
B. Moisture Anomaly (Z) Index
3.50+
Extremely Wet
2.5 to 3.49
1 to 2.49
Very Wet
Moderately Wet
-1.24 to 1
Near Normal
-1.25 to -1.99
-2.0 to -2.75
-2.75 and less
Moderate Drought
Severe Drought
Extreme Drought
C. Palmer Drought Severity Index
4.0+
Extremely Wet
3.0 to 3.99
Very Wet
2.0 to 2.99
Moderately Wet
1.0 to 1.99
Slightly Wet
0.5 to 0.99
Incipient Wet Spell
Near Normal
0.49 to -0.49
-0.5
-1.0
-2.0
-3.0
to
to
to
to
-0.99
-1.99
-2.99
-3.99
-4.0 and less
Incipient Dry Spell
Mild Drought
Moderate Drought
Severe Drought
Extreme Drought
94
-
Table 3. Climate Indices' Distributions
5%
-1.21
-2.67
-2.29
SPI
PDSI
Z-lndex
June 1995
15%
85%
-0.59
1.6
-1.41
2.93
-1.43
2.42
95%
2.03
3.68
4.1
1993
1994
1995
SPI
-1.29
(0.61)
-1.38
(0.49)
-1.47
(0.58)
PDSI
-0.29
-0.50
-0.71
Z-Index
(1.85)
-1.81
(0.96)
(1.86)
-2.05
(0.63)
(1.82)
-2.03
(0.77)
1.54
(0.69)
2.56
1.38
(0.65)
2.04
1.48
(0.52)
1.87
(1.88)
3.43
(2.02)
(1.83)
3.04
(1.77)
(1.67)
3.28
(1.59)
Minimum Monthly
Maximum Monthly
SPI
PDSI
Z-Index
Notes: The climate indices are: the Standardized Precipitation
Index (SPI), the Palmer Drought Severity Index (PDSI), and the
Palmer Moisture Anomaly (Z) Index (Z-Index). The climate
indices are calculated for each month of the year. The upper
panel contains the indicated percentiles of each index in June,
1995. The cells in the lower panel contain the cross-sectional
average of the minimum and maximum value across all 12
months. The standard deviations are in parentheses. See table
3 for the magnitude classification of each index.
95
Table 4. Characteristics of Disaster and Non-Disaster Counties
Disaster Criteria--Palmer Moisture Anomaly (Z) Index &
Palmer Drought Severity Index & Standardized Precipitation Index
Characteristic
Farm Size (acres)
Land Value ($/acre)
Sales ($)
Proportion of Cropland
Barley
Corn
Cotton
Soybeans
Oats
Sorghum
Wheat
Rice
Beans
Sunflowers
Other
Proportion of Farms
Cattle
Beef Cows
Milking Cows
Hogs and Pigs
Sheep and Lambs
Layers
Broilers
Proportion of Farmland
Rented Farmland
Irrigated Farmland
Farmer Characteristics
Experience
Age
Full Owner
Part Owner
Tenant
Primary Occupation Farming
Family Farm
Partnership
Family Corporation
Corporation
Disaster Payments
(1993)
NonDisaster
Disaster
Disaster Payments
(1994)
NonDisaster
Disaster
Crop Insurance (1995)
NonDisaster
Disaster
521.361
732.923
1,093.90 1,160.12
77,850.50 87,233.88
885.787
1,056.58
97,798.61
666.612
1,170.80
83,785.73
0.031
0.261
0.082
0.205
0.030
0.023
0.181
636.722
956.25
90,673.57
741.739
1,227.76
84,168.31
0.022
0.329
0.006
0.252
0.028
0.043
0.224
0.003
0.003
0.007
0.017
0.037
0.233
0.074
0.174
0.040
0.024
0.213
0.012
0.008
0.002
0.048
0.022
0.257
0.095
0.154
0.044
0.014
0.182
0.001
0.007
0.010
0.089
0.033
0.280
0.053
0.201
0.036
0.030
0.207
0.012
0.007
0.004
0.031
0.033
0.228
0.037
0.186
0.035
0.042
0.286
0.006
0.008
0.553
0.416
0.076
0.152
0.051
0.039
0.006
0.559
0.438
0.069
0.073
0.044
0.051
0.015
0.553
0.419
0.084
0.101
0.031
0.048
0.016
0.557
0.433
0.069
0.094
0.048
0.047
0.012
0.593
0.454
0.085
0.053
0.050
0.010
0.550
0.427
0.068
0.094
0.045
0.047
0.013
0.449
0.047
0.363
0.074
0.336
0.046
0.395
0.069
0.398
0.085
0.386
0.062
20.6
52.4
0.514
0.356
0.131
0.612
0.856
0.095
0.039
0.004
19.4
19.8
19.8
53.9
0.587
0.306
0.107
0.522
0.854
0.097
0.037
0.005
53.5
0.576
0.327
0.098
0.539
0.874
0.086
0.029
0.004
53.4
0.558
0.324
0.119
0.556
0.852
0.098
0.038
0.004
20.3
53.4
0.530
19.7
53.4
0.566
0.319
0.115
0.546
0.855
0.098
0.036
0.004
0.011
0.033
0.101
0.351
0.119
0.591
0.855
0.092
0.041
0.004
0.011
0.006
0.004
0.039
.
.
Notes: Data for all groups are from the 1992 Census of Agriculture, except data for crops data which are from the U.S.
Department of Agriculture-National Agricultural Statistics Service Crop County Files for each of the years listed. See text for
criteria used to assign counties to disaster and non-disaster groups.
96
.
Table 5. Quasi-Random Experiments:
The Effects of Disaster Payments and Crop Insurance on Farm Failure Rates
Disaster Criteria--Palmer Moisture Anomaly (Z) Index,
Palmer Drought Severity Index & Standardized Precipitation Index
NonDisaster/Non-Disaster
difference
Disaster Disaster
Regime
A. Per-Acre Disaster Payments
8.567
10.336
Disaster Payments (1993)
(0.429)
(0.449)
Disaster Payments (1994)
{795}
{2007}
5.113
(0.485)
2.904
(0.199)
{355}
{2440}
B. Proportion of Liabilities Indemnified
0.112
0.082
Crop Insurance (1995)
(0.003)
(0.007)
{465}
1.770
(0.621)
2.209
(0.524)
0.030
(0.007)
{2239}
C. Causal Effect of
Disaster Payments and Crop Insurance on Farm Failure Rates
-0.004
0.050
0.054
Disaster Payments (1993)
(0.001)
(0.002)
(0.001)
Disaster Payments (1994)
Crop Insurance (1995)
{795}
{2007}
0.059
(0.002)
0.065
(0.001)
{355}
{2440}
0.083
0.070
0.013
(0.005)
{476}
(0.002)
{2314}
(0.005)
-0.006
(0.002)
1993 - 1995 Relative Effect
0.017
(0.005)
1994 - 1995 Relative Effect
0.019
(0.006)
Notes: Standard errors are given in parentheses. The disaster classification corresponds to
the 'severe' classification of each climate index: Z-Index > 2.5 or < 2, PDSI > 3 or < -3, and
SPI > 1.5 or < -1.5. Panel A contains the average dollars per acre in disaster payments for
each group in the disaster payments regime. Panel B contains the ratio of indemnitied to
liabilities for each group in the crop insurance regime. The cells in panel C contain the
average farm failure rate for each the group during each regime.
97
Table 6. Treatment-Dummy Results
Across Regimes
Regime
Failure Rate
A. Quasi-Random Cross-sectional Experiments
Disaster Payments (1993)
-0.004
0.003
(0.002)
(0.002)
Disaster Payments (1994)
-0.006
(0.002)
-0.006
(0.002)
Crop Insurance (1995)
0.013
(0.005)
N
0.007
Covariates
(0.005)
Y
B. Relative Program Effects
1993 - 1995 Relative Effect
0.017
(0.005)
0.003
(0.005)
1994 - 1995 Relative Effect
0.019
(0.006)
0.012
(0.005)
Notes: White's heteroskedastic-robust standard errors are in
parentheses. The dependent variable is the county-level farm
failure rate. In panel A the coefficient is that on the 'disaster
county' indicator in the cross-sectional analysis for each
regime. Panel B contains the coefficient on the interaction
term when comparing the two years. The covariates are the
proportion of cropland planted to 12 different crop groups.
98
250.0
200.0
5 150.0
C 100.0
50.0
0.0
I
1
I
I
I
9
f
9
T
I
·
I
I
I9
·
I
r
9
I
93
I
995 11)
1
I
t\
T
I
I'
999
1
r
I
n;
Figure I . Total Acres Covered by Crop Insurance
8000
7000
'= 6000
_.
5000
4000
3000-
, 2000
1000
O
FT
I II
II
IF II
IrII 111 1
I
rITI
r 1o
11 1111 I
Figure 2. Number of Farms, 1910 - 2004
99
rII
11III IIII
1
-IIT
T
l0I Il
U.UY -
0.08 0.07
0.06
0.05
0.04 0.03 0.02
0.01
0O
I
1990
I
1991
1992
1993
1994
I
I
1995
1996
Figure 3. Farm Failure Hazard Rate
100
I
1997
1998
1999 2000
Figure 4. 1993 Geographic
A. Standardized
Distribution
Precipitation
Monthly-Minimum
-
c:::::J.
of Climate Indices
Index
Monthly-Maximum
............
&euer.ly
Or"
B. Moisture
Anomaly
(Z) Index
Monthly-Minimum
( ...
) .lnZlM>-a
c=J
_
UWa.1.-lly
Dr'au;tlt
Ib'1NI
(=:::J
_
a.--.
-.--
Monthly-Maximum
DrOUQht
.... Y.&n...
II!!!!!!!I ..........
(_J -.lIJlD.a
DrouQtIt
~
_____
Oruuvhl
"-
_"'ICllll)'
_&"-I",
C. Palmer Drought Severity Index
Monthly-Minimum
Monthly-Maximum
( __
I0 I
J __
PDRI~
c::::::J
a..--.
o...ought
_~."'I","'t
_1
as-
_NodI.r ....... u-
102
Figure 5. 1994 Geographic
A. Standardized
Distribution
Precipitation
Monthly-Minimum
of Climate Indices
Index
Monthly-Maximum
_~"""I
~~IYDry
B. Moisture Anomaly (Z) Index
(--)
.1ri"DSlec
Monthly-Minimum
Monthly-Maximum
Monthly-Minimum
C. Palmer Drought Severity Index
Monthly-Maximum
_t.~I~==1
( ...
_fJler-I~""
103
) _.t'D8IK
c::::J IUtr_
_"'I~II~
_ber-Iv
c:==J
Drougt1t
,
.
~.
_~.wl"
...,
Draught
104
Figure 6. 1995 Geographic
A. Standardized
Distribution
of Climate Indices
Precipitation
Monthly-Minimum
=--- ...,
B. Moisture
Index
Monthly-Maximum
Anomaly (2) Index
Monthly-Minimum
Monthly-Maximum
(_)
Monthly-Minimum
(_).1
....
1-=
_.llrm.c
_s--- ....~
_
.... ic.II"
......
C. Palmer Drought Severity Index
Monthly-Maximum
_~=I~~I
(_l_PD8I-=
_~"';:Ie---~I
_bw-I"
_bar-lif""
105
...,
106
Chapter 3
How Cost-Effective Are Land Retirement Auctions? Estimating the
Difference between Payments and Willingness to Accept in the
Conservation Reserve Program
Co-Authored with Ruben N. Lubowski and Michael J. Roberts
1
Introduction
The Conservation Reserve Program (CRP), established by the Food Security Act of 1985, offers
annual rental payments to farm operators who voluntarily retire environmentally sensitive
cropland under ten- to fifteen-year contracts.
and its environmental benefits.
The CRP is notable for the size of both its budget
The CRP currently pays about $1.7 billion per year to retire
about 34 million acres (equal to about 8% of U.S. cropland, or the size of Iowa).
For
comparison, Congressional appropriations for toxic waste cleanups under the Superfund
program, one of the nation's highest profile environmental initiatives, totaled $1.3 billion in
2003.
To date, the CRP has disbursed over $26 billion in direct payments.
Researchers have
estimated that the environmental benefits exceed the program's costs (Feather, Hellerstein,
Hanson 1999; Ribaudo et al. 1990).
This paper measures the cost-effectiveness of the CRP's bidding mechanism. Currently,
landowners submit bids that are ranked according to a score comprised of both an environmental
benefits index (EBI), which includes erodibility and other environmental factors, as well as the
landowner's proposed rental rate, subject to soil-specific maximums.30 Bids with the highest
In the literature from the Farm Service Agency (FSA), the total contract "score" is called the EBI and the
environmental components are termed the EEBI (environmental EBI). Here, we use EBI to refer to just the
30
107
scores are accepted into the program after each sign-up period. Landowners with especially high
EBI scores and especially low opportunity costs for their land can potentially submit rent bids
above their reservation rents and still have their lands accepted into the program, resulting in
windfall gains to these landowners. Our objective is to estimate the size of these premiums
based on observed bidding behavior using data on all individual CRP contract bids from regular
sign-ups since 1997.
We estimate the relationship between farmers' proposed rents net of costs and their bid
score, which is endogenous; farmers can increase the environmental component of their score via
their proposed conservation practice and the cost component of their score via their bid rent. To
deal with this issue, we use exogenous components of the EBI by farmers' proposed rent or practices -
those that cannot be influenced
as instruments for the total score.
In earlier research on CRP bidding behavior, Vukina, Levy, and Marra (2003) examined
CRP bid data to determine farmers' valuation of environmental benefits; Shoemaker (1989)
examined rental premiums obtained by farmers in the first two years of the CRP, before the
current EBI-based mechanism was established; and Cason and Gangadharan (2004) used
experimental auctions akin to CRP auctions to examine bidding behavior. These studies all
suggest that landowners adjust their bid rents upwards with increasing likelihood of acceptance
into the program.
2
Institutional Background
environmental components and "score" to refer to the overall total. In this way, we seek a more intuitive usage of
the term "environmental benefits index."
108
2.1
The CRP Bidding Mechanism
Initially focused on reducing soil erosion, CRP's environmental objectives were broadened with
the introduction of the EBI in 1990. In an effort to reduce costs, soil-specific maximum rental
rates also replaced earlier rate caps set at state and sub-state levels. Starting with sign-up 15 in
1997, the EBI was revised to include wildlife habitat and air quality benefits in addition to water
quality and soil erosion. Since then, CRP offers have been scored as follows.
Landowners offer a parcel of land with a set of proposed conservation practices for
enrollment in CRP. The offer receives an environmental score equal to the sum of six factors,
NI through N6 (collapsed to five factors beginning with sign-up 26 in 2003). The six factors
contain points delineated as follows: NI for grass covers, tree planting, location, and wetland
restoration practices that create wildlife benefits; N2 for water quality benefits from reduced
erosion, runoff, and leaching; N3 for on-farm benefits from reduced erosion; N4 for benefits that
will likely endure beyond the contract period; N5 for air quality benefits from reduced wind
erosion; and N6 for state or national conservation priority areas. 31 Landowners can influence the
number of environmental points they receive mainly via factors N1 and N4, depending on the
type of grass or tree covers and wetland restoration practices (if applicable) they choose to offer
for their enrolling land.
Points from the environmental factors are added to a cost factor to obtain the total score.
The cost factor penalizes higher proposed rents and offers that propose sharing the conservation
practice costs with the government. Specifically,
Costfactor = w(1-rl(HIGH)) + 10(l-s) + Min(15, rm-r),
31
(1)
Starting with sign-up 26 in 2003, the points for conservation priority areas were eliminated as a separate factor
and incorporated in the wildlife and water quality factors. The cost factor (described below) became N6.
109
where r is the proposed rent, r is the parcel's soil-based maximum rent (r must be less than or
equal to rm), HIGH is the highest allowed soil-specific rent for all contracts received, and s
denotes the farmer's decision to share costs (s=1 if share, s=0 if not). If farmers elect to share
costs, the government typically pays half the cost of establishing the proposed conservation
practice; if they forgo cost sharing, they receive an extra 10 points on their bid score. The last
term confers additional points for offering rent r below the maximum allowable rent r.
This
component was introduced starting with signup 16 and adds one point for each dollar offered
below the maximum rent up to a limit of 15 points.
The value of w is set at the government's
discretion. It was 175 for sign-up 15 and 125 for subsequent sign-ups.
In this paper, we use EBI to refer to the sum of the environmental EBI factors, N1
through N6 (N1 through N5 in sign-up 26), and SCORE as the sum of EBI plus the cost factor
N7 (renamed N6 in sign-up 26). The endogenous component of EBI refers to the sum of factors
N1 and N4, and the exogenous component of EBI refers to the sum of the other environmental
factors.3 2
3
Summary Statistics
Our data are from the Farm Service Agency's (FSA) CRP bid files.
We focus on five CRP
general sign-ups (15, 16, 18, 20, and 26) conducted between 1997 and 2003.3 3 Sign-up 15 is the
largest in terms of both the number of bids submitted (251,959) and the number accepted
(160,624). Sign-up 20 had the least number of bids submitted (56,093) and sign-up 26 had the
32 Subfactor
N5d for carbon sequestration, introduced in sign-up 26, depends on the choice of cover practice and is
treated as an endogenous part of the EBI in our analysis of this sign-up. Other N5 components are considered to be
exogenous.
33 Data
from the latest general sign-up, 29 held in 2004, were not available at the time of this study. "General" sign-
ups account for the bulk of CRP enrollment and are subject to the competitive bidding process discussed above.
FSA conducts smaller "continuous" sign-ups on a rolling basis that offer fixed rental payments for enrolling specific
high priority environmental practices.
110
least number accepted (38,621). In table 1 we report some summary statistics from these data,
cross-tabulated by quartiles of the exogenous component of the EBI and quartiles of the soilbased maximum payment (rm). For each subgroup, the table reports the average offered rent, the
average discount below the maximum rent, the proportion offering a discount below the
maximum, the average endogenous EBI, and the proportion of offers accepted in the program.
Several patterns emerge from this presentation of the data. Most notably, farmers appear
to offer higher rental rates if they have a high exogenous EBI score conditional on their
maximum payment quartile. This pattern is somewhat visible in the mean rent offered, which
generally increases across exogenous EBI quartiles within a maximum payment quartile. The
behavior is most clearly visible in the mean discount offered.
Since a farmer can increase his
total SCORE by offering to accept a rental rate below the soil-specific maximum, we uniformly
observe farmers with lower exogenous EBI scores offering a larger discount even after
conditioning on the maximum payment quartile.
For instance, in sign-up 16 among farmers in
the lowest quartile of maximum payments, farmers in the lowest exogenous EBI quartile offer a
discount more than twice as large as those in the highest exogenous EBI quartile (1.2 versus 0.5).
This pattern holds across all maximum payment quartiles in sign-up 16 and across virtually all
sign-ups. Discounts were lowest in sign-up 15 and greatest in sign-up 16. This may reflect
initial uncertainty about the new program structure in sign-up 15 and the influence of additional
points, introduced in sign-up 16, for bids below the maximum allowable rent.
A large share of bids have zero discounts, especially in the lower payment categories.
The proportion offering non-zero discounts tends to increase as the maximum rent increases and
tends to decrease for higher exogenous EBI scores. For example, in sign-up 26, the proportion
111
of non-zero discounts declines within the first maximum rent quartile from 30 to 19% and
declines from 85 to 62% in the highest maximum rent quartile.
Another way for farmers to increase their total EBI is by adopting more valuable
conservation practices, such as plantings of native grasses. The set of columns in table 1 labeled
Conservation Practice Points demonstrate the tendency to adopt a more valuable conservation
practice at lower exogenous EBI scores. Conservation practice points tend to increase with the
maximum payment quartiles, but the pattern is not universal. In general, the summary statistics
would seem to indicate the landowners increase the endogenous components of their bid as their
exogenously-determined EBI decreases. This pattern is consistent with the hypothesis that
landowners with higher exogenous EBI scores are able to command greater windfall gains from
the program.
4
A Model of Bidding Behavior
Although reservation rents are not directly observed, they are implicit in farmers' proposed rents
and conservation practices. Under the program's incentive structure, the exogenous components
of the EBI encourage farmers to propose higher rents and less costly covers and wildlife
practices, all else remaining the same.
In this section, we spell out these incentives using a
simple model.
Define F as the distribution function of farmers' subjective beliefs about E*, the critical
bid score.
All scores above E* are accepted into the program and all scores below E* are
rejected, but farmers do not know E* when they submit their bids. The Secretary of Agriculture
112
makes this decision after all bids have been submitted.34 Farmers choose their proposed rent, r,
conservation practice with annualized establishment cost, c, and whether or not they propose to
share costs with the government, s.3 5 Higher proposed rents negatively influence the SCORE.
The EBI is a function of exogenous factors and of the conservation practices chosen with
associated costs c.
Assume farmers choose r, c, and s in order to maximize the expected net benefit from
their proposed contract given their reservation rent
and the maximum rental payment rm.
Once a bid is accepted, the farmer's yearly net benefit is the rent minus their reservation rent (the
opportunity cost of their land) and the annualized value of establishing the proposed
conservation practice, (r - r - ac), for the duration of the contract. We call this net benefit the
farmer's premium, PREM.
The parameter a equals 1 if s=O and a equals /2 if s=1. Thus, the
farmer's objective is:
max V(r, c, s) = PREM x F + r
(2)
subject to: (rm - r) 2 0,
where F =Prob[E* < SCORE], and SCORE = g(c) + z - Or+ 1Os. The function g(c) gives the
endogenous component of the EBI as a function of cost, c; z is the exogenous EBI component,
and the parameter 0 scales the marginal reduction in SCORE from a $1 increase in the proposed
rental rate.
If r< rm (the maximum rent constraint does not bind), the first order conditions are:
34
The weight w in N7 is also uncertain in the eyes of landowners.
We do not explicitly examine this source of
uncertainty, though it should affect marginal incentives to bid in a similar way as uncertainty about E*. Although
this weight can be changed at the government's discretion, this value has been the same in sign-ups 16, 18, 20 and
26 so we'expect variation in beliefs about this weight to be small. While the weight is not explicitly reported, it
could be approximated from publicly available statistics on each sign-up.
35 Practices may also involve annual maintenance costs, which are not subject to the cost-sharing option. Instead,
the government typically pays a fixed maintenance allowance of $5 per acre. As long as the difference between the
actual maintenance costs and this allowance is small, this will not significantly influence the farmer's decisions.
113
- 0(r - r - ac)F'+F= O
(3)
g (r -
(4)
ac
? - ac)F'-aF =
These conditions imply that the marginal expected benefit from an increase in the probability of
acceptance equals the expected marginal cost from the decreased premium. If we define G =
FIF', an increasing function of SCORE, these conditions can be written as:
- (r-r--ac)+G=O
-(r
8c
Thus, PREM = G/ = aG/ ag
/
c'
- - ac) - aG = 0
(5)
(6)
where G is the distribution function divided by the density
function of farmers' beliefs regarding the critical EBI score. This implies farmers will balance
their offered bids and costs until 0-=
/a,
such that the marginal costs of raising the SCORE
by lowering the bid equal the marginal costs of doing so by increasing costs. The discrete cost
share decision (s) cannot be analyzed from these marginal conditions, but does influence the
marginal tradeoffs via a.
Both conditions imply PREM is an increasing function of SCORE, unless the constraint
(r' - r) > 0 binds, in which case PREM < G/0 =aG/ g . This is consistent with the intuition
/ ac
given in the introduction. Farmers with higher SCORES, all else the same, will have higher
rental premiums, unless these are constrained by the maximum allowable rent.
114
4.1
Empirical Model
Because we do not observe PREM, we rewrite the first order condition (5) in terms of rent net of
costs:
r - ac -G +r.
(7)
Reservation rents are not observable and are embodied by the intercept and residual in our
regressions.
The intercept reflects the average reservation rent and the residual embodies the
difference between this average reservation rent and that of the current observation.
Since
affects farmers' choice of SCORE, SCORE is endogenous. Dealing with the endogeneity of
SCORE is our main econometric challenge.
By using the exogenous component of the EBI as an instrument for SCORE, we solve the
problem that SCORE is a choice variable. However, a different form of endogeneity arises if r
is correlated with the exogenous EBI, not because one variable affects the other, but because
heterogeneous land attributes affect both land values and the exogenous components of the EBI.
In order reduce unobserved
heterogeneity
of land, we include the soil-specific rental rate,
measures of on-farm wind and soil erosion, and fixed effects for state.
Another issue concerns the functional form of G. In general, G has a positive slope. For
most parametric distributions, G will be a non-linear function of SCORE.
Of course, we have no
information about the shape of landowners' prior beliefs about E*. Rather than explicitly
modeling this unknown distribution of beliefs, we estimate (7) using a simple linear model:
(ri - aci ) =
+ AISCOREi + controlsi + ui,
115
(8)
In this formulation,
-
= l0 + ,l SCOREi, where SCORE, is the predicted SCORE from the first
stage. The estimated parameter
shape or whether landowners
]
G
gives the average slope of - regardless of the function's
have heterogeneous
beliefs about E*.36
Simplicity is also
necessary because we must estimate a censored regression model (with instrumental variables) to
account for the constraint, (rm - r) > 0. The reservation rent ()
is embodied by a combination
of the controls and the residual, ui. Costs (aci) are observed only for offers that propose costsharing with the government (where s=1 and a=1/2). We annualize these establishment costs by
multiplying them by 0.10. For offers not sharing costs, we approximate costs from offers in the
same county proposing the same practices that do share costs.
In order to estimate the effect of SCORE on the rent premium, we must account for the
constraint that bids cannot exceed the soil-specific maximum payment. But for this constraint,
farmers with very high EBI scores may offer a very high rental rate. In this sense, the farmer's
true offered rent is censored by the soil-specific maximum rent.
Farmers can circumvent the
constraint to some degree by offering less expensive conservation practices, thereby netting a
higher rental payment. A strictly constrained bid is one that offers to accept the maximum rental
rate and to provide the least expensive conservation practice. Our inability to observe the least
expensive conservation practice necessitates an estimation strategy that provides upper and lower
bounds to the effect of SCORE. We estimate these bounds using (1) a linear instrumental
variables (IV) model and (2) an IV-tobit model that treats all bids that offer to accept the
36
Besides the functional form of G and the maximum rental rate, the slope of the PREM and SCORE relationship is
influenced by the EBI cost function, whether or not a farmers chooses to share costs with the Federal government,
whether or not a farmer's offered rent is more than $15 below the maximum rental rate, and unobserved
heterogeneity in beliefs about E*. For these reasons it is important to interpret estimates from our linear model as
estimates for the average slope, which may not accurately reflect the slope associated with any particular offer.
116
maximum payment as constrained. The linear IV underestimates the effect of SCORE by
assuming that all bids are unconstrained and optimal, and the IV-tobit overstates the effect by
labeling too many bids as constrained and suboptimal, even though they may not be constrained
by the cost of the conservation practice.
We estimate reservation rents as predicted values from the estimated models with the
values for SCORE re-set to the lowest observed value. This prediction is based on the theory,
which tells us reservation rents equal the observed values minus G(SCORE). Since we only
approximate G(SCORE) with a linear relationship, it is inappropriate to extrapolate beyond the
region of our data by setting SCORE to zero. By setting SCORE to the minimum observed value
in each sign-up, we approximate the minimum observed value of G(SCORE), which is likely
close to zero. Essentially, this assumes those with the lowest bids believe they have a small
chance of being accepted into the program and have negligible premiums.
5
Results
Table 2 summarizes results from our IV and IV-tobit regressions.
For each sign-up, the table
reports the estimated coefficient and standard error on SCORE for each of the two regressions.
This coefficient gives the estimated increase in a landowner's premium for each additional
SCORE point, holding all other factors the same. For example, the first row of column 2 reports
the IV estimated coefficient on SCORE as 0.015.
This estimate implies farmers received a
premium of 1.5 cents per acre for each additional point they received in sign-up 15. If all
farmers were to bid their reservation rents, their bids would effectively be exogenous, and this
coefficient would be zero. In nearly every sign-up, the coefficient on the IV-tobit regression is
larger than the IV counterpart.
In line with the reasoning given above, the IV estimate
117
constitutes a lower bound on the marginal premium and the IV-tobit constitutes an upper bound.
The single exception is sign-up 15 where the difference between the estimates is negligible.
The coefficient on SCORE increases markedly between sign-ups 16, 18, and 20, during
which the configuration of the EBI remained unchanged. This pattern is consistent with a
progressive decline in bidders' uncertainty regarding the critical score needed to gain acceptance
into the program, enabling landowners to extract larger premiums. The estimated coefficients
for sign-up 26 are below those for sign-ups 18 and 20. This may reflect greater uncertainty over
the critical score in sign-up 26 due to the changes to the EBI introduced with this sign-up.
For
each regression, the table also reports the R2, and an F-test from the first-stage regression, which
indicate our instruments are strong predictors of SCORE.3 7
Table 3 reports average rents paid by the CRP and the average estimated premium for all
contracts accepted in the U.S. and for all those accepted in each of the eleven states enrolling
more than one million acres in total over the five sign-ups examined.
The estimates show how
much premiums have increased over time between sign-ups 15 and 20 before decreasing
somewhat in sign-up 26. In sign-up 20, we estimate an average premium per-acre of $15.1 to
$21.2 with an average rent paid of $52.8, implying that about $37 to $52 million of the $129
million in annual rent payments associated with this sign-up constitute windfall gains to
participating landowners.
6
Conclusion
In this paper we have estimated the premiums received by CRP participants above their
reservation rents. Estimated premiums have generally increased over time and constitute 10 to
37 Parameter
estimates for the controls and first-stage regressions are not reported for brevity but are available upon
request.
118
40 percent of the program's rental pay-outs under sign-ups 20 and 26, the two most recent signups for which we have data. However, it is not clear whether the same lands with the same
conservation practices could be retired for less money.
These premiums may be necessary to
induce landowners to reveal their reservation rents and enroll their land in CRP under alternative
systems.
Our estimates may begin to clarify the tradeoffs being made in the current bidding
scheme and inspire consideration of alternative payment mechanisms.
For example, a cost-effective alternative to the current program is one that institutes a
Pigouvian tax or subsidy. Barring large administrative costs, the CRP could reduce economic
costs by either offering rental rates on a schedule that pays a constant price per EBI point, or
taxing plantings of those who do not retire cropland at a rate equal to a constant price per EBI
point lost from non-retirement. 3 8 In both cases, this scheme would reduce societal costs, though
not necessarily government outlays. Economic efficiency would be achieved so long as the price
per EBI point was set equal to marginal environmental benefits in equilibrium. However, these
options may be politically infeasible because they would entail potentially large transfers of
wealth from taxpayers to farmers under the subsidy option and vice-versa under the tax.
The current system attempts to balance economic efficiency with limiting wealth transfer
to farmers.
By establishing
maximum rental payments, the current scheme seeks to limit
transfers to farmers, but it also inhibits farmers with reservation rents above maximum rents
from submitting bids, even if their land would have high EBI scores. Furthermore, under the
current scheme, those submitting bids with high exogenous EBI scores have little incentive to
adopt more valuable conservation practices, even at low private costs, since their bids are likely
to be accepted regardless.
3,
FSA currently offers a fixed price for specific conservation practices under the continuous CRP sign-ups.
119
Might it be possible to implement a bidding scheme that achieves a more efficient
balance between wealth transfer and environmental gain? Our findings suggest that potential
cost savings could be substantial. Future research might fruitfully explore the implications of
alternative bidding schemes and program structures.
120
References
Cason, T. N. "Auction Design for Voluntary Conservation Programs." American Journal of
Agricultural Economics 86 (2004): 1211-1217.
Feather, P., D. Hellerstein, and L. Hansen. 1999. "Economic Valuation of Environmental
Benefits and the Targeting of the Conservation
Programs: The Case of the CRP."
Agricultural Economics Report No. 778. U.S. Department of Agriculture Economic
Research Service. Washington, D.C
Ribaudo, M. O., D. Colacicco, L. L. Langner, S. Piper, G. D. Schaible.
1990.
"Natural
Resources and User Benefit from the Conservation Reserve Program." Agricultural
Economics Report No. 627. U.S. Department of Agriculture Economic Research Service.
Washington, D.C.
Shoemaker, R. 1989. "Agricultural Land Values and Rents under the Conservation Reserve
Program." American Journal ofAgricultural Economics 65 (1989): 131-137.
Vukina, T., A. Levy, and M. Marra. 2003. "Do Farmers Value the Environment?
the Conservation Reserve Program Auctions." Manuscript.
http://www.ag-econ.ncsu.edu/faculty/vukina/vukina.htm
121
Evidence from
-"D00I
1- -9 ,-'t' s OI
-t -t. r- CO1
t ON" 0OD
-' 'C
N
" -_ c0
£0
O
= - t N
O 0 0
C4
a(n
U
U
U
C)
-0
O 00
I'D
-
o
C 'I
oo
000000
00O
- -
Cl
0000
00: 00
0
00 0C
00
C &i
Ci
r) w kr ,I
"t
W) k
r-
NN'; .0
- r 'C
00-rli
ct "- k
C
r.r
10
0
1C
CII
CI
.
9 :a
C 0
C C)
C
I
in
o o6
r4C,, 00 "Cl
6666
r.
6
0
0
6
0
6
rN
r
o66
oo6
en 00 r)0
-INd -
t Cl "
o )o
C
O
l
rq C
C C C)0C>
I'\dC~
°ON
c
" " lC
ro
C>C)C>C
II
z
c-I
'IT)"00
0
,,I rx~ ,I
5
n
rcn
C
00
C
oo
-
00 - I'l
00
00 00 t
oo
-
-
O00
S
.S
e\000000
ON0000l0
&C l r00 000000C
elN'INkj lirN
0N,00 0000
C)..
C.)
C)Q
-l 4
"
M
o O0
,:
f)kf
tO
-_= C)
rn h)
C m O 0
h-I
n
Cl4
C,
O
W)
)
tn
W)
-D
(= (= ( c
cnr
n
ON 00
CN
clqlrq""
00
c
W)
V)
"C "C110 r
,C
in
-
T O en ,
^'t
mm
td- 00 n
^ Id
0 0
C R
o
ON O
't O
"C
'D00'.
to
n- '
eC
00oC
00 m '0D
6 6 )6
OC
moo o
ClC
II
II
r~( m
mm;
m;;
.
Q m
mv v
l
.mm0
X<
-
C
* oN
-
V
11
e
Ux
O
V
O'
'
N
V
00
00X
IM.
I
. 1
V)
o
-
w0o V
V
-IN
lbb
v v
am "
.r.cn
p v, "C
0\c
O
o
0- N
ON
V0 C
0
)mIXO
0
II
II mm m
O°°
k
1
o) k)- k,
0
,t
V V V-
X
Xl
\000V
I
II
V
6 VV
-
C)
C
N r-
sD0 N
00
II
~
n
enr
M en
C~
C~
.0 C'
C.)
- O
CN
C
O~C>ON
-
tn
O
C1
C
I-
rh
W) "
; Io0T o
nC
O.O
Co> (71 c
C)
_
~c M
c
ON \ :n W
- -'IN -<
-
o0\0\
-4 fNON
N
ONONONO,
Cl0Cl000
C
v0
f&~
00 00000N,
,t o r-- O
-
e~ o
-o
_n
[.,
O
("1
("',
,m-,
N
C -N
NoO
,4d
m
O
-
-
6 6 66
rc cr
·1
.,
o
t - O-
-4 -:
Cf
C)l-
0D
C
0
00
I
i dN _
·
_1
x
W
,,IN'INr
acl
-o
I-
I
'I
O
"
) (U
v
-
'C
d-
-m
0,C,
.
o .-
N-
rCCl
C
0000 C
0
O
o=
=o
O 0
cl tO
·C£
l
DVNN.0.
r 00 -00
00t
>
6
00
.o
o O cor
_N'.0'.
C - --
0
r M
rlo
6666 C
- cl
nC
i0 00 00
en N 'I
00
t 00 l t
^r
C>
t o
\C DCc 'IT
It I It
It
oo6 56
0
N
C- -- rl- 11
6oo
6'0
C)
0
ONOtr0t
"
-
00 'I
600 006 C5
6666
0
00
o
oo0
-e)
ONOa-,00
'
--
N"00
'IN
,t NNNr
O 'I0 C '
A
0=
oo
C)
-
o
r
00
6 ONON00
6N 6 6
I
00
alc o
oo 1D
oo
C,
DA
c-'IN
NN
O00Oal
C
C
m
O
O
c *-"o
- 0
0D
n o0 o 00
od
-4
: WN
u
'Ci 1~0'IC
Ci
o
Cj.
· ,
0
,o4- -
o
"T._-i
,,:I-t
I
ac
C
M
tg
aE
e
'IC
ON, O
00
en 00
00
On
00k0
'I _
ON O' 00 N
00 o4 t -Vr) -;
\n
9
ICN
1
-
l
cn
11o 'IC
C
0
0000
W) aE o
-
'.0 I
'.0'.0dd
-e~oOOO
ON
l0
oo~
k
C
r- ~t oe nrn
"Ir
C
'
en
r)
oo
r-
O~ llD
sr t o0
OC
O 0,N -
U I.c
co
a)
t
0000
O~ "
t
V) Ci 0o
It
e
t t
CO
lo
oCr
'C C
CE
r
I
oon00
6 666N
~t N--00 6a
N
rm'.00
Clh(nNO ~
o66
6 6 6 6o
Co6
O 00 ON
6 6 06
6
ro
0000
0000
_Un
O O O O
o0O
o6
"
C,
o,
r-
n
Cl
C.
C
U
O
OOO
-o
0
O
n. )tN00
CNNON
. .
t. . .
-0
". tnNrON
m
-Ci
-o
0 o -o
O
0
dddd~'I'N0I
0 VXbI
V
V Cl ,o
x
*" O 0
Xv w
w~
OO -
v c
oo00~
. pC
)a
\
3
~'
~C)
o ~ ~
__
'
cn O
c
>
'o
t~~~~
st <c,
oo
oo °o m-3crI
o M3
C
(a.
a,a,,
0 v,~o
- . _A3-v
o
c~~~~~k
)V1
C)a
r-
I
o
~~~~=
I -lt:~~~~
nI:T
kf)
>
O
rCD
tw
u
u 0 -\t:'A c
Iq E
c)~~
C3
(OI.
3;
t Ov
(00
=r
U
0
O~~~D~
vlO ~oo~o
00
o
0~
o~~~~~~~~~~~~~t
U
z
0
>l
f
oo
033
cl
(1)
(21
C~~~O
C13
CA
\o
000
o
0
1
k.
C7,
c~~~~~~~
a ~ O~
~~
r,
rc, 0
c
o
o
o;-
r=~~~~~~~~~~~~~~V
00
egg~~~~~~~~~~~~~~~D~~~"a
~~~~~~~~~~~~
S
0
d~
c,d O~~~~~
c~~"
' -
a
u,
dv,
O
" 0~~~~~~
~~~
C
C-1
C-_
.n
'"'~
_
Fw
",
" IZI
I
3
; -i5 ! nM
a
"8ol~~.
o:N~~~~~~~~'z
o\
C)
-
Oa,'-tr)
"
O
o o ~
t
>
4d
m.
-,
m
41.0
C's
4-4
0
E
C4
L
E
~
0
M
_n
W
z
-
F 4 P-
C13C
u
7" b
$,
-
`O
.D, I.
W
u
o~
°o
=
00
-vp
.
m
c,
~
o
to
z oo~~~~~~~~~~~~~~~~
En
m
~~
~~~~~;°
t-~~~
.
.
a, W aa ^"
C))
cn
]- .
"
0
,
)Q
rn
cn ~~~
° o owyOO
'
m
300~~3
o c~,'U ~e
c
CZ=
0
a, u3C.~::;
C's ~
a
I
~
--"-
- . c ~~
vl9~~
I
o 00
a-z~0u
L.
I
rj
¢
kn
0a)t)
m
t4
C)
o
i)
.)
C'
kncn
00
Cl
C
e
er Incrln~,v
Cl
_
0)
V
W) m
4
V)
t
p~
"t Cl Cl'
=
"
~~ Cl ~~l ~ l ~~l
C
l~l~~lC
lrI
0 )0
SuQ
p4<
E
0)
_
-e
c
O
o0
0
OC=
OOD
Oo=
o
I-
iI,
-rCl
.
_
d_
-0)
-
-
0N00
9
0)
lI
t- I
" <mw1
I)
tn n
" r- ,%
.-
I
I
c,
cnM kn
~ c lf A\00~ 0 ~~0~~~V
C
;
c.,r
-
0)
O
I-
QC>
S;
O
_ _ _
_ _ _ _ _
I-.c
pCcd
,CD- _
W
r.o
0 Cl
-n
.
av
n
-
rI
In
n;^
..
C
0
-n 0
o* _
- - -
- -
'ud
LQ
U
° 3
I00(
C5
V]
Hz e
00O kD
m00
N
*0)
oo
o
C0
Fa
-E
,
0
*
FI
00
kn Cn
W)
0
C1
m
00
r-:
r O
I~
I
00 ClC> 00 00 NCl
Cl
¢-
I
I
rI
0
1orod
0
6
IO
00
1-00t-g:
I0
o
6'
II AC
'
D
5PC5
E5 ¢.
c
0
~~~~~~~~~
00 .S 00 to
f~~~~~~~t~
I0 -
Cl
n
,
v
C,3'n- In
I ,
)
C)
aH
4-.
4-.
C l
-
Ct
¼
._
-^
0n
C_
0.0N0
oC~dC~0
a)
Ct
a
I-
.
Ev
ra
r.
C
-0) 0
_F t 1 0
cn
'n _
ud
0o
u
00
^.D
0
000 0r
0
00'ClN
m
0
~~4u
O*
k.
00
C-
w
C*
c~~v
z0)z ccio~c6~S