How Do Voters Retrospectively Evaluate Wasteful Government Spending? Evidence from Individual-Level Disaster Relief Jowei Chen and Andrew Healy Abstract: Why do voters often reward incumbents when they receive government spending? We develop a model in which distributive spending provides voters not just with a financial benefit, but also an opportunity to observe and judge the appropriateness of government decisions. Empirically, we test the model’s predictions using individuallevel data on FEMA disaster relief matched to voter turnout records, precinct-level election returns, and geographic data on hurricane severity. In accordance with the model, voters in areas experiencing severe hurricane conditions respond to the receipt of FEMA disaster aid with significantly higher turnout and electoral support for the incumbent administration. In contrast, voters show little response to aid in areas that experienced little damage and that audits identified as having received undeserved FEMA spending. Politicians thus appear to be constrained in their ability to use distributive spending to win elections since voters account for the merit of the aid they receive. Word Count: 9,025 Why do voters often reward incumbents when they receive government spending? Scholars have repeatedly found that the beneficiaries of distributive programs, such as veterans' benefits (Mettler and Stonecash 2008) and farm subsidies (Wolfinger and Rosenstone 1980), exhibit increased voter turnout (e.g., Ansolabehere and Snyder 2006) and stronger electoral support for incumbent politicians (e.g., Levitt and Snyder 1995). Under a classic view of distributive politics (e.g., Persson and Tabellini 2000), voters respond positively simply because they are motivated to maximize their share of government spending. They thus tend to reelect incumbents who have delivered distributive benefits to them. But these electoral rewards of distributive spending present an important theoretical puzzle: If distributive spending is so effective at winning votes, why do incumbents not simply provide the spending needed to guarantee reelection? Noting previous scholars’ repeated empirical finding that distributive spending garners votes, Golden and Min (2013) wonder: “It is perhaps surprising that any politician ever loses elected office given the impressive evidence that has been amassed showing the politicization of the public purse” (86). Given the existing literature, then, it seems puzzling why politicians do not engage in even higher amounts of electorally-motivated distributive spending to win more votes and guarantee reelection. We help to resolve this puzzle by arguing, both theoretically and empirically, that voters respond not merely to the quantity of distributive benefits they receive, but also to the appropriateness of such spending. In other words, voters use their experiences with distributive programs to evaluate the government’s quality. There are thus clear limits on policymaker incentives to increase distributive spending in advance of elections, as only spending that addresses a clear need may win votes. 2 A large literature on retrospective voting (e.g., Key 1966, Fiorina 1981; Markus 1988), has suggested that voters draw upon their past observations and experiences with government to retrospectively evaluate politicians' efficacy. As Fiorina (1981) explains, “In order to ascertain whether the incumbents have performed poorly or well, citizens need only calculate the changes in their own welfare” (5). By considering accessible metrics such as war casualties, voters can hold a government accountable for its foreign policy, for example (Fiorina 1981). More generally, voters can attempt to evaluate policies indirectly by observing their results. The receipt of distributive benefits actually offers voters an opportunity to directly observe government policy. Therefore, voters may utilize their receipt of distributive spending as information about a government’s performance, just as they consider the economy. Integrating distributive politics with retrospective voting, as we show, provides important insights into how voters appear to make decisions and the incentives that policymakers face. We develop and test a new theory in which voters view distributive programs as more than mere financial transactions. The intuition behind our argument, as illustrated in the paper's formal model, is as follows: Voters receive utility from distributive spending, but they also prefer highquality politicians who make appropriate policy decisions. Voters use their experiences with distributive programs to make inferences about the incumbent's intelligence and responsiveness. Suppose a government attempts to target aid to victims of a natural disaster. A government that appropriately delivers aid to victims signals both its responsiveness and intelligence. In contrast, a government that erroneously awards aid to non-victims will be perceived as responsive but lacking intelligence. Since voters want to re-elect politicians who are both responsive and intelligent, distributive spending wins more votes and motivates greater turnout when it is delivered to deserving victims rather than undeserving non-victims. 3 This theoretical framework brings together the literatures on the electoral consequences of distributive spending and the retrospective evaluation of politicians by voters. Our theoretical distinction between responding to distributive benefits and rewarding politicians for appropriate decisions has important implications for democratic accountability. If voters reward incumbents equally for any distributive spending, a government could simply provide as much spending to electorally important areas as needed to win the requisite number of votes. On the other hand, if voters respond primarily to spending when they are deserving recipients, then incumbents are constrained in their ability to win votes with pre-election spending. Under either theory, incumbents are incentivized to spend more resources before elections and in electorally important areas (e.g. Alesina and Cohen 1997; Kriner and Reeves 2013). 1 Only under the latter theory, however, are incumbents restricted in the kind of spending from which they will benefit. Our theory thus accords with canonical models in which retrospective voters use their observations to evaluate government performance. We build upon and extend this literature by explaining why governments do not spend even more than they do in electorally important areas. Undeserved spending may have little impact on their electoral fortunes. To empirically test our theory, we analyze individual-level data on Floridians who applied for FEMA disaster aid just before the November 2004 elections. In 2004, four separate hurricanes hit Florida in the three months before Election Day. We exploit this natural experiment to analyze voters’ responses to individually-targeted disaster assistance payments. Our data consist of all disaster relief applications made in response to the hurricanes and include recipients’ addresses, the amounts they received, and the dates of the payments. We link those 1 Much of the research on political business cycles has considered macroeconomic policy (e.g. Persson and Tabellini 1990; Rogoff 1990; Rogoff and Sibert 1988), although the logic is the same in terms of incentives to pursue expansionary policy before elections. 4 data to individual-level voter turnout records. Finally, we match each address to the storm severity that struck that location. As explained in the following section, victims and non-victims alike received significant disaster aid. The undeserved aid often occurred due to an excessive emphasis on haste and a lack of adequate oversight (Department of Inspector General 2005). 2 The results indicate that voter turnout increased in response to pre-election disaster aid in areas that experienced severe hurricane conditions. In hard-hit areas, even small amounts of disaster relief (less than $250) motivated voters to turn out. Amongst these voters, receiving a government check increased turnout by almost three percentage points. However, receiving a check of a similar size had no impact on voters’ behavior in areas that did not experience severe hurricane conditions. This paper proceeds as follows. First, we discuss the FEMA disaster aid process, explaining how FEMA inappropriately awarded aid to a significant number of non-victims in Florida. Second, to present our main theory, we develop a formal model of voter behavior, analyzing how voters respond upon receiving either deserved or undeserved disaster aid. Third, to test the model, we analyze individual-level data on FEMA applicants and voter turnout. We consider a variety of robustness checks, including placebo tests using payments received shortly after the election, that confirm the interpretation of the results. Finally, we also test the formal model by analyzing precinct-level election returns in November 2004, finding that disaster aid produced a significantly larger electoral payoff for the incumbent candidate in areas that experienced severe hurricane conditions. FEMA Disaster Aid and Government Audits 2 Part of the haste may have been motivated by electoral incentives. Research has found that disaster aid is higher and disaster declarations happen with greater frequency in election years (e.g. Downton and Pielke 2001; Garrett and Sobel 2003; Reeves 2011). 5 The FEMA Aid Decision Process: In August and September 2004, Florida was struck by four severe hurricanes: Charley, Frances, Ivan, and Jeanne. During the three months prior to the November 2004 election, all 67 counties in Florida were declared eligible by President Bush to apply for hurricane assistance under FEMA’s Individuals and Households Program (IHP). Formally, IHP provides up to $25,000 to compensate households for their disaster-related “necessary expenses and serious needs” not covered by insurance or other means (44 Code of Federal Regulations §206.110). To apply for aid, Florida residents had to contact FEMA by phone, by internet, or at a field office and request an inspection of their property. Applicants did not request a specific amount of aid. Instead, a FEMA inspector―generally an independent contractor and not a permanent FEMA employee―was tasked with identifying and assessing particular types of damage. FEMA then awarded a predetermined amount of aid for each category of damage using a standard but unpublished schedule. Inspectors marked damaged property on a checklist using a handheld computer, but they did not document damage photographically or with written descriptions. Hence, applicants did not have the opportunity to strategically misrepresent damage, but FEMA inspectors had the responsibility of making aid determinations with no documentation and no immediate oversight. In total, FEMA received 2.6 million assistance applications from 1.1 million Florida households. FEMA awarded a total of $1.2 billion to these households. Government Audits and Undeserved Aid: The Department of Homeland Security’s Office of Inspector General (OIG) issued a report in May 2005 that documented widespread payments to undeserving recipients after the four Florida hurricanes. The report looked at payments made under IHP and highlighted problems that led to erroneous payments. Some spending categories stood out as sources of questionable spending. In general, the report found 6 that speed took precedence over accuracy, leading to widespread spending going to undeserving recipients. It concluded that these issues it found in the one audited county cast “doubt about the appropriateness of IHP awards made to individuals and households in other counties of the state as a result of the four hurricanes, particularly those counties that had only marginal damage” (OIG 2005, 4, emphasis added). Some of the problems came from FEMA’s initial decision to override President Bush’s initial disaster declaration that omitted Miami-Dade county and 11 other counties from the group eligible for expedited assistance after Hurricane Frances. The initial declaration specified that Miami-Dade could only be eligible for aid after a Preliminary Damage Assessment (PDA), which FEMA overrode for unclear reasons. The OIG report highlighted that there was little evidence of substantial damage in Miami-Dade, a finding echoed in Senate hearings: In fact, the Miami-Dade County of Emergency Management described the damage from that hurricane as minimal, and the National Weather Service had no reports of flooding. Yet taxpayers bought Miami-Dade residents thousands of television sets, air conditioners, and other appliances, from microwave ovens to sewing machines. The taxpayers also bought rooms full of furniture, new wardrobes, and paid to repair or replace nearly 800 cars. It provided rental assistance to people living in undamaged homes. (Senate Committee on Homeland Security and Governmental Affairs 2005, 2). The FEMA decision to make Miami-Dade eligible paved the way for spending to flow quickly to residents, and the resulting reports of improprieties led to the Inspector General’s audits. According to the audit, the widespread availability of expedited funding combined with the four hurricanes in a short period of time to stretch FEMA’s resources thin. After Hurricane Frances, FEMA required its two contractors to increase the number of inspections they performed each day to around 15,000. To accomplish this task, each contractor hired about 1600 new inspectors, adding to an initial base of roughly 400. According to the report, “the new inspectors were not familiar with FEMA programs, received only 8 to 12 hours of basic training 7 on the FEMA inspection process, and their work was not closely monitored” (OIG, 29). The report noted at least the potential for conflicts of interest, as inspectors sometimes worked close to their own homes. More importantly, despite the speed with which applications were filed, it was official policy to issue payments with applications being checked only for completeness and not for errors. The report identified these errors and the undeserved spending that resulted as occurring with noticeable frequency in a number of particular spending categories. For these categories, FEMA provided inspectors with either vague guidance or gave them substantial latitude to determine eligibility. For example, funeral expenses were reimbursed for deaths with any hurricane-related connections, including stress-related causes, which could be broadly interpreted. The Florida Medical Examiners Commission found that more than 200 of the 319 cases of FEMA providing death benefits were for deaths unrelated to the hurricane, including a series of payments in areas where no deaths were disaster-related (Kestin and O’Matz 2005). FEMA also inappropriately awarded aid for cars. Vehicles worth as little as $1000 according their Blue Book value were reimbursed by FEMA for the maximum $6500 (OIG, 38). Some vehicles were replaced with no documented reason and others for flood damage despite no flooding occurring according to the National Weather Service. While funerals and automobile replacement appear to be clear cases of undeserved spending according to the report, these categories represent only 1-2% of total FEMA relief for the four hurricanes. Larger amounts of questionable spending occurred in the two categories FEMA identifies as expedited assistance and rental assistance, which together accounted for about 28% of awarded aid applications. General expedited assistance could include the kind of complete room replacement mentioned in the Senate report above. Applicants could receive the 8 replacement cost of an 11-piece bedroom suite if just one item in the room was found to be damaged (Senate Committee, 14). Likewise, the OIG audit documents found that “4,308 applicants who received rental assistance did not indicate a need for shelter at the time of registration or the $8.2 million they eventually received.” The audit further investigated a subsample of this group, finding that inspectors did not document reasons for assistance being granted, nor was there evidence that recipients resided elsewhere during the two months they received assistance worth $1452. Altogether, the evidence from the Inspector General’s audit indicates widespread undeserved spending after the four Florida hurricanes. Inadequate oversight resulted from the emphasis on delivering payments quickly. Questionable payments occurred most often for categories of spending, such as automobile replacement and rental assistance, with vague standards and limited required documentation. Moreover, the auditors found that FEMA spent the most on undeserved relief in those areas where residents incurred little damage. We use these findings to motivate our formal model in the following section, analyzing how voters respond when government inappropriately delivers disaster aid to undeserving non-victims. Theoretical Issues There are two key features of the formal model that drive our analysis. Incomplete Information: The incumbent's responsiveness and intelligence determine her likelihood of choosing a policy that benefits voters. As a result, voters prefer to elect a highquality candidate. Voters lack complete information about the incumbent's quality, however, and must infer her quality by observing and judging her policy choices prior to the election. 9 Hurricane Victimization and Aid: Nature determines whether a voter is victimized by a hurricane. While the voter knows whether he has been a victim, the government receives an imperfect signal of victim status, with a more intelligent government receiving a more accurate signal. Victims are assumed to derive the most benefit from receiving government disaster assistance, whereas aid directed to non-victims could be better spent in other ways. A responsive government's challenge is thus to provide disaster aid, but only to victims. Therefore, a voter can make inferences about the incumbent's quality by observing the government’s actions with respect to disaster aid. Formal Model Players: There is a single voter, V, who may be a victim of a hurricane in Period 1. There are two politicians. The incumbent, I, holds office in Period 1. The challenger, C, competes against I in an election held after Period 1. The election winner holds office in Period 2. For clarity, we use female pronouns for the Incumbent I and male pronouns for the Challenger C. Figure 1 illustrates the full sequence of play. [FIGURE 1 ABOUT HERE] Hurricane Victimization: At the outset of the game, Nature decides, with ½ probability, whether V is victimized by a hurricane. Let ω ∈ {victim, nonvictim} denote whether or not V was victimized. The realized value of ω is revealed privately to the Voter V, while each politician receives only a noisy signal about V's status. Strategies: In Period 1, the Incumbent I decides whether to deliver disaster aid to the Voter V by choosing s1 ∈ {0,1}. After observing I's choice, V chooses whether to vote in the election, and if so, whether to vote for I or the Challenger C. In Period 2, the election winner has to make a policy decision in an unrelated area in which an intelligent and responsive politician is 10 more likely to choose the correct policy. The politician decides whether to obtain a signal, θ 2 ∈ {A , B }, that can help her choose an appropriate Period 2 policy. The politician then chooses policy s 2 ∈ {A, B}. Politician Types and Information: Nature independently selects two qualities for each politician: Responsiveness, β I ,C ∈ {0,1}, and Intelligence, δ I ,C ~ U [1 2 ,1]. Responsiveness is a binary characteristic chosen with equal probabilities: Pr (β I = 1) = Pr (β C = 1) = 1 2 . A Responsive politician ( β = 1) internalizes the voter’s utility, while an Unresponsive one ( β = 0 ) does not. Intelligence is a continuous characteristic drawn from the uniform distribution U [1 2 ,1]. A politician's Intelligence, δ , determines the accuracy of the signals she receives about whether the voter is a victim (Period 1) and about the state of the world (Period 2). During Period 1, the Incumbent I receives a private signal, θ I ∈ {victim, nonvictim}, such that: Pr (θ I = victim ω1 = victim ) = Pr (θ I = nonvictim ω1 = nonvictim ) = δ I Hence, an Incumbent with the highest possible intelligence, δ I = 1, receives a perfect signal about whether the voter was victimized, whereas an Incumbent with the lowest intelligence, δ I = 1 2 , receives a completely uninformative signal. In Period 2, the election winner, p ∈ {I ,C}, chooses whether to obtain a noisy signal about the state of the world by paying a cost of ε > 0, assumed to be positive but negligibly small. If p pays this cost, she receives a signal, θ 2 ∈ {A , B }, that accurately represents the state of the world, ω2 , with a probability that matches her intelligence: Pr (θ= A ω= A= ) Pr (θ=2 B ω=2 B=) δ p . 2 2 Voter Utility: Voter V's payoff during Period 1 is: 11 m + n, if s1 = 1 & ω1 = victim U V ,1 = m, if s1 = 1 & ω1 = nonvictim , 0, otherwise (1) where m, n > 0. Thus, receiving aid provides positive utility to a nonvictim voter and an even greater utility gain when he is a victim. Voter V's payoff during Period 2 depends on the election winner's policy choice and is: K if s 2 = ω 2 UV ,2 = , − K if s 2 ≠ ω 2 (2) where K ∈ (0,1) is the positive utility the voter V enjoys when the policy choice matches the state of the world. If the politician chooses incorrectly, V incurs a cost of − K . Finally, Nature determines the cost of voting, γ , which is drawn from the uniform distribution U [0,1] and revealed to V. The Voter V chooses whether to incur this cost, choosing either the Incumbent I or the Challenger C if she does. We assume that V resolves indifference in favor of turning out. Combining the cost of voting with Eqs. 1 and 2, V's utility over the entire game is: m + n, if s1 = 1 & ω1 = victim γ if V turns out; K if s 2 = ω 2 U V = m, if s1 = 1 & ω1 = nonvictim − + , s ω if ≠ K 0 otherwise, − 2 2 0, otherwise (3) where the first term represents V's payoff from the Incumbent's Period 1 aid choice, the second term represents the cost of turning out to vote in the election, and the third term represents V's payoff from the election winner's Period 2 policy choice. Politician Utility: The politician serving in office during each period receives a payoff consisting of two components. First, a Responsive politician (β = 1) internalizes the Voter's 12 utility, while an Unresponsive one (β = 0) does not. Second, the politician incurs a cost if she chooses to deliver aid to the Voter. During Period 1, the Incumbent I holds office and enjoys the following utility: (m + n 2 ) if s1 = 1 U I ,1 = β I ⋅ U V ,1 − otherwise, 0 (4) where β I ∈ {0,1} denotes the I’s responsiveness, UV ,1 denotes the voter's utility during Period 1, and the final term represents the politician's cost of delivering aid ( s1 = 1) . The cost of delivering aid represents both the opportunity cost of spending public funds on hurricane aid as well as the bureaucratic cost of administering a disaster relief program. We set the cost of aid delivery as (m + n 2 ) for the following reasons: If the cost of delivering aid were lower than m, the model would have a trivial pure-strategy equilibrium in which responsive politicians always deliver aid, regardless of whether the Voter V is a victim. Analogously, if the cost of aid delivery were higher than (m + n ), then politicians would never deliver aid to either type of voter, regardless of victimhood. An intermediate cost focuses instead on the more interesting case in which it is cost-effective to aid victims but not non-victims. During Period 2, the election-winning politician p ∈ {I , C} holds office and enjoys the following utility: ε if p obtains a signal; U p , 2 = β p ⋅U V , 2 − otherwise, 0 (5) where β p ∈ {0,1} denotes the election-winner's responsiveness, U V , 2 denotes the Voter's utility during Period 2, and the final term represents the politician's cost of obtaining a signal about the state of the world. Since this cost is small, a responsive politician will always choose to obtain it to increase the likelihood of choosing the correct Period 2 action. 13 Equilibrium Results: For conciseness, we present only the subgame perfect Nash equilibrium results needed to calculate comparative statics for players' equilibrium-path behavior, omitting formal presentations of off-equilibrium-path strategies. Complete proofs appear in the Supporting Information (SI). Lemmas A and B describe politicians' strategies, while Lemmas C and D describe the Voter V's beliefs about the politicians. Lemma A: A Responsive politician (β p = 1) always obtains a signal in Period 2. An Unresponsive politician (β p = 0 ) never obtains a signal in Period 2. Conditional on obtaining a signal, θ 2 , the Period 2 politician always chooses the policy that matches her signal: s2 = θ 2 . Lemma A refers to the politician p's decisions in Period 2. First, p is only willing to pay the cost of obtaining a signal if she is responsive and internalizes the Voter's utility. If p does not obtain a signal θ 2 , then she has no information about the state of the world and thus randomizes, producing an expected payoff of zero. If p obtains a signal, she chooses the policy that matches this signal (s2 = θ 2 ) . Second, intelligent politicians get a stronger signal. As a result, voters are motivated to reelect a responsive and intelligent incumbent, since she has a greater chance of choosing the correct policy in Period 2. Lemma B: In Period 1, an Unresponsive Incumbent (β I = 0 ) will never deliver aid. A Responsive Incumbent ( β I = 1) will deliver aid if and only if she receives the signal that the voter is a victim. Lemma B describes the Incumbent's decision of whether to deliver hurricane assistance during Period 1. An Unresponsive Incumbent never delivers aid because she does not internalize the Voter V's payoff from aid. However, a Responsive Incumbent does internalize V's payoff, so she finds it worthwhile to deliver aid whenever she receives a signal that V was victimized, as such a signal is accurate with probability greater than ½. 14 Lemma C: If V is a victim, then receiving aid increases his estimate of the Incumbent’s probability of choosing the correct Period 2 policy by 11 45. Lemma D: If V is a non-victim, then receiving aid increases his estimate of the Incumbent’s probability of choosing the correct Period 2 policy by 1 21. Lemmas C and D describes how the delivery of disaster aid enhances the Voter V’s perception of the Incumbent I. If V is a victim (Lemma C), receiving aid causes him to positively update his beliefs about both I’s responsiveness and intelligence. Specifically, I’s delivery of deserved aid causes V to infer not only that she is responsive to the Voter, but also that she is intelligent enough to correctly identify him as a victim. These positively updated beliefs significantly increase V’s expectation that the Incumbent I would, if re-elected, select the correct policy in Period 2. But if V is a non-victim (Lemma D), receiving aid produces only a small net increase in his evaluation of the Incumbent. This increase is small because the delivery of undeserved aid sends mixed signals: I’s delivery of aid demonstrates that she is responsive, but the fact that she delivered this aid to an undeserving non-victim causes V to question I’s intelligence. Hence, although V still prefers I because of her demonstrated responsiveness, the undeserved aid delivery produces only a small net increase in V’s estimate of the probability that the Incumbent I would select the correct policy in Period 2. From the Lemmas, we derive the following two propositions, which we test empirically: Proposition 1: The increase in V's turnout probability from receiving aid during Period 1 is strictly larger if V is a victim than if he is a non-victim. Proposition 1, which follows directly from Lemmas C and D, states that the effect of aid on turnout depends critically on whether the aid was deserved. If V is a victim, then the delivery of aid sends a positive signal about both I's responsiveness and intelligence, thus unambiguously enhancing V's incentive to turn out and re-elect I. On the other hand, if V is a non-victim, the 15 delivery of aid sends a mixed signal, suggesting that I is responsive but relatively unintelligent. Therefore, the turnout effect from aid delivery is larger in magnitude when V is a victim deserving of assistance. Proposition 2: The increase in I's expected vote margin from delivering aid during Period 1 is strictly higher if V is a victim than if he is a non-victim. Proposition 2, which follows from Proposition 1 and Lemmas C and D, describes I's relative probabilities of receiving V's vote, according to whether he receives hurricane aid and whether he is a victim. I's expected vote share is significantly higher when V perceives her to have high responsiveness and intelligence. When V is victimized, the delivery of aid greatly enhances V's motivation to turn out and vote for I, thus giving her a significant increase in expected vote share. But when V is not victimized, aid delivery sends a mixed signal about I's quality and thus only slightly increases I's expected vote share. Together, Propositions 1 and 2 present the empirically testable predictions of our formal model, explaining how voters respond when they receive deserved and undeserved aid. In the following empirical sections, we test, at the individual level, whether the delivery of FEMA aid to hurricane victims increases voter turnout by a larger amount than the delivery of undeserved aid to non-victims (Proposition 1). We then analyze, at the precinct level, whether FEMA aid enhances the 2004 electoral vote shares of the President George Bush by a larger amount in victimized precincts than in non-victimized precincts, thus testing Proposition 2. Data and Summary Statistics We draw on a series of individual-level and geocoded data sources to test the propositions. Our data on natural disaster severity includes measures of flooding intensity based on newly available satellite imagery. We combine the disaster data with voter turnout records 16 and government spending data, matching these records to residential addresses using a new algorithm. The resulting dataset provides a precise picture of disaster conditions and aid payments. FEMA Disaster Aid: The FEMA disaster aid data cover all aid applications made by Florida residents after the 2004 hurricanes (Chen 2013). These data were made public after Freedom of Information Act (FOIA) requests from several Florida newspapers. Included in the data are the date of application, the date of a decision, the decision about whether to award aid, and the amount awarded. Additionally, the data include the category under which the application was made (e.g., housing assistance) and the general reason that an application was denied (e.g., insufficient damage). Based on the audit evidence described earlier, these variables provide insights into undeserved aid that we investigate below. We merge these data to Florida voter registration records, which are compiled by county election offices and report individual voter turnout in the 2000, 2002, and 2004 general elections. Most importantly, both the FEMA applicant data and the voter registration files report residential addresses. We used these residential addresses to merge the 1.1 million FEMA aid applicant records to the list of 10.3 million registered voters in Florida. We implemented a multi-stage process where we accounted for small differences in the reporting of address names between the two data sources. During this process, we merged addresses only where we obtained an exact match. We were able to merge the large majority of single-family addresses, but we did not merge the same share of apartment residences, whose addresses are sometimes listed with slight variations that could not be reconciled. In order to further eliminate cases of ambiguouslymatched addresses, we also restricted our analysis to addresses with no more than six voting members at the residence. 17 Disaster Intensity Data: We estimate hurricane severity for each registered voter’s residence in Florida. The National Oceanic and Atmospheric Administration (NOAA) deploys a large number of remote sensing devices to measure wind vectors in the regions surrounding each tracked storm. These observed wind vectors, combined with interpolated wind vectors in areas without a remotely sensed observation, are used to produce estimates of wind speeds across Florida throughout the duration of each hurricane. We then calculate, for each geocoded residence, the maximum wind speed observed across the four hurricanes in 2004. We also include in our measures of hurricane severity whether an address experienced flooding in the aftermath. To measure flooding, we utilize the satellite data collected and organized by the Dartmouth Flood Observatory. Those data contain GIS polygons that identify areas which were covered with surface water and carefully distinguish flood zones from other image interference such as cloud cover. We code an address as having experienced flooding if it is within any of the flood zones defined by the series of images for Florida in August and September of 2004. Summary Statistics: In Table 1, we report summary statistics for the key variables in the different datasets. The first set of means, standard deviations, and percentiles refers to voter turnout and voter demographics. Following that, we describe aid applications and awards. Finally, we report hurricane intensity and break aid down according to different levels of intensity. [TABLE 1 ABOUT HERE] In November 2004, approximately 70% of the registrants in Florida's voter file turned out to vote. About a quarter of voters experienced hurricane speeds of over 75 MPH at some point. A substantial share of registered voters (17.3%) applied for aid, with just over half of applicants 18 receiving some amount. Voters in high wind areas were much more likely to apply for aid. 46% of voters experiencing winds over 100 MPH applied for aid, while about 9% of those experiencing 60 MPH or less applied. Aid awards are thus much less frequent in low storm intensity regions. However, conditional on aid being awarded, the amount distributed is similar across low and high intensity areas. Evidence on Undeserving Aid Applicants: The two most common reasons for FEMA's rejection of an applicant are: 1) the applicant already having insurance are not carrying required insurance (43.5% of application rejections); and 2) insufficient damage to the applicant's property (31.2%). While FEMA approved a substantial number of applications in low-wind areas, FEMA also rejected a significant number of cases in areas that experienced significant hurricane damage, as Figure 2 illustrates. The map on the left side of Figure 2 depicts the maximum wind speeds observed across Florida during Hurricane Charley on August 13-14, 2014. The bright red line across this map follows the path of the storm's eye, tracing Charley's path as it made landfall on the Gulf Coast near Port Charlotte, traveled northeast through Orlando, and exited into the Atlantic near Daytona Beach. The map on the right of Figure 2 depicts the percent of Hurricane Charley applicants rejected because of insufficient damage, with the greenest areas having no such rejections and the reddest areas having over 70% of applicants rejected due to insufficient damage. [FIGURES 2 AND 3 ABOUT HERE] Figure 3 illustrates these same patterns at the individual level, with green dots representing applicants awarded at least $500 in aid and red dots representing applicants rejected 19 because of insufficient damage. In this map, the blue line depicts the path of Charley's eye as the storm made landfall in Port Charlotte, Florida on August 13, 2004. Together, these maps illustrate a striking geographic correspondence between hurricane wind intensity and applicant rejections. Applicants experiencing extremely high hurricane winds were virtually never found to have insufficient damage. Meanwhile, applicants who live far from the hurricane path nonetheless still often applied for FEMA aid and were, not surprisingly, rejected at relatively high rates because of insufficient damage. Nevertheless, the fact that people far from any hurricane path frequently applied for aid is consistent with rational behavior, given that applying for aid was virtually costless and applications were occasionally successful. 3 Analysis of Individual-Level Turnout Below, we consider a series of regression results that show the clear limits on legislators’ abilities to influence votes with distributive spending. We start by showing this result in Figure 4, which examines registrants who did not vote in 2000 or 2002. In this Figure, the blue points represent voters who lived in places that were severely hit by the hurricanes, while the red ones refer to people who lived in areas that were largely spared. 4 The picture thus omits voters who 3 In the supporting materials, we provide more details on the relationship between wind speed and rejections for insufficient damage. The kernel-weighted local polynomial regression results show the strong positive relationship between applying for aid and wind speed, along with the strong inverse relationship between being rejected for insufficient damage and wind speed. In addition, we collected data on a face-valid set of individuals who might be expected to be more likely to apply for aid despite not incurring damage: ex-convicts. We collected data on name and place of residence for 313,760 individuals who have been released from Florida prisons, finding that ex-convicts were more likely to be rejected for insufficient damage than other applicants, controlling for census block fixed effects. 4 The former category consists of voters living in locations where the maximum speed exceeded 100 MPH, they were less than 25 miles from a hurricane’s path, and the wind steadiness measure is less than 0.7. The latter category refers to voters from locations a maximum speed under 45 20 lived in areas of intermediate storm intensity, although they are included in the regression specifications that follow. In addition to potentially impacting incumbent vote share (Achen and Bartels 2004), the disaster could make it more difficult for voters to turn out (Gomez, Hansford, and Krause 2007; Hansford and Gomez 2010). Therefore, for each group, we consider the relationship between turnout and aid by comparing turnout to the baseline level amongst those who received no aid. [FIGURE 4 ABOUT HERE] Figure 4 illustrates that among disaster victims, aid recipients are substantially more likely than non-recipients to turn out. By contrast, aid receipt has no such effect for non-victims. The effect appears at relatively small amounts of aid, consistent with voters responding to their observation of government performance rather than the money enabling them to vote. The effect diminishes at higher levels of aid, consistent with increased disruption due to more intense storm conditions, which we account for in the regressions. Regression Results: In a series of regressions, we confirm this general result with models using more general measures of disaster impact. These regressions analyze all voters, covering the entire range of disaster intensity. In general, we consider regressions of the following form: Turnoutit = β0 + β1Reliefit + β2DisasterStrengthit + β3Turnoutit-1 + φControlsit + uit (6) In equation (6), Reliefit represents the FEMA relief spending that a household receives before Election Day. Since we observe an apparent effect of aid even at low levels in Figure 4, our main measure of relief is a dummy variable for the voter receiving positive aid before the election. We obtain similar results in terms of their interpretation when we control for the amount of spending that a person receives instead (see SI, Section 2). DisasterStrengthit captures the intensity of MPH, more than 50 miles from a hurricane’s path, and with a minimum wind steadiness exceeding 0.7. 21 disaster conditions that a voter experienced. We primarily utilize maximum wind speed at a voter’s location as the disaster intensity measure, considering alternative measures in the supporting information. Turnoutit-1 refers to whether the voter turned out in the previous election, so that the regression identifies the impact of aid conditional on a voter's past turnout. The control variables include voter characteristics such as Age and Gender, a dummy variable Applied indicating whether the individual applied for aid, and census block-group demographic variables. 5 As suggested by Angrist and Pischke (2009), we estimate the regressions as linear probability models. We obtain almost identical results in terms of interpretation with a logit specification, as we discuss in Section 2 of the SI. In all of the regressions, we correct the standard errors for clustering at the zip code level. Table 2 reports the results obtained by estimating equation (6) for registrants living in areas with different levels of maximum wind speed. 6 Column 1 includes registrants who experienced maximum wind speeds over 90 MPH. Columns 2 and 3 include registrants who experienced top wind speeds between 60 and 90 MPH, and under 60 MPH, respectively. 7 Column 4 includes all registrants. [TABLE 2 ABOUT HERE] The results indicate that higher levels of disaster aid lead to increased voter turnout, but only at high levels of hurricane wind speeds. Receiving aid before the election in an area that experienced wind speeds of at least 90 MPH predicts that a voter is 2.8 percentage points (p < 5 The regressions we report here include all registered voters. We obtain similar estimates and standard errors if we restrict our sample only to the applicants instead. Complete regression results are in the SI. 6 This regression includes controls for a cubic polynomial in wind speed. We obtain similar results for these coefficients if we include controls for duration of wind speeds and wind steadiness, as well. 7 We obtain similar results if we interact the aid variable with measures of disaster intensity. 22 .004) more likely to turn out to vote. However, this impact of disaster aid does not extend to the larger segment of the population that did not experience 90 MPH winds. The results indicate that aid has a precisely-estimated null effect both for voters living in areas with maximum wind speeds of 60-90 MPH and those from areas with maximum speeds below 60 MPH. Across all voters, receiving aid predicts the recipient is 0.7 percentage points more likely to turn out (p = .015), with this average effect driven entirely by the voters who experienced the most severe hurricane conditions. Spending Categories Identified by the Inspector General: As described earlier, the Department of Homeland Security’s Office of the Inspector General (OIG) identified several categories of spending that appeared to be particularly wasteful in its audit. OIG argued that the speed with which expedited payments were made and the lax oversight of various aid categories resulted in the awarding of significant amounts of undeserved aid. Thus, we consider the spending allocated under the four categories so identified by the OIG's report: expedited assistance (11.0% of approved aid applications), rental assistance (17.3%), transportation replacement (1.2%), and funeral support (0.04%). Columns 1-4 of Table 3 show regression results for categories of aid that the audit did not identify as containing questionable spending. Spending in these other categories constitutes the vast majority of spending in high-wind areas, and we correspondingly find a similar effect for those areas as we do for overall spending in Table 2. Even in lower wind areas, we find that these kinds of spending had a small, but positive effect. Altogether, spending outside the questionable categories was more likely to reflect real needs and accordingly appears to have a more positive effect on turnout even in places that also received wasteful spending. [TABLE 3 ABOUT HERE] 23 On the other hand, aid within the categories cited by the OIG's report has no positive effect on spending anywhere. Even in high wind areas, this kind of spending has a null effect. In the lower wind areas particularly mentioned in the report, these categories of spending enter with a negative sign. Hence, FEMA aid delivered to low disaster-damage areas within the four OIGidentified categories is the least likely to cause voters to turn out. In fact, such undeserved spending actually appears to make voters less likely to turn out. Robustness Checks and Placebo Tests: To establish the validity and robustness of our main empirical results, the SI (Section 2) presents a series of extensions and robustness checks. These additional tests use alternative definitions of hurricane victimization based on flooding maps, wind speeds, and geographic distance to storm paths. We also re-estimate our empirical models across voter partisanship and applicant status. Finally, to address possible omitted variable concerns, we conduct a series of placebo tests to see if there is any correlation between turnout and FEMA aid delivered just after Election Day. The SI discusses how these various checks affirm the validity and robustness of the main findings. Results for Precinct-Level Election Returns The previous results indicated that disaster aid only motivated turnout when that aid occurred in the areas hit hardest by the hurricanes. However, the formal model makes predictions about not only the effect of FEMA aid on voter turnout, but also its effect on the incumbent’s vote share. Specifically, Proposition 2 explains why the incumbent politician's electoral payoff from delivering deserved aid to hurricane victims is significantly higher than the electoral payoff from undeserved aid: The accurate delivery of aid to deserving victims sends a stronger, 24 unambiguously positive signal about the incumbent's quality and hence produces a larger boost for the incumbent's vote share in the subsequent election. To test this prediction, we compare George W. Bush’s precinct-level vote shares from the 2000 (pre-hurricane) and 2004 (post-hurricane) presidential elections. We analyze whether changes in precinct-level Bush vote share can be attributed in part to the relative amount of FEMA aid awarded in each precinct. We also analyze how this electoral effect of FEMA aid varies with the hurricane intensity that the precinct experienced. County election boards are permitted to redraw precinct boundaries between elections. Consequently, precinct-level vote counts cannot always be directly compared between different elections. In this section’s data, we include only precincts that satisfy the following three criteria: 1) the precinct's county provided vote counts for the November 2004 and 2000 presidential elections as well as the November 2002 Florida gubernatorial election; 2) the precinct's county made available maps of the 2004 precinct boundaries; 3) the precinct's boundaries in the 2000, 2002, and 2004 general elections are geographically comparable; and 4) the precinct had a nonzero population according to 2000 Census counts. Of Florida's 6,616 precincts as of November 2004, 5,073 (77%) satisfy all four of these criteria and are included in our analysis. Bush’s precinct-level vote shares in the 2000 and 2004 elections have a correlation of 0.92, suggesting that the 2000 precinct-level vote shares serve as an effective baseline measure of prehurricane partisanship across precincts. A second and more recent measure of precinct partisanship comes from the results from the 2002 gubernatorial election, in which the President’s brother, Republican Jeb Bush, defeated Democratic challenger Bill McBride by a margin of 56% to 43%. Jeb Bush’s 2002 precinct-level vote shares and President Bush’s 2004 vote shares exhibit a correlation of 0.93. 25 To estimate the electoral effects of FEMA awards, we regress Bush’s 2004 share of the two-party vote onto the 2000 and 2002 vote shares for George Bush and Jeb Bush, respectively, as well as two measures of FEMA aid distribution described below. The full model is: Bush04i = α + β1 ⋅ Reliefi + β 2 ⋅ DisasterStrengthi + β3 ⋅ LagGOPVotei + γ ⋅ Controlsi + εi , (7) where Reliefi is the measure of FEMA aid distributed to a precinct, which we measure in two ways. In one set of regressions, we utilize log ( Aidi + 1) , where Aidi is the FEMA aid dollars per capita directed to precinct i prior to Election Day. In a second set of models, we simply measure Reliefi as the percent of applicants who were awarded some amount of FEMA aid. As before, we control for DisasterStrength primarily with the maximum wind speed in a precinct. We control for a precinct’s pre-hurricane preferences by including George W. Bush’s share of the 2000 presidential vote and Jeb Bush’s share of the 2002 gubernatorial vote. Finally, Controls represents a vector of the following variables: the Mean Age of FEMA applicants in the precinct, the Female Proportion amongst applicants, the Hispanic Proportion for the precinct’s population, the precinct’s AfricanAmerican Proportion, and Median Household Income. All observations in our least-squares estimates of equation (7) are weighted by precinct population. [TABLE 4 ABOUT HERE] Table 4 estimates equation (7) across three subsets of precincts, grouped by their respective levels of hurricane severity during the 2004 hurricane season: precincts that experienced over 90 MPH wind speeds, precincts that had a maximum speed of 60-90 MPH, and precincts with less than 60 MPH maximum wind speeds. The results corroborate the theoretical predictions of Proposition 2. Model 1 estimates that for a highly victimized precinct with over 90 MPH winds, a 10% increase in the approval rate of FEMA applications causes a 1.0% increase in Bush’s 2004 vote share; this estimate has a 95% confidence interval of +0.4% to +1.6%. But in a moderately 26 victimized precinct with 60 to 90 MPH winds, a 10% increase in FEMA’s application approval rate causes a relatively smaller, 0.3% increase in Bush vote share (Model 2). And in the least victimized precincts with under 60 MPH winds, the effect is a negligibly small 0.1% increase in Bush vote share (Model 5). Hence, the results support the theoretical predictions concerning the electoral effects of disaster aid: FEMA aid boosts the incumbent’s electoral fortunes most strongly when awarded to hard-hit areas and most weakly when delivered to areas that were less affected. [FIGURE 5 ABOUT HERE] Figure 5 illustrates these same findings graphically by plotting the relationship between FEMA aid and Bush vote share separately across the three groups of precincts. In each plot, the vertical axis measures the difference within each precinct between George Bush’s (2004) vote share and Jeb Bush’s (2002) vote share. This measure thus identifies precincts where 2004 posthurricane electoral support for President Bush was high relative to the precinct’s pre-hurricane partisanship. The horizontal axis in each plot measures the percent of each precinct’s FEMA applications that were awarded aid. The dashed line in each plot represents the populationweighted least squares fit. These plots reinforce our main finding that FEMA aid’s positive effect on Bush vote share is conditional on the level of hurricane victimization. While all three groups of precincts exhibit a positive electoral responsiveness to FEMA aid, the slope is significantly steeper for the recipients who experienced the most severe hurricane conditions, consistent with the Proposition 2 prediction that distributive aid’s electoral impact depends on the deservedness of that spending. Conclusion 27 Our theory and empirical results have important implications both for electoral incentives and our understanding of the retrospective voter. First, distributive spending programs are more than simple financial transactions between government and voters. Voters also take advantage of their experiences with these programs to learn about the quality of government by judging the appropriateness of its spending decisions. Government audits found that FEMA often failed to make accurate assessments about the victimization of aid applicants. But the applicants themselves, with their personal knowledge of the hurricane conditions and damage at their respective residences, were well-placed to know whether they deserved any aid they received. Consistent with the empirical results, they may then make inferences about government efficacy and update their perceptions about the likelihood that politicians will make appropriate policy decisions in the future. Our findings thus build upon a rich, cross-national literature illustrating how voters learn from distributive programs. 8 For example, Manacorda, Miguel, and Vigorito (2011) found that Uruguayan recipients of cash transfers used their awards to learn about the policy preferences of the incumbent government. Thachil (2011) showed that, through receiving transfer payments, lower-caste voters in India learned about the incumbent government’s policy priorities and became more likely to support a party generally associated with the upper castes. In addition, Chong et al. (2010) used an experiment to conclude that simply receiving information about a government’s spending programs increased turnout and support for the incumbent government. 9 As with those findings, our results indicate that the relationship between distributive programs 8 In addition to voters learning from the personal experiences with receiving and not receiving distributive spending, they have shown the ability to credit and blame different levels of government for decisions that impact benefits (Gasper and Reeves 2011). 9 Also, De La O (2013) provides evidence that a Mexican conditional cash transfer program, Progresa, motivated voters to turn out to support the incumbent government. 28 and election outcomes is driven by more than merely a voter’s financial self-interest. 10 Distributive programs also provide important information about both the priorities and effectiveness of incumbent politicians. Moreover, consistent with Key (1966), voters appear to be doing much more than applying a heuristic of rewarding governments that provide them with benefits. Our results suggest that voters do not treat all kinds of spending equally. While such different treatment can result from voters not wanting the same things from government (e.g., Lazarus and Reilly 2010), our results suggest that different effects can also arise from variation in how voters learn from those spending decisions. In this sense, our theory builds upon and extends a significant literature on the policy feedback effects of social welfare programs. Soss (1999; 2002) argues that many income-based welfare programs, because of their paternalistic administration, may stigmatize clients and discourage their participation in politics. Meanwhile, universal, non-income-based programs, such as Medicare, Social Security, and veterans benefits (Campbell 2002; Mettler 2005; Mettler and Stonecash 2008), may encourage participation by giving recipients an enhanced stake in politics. Indeed, a voter’s interactions with a distributive spending program can do more than just change how voters feel about their own place in politics. These interactions offer a small, personal window through which a voter can judge the efficacy of government. Our results suggest that FEMA applicants retrospectively evaluate the government’s quality based on their respective individual interactions with the government. The results illustrate that an electionminded politician cannot expect to sway voters simply by gratuitously distributing financial 10 In addition to wanting to keep future benefits flowing, voters may also respond to distributive spending due to reciprocity. For example, Finan and Schechter (2012) find that politicians target spending to voters who surveys identify as likely to reciprocate. They also find that spending engenders a feeling of obligation in the recipient. 29 benefits to them. Rather, she must distribute benefits to voters in a manner that simultaneously reflects well upon her decision-making abilities, as voters evaluate not merely the distributive generosity of politicians but also the merit of their spending decisions. 30 References Achen, Christopher H. and Larry M. Bartels. 2012. “Blind Retrospection: Why Shark Attacks Are Bad For Democracy.” Vanderbilt University Center for the Study of Democratic Institutions Working Paper: 5-2013. Alesina, A., Roubini, N. and Cohen, G. 1997. Political Cycles and the Macroeconomy. Cambridge: MIT Press. Angrist, Joshua and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton, NJ: Princeton University Press. Ansolabehere, Stephen and James M. Snyder, Jr. 2006. “Party Control of State Government and the Distribution of Public Expenditures.” Scandinavian Journal of Economics 108(4): 547-569. Campbell, Andrea. 2002. “Self Interest, Social Security and the Distinctive Participation Patterns of Senior Citizens.” American Political Science Review 96 (3): 565-574. Chen, Jowei. 2013. “Voter Partisanship and the Effect of Distributive Spending on Political Participation,” American Journal of Political Science 57(1): 200-217. Chong, Alberto, Ana L. De La O, Dean Karlan, and Leonard Wantchekon. 2010. “Information Dissemination and Local Governments’ Electoral Fortunes: Evidence from a Field Experiment in Mexico.” Yale University working paper. De La O, Ana L. 2013. “Do Conditional Cash Transfers Affect Electoral Behavior? Evidence from a Randomized Experiment in Mexico.” American Journal of Political Science 57(1): 1-14. 31 Department of Homeland Security Office of Inspector General. 2005. “Audit of FEMA’s Individuals and Households Program in Miami-Dade County, Florida, for Hurricane Frances.” Office of Audits, OIG-05-20. Finan, Frederico and Laura Schechter. 2012. “Vote-Buying and Reciprocity.” Econometrica 80(2): 863-881. Fiorina, Morris P. 1981. Retrospective Voting in American National Elections. New Haven: Yale University Press. Garrett, Thomas A., and Russell S. Sobel. 2003. “The Political Economy of FEMA Disaster Payments.” Economic Inquiry 41(3): 496–509. Gasper, John T. and Andrew Reeves. 2011. “Make it Rain? Retrospection and the Attentive Electorate in the Context of Natural Disasters.” American Journal of Political Science 55(2): 340-355. Golden, Miriam and Brian Min. 2013. “Distributive Politics Around the World.” Annual Review of Political Science 16:73–99. Gomez, Brad T., Thomas G. Hansford, and George A. Krause. 2007. “The Republicans Should Pray for Rain: Weather, Turnout, and Voting in U.S. Presidential Elections.” Journal of Politics 69(3): 649-663. Hansford, Thomas G. and Brad T. Gomez. 2010. “Estimating the Electoral Effects of Voter Turnout.” American Political Science Review 104(2): 268-288. Kestin, Sally and Megan O’Matz. 2005. “Examiner Warns More FEMA Waste Is Likely.” The Sun-Sentinel. August 12. Accessed at: http://articles.sun-sentinel.com/2005-0832 12/news/0508111495_1_fema-payments-funeral-assistance-medical-examinerscommission. Key, Jr., V.O. 1966. The Responsible Electorate. Cambridge: Harvard University Press. Kriner, Douglas and Andrew Reeves. 2012. “The Influence of Federal Spending on Presidential Elections.” American Political Science Review 106(2): 348-366. Lazarus, Jeffrey and Shauna Reilly. 2010. “The Electoral Benefits of Distributive Spending.” Political Research Quarterly 63(2): 343-355. Levitt, Steven and James M. Snyder. 1995. “Political Parties and the Distribution of Federal Outlays.” American Journal of Political Science 39: 958-980. Manacorda, Marco, Edward Miguel, and Andrea Vigorito. 2011. “Government Transfers and Political Support.” American Economic Journal: Applied Economics 3(3): 1-28. Markus, Gregory B. 1988. “The Impact of Personal and National Economic Conditions on the Presidential Vote: A Pooled Cross-Sectional Analysis.” American Journal of Political Science 32(1): 137-54. Mettler, Suzanne B. 2005. Soldiers to Citizens: The G.I. Bill and the Making of the Greatest Generation. New York: Oxford University Press. Mettler, Suzanne B., and Jeffrey M. Stonecash. 2008. “Government Program Usage and Political Voice.” Social Science Quarterly 89(2): 273–293. Persson, Torsten & Tabellini, Guido, 1992. “The Politics of 1992: Fiscal Policy and European Integration.” Review of Economic Studies. 59(4): 689-701. 33 Persson, Torsten, and Guido Tabellini. 2000. Political Economics. Cambridge, MA: MIT Press. Reeves, Andrew. 2011. “Political Disaster: Unilateral Powers, Electoral Incentives, and Presidential Disaster Declarations.” Journal of Politics 73(4): 1142-1151. Rogoff, Kenneth. 1990. “Equilibrium Political Budget Cycles.” American Economic Review 80(1): 21-36. Rogoff, Kenneth and Anne Sibert. 1988. “Elections and Macroeconomic Policy Cycles.” Review of Economic Studies 55(1): 1-16. Samuels, David J. 2002. “Pork Barreling Is Not Credit Claiming or Advertising: Campaign Finance and the Sources of the Personal Vote in Brazil.” Journal of Politics 64(3): 845–63. Soss, Joe. 1999. “Lessons of Welfare: Policy Design, Political Learning and Political Action.” American Political Science Review 93(2). Soss, Joe. 2000. Unwanted Claims: The Politics of Participation in the U.S. Welfare System. Ann Arbor: University of Michigan Press. Thachil, Tariq. 2011. “Embedded Mobilization: Nonstate Service Provision as Electoral Strategy in India.” World Politics 63(3): 434-469. United States Senate Committee on Homeland Security and Governmental Affairs. 2005. “FEMA’s Response to the 2004 Florida Hurricanes.” May 18. Washington: U.S. Government Printing Office. Wolfinger, Raymond, and Steven Rosenstone. 1980. Who Votes? New Haven, CT: Yale Univ. Press. 34 Figure 1: Sequence of Play Nature determines the politicians' types: β I , β C ∈ {0,1} and δ I , δ C ~ U ( 1 2 ,1). Nature determines whether Voter V is victimized, ω1 ∈ {victim, nonvictim} I receives a signal about whether V is a victim: θ1 ∈ {victim, nonvictim} I decides whether to aid the voter, choosing: s1 ∈ {0,1} Nature determines V’s cost of turnout, γ ~ U [0, 1 2 ] V chooses whether to turn out. Turnout No Turnout V elects either I or C I Nature determines the election winner C I C Nature determines the state of the world, µ ∈ {A, B} Nature determines the state of the world, µ ∈ {A, B} I chooses whether to obtain signal C chooses whether to obtain signal I chooses policy: s 2 ∈ {A, B} C chooses policy: s 2 ∈ {A, B} 35 Figure 2: Hurricane Charley: Maximum Sustained Wind Speeds and the Percentage of FEMA Aid Applicants Rejected due to Insufficient Damage Highest Max. Wind Speeds (134 M.P.H.) 73% of Applicants Rejected for Insufficient Damage Lowest Max. Wind Speeds (10 M.P.H.) 0% of Applicants Rejected for Insufficient Damage Note: The left map depicts maximum sustained wind speeds measured throughout Florida during Hurricane Charley (August 13-14, 2004), with lower wind speeds shaded in blue and higher wind speeds shaded in red. The right map depicts the rate at which FEMA rejected aid applications due to insufficient damage. Areas with lower rejection rates are shaded in green, while areas with higher rejection rates (up to 73% rejected because of insufficient damage) are shaded in red. 36 Figure 3: FEMA Aid Applicants Rejected due to Insufficient Damage (Port Charlotte, Florida: Hurricane Charley Landfall on August 13, 2004) Hurricane Charley applicants awarded at least $500 in FEMA Aid Hurricane Charley applicants rejected because of insufficient damage Path of Hurricane Charley’s center eye during landfall (Port Charlotte, Florida, August 13, 2004) Note: The red line represents the center of Hurricane Charley's path during its August 2014 landfall. Green dots represent registered voters who applied for FEMA aid and were awarded at least $500. Red dots depict registrants who applied for FEMA aid but were rejected on the grounds of having insufficient damage. 37 Figure 4: Impact of Aid on Turnout for Victims vs. Non-Victims 0.1 2004 Turnout (relative to baseline) 0.08 0.06 0.04 0.02 Victim 0 Non-victim -0.02 -0.04 -0.06 -0.08 -0.1 Aid = $0 ΨϬфŝĚчΨϭ< Ψϭ<фŝĚчΨϮ< 38 ΨϮ<фŝĚчΨϯ< ŝĚхΨϯ< Figure 5: George W. Bush Precinct-Level Vote Share and FEMA Aid Awards 20% 40% 60% 80% 100% 0.2 0.1 −0.1 −0.2 −0.1 −0.2 0% Percent of Applicants Awarded Aid Non−Victimized Precincts (Under 60 M.P.H. Winds) 0.0 0.1 0.2 Moderately−Victimized Precincts (60−90 M.P.H. Winds) 0.0 0.1 0.0 −0.1 −0.2 George W. Bush 2004 minus Jeb Bush 2002 vote share 0.2 Heavily Victimized Precincts (Over 90 M.P.H. Winds) 0% 20% 40% 60% 80% 100% Percent of Applicants Awarded Aid 0% 20% 40% 60% 80% 100% Percent of Applicants Awarded Aid Note: Each point in this figure represents a single precinct. The vertical axes measure the difference between George W. Bush's November 2004 vote share and Jeb Bush's November 2002 gubernatorial election vote share. The dashed line in each plot depicts the least-squares fit. Observations are weighted by each precinct’s voting-age population. 39 Table 1: Summary Statistics Mean Standard deviation (25th percentile, 75th percentile) 0.703 0.564 0.587 52.45 0.681 0.544 0.457 0.496 0.492 17.882 0.466 0.498 (0,1) (0,1) (0,1) (38.781,65.951) (0,1) (0,1) 0.173 0.091 1759.818 1957.242 0.378 0.288 3065.773 3628.182 (0,0) (0,0) (835.97,1364.49) (835.97,1537.49) 60.561 0.016 56.608 18.624 0.125 38.535 (46.5,75.5) (0,0) (31.157,73.701) Aid in low intensity areas (max speed < 50 MPH) Did voter apply? Did voter receive aid? Pre-election aid received (recipients) 0.073 0.039 2006.428 0.26 0.194 3878.299 (0,0) (0,0) (835.97,1874.74) Aid in high intensity areas (max speed > 100 MPH) Did voter apply? Did voter receive aid? Pre-election aid received (recipients) 0.459 0.25 2056.031 0.498 0.433 3509.123 (0,1) (0,0) (852.87,1275) Voters Voted in 2004 Voted in 2002 Voted in 2000 Age White Female Aid Did voter apply? Did voter receive aid? Pre-election aid received (recipients) Total aid received (recipients) Hurricane intensity Maximum speed encountered Flood zone Distance from hurricane path 40 Table 2: Effect of Deserved vs. Undeserved Aid on Voter Turnout Dependent variable: Turnout in 2004 Maximum wind speed Over 90 MPH 60-90 MPH Under 60 MPH (1) (2) (3) Pre-election aid > 0 Maximum speed Applied for aid Age Age2 Female Local income Local share black Local share white Local mean age Voted in 2002 Constant Observations R-squared All voters (4) 0.0281*** (0.00958) -0.00132 (0.00141) 0.0737*** (0.0122) 0.00451** (0.00181) 0.000252 (0.00257) -0.00227*** (0.000774) 0.0328*** (0.00607) 0.00992*** (0.000816) 0.00166 (0.00257) -0.000422 (0.000383) 0.00750*** (0.00241) 0.0114*** (0.000212) 0.00729** (0.00299) -0.00190*** (0.000333) 0.0291*** (0.00512) 0.0103*** (0.000431) -4.53e-05*** (1.51e-05) 0.0171*** (0.00293) -0.00105 (0.0118) 0.129** (0.0533) 0.0360 (0.0642) 0.000761 (0.000863) 0.633*** (0.0412) 0.168 (0.162) -9.49e-05*** (6.62e-06) 0.0254*** (0.00113) 0.00945*** (0.00201) 0.0546 (0.0404) 0.0848** (0.0417) -0.000116 (0.000328) 0.491*** (0.0162) 0.223*** (0.0787) -0.000102*** (1.77e-06) 0.0290*** (0.00105) 0.0165*** (0.00120) 2.94e-05 (0.00789) -0.0111 (0.00843) 0.000561*** (0.000163) 0.391*** (0.00451) 0.139*** (0.0200) -9.49e-05*** (3.51e-06) 0.0266*** (0.000787) 0.0133*** (0.00132) 0.00778 (0.0158) 0.00404 (0.0169) 0.000328 (0.000211) 0.447*** (0.00901) 0.224*** (0.0318) 228,314 0.426 1,757,059 0.303 2,419,047 0.225 4,404,420 0.272 Notes: Standard errors, clustered at the zip-code level, are in parentheses. * p< 0.10, ** p< 0.05, and *** p< 0.01. 41 Table 3: Effect of Spending Outside and Within the Categories Identified by the Inspector General Dependent variable: Turnout in 2004 Pre-election aid > 0 Maximum speed Applied for aid Age Age2 Female Local income Local share black Local share white Local mean age Voted in 2002 Constant Observations R-squared Categories not identified as questionable Maximum wind speed All voters Over 90 MPH 60-90 MPH Under 60 MPH (1) (2) (3) (4) Categories that were identified as questionable Maximum wind speed All voters Over 90 MPH 60-90 MPH Under 60 MPH (5) (6) (7) (8) 0.0296*** (0.00764) -0.00138 (0.00141) 0.0760*** (0.0135) 0.00446** (0.00181) -4.49e-05*** (1.51e-05) 0.0172*** (0.00295) -0.00102 (0.0118) 0.131** (0.0535) 0.0363 (0.0642) 0.000769 (0.000865) 0.633*** (0.0413) 0.174 (0.162) 0.00570** (0.00252) -0.00227*** (0.000762) 0.0306*** (0.00644) 0.00994*** (0.000810) -9.50e-05*** (6.57e-06) 0.0254*** (0.00113) 0.00945*** (0.00200) 0.0534 (0.0393) 0.0833** (0.0405) -0.000114 (0.000327) 0.491*** (0.0161) 0.223*** (0.0765) 0.00904*** (0.00260) -0.000419 (0.000387) 0.00540** (0.00239) 0.0114*** (0.000212) -0.000102*** (1.77e-06) 0.0291*** (0.00106) 0.0165*** (0.00121) -0.000279 (0.00792) -0.0116 (0.00842) 0.000566*** (0.000164) 0.391*** (0.00452) 0.140*** (0.0200) 0.0126*** (0.00260) -0.00190*** (0.000332) 0.0278*** (0.00558) 0.0102*** (0.000431) -9.49e-05*** (3.50e-06) 0.0267*** (0.000789) 0.0133*** (0.00132) 0.00774 (0.0158) 0.00383 (0.0168) 0.000330 (0.000211) 0.447*** (0.00901) 0.224*** (0.0318) -0.0164 (0.0141) -0.00133 (0.00141) 0.0898*** (0.0135) 0.00448** (0.00178) -4.55e-05*** (1.49e-05) 0.0169*** (0.00297) -0.00123 (0.0119) 0.133** (0.0537) 0.0387 (0.0646) 0.000741 (0.000864) 0.633*** (0.0416) 0.171 (0.161) -0.0380*** (0.00495) -0.00225*** (0.000764) 0.0363*** (0.00611) 0.00992*** (0.000808) -9.48e-05*** (6.56e-06) 0.0255*** (0.00113) 0.00936*** (0.00200) 0.0556 (0.0394) 0.0848** (0.0406) -0.000116 (0.000327) 0.490*** (0.0162) 0.222*** (0.0767) -0.0531*** (0.00499) -0.000413 (0.000387) 0.0126*** (0.00271) 0.0114*** (0.000212) -0.000102*** (1.77e-06) 0.0291*** (0.00106) 0.0165*** (0.00120) 0.000135 (0.00793) -0.0114 (0.00842) 0.000564*** (0.000164) 0.390*** (0.00452) 0.140*** (0.0200) -0.0388*** (0.00495) -0.00189*** (0.000333) 0.0359*** (0.00559) 0.0102*** (0.000429) -9.49e-05*** (3.48e-06) 0.0267*** (0.000793) 0.0132*** (0.00132) 0.00874 (0.0159) 0.00468 (0.0169) 0.000325 (0.000211) 0.447*** (0.00903) 0.224*** (0.0318) 228,314 0.426 1,770,790 0.303 2,405,316 0.225 4,404,420 0.272 228,314 0.425 1,770,790 0.303 2,405,316 0.225 4,404,420 0.272 Notes: Standard errors, clustered at the zip-code level, are in parentheses. * p< 0.10, ** p< 0.05, and *** p< 0.01. 42 Table 4: Effect of FEMA Aid Awards on Electoral Support for George W. Bush Dependent variable: G.W. Bush Nov. 2004 vote share in the precinct Over 90 MPH Precincts (1) (2) FEMA Aid Dollars Per Capita (logged) 0.016*** (0.004) Percent of Applicants Awarded Aid Mean Applicant Age (100's of years) Female Proportion of Applicants Median Household Income ($100,000s) Hispanic Proportion of Population Black Proportion of Population G.W. Bush Nov. 2000 Vote Share Jeb Bush Nov. 2002 Vote Share Constant Observations R-squared Maximum wind speed 60-90 MPH Precincts (3) (4) 0.006*** (0.001) -0.082 (0.074) -0.377** (0.142) -0.021 (0.034) 0.029 (0.040) 0.001 (0.050) 0.461*** (0.057) 0.474*** (0.062) 0.171 (0.092) 0.102** (0.032) -0.058 (0.079) -0.454** (0.146) -0.025 (0.035) 0.026 (0.042) 0.012 (0.051) 0.437*** (0.059) 0.481*** (0.064) 0.240** (0.092) 0.847 245 0.84 245 Notes: Standard errors are in parentheses. * p< 0.10, ** p< 0.05, and *** p< 0.01. 43 Under 60 MPH Precincts (5) (6) 0.003*** (0.001) -0.114*** (0.023) 0.107* (0.043) 0.024** (0.008) 0.123*** (0.013) 0.081*** (0.014) 0.201*** (0.017) 0.702*** (0.018) -0.101*** (0.026) 0.032*** (0.008) -0.118*** (0.024) 0.118** (0.044) 0.021** (0.008) 0.118*** (0.013) 0.086*** (0.015) 0.212*** (0.017) 0.699*** (0.018) -0.096*** (0.026) -0.167*** (0.017) 0.009 (0.023) 0.006 (0.006) 0.094*** (0.005) 0.050*** (0.008) 0.379*** (0.013) 0.561*** (0.015) 0.025 (0.016) 0.013*** (0.004) -0.168*** (0.017) 0.009 (0.023) 0.002 (0.006) 0.100*** (0.005) 0.060*** (0.007) 0.379*** (0.013) 0.565*** (0.015) 0.022 (0.016) 0.908 1681 0.906 1681 0.948 3147 0.947 3147