Using Polls to Measure Pork

Transcription

1 Using Polls to Measure Pork Jaob Allen * Craig MIntosh ** August 2006 Abstrat Polls are naturally used by politiians and businesspeople to form expetations over politial outomes, but every poll omes with its margin of error. We suggest that the shok to expetations aused by voting realizations provides us with an innovation whih an be used to perform an event study. In partiular, forward-looking business behavior prior to the eletion is optimized relative to expetations formed through polls, and immediately after the eletion it adjusts aording to the new information revealed by the eletion itself. If there is no reation to the surprise vote, business behavior is not a funtion of politial behavior, and so we infer that no redible promise of pork has been made. Variation in the magnitude of this response aross eletoral distrits therefore gives us a way of identifying the loal-level harateristis that make it both desirable and redible for politiians to use pork. Using data from Ugandan mirofinane investors, we test the disontinuous response aross ounties of Uganda to the 2001 presidential eletion. Controlling for loal variation in violene, threatmaking, and expliit promises of pork, we find that both inumbent and opposition politiians are highly redible. Keywords: eletions, patronage, unertainty. JEL lassifiation: O16, N27, H42 * Graduate Student, International Relations/Paifi Studies, U.C. San Diego, 9500 Gilman Drive, La Jolla, CA, jaoballen@usd.edu. ** Assistant Professor, International Relations/Paifi Studies, U.C. San Diego, 9500 Gilman Drive, La Jolla, CA, tmintosh@usd.edu. Thanks to Gary Cox, Stephen Haggard, Ted Miguel, Devra Coren Moehler, Jeremy Weinstein, partiipants in the Politial Eonomy of Development Seminar at UCSD in May 2006, and the staff of FINCA/Uganda; thanks also to Joanitah Kirunda, Robert Apell, Mukalazi Deus, and Jennifer Bartlett for exellent researh assistane. 1

2 1. Introdution. Eletoral politis has long been viewed as a strategi interation between voters and politiians. Voters delegate governane with an expetation that the publi and private goods provided will inrease their welfare. Politiians, on the other hand, likewise seek the benefits they an arue through offie, both as a itizen-andidate and from the ability to extrat rents. As suh, voters hoose andidates based on their estimations of politiians poliy promises aording to mehanisms of seletion (Fearon 1999a) and santioning (Ferejohn 1986). In turn, politiians ourt voters in an attempt to minimize the distribution required to obtain offie (Alesina 1998; Dixit and Londregan 1996). In this paper we build on this ommon oneption of the eletoral game in two ways one theoretial, the other empirial. Theoretially, our model highlights the importane of the redibility problem that both politiians and voters have in promises to exhange votes for pork (see also Robinson and Verdier 2003). As Keefer (2004) points out, most analyses either assume that politiians promises are universally redible (f. Cox and MCubbins 1986; Lindbek and Weibull 1987; Dixit and Londregan 1996) or that they never are (f. Ferejohn 1986; Alesina 1987; Besley and Coate 1997). Along with a few reent attempts to reassess these assumptions, we argue that between these two extremes is a muh more ommon set of intermediate ases in whih politiians redibility varies aording to the nature of their relationship with voters. We model the irumstanes in whih both voters and politiians possess the attributes that will make possible redible patronage promises, and suggest an empirial test based on event studies from the finane literature. The test is similar in spirit to Fisman (2001), but the use of the surprise vote makes the tehnique generally appliable and we do not need speial data on the onnetions between politiians and speifi firms. The idea behind this test is that forward-looking business behaviors (suh as stok pries, business borrowing, housing starts, or investor onfidene indexes) will respond in a disontinuous fashion to the revelation of new information whih alters eonomi expetations. The gap between loal voting outomes as predited by polls and the atual vote on eletion day reveals new information about the extent of loal support for national politiians. The margin of error in eletion polls generates an 2

3 informational shok whih is (in theory) randomly distributed and observable for numerous eletoral units within a ountry. If any link exists between voting and distribution (Finan (2004) shows that suh a link exists in Brazil, and Miguel and Zaidi (2003) in Ghana), then business behavior will respond. Sine we an use the disontinuous shift in eonomi behavior observed just after the eletion, the statistial test laid out in this paper is both well-identified and new to the literature. If politiians annot redibly ommit to redistribute on the basis of voting, then business will not respond, and so an eletion won by an inredible politiian sees no alteration in eonomi behavior as a result of the surprise vote share. Our model agrees with both Keefer (2004) and Robinson and Verdier (2003) that the primary determinants of suh two-sided redibility are the extent of repeated interation between politiian and voter and the propensity for olletive ation on the part of voters. For Keefer this is ahieved either through investments aimed at building national (and thus long-lasting) parties or by buying into the redibility that underlies existing lientelisti networks. The existene of suh networks is the primary way in whih voters an oordinate patrons provide the oordination on the use of a retrospetive voting rule. 1 The fous on lientelism, as usually defined, presents empirial problems beause these networks are not in general observable to researhers. Given that these networks are relatively well-understood by loals, however, this suggests that we should be trying to infer from the way that loals respond to new politial information whether they foresee patronage. In this paper we suggest a definition of pork (namely, any politial transfer that is ontingent on voting behavior) whih is impreise in terms of hannels but whih opens up a variety of empirial tests. If a politiian is redible in any part of the ountry during an eletion, we an use ross-setional variation to measure the magnitude of the response to the surprise vote share. This allows us to test for the loal-level harateristis whih make the eletorate able to redibly ontrat with politiians. Using this ross-setional variation, we suggest and implement 1 Similarly, Robinson and Verdier (2003) argue that the groups with whom politiians an redibly exhange with will be determined by the soial network of individuals whose behavior they an observe relatively well perhaps beause they interat soially with them. They likewise onlude that lientelisti relationships are the most frequently-used mehanism for engaging in suh redible exhanges. 3

4 tests for suh diverse questions as whether politiians target ore or swing voters, whih ethniities believe a partiular politiian, and whether winning speifi sub-regions matters to politiians in a majoritarian system. We proeed to speify the model and empiris in greater detail. In Setion 2 we derive from a theoretial model an array of empirial tests that an be performed using the response in forward-looking business behavior to this surprise vote share. Using data whih span the 2001 Ugandan presidential eletion as desribed in Setion 3, we arry out these empirial exerises in Setion 4 and find signifiant mean effets using a variety of metris. Setion 5 provides a disussion of robustness of the estimators desribed in the empirial setion, and Setion 6 onludes with some suggested diretions for future researh using this way of measuring pork and ampaign redibility. 2. Model. As investors approah an eletion, they attempt to inorporate the likely results of that eletion into their intertemporal investment behavior. Hene information (suh as a pre-eletion poll) that allows them to foreast politial outomes is used to make forward-looking investments prior to the eletion. A politiian approahing this eletion has a similar informational problem whether and how to make ontingent promises diretly to onstituents in order to drive up votes; that is, whether to use pork. In assessing where promises of pork are most useful, the politiian also will use poll results to gauge how responsive the eletorate has been to reent ampaign promises in different regions of the ountry. In this sense a strategi dane takes plae between a national politiian and loal voters in whih poll outomes give both sides the most diret soure of information over the unertainties posed by the eletion. Following voting models as desribed by Lindbek and Weibull (1987) and Stromberg (2004) we model eletions as a two-period game between voters and a politiian in whih the first period is prior to the eletion wherein the politiian makes promises over post-eletion pork whih maximize his expeted 4

5 utility. 2 In this period opinion poll outomes are observed, and the equilibrium in the pre-eletion period is the point at whih voters, businesspeople, and the politiian have optimized their behavior onditional on their best guess of how others will play. In the seond period the eletion outome has been observed, and the error between the poll and the observed eletion outome represents a stohasti shok to whih equilibrium behavior may adjust. The loals understand the behavior of the politiian well enough to know what his ex post optimal payment will be, and so the politiian may fae a redibility onstraint that auses promises to be ignored by voters. If politiians ould ontrat with individuals, they would do so. Beause most modern voting systems are anonymous at the individual level, however, we assume that politiians attempt to ontrat with voters at the smallest unit at whih they an observe outomes (hereafter the ounty). 3 Therefore the politiian makes offers of pork to ounties as a ontrat ontingent upon voting behavior. We define pork broadly to inlude any form of distribution (whether patronage, subsidy, favoritism, or loal publi good) whih is offered strategially to voters in a given eletoral distrit in order to influene their voting. 4 The short-term, strategi use of pork as defined here is ontrasted with promises regarding issue positions and programmati poliies that are less malleable in the short term (Dixit and Londregan 1996). The model is similar in spirit to Stromberg (2004) and Ferraz & Finan (2005) in that politiians balane the benefits of diverting money against spending it to influene eletions, however we fous expliitly on whether promises that the politiian wishes to make ex ante will be redible ex post. Hene we fous our attention on expetations in period 1 over the politial transfers that will be made in period 2. 2 Payments to voters prior to the eletion annot be time-onsistent in the absene of ontrating; voters will simply take the money and then vote for whihever politiian they prefer (Robinson and Verdier 2003). 3 In systems suh as Kenya s where eletions line up voters to be ounted in publi, there is no reason to think that ontrating with individuals will not exist. Anedotes of the more olorful kind of individual ontrating mehanism abound, from the failure to oil one of the levers in the old voting mahines in Tammany Hall (Popkin) or the use of amera ell phones to transmit an image of the ompleted ballot in Mexio (Molinar). 4 For relevant disussions regarding the role of lientelism in suh behavior, see Keefer (2003) and Robinson and Verdier (2003). 5

6 Full Credibility. We begin by modeling the simple ase where the politiian has a ommitment devie available in period 1 whih makes him able to make time-onsistent promises over period 2 behavior The Politiian: If eleted, the politiian has a pool of resoures denoted by K at his disposal whih an be diverted to private uses or used to influene the behavior of voters. The politiian maximizes utility from K by balaning the utility from diversion against the inreased vote share that an be ahieved by spending on pork (f. Olson 1993). In order to affet the outome of the eletion, this pork must be promised (expliitly or impliitly) prior to the eletion in the form of ampaign promises. The politiian has a period 1 net present valuation Ω 1 of being in offie in period 2 (and 0 if not in offie) and uses transfers t to eah ounty to inrease the probability of re-eletion, F ( t 1,.., t N ). For simpliity, we model a majoritarian system where ounties have the same number of voters so that we an write this total probability as f ( t ) F(.) = N, where f ( t ) is the vote share reeived in ounty. Voters in a given ounty have a distribution of preferenes towards the politiian suh that, in the absene of any transfers, their voting outome will be f (0). 5 If there exists a large mass of voters who are ambivalent between the pork-paying politiian and the opponent, or if voters are easily bought, then we will see a high slope of the funtion f (.) as the politiian direts transfers towards a ounty. This derivative with respet to transfers to any given ounty is denoted by f ( t ), and we assume f '> 0 and f '' < 0. This implies that higher transfers to a ounty inrease the vote share from that ounty, but that there are diminishing marginal returns to pork. The justifiation for this is that a rational politiian will do first those things that are most effetive but as an inreasing amount of pork is direted to a given region it is pushed into projets to whih voters are less responsive. Even under full redibility, voters know they ' 5 This way of modeling the problem is onsistent with the presene of a seond politiian making fixed offers of patronage. 6

7 will only reeive the transfer if the politiian is eleted (whih the voters expet to our with probability V Fˆ ), and so the expeted value of a promised transfer is promise of t is thus f ( F ˆ V t ). 6 We assume that if vitory, differs from Fˆ V t. The atual number of votes ahieved by a P Fˆ, the politiian s estimate of his own probability of V Fˆ, the voters, that the true value F lies weakly between these estimates. Given a linear utility from diversion ω * K t and a utility of 0 if not eleted, the politiian s problem is to: Max t f ( Fˆ N V t ) Ω1 + ω ( K t ), whih has the first-order ondition Fˆ V f '(.) P ω( K t ) Fˆ 1 = ω N Ω +, meaning that the expeted marginal benefit of promising a transfer to a ounty will equal the expeted marginal opportunity ost of the transfer. This expression an be rearranged as: Fˆ f ' = Fˆ P V * Ω 1 ωn + ω( K Fˆ = t ) Fˆ P V ωn *. benefit of offie Fˆ Fˆ P V is the ratio of the politiian s to the voters predited probability of vitory in the national eletion. ˆ P F This term might be alled politiian s onfidene. Where > 1, the politiian is overonfident and ˆ V F ˆ P F thinks that the voters are more responsive to pork than they atually are. When < 1 ˆ V F the politiian is underonfident and promises too many transfers, being too pessimisti about his hanes and underestimating how easily votes an be bought. 6 For simpliity we assume that the number of ounties N is large enough that we an ignore the marginal effet of inreased V promises to a given ounty on voters estimated probability of overall vitory in the national eletion, Fˆ. The additional term substantially ompliates the derivative and adds little to the intuition of the loal response to promises of pork. 7

8 A primary role played by polls is the oordination of expetations, though the transpareny of the polling proess and freedom of the media in general obviously affets the extent of oordination between andidates, voters and businesspeople. 7 In any eletion onduted with ommonly observed national polls ˆ P F ωn we should see = 1, leaving us with a first-order ondition of: f ' =. ˆ V F Ω + ω( K ) 1 t The right hand side is a funtion only of the politiian s parameters, and so is a onstant aross ounties whih we will denote by W 1. Hene the interior solution to this problem is f ' = W1, and so any ounty that reeives transfers will have the same slope on the voting prodution funtion. Figure 1 illustrates this, where ounties 1 and 2 are at interior equilibria, and ounty 3 reeives no transfers in equilibrium. If ounties exist where even the first unit of transfers have lower probability effets than the opportunity osts (suh as ounty 3 in Figure 1), then these ounties will reeive no transfers in equilibrium. The optimal transfers are thus: (1) f '(.) = W1 s.t. f '(0) W1. t = 0 s.t. f '(0) < W Investors: Politis affets business outomes both through the provision of transfers and through the diret effets of poliy itself. The polls are used by investors to form expetations over both of these distint proesses to understand how loal voting will alter the level of transfers and to predit the outome of the eletion itself. We begin by forming rational expetations for voters given the model of provision of loal-level transfers outlined above. Investors understand the loal environment well enough to infer the funtion f Ft ˆ ) and engage in some intertemporal investment behavior whih is inreasing in t. Businesses read the urve that relates pork to probability of eletion in reverse, by using polls to foreast the extent of pork that their region will 7 Diaz-Cayeros, Magaloni and Weingast (2003b) argue that state ontrol of the media was important in maintaining the PRI s politial dominane in Mexio not beause it brainwashed people, but beause it portrayed a strong PRI and a weak opposition, thus inreasing the value of the PRI s promises of pork. ( 8

9 1 be provided (shown in Figure 2). We refer to the business metri as ( F, f ( )) π vote, where π is an inreasing funtion of t 1 = f ( vote ) and F indiates the outome of the national eletion. 8 We write the surprise vote share as ε, where vote = poll + ε. Figure 3 illustrates the linear ase using best responses to the model of pork developed in Setion 2.1; beause ounties 1 and 2 reeive positive transfers in equilibrium then dπ dε gives us the slope of 1 f π ( ) in those ounties, and in ounty 3 there is no slope beause there were no transfers in dπ equilibrium. For a ounty that does not expet to and does not reeive pork, = 0. For some dε ounties, however, the surprise vote share swithes it from one whih expeted to reeive pork to one that dπ did not, or vie versa. In this ase > 0, but it will also be true that dε dπ dε 1 < π f beause 1 f is a onvex funtion, meaning that the slope of the line whih onnets f (0) and the optimal transfer has a less steep slope than ( f ) evaluated at the optimal transfer. The swithers, then, have slopes on their 1 response funtions whih are intermediate between the zero slope of the no-pork ounties and the slope of a pork ounty. The seond distint use of the polls is to foreast the outome of the national eletion Fˆ. Beause the outome is binomial, the variane of the eletoral outome is Fˆ (1 Fˆ ). The mean and variane fully haraterize the binomial distribution, and so we model investment response to Fˆ in a mean-variane form: investment prior to the eletion is based on the expeted outome, and is depressed by the variane of the outome. An investor who is indifferent to the outome of an eletion displays no response to the 1 1 variane. Let = π ( F = 1, f ( poll )) π ( F = 0, f ( poll )) φ be the differene in optimal outomes that arises solely as a result of poliy effet, holding transfers onstant. From here we an model variane 8 Metris of π whih are readily observable and ould be used inlude business borrowing, loal investor onfidene indies, loal housing starts, and stok pries of ompanies based on the loation of their headquarters. 9

10 as entering the investor deision by suppressing pre-eletion investment through a term ρ φ Fˆ (1 Fˆ ), where ρ is a ounty-level oeffiient of risk aversion. 9 Immediately before the eletion, then, we will observe an outome based on the expeted mean and variane of optimal behavior under the winner and the hallenger: π ( ˆ 1 ) [ ˆ 1 ( ) ˆ F, f ( poll ), ρ, φ = Fπ p = 1, f ( poll ) (1 F) π ( p = 0, f ( ))] 1 poll ρ φ Fˆ (1 Fˆ ) If there is no risk aversion, the outome observed prior to the eletion is a simple weighted average based on optimal behavior under the two andidates and the information in the polls. As this pre-eletion investment outome depressed by the unertainty indued by the eletion. π. ρ inreases we see Immediately after the eletion we see an outome whih maximizes profits given the realization: 1 ( F, f ( )) eletion as: vote π = π =. From here we an form the disontinuous hange in outomes seen around the ( post eletion equilibrium) ( pre eletion equilibrium) π 1 1 ( F, f ( vote )) π ( Fˆ, f ( poll ), ρ, φ ) By adding and subtrating the outome that would have been observed after a vitory if the outome exatly equaled the poll, π F = 1, poll ) and olleting terms, we get: (. π = π ( F = 1, ε ) + φ (1 Fˆ )(1+ 1 dπ ( f (.)) = ε (1 Fˆ + φ )(1+ d( vote ) ρ Fˆ ) ρ Fˆ ) for s. t. φ < > 0. The first term on the RHS shows the jump in investor outomes that an be attributed to ε and not to the surprise in the overall eletion, whose effet is given by φ ( 1 Fˆ )(1+ ρ Fˆ ). The intuition for the ambiguous sign of the variane effet is that investors who preferred the hallenger ( φ < 0) have their 9 This way of expressing risk aversion is onsistent with a story wherein fixed investment deisions are influened by politial onerns, and so businesses refrain from sunk investment until the outome of the eletion is known. 10

11 outomes distorted by risk in the same diretion as the effet of the disovery of the eletion outome, while those who preferred the vitor see outomes distorted in the opposite diretion by risk. As F ˆ 1 there is no overall surprise present in the eletion, and as φ 0 outomes for business are idential under the two andidates. In these ases only the loal-level shok is informative and no meaningful innovation is ontained in the outome of the national eletion. In other words, we an deompose the surprise generated by the eletion into two omponents: who won? and did we bak the winner?. This deomposition of the hange in investment is useful beause it larifies the orthogonality assumptions that must made to draw ausal inferene from the orrelation between ε and π. In an eletion with a heavy favorite, F ˆ 1 and we an isolate the effets of ε. In this ase for an inremental 1 dπ ( π ( f (.)) ε, = poll dε ( vote ). Consequently, the hange in business behavior over the surprise vote share allows us to measure the marginal responsiveness of business to the loal vote. 10 In suh an eletion, the orrelation between π and ε allows us to ask the following question: do investors alter their behavior in regions that unexpetedly baked the winner of the eletion? If we have F ˆ < 1 and φ 0, there will be some disrete adjustment on average to the announement of eletion results. In this ase, use of the orrelation between π and ε requires the additional assumptions that ε φ (the surprise vote share is orthogonal to the loal benefit from vitory) and that ε ρ φ (the surprise vote share is orthogonal to the produt of loal-level risk aversion and loal benefit). Wlezien & Erikson (2006) haraterize the error in a poll as arising from three omponents: sample design mistakes, a bias in the house that onduts the poll, and random sampling error (inluding errors in the questionnaire). Our polls ome from a single surveyor, and so by demeaning 10 Note that a poll whih is inaurate at the ounty level generates more variation in ε than an aurate poll, and so improves the statistial identifiation of this relationship! 11

12 ε we an remove the house effet. The random error is orthogonal to ounty-level harateristis, and so the assumption required in eletions that ontain surprises an be stated as the following: the error that arises from sample design flaws must be unorrelated with variation in investor politial preferenes and in risk aversion. E We have a simple test for whether the eletion itself was not a shok to outomes: if the term ( φ ( 1 Fˆ )(1+ ρ Fˆ )) is non-zero, we will observe a disontinuous shift in outomes aross all units. This suggests that we test for the presene of a mean disontinuity around the time of the eletion. In setion 5.1 we perform this test and find no suh disontinuity. 11 Setion 5.3 onduts tests to verify the randomness of ε. Given Museveni s large lead in the polls during the 2001 Uganda presidential eletion, and the absene of a mean shok during the eletion, we proeed to investigate the influene of redibility in this three-sided game under the assumption that our tests are orretly isolating a the effet of ε Credibility in the absene of a ommitment devie. In the absene of a ommitment devie, behavior by both voters and the politiian must be timeonsistent. Appendix A3 disusses the irumstanes under whih the voters will tell the truth in polls. We argue that this will prove a dominant strategy. Credibility then beomes a funtion of two fators: first, the politiian s period 2 benefit from atually paying transfers, and seond, the ability of voters in a speifi ounty to punish the politiian in future eletions for not arrying through on promises. We refer to this seond harateristi as the memory of voters. 11 In developed-ountry two-party systems we might expet the primary politial leavage to be labor vs. apital, in whih ase investors are likely to share preferenes over politial outomes, and so φ has the same sign for most agents. In this environment if φ 0 we should see a mean shift in outomes as, for example, markets reat to the eletion of a pro-business president. In an Afrian ountry, where leavages follow less stritly soio-eonomi lines, it is more likely that the average national effet has a zero-sum omponent, where the responses of the winners and the losers to the eletion balane on average. Thus failing to find any average shok to outomes annot be onsidered to have eliminated the possibility of a orrelation between φ and ε in the Ugandan ontext. 12

13 Voters memory refers to their ability to ensure that eleted politiians are held aountable for their performane in government. Aountability, of ourse, is a muh-studied topi in politial siene. In the ontext of our analysis, voters memory may be affeted by many fators. First, the frequeny of interation between voter and politiian diretly affets the latter s future payoffs insofar as repeated interation an reate reputational osts if promises are not kept (f. Alesina 1987). Seond, in order to redibly ommit to retrospetive santions voters must find ways to overome olletive ation problems. As in other areas, issues suh as group size (Olson 1965), density (Bates 1981) and identity or ethniity (Fearon 1999b) might all affet the redibility of voters threats to at olletively. As noted in the introdution, many authors argue that voters are severely onstrained in their ability to redibly threaten to punish ex post (Fearon 1999a; Alesina 1987; Besley and Coate 1996). Third, the importane of pork for voters (either individually or as a group) an be signifiant. In keeping with a long aademi tradition, voters are ommonly pereived as aruing both ideologial and material or partiularisti benefits from governmental poliies (Cox and MCubbins 1986; Lindbek and Weibull 1987; Dixit and Londregan 1996). The relative weights of these two omponents will likely affet the extent to whih voters an redibly threaten to santion the failure of politiians to deliver pork. Many authors show how inome levels affet the relative importane of pork a bag of rie matters more for poor voters than for rih ideologues (Dixit and Londregan 1996; Calvo and Murillo 2004; Keefer 2003). Similarly, Stokes (2005) argues that with regard to ore supporters ideology typially dominates and attention to material benefits is minimal. As a result, threats from the ore to swith to another politiian in response to promises of inreased pork are inredible. Finally, Wanthekon (2003) finds that pork matters more than ideology or other publi goods for voters with only provinial interests. The relative saliene of pork is losely related to a fourth potential fator the existene of attrative seond-best alternatives. In general, when failing to arry out threatened santions involves signifiant opportunity osts, suh threats are inreasingly redible. Diaz-Cayeros, Magaloni and Weingast (2003a) demonstrate this intuition in analyzing the support of the PRI in the various states of Mexio. Those 13

14 regions with lose eonomi ties to the United States had a redible exit option from PRI support and the resulting pork. The previous examples identify a few of the fators that might affet voters memory or ability to redibly threaten santions ex post. Eah is theoretially plausible and is therefore a matter of empirial analysis. We ondut several suh tests in Setion 4. We now wish to add ounty-level heterogeneity in this harateristi into our theoretial model. One the eletion has passed the only reason to make the transfers is the need to retain the redibility of promises regarding suh transfers for future eletions. One in offie (period 2), the politiian has a net present valuation for future offie equal to Ω 2 < Ω1, and a disount fator δ over the interval until the next eletion. Reputation is ahieved by making transfers as promised subsequent to the eletion, and the benefit of maintaining a reputation for truthfulness is that pork promises will remain effetive in future eletions. We write the probability effet of period 2 transfers on future voter behavior as h t ). Voters in a given ounty remember only a share α of the transfer made in period 2, and so h t ) = f ( αt ). This ( ( implies that in the absene of any transfers that h 0) = f (0) (meaning that politial preferenes are ( stationary between eletions), and that h' = α f '. In a ounty with a perfet memory α = 1 and h (.) = f (.). Counties with no memories have h (.) = h (0) t. 12 The politiian s utility-maximizing transfer in the seond period is the largest transfer that is redible in the first period. This is given by: ωn W2 ωn (2) f ( t ) = W1. α [ δω + ( K t )] α Ω + ( K t 2 1 ) 12 As indiated previously, we an also think of α as representing the ability of politial organizers in ounty to overome the olletive ation problem inherent in threatening to hold the behavior of politiians against them in upoming eletions. One the next eletion arrives voters may want to revert to their preferene-based voting outome h (0), but an organized ounty has α 1, and is thus able to redibly threaten politiians with reperussions of failing to pay pork. 14

15 Define * t as the optimal transfer to a ounty based on the eletion result in the full-redibility equilibrium given by (1), and the optimal transfers given by (2) as tˆ. This says that * ˆ W2 t t (or > W1 ), α meaning that the redible transfer without a ommitment devie is smaller. This is true for several reasons. The first is that disounting pushes up the opportunity osts of diverting funds ( δ < 1). 13 Seondly, Ω 2 Ω 1, sine the former is a NPV of future offie with one fewer potential terms in offie than the latter. These relationships are ommon to many modeling environments, and show redibility to be dereasing in the presene of impatient politiians with a probability of overthrow or politiians under term limits (in the absene of redible parties) (Keefer 2003; Wanthekon 2003). These two soures of heterogeneity pertain to the politiian, and so their influene on the differene between tˆ and onstant within a given eletion. * t will be Only α reates ounty-level differentiation between the ex-post optimal transfers with and without a ommitment devie. Under limited redibility, variation in ounties ability to threaten the politiian with reperussions means that even ounties reeiving non-zero transfers may have different slopes in the voting prodution funtion. Those ounties with low α reeive a lower transfer (beause f has to be driven up by dereasing transfers). This smaller transfer relative to the full-redibility transfer auses the marginal returns to pork to be inreasing in the redibility onstraint. Beause business is responding to the inverted funtion 1 h, this means that eteris paribus dπ dε, the response of the business outome to the surprise vote share, will be smaller in ounties that make less redible threats to hold politiians aountable in future eletions. This ounty-level redibility onstraint also has an extensive margin, beause there may exist ounties in whih f '(.) t= 0 W1 but in whih α f '(.) t= 0 < W2, and so while these ounties would have 13 The politiian does not disount over the disontinuous period from before to after the eletion, but having won the eletion is now onsidering the multi-year period between eletion yles when deiding whether to make transfers. 15

16 reeived positive transfers under full redibility, they reeive none when redibility is limited. Figure 4 shows the differential slope that will be observed aross ounties depending on the degree to whih the politiian is redibility onstrained in making promises to them. This onstraint differs fundamentally from the yield onstraint onerning the basi effiay of pork whih defines (1), and for suffiiently dπ small α we see = 0 dε under limited redibility beause no promises are redible. This theoretial model generates several quite general propositions: dπ P1. A finding that = 0 dε indiates that no pork transfers are redible. dπ P2. A finding that = k > 0 indiates that the politiian has a ommitment devie, making dε ounty-level harateristis unimportant to the slope in equilibrium. dπ P3. If > 0 dε in some ounties and varies aross, then the politiian has limited redibility. As a result, harateristis whih are assoiated with a steeper slope of the response desribe the attributes that make an eletoral unit redible. We now proeed to test these propositions with the Ugandan data. 3. Data. There are four distint soures of data brought together in this paper. The first is data from the lient reords of FINCA/Uganda, a major mirofinane ageny. From these reords, we have the following data: loans are for a 16-week term, to be repaid weekly at an 87% effetive annualized rate, and flutuations in borrowing an be seen as lients subjetive short-term foreasts of the business environment. We see savings in FINCA; these are physially held at Standard Chartered Bank in group aounts (all FINCA lending is done in groups of 30) and are not demand deposits. As these savings are 16

17 seized by FINCA in the event of failure to repay and are usually only available at the end of the borrowing yle, they are one of the less liquid forms of saving available to these lients. Our seond soure of data is ounty-level eletion outomes, inluding votes ast for eah of the andidates and the total vote. We use the perentage of the vote in eah ounty ast for the inumbent Movement andiday of Yoweri Museveni and the total perentage ast for the opposition andidates ombined (Kizza Besigye, along with Moody Awori, Chapaa Karuhanga, Kibirige Mayanja, and Franis Bwengye). We have aounting data for FINCA lients loated in 71 out of Uganda s 214 ounties. Third, we have a newspaper survey whih overs the months of November 2000 through July Uganda has two major newspapers the New Vision, owned by the Movement government, and the Monitor, owned by the Aga Khan s Nation Media Group. Only events whih ould be loated at a speifi plae and time were inluded in order to fous the identifiation. Summary statistis of the events taken from the newspaper survey an be found in Tables 1 & 2, and a graphial summary of the data on pork promises and threats is found in Figure 5. We see sharply divergent patterns for pork promises made by Museveni and the opposition politiians; while the Movement made suh promises throughout the politial spetrum, the opposition made suh promises only to their own base. The final soure of data is a pair of opinion polls released by the New Vision on February 3 rd and February 28 th These polls report outomes only at the distrit level (distrits are omprised of several ounties), and results exist for only 19 of Uganda s 56 distrits. FINCA groups exist in 22 distrits, of whih only 11 overlap with those overed by the opinion polls. From these we have onstruted a ounty-level predition of voting based on that in found in the distrit as a whole, or that in the losest distrit for whih data exist. While this measure is very imperfet, it also represents the best information available to voters in the run-up to the eletion as to likely voting outomes; no other polls existed, and it is not unreasonable to think that voters were using polls to form expetations in a similar manner. 17

18 3.1. Changes in the Outomes. We use two outomes, eah observed at the individual level for 21,050 small business-women during the months around the eletion (April 2000-July 2001). The first is the sale of business-lending; sine all loans are on a four-month yle this gives us a forward-looking measure of business expetations. As expetations of business opportunities expand, we expet to see business lending inrease, giving us a metri of the investment opportunities as pereived by the businesswomen of FINCA for the four months following the date of the loan. Our seond outome is the savings whih lients hoose to leave with FINCA; these are partially ollateralized savings whih are not easily withdrawn during the four-month yle and thus represent the businessperson s willingness to take liquidity out of the mattress and plae it beyond their reah for a four-month period. Loan quantities an be linked to savings; we often see these two move together but our results show that in irumstanes where itizens are diretly threatened by politiians redit demand inreases and savings falls, meaning that households are trying to provide themselves with the maximum short-term liquidity. In general, then, the loan volume is a metri of business expetations and the savings volume gives a measure of pereptions of short-term finanial seurity. A shok that is positive both in terms of business expetations and short-term seurity will see both redit demand and willingness to save inrease. The FINCA data is a rolling panel in whih eah individual features only ever four months (or more) as they reapitalize their loans. For this reason, it is diffiult to interpret disontinuous hanges between time periods beause they onsist of different individuals. The temporal sorting proess is non-random, a fat whih is easily verified by piking plaebo disontinuity dates, of whih nearly 50% are signifiant. In order to minimize differenes between ohorts we differene outomes. The dependent variable used throughout the study is & π t = π it π ; individuals for whom only one observation are available were i dropped from the study. We thus analyze how explanatory variables relate to deviations from individual mean outomes for eah individual ontrolling for the strongly inreasing and onave pattern that both loans and savings display. 18

19 4. Empiris. 4.1 The mean response to the surprise vote share. Beause our dependent variables are defined in differenes from individual means, the random effets estimator is idential to pooled OLS. The use of fixed effets in this ontext would only demean the RHS, and hene the FE estimate is very similar to POLS in all speifiations. We therefore use pooled OLS with standard errors lustered at the ounty level to estimate & π igt 8 = β0 + δ t + β1χigt + s= 1 β S + γε + u s st t igt for the differened outome π& & igt for individual i in group g and ounty at time t. χ igt is a vetor of ontrols (loan yle number and yle squared) that vary at the individual level over time, and S t is a set of eight dummies equal to one in the month after new politial events are revealed to have ourred in a ounty. Given the time dummies, these indiator variables measure how outomes in ounties in the month after eah kind of shok has ourred differ from the average differened outomes in that month. ε t equals the surprise vote share in a ounty in the month after the eletion and zero all other times, so γ measures the extent to whih ounty-speifi shoks to eonomi outomes in the month after the eletion are orrelated with the surprise vote share. The additional eight ovariates, taken from the newspaper survey, measure the following: 1. Ats of violene (any attaks, beatings, shootings, et. whih are politially motivated). 2. Threats against the opposition, whih inludes the following ategories from the newspaper survey: Government threats to opposition andidates, Government Arrests opposing supporters, Government threats to opposing offiials, Government supporters threaten opposition, and Physial attak on opposition offiials/supporters. 3. Threats against the Movement, inludes Opponent supporters threaten government offiials, Citizens threaten government, and Physial attak on government offiial/supporters. 4. Threats against itizens, inludes Opponent supporters threaten itizens, Government/polie attak/arrest itizens, Citizens struggle with itizens, and Inreased seurity (inluded beause it was an implied threat, and a response to violene). 5. Eletion results ontested. 6. Eletion was lose: This is a dummy whih swithes on in the month after the eletion that equals one if the poll was between 45% and 55%; this variable aptures whether the response to the eletion is a funtion of loal-level unertainty over the outome. 19

20 7. Movement promise of pork made to a ounty by a Movement politiian. 8. Opposition promise of pork made to a ounty by any opposition politiian. Table 3 gives these estimates (non-politial variables are suppressed in the table). Both threats and promises have measurable effets on borrower behavior. There is some evidene that threats against the Movement drive up loan sizes, but threats against itizens have a stronger effet; here we see inreased borrowing and sharply dereased savings as households maximize their short-term liquidity. The response to a pork promise is positive, signifiant, and resembles the effets of the surprise vote share in signifiane although the oeffiients indiate that the surprise vote share would have to be on the order of 20-40% to rival the quantity effets of a diret promise of pork. Intriguingly, the effets of pork promises by opposition politiians seem to be larger in both magnitude and signifiane than those made by the Movement. It seems diffiult to argue that they would have been more redible, and so it may be instead that they were more unexpeted, ausing a bigger disontinuous shift. There is strong support here for the idea that hanges in business outomes are related to the surprise vote share. Figure 6 plots the response to the surprise vote share as we inrease the length of the post-eletion window used to measure the shok. Supporting the theory that this informational shok is both immediate and transitory we see the largest effet using a two-week window and find that by the time the window has expanded to two months the measured effets beome insignifiant. Sine we are seeing different lients show up to take loans over this time interval and that those oming to FINCA between two and four months after the eletion are taking their first post-eletion loan, the impliation is that they have already ome up with other ways of responding to the new information provided by the eletion and hene do not respond differently Heterogeneity in the response to surprise vote share. Now we wish to measure the onditional distribution of π ε, whih an be aomplished through the use of interations. For some fixed (and demeaned) ounty-level harateristis X, we an use OLS 20

21 8 to estimate π& & igt = β0 + δ t + β1χigt + β ss st + β 2 X + γε t + η( X * ε t ) + uigt, where s= 1 η measures the slope of π aross d dε X. Assuming that the responsiveness of π to pork does not differ aross ounties, differenes in the slope of π aross d dε X give the determinants of α. Table 4 shows that strong differentiation in slopes exists aross X. The ounties that show a strong response to the surprise vote share are eduated, sparsely populated, and female. This implies that eduated rural ounties with a high perentage of females have the values of α losest to 1, and so have the punishment apabilities whih make pork redible. Conversely, dense, poorly eduated areas with a high onentration of males show the smallest response to the surprise vote share whih implies that politial redistribution onditional on voting behavior is least redible in these areas. Given the strongly ethni nature of many Afrian eletoral ontests, another interesting dimension along whih to examine the onditional distribution of dπ dε is ethniity. Using the language spoken in group meetings as the best proxy we have for ethniity we again find strong differentiation in slopes. Table 5 presents the results of interations performed in the same way as desribed above where we dummy out seven language ategories, leaving Kiswahili-speakers as the omitted ategory. Swahili is the language of trade in East Afria but is not widely spoken in Uganda, and so these groups are likely to be ethnially heterogeneous and to be trading over larger distanes than other groups. We find that neither English-speaking groups (whih may be the most highly eduated) nor the major ethniities of southern Uganda (the Baganda and Basoga) have slopes whih differ from Kiswahili speakers. However, two kinds of ethniities have sharply divergent responses. On one hand ethniities from the war-torn north of Uganda, where Museveni has struggled to establish authority, show muh lower response to the surprise vote share. On the other hand the four small, non-northern ethniities show muh larger responses. The impliation is that Museveni is inredible in the North but that he is 21

22 strongly redible among small ethni groups. The former onlusion makes sense to anyone familiar with the past deade in Uganda and the latter may be evidene for the fat that a small ethni group an more easily overome the olletive ation problem inherent in trying to redibly threaten to hold politiians aountable for urrent pork distribution in future eletions. We an also use individual-level harateristis to measure interation effets: π& & igt δ t + γε t + β1χigt + β s st + β 2 X ig + η( X ig * ε t s= 1 = β S ) + u igt. These results are given in Table 6. Despite the strong differentiation in effets at the ounty and ethniity level, we see very little evidene that individual harateristis matter for the response to the surprise vote share. The only signifiant harateristi is the response to the survey question (asked several months before the eletion) Do you think that your loal eonomi limate is worsening (-1), unhanged (0), or improving (1)? This variable is itself measuring a loality effet rather than an individual effet (unless we onsider the answer to ome from idiosynrati optimism). Sine we have modeled heterogeneity in the response to the surprise vote share as arising from loal-level redibility, it is unlear what individual heterogeneity in the response would mean, unless it arose from harateristis whih make a given individual more likely to reeive pork diretly. The fat that we fail to find idiosynrati differenes in slopes is seen as onfirmation of the fat that it is indeed heterogeneity at the level of α (and not π igt or X ig ) that determines the slope of the response funtion. Using the mahinery developed up to this point, we an also test several of the fundamental hypotheses in the literature on eletions Core versus Swing voters. A major theoretial debate has taken plae over whether politiians will target pork to ore or swing onstituenies. Cox and MCubbins (1986) assume risk-averse inumbents, onluding that they will steer transfers disproportionately to their ore supporters in order to maintain oalitional stability. Dixit and Londregan (1996), on the other hand, argue that the inumbent s ore supporters will only benefit when 22

23 the inumbent has an organizational advantage in direting favors to the ore. Otherwise, welfare transfers will be direted to voters whose value for material utility is high relative to their ideologial persuasions. By definition these voters are typially swing voters. Lindbek and Weibull (1987) reah a similar onlusion. Empirial support for these formal models is mixed. Case (2001), Miguel (2003), Levitt and Snyder (1995), Stokes (2005), and Magaloni, Estévez and Diaz-Cayeros (2000) find evidene supporting transfers to the ore. The evidene of Dahlberg and Johansson (2002), Kehmani (2003) and Kasara (2005), all support the Dixit-Londregan swing model. The test suggested here provides a new way of addressing this question. In terms of our model, this is a debate over the responsiveness of voters to pork, and hene relates to f '( Ft ˆ ). On both the intensive and extensive margins, we will see less business response if politiians are making fewer redible promises of pork to a given ounty, and so we an ask whether the slope dπ dε differs aross ore Movement, ore opposition, and swing ounties. We define dummies that define ore Movement and swing ounties (Movement vote perentages of >60%, and >40% and <60%, respetively), and interat these dummies with the surprise vote share variable, using the same speifiation as above. The last two rows of Table 7.1 shows these interations. We find that Movement ore ounties show a signifiantly larger response in loans to a surprise vote share of a given size than opposition ore ounties while swing ounties do not respond differently. The savings response does not seem to be different aross ounties, whih is reinfored by Table 7.2 whih partitions the ounties aording to their status and runs three separate sets of regressions. In all three groups we see a savings response whih is positive but marginally signifiant with very similar point estimates. The loan response is highly divergent, however, with a strongly signifiant positive response in Movement ore ounties and highly insignifiant responses in the other groups. The impliation is that while the seurity effets of a positive vote share (measured by savings) are not divergent aross ore and swing ounties, the disontinuous improvement in business opportunities engendered by redible pork is stritly limited to Movement ore ounties. 23