Printer Friendly

DOES THE SAMARITAN'S DILEMMA MATTER? EVIDENCE FROM U.S. AGRICULTURE.

I. INTRODUCTION

The state periodically steps in as the "insurer of last resort" during systemic shocks, such as natural disasters (e.g., Hurricanes Katrina and Sandy) and economic crises (e.g., the financial crisis of 2007-2008). In doing so it faces the Samaritan's dilemma, first described by Buchanan (1975): victims who expect to be bailed out may take on additional risk in response. For example, potential bailout recipients may purchase less flood insurance or invest in riskier securities. (1) This type of moral hazard increases the economic cost of the shock and decreases overall welfare due to the variability in the marginal utility of income across states of nature (Coate 1995; Kaplow 1991; Kim and Schlesinger 2005). In other words, bailout expectations lead to unequal marginal utilities of income across states of nature, which in turn creates a welfare loss even if real outcomes such as output are unaffected. If real outcomes are affected as well, the welfare loss is larger.

We provide some of the first evidence on the empirical importance of the Samaritan's dilemma with respect to ad hoc aid. Anecdotally, the expectation of free assistance is thought to be an important explanation for the relatively low rates of insurance take-up and self-protection measures in several important settings, including natural disaster insurance, foreign aid, and financial markets. However, empirically estimating the Samaritan's dilemma is incredibly challenging for at least two reasons. First, bailouts and risk exposure are simultaneously determined: the size of a bailout depends on economic agents' risk exposure, and agents' risk exposure depends on their expectation of a bailout. Moreover, changes in background risk that are unobservable to the econometrician may affect both the size of a bailout and the agents' choice of risk exposure (Gollier and Pratt 1996; Harrison, List, and Towe 2007). Few empirical studies have attempted to tackle the simultaneity and confounding variables issues. (2) Second, shocks during which agents can reasonably expect to be bailed out, such as Hurricane Katrina or the financial crisis of 2007-2008, are rare in most settings, and rare shocks do not easily lend themselves to systematic statistical examination.

To credibly gauge the relevance of the Samaritan's dilemma for the provision of social insurance, we would need a setting with fairly frequent shocks, extensive insurance availability, and relatively frequent government bailouts that, to some extent, vary exogenously. U.S. agriculture provides such a setting. Agricultural producers can purchase heavily subsidized crop insurance, but the government appears unable to withhold ex post aid: Congress provided ad hoc disaster payments every year between 1990 and 2010, the period of our analysis, at an average of $1.8 billion per year. (3) Politics has long been thought to play a role in agricultural disaster aid allocation, both in the United States and elsewhere (e.g., Chang and Zilberman 2014; Cole, Healy, and Werker 2012; Garrett, Marsh, and Marshall 2006; Goodwin and Vado 2007), creating plausibly exogenous variation in aid that is not directly related to farmers' insurance decisions.

Motivated by these facts, we rely on political variation to identify the causal relationship between aid expectations and insurance decisions. Our choice of instrument is guided by the theory of tactical redistribution in which politicians make pre-election promises in a bid to gain votes (Dixit and Londregan 1996, 1998). Specifically, we employ the "swing voter" model, which is the most commonly used model in this literature (Dahlberg and Johansson 2002; Lindbeck and Weibull 1987). We use changes in the percent of a county's voters who voted for a third-party candidate in the most recent prior presidential election as an instrument for disaster aid. As we discuss later, third-party voters are easier to sway than someone voting for a Republican or Democrat, making them excellent targets for any politician who is trying to gain voters in county, congressional, state, or even national elections. Likewise, agricultural disaster aid is a cost-effective way to target voters, because the majority of Americans of both parties favor financially supporting farmers, especially in bad years (see, e.g., Kull et al. 2004). At the same time, farmers represent a small share of the population, reducing the likelihood of instrument endogeneity with respect to crop insurance.

We use county fixed effects to account for unobserved cross-sectional heterogeneity, such as the inherent riskiness of an area for crop production. We account for macro-level shocks, such as price variation or policy changes, with year fixed effects. Thus, our identification comes from within-county changes in voting patterns, disaster aid, and insurance coverage. We also control for a number of time-varying county characteristics, including farm and non-farm incomes, total employment, population, the share of population employed in agriculture, and the number of farm proprietors. Our identifying assumption is that, conditional on these controls, recent voting behavior in a county is related to the crop insurance decisions of a county's farmers only through the disaster aid channel. We argue that our instrument is likely to meet the exogeneity requirement, in part because farmers make up a small fraction of the electorate in the modern United States. We also show that our estimates are robust to employing additional instrumental variables based on other political theories.

We find that the elasticity of farmers' out-of-pocket expenditure on insurance with respect to expected disaster payments is about -0.2. That is, a 10% increase in expected disaster payments reduces the premiums farmers pay by 2%. We confirm this result by using alternative measures of coverage, such as total liability, total number of policies, and premium subsidies. Consistent with farmers reducing insurance coverage rather than foregoing it altogether, we find evidence that farmers are choosing less generous insurance plans. Finally, we find that bailout expectations result in reduced spending on farm labor and fertilizer, lower price-weighted yields, and lower revenue from crop sales.

Theoretical literature predicts that subsidizing risk-reduction activities such as insurance reduces agents' reliance on bailouts (Coate 1995); the theory can be extended to show that increasing the uncertainty of a bailout also reduces the Samaritan's dilemma. Crop insurance is heavily subsidized--the government currently pays about two-thirds of the premiums--and bailouts are ad hoc and thus inherently uncertain, especially from the point of view of an individual farmer. Yet we find that the Samaritan's dilemma still exists and is non-trivial in magnitude, suggesting that it is a more pervasive phenomenon than expected.

Our findings have important implications for a number of other settings. Two that are particularly similar are domestic disaster aid more generally and foreign aid. (4) The United States spent about $100 billion on non-agricultural domestic disaster relief in the 2011-2013 fiscal years (Weiss and Weidman 2013) and about $31 billion in foreign economic assistance in the 2012 fiscal year (United States Agency for International Development 2014). In both cases, the aid is discretionary and thus uncertain. At the same time, it is awarded fairly regularly, making it more likely that potential recipients will expect it. The similarities between these settings and ours make the existence of the Samaritan's dilemma in the former very likely.

Our results also empirically validate the idea that the Samaritan's dilemma and, more generally, ex ante moral hazard--where recipients expose themselves to a higher risk of income loss because of the presence of some safety net--are present in social insurance settings such as unemployment insurance, Temporary Assistance for Needy Families, or Supplemental Nutrition Assistance Program (Buchanan 1975). In contrast to private insurance markets where premiums, deductibles, and co-payments can be adjusted to internalize ex ante moral hazard (Chiappori 2000; Dave and Kaestner 2009), the Samaritan's dilemma implies long-run, persistent welfare losses in social insurance programs that cannot easily be tailored to individual behavior.

In addition, our findings are relevant for gauging the effects of agricultural subsidies, which are prevalent in developed nations. In their theoretical work on the Samaritan's dilemma, Bruce and Waldman (1991) and Coate (1995) suggest that replacing ex post disaster aid with an ex ante in-kind transfer in the form of full insurance coverage eliminates the Samaritan's dilemma. Although some U.S. agricultural subsidies are independent of production or prices, many are effectively partial insurance programs where payments depend on market conditions. Indeed, direct (unconditional) payments to U.S. producers have been shrinking over time, while subsidies for crop insurance have grown substantially and now account for a large share of agricultural support. Our findings suggest that even these large subsidies appear to not eliminate the Samaritan's dilemma entirely.

Finally, our results also provide insight into how farmers alter their risk-management behavior in anticipation of future government payments. A substantial body of research examines the relationship between farmers' risk-management behavior and land-specific subsidies that are known to the farmer ex ante, for example, Direct Payments (see Weber and Key 2012, for an overview of this literature). Surprisingly little work, however, provides carefully identified empirical estimates of the risk-management response to ex ante unknown subsidies, such as Counter-Cyclical Payments or Loan Deficiency Payments. (5) Our research, therefore, provides an innovative way to examine the effect of expected government benefits on farmers' risk-management behavior.

To our knowledge, only two working papers have attempted to credibly estimate the importance of the Samaritan's dilemma in the areas of foreign aid and domestic disaster aid. Raschky and Schwindt (2009) estimate the impact of foreign aid on recipient countries' death tolls from natural disasters, a proxy for disaster preparedness. To get around the endogeneity problem, they use voting patterns in the U.N. General Assembly and the aid recipient's oil reserves and natural gas production as instruments for foreign aid. More foreign aid leads to higher death tolls from storms, which provides some evidence for the Samaritan's dilemma. However, they cannot measure disaster preparedness efforts directly and do not find any effect of foreign aid on death tolls from earthquakes or floods. With respect to domestic disaster assistance, Kousky, Michel-Kerjan, and Raschky (2015) use several measures of political variation as instrument for aid. They find that higher disaster aid in the previous year leads to lower flood insurance takeup on the intensive but not the extensive margin.

The rest of the paper is organized as follows. In Section II, we outline the basic intuition for the inefficiency of ex post relief, which has been shown formally in previous theoretical literature. In Section III, we provide background on crop insurance and disaster payments in the United States. In Section IV, we describe our data and empirical strategy. Section V presents the results, and Section VI concludes.

II. THE INEFFICIENCY OF EX POST AID

Numerous theoretical papers have demonstrated the inefficiency behind the Samaritan's dilemma (Bruce and Waldman 1991; Coate 1995; Dijkstra 2007; Kaplow 1991; Kim and Schlesinger 2005). To frame our empirical work, we highlight the salient intuition from this literature. Altruism is a fundamental tenet of the Samaritan's dilemma--it is the altruism of some economic agents (the "Samaritans") that leads to the recipients' inefficient behavior. In our setting, non-farmers are altruistic toward farmers, which is a well-documented phenomenon (Ellison, Lusk, and Briggeman 2010a; Kull et al. 2004; Lusk 2012; Variyam, Jordan, and Epperson 1990). As a consequence, farmers' consumption is a public good for non-farmers, and private charity will be inefficiently low due to the free-rider problem. In the theoretical literature, the government acts to address the free-rider problem with ex post transfers, which in our empirical setting corresponds to providing farmers with ad hoc aid following a negative shock.

Despite solving the free-rider problem, the socially optimal level of the public good is unlikely to be privately optimal for farmers. The Samaritan's dilemma has adverse efficiency effects stemming from the fact that the government acts ex post rather than ex ante and acts in the interest of the Samaritans rather than the farmers. Coate (1995) shows that ex post aid will be less than the net indemnity under full insurance in the loss state. This outcome is ex ante inefficient because farmers will have unequal consumption in the "loss" and "no-loss" states of nature. And by under-insuring in the first period, risk-averse farmers take on too much risk.

Finally, Kaplow (1991) and Bruce and Waldman (1991) show that ex post aid is not cost-effective because it affects agents' self-insurance. This fact is also potentially relevant in our setting, as farmers have multiple means of self-insurance. For instance, they can ameliorate the consequences of an adverse production shock through irrigation, pesticides, or increased labor; and savings and inventory can ease the burden of a price shock. When the amount of ex post aid depends on the size of the loss, farmers have incentive to reduce self-insurance, which exacerbates the cost of a bailout.

III. CROP INSURANCE AND DISASTER PAYMENTS

Federal crop insurance and agricultural disaster payments have provided overlapping risk protection to farmers for over 40 years. The Agricultural Adjustment Act of 1938 established the Federal Crop Insurance Corporation (FCIC) to administer what was essentially an experimental crop insurance program until 1980. (6) In 1973, while crop insurance was in this experimental phase, Congress established a standing Crop Disaster Payment (CDP) program that was akin to free insurance coverage for a select group of crops. When yields fell below two-thirds of normal, low-yield payments were made to farmers who participated in income- and price-support programs. The Government Accountability Office (GAO) suspected that there was a conflict between these programs--the Samaritan's dilemma--when it reported that, where crop insurance was offered, "[disaster] payments actually compete with crop insurance because they require no premiums" (U.S. Government Accountability Office 1980). Aware of the disincentive effects disaster payments have on crop insurance demand, some crop insurance demand models have included proxies for disaster payments (e.g., Barnett and Skees 1995; Niewuwoudt and Bullock 1985). These models typically reveal a negative correlation, but they cannot identify the direct effect of disaster payments on crop insurance demand.

In 1980, Congress ended the standing CDP program and greatly expanded the Federal Crop Insurance (FCI) program. In spite of this expansion, Congress continued the pattern of having two parallel mechanisms for dealing with crop-loss risk by providing $6.9 billion in disaster payments on top of $4.3 billion in crop insurance indemnities in 1980-1988 (U.S. General Accounting Office 1989). At the same time, FCI participation stagnated at 50 million acres, less than 25% of insurable land (Glauber and Collins 2002). In 1989 the GAO reported that "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." Despite GAO warnings, Congress continued to frequently authorize ad hoc disaster aid throughout the 1990s and 2000s, allocating a total of $40.2 billion (2011 dollars) to CDP programs in 1990-2011. (7)

A. Crop Insurance

Starting from the mid-1990s, farmers have had a lot of choice when it comes to crop insurance. Importantly, farmers can choose the generosity of the insurance plan they purchase. The options typically range from a 50% coverage plan, which only pays indemnities after the farmer's yield or revenue has fallen to 50% or less of its expected value, to a 90% coverage plan, which begins paying after only a 10% drop. Farmers can also choose how much money they are paid per unit of shortfall from a pre-specified range. (8)

In addition, if a farmer owns multiple plots growing the same crop in the same county, he can choose to insure them jointly and pay a lower insurance premium. If crop insurance were actuarially fair, economic theory predicts that a farmer would want to combine all his plots under a single insurance policy, as he should care about his aggregate income rather than income from any single plot. However, because crop insurance is heavily subsidized, farmers sometimes find it advantageous to insure plots under different policies in order to maximize the expected return per dollar of premium, even if doing so raises the overall variance of their income. Importantly, because of farmers' ability to insure plots separately, the number of policies can reflect both the extensive and intensive measure of insurance takeup.

Unlike many other insurance markets, providers of crop insurance cannot set their own prices or offer customized insurance plans. However, the federal government reinsures the providers and reimburses them for administrative expenses. The prices and plans are determined by the Risk Management Agency (RMA) of the United States Department of Agriculture (USDA) and are typically made public near the end of the preceding calendar year. The rating methodology used to set prices has been fairly consistent and largely formulaic throughout our sample period. (9) This is another important feature of our setting, as it rules out the possibility that insurance prices might be changing for political reasons or in anticipation of greater disaster aid from the government.

Since Congress ended the CDP program in 1980, it has subsidized crop insurance premiums to encourage farmers to purchase more coverage and thereby reduce the need for ad hoc disaster payments. Figure 1 illustrates the evolution of premium subsidy rates from 1990 to 2011. Despite premium subsidies ranging from 17% (for 75% coverage) to 30% (for 50% and 65% coverage) during the 1980s and early 1990s, voluntary participation in the FCI program remained low. The Federal Crop Insurance Reform (FCIR) Act of 1994 greatly expanded the crop insurance program by requiring farmers who received other government support to adopt fully subsidized catastrophic-level (50%) coverage. The FCIR also increased premium subsidies for higher coverage levels. The insurance requirement was removed in the following year, but the higher subsidy rates remained. The Agricultural Risk Protection Act of 2000 (ARPA) further increased premium subsidies, especially for higher coverage levels. The subsidy rate increased by half for the 65% coverage level, more than doubled for the 75% coverage level, and nearly tripled for the 85% coverage level. It is important to note that the premium subsidy rates do not vary geographically. Thus, they cannot be manipulated by politicians to target specific areas in the same way that disaster payments can.

Predictably, the above-mentioned reforms raised insurance coverage. Figure 2 illustrates the share of total acres insured by FCI by coverage level from 1990 to 2011. Nearly all of the increase in participation in 1995 came from an increase in the mandated (and most heavily subsidized) 50% coverage. In 2001, participation levels increased further when premium subsidies were raised even more under ARPA. With these dramatically increased subsidy rates, participation returned to the 1994-mandated level in 2004 and has hovered around 80% of eligible acreage since then. However, insurance coverage among the insureds is far from full--most acres are insured under plans with at least a 25% deductible (coverage level of 75% or lower). About a third of insured acres are covered by plans with at least a 35% deductible (coverage level of 65% or lower).

B. Disaster Payments

Unlike crop insurance indemnities, which are known for a given loss level and plan choice, disaster payments are not perfectly predictable, especially from an individual farmer's point of view. The disaster designation process adds to the uncertainty. First, a state's governor requests a disaster designation for the affected counties in the state. The Secretary of Agriculture then determines whether a natural disaster has caused a 30% or more production loss of at least one crop in the county. Once the Secretary of Agriculture issues a disaster designation, farmers in the primary and contiguous counties become eligible for emergency loans. Farmers in these counties may also receive disaster payments if Congress passes legislation funding an ad hoc disaster program. Disaster payments are usually calculated in a way that is very similar to a not-very-generous crop insurance plan. (10)

Figure 3 shows the pattern of indemnity payments, made by insurance companies, and crop disaster payments, made by the government, over the same time period. To control for the growth of insurance coverage, we show these quantities as a percent of total liability. On average, disaster and indemnity payments are similar in magnitude. In several years, disaster payments exceed indemnity payments. In recent years, disaster payments have been relatively low, potentially due to increasing coverage. Disaster payments were made in every year, although in some years the amount is very small. Consistent with their ad hoc nature, disaster payments are much more volatile than indemnity payments.

Disaster aid programs are administered in such a way that disaster payments are a de facto supplement to indemnity payments. In an effort to be equitable and not discourage crop insurance purchase, Congress typically mandates that "there should not be discrimination, in making payments, against persons who had acquired federal crop insurance" (2000 Crop Disaster Program 2001). In other words, both insured and uninsured farmers can qualify for disaster payments. Only the U.S. Troop Readiness, Veterans' Care, Katrina Recovery, and Iraq Accountability Appropriations Act of 2007 (2007) and the Food, Conservation, and Energy Act of 2008 (2008) have limited disaster payments to farmers who purchased insurance or who did not have the option to purchase insurance. (11) For insured farmers, disaster payments "top up" indemnity payments. However, insurance payments are not ignored completely; once the sum of indemnity and disaster payments reaches 95% of the farmer's expected revenue, the farmer is not eligible for more disaster payments. (12) Allowing disaster payments to be given in addition to crop insurance creates a strong incentive for farmers to respond on the intensive margin by purchasing less insurance than they otherwise would, rather than foregoing crop insurance entirely.

Although Congress has regularly responded to agricultural disasters with CDP programs, it has not been without reluctance. Over the period of our analysis Congress attempted to move away from CDP programs by strengthening the FCI program and weakening its own ability to pass disaster-assistance legislation by tightening budgetary constraints. In 1990-1994, disaster payments came from emergency supplemental appropriations that were exempted from discretionary spending caps. The 1994 FCIR eliminated the use of emergency legislation for agricultural crop disaster assistance, thereby making future disaster payments subject to discretionary spending caps. Together with mandatory catastrophic coverage, these requirements were meant to send a signal that future disaster payments were unlikely (see Jose and Valluru 1997). (13) Congress, however, rescinded the catastrophic-coverage mandate after just 1 year. In 1998, it also reverted disaster spending to "emergency" status and implemented a multi-year CDP program--something it said it would not do 4 years earlier. (14)

In an attempt to reduce disaster payment uncertainty, the 2008 farm bill established a standing disaster program called the Supplemental Revenue Assistance Program (SURE) (Food, Conservation, and Energy Act of 2008). The program, however, failed to reduce uncertainty; according to USDA officials it was "the most complex program USDA's Farm Service Agency has undertaken" (Shields 2010). Moreover, despite the standing disaster program, Congress passed ad hoc disaster payment legislation in 2009, and in a rare move the president sidestepped Congress and implemented a CDP program in 2010. SURE expired in 2011 and was not renewed in the 2014 farm bill. Thus, the current pattern of a heavily subsidized insurance market combined with relatively frequent ad hoc disaster aid can be expected to continue.

It is important to note that crop insurance and disaster payments are part of a larger safety net that includes price supports, production subsidies, and input-specific subsidies. Subsidy programs are unlikely to confound our analysis, however. They are determined by the federal farm bill, which only changes every 6 years and applies uniformly to all farms in the United States. Because of this uniformity, it is unlikely that politicians use general agricultural subsidies to target counties based on changes in their third-party voting patterns, although they can use such subsidies to target agricultural counties more generally. In contrast, crop insurance-coverage decisions and disaster-payment legislation occur annually and apply differentially depending on a county's disaster designation. Thus, politicians can deliver on their election-year promises more quickly and in a more targeted fashion with disaster payments than with changes to the farm safety net.

C. The Likely Importance of the Samaritan's Dilemma in U.S. Agriculture

Despite circumstances that seem to favor the Samaritan's dilemma in our setting, it is not a foregone conclusion. Four potentially offsetting factors make it difficult to determine, ex ante, the extent to which government bailouts affect insurance decisions. First, free disaster payments may seem preferable to costly insurance, but due to heavy premium subsidies, crop insurance is cheap, which should reduce the amount of crowd out. Second, although disaster payments have been made fairly regularly, they are still more uncertain than insurance payments, especially from the point of view of an individual farmer. Third, because disaster payments can supplement indemnity payments, we expect lower crowd out than if insured producers could not receive disaster payments. Finally, if the conditions that trigger crop insurance and disaster payments are very similar, then the latter might be a good substitute for the former. However, as we show in later sections, disaster payments are heavily influenced by politics, and thus might be a poor replacement for market insurance.

To gain some insight into the likely importance of the Samaritan's dilemma in U.S. agriculture, we simulate farmers' insurance choices, taking into account insurance subsidies, the uncertainty of disaster payments, and the fact that the correlation between disaster payments and losses may be low. Specifically, we model the farmers' choice of coverage level in a plan that insures individual revenue and calculate the corresponding out-of-pocket payments for a range of realistic loss and disaster payment parameters. The details of the simulation are presented in Appendix S1, Supporting Information. As expected, increasing the uncertainty of disaster payments or decreasing their correlation with losses reduces the amount of crowd out. Overall, we find a substantial amount of crowd out across a variety of scenarios, suggesting that the Samaritan's dilemma is likely to be important in this setting.

IV. EMPIRICAL STRATEGY

A. Data

We identify the effect of expected disaster payments on crop insurance coverage with county-level administrative data. Despite the absence of individual-level data connecting disaster payments to crop insurance decisions, the designs of these two programs allow us to estimate the magnitude of the Samaritan's dilemma at the county level. Notably, all farms in a county face similar incentives because both the disaster designation process and the crop-insurance base premium calculation occur at the county level.

Crop insurance takeup information is publicly available from the RM A. (15) For each year between 1990 and 2011, the dataset reports the number of insurance policies purchased, the amount of premiums and premium subsidies paid, total liability, the number of acres insured, and the total indemnity payments. (16) We measure county-level crop-related disaster payments with USDA Farm Services Agency (FSA) data, obtained through a Freedom of Information Act request. After 1994, uninsurable crops received disaster payments through the "Non-insurable Crop Disaster Assistance Program" (NAP). We eliminate NAP payments from our data and focus on the disaster payments that may directly affect farmers' insurance decisions. The county-level characteristics we control for in our estimation come from the Regional Economic Information Systems (REIS) and the County Business Patterns (CBP) databases. Expenditure on farm labor and fertilizer, as well as crop revenue are also reported by REIS. Finally, data on production and average farm price received are from the National Agricultural Statistical Service. Total production and acres harvested are observed at the crop-county-year level, while prices received are summarized at the crop-year level.

Table 1 shows key summary statistics for our main regression sample. As we discuss below, it is important for our identification strategy that only a small fraction of a county's population is composed of farmers and their employees. Indeed, we see that less than 4% of the average county's population consists of farm proprietors. About 1% of total employment is in the forestry/agriculture sectors, and farm income represents only about 3.5% of total personal income on average. (17) Of course, while agriculture is a small share of economic activity on average, in some counties it makes up a much larger share. We later show that excluding these counties from our sample does not affect our results.

The next few variables in Table 1 summarize the insurance coverage in our sample. On average, there are about 420 crop insurance policies issued per county in each year, covering about 66,000 acres. Farmers in the average county spend about $677,000 on insurance, with the government contributing an additional $889,000 in premium subsidies. Because of the heavy premium subsidies, we distinguish between premiums that are paid by the farmers themselves and total spending on insurance coverage. Specifically, we refer to the former as "out-of-pocket" insurance expenditure and premiums that include subsidies and out-of-pocket payments as "gross premiums."

The mean total liability in a county is about $18 million. Over our sample period, insurers paid $1.24 million in indemnity payments in the average county each year, while the government disbursed an additional $464,000 in disaster payments to producers of insurable crops. Thus, disaster payments are over a third of the size of indemnity payments, while premium subsidies are two-thirds as large as indemnity payments. Taken together, premium subsidies and disaster payments exceed indemnity payments, reinforcing the idea that farmers enjoy substantial government support in this area.

Finally, as we discuss in detail below, we use the percent of voters casting ballots for a third-party candidate in the most recent previous presidential election as our instrument. The 2004 and 2008 county-level data come from Dave Leip's Atlas of U.S. Presidential Elections (Leip 2014), while earlier data were generously shared by James Snyder. For non-election years, we use votes from the most recent past presidential election. In the average county, about 33,000 votes were cast, with about 7% of those votes going to a third-party candidate. The standard deviation of 8.2 suggests that there is substantial variation in third-party voting in our sample.

Figure 4 shows the spatial distribution of third-party voting for the counties in our preferred regression sample. To illustrate the variation used in subsequent analysis, we subtract the county-level mean and account for year fixed effects. We then take the absolute values of these deviations from geographic and temporal trends and average them by county. The resulting map thus demonstrates the locations of the largest sources of variation in third party voters (darker areas). Although we see some geographic concentration in parts of the South and Midwest, there is substantial idiosyncratic variation outside these areas, suggesting that our results will not be driven by a particular part of the country.

B. Regression Specification

The central empirical question examined in this paper is "Is the Samaritan's dilemma relevant in US agriculture?" We answer this question by testing whether farmers in county c and year t purchase less crop insurance--[Insurance.sub.c,t], as measured by one of the metrics discussed below--when they expect more ad hoc disaster payments conditional on a) county fixed effects ([a.sub.c]) that account for the underlying soil type, climate, and other characteristics that determine the inherent riskiness of producing in each area and b) year fixed effects ([a.sub.t]) that account for macroeconomic shocks such as annual crop price variation and broad changes in the crop insurance program over time (e.g., the premium subsidy rates).

If the Samaritan's dilemma holds, we would expect estimates of [gamma] in the following equation to be negative:

(1) [Insurance.sub.c,t] = [gamma]E [[Disaster.sub.c,t]] + [X'.sub.c,t-1][phi] + [a.sub.c] + [a.sub.t] + [[epsilon].sub.ct].

The key variable in Equation (1) is farmers' expectation of disaster payments in county c and year t, E[[Disaster.sub.c,t]]. The county-level control variables, [X.sub.c,t-1], include population, the number of farm proprietors, total farm income, and per capita income from REIS, as well as the fraction of total employment in forestry and agriculture sectors from CBP. The characteristics are lagged throughout because the insurance decision must be made by March of each year in most cases.

Several metrics of the multifaceted insurance decision are available to us. Given the institutional background, much of the response to disaster aid expectations may be on the intensive margin, with farmers reducing their insurance coverage rather than foregoing it altogether. A variable that captures both the intensive and extensive margins, farmers' out-of-pocket expenditure on insurance, is arguably the most relevant measure of the Samaritan's dilemma in our setting. Total liability and subsidy payments made by the government provide alternative measures of both the intensive and extensive margins.

The total number of policies seemingly provides a clear measure of the extensive margin of the insurance decision. However, because farmers may consolidate multiple plots under one policy or insure them separately, a drop in the number of policies is not straightforward to interpret. Instead, the number of acres insured provides a cleaner measure of the extensive margin.

Three challenges in estimating and interpreting Equation (1) are apparent. First, insurance decisions are based on expected disaster payments, E[[Disaster.sub.c,t]], which are unobservable to us. Instead, we observe actual disaster payments, [Disaster.sub.c,t]. To the extent that the latter is a noisy estimate of farmer expectations, the potential for measurement error and attenuation bias arises. Thus, our estimates of the extent to which the Samaritan's dilemma matters in agriculture should be viewed as lower bounds. Second, realized disaster payments are likely themselves affected by farmers' insurance decisions, as outlined in the theoretical models of the Samaritan's dilemma: areas that buy less insurance coverage may receive more aid. In this case, the simultaneity of the insurance and bailout decisions will cause estimates of y to be meaningless from a causal point of view. Third, unobservable (to us) changes in risk may affect both insurance coverage and disaster payments, again introducing bias to estimates of [gamma] (e.g., Gollier and Pratt 1996). Over the period of our analysis several innovations, for example, climate change and genetically modified seed, have changed the pattern and practice of crop production in ways that may have affected both the crop insurance decision and disaster payments but remain unobserved and unaccounted for in the analysis. We discuss how we try to overcome these challenges in Section IV.C.

Table 2 shows the results of estimating Equation (1) with ordinary least squares where our measure of [Insurance.sub.c,t] is 1n([Premiums.sub.c,t]), the log of farmers' out-of-pocket expenditure on insurance (i.e., not counting subsidies) in county c in year t. We substitute the log of realized disaster payments, 1n([Disaster.sub.c,t] + 1), for E[[Disaster.sub.c,t]]. (18) All specifications include county and year fixed effects, while Columns 4-6 also control for lagged county-level characteristics. Standard errors are clustered by county.

We find a positive and highly significant relationship between contemporaneous disaster payments and insurance expenditure, possibly because both variables are responding to an unobservable shock. For example, low precipitation prior to the growing season may be indicative of adverse growing conditions, prompting farmers to take out more insurance to protect themselves and leading to higher disaster payments. Additionally, the adoption of a high-value crop might prompt more coverage and increase the size of disaster payments.

We also find a positive relationship between lagged disaster payments and insurance expenditure; when we include both lagged and contemporaneous disaster payments, each is significant. Specifically, a 1% increase in disaster payments is associated with a 0.007% to 0.009% increase in insurance expenditure in the current year and a 0.011% to 0.012% increase in the following year. The lagged positive relationship can arise for a number of reasons. First, farmers who receive disaster payments are typically required to purchase crop insurance in the next 1 or 2 years. Second, an adverse event can trigger disaster payments and change farmers' beliefs about risk to their crops, resulting in more insurance in future years. More generally, simultaneity confounds the ordinary least square (OLS) estimate of y in Equation (1). Thus, without a valid instrument for disaster payments, we cannot say much about the Samaritan's dilemma.

To address potential sources of bias in estimating [gamma], we need to isolate variation in disaster payments that is correlated with farmers' disaster aid expectations but uncorrected with the risk environment or insurance decisions more broadly. Consonant with the idea that the government provides ad hoc disaster payments because of voters' altruistic preferences toward farmers, our instrument is derived from county-level voting patterns, which plausibly affect the benefits of crop insurance only through their effect on disaster payments. We discuss this assumption in more detail in the next section.

C. The Political Determinants of Disaster Aid

To identify the effect of aid expectations on the insurance decision, we exploit the political determinants of agricultural disaster aid over a 20-year period, using the tactical redistribution theory as our guide (Dixit and Londregan 1996, 1998). The most commonly used model of tactical redistribution is the "swing voter" model, which posits that elected officials cater to easily persuadable voters with pre-election promises (e.g., Dahlberg and Johansson 2002; Lindbeck and Weibull 1987). (19) This model is typically formulated as two competing political parties promising transfers in exchange for votes. The parties have limited resources and must thus direct transfers to places or voters where they get the most "bang for their buck." Rationally, the parties promise the marginal dollar in a way that maximizes the number of votes they subsequently receive. The easiest-to-persuade voters that are targeted by the marginal dollar of political funds are then referred to as "swing voters."

Although the theory behind the swing voter model is clear, the empirical literature in political science and political economy has surprisingly little to say about the characteristics of actual swing voters in the United States or even how to measure whether someone is easily persuadable. (20) In one of only two systematic studies, Mayer (2007) defines a swing voter as one who equally likes or dislikes the two major parties. (21) Using National Election Studies data from 1972 to 2004, he finds that approximately 9% of voters view the two major parties equally favorably or equally unfavorably and that these voters are almost equally likely to vote for Democrats and Republicans. Expanding this definition to include voters who very slightly favor one party over another, he finds that about 23% of the electorate can be classified as a swing voter in each presidential election during this time period, on average. Surprisingly, there are few systematic demographic differences (e.g., age, race, or gender) between swing and non-swing voters. However, swing voters are less partisan, are more likely to be moderates, and care less about who wins the election.

No county-level surveys tell us how many voters are indifferent or close to indifferent between the major parties. Our measure of easily persuadable voters in a county is the percentage of votes cast for a third-party candidate in the most recent previous presidential election. It is generally agreed that third-party voters are dissatisfied with the major parties and/or the government, feeling alienated from or perceiving little difference between the two major parties (e.g., Allen and Brox 2005; Donovan, Bowler, and Terrio 2000; Rosenstone 1996). Gold (1995) attributes the fact that 19% of voters cast their ballot for Perot in the 1992 election to a "large base of weak partisans." Contrary to popular belief. Herron and Lewis (2007) predict that at least 40% of Nader voters in Florida would have voted for Bush, not Gore, if Nader were not running. (22) Both these findings support the idea that third-party voters are promising targets for both major parties. Overall, the characteristics of third-party voters correspond nicely to those we would expect swing voters to have.

Of course, if third-party voters' expression of dissatisfaction were permanent, they would not be easy to persuade. However, the national share of votes going to a third party in a presidential election varies widely: between 1988 and 2012, it ranged from under 2% in 2008 to almost 19% in 1992. At the county level, the variation is even larger. Because we use county fixed effects, our identification comes from the changes in the share of votes going to a third party, which corresponds to voters who switch between (a) voting for a major party or not voting at all and (b) voting for a third-party candidate. Year fixed effects flexibly account for the overall trend in third-party voting over this time period. As discussed in Section IV.A, the geographic variation in the residual changes in third-party voting patterns is substantial.

Politicians appear to be aware of the perils and opportunities that third parties represent. For example, the Nader candidacy appears to have affected how and where Al Gore campaigned in 2000 (Ceaser and Busch 2001). More generally, Hirano and Snyder (2007) find that much of the 20th-century decline in third-party voting in the United States was due to the Democratic Party adopting left-wing third parties' agendas. More recently, the emergence of the Tea Party movement seems to have caused the Republican party to shift to the right in order to attract disaffected voters (Abramowitz 2011; Jacobson 2011; Williamson, Skocpol, and Coggin 2011).

We do not claim that the swing voter channel is the only one through which politicians direct agricultural funds to their advantage. Politicians may also allocate aid in response to their core constituents' preferences--the "core voter" theory (e.g., Cox and McCubbins 1986; Levitt and Snyder 1995)--or to increase voter support by a combination of increasing turnout of loyal voters and decreasing turnout of non-loyal voters (Chen 2013). Because our main goal is to estimate the effect of disaster aid on the decision to insure, we require only one credible instrument for disaster aid, and fully explaining the political process behind the allocation of funds is outside the scope of our inquiry.

Counties are natural geo-political units at which to direct disaster payments, and politicians need not be trying to win elections at the county level to want to direct disaster payments to specific counties. Rather, politicians competing (or expecting to compete) in elections at the congressional district, state, or even national level may find targeting specific counties appealing because counties are the "units" of disaster declarations and because certain counties contain more easily persuadable voters. However, we do not use the number of third-party voters as our instrument to avoid capturing variation driven by unobservable county growth patterns that may be, for example, reducing the amount of farmland.

Despite being a small share of the voting population, targeting farmers can yield many votes at the margin, because voters who are concerned about farmers are likely to respond to disaster payments to farmers. Previous research has shown that the majority of Americans favor agricultural subsidies for small farms (e.g., Ellison, Lusk, and Briggeman 2010a, 2010b; Kull et al. 2004; Lusk 2012). At the same time, most Americans believe that small farmers get an equal or greater share of agricultural subsidies than large farmers, while in reality the former receive only 20% (Kull et al. 2004). Moreover, the second most common reason for favoring subsidies is the unpredictability of farmers' incomes, due to weather and other factors (Ellison, Lusk, and Briggeman 2010a). (23) Relatedly, the majority of subsidy proponents prefer to give farmers subsidies only in "bad years" (Kull et al. 2004). Finally, although an earlier study finds higher levels of support for farm subsidies among Democrats (Variyam, Jordan, and Epperson 1990), more recent studies find no relationship between a Republican versus Democratic party affiliation and the level of support for farm subsidies (Ellison, Lusk, and Briggeman 2010a; Lusk 2012). Thus, it is rational for Congressmen to allocate agricultural disaster spending strategically, including to areas where farmers are not a large fraction of the voting population.

While it is true that farmers make up a small share of the average county's population, there are a few counties where farm proprietors make up a quarter or more of the population and where farm employment is a very large component of total employment. In order for third-party voting to remain a valid source of exogenous variation in these counties, it must be true that (a) variations in third-party voting are largely driven by non-farmers and non-farm workers OR (b) the factors that drive farmers'/farm workers' voting patterns (and are unobservable to us) do not also affect farmers' crop insurance decisions directly. While the first condition almost certainly does not hold in counties where farmers and their employees make up a large share of the population, the second condition is more plausible, especially once we control for time-varying characteristics such as farm income and the agricultural share of employment. However, because identifying assumptions are fundamentally untestable, we also replicate our analysis using only the sample of counties where agricultural employment never exceeds 5% of total employment, with little changes to our results.

Other commonly considered determinants of disaster payments, such as Representatives' membership on the Agriculture or Appropriations Committee (see, e.g., Garrett, Marsh, and Marshall 2006; Goodwin and Vado 2007), are less likely to be appropriate instruments because they relate to farmers' insurance purchases through more channels that just disaster payments, that is, they violate the instrument exclusion restriction. Assignment to these committees is not random. Unobservable (to the econometrician) changes in the agricultural sector in their jurisdictions that might directly affect insurance purchase decisions also could cause a Congressperson to pursue these committee assignments. These unobservable changes are precisely the reason an instrumental variables strategy is necessary, so using these disaster payment determinants would not solve the problem. As a robustness check, however, we will explore the impact of using Congressional committee membership as instrumental variables.

In the next section, we show that recent third-party voting behavior is correlated with realized disaster aid. However, a reasonable concern is whether such behavior is correlated with farmers' expectations about disaster aid, which we cannot observe. As discussed in the previous section, to the extent that the first stage reflects a noisy measure of farmer expectations, our estimates will be lower bounds. It is unlikely that farmers use observed third-party voting results directly in forming expectations about future disaster payments. However, third-party voting behavior is likely correlated with other manifestations of discontent that predict future disaster payments and are easier for the farmer to observe (but that are impossible for us to observe). For example, what farmers may actually observe and use to form their expectations is discontent with elected leaders expressed through conversations, town hall meetings, bumper stickers, third-party promotional material, and so on. This discontent leads both to more third-party voting and more disaster payments. In this case, third-party voting is a relevant instrument because it is a good proxy for voter discontent. Alternatively, it could be that past voting patterns lead politicians to make promises about future disaster aid in the event of a disaster, in which case third-party voting is again an appropriate instrument. Election-year promises aimed at farmers appear to be fairly common (e.g., Ganzel 2007; Nosowitz 2016; Seidl 2010), although systematic data on this phenomenon are not available.

To summarize, it seems plausible that politicians would target counties where a third-party candidate had recently won a surprisingly large number of votes by promising to allocate more agricultural disaster payments to farmers in that county. The identifying assumption is that county-level voting outcomes are only related to the insurance decision through the disaster aid channel. The summary statistics in Table 1 demonstrate that farmers represent a small fraction of the electorate in most counties, and are thus unlikely to be driving the political trends. Thus, the exclusion restriction is likely to hold because we do not expect county-wide voting changes to directly affect or be affected by an individual farmer's incentives to insure. Nonetheless, in subsequent analysis we control for time-varying county characteristics that could potentially affect both the crop insurance decision and political attitudes.

V. THE EFFECT OF DISASTER AID ON CROP INSURANCE

A. Swing Voters and the Allocation of Disaster Aid

We proceed by estimating the swing voter model, which makes up the first stage of our two-stage approach:

(2) 1n([Disaster.sub.c,t]) = [beta][PctInd.sub.c,t-1] + [X'.sub.c,t-1] [theta] + [a.sub.c] + [a.sub.t] + [v.sub.c,t],

where In ([Disaster.sub.c,t]) is the log of total payments made to county c for a disaster in year t, with 1 added to avoid dropping zeroes. The variable [PctIndc.sub.c,t-1] measures the percentage of the electorate that voted for a third party candidate, based on the most recent previous presidential election. We use all the years for which we have crop insurance and disaster data for the estimation, including ones which did not follow a presidential election. For example, if X% of the county's electorate voted for a third-party candidate in 2004, we set [PctIndc.sub.c,t-1] = X for t = 2005, 2006, 2007, and 2008. Finally, as in Equation (1), [X.sub.c,t-1], represents the time-varying control variables that could potentially affect both the crop insurance decision and political attitudes and ac and [a.sub.t] represent county and year fixed effects. Standard errors are clustered by county.

Table 3 reports OLS regression estimates of Equation (2). When no controls or only county fixed effects are included (Columns 1 and 2), there is no significant relationship between third-party voting and disaster payments, suggesting that there are important fixed determinants of third-party voting and/or disaster payments. For example, a charismatic third-party candidate may attract a lot of votes, but if politicians understand that this is a temporary "shock," they are less likely to respond to such changes in voting patterns with disaster payments. Similarly, widespread crop devastation in some years (e.g., the Great Flood of 1993) could lead to large disaster payments regardless of voting patterns, again leading to a weak correlation. For these reasons, relying on more idiosyncratic variation is more appropriate.

The results when county and year fixed effects are included (Columns 3 and 4) show a strong relationship between disaster payments and political changes in the county. Specifically, a one standard deviation increase in the percent of people who voted for a third party candidate in the last presidential election increases disaster payments in that county by 33-35%, suggesting that disaster payments are indeed being used to sway independent voters. The F-statistic in the specification that includes controls for county characteristics, as well as year and county fixed effects (Column 4), is well above the conventional threshold of 10.

All else equal, counties with higher populations, lower per-capita income, and lower total employment receive more disaster payments. Perhaps surprisingly, changes in the number of farm proprietors and the share of agricultural employment are not significant predictors of disaster payments. However, this pattern is consistent with our earlier hypothesis that agricultural disaster payments are being used to sway the non-farming portion of the constituency. In this case, it would not be unreasonable for changes in the local agricultural sector to make little difference for disaster aid.

B. IV Regression Results

We next estimate the importance of the Samaritan's dilemma for out-of-pocket spending on insurance, In([Premiums.sub.c,t]), by instrumenting for log disaster payments, In([Disaster.sub.c,t]), with the percentage of the electorate that voted for a third party candidate, [PctIndc.sub.c,t-1]. Specifically, the second stage of our two-stage approach is:

(3) [mathematical expression not reproducible]

The variable In([[??].sub.c,t]) the predicted value of the log of disaster payments from Equation (2), the first stage. As above, [X.sub.c,t-j] represents the time-varying control variables that could affect the crop insurance decision. In this specification, [gamma] < 0 indicates the presence of the Samaritan's dilemma.

Table 4 shows the effect of disaster payments on out-of-pocket crop insurance expenditure in a county, as estimated by Equation (3). Column 2 shows our preferred specification, which includes controls for lagged county characteristics. A 1% increase in expected disaster payments causes spending on insurance to drop by 0.20%. This estimate is highly significant. Without controlling for county characteristics, we get a slightly lower but still highly significant estimate of -0.14 (Column 1). Furthermore, the results are even stronger if we include observations where no out-of-pocket premiums are paid by adding 1 to net premiums prior to taking the log (Columns 3-4). In dollar terms, our preferred estimate roughly corresponds to a decrease in out-of-pocket insurance spending of about $25,000 per county for every percent increase in expected disaster aid. (24)

Other measures of the insurance coverage are available to us. In Table 5, we estimate how disaster payment expectations change total liability (Column 1). We again find evidence of the Samaritan's dilemma: a 1% increase in expected disaster payments lowers total liability by 0.19% or about $570, 000 per county. Column 2 in Table 5 shows the causal relationship between disaster payments and the total number of policies (in logs). As with out-of-pocket premiums and liability, the number of policies falls as expectations of disaster payments increase, with an elasticity of about -0.2. However, without additional analyses, we cannot tell whether farmers are completely dropping insurance coverage for some plots or simply consolidating multiple plots into a single policy.

Total out-of-pocket premiums, liability, and (to some extent) the number of policies capture both the extensive margin of the insurance decision (choosing whether or not to have crop insurance) and the intensive margin (choosing how much crop insurance to purchase). We cannot estimate the intensive margin separately by looking at premiums per insured acre, for example, because there could be differential selection out of insurance. We can, however, look at the extensive margin. In Column 3, we look at the number of insured acres as the outcome. Here, we find no evidence that the number of acres insured declines, suggesting that farmers respond to disaster payment expectations by reducing their coverage level rather than foregoing insurance altogether. This response is sensible for several reasons. First, completely dropping coverage is risky because disaster payments may end up not being given. Second, the 50% coverage level plans are almost fully subsidized. Finally, disaster payments top up insurance indemnity payments until the sum of the two reaches 95% of the farmer's expected income, which means that farmers with the least generous insurance plans will not lose out on most disaster payments.

Finally, we consider the extent to which the government crowds out its own premium subsidy payments (Column 4). We find that insurance subsidy payments decrease by 0.34% for every percent increase in expected disaster payments. In dollar terms, this corresponds to about $43, 000 per county. Thus, for every dollar the government crowds out in farmer out-of-pocket spending, it crowds out about $1.72 in subsidy spending.

One potential worry is that premium subsidies are changing with politics and/or disaster aid expectations. However, as discussed earlier, premium subsidy rates do not vary geographically. Unlike agricultural disaster payments, they are not ad hoc, do not change frequently, and, because of their national nature, cannot be used to target specific geographic areas. Thus, politicians cannot use the premium subsidy channel to deliver funds to particular areas, and any relationship between premium subsidies and national politics will be accounted for by year fixed effects.

The facts that (a) the government pays for the majority of the costs of the crop insurance program and (b) government spending on disaster aid crowds out government spending on insurance subsidies may make it seem like the crowding out of insurance by disaster aid is mostly semantics and should not create an efficiency loss. However, the crowding out of subsidies only implies that the total cost of public funds used for disaster aid is smaller than it would be if crop insurance were not subsidized. The real source of inefficiency is that farmers are exposing themselves to more risk in response to disaster aid expectations, as explained in Section II. This inefficiency will arise as long as ad hoc disaster aid reduces insurance coverage, regardless of whether crop insurance is publicly or privately funded.

Next, we investigate whether the expectation of disaster payments causes some farmers to switch to a less generous insurance plan. Recall that a farmer's combined payments from crop insurance and disaster programs cannot exceed 95% of his baseline income. The probability that this happens is increasing in the plan's coverage level. Thus, instead of dropping insurance coverage altogether, which could be devastating if disaster aid is not given, a farmer may choose a plan with a lower coverage level.

To see if farmers are selecting out of more generous insurance plans, we look at changes in the number of policies in different coverage levels.

As in the case of liability and premiums paid, these estimates will include both changes in who purchases insurance and changes in insurance decisions among those who continue to insure. For simplicity, we combine coverage levels into four groups: 50%, 55%-65%, 70%-75%, and 80%-90% coverage levels. To avoid missing values, we add 1 to each variable prior to taking the log. (25) The results are shown in Table 6. We find that for a 1% increase in expected disaster payments, the number of farmers choosing the most generous set of plans falls by about 1.2%, while the number of plans with a 70% or 75% coverage level falls by 0.4%. Correspondingly, there is a rise in the number of the less generous plans, by about 0.2% in both the 50% and 55%-65% coverage levels. The latter results confirm that some farmers are responding to higher expected disaster payments by switching to less generous insurance plans. (26)

Finally, it is important to recognize that farmers may also respond to disaster payments by switching to a higher-risk crop, adopting a riskier farming strategy without switching crops (e.g., lowering pest control expenditure), or exerting less effort in maintaining crop yields if they anticipate that a disaster payment is likely. For example, extant literature suggests that in some cases increased fertilizer applications reduce risk (Sheriff 2005) or are perceived by farmers to reduce risk (Osmond et al. 2015; Stuart, Schewe, and McDermott 2014). Alternatively, crop-switching can have an indirect effect on farm inputs if the newly adopted crop requires a different amount of them.

How changes in farming strategies affect average yields is unclear. It is possible that average yields are unchanged if the riskier strategy simply raises the yield variance. However, the riskier strategy may also result in a higher expected yield and higher yield variance. Finally, because disaster payments are higher when yields are lower, farmers have an incentive to reduce yield-enhancing efforts in times when disaster payments are likely. However, because historic yields also affect payments, both under crop insurance and disaster aid, consistently targeting lower yields is costly. It is thus theoretically unclear whether yields should increase, decrease, or remain unchanged with higher expectations of disaster payments. To construct a single measure of yields, we use data on prices and production of eight major crops (barley, corn, cotton, oats, rice, sorghum, soybeans, and wheat), and calculate the price-weighted yield in a county by summing price times production across these crops, dividing by the total number of acres harvested, and taking the log.

In Table 7, we look at how disaster payments affect expenditure on farm labor and fertilizer, price-weighted yields, and crop revenue. The results point to further welfare losses from farmers altering their farming strategies in response to disaster payment expectations. We find that farm labor costs fall by a small but significant amount (0.06% for a 1 -% increase in disaster payments), and fertilizer expenditure falls by about 0.12%. Price-weighted yields also fall by a small but significant 0.03%, suggesting that farmers are switching to lower-value crops and/or realizing lower crop yields. Relatedly, we find no evidence that farmers are switching to higher-value crops, as receipts from crop sales fall by 0.35% for every 1-% increase in expected disaster payments.

C. Robustness

If our first stage yields a noisy measure of farmer expectations, then our second stage estimates will be attenuated. While we cannot directly evaluate the seriousness of this problem because we do not observe farmers' expectations, we can check the robustness of our results by estimating a model where farmer expectations are based on past disaster payments and instrument for these with a deeper voting lag. Past disaster payments may affect farmer expectations about future disaster payments more than local political activity.

An issue with using lagged disaster payments to study the Samaritan's dilemma is that farmers are often required to purchase crop insurance for 1-2 years after they receive disaster payments, resulting in a mechanical positive relationship between past disaster payments and current insurance holdings. Unlike measurement error, which only attenuates estimates toward zero, such a requirement could in principle cause us to find a positive relationship between past disaster payments and current insurance holdings, even if farmers do reduce insurance coverage in expectations of disaster payments when they are not constrained by such requirements. Nonetheless, the results we obtain from this model (shown in Table S3) are broadly similar to our preferred specification. One exception is that we find a significant decrease in acres insured when using the lagged model. This similarity suggests that post-aid insurance requirements are not important for the dynamics we study here, possibly because the responses we measure appear to be happening on the intensive rather than extensive margin.

In the mid-1990s, the Federal Crop Insurance Reform and Department of Agriculture Reorganization Act of 1994 substantially reformed crop insurance. It is worth asking whether the Samaritan's dilemma phenomenon was affected by this regime change. While we do not have enough data to consider the pre-1996 period, we can easily restrict our sample to years 1996 and later. The results, shown in Table S4, are generally stronger than what we obtain by using the entire sample period. Importantly, we now find a significant extensive margin effect, with insured acres falling in response to disaster payment expectations. These results demonstrate that the Samaritan's dilemma is a highly relevant phenomenon under the current crop insurance regime.

Our estimates could also be affected by dynamics where past third-party voting affects disaster payments (and crop insurance choices) and disaster payments subsequently affect third-party voting. In another robustness check, we control for third-party voting two elections ago, which should reduce the influence of these dynamics, and use third-party voting in the most recent election as the instrument. The results, shown in Table S5, are not substantially affected by the inclusion of this additional control variable.

As discussed earlier, while farming is a small fraction of the average county's economy, there are counties that rely heavily on agriculture. In these areas, overall agricultural outcomes may drive voting patterns, implying that the instrument exclusion restriction no longer holds. To see whether this affects our results, we exclude counties that report more than 5% agricultural employment at any point in the sample. This restriction does not have a meaningful impact on the results, except that the drop in labor costs is no longer significant (Table S6). In another robustness check, we exclude counties that are not in our data for the full sample period (Table S7). Our results are almost unchanged.

We also explored the impact of alternative measures of political influence as instrumental variables. Garrett, Marsh, and Marshall (2006) and Goodwin and Vado (2007) report that states receive more disaster payments when they have a Representative or Senator on the Agriculture or Appropriations committee. A concern is that congress people choose membership on these committees when their state experiences unobservable changes in the agricultural sector that might directly affect insurance-purchase decisions. Nevertheless, we examine the effect of including indicator variables of whether a county's congressperson is on the House Agriculture or Appropriations committee, committee membership interacted with majority party membership, and whether the congressperson is chair of the committee in the set of instrumental variables (Table S8). (27) The results remain essentially the same.

Our results are generally robust to a number of other assumptions. Adding 1 to the outcome variables prior to taking the log generally increases the magnitude of our estimates and strengthens our conclusions (Table S9). In addition, we find a marginally significant drop in the number of acres insured in this case. Similarly, our results are invariant to adding larger or smaller numbers (from 0.0001 to 100) to disaster aid payments prior to taking the log. (28) Normalizing insurance coverage and disaster aid measures by population, by the number of farm proprietors, or by total cropland (as reported in the Census of Agriculture, with linear interpolation between Census years) prior to taking the log also yields similar conclusions.

Finally, we have also probed the robustness of our instrument and second-stage results to including other controls that could affect both the insurance decisions of farmers and the population's decision of whether to vote for a third-party candidate. For example, the occurrence of extreme events could plausibly lead farmers to increase their insurance coverage and cause voters to prefer or stay away from a third-party candidate. Using data from the Spatial Hazard Events and Losses Database for the United States (SHELDUS), we added flexible controls for up to 4 years of past extreme events to our preferred specification. Our point estimates become slightly (but not significantly) larger in absolute terms, but our overall conclusions are again left unchanged.

VI. CONCLUSION

The Samaritan's dilemma was described by James Buchanan 40 years ago. This type of moral hazard may exist in many areas of the economy, from banks taking on excessive risk to homeowners foregoing flood insurance because they expect to be bailed out. Its existence and magnitude both have important implications for efficiency. However, few empirical papers confirm or disprove its existence.

We test for the existence of the Samaritan's dilemma in U.S. agriculture, an area in which it has long been posited to be a problem: since the establishment of modern crop insurance in 1980, Congress has passed ad hoc bills granting disaster aid to farmers who suffered crop losses, even if they had insurance. We instrument for disaster payments using political variation at the county level. We then estimate how expected disaster payments affect farmers' crop insurance decisions.

We find that the Samaritan's dilemma exists and is of a non-trivial magnitude. Out-of-pocket insurance expenditure is moderately sensitive to disaster payments, decreasing by 0.2% for every percent increase in expected disaster payments. This is largely driven by farmers switching to less generous insurance plans and consolidating multiple plots under the same policy. Furthermore, bailout expectations also affect real outcomes, as farmers reduce expenditure on farm labor and fertilizer and subsequently realize lower yields and lower revenues from crop sales.

Comparing our results to the previously estimated distortions from crop insurance, we find that the distortions from ad hoc assistance appear to be greater. In contrast to disaster payments, crop insurance deters moral hazard through deductibles, experience rating and nonlinear pricing as coverage increases. Consequently, the moral hazard associated with crop insurance appears to be small. Goodwin, Vandeveer, and Deal (2004) find a small relationship between crop insurance and corn acreage, and Weber, Key, and O'Donoghue (2016) find no effect of expanded crop insurance coverage on farm revenue or fertilizer expenditure. This contrast underscores the unique nature of the ex ante moral hazard associated with the Samaritan's dilemma. While insurance can be priced to minimize moral hazard, ad hoc disaster relief addresses losses without distinguishing loss due to moral hazard.

Overall, our estimates imply that eliminating disaster payments would significantly raise crop insurance coverage and reduce inefficiencies in farm investment decisions. Of course, eliminating disaster payments would require the government to be able to commit to not grant them ex post, something that it has not been able to do thus far.

If the government is unable to commit to not bail out farmers, why does it not give farmers completely free insurance, as Coate (1995) suggests? One possibility is that it is politically advantageous to target ex post disaster payments to particular constituencies. Congress members may get more "credit" from their constituencies for voting for disaster payments each year than for a one-time passage of a free crop insurance bill. Alternatively, disaster payments might be useful as a bargaining chip for Congressmen from non-disaster counties, who may use them to garner support for their own policies. Finally, it is possible that an insurance system where farmers make no out-of-pocket payments is politically infeasible or would result in even greater moral hazard on other dimensions than those affected by ex ante uncertain ex post aid. Although the definitive answer is outside the scope of this paper, it is a promising area for future research.

An important caveat to our analysis is that we assume farmers form expectations about future disaster payments based on phenomena correlated with recent third-party voting in their county. More generally, because the Samaritan's dilemma is an ex ante form of moral hazard, expectations play a critical role in its presence and extent. Unfortunately, expectations are often difficult or impossible to directly observe. While our results are consistent with our assumption about farmer expectations, it would be fruitful for future research to test it directly.

ABBREVIATIONS

ARPA: Agricultural Risk Protection Act of 2000

CDP: Crop Disaster Payment

CBP: County Business Patterns

FCI: Federal Crop Insurance

FCIC: Federal Crop Insurance Corporation

FCIR: Federal Crop Insurance Reform

FSA: Farm Services Agency

GAO: Government Accountability Office

NAP: Non-insurable Crop Disaster Assistance Program

OLS: Ordinary Least Squares

RMA: Risk Management Agency

REIS: Regional Economic Information Systems

SHELDUS: Spatial Hazard Events and Losses Database for the United States

SURE: Supplemental Revenue Assistance Program

USDA: United States Department of Agriculture

REFERENCES

2000 Crop Disaster Program. 66 Fed. Reg. 15979, 2001.

2005-2007 Crop Disaster Program. 7 C.F.R. $760.805,2008.

Abramowitz, A. "Partisan Polarization and the Rise of the Tea Party Movement." Working Paper, 2011.

Agriculture. Rural Development, Food and Drug Administration, and Related Agencies Appropriations Act, 1999. Pub. L. No. 105-277, 1998.

Allen, N., and B. J. Brox. "The Roots of Third Party Voting: The 2000 Nader Campaign in Historical Perspective." Party Politics, 11(5), 2005, 623-37.

Barnett, B. J., and J. R. Skees. "Region and Crop Specific Models of the Demand for Federal Crop Insurance Insurance." Journal of Insurance Issues, 18(2), 1995, 47-65.

Brown, J. R., and A. Finkelstein. "The Interaction of Public and Private Insurance: Medicaid and the Long-Term Care Insurance Market." American Economic Review, 98(3), 2008, 1083-102.

Bruce, N., and M. Waldman. "Transfers in Kind: Why They can Be Efficient and Nonpaternalistic." American Economic Review, 81(5), 1991. 1345-51.

Buchanan, J. "The Samaritan's Dilemma," in Altruism, Morality, and Economic Theory, edited by E. S. Phelps. New York: Russell Sage Foundation, 1975, 71-85.

Ceaser, J. W., and A. Busch. The Perfect Tie: The True Story of the 2000 Presidential Election. Lanham, MD: Rowman & Littlefield, 2001.

Chang, H.-H., and D. Zilberman. "On the Political Economy of Allocation of Agricultural Disaster Relief Payments: Application to Taiwan." European Review of Agricultural Economics, 41(4), 2014, 657-80.

Chen, J. "Voter Partisanship and the Effect of Distributive Spending on Political Participation." American Journal of Political Science, 57(1), 2013, 200-17.

Chiappori, P.-A. "Econometric Models of Insurance under Asymmetric Information," in Handbook of Insurance, edited by G. Dionne. Dordrecht, The Netherlands: Springer, 2000, 365-93.

Coate, S. "Altruism, the Samaritan's Dilemma, and Government Transfer Policy." American Economic Review, 85(1), 1995,46-57.

Coble, K. H., T. O. Knight, B. K. Goodwin, M. F. Miller, and R. M. Rejesus. "A Comprehensive Review of the RMA APH and Combo Rating Methodology: Final Report." Technical Report, U.S. Department of Agriculture, 2010.

Coble, K. H., M. F. Miller, R. M. Rejesus, R. Boyles, T. O. Knight, B. K. Goodwin, and G. Duffield. "Methodology Analysis for Weighting Historical Experience--Implementation Report." Technical Report. U.S. Department of Agriculture, Washington, DC, 2011.

Cole, S., A. Healy, and E. Werker. "Do Voters Demand Responsive Governments? Evidence from Indian Disaster Relief." Journal of Development Economics, 97(2), 2012, 167-81.

Cox, G. W., and M. D. McCubbins. "Electoral Politics as a Redistributive Game." Journal of Politics, 48(2), 1986, 370-89.

Cutler, D. M., and J. Gruber. "Does Public Insurance Crowd Out Private Insurance?" Quarterly Journal of Economics, 111(2), 1996, 391-430.

Dahlberg, M., and E. Johansson. "On the Vote-Purchasing Behavior of Incumbent Governments." American Political Science Review, 96(1), 2002, 27-40.

Dave, D., and R. Kaestner. "Health Insurance and Ex Ante Moral Hazard: Evidence from Medicare." International Journal of Health Care Finance and Economics, 9(4), 2009, 367-90.

Dijkstra, B . R. "Samaritan versus Rotten Kid: Another Look." Journal of Economic Behavior & Organization, 64(1), 2007, 91-110.

Dixit, A., and J. Londregan. "The Determinants of Success of Special Interests in Redistributive Politics." Journal of Politics, 58(4), 1996, 1132-55.

--. "Ideology, Tactics, and Efficiency in Redistributive Politics." Quarterly Journal of Economics, 113(2), 1998,497-529.

Donovan, T., S. Bowler, and T. Terrio. "Support for Third Parties in California." American Politics Research, 28(1), 2000, 50-71.

Ehrlich, I., and G. S. Becker. "Market Insurance, Self-Insurance, and Self-Protection." Journal of Political Economy, 80(4), 1972, 623-48.

Ellison, B., J. L. Lusk, and B. Briggeman. "Other-Regarding Behavior and Taxpayer Preferences for Farm Policy." The BE Journal of Economic Analysis & Policy, 10(1), 2010a, Article 96.

--. "Taxpayer Beliefs about Farm Income and Preferences for Farm Policy." Applied Economic Perspectives and Policy, 32(2), 2010b, 338-54.

Food, Conservation, and Energy Act of 2008. Pub. L. No. 110-246, $12033, 122 Stat. 1663,2008.

Ganzel, B. "JFK's Farm Programs." Technical Report, Wessels Living History Farm, 2007. http://www .livinghistoryfarm.org/farminginthe50s/money_06 .html.

Garrett, T., T. Marsh, and M. Marshall. "Political Allocation of US Agriculture Disaster Payments in the 1990s." International Review of Law and Economics, 26(2), 2006, 143-61.

Glauber, J. W., and K. J. Collins. "Crop Insurance, Disaster Assistance, and the Role of the Federal Government in Providing Catastrophic Risk Protection." Agricultural Finance Review, 62(2), 2002, 81-101.

Gold, H. J. "Third Party Voting in Presidential Elections: A Study of Perot, Anderson, and Wallace." Political Research Quarterly, 48(4), 1995, 751-73.

Gollier, C., and J. W. Pratt. "Risk Vulnerability and the Tempering Effect of Background Risk." Econometrica: Journal of the Econometric Society, 64(5), 1996, 1109-23.

Goodwin, B. K, and L. A. Vado. "Public Responses to Agricultural Disasters: Rethinking the Role of Government." Canadian Journal of Agricultural Economics/Revue canadienne d'agroeconomie, 55(4), 2007, 399-417.

Goodwin, B. K. B., M. L. M. Vandeveer, and J. J. L. Deal. "An Empirical Analysis of Acreage Effects of Participation in the Federal Crop Insurance Progam." American Journal of Agricultural Economics, 86(4), 2004, 1058-77.

Grossman, G. M., and E. Helpman. Special Interest Politics. Cambridge, MA: MIT Press, 2002.

Gruber, J., and K. Simon. "Crowd-Out 10 Years Later: Have Recent Public Insurance Expansions Crowded Out Private Health Insurance?" Journal of Health Economics, 27(2), 2008,201-17.

Hagen, R. J. "Samaritan Agents? On the Strategic Delegation of Aid Policy." Journal of Development Economics, 79(1), 2006, 249-63.

Harrison, G. W., J. A. List, and C. Towe. "Naturally Occurring Preferences and Exogenous Laboratory Experiments: A Case Study of Risk Aversion." Econometrica, 75(2), 2007,433-58.

Herron, M., and J. Lewis. "Did Ralph Nader Spoil a Gore Presidency? A Ballot-Level Study of Green and Reform Party Voters in the 2000 Presidential Election." Quarterly Journal of Political Science, 2(3), 2007, 205-26."

Hirano, S., and J. M. Snyder. "The Decline of Third-Party Voting in the United States." Journal of Politics, 69(1), 2007, 1-16.

Jacobson, G. C. "The President, the Tea Party, and Voting Behavior in 2010: Insights from the Cooperative Congressional Election Study." 2011. Available at SSRN: http://ssrn.com/abstracts1907251 or https://doi.org/10 .2139/ssrn.1907251.

Jose, H. D., and R. S. K. Valluru. "Insights from the Crop Insurance Reform Act of 1994." Agribusiness, 13(6), 1997,587-98.

Kaplow, L. "Incentives and Government Relief for Risk." Journal of Risk and Uncertainty, 4(2), 1991, 167-75.

Kelley, S. Interpreting Elections. Princeton, NJ: Princeton University Press, 1983.

Kim, B. J., and H. Schlesinger. "Adverse Selection in an Insurance Market with Government-Guaranteed Subsistence Levels." Journal of Risk and Insurance, 72(1), 2005, 61-75.

Kousky, C, E. O. Michel-Kerjan, and P. A. Raschky. "Does Federal Disaster Assistance Crowd Out Private Insurance?" Working Paper, 2015.

Kull, S., C. Ramsay, S. Subias, and E. Lewis. "Americans on Globalization, Trade, and Farm Subsidies." Program on International Policy Attitudes, University of Maryland, 2004.

Leip, D. "Dave Leip's Atlas of US Presidential Elections," 2014. Accessed October 19, 2017. https://uselectionatlas.org/

Levitt, S. D., and J. M. Snyder. "Political Parties and the Distribution of Federal Outlays." American Journal of Political Science, 39(4), 1995, 958-80.

Lindbeck, A., and J. W. Weibull. "Balanced-Budget Redistribution as the Outcome of Political Competition." Public Choice, 52(3), 1987, 273-97.

Lusk, J. L. "The Political Ideology of Food." Food Policy, 37(5), 2012,530-42.

Mayer, W. G. "The Swing Voter in American Presidential Elections." American Politics Research, 35(3), 2007, 358-88.

McDonald, J. D., and D. A. Sumner. "The Influence of Commodity Programs on Acreage Response to Market Price: With an Illustration Concerning Rice Policy in the United States." American Journal of Agricultural Economics, 85(4), 2003, 857-71.

Niewuwoudt, W. L., and J. B. Bullock. "The Demand for Crop Insurance." International Association of Agricultural Economists Triennial Conference, Malaga, Spain, 1985, 655-67.

Nosowitz, D. "Here Are 5 Issues That Caused Farmers to Vote for Trump." Modern Farmer, 2016. http:// modernfarmer.com/2016/11/5-issues-caused-farmersvote-trump/.

Osmond, D. L., D. L. Hoag, A. E. Luloff, D. W. Meals, and K. Neas. "Farmers' Use of Nutrient Management: Lessons from Watershed Case Studies." Journal of Environmental Quality, 44(2), 2015, 382-90.

Pedersen, K. R. "Aid, Investment and Incentives." The Scandinavian Journal of Economics, 98(3), 1996, 423-37.

--. "The Samaritan's Dilemma and the Effectiveness of Development Aid." International Tax and Public Finance, 8(5-6), 2001, 693-703.

Raschky, P.. and M. Schwindt. "Aid, Natural Disasters and the Samaritan's Dilemma." World Bank Policy Research Working Paper No. 4952, 2009.

Raschky, P., and H. Weckhannemann. "Charity Hazard--A Real Hazard to Natural Disaster Insurance?" Environmental Hazards, 7(4), 2007, 321-29.

Rep. Combest (Texas). Concurring in the Senate Amendment to h.r. 4217, Federal Crop Insurance Reform and Department of Agriculture Reorganization Act of 1994, with an Amendment, 1994, October.

Rosenstone, S. J. Third Parties in America: Citizen Response to Major Party Failure. Princeton, NJ: Princeton University Press, 1996.

Seidl, J. "Obama Keeps Promise: USDA Gives $630 million to Farmers." 2010. theblaze.com. http://www.theblaze .com/news/2010/09/15/usda-distributes-election-yearfarm-aid/.

Sheriff, G. "Efficient Waste? Why Farmers Over-Apply Nutrients and the Implications for Policy Design." Review of Agricultural Economics, 27(4), 2005, 542-57.

Shields, D. A. "A Whole-Farm Crop Disaster Program: Supplemental Revenue Assistance Payments (SURE)." Technical Report R40452, Congressional Research Service, Washington, DC, 2010.

Stuart, D., R. Schewe, and M. McDermott. "Reducing Nitrogen Fertilizer Application as a Climate Change Mitigation Strategy: Understanding Farmer Decision-Making and Potential Barriers to Change in the US." Land Use Policy, 36, 2014, 210-18.

Svensson, J. "When Is Foreign Aid Policy Credible? Aid Dependence and Conditionality." Journal of Development Economics, 61(1), 2000, 61 - 84.

--. "Why Conditional Aid Does Not Work and What Can Be Done about It?" Journal of Development Economics, 70(2), 2003, 381-402.

U.S. General Accounting Office. "Disaster Assistance: Crop Insurance Can Provide Assistance More Effectively Than Other Programs." Technical Report, United States General Accounting Office, 1989.

U.S. Government Accountability Office. "Federal Disaster Assistance: What Should the Policy Be?" Number PAD80-39. Washington, DC: Government Printing Office, 1980, June.

U.S. Troop Readiness, Veterans' Care, Katrina Recovery, and Iraq Accountability Appropriations Act of 2007. Pub. L. No. 110-28, $9001, 121 Stat. 213, 2007.

United States Agency for International Development. "Foreign Assistance Fast Facts." 2014. https://eads .usaid.gov/gbk/data/fast_facts.cfm. [Online; accessed September 25, 2014],

Variyam, J. N., J. L. Jordan, and J. E. Epperson. "Preferences of Citizens for Agricultural Policies: Evidence from a National Survey." American Journal of Agricultural Economics, 72(2), 1990, 257-67.

Weber, J. G., and N. Key. "How Much Do Decoupled Payments Affect Production? An Instrumental Variable Approach with Panel Data." American Journal of Agricultural Economics, 94(1), 2012, 52-66.

Weber, J. G., N. Key, and E. O'Donoghue. "Does Federal Crop Insurance Make Environmental Externalities from Agriculture Worse?" Journal of the Association of Environmental and Resource Economists, 3(September), 2016, 707-42.

Weiss, D., and J. Weidman. "Disastrous Spending: Federal Disaster-Relief Expenditures Rise Amid More Extreme Weather." Center for American Progress Report, 2013.

Williamson, V., T. Skocpol, and J. Coggin. "The Tea Party and the Remaking of Republican Conservatism." Perspectives on Politics, 9(01), 2011, 25-43.

SUPPORTING INFORMATION

Additional Supporting Information may be found in the online version of this article:

Appendix S1. Additional Details about Crop Insurance and Disaster Aid; A Simulation of the Samaritan's Dilemma in Crop Insurance; Supplementary Tables with Additional Results

TATYANA DERYUGINA and BARRETT KIRWAN *

* We thank the editor, two anonymous referees, David Albouy, Jeff Brown, Don Fullerton, Nolan Miller, Julian Reif, Michael J. Roberts, and Kent Smetters for helpful discussions and comments. We are grateful to seminar participants at the American Economic Association Meetings, the Institute of Government and Public Affairs, NBER Insurance Working Group, Kansas State University, the Midwestern Economics Association Meetings, the University of British Columbia, and the University of Illinois. Xian Liu provided excellent research assistance.

Deryugina: Assistant Professor of Finance, NBER, University of Illinois at Urbana-Champaign, Champaign, IL 61820. Phone 217-333-9498, Fax 217-244-7969, E-mail deryugin@illinois.edu

Kirwan: Adjunct Research Assistant Professor, University of Illinois at Urbana-Champaign, Urbana, IL 61801. Phone 202-615-4053, Fax 217-244-7088, E-mail bkirwan@iIlinois.edu

doi: 10.1111/ecin.12527

(1.) Terms that describe phenomena similar to the Samaritan's dilemma include "ex ante moral hazard" and, more generally, "crowd out." Ex ante moral hazard typically refers to market insurance crowding out self-protection activities (Ehrlich and Becker 1972). Papers that deal with crowd out more generally typically consider the relationship between a permanent public insurance program and private insurance (e.g., Brown and Finkelstein 2008; Cutler and Gruber 1996; Gruber and Simon 2008). By contrast, the ad hoc nature of bailouts makes them more similar to charity than to public insurance.

(2.) Exceptions are Raschky and Schwindt (2009) and Kousky, Michel-Kerjan, and Raschky (2015). See Raschky and Weckhannemann (2007) for an overview of the literature.

(3.) All dollars are inflation-adjusted to 2011.

(4.) For theoretical considerations of the Samaritan's dilemma in foreign aid settings, see Hagen (2006), Pedersen (1996, 2001), and Svensson (2000, 2003).

(5.) McDonald and Sumner (2003) review the shortcomings of this literature.

(6.) For a more detailed history of the early crop insurance program see Glauber and Collins (2002).

(7.) See Appendix SI for a list of public laws passed between 1989 and 2009 that authorize crop disaster payments.

(8.) For more details on how indemnity payments are deter mined, see Appendix S1.

(9.) See Appendix SI, Coble et al. (2010), and Coble et al. (2011) for more details on how prices are set.

(10.) We provide more details on how crop insurance and disaster aid payments are typically calculated in Appendix S1.

(11.) The latter group has access to a separate disaster assistance program called Noninsured Crop Disaster Assistance Program (NAP), which we do not consider here.

(12.) Typically, a CDP program stipulates "the sum of the value of the crop not lost, if any; the disaster payment received under this part; and any crop insurance payment ... for losses to the same crop, cannot exceed 95% of what the crop's value would have been if there had been no loss" (2005-2007 Crop Disaster Program 2008).

(13.) Speaking just before passage of FCIR, Rep. Larry Combest declared, "This means an end to emergency spending for agricultural disasters." (Rep. Combest (Texas) 1994)

(14.) Agriculture, Rural Development, Food and Drug Administration, and Related Agencies Appropriations Act, 1999(1998)

(15.) Available at http://www.rma.usda.gov/data/sob.html

(16.) We exclude rangeland, which became insurable in the middle of our sample period, from the crop insurance sample. Because of its low value, it is not likely to be receiving a substantial amount of disaster payments. However, it makes up a significant fraction of insured acres (but not of premiums or liabilities), and its inclusion may obscure farmer responses to disaster payments on that margin.

(17.) Farm income is reported net of costs, and it is not unusual for county-level farm income to be negative. It is also possible for farm income to exceed 100% of personal income in some cases because of some methodological differences in calculating personal and farm incomes. For further details, see http://www.bea.gov/regional/pdf/lapi2010.pdf.

(18.) We add 1 to disaster payments prior to taking the natural log due to the presence of many zeros. Our results are robust to adding other positive numbers to all disaster payments and to replacing the zeros with small positive numbers. Adding 1 to the dependent variables prior to taking the log, which we do not do in our preferred specifications, increases the magnitude of most of our estimates, making our conclusions even stronger.

(19.) A full review of the literature on models of tactical redistribution is beyond the scope of this paper. For an overview of special interest politics, see Grossman and Helpman (2002).

(20.) For example, considering only "undecided" voters as swing voters might miss a substantial fraction of the electorate with a very weak and easily changeable preference (Mayer 2007).

(21.) Also see Kelley (1983).

(22.) To do this, they analyze actual ballots, using individual voting patterns in non-presidential races to estimate the counterfactual in the presidential race.

(23.) The most common reason is to maintain a secure food supply for U.S. citizens, which might also lead voters to support disaster aid.

(24.) We arrive at this approximation by calculating [[mu].sup.[mu]+[??]], [e.sup.[mu]] where [mu] is the mean of the log of out-of-pocket premiums and [??] is the estimated impact of additional disaster payments.

(25.) Our results are similar if we do not add 1 to the number of policies prior to taking the log, although the number of observations becomes unbalanced and the first stage F-statistics fall.

(26.) If we look at the number of acres insured at each coverage level, our results are similar with the exception that the number of acres covered at 70-75% coverage levels increases. This further points to farmers consolidating multiple plots under the same insurance plan in response to disaster payment expectations.

(27.) Population-weighted shares are used when a county contains multiple congressional districts.

(28.) These and other results not in Appendix S1 are available upon request.

Caption: FIGURE 1

Crop Insurance Premium Subsidy Rates from 1990 to 2011

Caption: FIGURE 2

Share of Insurable Acres Covered by Crop Insurance from 1990 to 2011

Caption: FIGURE 3

Indemnity and Disaster Payments over Time

Caption: FIGURE 4

Absolute Mean Changes in Percent Voting for Third-Party Candidate
TABLE 1

Summary Statistics

                           (1)     (2)    (3)      (4)      (5)
                          Mean     sd     Min      Max      Obs

Percent of population     3.71    3.96     0       37      60,592
who are farm
proprietors

Percent employed in       1.03    2.35     0       100     60,590
forestry or agriculture

Farm income as percent    3.50    8.22    -314     125     60,592
of total income

Number of policies         420     602     0      7,304    60,592

Acres insured              66      98      0      1,036    60,592
(thousands)

Premiums net of            673    1,233    0     22,455    60,592
subsidies (thousands)

Subsidies (thousands)      883    1,775    0     32,163    60,592

Liability (millions)       18      34      0       890     60,592

Indemnity (thousands)     1,236   3,385    0     152,862   60,592

Disaster payments          460    1,539   -95    109,931   60,592
(thousands)

Notes: Unit of observation is a county-year. All monetary amounts
are in 2011 dollars. Excludes counties with fewer than 18 observations
over the sample period and observations that are missing control
variables. Total number of counties in the sample is 2,916.

Sources: Regional Economic Information Systems, County Business
Patterns, and David Leip's Atlas of U.S. Presidential Elections.

TABLE 2

The Relationship between Insurance Expenditure and
Disaster Payments, OLS

                                  (1)         (2)         (3)

Disaster aid this year (log)   0.009 ***               0.007 ***
                                (0.001)                 (0.001)

Disaster aid last year (log)               0.012 ***   0.012 ***
                                            (0.001)     (0.001)

Farm proprietors (log)

Pet. employed in ag.

Population (log)

Per capita pers. inc. (log)

Total employment (log)

Dep. var. mean                  11.837      11.861      11.861

Observations                    57,879      55,105      55,105

R-squared                        0.916       0.920       0.920

                                  (4)          (5)          (6)

Disaster aid this year (log)   0.009 ***                 0.007 ***
                                (0.001)                   (0.001)

Disaster aid last year (log)                0.012 ***    0.011 ***
                                             (0.001)      (0.001)

Farm proprietors (log)         -0.283 ***   -0.217 **    -0.216 **
                                (0.093)      (0.096)      (0.096)

Pet. employed in ag.             -0.004       -0.004       -0.004
                                (0.004)      (0.004)      (0.004)

Population (log)                 -0.177     -0.260 **    -0.266 **
                                (0.124)      (0.129)      (0.129)

Per capita pers. inc. (log)    -0.676 ***   -0.649 ***   -0.646 ***
                                (0.115)      (0.112)      (0.112)

Total employment (log)           -0.035       -0.014       -0.011
                                (0.056)      (0.057)      (0.057)

Dep. var. mean                   11.837       11.862       11.861

Observations                     57,879       55,210       55,105

R-squared                        0.917        0.920        0.920

Notes: Standard errors (in parentheses) clustered by county.
Outcome variable is log of farmers' out-of-pocket expenditure on
crop insurance in the county. All regressions include county and
year fixed effects. Specifications with controls also include farm
income decile indicators, which are omitted for readability. All
control variables are lagged by 1 year.

* Significant at 10%; ** significant at 5%, *** significant at 1%.

TABLE 3

The Effect of Politics on Disaster Payments

                       (1)       (2)         (3)            (4)

Pet. voting for       0.029    -0.014     0.319 ***      0.291 ***
third party (SD)     (0.019)   (0.018)     (0.049)        (0.049)

Number of farm                                             -0.044
proprietors (log)                                         (0.173)

Pet. employed in                                           -0.012
forestry/                                                 (0.009)
agriculture

Population (log)                                          0.557 **
                                                          (0.249)

Per capita                                               -0.590 **
personal income                                           (0.272)
(log)

Total employment                                         -0.395 ***
(log)                                                     (0.134)

Fixed effects         None     County    County, Year   County, Year

F-statistic           2.347     0.654       41.846         35.262

Dep. var. mean        7.681     7.681       7.681          7.681

Observations         60,475    60,475       60,475         60,475

R-squared             0.000     0.103       0.663          0.664

Notes: Standard errors (in parentheses) clustered by county. The
outcome variable is log of (total disaster payments in the county
+1). Specifications with controls also include farm income decile
indicators, which are omitted for readability. All characteristics
controls are lags.

* Significant at 10%; ** significant at 5%, *** significant at 1%.

TABLE 4

The Effect of Disaster Payments on Out-of-Pocket
Insurance Expenditure, IV

                          (1)         (2)          (3)         (4)

                         Net Premium (log)        Net Premium +1 (log)

Disaster payments      -0.137 **   -0.197 ***   -0.285 **   -0.371 ***
(log)                   (0.058)     (0.066)      (0.119)     (0.136)

Number of farm                     -0.293 ***               -0.372 **
proprietors (log)                   (0.101)                  (0.171)

Pet. employed in                     -0.005                   -0.010
forestry/agriculture                (0.005)                  (0.007)

Population (log)                     -0.023                   -0.146
                                    (0.145)                  (0.251)

Per capita personal                -0.817 ***               -1.054 ***
income (log)                        (0.137)                  (0.224)

Total employment                    -0.124 *                  -0.161
(log)                               (0.069)                  (0.115)

F-statistic             52.195       44.827      43.967       37.050

Dep. var. mean          11.837       11.837      11.328       11.328

Observations            57,845       57,845      60,455       60,455

Notes: Standard errors (in parentheses) clustered by county.
Disaster payments are instrumented for with the share of voters
voting for a third-party candidate in the most recent presidential
election. Outcome variables are specified at the top of each
column. The net premium is the farmers' out-of-pocket expenditure
on crop insurance in the county. All regressions include county and
year fixed effects. Specifications with county characteristics
controls also include farm income decile indicators, which are
omitted for readability. All control variables are lagged by 1
year.

* Significant at 10%; ** significant at 5%, *** significant at 1%.

TABLE 5

The Effect of Disaster Payments on Other Measures of
Insurance Coverage, IV

                      (1)          (2)            (3)           (4)
                   Liability     Policies        Acres        Subsidy
                     (log)        (log)      Insured (log)     (log)

Disaster           -0.185 ***   -0.186 ***       0.019       -0.336 ***
payments (log)      (0.067)      (0.052)        (0.041)       (0.077)

Number of farm       0.114        0.088       -0.284 ***     -0.206 **
proprietors         (0.100)      (0.079)        (0.071)       (0.099)
(log)

Pet. employed        -0.006       -0.004         0.000         -0.007
in forestry/        (0.005)      (0.003)        (0.003)       (0.005)
agriculture

Population (log)   0.704 ***    0.341 ***      0.572 ***      0.287 *
                    (0.151)      (0.119)        (0.104)       (0.150)

Per capita         -0.406 ***   -0.530 ***    -0.292 ***     -0.538 ***
personal income     (0.133)      (0.104)        (0.087)       (0.148)
(log)

Total employment   -0.194 ***   -0.142 ***    -0.214 ***     -0.214 ***
(log)               (0.071)      (0.051)        (0.047)       (0.078)

F-statistic          37.816       37.816        38.327         37.816

Dep. var. mean       15.026       4.698          9.470         11.925

Observations         60,004       60,004        59,396         60,004

Notes: Standard errors (in parentheses) clustered by county.
Disaster payments are instrumented for with the share of voters
voting for a third-party candidate in the most recent presidential
election. Outcome variables are specified at the top of each
column. All regressions include year and county fixed effects as
well as farm income decile indicators, which are omitted for
readability. All control variables are lagged by 1 year.

* Significant at 10%; ** significant at 5%, *** significant at 1%.

TABLE 6 The Effect of Disaster Payments on Coverage
Level Choice, IV

                        (1)         (2)         (3)          (4)
                     50% (log)    55%-65%     70%-75%      80%-90%
                                   (log)       (log)        (log)

Disaster payments    0.203 ***   0.178 ***   -0.374 ***   -1.150 ***
(log)                 (0.064)     (0.057)     (0.100)      (0.199)

Number of farm         0.115     0.630 ***   -0.437 ***   -2.501 ***
proprietors (log)     (0.083)     (0.080)     (0.126)      (0.230)

Pet. employed in       0.004     0.007 **    -0.037 ***   -0.037 ***
forestry/             (0.003)     (0.003)     (0.008)      (0.012)
agriculture

Population (log)     0.552 ***   0.602 ***   -1.094 ***   -1.768 ***
                      (0.134)     (0.122)     (0.203)      (0.350)

Per capita           0.286 **     -0.099     -0.388 **    -2.445 ***
personal income       (0.122)     (0.109)     (0.198)      (0.366)
(log)

Total employment       0.023      -0.047       -0.034       0.043
(log)                 (0.064)     (0.054)     (0.097)      (0.185)

F-statistic           37.050      37.050       37.050       37.050

Dep. var. mean         3.237       3.579       2.912        0.997

Observations          60,455      60,455       60,455       60,455

Notes: Standard errors (in parentheses) clustered by county.
Disaster payments are instrumented for with the share of voters
voting for a third-party candidate in the most recent presidential
election. Outcome variables are the logged number of policies in
the coverage levels indicated at the top of each column. The 1 has
been added to all outcome variables prior to taking the log. All
regressions include year and county fixed effects, as well as
controls for county characteristics. Farm income decile indicators
are included in the regression but are omitted from the table for
readability. All control variables are lagged by 1 year.

* Significant at 10%; ** significant at 5%, *** significant at 1%.

TABLE 7

The Effect of Disaster Payments on Input Spending and
Real Outcomes, IV

                                       (1)            (2)
                                    Labor Cost     Fertilizer
                                       dog)      Spending (log)

Disaster payments (log)             -0.060 **      -0.121 ***
                                     (0.024)        (0.029)

Number of farm proprietors (log)    0.140 ***        -0.041
                                     (0.033)        (0.035)

Pet. employed in forestry/            -0.001       -0.006 **
agriculture                          (0.002)        (0.002)

Population (log)                    0.211 ***      -0.398 ***
                                     (0.052)        (0.061)

Per capita personal income (log)    0.191 ***        -0.063
                                     (0.050)        (0.057)

Total employment (log)                0.018          0.011
                                     (0.025)        (0.028)

F-statistic                           34.966         36.758

Dep. var. mean                        8.434          8.375

Observations                          60,881         60,704

                                          (3)               (4)
                                     Price-Weighted    Cash Receipts
                                    Mean Yield (log)       (log)

Disaster payments (log)                -0.034 **        -0.347 ***
                                        (0.015)           (0.063)

Number of farm proprietors (log)       -0.057 ***        0.248 ***
                                        (0.019)           (0.069)

Pet. employed in forestry/               -0.001           -0.004
agriculture                             (0.002)           (0.004)

Population (log)                       -0.156 ***         -0.135
                                        (0.029)           (0.109)

Per capita personal income (log)         -0.033           -0.108
                                        (0.030)           (0.113)

Total employment (log)                   0.004            -0.079
                                        (0.015)           (0.057)

F-statistic                              28.656           35.172

Dep. var. mean                           1.496             9.876

Observations                             51,544           60,942

Notes: Standard errors (in parentheses) clustered by county.
Disaster payments are instrumented for with the share of voters
voting for a third-party candidate in the most recent presidential
election. All regressions include year and county fixed effects, as
well as controls for county characteristics. Farm income decile
indicators are included in the regression but are omitted from the
table for readability. All control variables are lagged by 1 year.

* Significant at 10%; ** significant at 5%, *** significant at 1%.
COPYRIGHT 2018 Western Economic Association International
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2018 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Deryugina, Tatyana; Kirwan, Barrett
Publication:Economic Inquiry
Date:Apr 1, 2018
Words:16219
Previous Article:MULTIREGIONAL FIRMS AND REGION SWITCHING IN THE U.S. MANUFACTURING SECTOR.
Next Article:THE IMPACT OF STOCK MARKET FLUCTUATIONS ON THE MENTAL AND PHYSICAL WELL-BEING OF CHILDREN.
Topics:

Terms of use | Privacy policy | Copyright © 2022 Farlex, Inc. | Feedback | For webmasters |