ABSTRACT Title of Dissertation: POLITICAL INFLUENCES ON MONETARY AND FISCAL POLICY Marcela Eslava, Doctor of Philosophy, 2004 Dissertation directed by: Professor Allan Drazen Department of Economics This dissertation studies the influence of some political factors on economic policy. Chapter 1 studies the choices of central bankers when: (i) central bankers are government appointees; (ii) monetary policy is decided by a committee. In this framework each central banker faces incentives to protect the Central Bank?s reputation and incentives to show loyalty to the government in order to be reap- pointed. I show that collective policy-making can be better than having a single central banker at achieving low inflation and isolating policy from government pressures, because the committee can reduce the relative value central bankers assign to reappointment. For this to hold, the committee must be small so that each member values the impact of his vote on policy, while the rate of turnover of members must be small so that the risk of being replaced for deviating from the government?s preferred policy is small. Chapter 2 uses the view that politicians favor the interests of specificgroups of voters to explain political budget cycles in a model with rational voters. Vot- ers use past fiscal policy to learn information about which types of spending the incumbent is likely to favor if re-elected. The result is a pre-electoral shift of gov- ernment resources from non-targeted types of expenditures toward goods specific groups of voters care more about. Pre-electoral transfers are mostly directed to undecided groups of voters. Even though voters are rational and predict this be- havior, they respond to electoral manipulation, since theycannot observe whether they are being targeted because they are crucial to the incumbent?s re-election or because they are genuinely liked. In chapter 3, I use data on government expenditures and electoral outcomes in Colombia to analyze both voting behavior and the pre-electoral dynamics of government spending. I suggest a correspondence between fiscal data and the conceptual division of expenditures into targeted and non-targeted spending. I find that targeted spending grows and non-targeted spending contracts in the year leading to an election. Consistent with this result, I find that Colombian voters reward pre-election increases in targeted spending, but punish incumbents who run high deficits when the election approaches. POLITICAL INFLUENCES ON MONETARY AND FISCAL POLICY by Marcela Eslava Dissertation submitted to the Faculty of the Graduate School of the University of Maryland, College Park in partial fulfillment oftherequirementsforthedegreeof Doctor of Philosophy 2004 Advisory Committee: Professor Allan Drazen, Chairman Professor John Haltiwanger Professor Deborah Minehart Professor Irwin Morris Professor John Shea c? Copyright by Marcela Eslava 2004 DEDICATION AMarc A la memoria de Gerardo Eslava y Olga Cobos ii ACKNOWLEDGEMENTS I want to express my gratitude to a very special group of people. First, my family; without their support, care and encouragement, this project would have sunk at some rough spots. Allan Drazen helped shape this research from a very early stage, and provided enlightening discussions. He and John Haltiwanger have been my mentors and a great model to look up to. Working with them has been a fascinat- ing experience. John Shea very generously dedicated many hours to reading earlier drafts of these essays and listening to my presentations, and provided a myriad of extremely useful comments. Deborah Mine- hart was also a dedicated and sharp reader, whose suggestions helped keep these essays disciplined. Irwin Morris enthusiastically joined the project when it was nearing an end, and provided a very useful fresh look. My work also owes a great deal to Juan Carlos Echeverry, who first sent me down the path of research. He and Mauricio C?rdenas have been an inspiring example. iii These essays have greatly benefited from the comments of all mem- bers of my advisory committee, as well as seminar participants at the University of Maryland, the 8th meeting of LACEA, the 4th and 5th meetings of the Political Economy Network of LACEA, the Colom- bian Ministry of Finance and National Planning Department, and the 2004 meeting of the Public Choice Society. The comments of two discussants, Fabio S?nchez and Eduardo Ganapolsky, are specially ac- knowledged. The excellent assistance of Lina Ladino and the advise of Gabriel Piraquive were very helpful in putting together part of the data used for Chapter 3. I thank the financial support of the Colom- bian Ministry of Finance, and the logistical support of the National Planning Department. The views expressed here may not reflect those of these institutions. iv TABLE OF CONTENTS List of Tables ix List of Figures xi I Central Bank Institutions and the Government?s In- fluence Over Monetary Policy 1 1 Central Banks as Government Appointed Committees 2 1.1 Motivation............................... 2 1.2 A little perspective: literature on government appointments and committeecentralbanks....................... 6 1.3 Model:Generalsetting........................ 10 1.3.1 Thepreferencesofcentralbankers ............. 10 1.3.2 Timingandinformation ................... 16 1.3.3 Solvingtheproblemofadovecentralbanker........ 17 1.3.4 Expectationsandequilibriumchoices............ 22 1.4 TheinstitutionaldimensionI:collectivedecisionmaking ..... 24 1.4.1 Singlecentralbanker..................... 25 1.4.2 AcommitteeCentralBank.................. 27 v 1.5 TheinstitutionaldimensionII:governmentappointments..... 36 1.5.1 Government appointments and the case of a single central banker............................. 39 1.5.2 Collective policy-making and government appointments . . 41 1.6 Concludingremarks.......................... 49 II Political Budget Cycles 51 2 Political Budget Cycles When Politicians Have Favorites 52 2.1 Introductionandliteraturereview.................. 52 2.2 Athree-periodmodel......................... 58 2.2.1 Voters............................. 59 2.2.2 Politicians........................... 60 2.2.3 Votingbehaviorandelectionoutcomes........... 65 2.2.4 Equilibrium .......................... 70 2.2.5 Asymmetric information about the electoral environment . 74 2.3 Ageneralization:lettingpoliticianshavefavorites......... 77 2.3.1 Voters............................. 79 2.3.2 Theincumbent?sproblem .................. 81 2.3.3 Voters? expectations and the impact of fiscal policy on vote shares ............................. 91 2.3.4 Anexample.......................... 93 2.4 Concludingremarks..........................102 3 On Political Budget Cycles and Voters as Fiscal Conservatives: Evidence From the Colombian Experience 103 vi 3.1 Introduction..............................103 3.2 Adiscussionofpreviousliterature..................105 3.3 TheColombiancase .........................109 3.4 Theelectoraldynamicsofgovernmentspending ..........112 3.4.1 Data..............................115 3.4.2 Regression specification....................122 3.4.3 Regressionresults.......................125 3.5 Dovotersrewardtargetedspending? ................132 3.5.1 Data..............................133 3.5.2 The e?ect of fiscalpolicyonvoteshares ..........135 3.6 The e?ectofideologicalpolarization:PBCsinswingcities ....139 3.7 Thepersistenceofexpenditurechoices ...............146 3.8 Electoral cycles in the central government?s budget . . ......151 3.9 Concludingremarks..........................154 A Technical Appendix to Chapter 1 156 A.1 Proofofproposition1 ........................156 A.2 Derivationofequation(1.11).....................157 A.3 Explicitformofcondition(1.16)...................158 B Appendix to Chapter 2 159 B.1 Proof of proposition 9. ........................159 C Appendix to Chapter 3 161 C.1 Sourcesanddetailsondata .....................161 vii C.1.1 Dataforlocallevelestimations ...............161 C.1.2 Datafornationallevelestimations .............162 C.2 Matchingpoliticalmovementsandthemainparties........163 C.3 Electiondates.............................165 C.3.1 Localelections ........................165 C.3.2 Nationalelections.......................165 Bibliography 167 viii LIST OF TABLES 1.1 Casewithnogovernment....................... 35 3.1 Compositionofspending.......................116 3.2 Summary statistics for di?erenttypesofexpenditure .......119 3.3 Listofcontrolvariables........................122 3.4 E?ectofelectionsonthecompositionoftotalexpenditures....127 3.5 E?ect of elections on di?erenttypesofexpenditure ........131 3.6 Summarystatisticsforelectionoutcomes..............135 3.7 E?ect of fiscalperformanceonvoteshares .............138 3.8 Summary statistics for fittedvoteshares ..............141 3.9 Samplecorrelationsforswingdistrictdummies...........142 3.10 E?ect of elections on government expenditure: swing versus non- swingdistricts(time-invariantdummies) ..............144 3.11 E?ect of elections on government expenditure: swing versus non- swingdistricts(time-varyingdummies)...............145 3.12 Persistence of the composition of government expenditure (basic instruments)..............................149 3.13 Persistence of the composition of government expenditure (all in- struments)...............................150 ix 3.14 E?ect of elections on di?erent types of expenditure. Central Gov- ernment.................................153 C.1 Partycorrespondences ........................164 C.2 DatesofelectionsforMayors.....................166 C.3 DatesofelectionsforPresident ...................166 x LIST OF FIGURES 1.1 Timingofevents ........................... 17 1.2 Reputation-buildingincentive,nogovernment ........... 34 1.3 Reappointment incentive if vote a?ectspolicy ........... 44 1.4 Overallreappointmentincentive................... 47 1.5 Benefitofchoosingv i =0 ...................... 47 2.1 Case 1: ? h 1 (g t+1 ) close to ? h2 (g t+1 ) ................. 89 2.2 Case 2: ? h 1 (g t+1 ) far from ? h 2 (g t+1 ) ................ 89 2.3 f h (?) for di?erent values of ?? h .................... 94 2.4 ? 0 (g h 1 t ) and ? 0 (g h 2 t ) .......................... 99 2.5 Incumbent?s firstorderconditions..................100 2.6 g h t (? h I ) and g h t+1 (? h I ) ..........................101 3.1 Evolutionofspending:broadcategories...............118 3.2 Evolutionofspending:investmentcategories............120 3.3 Evolution of financialindicators...................123 3.4 Evolution of fiscaldependence....................123 xi Part I Central Bank Institutions and the Government?s Influence Over Monetary Policy 1 Chapter 1 Central Banks as Government Appointed Committees 1.1 Motivation The question of what is the optimal institutional design of an independent Cen- tral Bank (CB) has been of great interests to legislators, central bankers, and academics alike. One of the sticky points is how much independence should the CB be given. Although the importance of CB independence is widely recognized, there is also ample agreement around the idea that the CB should not be com- pletely isolated from the government. This is in part due to the need for harmony between monetary and fiscal policy, but also because the participation of publicly elected o?cials in all policy decisions is at the core of democratic systems. The dilemma is how to maintain the desired link with the government, while giving the CB enough independence to alleviate ties between monetary policy and the electoral cycle, and to reduce inflation bias. In practice, such dilemma is frequently addressed by placing the design of monetary policy in the hands of a committee of central bankers, and giving the government some power over the appointment of its members. The idea is that, while these appointment powers create a tie between the government and 2 the CB, the larger set of decision-making mechanisms and appointment rules available under a committee arrangement can be exploited to make policy more independent from government?s pressures than it would be under a single central banker. In other words, for any given position of each individual member, the mechanisms through which positions are aggregated can be optimally designed to reduce the influence of the government. Although this argument points to certainly important advantages of commit- tee designs, it fails to recognize that central bankers embedded in this type of setting may in fact not behave as they would in the absence of government ap- pointments, or within monolithic institutions. That is, the positions of individual members are not ?given?. This paper brings into the picture the potential e?ects of collective decision-making and government appointments on the choices of each individual central banker. Accounting for this dimension makes clear that not under all circumstances a committee structure indeed helps in isolating mone- tary policy from the electoral cycle and reducing inflation bias. In particular, the size of the committee and the fraction of members over whose appointment the government has influence are key determinants of the e?ects of collective decision- making. My results contribute to the debate on the optimal institutional design of the CB, shedding light on the question of how to obtain the benefits of CB independence without fully isolating the CB from the government. The introduction of collective decision-making and government appointment powers can a?ect the incentives faced by central bankers in ways that are par- ticularly acute when informational asymmetries give rise to reputation concerns. As previous literature has pointed out, central bankers may want to build a rep- utation for the CB being tough on inflation. But this is not the only source for 3 reputation building incentives when appointments to the CB are a privilege of the government. In fact, a central banker may care about the perception the gov- ernment has about him (how close are his preferences to those of the government, how loyal to the government he is) because such perception a?ects his chances of being reappointed 1 . Hence, a central banker may want to build a reputation of being close to the government. Moreover, how these reputation building incentives play into acentral banker?s decisions depends crucially on whether decisions are monolithic or collective. In particular, reappointment probabilities are a?ected by the government?s percep- tions about each individual central banker, while inflation expectations are af- fected by the public?s perception about the CB as a whole. In a monolithic environment, the central banker and the CB are the same entity, so there is no di?erence between the CB reputation and that of the policy-maker. But under a committee there is a dichotomy between the individual reputation of each member and the collective reputation of the CB. Furthermore, since the vote of a central banker may not be reflected in the actual policy, each member knows that his vote will only have a limited e?ect on the collective reputation of the CB. Hence, under committee decisions the incentive to show closeness to the government can override the incentive to build a reputation of toughness for the CB, as well as other considerations. If this is the case, the idea that a committee design limits the influence of the government on monetary policy may be wrong. Motivated by these considerations, I study how the reputation building in- 1 As an example, Ersenkal et al. (1985), find that monetary base in the U.S. expands faster in the months that precede the decision of whether a Federal Reserve Board chairman will be reappointed. 4 centives faced by central bankers are a?ected by the government?s appointment powers and by collective decision-making. I frame my analysis in a two-period model that falls in the tradition of Barro and Gordon (1983), in which I introduce the possibility of endogenous reappointments of CB members. In this setting I address the tension between the CB collective reputation and each member?s in- dividual reputation, as discussed above. While a central banker?s concern about the reputation of the CB can reduce his temptation to inflate, his concern to show loyalty to the government in order to be reappointed creates a channel for the government to influence policy. I show that, compared to a single central banker, a committee design has the potential to both increase the importance of building a reputation for the CB being tough on inflation, and reduce the importance of reappointment consider- ations. The realization of this potential, however, depends crucially on specific features of the committee design, in particular its size and the fraction of mem- bers the government can replace from one period to the next. An increase in the fraction of members to be replaced reduces the relevance of past inflation for future expectations of inflation, diminishing the ability of reputation consider- ations to reduce inflation bias. It also increases the risk that a central banker will be removed from o?ce if believed to be of a di?erent type than the govern- ment, increasing the incentives to follow the government?s preferred policy. On the other hand, an increase in the size of the committee reduces the ability any single central banker has to a?ect actual policy, thus increasing the importance central bankers give to influencing reappointment rather than choosing policy. The paper is divided in six sections , including this introduction. In section 1.2, I review some relevant literature, and discuss the contributions of this paper. 5 Section 1.3 discusses the general setting of my model, and solves the problem of a central banker. Section 1.4 compares how that solution di?ers between the cases of a single central banker and a board. Section 1.5 introduces government appointments. Finally, section 1.6 concludes. 1.2 A little perspective: literature on govern- ment appointments and committee central banks The e?ects of government appointments on monetary policy, as chosen by a com- mittee, have been previously studied by Lohman (1997) and Waller (1989). These papers extend Alesina?s (1987) model of partisan cycles in monetary policy and output, to account for committee decision-making and government appointments to the CB. A central assumption is that each government can change some mem- bers of the CB during its term, and it appoints central bankers who belong to its own party. The only source of uncertainty is which party will win the following elections; the types of central bankers are not private information, and as a result each member of the committee always votes for the policy preferred by his politi- cal principal. The basic result of this literature is that a well designed committee can reduce inflation cycles because it renders more stable CB preferences than a single central banker arrangement 2 . This implies less uncertainty about future decisions of the central bank, and therefore ends up reducing output cycles as 2 This literature does not have much to say about inflation bias: since there is perfect information about the preferences of all policy-makers, inflation always reflects the ?median type? in the committee. 6 well. Waller (1989) shows that lengthening the terms of central bankers (reduc- ing the fraction of members changed at each point) further reduces those cyclical variations, because it implies even less uncertainty about the future preferences of the CB. The decentralization of appointments is shown to have a similar e?ect by Lohman (1997). The results in these papers support the informal arguments that, in many countries, led to the adoption of boards to direct the CB. Both o?cial statements and other accounts about the reasons that led to this type of institutional design in many countries argue that the preferences of a committee CB will evolve more slowly than those of its single-member counterpart, and will therefore be less subject to the influence of the political cycle 3 . The absence of informational asymmetries in this literature implies that a central banker always behaves like his political principal. As a result, the only di?erence with the case where the government chooses monetary policy is the committee nature of policy-making. In other words, CB independence plays no role in these models, and the only e?ect of a committee structure is to aggregate the otherwise unchanged votes of central bankers. My approach in this paper is fundamentally di?erent, although not necessarily conflicting: I focus on how a committee a?ects the decisions of central bankers, rather than on how those 3 For instance, the committee design of the Colombian Central Bank is explained in the Bank?s Website as follows : ?This system guarantees continuity in Bank policy while safe- guarding it from the influences of political change, thus ensuring planning more in view of the long-term and garnering greater credibility with the public? (El Nuevo ordenamiento del Banco ysuJuntaDirectiva, in www.banrep.gov.co). On other front, Waller (1989) documents how several historians of the U.S. Federal Reserve ?argue that the board structure of the Federal Re- serve Board of Governors was, in fact, chosen specifically to minimize the influence of partisan politics on the setting of monetary policy? (pp 422-423) 7 decisions are aggregated. Moreover, there is a rationale for CB independence in my model, since the government is assumed to be immune to the reputation concerns that restrain the inflationary temptations of central bankers. Another strand of literature (Faust, 1996, Dal B?, 2002) documents the ability of a committee structure to reduce inflation bias, in the absence of reputation concerns. These papers take advantage of the fact that a committee provides flexibility to aggregate the positions of its members in di?erent ways. The basic starting point is that the inflation rate preferred by the median voter is too high, in the sense of imposing inflation costs without generating benefits from surprise inflation. They show that an optimal decision mechanism can be designed, such that the monetary policy chosen by a board is less subject to inflation bias than the one preferred by the median voter. In Faust?s model, monetary policy is the result of bargaining between members of the CB. Inflation bias is reduced if the representation of anti-inflation groups is disproportionately large, relative to their representation in society. Meanwhile, in Dal B??s model central bankers vote over policy choices. He shows that a less inflationary balance of preferences canbeachievedifanoptimalsupermajorityruleisusedtochoosepolicy 4 .The 4 Dal B??s discussion relies on the assumption that CB committees are large enough to be representative of the distribution of types in society. Given my results this assumption not only is unrealistic, but it also ignores that a large committee creates perverse incentives for central bankers. Moreover, in the supermajority rule he proposes, central bankers vote over status- quo optimal inflation and a ?reform inflation? that is chosen by the majority. The majority is supposed to choose a reform inflation level di?erent from the one preferred by the median voter, realizing that only in this way will a supermajority be obtained in the second round. This assumption requires a large amount of coordination among voters, that I find di?cult to justify if the committee, as assumed in the paper, is large. 8 supermajority rule ensures that the pivotal voter is more conservative than the median voter. My results di?er from these papers in that, in my model, an appropriately designed board can reduce the inflationbiasevenwhenthepro- inflation types are in command of the CB. This arises because, in my model, reputation concerns a?ect the choices of these ?weaker? types, and a committee structure can make reputation more relevant. Finally, the e?ects of a committee design on the reputation building incentives of central bankers are addressed by A. Sibert (1999). In her model, as is the case in this paper, the fact that a central banker?s vote may not be reflected in the policy outcome changes the incentives to build a reputation. However, she abstracts from the mechanisms for appointing central bankers which, as this paper shows, are a key determinant of the importance central bankers give to reputation-building considerations. Moreover, Sibert?s model assumes that the public obtains and uses information about individual votes in the CB, while in this paper the idea that individual votes are observed only imperfectly is pivotal to the concept of ?collective reputation?. My contention is that people may form expectations based on the actual policy even if individual voting records are available, due to reasons such as the lower cost of observing policy compared to investigating individual votes. Sibert?s paper does analyze the case in which voting records are not published, but my results di?er from hers even in this case, given my consideration of the reappointment process. My idea of collective reputation is closer to that proposed by Tirole (1996), for whom the reputation of an organization di?ers from the reputation of each member because the output of the latter is observed only imperfectly. 9 1.3 Model: General setting The model I use follows the tradition of the Barro-Gordon model (1983), which is probably the most standard framework to model the choices of the Central Bank. My analysis relies on previous models of reputation building by the monetary authority (Backus and Dri?ll, 1985; Barro, 1986). I begin by studying the choice of monetary policy without imposing a specific institutional structure, and derive a general solution to the problem of a central banker. Even without further institutional detail, my general approach di?ers from the traditional setting in that, as in many CB?s around the world, a central banker?s term can be extended. That is, a central banker can be reappointed. I will later I specialize the general solution obtained in this section, analyzing the form it takes under di?erent specifications of the institutional setting: board vs single central bankers, and government appointments vs. random appointments. 1.3.1 The preferences of central bankers Consider a two-period economy with a Central Bank (CB) that rules over mone- tary policy. The CB is assumed to be independent from the government, in that the latter does not participate in the choice of monetary policy. However, under some of the institutional arrangements I consider, the government is given the role of choosing central bankers. I assume that the CB has perfect control over the inflation rate, ?,which summarizes monetary policy in the model. The preferences of central bankers reflect the costs and potential benefits of inflation. Inflation and inflation vari- ability impose costs on society, for instance because they are associated with distortions of relative prices. At the same time, unanticipated inflation generates 10 benefits, which can reflect di?erent factors. On one side is the familiar Phillips curve mechanism which, in its simplest form, points that unexpected inflation reduces real wages in the presence of inflexible nominal wage contracts, stimulat- ing the demand for labor. Moreover, unexpected inflation reduces the real value of outstanding nominal liabilities, generating gains to borrowers such as the gov- ernment. As a result, central bankers derive utility from low and stable inflation, but they also value inflation surprises. Furthermore, central bankers may not be purely altruistic, in which case they also give some value to being in o?ce. The importance given to reducing inflation, relative to other objectives, di?ers across central bankers. There are many reasons why such heterogeneity can arise. In Alesina?s (1987) model, for instance, it is the reflection of ideological di?erences and/or party a?liation. It can also capture skepticism on the part of some policy makers about their ability to exploit inflation-output trade-o?s (which could in turn give them the ability to commit to a monetary rule, as in Barro?s 1986 model); or it could result from central bankers representing interests that are a?ectedby inflationindi?erent ways, suchasborrowersand lenders(Faust, 1996). The implication, in any case, is that some central bankers have more incentives than others to reduce inflation, relative to generating inflationary surprises. I assume that central bankers come in two types, which I call hawks and doves. Hawks give more importance than doves to fighting inflation. I assume that the type of a central banker is his private information, which will raise reputation considerations. The preferences of a dove central banker identified as i are reflected in the following loss function: 11 L d i = E 2 X t=1 ? t?1 i ? ? 2 t ?c(? t ?? e t )?b?o it ? (1.1) where the superindex d refers to a dove, ? t is the inflationrateinperiodt, ? e t is the public?s expectation of inflation for period t, ? i ? [0,1] is the discount factor, and o it is 1 if i is in o?ce in period t, and 0 otherwise. I allow the discount factor ? i to vary across individual bankers, so that di?erent dove central bankers may care more or less about the future implications of their choices; ? i is characterized by a cumulative density function F(? i ). The parameters c and b (where c>0 and b ? 0), capture the value given by a dove policy maker to, respectively, generating inflation surprises and being in o?ce (relative to reducing inflation). Central bankers face uncertainty that stems from two sources. First, if mon- etarypolicyischosenbyacommittee,eachcentralbankerfacesuncertainty about the types of his fellow committee members. The other source of uncer- tainty for a central banker comes from the possibility that the central banker is not reappointed for period 2. Although I will consider di?erent scenarios for how reappointment is decided, in all of them reappointment is uncertain from the perspective of period 1. The first term of loss function (1.1) implies that a dove central banker cares about driving inflation to its target value of zero (a di?erent target level could be assumed without changing the basic results). The convex functional form of this term captures inflation stabilization incentives. The second term reflects a dove?s incentives to generate inflation surprises. As in the original formulation of the Barro-Gordon model, I assume a linear func- tional form. This assumption has the shortcoming of ignoring potential output stabilization incentives, which could be a?ected by the institutional environment. 12 However, it makes the problem of the central banker more tractable; this is im- portant in the context of this paper, because the simultaneous introduction of government appointments, committee decision-making, and possible reappoint- ments expands considerably the set of possible states of the world that a central banker must consider 5 . Notice that, for any individual i serving as central banker in period one, o it=1 =1.Moreover,defining r i such that r i =1if i is reappointed for period 2 and r i =0otherwise 6 ,onecanwriteE(o it=2 )=Pr(r i =1). To simplify the notation, I will drop the time subindexes, and indicate period 2 variables with a 0 mark. A dove?s loss function can, therefore, be rewritten as: L d i = E ? ? 2 ?c(? ?? e )+? i ? (? 0 ) 2 ?c(? 0 ?? e0 ) ?? ?b[1 + ? i Pr(r i =1)] (1.2) What is important to note from this loss function is that the choices of a dove central banker in the first period a?ect not only his current losses, but also his expected losses for period 2. One channel for these intertemporal e?ects is the well known incentive to build a good reputation for the CB (Backus and Dri?ll, 1985; Barro, 1986). High inflation in the first period sends the signal to the public that the members of the CB are dove with high probability, and are therefore likely to chose high inflationinperiod2.Inotherwords,high? a?ects the reputation of the CB, increasing future inflation expectations (captured by ? e0 ). Since ? e0 enters loss function (1.2), consideration of the CB?s reputation 5 Some interesting consequences of CB collective decision-making for the ability of the mon- etary authority to stabilize output, although in more restricted contexts, are analyzed by Dal B? (2000) and Waller (1989). 6 Later I allow the government to choose r. 13 makes a dove central banker wary of increasing ? in the first period. Consider, for instance, the case of a single central banker. The public uses ? to make inferences about the type of the central banker who was in o?ce in period 1. If high ? is observed, they assign a higher probability to the central banker being dove than if ? waslow.Incasethecentralbankerisreappointed for period 2, these inferences a?ect ? e0 . Importantly, this paper?s focus on government appointments to the CB cre- ates an additional link between today?s policy and tomorrow?s expected losses. In particular, if the decision to reappoint the central banker lies with the gov- ernment, the central banker?s choice of policy in period 1 is likely to a?ect his chances of reappointment. Reappointment obviously a?ects the central banker?s expected utility from being in o?ce in period 2, captured by b. But further than that, his assessment of the probability of staying in o?ce also a?ects his expecta- tions about both ? 0 and ? e0 (that is, ? 0 and ? e0 also depend on r i , as will become clear). As a result, when choosing ?, the central banker must take into account the potential e?ects of his choice on his chances of reappointment. Under the assumption that the government favors partisan appointments, central banker i has incentives to choose the policy preferred by the government to increase his chances of remaining in o?ce. In short, when choosing policy in period 1, a dove central banker takes into account the potential e?ect of his choice on three di?erent factors: 1) current inflation, and therefore the central banker?s current welfare, 2) next period ex- pected inflation, and 3) his chances of reappointment. Simply put, a dove central banker has a preference for generating an inflation surprise in period 1, but may be discouraged from doing so because high inflation damages the reputation of 14 the CB, increasing future expectations of inflation. He may also avoid high infla- tion if this damages his chances of being reappointed, or stick to it if the opposite is true. As for hawk central bankers, I assume they care solely about driving inflation to zero in the given period, so that a hawk always chooses zero inflation. This assumption is, obviously, extreme, and it is adopted for simplicity. As will become clear later, together with the absence of incentives for output stabilization, it reduces the space of possible policies to a binary one. Although it has the cost of precluding signaling by hawks, it should not a?ect my conclusions since the focus of this paper is how the institutional environment shapes the incentives of those central bankers who are subject to reputation considerations. A central banker?s choice variable is his vote over ?,whichisequivalenttothe actual policy in the case of a single central banker, but not necessarily so under a committee arrangement. Given the preferences outlined before a hawk central banker votes for ? =0in both periods. However, the problem of a dove is more involved. Solving backwards, in period 2 a dove central banker votes for the rate that minimizes his intratemporal loss function, which I will denote as ? d in what follows: ? d = c 2 (1.3) In period 1, meanwhile, he also considers the implications of his vote on the perceptions of others about his type. These a?ect both future expected inflation and, potentially, his chances of being reappointed. Notice that, since hawks only vote for ? =0, voting for any rate di?erent from ? =0will reveal i as a dove 15 to those observing his vote 7 . As a result, a dove central banker will consider two possible votes for period 1: ? =0and ? = ? d , where the latter is his preferred vote among those that reveal his type. Inflation is thus ? =0or ? = ? d in both periods. Before solving for a dove?s optimal choice, let us define the structures of time and information. 1.3.2 Timing and information The timing of events, summarized in Figure 1.1, is as follows. There are two periods, t =1,andt =2.Eachperiodt is divided into three ?subperiods?, t ? ,t,t + .Att ? central bankers are appointed (e.g. by the government): they can be newcomers, or, for t =2, central bankers that were in o?ce in t =1and were reappointed. At t the public forms inflation expectations 8 .Att + the CB chooses ?, which is observed by all players. As for the information structure, in any given period there is uncertainty about the type of the central banker in o?ce (central bankers, if the CB is a committee), which is known only to himself. In period 1, only the unconditional probability of a central banker being a dove (which I denote by ?)isknowntothe 7 Even in the case of a committee, where his individual vote may not be observable to others, voting for ? 6=0could result in this alternative rate being chosen, which would reveal that there are at least some dove members in the committee. 8 Notice that, by assuming that expectations are formed only once within a central banker?s term, I am implicitly imposing an equivalence between the length of a central banker?s term and the length of nominal contracts. Although relaxing this assumption could have interesting implications for the cycle of monetary policy within a central banker?s term, I will keep it throughout the paper for simplicity. 16 t=1 central banker(s) appointed Public forms ? e CB chooses ?, people observe ? t - tt + t=1 t - tt + t=2 central banker(s) appointed, any reappointments announced Public forms ? e ? t=2 CB chooses ?? people observe ?? Figure 1.1: Timing of events public. Given the timing described above, however, in period 2 all players know what policy was chosen in period 1. It is also known if a central banker is an incumbent reappointed at the end of period 1, or a newcomer. This information is used by the public to update beliefs about the types of period 2 central bankers, and ultimately to form ? e0 . 1.3.3 Solving the problem of a dove central banker Remember that in period 2 a dove central banker always chooses the policy that minimizes his current losses, ? d = c 2 .Igonowbacktodiscussingadove?sproblem in period 1. Without loss of generality, I will address the problem of a specific dove central banker, whom I index by i. I denote i?s vote (which equals actual policy in the single central banker case) as v i ? {0,? d }. Since this vote can only take the values of 0 and ? d , the problem of dove central banker i in period 1 can be rewritten as 9 : 9 Here I use the facts that c? d = c 2 2 and ?c? d + ? ? d ? 2 = ? c 2 4 17 Min {v i } Pr(? = ? d | v i ) ?? ? c 2 4 ? + ? i E ? L 0 | v i ,? = ? d ? ? +Pr(? =0| v i )[? i E (L 0 | v i ,? =0)] where E(L 0 | v i ,?)= ? ? c 2 4 ? ? Pr(r i =1| v i )Pr(? 0 = ? d | r i =1)+Pr(r i =0| v i )Pr(? 0 = ? d | r i =0) ? + ? c 2 2 ? E ? Pr p (? 0 = ? d | ?) | v i ? ?bPr(r i =1| v i ) and I have abstracted from constant terms. I use Pr p (? 0 = ? d | ?) to designate the probability assigned by the public to ? 0 = ? d , given the realization of ?.I use this notation to di?erentiate the probability assigned by the public to a given outcome from the probability assigned by the central banker, who knows his type and can therefore use this additional information. The term in Pr p (? 0 = ? d | ?) comes from the fact that ? e0 = ? d ?Pr p (? 0 = ? d | ?). Note that i is solving this problem before observing the votes of his colleagues (in the committee case), and before knowing whether he will be reappointed. Moreover,hedoesnotknowthetypesofhisfellowcommitteemembers.These sources of uncertainty explain why a member of a committee does not know with certainty what ? will be given his vote. They also explain the expectation operator in front of Pr p (? 0 = ? d | ?), since the public assigns Pr p (? 0 = ? d | ?) after observing reappointments 10 . 10 That is, 18 Before proceeding with the discussion, I will introduce some notation that facilitates reading the rest of the paper. For any outcome x, I will use ?Pr(x) ? Pr(x | v i = ? d )?Pr(x | v i =0)to denote the impact of i?s vote on that outcome. More specifically, I define: ?Pr(? = ? d ) ? Pr(? = ? d | v i = ? d )?Pr(? = ? d | v i =0), ?Pr(r i =1)? Pr(r i =1| v i = ? d )?Pr(r i =1| v i =0),and ?Pr(? 0 = ? d ) ? Pr(? 0 = ? d | v 0 i = ? d )?Pr(? 0 = ? d | v 0 i =0). Note from the last line that ?Pr(? 0 ) is defined as a function of v 0 i rather than v i ,wherev 0 i is period 2?s vote of i or his replacement if he is not reappointed. Moreover, I also denote E ? Pr p (? 0 = ? d | ?) | v i ? ? E v i Pr p (? 0 = ? d | ?). Given the above discussion, dove central banker i chooses v i =0if and only if L(v i =0) ?Pr(? = ? d ) (1.4) Proof. See appendix A Condition (1.4) is key to understanding the incentives faced by a dove central banker. It indicates that a dove may vote for zero inflation in period 1 if the benefits from this choice exceed the costs. The top row reflects the benefits of voting for ? =0in terms of CB reputation, derived from the fact that low current inflation would reduce future expectations of inflation and therefore open the door for future inflation surprises. This term is multiplied by ?Pr(? = ? d ), indicating that v i =0will only generate reputation gains if i?s vote is reflected in actual policy. I will call Pr p (? 0 = ? d | ? = ? d ) ? Pr p (? 0 = ? d | ? =0)the reputation factor, which generates an incentive to vote for low inflation. On the other hand, I will refer to i?s ability to a?ect policy, captured by ?Pr(? = ? d ), as the representation factor. This factor indicates that i will be concerned about the costs his vote may impose in terms of the CB?s reputation only if that vote is represented by the policy . Themiddlerowcapturesthebenefits of v i =0in terms of reappointment, since ??Pr(r i =1)is the gain in reappointment probability from choosing v i =0.Notethat?Pr(r i =1)can have any sign, depending on how reap- pointment is chosen. Although reappointment will be discussed in detail below, it is important to note that, if appointments are in the hands of the government, 20 this term is likely to imply an incentive to vote for the government?s preferred rate (that is, ?Pr(r i =1)is negative if the government is hawk and positive it if is dove). The term in parentheses, ? 4b c 2 +(1??)?Pr(? 0 = ? d ) ? ,givesthevalue of reappointment, where ? is the unconditional probability of a central banker being a dove. As noted before, this value is not only given by the opportunistic reward to being in o?ce, b, but also by i?s desire to shape policy according to his preferences. This is key, as central bankers? incentives to increase their chances of reappointment are frequently dismissed with the argument that central bankers are not opportunistic individuals. This model makes the point that even an al- truistic central banker wants to remain in o?ce, so that his beliefs about what is optimal for social welfare are taken into account in the design of future policies. Finally, the term to the right of the inequality sign captures the fact that v i =0is costly from the point of view of period 1, since in the absence of considerations about the future a dove central banker would prefer ? = ? d to ? =0. As was the case with the reputation incentive, this current loss is only incurred if i?s vote is reflected in the actual policy, so representation also plays a role in this case. Inshort, v i =0can only result if reputation and reappointment considerations for the future overcome the current costs zero inflation imposes on a dove central banker. This requires high reputation (and/or reappointment) benefits from v i = 0, and also that central banker i assign a high value to the future. The latter requirement is reflected in the fact that condition (1.4) can only be satisfied for large enough values of ? i . 21 1.3.4 Expectations and equilibrium choices When forming expectations of inflation for period 2, individuals know that central bankers correctly represent their types in that period. That is, they know that doves will vote for ? 0 = ? d and hawks will vote for ? 0 =0,andthereforePr p (? 0 = ? d | ?) depends on people?s beliefs about the types of period 2 central bankers. For a newly appointed central banker, past policy provides no information, so the public assigns a probability ? that he is dove (where ? was defined above as the unconditional probability that a central banker is dove). For central bankers that were reappointed from period 1, however, the public updates beliefs using Bayesian updating. If i is reappointed for period 2, the public will assign: Pr P (i is dove | ?)= Pr(? | i is dove)Pr(i is dove) Pr(?) (1.5) Moreover, letting w =Pr(v i = ? d | i is dove) Pr(? | i is dove)=w?Pr(? | v i = ? d )+(1?w)Pr(? | v i =0) (1.6) In equilibrium, w should reflect the optimal strategy of dove central bankers, as presented in proposition 1. Hence, using condition (1.4), w is given by w = ? Pr(? i < ?)=F(?) if 1 > ? > 0 1 otherwise ? (1.7) where ? = 1 E v i ( Pr p (? 0 =? d |?=? d )?Pr p (? 0 =? d |?=0) ) 0.5 ? ?Pr(r i =1) [ 4b c 2 +(1??)?Pr(? 0 =? d ) ] ?Pr(?=? d ) 22 Equation 1.7 reflects the fact, already discussed, that for small values of ? i i chooses v i = ? d . It is important to note that w summarizes the solution to this problem, since it represents the choices of dove central bankers (and the choices of hawks are trivial). Also, the solution to this problem involves the public forming expectations rationally given the optimal strategies of central bankers, and central bankers optimally choosing their votes on policy given the public?s beliefs. The former is reflected in (1.5). The latter is satisfied because central bankers know the mechanism of expectations formation just described, and take it into account when evaluating Pr p (? 0 = ? d | ?). It is important to point out that any given central banker does not play a mixed strategy. That is, i either chooses v i =0with probability 1 or v i = ? d with probability 1. Although this and the solution to i?s problem are known to the public, the public can only make an evaluation of the probability that a given member chooses v i = ? d , because they do not observe the individual ? i .Di?erent discount rates capture sources of heterogeneity across central bankers other than their relative preferences towards fighting inflation 12 .Notealsothatanygiven central banker faces a problem similar to that of the public, in that he cannot observe the types of his colleagues. Therefore, he also uses w to evaluate the probability that any other dove central banker chooses v i = ? d . The discussion above presents a general characterization of the equilibrium choices of central bankers. I will now analyze what these conditions imply in terms of the optimal institutional design of the CB. I will present alternative 12 If all dove central bankers were identical, w would be either 1 or 0.Inthiscase,reappoint- ment incentives are the only reason a dove central banker would ever choose low inflation (if w =0all members are voting for the same rate, so any given member knows his vote does not define policy, i.e. the representation factor is 0). 23 institutional settings and study the choices of central bankers under each of these environments. The strategy will be to focus on condition (1.4), and study how institutional design a?ects the di?erent incentives faced by central bankers (sum- marized by the three key concepts of reputation, reappointment, and representa- tion, discussed above), and ultimately how they a?ect w. When discussing the convenience of one or another institutional design, I will be implicitly assuming that CB institutions should minimize both inflation bias and the government?s influence over monetary policy. These two goals are frequent motivations for placing monetary policy in the hands of an independent CB, and I here simply take them as given 13 . 1.4 The institutional dimension I: collective de- cision making Consider first the case for a committee-based CB. As a first approximation, I will isolate the analysis of collective decision-making from the consideration of government appointments. I do so by assuming throughout this section that a central banker?s reappointment is exogenously given and independent of his vote in the first period, so that ?Pr(r i =1)=0. The condition under which a central 13 Although I do not provide formal proof that reducing inflation and the gvoernment?s influ- ence is optimal, my assumptions about the preferences of central bankers are in fact consistent with social losses that are increasing in inflation and inflation variability, under the very plausi- ble assumption that the preferences of central bankers are similar to those of other individuals in society. Note that political pressures constitute a source of undesirable variability of mone- tary policy, in that the political cycle would impose changes in policy not warranteed by shocks to economic fundamentals. 24 banker votes for ? =0(see proposition 1) can thus be written: 2? i E v i ? Pr p (? 0 = ? d | ? = ? d )?Pr p (? 0 = ? d | ? =0) ? > 1 (1.8) Note that not only the terms involving reappointment incentives disappear, but the ?Pr(? = ? d ) term does not enter (1.8) either. The reason for this is that, without government appointments, the votes of central bankers only mat- ter if they a?ect policy; they do not have any costs or benefits by themselves. As a result, any ?Pr(? = ? d ) > 0 su?ces for policy considerations to be the only factors in a central banker?s choice. This does not mean that representation ceases to play any role, as we will see when discussing the implications of a com- mittee design, but it does imply that representation does not a?ect the relative importance of reputation-building incentives, compared to other considerations. Given condition (1.8), reputation building is at the heart of a central banker?s choice in the case without government appointments: v i =0is chosen if and only if the reputation gain from low inflationishighenoughtoovercomethecurrent loss ? =0imposesonadovecentralbanker.Itisthusthecasethatinflation bias is minimized when the impact of inflation on the CB reputation is maximized. With this in mind, I will analyze how reputation considerations are a?ected by specific characteristics of a committee-based CB, such as its size and the fraction of members being replaced in each period. As a benchmark, I start by studying the case of a single central banker, and then move to the committee case. 1.4.1 Single central banker Suppose there is a single central banker who, after his first term in o?ce, is reappointed with exogenous probability p (Pr(r =1)=p). One may see 0 25 and 1 as the only plausible values for p, but I will allow intermediate values for the purpose of comparing with the committee case, and later with government appointments. With monolithic decisions, a central banker?s vote is perfectly reflected in the policy choice and Pr p (? 0 = ? d | ?,r i =1)=Pr(i is dove| ?). Public beliefs about period 1?s central banker, who I will still call i,aregivenby Pr(i is dove | ?)= ? 1 if ? = ? d (1?w)? 1?w? if ? =0 ? (1.9) where I have used the Bayesian updating mechanism captured by equation (1.5). Hence, the reputation cost of choosing ? d is reflected in the fact that Pr(i is dove| ?) is larger when period 1 inflation is high. Moreover, this cost is only perceived if the central banker is reappointed. If he is replaced by a newcomer, people do not give any importance to past policy in setting inflation expectations, and ? Pr p (? 0 = ? d | ? = ? d )?Pr p (? 0 = ? d | ? =0) ? =0. Given these elements, during period 1 this dove single central banker can produce zero inflation in equilibrium if: p2? i ? 1? (1?w)? 1??w ? > 1 (1.10) where, again, the left hand side reflects the incentive to build a reputation for the CB being tough on inflation. Zero inflation can only be supported by the dove central banker if ? i is large enough, ? is low enough and the probability people assign to doves voting for high inflation, w, is high. ? i increases the value 26 given to reducing future expected inflation. Meanwhile, with large ??s, people assign a high probability that the central banker is dove to begin with, so it is hard to lead them to believe the opposite through the choices of ?. Finally, a high value of w implies large reputation gains from choosing low inflation, since people know doves are very unlikely to choose ? =0,andthereforeassignalow probability that someone who chose low inflation is dove. The key element to keep in mind, however, is the fact that i?s reputation is only relevant for next period?s expected inflation if i is reappointed. This is the reason why the reputation incentive, given by the LHS of condition (1.10), is increasing in p. A single central banker thus only cares about building reputation if there is some positive probability that he will be reappointed for next period. Interestingly, a characteristic of single central banker arrangements is precisely that the term of the CB expires at some point, di?erent from committee designs where the terms of central bankers can be staggered to ensure that at every point in time there is some continuity. If we considered period 1 to be the last (only) period in o?ce of our dove single central banker (such that p =0), he would not choose ? =0no matter what values other parameters took. 1.4.2 A committee Central Bank Suppose now that monetary policy is decided by a committee, whose members vote over the possible policy choices. What makes collective and monolithic decision-making di?erent from the point of view of a central banker? Without government appointments, there are three major dimensions along which di?er- ences exist: the power of each central banker to a?ect policy, the potential for reputation building and the importance central bankers assign to it, and the 27 available sets of rules for appointments to the CB. First, when policy is decided by the votes of several individuals, the vote of any given member may not be reflected in the actual policy choice. This is the representation concept discussed above: a given member may not be represented by policy. Although, as captured by condition 1.8, representation does not a?ect the relative importance of reputation when reappointments are random, it does have an e?ect on how past inflation a?ects expectations of future inflation. Put simply, past policy does not reveal individual votes, and therefore contains less information about the types of each member than it does in the single central banker case. One can think of this in terms of policy revealing only the votes of the majority. As a result, the amount of information relevant to learn the types of central bankers that is contained in past policy is decreasing in the level of representation. Second, the incentives faced by central bankers in terms of building a rep- utation for the CB can be fundamentally di?erent when policy is chosen by a committee. The key assumption in this paper is that the public forms expecta- tions of future inflation based on past policy, rather than on individual votes of central bankers. This is captured by the assumption that the public observes ?, but not the individual v i . In some of the literature (e.g. Sibert, 1999), a similar feature appears in the assumption that individual votes are not disclosed, or are disclosed only with a lag. However, the public may base expectations only on past policies even if individual votes are disclosed in the CB?s public statements, and it is my view that this is a better representation of reality. On one side, while information about monetary policy choices makes it to the headlines of the mass media outlets, the positions of individual central bankers receive less atten- 28 tion, often restricted to specialized media. This creates costs that may prevent the public from getting information about individual votes at the CB. Moreover, tracking the behavior of each individual member adds one more layer to the, as we will see, already sophisticated calculations involved in forming inflation ex- pectations. It seems relevant to question whether the general public, or even the players involved in wage setting, analyze every policy decision with such high level of sophistication. In terms of reputation building incentives, the idea that the public forms expectations based on policy outcomes rather than individual votes implies that inflationexpectationsdependonthecollective reputation of the CB. As a result, central banker i?s vote can damage the CB?s reputation and a?ect future expected inflation even if he is not reappointed. A committee design has therefore the potential, through this channel, to increase the importance central bankers give to protecting the CB?s reputation. Finally, in terms of reappointment mechanisms, a committee arrangement provides a flexibility that cannot be matched by single central banker designs. The point is simply that for any maximum period a single central banker can stay in o?ce, a committee can be designed to outlive that maximum length 14 . In particular, the rules of the committee can be such that at any given point in 14 More generally, the key issue is that committee appointment rules can always imitate their single central banker counterparts, but the opposite is not true. At any given point in time a single central banker either stays or goes, so that p is either 0 or 1. A single central banker that faces retirement (p =0) can be imitated by a committee where the terms of all members expire simultaneously. I the single central banker faces p =1, an analogous committee would have all members staying. However, a committee can also be designed to have some members outlasting others, and this is a feature monolithic institutions cannot imitate. 29 time at least some of the members remain in o?ce. As discussed in the intro- duction, this possibility is considered an important tool to guarantee continuity in monetary policy-making, and it has been a key motivation for the adoption of committee CB designs in many countries. I will therefore consider a committee design that exploits it. Let us now solve the problem of a dove central banker under a committee arrangement. I assume there are n =2z +1members in the committee, and policies are decided by a simple majority rule, so that the inflation rate that obtains z +1or more votes is adopted 15 . In particular, I assume that at the end of period 1 exactly m members are removed from o?ce, while the remaining n?m are reappointed 16 .Asaresult,Pr(r i =1)=1? m n . The public knows that in period 2 all doves in the committee vote for ? = ? d while all hawks vote for ? =0. The probability that ? 0 = ? d is therefore given by the probability that z+1or more members of the t =2committee are doves. The public?s estimation of this probability is influenced by period 1 inflation because ? 1? m n ? of period 2 central bankers helped choose ?. In other words, ? contains information about a fraction ? 1? m n ? of the committee that serves in t =2. Letting i be a member in period 1, and public posterior beliefs be denoted by Pr(i is dove| ? = ? d ) ? ? + and Pr(i is dove| ? =0)? ? ? ,weobtain 15 Notice the assumptions imply an odd-numbered committee, where we can abstract from tie- breaking rules. I do not address here the question of optimal voting rules, which the literature has now also started to look at (see, for instance, Dal B??s 2000 paper). 16 This assumption closely matches the arrangement that characterizes the Colombian CB. It is also close to other designs with staggered terms (like the cases of Mexico and Venezuela). In many cases, these arrangements are mixed with caps on the number of times a given member can be reappointed. 30 Pr p (? 0 = ? d | ? = ? d )?Pr p (? 0 = ? d | ? =0) = ? 1? m n ?? ? + ?? ? ? n P x=z+1 h ? n x ? ? x (1??) n?x ? x ? ? n?x 1?? ?i (1.11) where I have used an approximate form of Pr p (? 0 = ? d | ? = ? d ) to simplify the analysis (a formal derivation of equation (1.11) can be found in appendix A). People?s beliefs about the ?representative? central banker (i.e. ? + and ? ? )are ruled by: Pr(i is dove | ?)= ?w?Pr(? | v i = ? d )+?(1?w)Pr(? | v i =0) Pr(?) (1.12) Expression (1.11) captures the cost of high period 1 inflation in terms of the collective reputation of the CB. This cost is increasing in the fraction of period 2 central bankers who also served in period 1,because? contains information only about those n?m members. It also depends on a modified version of the unconditional probability that the majority of members of the committee are doves (the term preceded by the summation sign). Finally, it is proportional to the e?ect that choosing high inflation has on the public?s beliefs about the probability that any given member is dove, captured by ? ? + ?? ? ? . Those beliefs are generated with the Bayesian updating mechanism captured by equation (1.12) (derived from (1.5) and (1.6)). Note that the reputation loss captured by (1.11) depends crucially on the size of the committee and the number of members who are replaced from one period to the next, represented by the parameters n and m.Anincreaseinthe fraction of members replaced, m n , reduces the amount of information relevant for period 2 that is contained in period 1 policy. As a result, the reputation gain 31 from avoiding high inflationinthefirst period is decreasing in m n ,asthefirst factor in equation (1.11) shows. On the other hand, representation is lower in larger committees, as the impact of any individual member?s vote on the policy choice decays when the size of the committee grows. The public therefore learns less from policy about the types of individual members when the committee is large, and this reduces the potential reputation gain to be obtained from choosing low inflationinthefirst period. This point may be easier to understand if one thinks of policy as revealing the votes of the majority, where the fraction of votes that constitutes a majority is decreasing in the size of the committee. In this sense, policy reveals less about the ?average member? if the committee is larger. In terms of expression 1.11, this e?ect of representation is reflected in the fact that ? ? + ?? ? ? is decreasing in the size of the committee 17 . Finally, n also has a positive e?ect on the last term of (1.11). Notice an overall conflicting e?ect of the committee size: while increases in n increase this last term and (1 ? m n ), increasing the reputation cost of high ?, they also generate an opposite e?ect by reducing the amount of information contained in policy about the ?average member?. Consider, for instance, the illustration of ? Pr p (? 0 = ? d | ? = ? d )?Pr p (? 0 = ? d | ? =0) ? for di?erent values of m and n in Figure 1.2. The reputation loss is depicted as a function of w, the probability that a dove votes for low inflationinthefirst 17 One can write ? ? + ?? ? ? = ?w(1??)?Pr(?=? d ) Pr(?=? d )(1?Pr(?=? d )) , which makes the importance of the representation factor(appearing explicitly in the numerator) evident . Note also that this term does not depend on m, because it refers to the probability that a member who is known to have been reappointed is dove. 32 period, under the assumption that ? =0.5 (i.e. an ex-ante equal proportion of doves and hawks). For this and all other examples, using the notation (n, m n ),I will compare the case of (3, 1 3 ) to the cases of (9, 1 9 ) and (9, 3 9 ). These cases are interesting because in any pair of them n, m or m n are kept constant. In this and all other figures I will use crosses to depict (3, 1 3 ), a solid line for (9, 1 9 ),andboxes for (9, 3 9 ). Going back to figure 1.2, note that if n grows keeping m constant, so that the fraction of members replaced from one period to the next is reduced, and ? becomes more important as a signal about ? 0 (which is the reason why the solid line is always above the crosses). This dominates other potential e?ects of increases in the committee size, and as a result larger committees increase the reputation loss from high inflation. However, if m n is kept constant a potential negative e?ect of size on reputation shows up, as can be seen by comparing the (3, 1 3 ) and (9, 3 9 ) cases when w is large. The negative e?ectmaydominateonlyif representation is su?ciently important. This is the case when w is large, because then the separation between hawks and doves in maximized, and the public gives greater importance to the choices of central bankers as reflection of their types. Comparing this reputation loss to the case of a single central banker discussed above shows both advantages and disadvantages of the committee design. On the darker side, because representation is lower in the committee case, the impact of policy on beliefs about individual central bankers is reduced, so that ? ? + ?? ? ? in the committee case is smaller than the corresponding e?ect on beliefs in the single case 18 . On the other hand, as mentioned above, the variety of appointment arrangements available for a committee structure implies that it can always be 18 This cost of committee design has also been pointed out by Sibert (1999). 33 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0.4 0.6 0.8 1 w Figure 1.2: Reputation-building incentive, no government designed to outperform the single central banker in terms of the continuity of policy from one period to the next (i.e. m n canbechosenassmallasdesired, and turnover is smooth over time, while for a single central banker r must be 0 in some periods). In particular, if we compare the decision of a dove who knows he will not be reappointed under the two arrangements, only in the committee case is it possible that he will vote for low inflation. This stems from the fact that even if he is not reappointed, his vote will impact the CB?s reputation. The benefit of committee designs in terms of continuity, or ?inertia?, of monetary policy has been stressed before (e.g. Blinder 1998). This model points out that the benefits of this greater persistence do not come only from welfare gains of reducing the variability of inflation over time, but also from greater incentives for policy makers to keep low inflation to protect the reputation of the CB. In net, a committee can imply more reputation building than the single central banker case, and increase the chances that low inflation is achieved. However, this is only possible if the committee is appropriately designed in terms of the choice of m and n; a design that maximizes reputation-building incentives implies choosing 34 nm 31 1 10.94 21 10.92 72 1 31 10.90 206 11 3 1 41 51 Probability that dove central banker votes for high inflation 5 Table 1.1: Case with no government small m n to maximize the persistence of policy, but may also require a small enough committee that past policy is considered informative about the types of individual members. Using equation (1.11) for the reputation loss, the condition under which cen- tral banker i votes for low inflationisinthiscase: 2? i ? 1? m n ? ? ? + ?? ? ? n X x=z+1 ?? n x ? ? x (1??) n?x ? x ? ? n?x 1?? ?? > 1 (1.13) It is clear from this expression that a necessary condition for a dove cen- tral banker to choose low inflation is that the reputation loss in equation (1.11) exceeds 1 2 . From Figure 1.2, when ? =0.5 this is only possible for some commit- tee arrangements that exhibit low proportions of members replaced from period to period. As further illustration, Table 1.1 shows the probability that a dove chooses v i = ? d for an example with ? =0.5 and ? i ? U(0,1) (this probability is equal to w, and it is calculated as in equation (1.7)). Only for low m n cases will high ? i doves vote for low inflation. It is important to point that these results imply some benefits from the fact 35 that the public does not observe individual votes. In particular, because inflation expectations are based on past policy, the actions of a central banker can a?ect the CB?s reputation even if he is not reappointed. This makes a central banker wary of increasing inflation, even if he knew he would not be reappointed 19 . This conclusion runs against A. Sibert?s (1999) result that publishing the votes of central bankers reduces inflation bias. The reason for this di?erence is that in her model the terms of central bankers are of fixed length, and dove central bankers optimize solely over their horizon in o?ce. Hence, either the central banker knows he will be in o?ce in period 2 for sure, in which case considerations of reputation ?even if he is not reappointed? are irrelevant, or he does not care about the following period and hurting the CB?s reputation does not concern him. 1.5 The institutional dimension II: government appointments Consider now the possibility that appointments (and reappointments) of central bankers are not an exogenously determined event, but rather a choice of the gov- ernment. The government appoints central bankers for period 1,decideswhether a central banker is reappointed at the end of that period, and chooses who will replace a central banker who is not reappointed. The president has a seat (but notavote)ontheCentralBankcommittee,sothathecanobservethevotesof 19 I believe this is a fair representation of the actual decisions of central bankers, who are usually highly concerned above eroding the credibility of the CB, even in their last terms in o?ce. 36 the central bankers 20 . As a result, the reappointment of an incumbent central banker is based on his vote for period 1 policy. As for the appointment of new- comers (for period 1, or to replace a retired incumbent in period 2), there is no relevant information available to distinguish one potential central banker from another, so the government randomizes among all newcomers. In practice, obviously, previous choices by incumbent central bankers are not the only information governments use to choose central bankers. Governments are likely to appoint newcomers on the basis of their party a?liation and their previous performance in other areas of public service, or in the private sector. Although this makes my assumption that new appointments are uninformed de- cisions, it reflects the fact that this paper focuses on the e?ects of appointments on the decisions of incumbent central bankers, which translates into a focus on reappointments (as opposed to appointments of newcomers) 21 . My modeling of the government is extremely simple. I abstract from the political process that determines who the president is, and assume that the gov- ernment?s type is known to everybody. I also focus on partisan appointments by assuming that governments want to appoint members of their own type. Since members who voted for high inflation reveal themselves as doves, a dove government tries, subject to the institutional constraints in place, to reappoint these members. For the same reason, a hawk government tries to replace central 20 It is indeed frequent that the government participates in the meetings of the CB committees. In many countries, this takes the form of the Finance Minister or a member of his sta? being also a member of the CB committee. 21 For an analysis of the e?ects of partisan appointments of newcomers ?in the absence of reputation concerns?, see A. Alesina?s (1987) model, and subsequent extensions to the case of committees (e.g. Waller, 1989; Lohman, 1997). 37 bankers who voted for high inflation. This strategy on the part of the govern- ment is consistent with previous models of government appointments to the CB. For instance, given the abstraction from presidential elections and congressional approvals of appointments to the CB, partisan appointments will arise in both Waller?s (1992) and Havrilesky?s (1995, chapter 9) models. How do government appointments a?ect the vote of a dove central banker? Consider again the condition under which i chooses v i =0, stated in proposition 1(andwritteninaslightlydi?erent form): ? i ? E v i ( Pr p (? 0 =? d |?=? d )?Pr p (? 0 =? d |?=0) ) 0.5 ? ?Pr(r i =1) [ 4b c 2 +(1??)?Pr(? 0 =? d ) ] ?Pr(?=? d ) ? > 1 (1.14) As discussed, the firsttermontheLHSreflects the benefits in terms of the CB reputation of choosing v i =0, while the second term captures potential benefits or losses stemming from the e?ect of v i on i?s chances of being reappointed. Given the government?s reappointment strategy, voting for low inflation increases one?s chances of being reappointed if the government is hawk, and decreases them if the government is dove. This is reflected in the fact that ?Pr(r i =1)is negative (positive) when the government is hawk (dove). The reappointment factor thus creates incentives to vote for the rate preferred by the government. Furthermore, notice from the second term that the value given to reappoint- ment relative to other objectives is decreasing in ?Pr(? = ? d ).Thisreflects the fact that the probability of being reappointed depends on the central banker?s vote while his other objectives are only a?ected by the actual choice of policy. Thus, representation now plays the role of determining how much each central banker values reappointment: if each vote has very low impact on the actual 38 choice of policy, the decisions of central bankers will be mainly driven by reap- pointment incentives. As a result, the influence of the government on monetary policy is greatest when representation is lowest. I now return now to the analysis of specific institutional designs, with em- phasis on how they a?ect the balance between reappointment and reputation considerations. 1.5.1 Government appointments and the case of a single central banker Let us start again with the case of a single central banker. Remember first that the single central banker has perfect power over the choice of policy, so that ?Pr(? = ? d )=1. Also, if g = {h,d} represents the type of the government, then the general reappointment strategy discussed above implies: ?Pr(r =1| g = d)=1 ?Pr(r =1| g = h)=?1 Thestructureofreappointments a?ectstheincentivesfacedby central bankers, not only because they value reappointment, but also because current inflation a?ects the reputation of the CB in the future only if the central banker is reap- pointed. The condition under which a dove central banker chooses low inflation is now: 39 ? i ? 2(1??)?w (1??w) + 4b c 2 +(1??) ? > 1 if g = h (1.15) ? i ? 2(1??)? 4b c 2 ?(1??) ? > 1 if g = d where I have used the fact that, if the central banker is not reappointed, Pr p (? 0 = ? d | ?)=? for any ?. Also, to keep matters as simple as possible, I have assumed that the public?s beliefs about the type of the central banker are not a?ected by the type of the government, so that the formula for Pr(i is dove | ?) is as in the case without government 22 (equation (1.9)). In each case the first term in the LHS of condition (1.15) represents the reputation loss from choosing high inflation, while the rest of the LHS represents reappointment considerations. As already discussed, the latter increase the incentives to choose v i =0if the government is hawk and decreases themin the opposite case, since central bankers value being reappointed. It is clear from this expression that government appointments tie the choices of dove central bankers to the preferences of the government, implying greater probability of low inflation under hawk governments than under dove ones 23 .In 22 If the public takes the decisions of the government into account, it will assign a higher probability that members reappointed to a second period are of the same type of the govern- ment who kept them in o?ce. This diminishes the incentives to central bankers to follow the government?s preferred policy (because the cost of high inflation in terms of future expectations is higher under a dove government), but does not completely o?set them. 23 To see this clearly, rewrite the LHS of the condition as (1+?w)(1??) 1??w + 4b c 2 when g = h and (1 ? ?) ? 4b c 2 when g = d. Note that the expected reputation loss from choosing high inflation may be larger under a dove government because the central banker will be reappointed 40 fact, under the specific assumptions made here, a dove central banker will never choose low inflation if the government is dove. From a broader perspective, when appointments to the CB are a choice of the government, the political cycle can be a source of variability for monetary policy. I turn now to the question of how collective decision-making a?ects this link between monetary policy and the preferences of the government. 1.5.2 Collective policy-making and government appoint- ments Consider now government appointments under in the committee case. As is by now familiar, we need to worry about the e?ect of v i on the reputation-building and reappointment incentives, represented respectively by the first and second terms in the LHS of E v i ( Pr p (? 0 =? d |?=? d )?Pr p (? 0 =? d |?=0) ) 0.5 ? ?Pr(r i =1) [ 4b c 2 +(1??)?Pr(? 0 =? d ) ] ?Pr(?=? d ) > 1 ? i (1.16) Since we are assuming that people?s beliefs about the types of period 1 cen- tral bankers are not a?ected by the type of the government, reputation building incentives are exactly as in the case without government. Future expected in- flation does not depend on what members were reappointed, as the public has no information on individual votes to distinguish between the types of those reappointed and those replaced. Since reappointment does not a?ect expected inflation, reputation-building incentives do not change with the type of the gov- ernment; they did change in the single CB case because then whether the central precisely after choosing high inflation . This diminishes the power of the government to dictate monetary policy but does not completely overcome the reappointment incentive, as reflected in the condition required for low inflation. 41 banker was reappointed had an e?ect on expectations, and reappointment proba- bilities for a given vote changed with the government?s type. Since the reputation building incentives are not modified by the inclusion of the government, I will start by studying the reappointment incentive, and ask how should the committee be designed to minimize the power of the government to dictate monetary policy. The overall e?ect on the incentive to choose v i =0will be addressed later. As before, the committee has n members and m of them are to be changed from period 1 to period 2. The president chooses which members will be replaced and, given his preference to keep central bankers of his same type, he starts by replacing members who voted for the policy he likes less. For instance, suppose the government is hawk, and let V ?g be the number of members who voted for ? d .IfV ?g ? m, the president randomly chooses m members among those who voted ? d .IfonthecontraryV ?g 0, V 00 (?) < 0,andg t also depends on P, given that it is chosen by the government. A voter j is thus characterized by ? j . (To help in following the exposition, note that V (g t ) does not depend on j, since all voters receive the same level of targeted expenses. Hence in discussing the problem of inferring g t+1 from g t , we may ignore the index j.) An individual?s only choice is whether to vote for the incumbent or the chal- lenger, and only in an election period. We therefore focus on utility as of period t, when the election takes place. The present expected discounted utility of indi- vidual j as of period t is W j t = U j t (I)+?E t U j t+1 (P) (2.2) where ? is the discount factor, and P = I,C. E t U j t+1 (P) is the utility expected for period t +1from the perspective of t.Uncertaintyaboutt +1stems from two sources: uncertainty about who will be in o?ce after the elections, and uncertainty about the preferences of the candidates. The latter will be made clear in the next section. 2.2.2 Politicians We assume fixed tax revenue and zero budget deficit, so that government fiscal policy consists only of expenditure decisions. There are two government provided goods: g t to voters and K t , a good which the politician values but which voters 60 do not. One may think, for example, of politicians who value managing a large bureaucracy. (For simplicity of exposition we call K t ?desks?.) However, the idea we have in mind is more general: voters may value some government services less than others for many reasons, such as the low visibility of some types of expenditure or voters? failure to recognize the positive externalities these services produce. The characterization of K t as total waste in the eyes of voters is sim- plyanextremewaytocapturethosedi?erences in the value assigned by voters to di?erent goods and services provided by the government. Each period, the government thus faces a budget constraint: T = K t + g t (2.3) where T is a fixed an exogenous level of tax revenue. The politician?s objective is to maximize a weighted sum of transformed vot- ers? utilities, the fixed value ? of being in o?ce, and the value of non-targeted expenses. We let ? P,t denote the weight a politician P puts on voters, as op- posed to non-targeted expenses, in period t. A politician P 0 s objective in the post-election period can then be written ? P t+1 = ? P,t+1 " V (g t )+ N X j=1 ? ? j ?? P ? 2 N # + a(K t+1 )+? (2.4) where N is the size of the voting population, which we assume to be constant, K is non-targeted expenditure and a(K) is an increasing and (weakly) concave function. We have written this objective in per-capita terms for simplicity. The crucial assumption in our argument about the e?ectiveness of election year expenditures to produce votes is that ? Pt is not observed by the voters, 61 but is correlated over time. Voters can only try to infer the value of ? I,t+1 from observations on g t , that is, on expenditures before an election. Hence, in this model the reason why voters prefer some types of politicians over others has to do with the politicians? preferences, rather than their ability to provide public goods. Information asymmetries do not refer to a lack of ability, on the part of voters, to observe parts of the budget, but to the politician?s preferences being his private information. We make assumptions such that ? I,t provides all relevant information about ? I,t+1 , which is the reason why voters focus solely on g t rather than incorporating policies further into the past (i.e. g t?1 ). More generally, the idea is that is ? I,t+1 is closest to ? I,t , so that voters would take into account all past policy realizations, but give more importance to more recent observations. We abstract from this possibility for simplicity, and therefore will make no further reference to period t?1, which is introduced only to show why manipulation occurs right before the election, rather than over the whole horizon of the incumbent in power. We assume the process governing the evolution of ? Pt for P ? {I,C} takes the simplest possible form that satisfies the conditions discussed above. In par- ticular, for any politician P, ? Pt cantakeontwovalues:? Pt = {?,?} with prior probabilities Pr(? = ?)=?p and Pr(? = ?)=(1? ?p).Wesuppose? > ?,so that a politician endowed with ? cares more about targeting expenses to people (a people politician), while ? Pt = ? makes the politician more interested in bu- reaucracy than targeting (a desks politician). Moreover, also for simplicity, ? Pt only changes from t?1 to t,butnotfromt to t+1.Thatis,? Pt = ? P,t+1 ? ? P for P ? {I,C}. Abstracting from constant terms, we write the incumbent?s objective in an 62 election year as ? I t = ? I V t (g t )+a(K t )+??(g t )? I t+1 + ? (1??(g t ))? I(P=C) t+1 (2.5) where ? is a discount factor and ? is the probability of re-election, which is a function of g t given that voters use g t to make inferences about the incumbent?s preferences. ? I(P=C) t+1 represents the utility the incumbent would obtain if the challenger won the election, given by his post-election objective function evalu- ated (in expected terms) at ? P = ? C , ? =0, and the levels of g t+1 and K t+1 that the challenger provides. Note that in (2.5) we dropped the term concerning ideology because, since ideological positions of both voters and candidates are given, this argument is a constant in period t. One solves the problem backwards, starting with the post-election period. A government P maximizes (2.4) by choice of g t+1 subject to the budget constraint (2.3). The solution is given by the following first-order condition: ? P V 0 (g t+1 )=a 0 (T ?g t+1 ) (2.6) The assumption of concavity of V (g) and a(K) implies that the post-election transfers to people are increasing in the weight the politician gives to such trans- fers, so that g t+1 (?) >g t+1 (?). We will denote g t+1 (?)=g, and suppose that ?,thevalueof? for the people-type politician, tends to ? ,sothathealways chooses the maximum level of transfers possible, g t+1 (?)=T. This assumption simplifies the solution but, as we discuss later, we could dispense of it and still prove that politicians are expected to engage in pre-electoral increases in targeted expenses. 63 In the election period, the incumbent chooses g t to maximize his objective (2.5), subject to the budget constraint (2.3). The fact that g t a?ects his chances of reelection, implies that a politician may then choose a value of g t di?erent from what he chooses in a non-election period. Given our assumption that ? ?? ? ,a ?people-type? politician would provide the maximum possible g t even in the non- election period, so he would not change his policy in the election period. A desks policy-maker (one characterized by ?), however, could choose g t (?) >gif that would get him more votes. In particular, from the solution to the post-election fiscal problem and voters? preferences, notice that voters prefer a people policy- maker over a desks one, so the latter may find worth pretending he is of the people type. There are, therefore, two possible choices of policy for an incumbent with ? = ?: his own preferred non-election spending (that is, g t (?)=g t+1 (?)=g), or g t (?)=T, the transfer provided by the high type in any period 2 . He will choose high transfers if the current utility benefit from choosing his preferred policy (low g t ) is smaller than the benefit derived from increasing his reelection chances through high targeted expenses. More formally, the desk-type incumbent will choose high targeted spending in the election period if ?? < ? ? ?(T)?? ? g ?? (???p + ?) (2.7) where ?? is the current utility gain to a policy maker of ? type of choosing his own preferred policy: ??? ? ? V (g)?V (T) ? + ? a(T ?g)?a(0) ? > 0 2 In fact, we chose ? ?? ? precisely to constrain the choice space to this binomial configu- ration. Notice that by choosing any policy di?erent from g t = T the politician reveals himself as the low type, which is the reason why no other policy generates electoral benefits. 64 (where the last inequality follows from g being policy that maximizes ?(V (g))+ a(T ?g)). Note that the benefit from choosing high transfers in t, given by the right hand side of condition (2.7), depends on both the gain in reelection probability, ? ?(T)?? ? g ?? , and the value of being reelected. The latter includes not only the exogenous value given to o?ce, ?, but also the value of having in t+1one?s preferred policy rather than the challenger?s, which is non-zero only in the event the challenger is of a di?erent type. This explains the presence of ?p =Pr(? P = ?) in condition (2.7). The following proposition summarizes the behavior of the incumbent: Proposition 2 In the election period, the incumbent?s optimal choice of targeted expenses g t (? I ) is characterized by the following policy rule: g t (? I = ?)=T and g t (? I = ?)= ? T if ?? < ? ? ?(T)?? ? g ?? (???p + ?) g otherwise ? To find the equilibrium choice of policy, we still need to characterize the re-election probabilities ?(T) and ? ? g ? , which depend on voters? choices. We therefore turn now to characterizing those choices. 2.2.3 Voting behavior and election outcomes We consider now the problem of voters. Let E [V (g t+1 ) | P,g t ] be voters? expec- tation of his utility from government transfers if politician P is elected for t+1, 65 conditional on g t . An individual j votes for the incumbent if he expects to receive higher utility in t +1under the incumbent, relative to the challenger. This is, E [V (g t+1 ) | I,g t ]?(? j ?? I ) 2 >E[V (g t+1 ) | C]?(? j ?? C ) 2 (2.8) The voter knows the ideological positions of both challenger and incumbent, ? I and ? C , as well as the politicians? first order condition for the post-election period, (2.6), which determines g t+1 . However, he has imperfect information about both ? I and ? C . To infer the challenger?s position, he has no other in- formation than the knowledge of the ex-ante distribution of ?, summarized by Pr(? C = ?)=?p. However, voters can use the realization of g t to extract addi- tional information about the incumbent?s type. Using Bayes? rule and the prior ?p, voters adjust their beliefs about the incumbent according to following expression: ?p 1 (g t ) ? Pr(? = ? | g t )= Pr(g t | ? = ?)?Pr(? = ?) Pr(g t ) (2.9) where we have denoted the posterior probability voters assign to ? =?? as ?p 1 (g t ). Equation (2.9) captures the rational essence of voters in this model. Specifi- cally, it implies: ?p 1 (g t = g)=0 That is, since voters know a people type politician never chooses low transfers (Pr(g t = g | ? = ?)=0), upon observing g t = g they assign a zero probability to the incumbent having ? I = ?. On the other hand, ?p 1 (g t = T)= ?p ?p +(1? ?p)q (2.10) 66 where q =Pr(g t = T | ? = ?) ? 1 is the probability that a low-type politician will choose g t = T in the electoral period. The nature of voters? posterior beliefs reflects an essential characteristic of this political game: while high pre-election transfers indicate the incumbent would also choose high targeting after the election, voters only take this signal seriously if the political incentive is not so large that in the election period any politician would provide high electoral transfers, no matter what his preferences are. This explains why observing high pre-election transfers does not give the incumbent any advantage over the challenger if q =1(i.e. ?p 1 (g t = T)=?p if q =1). We can now rewrite the condition under which voter j prefers the incumbent over the challenger, condition (2.8), as: (?p 1 (g t )? ?p) ? V (T)?V (g) ? > (? j ?? I ) 2 ?(? j ?? C ) 2 (2.11) where the left hand side represents the expected gain in utility from consump- tion if the incumbent is reelected, and the right hand side represents the expected loss in utility from ideological issues if reelection occurs. Since the vote of any individual depends on his position in the ideological space, we need to give more structure to the distribution of voter preferences to determine the share of votes an incumbent obtains (and hence his re-election probability ?) for a given choice of election period fiscal policy. Suppose, without loss of generality, that smaller ? j represent positions more ?to the left?, and ? I < ? C . Assume also that there are three ideological positions voters hold: e? j = {?? I ,? M = ? I +? C 2 , ?? C }.Voterswith? =?? I are the incumbent?s core voters, and they are so far left of center that they vote for the incumbent even if he is known to be of the desks type, that is, even if ?p 1 =0. Analogously, 67 voters with ? =?? C are the challenger?s core voters, and they are so far right of center that they vote for the challenger even if the incumbent is known to be of the people type 3 . In the middle are voters with ? = ? M , swing voters in that they are ideologically as close to one candidate as they are to the other. They therefore vote on the basis of the fiscal policy they expect to see from the candidates. They vote for the incumbent if and only if they believe he is more likely than the challenger to have high ?,thatis,i? ?p 1 (g I 1 ) > ?p.(If?p 1 (g I 1 )=?p,swingvoters are indi?erent between the two candidates, and vote to reelect the incumbent with some probability r. We postpone further analysis of this case for section 2.2.4, where we study the equilibrium.) The crucial point is that swing voters may be led to vote for the incumbent by high pre-election targeted expenditure, even though they know that he may be a desks politician who provides such expenditure solely to increase his chances of being reelected. We summarize the behavior of voters in Proposition 3 In an election between the incumbent and a challenger, the op- timal voting strategy of an individual j is given by: 1) If ? j =?? I individual j votes for the incumbent with probability 1 2) If ? j =?? C individual j votes for the challenger with probability 1 3) If ? j = ? C +? I 2 individual j votes for the incumbent with probability r(g t ), 3 Formally, using (2.11), one may derive ?? I < ? M ? ?p[V(T)?V(g)] 2(? C ?? I ) and ?? C > ? M + (1??p)[V(T)?V(g)] 2(? C ?? I ) 68 where r(g t )=1 if ?p 1 (g t ) > ?p r(g t ) ? [0,1] if ?p 1 (g t )=?p r(g t )=0 if ?p 1 (g t ) < ?p where ?p 1 (g)=0,and?p 1 (T)= ?p ?p+(1??p)q Given the voting strategies in proposition 3, election outcomes are easy to characterize. Let ? I, ? C and ? M , be the fraction of voters with ? j equal to ?? I , ?? C , and ? M , respectively, where we assume that ? I < 1 2 and ? C < 1 2 .Theelection is decided following a simple majority rule 4 . Hence, no candidate can win the election without getting the votes of at least some swing voters 5 , and a candidate supported by all swing voters wins the election. The incumbent obtains ? I of the votes if ?p 1 < ?p, ? I +r? M if ?p 1 =?p,and? I +? M of the votes otherwise. In other words, the incumbent is re-elected if ?p 1 > ?p or if ?p 1 =?p and ? I + r? M ? 0.5. For the time being, we assume that both voters and politicians have perfect information about ? I , ? M ,and? C . Translating this into election outcomes as a function of g t ,theassumption that no group of core voters is a majority implies that an incumbent who chooses 4 For completeness, we will suppose that a tie is resolved in favor of the incumbent, but this is not crucial to our results. 5 This follows from our assumption that ? I < 1 2 and ? C < 1 2 . The solution is trivial if any group of core voters is a majority: the election is decided beforehand, and therefore there is no incentive for the incumbent to increase transfers in the pre-electoral period. Fiscal policy is given by g t (?)=g t+1 (?), and swing voters vote for the incumbent if and only if they observe high pre-election transfers, but their vote does not a?ect the outcome of the election. 69 low pre-electoral targeted spending will not be reelected, since voters recognize him as being of the desks type. Hence, ?(g t = g)=0.Moreover,?(T)=1if swing voters prefer the incumbent (?p 1 (g t ) > ?p),since? I + ? M =1? ? C ? 1 2 . Finally, if swing voters are indi?erent between the two candidates (?p 1 (g t )=?p) then ?(T)=1if and only if indi?erent voters choose the incumbent with high enough probability, and there are enough swing voters. That is, if and only if ? I + r(g t )? M ? 0.5. 2.2.4 Equilibrium Armed with this knowledge of optimal strategies, we can now study the possible political-economic equilibria. We use the concept of Perfect Bayesian Equilibria. A pair of strategies (for the incumbent and voters) is an equilibrium if the voter?s strategy satisfies proposition 3 and the incumbent?s choice of g t satisfies propo- sition 2 given the voters? behavior. Note that these conditions are su?cient and necessary for an equilibrium since: 1) posterior beliefs in proposition 3, repre- sented by ?p 1 (g t ), obey Bayes? rule, 2) proposition 3 summarizes the strategy that maximizes a voter?s utility given his beliefs and the incumbent?s strategy, g t (? I ), and 3) proposition 2 summarizes the strategy that maximizes the incumbent?s utility. For simplicity of notation we describe the incumbent?s strategy in terms of the probability of choosing g t = T given his type, Pr(g t = T | ? I ),andswingvoter j?s strategy in terms of r(g t ), the probability that he will vote for the incumbent given g t . We have already discussed the voting strategies of core voters. Given the strategies in propositions 2 and 3, there are three possible types of equilibria: 70 Pooling equilibria: incumbents choose Pr(g t = T | ? I )=1independently of their type, and voters choose r(g t )= ? r ? 0.5?? I ? M if g t = T 0 otherwise ? . A separating equilibrium: each type of politician chooses the same policy they will choose in the post election period, and swing voters vote for the incumbent if and only if g t = T. Equilibria in mixed strategies: a desks incumbent chooses Pr(g t = T | ? I )= q ? (0,1) and swing voters vote for the incumbent if and only if g t = T. These three sets of strategies indeed constitute equilibria, since no player wants to deviate from his strategy, given the other?s. Note also that there are no pooling equilibria with r(g t = T) < 0.5?? I ? M , since then incumbent would be better o? deviating to Pr(g t = T | ? I = ?)=0. Proposition 4 describes the equilibrium outcomes for the case where swing voters shift the outcome of the election, depending on whether a desks politician gives more value to reelection or to providing low transfers (that is, whether ?(?p??+ ?) > ??). Proposition 4 Given ? I + ? M ? 0.5, there are three cases in equilibrium: Case 1) ?(?p??+ ?) > ?? The optimal strategy for the incumbent is Pr(g t = T | ?)=Pr(g t = T | ?)=1 The optimal strategy for swing voters is r(g t )= ? r ? 0.5?? I ? M if g t = T 0 otherwise ? Case 2) ?(?p??+ ?)=?? 71 The optimal strategy for the incumbent is Pr(g t = T | ? I )= ? 1 if ? = ? q ? [0,1) if ? = ? ? The optimal strategy for swing voters is r(g t )= ? 1 if g t = T 0 otherwise ? Case 3) ?(?p??+ ?) < ?? The optimal strategy for the incumbent is Pr(g t = T | ? I )= ? 1 if ? = ? 0 if ? = ? ? The optimal strategy for swing voters is r(g t )= ? 1 if g t = T 0 otherwise ? Proof. Note first that all of these sets of strategies constitute equilibria, since given the voters? strategy the incumbent?s satisfies proposition 2, and given the incumbent?s strategy the voters? satisfies proposition 3. Second, to prove that in each case only the type of equilibrium described exists, note that a separating (resp. pooling) equilibrium cannot be supported if ?(p?? + ?) > ?? (resp. < ??) because the incumbent would deviate to g t (?)=T (resp.g t (?)=g). Moreover, an equilibrium where the incumbent plays mixed strategies can only exist if he is indi?erent between the two policies, which happens i??(p??+?)= ??. Proposition 4 implies that, provided reelection is valuable enough to incum- bents, desk type politicians will choose g t = T with some positive probability. Meanwhile, in the post election period desk politicians choose g t = g with cer- tainty. As a result, the unconditional expectation of government?s targeted ex- penses is higher in the pre-election period, compared to their expected value for 72 other periods 6 . Conversely, non-targeted expenses are expected to be lower prior to an election than in other periods. In other words, fiscal policy exhibits cy- cles with the timing of elections. These cycles take the form of a change in the composition of expenditures, which shift towards targeted expenses in election times. Of course, a political budget cycle of this formwill only appear if the incentives toinfluencetheelectionarelargeenough. Therearetwo partsto this requirement. The first refers to the preferences of politicians: electoral manipulation of the budget will only arise if ?(p??+?) ???, so that the incumbent assigns a large value to being reelected. There is, however, an additional necessary condition, namely that swing voters (those whose voters depend on fiscal policy) can change the outcome of the election (? I +? M ? 0.5). The existence of a political budget cycle therefore depends on the political environment, in particular in the potential electoral benefit from increasing transfers. What is interesting about the apparently obvious need for a large fraction of impressionable voters is that, given the rational character of voters in our model, fiscal manipulation is less e?ective to ?buy? the vote of any single individual precisely in the cases where there are most swing voters. In this simplified setting, where our assumptions imply ?(g t ) is either 0 or 1, this is reflectedinthefact that ?p 1 (T | ? I + ? M < 0.5) = 1 ? ?p 1 (T | ? I + ? M ? 0.5). Note further that the assumption that ? = ? (and the implication that a fiscalexpansioninanelectionyearreflects mimicking by the ? politician, whom 6 The unconditional expectation of targeted expenditure is given by E(g t )= T [?p +(1? ?p)Pr(g t = T | ?)]+g(1??p)Pr(g t = g | ?) in the preelection period, and E(g t+1 )= E(g t?1 )=T ?p +g(1? ?p) in non-election times. 73 voters would not prefer if his type were known) is a convenient modeling device, rather than essential to the existence of the political cycle. If ? < ? , a cycle might take the form of signaling, in that the ? type would choose g in both elec- tion and non-election periods, while the ? type would choose g t just high enough to separate himself in an election period. Under the now familiar conditions that re-election is valuable enough, this g t is higher than this type?s g t+1 , and we have (qualitatively, at least) the same type of cycle. For instance, in Rogo??s (1990) model, pre-electoral manipulation arises from signaling by the good type (compe- tence signaling, since competence is the focus of his model rather than targeted spending). His approach has been criticized (unfairly, we think) in that it is the more competent candidate who engages in fiscal manipulation, rather than the less competent one. In a mimicking model it is the less desirable candidate who engages in fiscal manipulation. What is essential to our model (and to Rogo??s as well, from our perspective) is that at least some types of o?cials engage in this manipulation. Which types are susceptible to these incentives changes with the specific assumptions of the model, and is therefore not substantive or robust. In fact, in section 2.3 we extend the model to continuous types, and show that under certain conditions all types may engage in election-year manipulations of the budget. 2.2.5 Asymmetric information about the electoral envi- ronment So far we have assumed that all players have symmetric and perfect information about ? I , ? M ,and? C . This assumption is clearly not realistic, as the electoral e?ectiveness of providing targeted spending to voters is not known with certainty, 74 and candidates frequently hold more information about it than the public does. We will now relax this assumption, and show that the existence of asymmetric information about the political environment reinforces the incentives faced by in- cumbent o?cials to a?ect election outcomes through changes in fiscal policy. This will also eliminate the unsatisfactory feature that in some of the equilibria with electoral transfers (in particular the pooling equilibrium) voters are indi?erent between the challenger and an incumbent that targets them with spending. This is of course a result of the technical approach and the simplifying assumptions we have made, so we do not take the model as predicting that voters will strictly be indi?erent. However, it does open the question of how do individuals actually vote when they are ?indi?erent?, since one would not expect them to simply toss acointodefine which way to vote. In other words, one may question whether r(g)?(0,1) is meaningful in terms of the actual behavior of voters. To account for the possibility that candidates running for election know more than voters about how e?ective are targeted expenses to generate votes, we as- sume that the shares of core and swing voters are only known to the politician. In particular, we assume that voters assign a probability z that ? I +? M ? 0.5. In other words, voters assign a probability 1?z that the challenger?s core voters are a majority, in which case a desks incumbent would have no incentive to choose g t = T. Voters now characterize the behavior of the incumbent by Pr(g t = T)=?p + z(1? ?p)Pr(g t = T | ? = ?,? I + ? M ? 0.5) where, as before, ?p is the unconditional probability that the incumbent is not a desks politician. 75 After observing fiscal policy, voters update their beliefs about the incumbent?s type following Bayes? rule, as captured by equation (2.9). Their beliefs on the probability that a policy maker who chose high transfers is of the high type are now: ?p 1 (g t = T) ? Pr(? = ? | g t = T) (2.12) = ?p ?p + z(1? ?p)Pr(g t = T | ? = ?,? I + ? M ? 0.5) Given z<1 ,itisnowthecasethat?p 1 (g t = T) > ?p even if Pr(g t = T | ? = ?,? I + ? M ? 0.5) = 1. That is, it is now true the incumbent can lead swing voters to prefer him over the challenger by choosing high transfers, even if desks politicians are as likely to choose high election transfers as high-type politicians whenever ? I + ? M ? 0.5. The reason is simply that voters do not know whether the latter holds. The equilibria for this case are described in proposition 5 Proposition 5 In equilibrium, the optimal strategy for swing voters is r(g t )= ? 1 if g t = T 0 otherwise ? The optimal strategy for the incumbent is Pr(g t = T | ? I )= ? 1 if ? = ? q if ? = ? ? 76 where q =0if ? I + ? M < 0.5.If? I + ? M ? 0.5 then q =1if ?(p??+ ?) > ?? q =0if ?(p??+ ?) < ?? q ? [0,1) if ?(p??+ ?)=?? This type of imperfect information makes the problem more interesting, by capturing an additional inference problem for voters. Voters now need to make inferencesabout whethertheyarebeing targetedwithtransfersbecausethe politi- cian likes providing such transfers, or because transfers are very e?ective to raise votes. The fact that they assign some probability that the latter is not true gives more room for the politician to influence the outcome of elections by providing more targeted expenses prior to elections. 2.3 A generalization: letting politicians have fa- vorites The simple model discussed above illustrates the ideas of voters rationally re- sponding to pre-electoral manipulation and of the electoral shift of government resources toward targeted spending. This simplified setting, however, also raises questions that need to be addressed in a more general model. First, an unattrac- tive feature of that simple model is that the categorization of politicians into two types renders an equilibrium in which only one of the types engages in pre- electoral manipulation. Second, if electoral manipulation arises from targeting 77 spending toward projects most favored by voters, it would be more natural to represent voters as having heterogeneous fiscal preferences. In fact, our focus on targeted spending comes from the belief that voters have narrowly defined prefer- ences over government expenditure, and that they want fiscal policy in the hands of an o?cial who shares those preferences with them. It is our view that voters use pre-election fiscal policy not only to learn about the incumbent?s preferences in terms of targeted and non-targeted spending, but also to identify which candi- date should be expected to devote larger resources to the projects a given group of individuals is most interested in. Put in a di?erent way, an individual?s vote also depends on his beliefs about which groups of voters are most favored by the incumbent. A relevant question is, therefore, how does the incumbent allocate targeted spending across di?erent groups of voters. We present in this section a more general version of the model discussed above, in which we address the allocation of spending across groups of voters and consider a continuum of types of politicians. We assume that each voter has a taste for only one type of government ex- penditure. Groups of voters are defined according to preferences over types of expenditure: a group h is made up of people who prefer the same type of ex- penditure g h , and everyone in group h receives the same per-capita level of the expenditure. For simplicity, we consider only two di?erent types of targeted spending. This defines two groups of voters, h 1 and h 2 . Within each group, voters di?er in their preferences toward non-fiscal policies. That is, for each group there is a non-degenerate distribution of preferences over ideology; we denote this distribution as f h (?) for group h.Aswedidattheend of section 2.2, we assume there is asymmetric information about the e?ectiveness 78 of di?erent types of government expenditure in raising votes (i.e. how ?swing? specific groups of voters are). In particular, we assume that the incumbent knows f h 1 (?) and f h 2 (?), while voters only have imperfect information about them. We will first focus on the most general form of the problem, and will only later make specific assumptions about the amount of information about the f 0 s that is available to voters. 2.3.1 Voters Intermsofa voter?sproblem, theonly di?erence with respect tothe simpler model of previous sections is that now his consumption is indexed by h.Therefore,as in equation (2.8), a voter j in group h prefers the incumbent over the challenger if E ? V ? g h t+1 ? | I,g h t ? ?(? j ?? I ) 2 >E ? V ? g h t+1 ? | C ? ?(? j ?? C ) 2 (2.13) The indi?erent voter in group h may therefore be represented by the position e?(g h t ),defined by e?(g h t )= E ? V ? g h t+1 ? | I,g h t ? ?E ? V ? g h t+1 ? | C ? +(? C ) 2 ?(? I ) 2 2(? C ?? I ) (2.14) Since g h t a?ects the utility voters expect to receive if the incumbent is re- elected, the indi?erent position is a function of g h t . Suppose, without loss of generality, that ? I < ? C . Then, within group h, all individuals characterized by ? j < e?(g h t ) vote for the incumbent, while those with ? j > e?(g h t ) vote for the challenger. 79 We can then measure the fraction of group h votes obtained by the incumbent as a function of the pre-election expenditure observed by voters. Denoting this fraction as ? h (g h t ) and the lower bound of ? j as ?, we obtain: ? h (g h t )= Z e?(g h t ) ? f h (?)d? = F h ? e?(g h t ) ? (2.15) so that ? 0 h (g h t )=f h ? e?(g h t ) ? ?e?(g h t ) ?g h t (2.16a) = f h ? e?(g h t ) ? ? " ?E ? V ? g h t+1 ? | I,g h t ? ?g h t ? 1 2(? C ?? I ) # (2.16b) where the last equality uses equation (2.14). Note that groups di?er in the level of spending they receive, and as a result in the ideological position of the indi?erent voter e?(g h t ), as well as in the distribution f h .Wechoosethef h sto be smooth functions, so that at no point a marginal increase in e?(g h t ) can bring large masses of additional voters to the incumbent?s side. As a result of this and the concavity of V (), ? h is also concave. ? 0 h (g h t ) measures the electoral benefit to the politician from directing an addi- tional dollar to voters in group h. The size of this benefit depends on how much that additional dollar expands the range of ideological positions where voters prefer the incumbent, characterized by the position of the indi?erent voter e?(g h t ). If the utility voters expect to obtain under the incumbent in t +1increases, e?(g h t ) increases (that is, moves closer to ? C ) and the range of supporters for the incumbent expands. For a given change in expected utility, the increase of e?(g h t ) is smaller the further apart ? C and ? I are, as the cost to voters from having their least preferred ideological position in power becomes larger. 80 Besides the e?ect of an additional dollar on e?(g h t ), ? 0 h (g h t ) also depends on the mass of h voters at point e?(g h t ), which determines how many additional votes the incumbent obtains from increasing e?(g h t ). Thismassofvotersismeasured by f h ? e?(g h t ) ? . An important assumption here is that voters observe only the expenditures targeted to their own group. This is close to the visibility argument: voters care about certain types of government expenditure because these are the ones visible to them. Probably the most familiar expression of this is regional targeting of expenditures, where voters in one region do not directly observe the goods provided to another. 2.3.2 The incumbent?s problem We now write the incumbent?s objective in an election year as ? I t = 2 X h=1 ? h I V ? g h t ? + a(K t )+?? ? N I ? ? I t+1 + ? ? 1?? ? N I ?? ? I(P=C) t+1 (2.17) where ?, the probability of re-election, is now expressed as function of the fraction of votes the incumbent receives, denoted as N I . The incumbent assigns potentially di?erent weights to voters in di?erent groups, which is the reason why those weights are now indexed by h.Forany group h, we assume that the weight ? h I is drawn from a distribution m(? h ),and ? h can take any value in a continuous interval [? l ,? u ] (where all values are pos- itive). We assume that m(? h ) is the same for both incumbent and challenger (that is, ? h C is also distributed according to m(? h )). Wealsoassumethatthe values of ? l and ? u ,aswellastheformsofV () and a(),aresuchthatinthe post-election period politicians always choose interior solutions for K, g h 1 andg h 2 . Without this assumption we would need to address limit cases, but the intuition 81 of electoral manipulation and resource allocation across groups would still be the same. For tractability, we consider ?(N I ) as a continuous increasing function. The continuity of ?(N I ) is clearly inexact in a setting where elections are decided by some majority voting rule, but it simply implies that candidates try to maximize the number of supporters they have 7 . Moreover, we actually assume that ?(N I ) is a linear function of the form ??N I . Both of these assumptions imply that all votes are equally important, and ignore the fact that candidates mainly want to obtain a majority. This is certainly extreme, but it is not a crucial force behind our results about how electoral transfers are allocated across groups of voters. Assuming equal sized groups, the fraction of votes received by the incumbent, N I , is given by: N I = 2 X h=1 ? h (g h t ) Each period, the government faces the budget constraint: T = K t + 2 X h=1 g h t (2.18) 7 One of many reasons why this behavior can be optimal is a candidate?s uncertainty about turnout: some voters that prefer him over his opponent may not turn out to vote, so that it is optimal to have ample advantage. Another potential reason to justify the interest to maximize votes is that larger political support increases the outside value of the candidate. We do not formally model any of these reasons, but they provide a motivation for the assumptions regarding ?. 82 As before, we solve the politician?s problem backwards. Without loss of gener- ality, we focus on his targeting of a generic group h. If the incumbent is re-elected for the post-election period, he chooses g h t+1 accordingtothefirst order condition (FOC): ? h I V 0 ? g h t+1 ? = a 0 (K t+1 ) (2.19) A re-elected incumbent therefore chooses spending good h as an increasing function of the weight ? h I . As a result, the post-election utility an h voter in receives if the incumbent is re-elected is also increasing in ? h I . Let ? = ??? ? ? I t+1 ?? I(P=C) t+1 ? be the discounted value the incumbent expects to obtain from an additional vote from group h, which we take as given, and remember ? h (g h t ) is the share of group h?svotesthatgoestotheincumbent. For the election period, the incumbent?s optimal choices are summarized by the following proposition: Proposition 6 The incumbent?s choice of policy in the pre-election period sat- isfies the budget constraint (2.18) and ? h I V 0 ? g h t ? + ? 0 h ? g h t ? ? = a 0 (K t ) (2.20) for h = h 1 ,h 2 . The left hand side of FOC (2.20) represents the benefit from a marginal in- crease in g h t . As in the post-election period, this benefit includes the utility gain this change induces for group h voters. However, prior to an election the politi- cian potentially derives an additional benefit from targeting group h,namely obtaining more votes from this group?s voters. 83 Note that a similar FOC applies for spending targeted to the other group. Optimal choices of g h 1 t and g h 2 t therefore also satisfy: ? h 1 V 0 ? g h 1 t ? + ? 0 h 1 ? g h 1 t ? ? = ? h 2 V 0 ? g h 2 t ? + ? 0 h 2 ? g h 2 t ? ? (2.21) The first important result is that targetedspending increases the share of votes that goes to the incumbent, despite the fact that voters recognize the electoral incentives faced by the incumbent. The following proposition states and proves this result. Proposition 7 For any group h, ? 0 h ? g h t ? > 0. Proof. Suppose ? 0 h ? g h t ? ? 0. The incumbent would then get more votes by reducing, or at least not increasing, targeted spending to group h. Larger g h t in this case cannot be driven by electoral motives, but by ? h I being high. Increases in g h t then lead voters in h to perceive higher ? h I and expect higher post-election utility. As a result, more group h voters want to vote for the incumbent, that is, ? 0 h ? g h t ? > 0. This contradicts the initial assumption. The intuition behind this result is as in the simple model of previous sections. High spending to a group could reflect electoral motives or a genuine preference of the incumbent for that group. Voters cannot distinguish the true force behind high observed transfers, and therefore assign some probability to being targeted due to actual fiscal preferences rather than electoral reasons. With respect to the post-electoral allocation of expenditures there is a pre- electoral shift of government resources away from desks and into targeted spend- ing. In other words, K t 0 with the fact that K t+1 satisfies the post-election FOC (2.19). Given these two 84 elements, if the incumbent chose K t = K t+1 the pre-election marginal benefit of targeted spending would exceed that of desks.Since? h (), V (), and a() are all concave (with a() possibly weakly concave), satisfying the pre-election FOC (2.20) therefore requires lower non-targeted expenditure before the election. The pre-electoral shift of resources toward targeted spending holds for any realization of ? h 1 I and ? h 2 I , so that all types of politicians have incentives to change the composition of expenditures prior to an election 8 . The main reason for the di?erence with respect to the results above is that m(?) is continuous, so that all politicians want to increase targeting to separate from marginally lower types. In this sense, the result of electoral manipulation by all types of politicians reflects a more general representation of fiscal behavior. Note that another di?erence with the results from above is that here ?, the value of re-election, plays no role in determining whether an incumbent follows electoral incentives in designing the budget (although it determines the extent of the manipulation). This results from ignoring the discontinuous character of the ?(g h t ) function, and should therefore not be taken literally. The more interesting question to address with this generalized framework is how electoral motives change the allocation of resources across groups in the pre-election period, compared to non-election periods. We will provide here an intuitive discussion of how these resources are allocated. For this purpose, it is useful to establish the post-election choice of fiscal policy as a benchmark against which pre-election outcomes can be compared. This is the relevant benchmark 8 This is true as long as incumbents always choose interior solutions for K, g h=1 and g h=2 in the post-election period. As mentioned above, we have assumed that this is the case to avoid dealing with the limit cases. 85 because we are interested in comparing the composition of spending chosen before the election to what would be chosen in the absence of electoral incentives. Since we have designed the t+1period to di?er from t only in the electoral incentive, the political budget cycle is given precisely by the di?erence between the t and t +1fiscal choices. We will therefore start from the possibility that the t +1 composition of spending is imposed in the t period. The key determinant of how pre-electoral transfers are split between the two groups is their relative potential to generate additional votes to the incumbent. That is, the ratio of ? 0 h 1 ? g h 1 t+1 ? to ? 0 h 2 ? g h 2 t+1 ? is crucial. Suppose also that, at the t+ 1 levels of targeted spending, the incumbent would derive more electoral benefits from targeting h 1 voters than from targeting h 2 voters. That is, ? 0 h 1 ? g h 1 t+1 ? > ? 0 h 2 ? g h 2 t+1 ? . Since K t+1 ,g h 1 t+1 and g h 2 t+1 satisfy the FOC (2.19), and ? 0 h (g) > 0, the following relations hold: ? h V 0 ? g h t+1 ? + ? 0 h ? g h t+1 ? ? >a 0 (K t+1 ) for h = h 1 ,h 2 and ? h 1 V 0 ? g h 1 t+1 ? + ? 0 h 1 ? g h 1 t+1 ? ? > ? h 2 V 0 ? g h 2 t+1 ? + ? 0 h 2 ? g h 2 t+1 ? ? That is, if the t +1composition of spending is imposed in t,themarginal benefit of expenditures targeted to any group exceeds that of K,andthebenefitof directing one more dollar to g h 1 t exceeds that of directing it to g h 2 t . The incumbent then has incentives to take one dollar out from non-targeted expenditures K, and put it into g h 1 , the most valuable form of targeted spending, while keeping g h 2 unchanged. This will increase the marginal benefit of desks (non-targeted 86 spending), given the concavity of a(K). What happens to g h 2 t and the final e?ect on K t depend on the relative distance between ? 0 h 1 ? g h 1 t+1 ? and ? 0 h 2 ? g h 2 t+1 ? . Case 1: ? 0 h 1 ? g h 1 t+1 ? and ? 0 h 2 ? g h 2 t+1 ? are ?close?. Then resources will only be shifted from K into g h 1 until g h 1 = g 0 ,whereg 0 satisfies ? h 1 V 0 (g 0 )+? 0 h 1 (g 0 )? = ? h 2 V 0 ? g h 2 t+1 ? + ? 0 h 2 ? g h 2 t+1 ? ? From this point, resources from K wouldberedirectedtowardbothtypesof targeted expenditure, now equally valuable, until the FOC?s (2.20) and (2.21) are simultaneously satisfied. Compared to the post-election period, the equilibrium composition of spending before the election would involve lower K t and higher targeted transfers to both groups. Case 2: ? 0 h 1 ? g h 1 t+1 ? and ? 0 h 2 ? g h 2 t+1 ? are ?far ? from each other.Inthis case resources are shifted from K t towards g h 1 t until K = K 1 ,whereK 1 satisfies ? h 2 V 0 ? g h 2 t+1 ? + ? 0 h 2 ? g h 2 t+1 ? ? = a 0 (K 1 ) Then, g h 2 t and K t will both contract until the marginal benefits of all types of spending are equated. In this case, relative to the post-election, only the group that produces high electoral benefits will obtain larger targeted expenditure before the election. The other group will actually receive g h 2 t ? 0 h 2 (g t+1 ) in the graph, e f h 1 (g t+1 ) > e f h 2 (g t+1 ). In case 1 above, represented by Figure 2.1, while keeping g h 2 at g t+1 , resources are shifted from K into g h 1 until a point such as g 0 . At this point, the distance between f h 2 (g t+1 ) and f h 1 (g 0 ) has become quite modest, and it just compensates the utility gap that has been created between the two groups. As a result, it is valuable to start pumping resources into group h 2 as well. In case 2, however, the distance between the two distributions is much larger (Figure 2.2). The point g 1 , at which the marginal benefits of g h 1 = g 1 and g h 2 = g t+1 are equal, is much larger than g 0 . As a consequence, the incumbent is unwilling to reduce desks as much as necessary to provide g 1 to group h 1 and g t+1 to group h 2 . Rather than reducing desks until g h 1 =g 1 ,atsomegE ? ln? h C ? ? ? ? ? ? ? ? ? ? ? ? (2.25) Note that Y (g h t ) is the component of ? h0 ? g h I ? a?ected by voters?s expectations, so our analysis of their beliefs will focus on Y (g h t ). Also, ex-ante incumbent and challenger are identical, so ? h C follows the same unconditional distribution that characterizes ? h I . E ? ln? h C ? is formed according to that unconditional distribu- tion. Voters unveil the relationship between ? h I and g h I from the FOC (2.24), and use it to form expectations about the future. That relationship is given by: ? h I = ? ? ? ? ? ? ? ? ? g h t ? ? ?? h ?Y 0 (g h t ) ? if E ? ln? h I | g h t ? ? E ? ln? h C ? g h t ? ? + ? h ?Y 0 (g h t ) ? if E ? ln? h I | g h t ? ? E ? ln? h C ? ? ? ? ? ? ? ? ? ? (2.26) 95 It is clear from this expression that one key reason why voters respond to pre- electoral manipulation is their lack of information about ? h , which determines how attractive from the electoral point of viewis a given group. In this example, if the ? h were known to voters, they could perfectly infer ? h I from their observation of g h t ,andincreasesing h t would generate no electoral benefits to the incumbent. Voters form E ? ln? h I | g h t ? by taking logs of both sides of (2.26), and using Pr(? h =??)=p ?? h . Writing these expectations in terms of Y (g h t ),weobtain: Y (g h t )= ? ? ? ? ? ? ? ? ? e ?E ( ln? h C ) g h t ? ? 1?? ? ? Y 0 (g h t ) ? p ? h ? 1?? ? ? Y 0 (g h t ) ? (1?p ? h ) if g h t ? ?g e E ( ln? h C ) ? g h t ? ? 1+? ? ? Y 0 (g h t ) ? p ? h ? 1+? ? ? Y 0 (g h t ) ? (1?p ? h ) ? ?1 if g h t > ?g ? ? ? ? ? ? ? ? ? (2.27) where ?g is such that E ? ln? h I | g h t ? ? E ? ln? h C ? if and only if g t ? ?g. 9 This is the first order di?erential equation that characterizes voters? beliefs. Note that expression (2.26) represents the incumbent?s optimal choice of g h t given voters? ex- pectations, while expression (2.27) represents voters? rational expectations, given the incumbent?s actions. Equilibrium outcomes are therefore represented by a function Y (g h t ) that solves expression (2.27), and the choice of g h t that satisfies (2.26) for that Y (g h t ). Those equilibrium outcomes, which we illustrate below, are summarized in proposition 9. 9 The fact that E ? ln? h I | g h t ? is increasing in g h t was proved for the general case in previous sections. This example is, in any case, self-contained: we can consider the positive slope of E ? ln? h I | g h t ? as a conjecture, which will then prove consistent with the politicians? choices. 96 Proposition 9 In this example, voters? equilibrium expectations about the future are characterized by E(ln? h I | g h t ) ? ? ? ? ? ? ? ? ? ? =ln(g h t ?c 0 ) if g h t < ?g ? ? ?? ? g h t ? 2 c 3 ? ln h c 1 + c 2 R exp ? ? ? g h t ? 2 c 3 ? dg i if g h t > ?g ? ? ? ? ? ? ? ? ? ? (2.28) where c 0, c 1 ,c 2 and c 3 are constants which depend on ??, ? and p ?? h ,and ?g = e E(ln? h C ) ?c 0 Meanwhile, the incumbent?s optimal choice for g h t is given by ? h g h t = ? ? ? ? ? ? ? ? ? ? ? ??? h ??e ?E(ln? h C ) c 0 if E ? ln? h I | g h t ? ? E ? ln? h C ? ? + ? h ? ? c 1 ?2?g h t Y (g h t )c 3 ? if E ? ln? h I | g h t ? >E ? ln? h C ? ? ? ? ? ? ? ? ? ? ? ? (2.29) Proof. See appendix B. We can now illustrate this solution 10 . Take the following set of parameters: ? ? U [0.2, 1], T =1, ? =1.3, ? h 1 =1.93 (or ?? h 1 =0.3), ? h 2 =0.79 (or ?? h 1 =1), p ? h 1 =?? =0.5,and? =0.1.Thechoiceof? is consistent, for instance, with 11 ? =0.99, ? =1and ? I t+1 ?? I(P=C) t+1 =0.11, where the latter would be 10 Note that the solution for the upper branch of E(ln? h I | g h t ) is an approximation, since it involves linearizing the di?erential equation around the E(ln? h I | g h t )=E(ln? h C ) point (see appendix). 11 ? =0.99 corresponds to a discount rate of 0.01, which is consistent with historical records of quartely interest rates. 97 satisfied by combinations of ? h 1 I and ? h 2 I such as 0.3 and 0.9 or 0.5 and 0.45. These parameters imply ?g =0.53. The solution to the problem can be summarized by ? 0 (g h t ),thefirst order condition (2.24), and the resulting choice of g h I as a function of ? h I and ? h .We depict them in the following three figures. Figure 2.4 shows ? 0 (g h t ) forthetwogroups.Keepinmindthat? 0 (g h t ) repre- sents the additional h votes the incumbent can obtain from raising g h t one dollar. Thetoplineinthatfigure corresponds to the group with more swing voters, whichinthiscaseish 1 since it has the larger ? h .Thelargere?ect on votes for the more swing group is consistent with our previous result that electoral incentives to target swing groups are large, compared to more core groups. Note also that ? 0 (g h t ) is positive and (weakly) decreasing everywhere, reflecting the fact that the incumbent can always obtain more h votes by increasing g h t , but the electoral gain tends to decrease as g h t grows. In other words, the share of group h votes the incumbent obtains, given by ?(g h t ), is increasing and (weakly) concave. The positive slope shows the incentive for electoral increases in targeted spending. The concavity is a consequence of decreasing marginal utility, and less concentration of voters in the tails of the ? h distribution. In fact, note that the decreasing pattern of ? 0 (g h t ) is less pronounced for group h 2 (bottom line), which exhibits a ? h distribution with fatter tails. The incumbent?s choice of g h t is characterized by the FOC (2.24), which can be written as: ? h I g h t = ???? h0 ? g h t ? (2.30) 98 Figure 2.4: ? 0 (g h 1 t ) and ? 0 (g h 2 t ) This representation is useful because the FOC then looks very similar to the FOC for the post-election period. The only di?erence is the last term of the right hand side. We depict both the pre-election and the post-election FOC?s in Figure 2.5. The left hand side, ? h I g h , is given by the decreasing dotted curves for di?erent values of ? h I . From bottom to top, these curves correspond to ? h I =0.2, ? h I =0.4, ? h I =0.6 and ? h I =0.8. Meanwhile, the dashed horizontal line corresponds to the right hand side of the t+1FOC (which is given simply by ?). The two solid curves represent the right hand side of the period t FOC for the two groups: the bottom one represents the case of the more swing group (h 1 ),whichwealready noted exhibits the larger ? h0 for any g h t . Take, for instance, group h 1 . The incumbent?s optimal choice of g h 1 t is given by the intersection between the ? ? ?? h 1 0 (g) line (bottom solid line) and the ? h I g h curve. Meanwhile, his optimal choice of g h 1 t+1 is at the intersection of the dashed horizontal line and the same ? h I g h curve (since ? h I does not change between t and t +1). Note that, for any given ? h I , both groups receive larger targeted expenditures before the election than after it (g h t >g h t+1 for both h). In this case, 99 Figure 2.5: Incumbent?s first order conditions as discussed above, the constant marginal utility of desks precludes the possibility that one of the groups receives less targeted spending before the election than in t +1. Thesizeofpre-electoraltransfers(thedi?erence between g h t and g h t+1 ) is larger for group h 1 , characterized by a larger mass of swing voters. The di?erences between the two groups, however, become smaller for larger values of g h t ,since at these levels voters already perceive high benefits of choosing the incumbent (note that the two curves grow closer as g increases). In other words, given decreasing marginal utility, providing voters with additional expenditures in the high g region has only small e?ects in the well-being they expect to enjoy if the incumbent is re-elected. These results are reflectedinFigure2.6,whichshows the optimal choice of g h t as a function of ? h I . The extent to which pre- and post-electoral policy di?er (i.e. the size of the political budget cycle) obviously depends on the specificparameterschosen.For instance, larger values of ? imply a larger value of re-election, and therefore 100 0.2 0.4 0.6 0.8 0.2 0.4 0.6 ? h 0.3 0.5 0.7 g t h (a h =? ) g t h (a h =a) g t+1 h Figure 2.6: g h t (? h I ) and g h t+1 (? h I ) lead the incumbent to chose larger g h t . Small values of ? imply that the post- election level of targeted expenditure is already high (for any candidate) and, given decreasing marginal utility, reduce the potential di?erences between one and another candidate in terms of provision of targeted goods. This reduces the incentives for electoral increases of g h t . Larger ideological gaps between the di?erent candidates reduce the importance voters give to fiscal policy in choosing the candidate, and therefore reduce the incentives for electoral increases of g h t . Di?erent choices of ? h 1 and ? h 2 will change the electoral benefit the incumbent canobtainfromincreasingg h t , as can be deduced from the figures above. The general patterns of electoral changes for g h t , however, are quite robust to changes in the parameters. 101 2.4 Concluding remarks This paper has used the widely held view that politicians favor the interests of certain specific groups of voters to explain the emergence of political budget cycles in a model with rational voters. In the model, voters use past fiscal policy to learn information about which groups and types of spending the incumbent is likely to favor after the election, if he is re-elected. The result is a pattern of pre- electoral shifts of government resources from non-targeted types of expenditures toward goods specific groups of voters care more about. Pre-electoral transfers are mostly directed to undecided groups of voters, while groups that are committed to one or other candidate may even face negative transfers. A key feature of the model is that, even though voters are rational and predict this behavior, they respond to electoral manipulation, since they cannot observe whether they are being targeted because they are crucial to the incumbent?s re-election or because they are genuinely liked. Our view di?ers from other models of political budget cycles in that voters? care about the preferences of incumbents over di?erent interest groups, rather than his competence. This focus is motivated by traditional political practices in the management of the government?s budget, which suggest an important role for special interests in electoral budget manipulation. Although the idea of pork barrel politics is no strange to the political economy arena, it has not been previously incorporated in intertemporal models of fiscal policy-making. Furthermore, previous literature does not address the question of why providing pork spending would a?ect the votes of rational, forward-looking, individuals. 102 Chapter 3 On Political Budget Cycles and Voters as Fiscal Conservatives: Evidence From the Colombian Experience 3.1 Introduction Common wisdom is that, as elections approach, elected o?cials increase govern- ment spending to improve the chances that they or their parties will be reelected. It is not unusual to hear reports of pork spending growing in election years, and the public seems to expect popular spending projects when elections are im- minent. For economists, however, the debate about the existence, extent and characteristics of this manipulation is hardly settled. Among the many sticky points, there are inconsistencies between the idea of pre-electoral increases in the government?s budget and both the actual dynamics of government spending and the behavior of voters. First, evidence of election-year increases in govern- ment spending and the deficit is at best mixed. Second, increases of the size of the budget seem to hurt, rather than improve, an incumbent?s chances of being re-elected. Partly motivated by this empirical evidence, the model introduced in previ- ous pages presents a view where political budget cycles (PBC) take the form 103 of a change in the composition, rather than the size, of government spending. Before elections, specific groups of voters are targeted with spending, while non- targeted types of expenditure contract. Behind the model is the idea that voters dislike deficits, but like receiving goods specifically targeted to them. Voters re- act to fiscal policy because it provides information about the amount of targeted expenditure the incumbent will provide in the future if re-elected. This view of electoral cycles suggests that empirical analyses of the PBC that focus solely on the dynamics of the overall budget ?as is the case in most of the literature? are at risk of misinterpreting the evidence, besides missing an impor- tant part of the action. For one, electoral manipulation of government expendi- tures may occur without impacting the overall budget or the deficit. Moreover, the e?ect of elections on the dynamics of government spending is inextricably linked to the preferences of voters. These may vary from country to country, and learning about them is crucial for an adequate interpretation of the evolution of the budget around election times. Put simply, even contractions of government expenditure may be consistent with electoral incentives, if they are what the preferences of voters dictate. In this chapter, I use data on government expenditures and electoral outcomes in Colombia to examine the characteristics of PBCs, in terms of both voting be- havior and government spending 1 . I ask whether election outcomes are consistent with voters having di?erent preferences toward di?erent types of government ex- penditure. I also examine if it is rational for voters to believe that fiscal choices reflect persistent characteristics of o?cials, an assumption behind most models 1 This paper focuses solelyon the electoral manipulation of government expenditures, without dealing with potential electoral cycles in the revenue side. 104 of PBCs with rational voters. From the point of view of policy, meanwhile, I look at evidence on pre-electoral changes of government expenditures. Special emphasis is put on the separate analysis of di?erent types of government expen- diture, since both the popular Rogo? (1990) model and the model introduced in chapter 2 suggest the potential importance of composition e?ects. I finally tackle the question of whether the extent of electoral e?ects on fiscal policy depends on the degree of party polarization among voters. This chapter is organized as follows. I begin by discussing some relevant em- pirical literature in section 3.2. Section 3.3 comments on some interesting features of the Colombian case. The pre-electoral dynamics of government spending are analyzed in section 3.4, while section 3.5 studies the e?ect of fiscal behavior on election outcomes. These two sections introduce the data and tools necessary to examine the e?ect of party polarization on the electoral budget cycle, which is done in section 3.6. In section 3.7 I analyze whether the dynamics of expenditure are consistent with the idea that the fiscal preferences of o?cials exhibit some in- ertia. While all other empirical sections focus on Colombian municipalities, some evidence at the central government level is introduced in section 3.8. Finally, section 3.9 provides some concluding remarks. 3.2 A discussion of previous literature Two bodies of past empirical work are relevant for the questions I address in this chapter: literature on fiscal preferences of voters, and literature on the dynamics of fiscal policy around election times. The first of these branches of work examines how fiscal policy a?ects the incumbent?s chances of being reelected. Brender (2003) uses data on the elections 105 of local mayors in Israel, and estimates a model where the probability of reelection of the mayor is a function of his budget choices. He finds that voters penalize election year increases in deficits, but also that they reward high expenditure in development projects. Peltzman (1992) shows that, in the US, the share of votes received by the incumbent?s party is decreasing in government current (as opposed to capital) expenditures. This result, however, loses power if investment in roads, an important component of public investment, is included in the policy variable 2 . For OECD countries, Alesina, Perotti, and Tavares (1998) find that governments that adopt tight fiscal policies do not su?er falls in popularity and are not penalized by voters in the polls. In short, this literature suggests that elected o?cials do not receive electoral benefits from boosting spending before elections. If anything, the opposite seems to be true. However, not all types of government spending generate the same opposition: some development projects actually appear to increase political support for the incumbent. A second group of relevant empirical papers examines the dynamics of fiscal policy, looking for systematic changes that coincide with election times. The most comprehensive studies are those by Schuknecht (1994), Shi and Svensson (2000), Persson and Tabellini (2002), and Brender and Drazen (2003). Except for Schuknecht?s paper, this work focuses solely on aggregate measures of fiscal policy: total spending 3 ,taxrevenue,anddeficits. Using data for 35 developing 2 The author interprets the ?odd findings? obtained when including expenditure in roads as a result of the high lumpiness of this component. An alternative interpretation, consistent with the model introduced above, is that roads are clearly targeted public goods and voters react favorably to being targeted by the incumbent. 3 The measure of spending used in Shi and Svensson?s paper is actually government consump- tion spending from national accounts. I find results with this measure di?cult to interpret as 106 countries over the years 1970-1992, Schuknecht (1994) finds that overall fiscal balances worsen during an election year, due to an increase in total government spending. Shi and Svensson (2000) study a sample of 123 developed and underde- veloped countries over a horizon of 20 years. They find that deficits grow before elections in both developed and developing countries, but the electoral e?ects are particularly strong in developing economies. Moreover, spending shows pre- electoral increases in their sample of developing countries, but not in developed economies. Later results, however, are not so favorable to the existence of electoral cycles in total government spending. Brender and Drazen (2003) show that the find- ing of PBCs in total spending is driven solely by new democracies: there is no evidence of electoral manipulation in countries with a long history of democratic institutions. They argue that fiscal cycles in new democracies reflect underdevel- oped media and poor accounting practices. Persson and Tabellini (2002), with data for 60 countries for 1960-1998, study if PBC?s vary across political systems. They find no e?ect of elections on government expenditure or surplus for the over- all sample, and a pre-electoral contraction of government spending in countries with majoritarian electoral rules. In sum, the evidence in favor of pre-electoral increases in total government spending and government deficits is, at best, mixed. Furthermore, it seems that politicians only engage in spending hikes when vot- ers cannot e?ectively monitor government balances, a behavior that is consistent with voters being opposed to raising overall spending. In terms of the composition of expenditures, Schuknecht finds that, prior to it includes di?erent levels of government (local, national, publicly owned enterprises), which should respond to di?erent types of elections and in di?erent ways. 107 elections, capital expenditures rise as a share of both GDP and overall expendi- ture. Kneebone and McKenzie (2001) find no pre-electoral increases in aggregate spending for Canadian provinces, but do find that what they call ?visible expen- ditures?, mostly capital expenses such as construction of roads and structures, grow in election periods. They also find that spending in social services, in- dustrial development, and health actually contract before elections. Very similar findings are reported for Mexico by M. Gonz?lez (2001), who also finds that other categories of spending, such as current transfers, contract prior to elections. In short, pre-electoral manipulation of the budget is concentrated in some specific categories of government spending. For the countries covered in these studies, it would appear that capital expenditures are seen by politicians as an e?ective way to impress voters, and that o?cials attempt to take advantage of this fact 4 . Moreover, o?cials seem to find ways to increase these expenditures prior to elec- tions without increasing overall spending, thus avoiding being penalized by voters who are fiscal conservatives. The picture that emerges from this findings is a confusing one, at least in light of the widespread idea that o?cials expand government expenditure prior to elections. On one hand, voters seem to penalize such increases, and there is no robust evidence that overall spending is manipulated in this way. On the other, expenditure increases are actually observed for some types of spending. The main contribution of this paper is to put these apparently contradictory pieces together and show how they can be reconciled. In contrast to previous work, I analyze 4 Alesina et al. (1998) assert that ?cuts in public invesment are less visible and [politically] costly [than cuts in other spending] ?, but provide no evidence that this is the case. The evidence just discussed, as well as the results I present in this paper, point in the opposite direction, at least for the countries covered by these studies. 108 both voting behavior and the dynamics of government budgets as two parts of the same problem, and focus on the idea that not all types of government spending should be treated equally in this analysis. My results suggest that the model put forward in the previous chapter can serve as a unifying framework with which to understand the diverse findings from these branches of the empirical litera- ture. This paper also proposes that some specific components of the government accounts are more likely than others to reflect what the previous chapter calls targeted expenditure, that is, expenditure that generates large e?ects in electoral support. In that spirit, another contribution of this paper is a systematic analy- sis of di?erential e?ects of elections on the various components of overall public spending. 3.3 The Colombian case Colombia o?ers an interesting case for the study of these issues, as an example of a developing economy with a relatively well established democracy 5 . Electoral cy- cles in the overall budget have long been considered a phenomenon of developing countries. This is partly because of the idea that incumbents increase spending to boost economic activity, a tool that is considered specially powerful in less de- veloped economies. However, if the engine behind PBC is the use of fiscal policy 5 The statement that Colombia is a ?well established democracy? may be puzzling for the reader, in view of the intense armed conflict that has bled the country for years. However, from the point of the institutional regime, Colombia has enjoyed the rule of democracy practically without interruption since the 19th century. This is not a minor achievement in the Latin American context, where most countries went through long and painful periods under the rule of dictators, even in the last decades of the 20th century. 109 as a signal about characteristics of the incumbent (as theoretical models suggest) the crucial distinction is not between richer and poorer countries, but between older, well established, democracies and their younger counterparts. This is so because voters in an established democracy have greater ability to monitor the fiscal choices of o?cials. In such democracy, therefore, the fiscal preferences of voters should impose constraints on the specific form the PBC takes. If vot- ers are fiscal conservatives, we should not observe pre-election deficit increases in an old democracy, even if the country lags in economic development (this is consistent with the evidence in Brender and Drazen, 2003). The contrast with the traditional view of ?unconditional? manipulation of the budget in developing economies, makes countries like Colombia a particularly interesting ground to examine the link between the preferences of voters and electoral manipulation of government spending. Although I do present some evidence on the central government?s budget, the main focus in this paper is on spending behavior by local governments. I choose this ?cross-district? approach, rather than the more usual cross-country strategy for two reasons. First, the PBC model of chapter 2 suggests the importance of distinguishing between targeted and non-targeted types of expenditure. This distinction is most relevant at the local level, where expenses can be targeted most e?ciently. Second, the extent of cross sectional di?erences in institutional settings is much larger in a multi-country setting than it is for cross sectional units within the same country. This is specially clear with respect to constitutional rules, national laws, electoral and judicial systems, and monetary policy, all of which are important determinants of the existence and strength of political budgetary cycles. Most important of all, the view in this paper is that the fiscal preferences 110 of voters play a key role in determining the PBC. This creates the potential for widely di?erent forms of electoral budget manipulations from country to country, and suggests the convenience of limiting the analysis to a single country, where general features of the political system that are di?cult to control for do not vary. Note that the within-country strategy does not wipe out the sources of variation in the institutional environment that are necessary to identify key characteristics of the PBC; in particular, the degree of ideological commitment to one or other party does exhibit wide variation across districts of the same country. I should point that, although in Colombia the direct reelection of incumbent executive o?cials is banned, pre-electoral manipulation of fiscal finances is re- garded as a usual political practice. PBCs are thought to arise in Colombia largely due to the actions of the legislative bodies, whose members are in fact subject to direct re-election (in the case of city councils), or at least have found ways to circumvent formal restrictions to run for direct re-election (as in the national Congress). There are also reasons why even an incumbent mayor, who cannot run for re-election immediately, would want to manipulate fiscal policy at the end of his period in o?ce. Most importantly, voters identify the preferences of the o?cial with those of his party, and therefore the policy decisions of an o?cial are interpreted as signals of party preferences and competence. Hence, the incumbent knows that his decisions a?ect his party?s re-election chances. 111 3.4 Theelectoral dynamicsof government spend- ing In this section, I analyze evidence of pre-electoral manipulation of public spending in Colombian municipalities, with special emphasis on contrasting di?erent types of expenditure. In particular, the model introduced in previous pages suggests important di?erences between targeted expenditures and other components of the budget. I try to unveil such di?erences using the disaggregate components of government accounts. Aclassification of government expenditure into targeted and non-targeted expenses is not readily available, or straightforward. In fact, all government expenses (probably with the exception of interest payments on external debt) generate benefits for at least some groups in society, even if it is only to those individuals who provide the services and goods to the government. However, my view is that some of the components of expenditure that governments report separately, in particular most categories of investment expenditures, are more likely to reflect what I call targeted expenses than others. Opportunistic targeted expenditures, close to the familiar concept of pork spending, are most often associated with projects of infrastructure development: construction of roads, schools, water plants. These are highly visible expendi- tures that benefitspecific (yet potentially large) groups of voters. In Kneebone and McKenzie?s (2001) words, infrastructure spending fits the ?caricature of the opportunistic politician building roads, hockey rinks and schools just prior to elections?. On the other hand, some current expenditures, such as purchases of supplies and services and payments to other governmental entities, can be pre- 112 sumably cut without visibly hurting large groups of voters. Data that separates current government spending from expenditures linked to development projects could therefore fit the need for distinguishing targeted from non-targeted expen- ditures, and would also distinguish visible from non-visible spending, which is relevant in some models of PBC (e.g. Rogo??s 1990 model). I use a panel of yearly data on government accounts for all municipalities in Colombia over the 1984-2000 period. Expenditures are reported at a relatively high level of detail, allowing me to discern the behavior of di?erent types of spending. Following much of the literature, I estimate equations in which the policy variables are represented as functions of the timing of elections, as well as other controls. The basic relationship can be written as: y it = a i + b?y it?1 + X k c k ?x k,it + d?elecdum it + ? it (3.1) where i is an index for districts, y it is some specifictypeofspendingbythelocal government of city i in period t, a i is a district e?ect, and the x are control variables (indexed by k to allow the use of more than one control). The variable elecdum is a political dummy that captures the timing of elections, and it is the central variable in the analysis. This variable takes a value of 1 in periods preceding local elections, and 0 in all other periods. The error term is white noise. The autoregressive form is used in the literature on political cycles as a parsimonious representation of the policy choices, given the lack of elements to incorporate a fully structural model of fiscal policy. However, I also include additional controls to account for as much variability in the data as I can. I estimate a separate regression for each type of government expenditure (that is, each type of government expenditure is a di?erent y). In all regressions, the main 113 interest is d,thecoe?cient that captures the e?ect of elections. The traditional view of political budget cycles is that we should observe pre- electoral increases in overall spending and at least some of its disaggregate cat- egories (d>0 for at least some y 0 s, d<0 for no category). This is the kind of evidence previous empirical studies have attempted to find. Theoretical models, however, suggest the possibility that the PBC takes the form of a change in the composition, rather than the size, of the budget. In Rogo??s (1990) model of competence signaling, as an election approaches incumbents expand the provi- sion of visible goods and contract that of less visible goods. For a given level of competence of the incumbent, this implies more spending in visible projects and less spending in other government tasks. Meanwhile, the model introduced in the previous chapter postulates a pre-electoral shift in government resources away from non-targeted spending and into targeted projects. Given the discus- sion above, both models would be consistent with some components of current spending contracting prior to elections, with a simultaneous expansion of cate- gories related to development projects 6 . 6 In Rogo??s paper visible spending is called government ?consumption?, while the less vis- ible good, which the author identifies as spending in national defense and financial activities, is referred to as ?government investment?. When taking the model to the data, this choice of words may be misleading, at least for some countries. In the Colombian case, less visible expen- ditures such as defense, payments to pensioned employees, and o?ce supplies are all recorded under the consumption or ?current spending? categories. Highly visible types of projects, like the construction of bridges, schools, and water plants, are all under the ?investment? heading. The multi-period character of these projects raises a question about whether politicians are able to time them so that voters observe them before the election. Common wisdom, the existing empirical evidence, and the evidence I present here all seem to suggest that they are. 114 3.4.1 Data The data consist of annual observations for each Colombian municipality (close to 1100 cross-sectional units) for the period 1984-2000. Of the 18 years in the sample, 6 are local election years, when mayors and city councils are elected 7 . Elections occur at predetermined dates, and all cities hold elections the same day. Mayors have been elected by popular vote only since 1988. In previous years, the Colombian administrative system was highly centralized, with the president appointing governors and governors appointing mayors. From the late eighties, however, a gradual decentralization process has taken place both in the fiscal and the administrative fronts. Since 1988 mayors and governors have been elected by the people, and local o?cials now have many fiscal responsibilities that in previous decades were in the hands of the central government. The timing of the election dummy is such that the pre-election period is the year previous to the election if the election takes place in the first semester, and the year of the election if the election is held in the second semester. In terms of government spending, I use data from the Colombian Contralor?a General, a public entity with the task of monitoring public finances. Since the use of these data is a novel feature of this paper, I will discuss features of the database in some detail. The financial reports local governments file with the Contralor?a contain a detailed description of the revenues and expenditures of the regional government, so disaggregate measures are available. The general structure of the expenditure accounts available is summarized in the first column of Table 3.1 (the label I use for each account below is listed in the second column). Note 7 Mayors and councils are elected simultaneously. A list of the local elections held in the 1988-2000 period is presented in the appendix. 115 Type of Expenditure Name Total Expenditure TOTAL 1. Current Expenditure CURRENT 1.1. General Payments GENERAL 1.2. Personnel Expenditure PERSONNEL 1.3. Current Transfers TRANSFERS 2. Investment INVESTMENT 2.1. Infrastructure INFRASTRUCTURE 2.2. Water, Energy, and Comunications POWER 2.3. Housing HOUSING 2.4. Education EDUCATION 2.5. Health HEALTH 2.6. Others OTH_INV 3. Interest payments INTEREST Table 3.1: Composition of spending that the Colombian government accounts designate as ?Investment? items that in many other countries are known as capital expenditure 8 . Besides the categories listed in Table 3.1, I also examine investment in roads, which is asubcomponent of infrastructure investment. The reason for highlighting this specificsubcomponent is that construction of roads is regarded as a particularly popular way to raise votes prior to an election 9 . Infrastructure development projects show up mostly in the investment cat- egories. Given the discussion above, I associate these categories with targeted expenditures. Not all districts report all levels of disaggregation, or in all years. I use all 8 In the category ?Infrastructure? I include expenditures listed under roads, urban infras- tructure, and construction of market places. 9 For instance, there are several anecdotes in Colombia about mayors who insisted in inau- gurating bridges not fully finished because their terms were nearing an end. The result were poor quality bridges that ended up causing waves of accidents, or even collapsing. 116 the available information, with numbers of observations as reported in Table 3.2. Current expenditure and its broader subcategories, as well as total investment, are available for more than 90% of the districts in years prior to 1997, and for close to 80% of the districts after that. The disaggregation of investment is only available since 1990 10 . Table 3.2 presents summary statistics for the di?erent categories of spending I analyze. For each type of expenditure, statistics in the first row refer to all periods, statistics in the second row are for pre-electoral periods, and statistics in the third row are for other periods. All measures are in 1998 prices. Notice that most current expenditure categories display lower averages in pre-electoral periods that in other periods, while the opposite is true for most investment categories, in particular those associated with the development of infrastructure (infrastructure, water and energy, housing). These observations suggest pre- electoral changes in the composition of spending, in directions consistent with the predictions of both the Rogo? (1990) model and the budget composition model of chapter 2. A more formal analysis is undertaken in the following section. The dynamics of di?erent categories of government spending can be seen in Figures 3.1 and 3.2. The former contrasts current spending, and its subcompo- nents, with investment. The latter shows the di?erent categories of investment, for the period in which disaggregated data is available (1990-2000). Vertical dot- ted lines indicate pre-election years, as defined by the variable elecdum.Note, in particular, peaks in pre-election periods for infrastructure, road construction, 10 When disaggregations are reported, I check for consistency between reports of total invest- ment and the sum of its disagregate categories. I do the same for current spending and its subcategories. In both cases, inconsistencies arise in less of 1% of observations, and I discard those observations with inconsistencies. 117 4 5 6 7 8 9 10 11 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 L og e xpe ndi t ure Current Current transfers Investment Personnel General Figure 3.1: Evolution of spending: broad categories housing, and investment in energy and water plants. 118 Type of Number of Standard expenditure observations deviation All 12,318 56,502 611,645 TOTAL Pre-electoral 5,289 53,490 586,284 No pre-electoral 7,029 58,769 630,088 All 12,317 19,872 185,560 CURRENT Pre-electoral 5,288 18,535 177,053 No pre-electoral 7,029 20,877 191,717 All 12,248 4,071 21,020 GENERAL Pre-electoral 5,239 3,853 21,924 No pre-electoral 7,009 4,234 20,317 All 12,249 9,767 82,734 PERSONNEL Pre-electoral 5,240 9,202 80,848 No pre-electoral 7,009 10,190 84,120 All 12,230 5,900 91,405 TRANSFERS Pre-electoral 5,229 5,074 78,942 No pre-electoral 7,001 6,518 99,703 All 12,318 30,151 382,388 INVESTMENT Pre-electoral 5,289 29,187 365,595 No pre-electoral 7,029 30,876 394,578 All 5,265 3,173 8,257 INFRASTRUCTURE Pre-electoral 2,005 3,526 7,375 No pre-electoral 3,260 2,955 8,750 All 7,451 2,617 7,264 ROADS Pre-electoral 2,837 2,943 6,865 No pre-electoral 4,614 2,417 7,492 All 5,563 3,707 6,168 POWER Pre-electoral 2,111 4,273 7,103 No pre-electoral 3,452 3,360 5,490 All 7,351 762 4,073 HOUSING Pre-electoral 2,794 882 4,622 No pre-electoral 4,557 688 3,695 All 7,455 3,612 5,523 EDUCATION Pre-electoral 2,840 3,882 5,827 No pre-electoral 4,615 3,446 5,322 All 7,455 2,710 5,010 HEALTH Pre-electoral 2,840 2,934 5,262 No pre-electoral 4,615 2,572 4,844 Periods Mean Table 3.2: Summary statistics for di?erent types of expenditure 119 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 L og e xpe ndi t ur e Roads Power & Water Infrastructure Health Education Housing Figure 3.2: Evolution of spending: investment categories Table 3.3 lists the di?erent control variables I use in alternative specifications. Note that most controls are timed in the previous year, since I expect these past outcomes to influence the incumbent?s fiscal choices in the current period. This timing should also address concerns relating potential endogeneity of these regressors. Iusedi?erent specifications, with alternative sets of controls, to analyze the robustness of the results. Specification (1) includes state per capita GDP to con- trol for economic activity (GDP_PC), a quadratic time trend (T) 11 ,andsome 11 I use trend variables, rather than time-specific dummies, because of the di?culty of con- ceptually separating PBC e?ects from other year e?ects (besides, obviously, the mechanical impossibility of identifying separate coe?cients for all years plus the electoral dummy). For instance, there are arguments, beyond the scope of this paper, that would imply PBC-related e?ects on spending in years after elections. The di?culty with this data is that, because of 120 social indicators that could be used as inputs in fiscal policy decisions. The lat- ter include population and a poverty indicator known as Unsatisfied Basic Needs (UBN). Specifications (2) and (3) use alternative financial indicators, trying to account for the financial constraints faced by local governments. These con- straints are particularly important in later years, when the law has required that regional governments in Colombia obtain authorization from the central level to increase expenditure if they have been running deficits in previous years. I use deficit, debt, and fiscal dependence indicators, which were all constructed by me from the Contralor?a data. The Fiscal Dependence indicator, included in both specification (2) and (3), accounts for the level of fiscal decentralization in the country, which grew dramatically over this period. It is increasing in the share of revenues represented by transfers from the central government (as opposed to the local government?s own fiscal e?ort). Finally, in specification (4) I include Incumbent Advantage, measured by the percentage share of votes received by the incumbent o?cial in the last election. I try to account in this way for the greater degrees of freedom that a popular incumbent has when choosing fiscal policy. Appendix C provides sources and more details on how these controls were generated. Figure 3.3 depicts the evolution of per-capita spending, deficits, and debt, while Figure 3.4 shows the fiscal dependence index. The increasing trend in all of these variables reflects the growing level of decentralization in the country. the timing of elections in Colombia, most years are either a pre-election year or a post-election year. 121 Specification Control (1) (2) (3) (4) T(t) xxxx y(i,t-1) xxxx GDP_PC(i,t-1) x x x x UBN(i,t-1) x x x x POPULATION(i,t-1) x x x x DEFICIT(i,t-1) x x DEBT_84(i,t-1) x T*FISCAL_DEP(t) x x x VOTE SHARE(i, prev.elect) x Table 3.3: List of control variables 3.4.2 Regression specification Since I have more than 1,000 cross-sectional units, estimating the city-specific e?ects (a i ) separately is not e?cient. I therefore use the first-di?erenced version of relationship (3.1). 12 The specification can be written as: ?y it = b??y it?1 + X k c k ??x k,it + d??elecdum t + u it (3.2) where u it = ?? it Note that I have dropped the i subindex for elecdum, since elections occur at the same time in all districts. Also, the model in di?erences does include a 12 Due to econometric problems related to the autoregressive component, a fixed-e?ects es- timator is not appropriate, as it would yield biased estimates. While lags of the endogenous variables can be used as instruments in the first-di?erences specification to address this problem, thesameisnottrueinafixed-e?ects specification. 122 0 1 2 3 4 5 6 7 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 D e fi c i t , e xpe ndi t ure ( a l l c a t e gor i e s ) 0 5 10 15 20 25 D e bt a c c um ul a t e d s i n c e 19 84 Deficit per capita Expenditure per capita Debt per capita Figure 3.3: Evolution of financial indicators -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 D e vi a t i ons from 1991 fi s c a l de p e nde nc e Figure 3.4: Evolution of fiscal dependence 123 constant and a linear trend, since the controls include a quadratic trend (with both linear and quadratic terms). Given that the first local election is dated in 1988, in the regressions the initial t is 1987 (the first pre-election period). Data for previous years provide instruments in the estimations. Regression (3.2) is a?ected by the familiar endogeneity problem: the error term now includes ? i,t?1 which is correlated with ?y t?1 , and also includes ? i,t which is correlated with the x t if these outcomes are a?ected by contemporane- ous fiscal policy. To address this problem, I follow the approach suggested by Anderson and Hsiao (1982), and estimate (3.2) by 2SLS, using y i,t?2 and y i,t?3 to instrument the ?y i,t?1 ,andx i,t?1 and x i,t?2 to instrument the ?x i,t 13 .Un- der my assumption that ? i,t is white noise makes this instruments valid, in the sense of being orthogonal to the innovations. As a robustness check, I re-do these estimations using a GMM approach with the same matrix of instruments. I also run regression (3.1) in levels by OLS, replacing the city-specifice?ect with time-invariant variables for each municipality. This last model is specified as follows: y it = a 1 ?surf i + a 2 ?dist i + 6 X r=3 a r ?regiondum i (3.3) + b?y it?1 + X k c k ?x k,it + d?elecdum t + ? it 13 A widely used alternative methodology is the one suggested by Arellano and Bond (1991). I do not use this approach because, with the relatively large numbers of periods (15) and endogenous variables (up to 5) in my estimations, the matrix of instruments suggested by these authors would contain a minimum of 60 columns, even if we only use two lags as instruments for each period. Besides being computationally consuming, GMM estimators with such a large number of overidentifying restrictions are known to have poor finite sample properties (see Wooldridge, 2002, for a discussion). 124 where I include the surface covered by the city or state, its average distance with respect to main markets in the country, and dummies for the four main regions in the country. I report robust standard errors. I refer to the results from OLS and GMM estimations in the text, but do not report these results to avoid overcrowding the paper with tables. The dependent variables are expressed in logs, as are population, GDP, and UBN. Note from Table 3.2 that I have a total of 12,318 observations, 5,289 of which fall in pre-electoral periods. All reported regressions are weighted by expen- ditures; results are largely robust to eliminating these weights. I also conducted some robustness tests varying in the sample (both in terms of time periods and of districts included in the regression). These changes did not a?ect results im- portantly. 3.4.3 Regression results Results for the political dummy in which we are interested, d,arepresentedin Tables 4 and 5. To facilitate reading, estimates for other coe?cients are not reported, but are available from the author upon request. In these tables, each of columns (1) through (4) represents a di?erent set of controls, as detailed in Table 3.3. Each row corresponds to a di?erent regression, and the dependent variable for that regression is recorded in the first column. Throughout the paper, results in bold letters are significant at the 5% level, while results in bold and italics are significant at 10%. When interpreting the size of estimated coe?cients, keep in mind that expenditure variables are expressed in logs. I run two versions of equation (3.2). Table 3.4 presents results for the first of those versions, where y i,t corresponds to the share of total expenditure repre- 125 sented by a specifictypeofspending.Forinstance,therowmarkedCURRENT reports the estimate of d when the dependent variable is current expenditure as a share of total expenditure (including interest payments, i.e. total spending is not simply the sum of current spending and investment, which is why conducting the exercise for both categories is not redundant). This is motivated by theoretical models that, as mentioned, suggest PBCs should take the form of changes in the composition of the budget, rather than its size. The results point to a change in the composition of expenditure away from current expenditures and into capital spending. The categories of investment as- sociated with most visible infrastructure projects, namely construction of roads, infrastructure, housing, and water, power, and communications (POWER), all show pre-electoral expansions. These are significant, both statistically and eco- nomically; for instance, in some of the estimates expenditure in infrastructure grows by around 60% prior to elections. Other categories of investment, more related to the provision of universal (as opposed to targeted) goods, as health and education, do not show similar pre-electoral cycles. At the same time, there is a contraction of current spending, which can be attributed to a decrease in the share of the budget represented by current transfers 14 .Thesefindings are consis- tent with an opportunistic pre-electoral expansion of targeted expenditures. 14 This may seem puzzling, since the theoretical literature frequently refers to targeted expen- ditures as transfers to the targeted groups. However, the reader should not confuse these with ?Current Transfers?, as defined in the Colombian government accounts. These cover benefits to retired and temporary employees, and transfers to other levels of government. None of these is likely to constitute a group of voters worthy of pre-electoral targeting. As argued above, in the government accounts pre-electoral opportunistic transfers are more likely captured by some investment categories. 126 Dependent variable: Type of expenditure Electoral effect (coefficient d ) (as a fraction of total expenditure) (1) (2) (3) (4) CURRENT -0.143 -0.026 -0.059 0.001 (0.016) (0.010) (0.009) (0.010) GENERAL -0.059 0.011 -0.003 0.079 (0.028) (0.018) (0.017) (0.018) PERSONNEL -0.036 0.005 -0.019 0.039 (0.032) (0.020) (0.018) (0.021) TRANSFERS -0.413 -0.290 -0.334 -0.266 (0.035) (0.023) (0.023) (0.024) INVESTMENT 0.206 0.110 0.134 0.101 (0.088) (0.031) (0.025) (0.016) INFRASTRUCTURE 1.277 0.630 0.818 0.526 (0.237) (0.157) (0.243) (0.109) ROADS 1.112 0.693 0.639 0.820 (0.160) (0.130) (0.128) (0.153) POWER 0.782 0.466 0.373 0.517 (0.140) (0.114) (0.090) (0.124) HOUSING 0.915 0.717 0.772 0.420 (0.919) (0.760) (0.830) (0.302) EDUCATION 0.346 0.008 -0.008 0.165 (0.085) (0.073) (0.069) (0.068) HEALTH 0.227 -0.011 -0.005 0.191 (0.072) (0.075) (0.075) (0.076) Notes: This table presents estimates for coefficient d in regression (3.2). Estimation is done by 2SLS. Standard errors in parentheses. Bold characters denote significance at 5%. Bold and italics denote significance at 10%. Each row corresponds to a different regression, where the dependent variable is the share of total expenditure represented by a different category (current, general, etc). Each column corresponds to a different set of controls as detailed in Table 3.3. Table 3.4: E?ect of elections on the composition of total expenditures 127 My choice of instruments is greatly restricted by data availability, since not many other variables are measured at the local level. However, the set of in- struments I use, consisting of the lags of potentially endogenous variables, does seem to perform relatively well. The relevance of the instruments should not be of much concern, since most of these variables exhibit persistence: economic activity, poverty, and the size of the population are all variables that change only slowly over time. Below, I show that government spending also shows inertial dynamics. The first stage R-squared for endogenous variables are in most cases around 0.3. To test the exogeneity of the instruments, I run regressions of the residuals from equation (3.2) on the matrix of instruments used for each case. Intuitively, the R-squared from these regressions should be low if the instruments are indeed exogenous. These R-squared are indeed quite small, ranging in size from 0.005 in the regressions from transfers to around 0.04 in the regressions for personnel spending. However, they are not small enough to support the hypothe- sis that instruments are exogenous in a formal Chi-squared test 15 ,asisfrequently the case when using large panels. Regressions with other methods, and/or without weights, produce similar results. OLS regressions of the specification (3.3) actually show significant pre- election increases in all the reported components of investment, accompanied by contractions of current transfers and some subcomponents of personnel spending. GMM estimates of (3.2) are mostly similar. The main exception is in the weighted estimations, where GMM estimates of the electoral e?ect on total investment lose 15 Under the null hypothesis that instruments are exogenous, the product of this R-squared and the number of observations follows a Chi-squared distribution with degrees of freedom equal to the number of overidentifying restrictions. In my case, these statistics are frequently larger than the corresponding critical value. 128 significance. I repeat these regressions setting y it equal to the log level of expenditure in a given category. That is, I look for e?ects on spending amounts, rather than shares of overall spending. Table 3.5 presents the results of this approach. As in Table 3.4, spending on total investment, infrastructure, power, and roads all show pre- electoral increases. The level of current transfers, meanwhile, contracts. These e?ects are large in size, approaching 20% in absolute value for infrastructure, roads, water and power plants, and transfers. An interesting result from these regressions is that payments to personnel increase by about 8% prior to elections 16 . Education expenditure also increases significantly before elections. Overall expenses do not show any economically significant change in the pre-election periods (even when statistically significant, the change is smaller than 2%). OLS (for equation 3.3) and GMM estimations yield very similar results. The main di?erence is that the e?ect on total investment loses significance in those two specifications, although the e?ectsonitsmainsubcomponentsremainsignificant and large. The GMM estimation also fails to identify any significant e?ect on housing expenditures. In sum, the results indicate a pre-election shift of government resources from current spending into investment types of expenditures, while the overall budget does not experience significant changes. This is consistent with the view that incumbents try to obtain voters? support by increasing the provision of goods that are most valuable and/or visible to them, while avoiding large deteriorations 16 The finding of a pre-electoral expansion of personnel expenditures would be consistent with the widespread idea that politicians in Colombia trade government jobs in exchange for political support. 129 of the overall balance. Is this strategy optimal to tilt election outcomes in favor of the incumbent? Trying to answer this question I now examine some empirical evidence on the link between the government?s budget and election outcomes. 130 Dependent variable: Type Electoral effect (coefficient d ) of expenditure (1) (2) (3) (4) TOTAL -0.089 0.015 0.018 -0.013 (0.072) (0.008) (0.008) (0.009) CURRENT -0.030 -0.014 -0.051 -0.035 (0.028) (0.017) (0.012) (0.013) GENERAL 0.008 0.058 0.050 0.054 (0.048) (0.009) (0.009) (0.009) PERSONNEL 0.112 0.067 0.077 0.067 (0.012) (0.006) (0.006) (0.006) TRANSFERS -1.778 -0.213 -0.278 -0.261 (1.251) (0.025) (0.025) (0.028) INVESTMENT 0.069 0.051 0.090 0.046 (0.042) (0.012) (0.010) (0.011) INFRASTRUCTURE 0.249 0.224 0.199 0.197 (0.049) (0.043) (0.059) (0.047) ROADS 0.331 0.265 0.261 0.252 (0.042) (0.041) (0.039) (0.045) POWER 0.325 0.287 0.275 0.257 (0.052) (0.042) (0.039) (0.055) HOUSING 0.632 0.615 0.604 0.237 (0.329) (0.316) (0.320) (0.287) EDUCATION 0.089 0.067 0.071 0.033 (0.029) (0.020) (0.019) (0.035) HEALTH -0.078 -0.014 0.028 -0.136 (0.057) (0.044) (0.041) (0.093) Notes: This table presents estimates for coefficient d in regression (3.2.). Estimation is done by 2SLS. Standard errors in parentheses. Bold characters denote significance at 5%. Bold and italics denote significance at 10%. Each row corresponds to a different regression, where the dependent variable is a given type of government expenditure. Each column corresponds to a different set of controls as detailed in Table 3.3. Table 3.5: E?ect of elections on di?erent types of expenditure 131 3.5 Do voters reward targeted spending? The evidence presented above is consistent with the view that elected o?cials face incentives to manipulate the budget before elections, in order to improve the chances that their parties will be re-elected. The implicit foundation for this view is that voters? support for the incumbent is a?ected by his previous fiscal choices. One argument is that all increases in spending should boost popular support for the government, because of their expected favorable e?ects on eco- nomic activity. According to theoretical models of the PBC, however, the reason why rational voters care about past fiscal policies when deciding whether to re- elect an incumbent is that the fiscal choices of an incumbent reflect persistent personal characteristics. Past policies are therefore an indication of the policies the incumbent would enact if re-elected, and because of this reason they influ- ence voters? perceptions about the incumbent. In Rogo??s (1990) model voters? support depends positively on the provision of goods visible to voters, which they associate with higher competence. Meanwhile, the model presented in chapter 2 suggests that voters react favorably to increases in targeted spending, even if they dislike incumbents who run deficits 17 . Three obvious questions come to mind. Does recent fiscal policy indeed a?ect the choices of voters? If so, what are the directions of those e?ects? Finally, is there evidence that an incumbent?s fiscal choices are indeed persistent? I devote this section to answering the two first questions, while the last one is addressed 17 Models of competence inference (e.g. Rogo? 1990 and Shi and Svensson 2001) are also consistent with voters punishing high deficits, under the assumption thatif at least some voters observe the overall budget. Informed voters associate higher deficits, for any given level of spending, with lower competence of the incumbent. 132 in the section 3.7. 3.5.1 Data As argued before, the relevant definition of ?incumbent? for the Colombian case is the incumbent party, since o?cials cannot run for direct re-election. I therefore use data on the share of votes obtained by each party in the local mayor elections of 1992-2000 (four elections). Unfortunately, for previous elections only the share of votes obtained by the winner of the election is available, so that full party shares cannot be calculated 18 . Politics in Colombia have been traditionally dominated by two major parties, Liberal and Conservative. While some candidates, particularly in the 1990?s, ran under the banner of a myriad of di?erent parties or political movements, many of these movements can be traced back to the traditional parties, and voters in each locality are frequently aware of those ties 19 . In that sense, elections are still mainly a contest between these two major parties, although there are also two smaller left-wing parties and some truly independent political groups. The challenge is to identify in the data which candidates are associated with one of the major parties, in order to calculate the appropriate shares of party 18 It is often the case that several candidates run for the same party, so that the votes received by one candidate cannot be assumed to equal the votes received by his party in that election. 19 There are two reasons why candidates prefer to run for movements linked to the parties, rather than the parties themselves. The first is that by creating a new group they can access funding that is available for each political organization in the race. The second is that voters have grown suspicious of the political practices (not necessarily the ideals) of the traditional parties. Candidates then try to avoid being associated with those practices by running outside the structure of the party. 133 votes. I use information from external sources, including informal accounts, to match the di?erent movements with the traditional party division between liber- als and conservatives. Appendix C contains a list of movements and parties that I have been able to match with the larger parties. All movements not listed in appendix C are considered ?independents? in my analysis. I calculate the share of votes obtained by, for instance, the Liberal party, as the sum of the shares obtained by all the smaller organizations linked to the Liberal party in that list. Since some apparently independent groups may indeed also be Liberal or Conservative, even if I am not able to identify them as such, the share of votes my calculations assign to a given major party is a lower bound. Table 3.6 presents some summary statistics; panel 1 refers to vote shares, and panel 2 to number of elections won. Columns (1) and (2) record statistics for the Liberal and Conservative party, respectively, while column (3) shows statistics for the predominant party in each election (between conservatives and liberals). Figures in column (4) correspond to the winning candidate. The high frequency of zero shares is due to the fact that these measures are lower bounds. Note that, despite this fact, in 50% of the elections at least one of the major parties receives 65% of the votes (column (3)). Moreover, the predominance of the two parties is confirmed by the fact that, out of 3880 total elections, 2880 are won by a candidate that I can tie to one of these parties. Votes to the predominant party tend to exceed those obtained by the winning candidate, since often more than one candidate runs for each party. 134 Party (1) (2) (3) (4) Liberal Conservative Max(lib,cons) Max (candidates) Mean 39.69 30.79 63.81 58.68 Median 37.91 32.39 64.93 54.51 Panel 1: 25 percentile 0 0 45.11 47.79 vote shares 75 percentile 74.50 57.39 95.63 66.60 Maximum 100 100 100 100 Minimum 0 0 0 20.38 Wins 1650 1230 Panel 2: elections won Wins running within the party 1597 1064 Wins runing for associated movement 53 166 Table 3.6: Summary statistics for election outcomes 3.5.2 The e?ect of fiscal policy on vote shares I study here the relation between the share of total votes obtained by each of the two major parties and pre-electoral fiscal policy. As in previous studies, the e?ect of the overall budget is captured by the government?s deficits. However, I have already highlighted that theoretical models suggest that voters see targeted, more visible, expenditures with a di?erent eye than they see the rest of spend- ing. I therefore attempt to distinguish the e?ects of these spending categories. Following the previous discussion, I treat investment spending as targeted expen- diture, and current spending as non-targeted expenditure. I run a regression of the following form: votes pit = ? 0 + ? 1 votes pit?1 (3.4) +(? 2 invest it + ? 3 current it + ? 4 deficit it + ? 5 gr it?1 )?inc pit?1 + ? itp The time indices here refer to election periods, so that t is the current election 135 and t ? 1 the previous election. votes pit is the share (in percentages) of votes obtained by party p in city i during the t election. For this analysis, I treat zero vote shares as missingvalues, sinceI suspectthat most of these cases donot reflect that a major party did not present any candidate, but rather that I cannot tie a candidate to the party he belongs to. Vote shares are modeled as a function of the interaction between fiscal variables and the discrete variable inc pit?1 ,which takes the value of 1 if party p is in o?ce at the time of the election and ?1 otherwise. The fiscal variables correspond to the pre-election year; I include the log of investment spending (invest it ), the log of current spending (current it ), and the per capita government deficit (deficit it ). Average GDP growth between t? 1 and t (gr it?1 ) is also considered to control for other observables that may a?ect voters? perceptions about the incumbent. Under the assumption that ? itp captures the part of voting behavior that the politician cannot predict, fiscal policy decisions cannot be based on those innovations, and the policy variables included in the regression should satisfy the restriction of being orthogonal to the error term. Results are reported in Table 3.7; column (1) reports estimates of (3.4) with the expenditure variables (invest and current) in per-capita terms, while for column (2) the expenditure regressors are expressed as shares of total expenditure (these shares do not add up to one, since interest payments is not included in any of these categories). Note that the dependent variable is expressed as percentages, while the spending measures are in logs. As previous studies have found for other countries (e.g. Brender, 2003, for Is- rael, and Peltzman, 1992, for the US), and contrary to traditional interpretations of the PBC, my results indicate that Colombian voters penalize the incumbent 136 party for running high deficits. Furthermore, high capital expenditures (which I interpret as targeted spending) increase the share of votes obtained by the incumbent party, while current (?non-targeted?) expenditure has no significant e?ect 20 . A one percent increase in per capita investment increases the fraction of votes obtained by the incumbent party by about 0.03%, while a two standard deviation increase in the deficit per capita (about $4,000 pesos of 1998) decreases the share of votes to the incumbent party by close to 0.08%. These results are consistent with the view that voters dislike incumbents who run high deficits, while they value specific types of expenditures. They are also consistent with my results on electoral changes in the composition of spending, that show incum- bents increasing targeted spending before the elections, while they try to avoid concomitant increases in the overall budget. According to findings in this section, this fiscal strategy is optimal in terms of maximizing the share of votes that go to the incumbent party in the upcoming election. 20 GDP growth does not have any significant e?ect. This result is not surprising result since my measure of GDP is at the state level. Voters probable do not ?blame? the local mayor for the state economy?s performance. 137 Dependent Variable: Votes to party P Regressor (1) (2) Constant 32.801 31.694 (1.606) (1.761) Votes to P in past election 0.485 0.507 (0.026) (0.030) Deficit * incumbent -0.023 -0.019 (0.009) (0.009) Investment Expenditure 2.757 5.282 *incumbent (0.805) (1.414) Current Expenditure 1.312 -0.070 *incumbent (0.787) (0.660) GDP growth*incumbent -6.464 1.465 (16.070) (16.553) Observations 2032 2032 R-square 0.222 0.221 Notes: this table presents the results of estimating equation 3.4. Bold characters denote significance at 5%. Bold and italics denote significance at 10%. The (log of) expenditure variables are either in per capita terms (column (1)) or fractions of total expenditure (column (2)) Incumbent is 1 if party P is in power at the time of the election, -1 otherwise Table 3.7: E?ect of fiscal performance on vote shares 138 3.6 The e?ect of ideological polarization: PBCs in swing cities The model introduced in chapter 2 also suggests that electoral e?ects on the budget should vary depending on the level of loyalty of voters to one party or the other. The incumbent?s core voters will support him even if they do not expect to receive targeted transfers from him in the future, and the opposite is true for the challenger?s core voters. In this sense, pre-electoral manipulation of targeted versus non-targeted spending should be more of an issue in societies with a large fraction of uncommitted, or swing, voters. This section examines how the extent of the electoral e?ect depends on the degree of polarization in a given district. I begin by characterizing districts as swing and non-swing, on the basis of five di?erent measures of degree of commit- ment to one of the two major parties. In all of the cases, the swing variable is a dummy equal to 0 if the district is not swing, 1 if the district is swing. The three first measures are time invariant, while I allow the other two to vary from election to election. The definitions are as follow: 1. Swing_freqelec: district is not swing if the same party won at least 5 of the 7 elections in the period 21 . 2. Swing_actual_vote: district is not swing if one of the parties obtained at least 70% of the votes in every election in the 1992-2000 period 22 . The 70% threshold is based on the summary statistics of vote shares (see Table 3.6): in 21 Even though I only have voting shares for 1992-2000, my data base does contain the winning candidate?s party a?liation for all the elections. 22 The period limitation responds to data availability. 139 half of my sample elections, one major party obtained at least 64% of the votes, and that number rises to 71% if I omit elections in which I could not tie any candidate to a major party (as will be the case in this section). 3. Swing_fitted_vote: same as swing_actual_vote, but constructed using projected rather than actual vote shares. The fittedvotesharescomeforare- gression identical to (3.4), except that I allow the coe?cients to vary for the Liberal party with respect to the Conservative party. Vote shares are projected by fixing investment at 0, since the ideological polarization of a group in the model of chapter 2 (the degree to which it is not ?swing?) represents its coin- cidence with one party on issues other than targeted spending.Animportant conceptual advantage of these projected measures is that they eliminate the po- tential endogeneity of swing measures with respect to government investment. 4. Swingtemp_actual_vote: time-varying measure. District is not swing in election period t if in election t?1 the ruling party obtained at least 70% of the votes. 5. Swingtemp_fitted_vote:sameasswingtemp_actual_vote but using pro- jected vote shares (again, assuming investment is zero) rather than actual. The fraction of districts identified as swing is 57% for the first measure (swing_freqelec), around 75% for the time-invariant measures based on vote frac- tions, and around 60% for the time-varying measures. Some summary statistics for the projected vote shares are recorded in Table 3.8. The projected vote shares exhibit less dispersion than the actual shares, and have higher means (the actual shares are summarized in Table 3.6); those means exceed 50% for both parties due to the presence of missing values, more prevalent in districts where the party is not dominant. For each party, the projected share is positively correlated 140 Mean Std Dev Minimum Maximum Projected vote share Liberal party 63.912 13.536 36.550 80.704 Projected vote share Conservative party 60.006 19.714 23.252 86.440 Sample Correlations Projected vote share Liberal party Actual vote share Liberal party Projected vote share Cons. party Actual vote share Cons. party Projected vote share Liberal party 1 0.394 -0.624 -0.349 Actual vote share Liberal party 1 -0.426 -0.175 Projected vote share Conservative party 10.546 Actual vote share Conservative party 1 Table 3.8: Summary statistics for fitted vote shares with the actual share, and negatively correlated with the actual share for the other party. Table 3.9 presents sample correlations between the di?erent swing dummies, which are low in many cases. The only high correlations are between the two time-varying dummies (0.995), and between the time-invariant measures obtained from projected votes and from the frequency of wins. I place most con- fidence in the swing dummy based on the frequency of wins, and the time-varying measure based on actual votes. This is because the di?culties in correctly iden- tifying which candidates are tied to which parties, and the resulting proliferation of missing values for the vote shares, place doubts on swing measures that de- pend on being able to follow vote shares within a district over time. The three other dummies for swing districts (swing_actual_vote, and the two dummies that depend on projected vote shares) are a?ected by this problem. To analyze whether the electoral change in the composition of spending di?ers 141 freqelec actual_vote fitted_vote temp_actual temp_fitted swing_freqelec 1 0.058 0.452 0.056 0.053 swing_actual_vote 1 0.083 0.014 0.013 swing_fitted_vote 1 0.011 0.008 swingtemp_actual_vote 10.95 swingtemp_fitted_vote 1 Table 3.9: Sample correlations for swing district dummies between swing and non-swing districts, I add an interaction between the election dummy and the swing district dummy to the basic regression (3.2). The modified empirical model is: ?y it = b??y i,t?1 + X k c k ??x k,it (3.5) + d??elecdum t + s??(elecdum t ?swing ti )+u it where swing it is one of the swing district dummies defined above. The d co- e?cient captures now the electoral e?ect in non-swing districts, while the cor- responding e?ect in swing districts is given by d + s. I express the dependent variables as shares of total spending, since this exercise is motivated by the ex- penditure composition model of chapter 2. For non-swing districts, the electoral e?ect reflects possible electoral motives outside the expenditure composition incentive. Given the above discussion, one would expect those electoral incentives to capture the need for reducing the deficit in order to attract votes 23 .Thed coe?cient should therefore be negative. The 23 As mentioned, loyalty to one party is defined in terms of the coincidence with that party in issues other than targeted spending. Hence, non swing districts cannot be swayed by targeting expenditures to them, but can change their support to one or another candidate as a result of shifts in other issues. These other issues include the overall budget. Given the preference of voters for smaller overall budgets, increases in spending should negatively a?ect the incumbent party in non-swing districts. 142 electoral incentives to change the composition of expenditure toward targeted spending are reflected in d + s.Iwouldthusexpectd+ s>0 for components of expenditure related to targeted spending (such as investment), and d+s<0 for at least some of the other components. The results of estimating equation (3.5) by 2SLS, using the same instruments as in previous tables, are presented in Tables 3.10 and 3.11. Table 3.10 reports results using the time-invariant definitions of swing districts (definitions 1, 2 and 3), while Table 3.11 reports results for the time-varying measures (definitions 4 and 5). The set of controls included in the estimation is the one listed under column (4) in Table 3.3. Besides the coe?cients? estimates and their standard errors, I include the Wald statistic for the null hypothesis that d+s =0, that is, that the electoral e?ect is not significant for swing districts 24 . The results are consistent with the preceding discussion. In non-swing dis- tricts several types of expenditures exhibit a pre-electoral contraction. This result is robust to di?erent measures of swing districts, except for the time-invariant dummy based on actual vote shares, under which I find no electoral e?ectatall in non-swing districts. Swing districts, however, exhibit changes in the composi- tion of spending consistent with the pattern described above: current spending contracts, mostly due to a reduction in current transfers, and this contraction is larger than the e?ect observed for non-swing districts. The infrastructure-related category of road construction shows an expansion, and the other investment cat- egories, education and health, do not exhibit the contraction they display for 24 Note that less categories of spending are listed in this table due to data problems: since election results for major parties are only available for some districts, and the coverage of the spending data also varies across districts and categories of expenditure, for some categories the set of districts with recorded data contains only swing districts. 143 Swing district dummy Dependent variable: type of expenditure swing_freqelec swing_actual_vote swing_fitted_vote (as fraction of total expenditure) Pre-election dummy Pre-election * swing Pre-election dummy Pre-election * swing Pre-election dummy Pre-election * swing -0.295 0.109 1.505 -1.716 -0.283 0.086 CURRENT (0.057) (0.035) (1.001) (0.998) (0.064) (0.041) Wald stat: joint effect 8.504 11.911 10.039 -0.006 0.148 1.478 -1.382 -0.008 0.131 GENERAL (0.118) (0.063) (1.816) (1.810) (0.131) (0.074) Wald stat: joint effect 1.206 0.617 0.971 -0.182 0.086 2.192 -2.296 -0.195 0.096 PERSONNEL (0.083) (0.055) (1.609) (1.606) (0.095) (0.065) Wald stat: joint effect 1.140 1.471 1.265 -0.792 -0.159 2.830 -3.747 -0.706 -0.244 TRANSFERS (0.183) (0.091) (2.446) (2.429) (0.206) (0.113) Wald stat: joint effect 27.147 26.373 28.060 0.227 0.970 -2.879 3.655 0.013 1.044 ROADS (0.294) (0.311) (5.588) (5.611) (0.349) (0.366) Wald stat: joint effect 10.813 7.129 9.523 -0.485 0.566 0.047 -0.200 -0.493 0.456 EDUCATION (0.110) (0.107) (2.168) (2.171) (0.128) (0.124) Wald stat: joint effect 0.437 2.065 0.099 -0.444 0.573 1.576 -1.722 -0.403 0.366 HEALTH (0.107) (0.107) (2.175) (2.177) (0.128) (0.124) Wald stat: joint effect 1.140 1.875 0.105 Notes: this table presents the results of estimating equation (3.5) by 2SLS. Columns correspond to different regressions, changing the swing district dummy. Standard errors in parentheses. The Wald statistic corresponds to H0: d+s=0. Bold characters denote significance at 5% level. Bold and italics denote significance at 10%. Table 3.10: E?ect of elections on government expenditure: swing versus non- swing districts (time-invariant dummies) non-swing districts. 144 Swing district dummy Dependent variable: type of expenditure swingtemp_actual_vote swingtemp_fitted_vote (as fraction of total expenditure) Pre-election dummy Pre-election * swing Pre-election dummy Pre-election * swing -0.138 -0.324 -0.096 -0.383 CURRENT (0.052) (0.067) (0.050) (0.062) Wald stat: joint effect 19.871 25.522 0.103 0.358 0.095 0.342 GENERAL (0.122) (0.205) (0.114) (0.176) Wald stat: joint effect 2.270 2.719 -0.115 -0.224 -0.093 -0.247 PERSONNEL (0.075) (0.101) (0.074) (0.099) Wald stat: joint effect 5.092 5.712 -0.888 -0.340 -0.854 -0.409 TRANSFERS (0.172) (0.231) (0.163) (0.211) Wald stat: joint effect 10.984 14.098 0.583 1.557 0.683 0.992 ROADS (0.259) (0.660) (0.276) (0.616) Wald stat: joint effect 7.843 5.457 -0.143 -0.100 -0.167 0.179 EDUCATION (0.103) (0.189) (0.105) (0.206) Wald stat: joint effect 1.252 0.003 -0.130 -0.036 -0.145 0.073 HEALTH (0.101) (0.227) (0.105) (0.229) Wald stat: joint effect 0.420 0.085 Notes: this table presents the results of estimating equation (3.5) by 2SLS. Columns correspond to different regressions, changing the swing district dummy. Standard errors in parentheses. The Wald statistic corresponds to H0: d+s=0. Bold characters denote significance at 5% level. Bold and italics denote significance at 10%. Table 3.11: E?ect of elections on government expenditure: swing versus non- swing districts (time-varying dummies) 145 3.7 The persistence of expenditure choices All models in which voters rationally respond to electoral manipulation of the budget rely on some mechanism that relates pre- and post-election fiscal policy. This is so because forward looking individuals vote on the basis of what they expect from the candidates once in o?ce. The empirical implication is that an o?cial?s fiscal choices should exhibit some inertia. I analyze here whether the dynamics of government spending in Colombian municipalities are consistent with this implication. If so, we should observe that government spending exhibits some persistence, but the inertial component should be less pronounced in periods in which a power transition occurs 25 .Inpar- ticular, in light of models discussed previously, I examine whether the composition of spending is persistent over time, and whether that persistence diminishes at times of power transitions. Since mayors cannot be directly re-elected in Colom- bia, I look at party transitions. If fiscal policy reflects inertial characteristics of both parties and o?cials, the pattern of autocorrelation of spending should be less pronounced in periods of transition from one mayor to another, but this di?erence should be less important when the transition occurs within the same party. The empirical model is a modification of the one used in previous sections, and takes the following form: 25 The persistence of government spending is not driven solely by persistent characteristics of o?cials. Other factors that contribute to persistence are stable fundamental characteristics of the economy, relatively rigid spending items (e.g. public employment), and multi-period spending commitments. 146 ?y it = b??y it?1 + X k c k ??x k,it + d??elecdum t (3.6) + b 1 ??(y it?1 ?elecdum t+1 )+b 2 ??(y it?1 ?elecdum t+1 ?reel it )+u it where reel it is a dummy equal to 1 if the same party remains in power between t?1 and t, 0 otherwise, and the rest of the notation is as above. Note that the transitions of o?cials in power occur at periods given by elecdum t+1 .Theterms in the top line of equation 3.6 are the focus of this section. The serial correlation is given by b in ?normal? times, b+b 1 in periods in which the incumbent party is replaced by a challenger, and b + b 1 + b 2 when the same party remains in o?ce, but the mayor changes. As the previous discussion suggests, the assumption of persistent o?cial characteristics should be reflected in b>0, b 1 < 0,andb 2 > 0 to (partially) o?set the negative e?ect of b 1 . As before, I carry this estimation by 2SLS, adding the first and second lags of y it?1 ?elecdum t+1 and (in some cases) y it?2 ?elecdum t+1 to the list of instruments introduced in section 3.4. Tables 3.12 and 3.13 summarize the results. The dependent variable for the results reported is the ratio of current to investment spending, since the focus is on the composition of expenditures between targeted and non-targeted categories. However, the estimates are qualitatively similar if I use instead the shares of total spending represented by these categories, or the levels of investment and current spending. Note that, di?erent from previous tables, here the rows represent di?erent regressors, while the dependent variable is the same in all cases. I only report results for the b, b 1 , b 2 and d coe?cients, and the Wald statistic for the null hypothesis that b 1 + b 2 =0(i.e. that all persistence is due to party characteristics). Table 3.12 presents results using 147 what the table calls the ?basic set of instruments?, consisting of the instruments used for previous tables plus the first and second lags of y it?1 ?elecdum t+1 ,while Table 3.13 reports results adding y it?2 ?elecdum t+1 to the set of instruments (for what the table calls the expanded set of instruments). In both tables, the results in column (1) come from a specification with the basic set of controls (Table 3.3, column 1), while in column (2) the full set of controls (Table 3.3, column 4) is used. The first-stage R-squared is included for each regressor. Results are consistent with the patterns expected. The ratio of current to investment spending exhibits high serial correlation, but this correlation falls by about 10% when the party in power changes. The reduction in the correlation is only around 2% when the mayor is replaced by someone from the same party. Note that, consistent with results in previous sections, the ratio of current to investment spending falls in pre-electoral periods. Also, note that results are very similar in all columns of Tables 3.12 and 3.13. 148 Dependent variable: current expenditure / investment (1) (2) Regressors Basic controls, basic instruments All controls, basic instruments Estimates 1st stage R-squared Estimates 1st stage R-squared Lag dependent variable 0.920 0.730 0.922 0.683 (0.013) (0.019) Lag dependent * power transition -0.140 0.839 -0.249 0.919 dummy (0.028) (0.034) Lag dependent * power transition 0.118 0.827 0.288 0.893 dummy* party reelection dummy (0.027) (0.034) Pre-election dummy (electoral effect) -0.299 1 0.177 1 (0.030) (0.053) Wald statistic (H0: b 1 +b 2 =0) R-square Observations Notes: this table presents the results of estimating equation (3.6) by 2sls. Columns 1-2 correspond to different sets of controls. Standard errors in parentheses. Bold characters denote significance at 5% level. Bold and italics denote significance at 5%. 48.330 0.525 6651 6471 0.162 107.070 Table 3.12: Persistence of the composition of government expenditure (basic instruments) 149 Dependent variable: current expenditure / investment (1) (2) Regressors Basic controls, all instruments All controls, all instruments Estimates 1st stage R-squared Estimates 1st stage R-squared Lag dependent variable 0.923 0.741 0.920 0.744 (0.013) (0.013) Lag dependent * power transition -0.130 0.887 -0.134 0.892 dummy (0.022) (0.024) Lag dependent * power transition 0.109 0.862 0.113 0.864 dummy* party reelection dummy (0.021) (0.023) Pre-election dummy (electoral effect) -0.295 1 -0.294 1 (0.031) (0.040) Wald statistic (H0: b 1 +b 2 =0) R-square Observations Notes: this table presents the results of estimating equation (3.6) by 2sls. Columns 1-2 correspond to different sets of controls. Standard errors in parentheses. Bold characters denote significance at 5% level. Bold and italics denote significance at 5%. 48.790 0.468 6651 6471 0.515 49.630 Table 3.13: Persistence of the composition of government expenditure (all instru- ments) 150 3.8 Electoral cycles in the central government?s budget Although I have argued that electoral cycles derived from the targeting of ex- penditures are most relevant at the local level, where targeting is most e?cient, this phenomenon is no stranger to national level politics. In fact, the idea of pork projects is most often associated by the public with Congress politics. One question is, therefore, whether we also observe this type of e?ectatthenational level. In this section, I take an exploratory look at that question, focusing on the dynamics of some components of the Colombian central government?s budget. One problem for this exploration is the short length of the o?cial quarterly time series on fiscal policy, which begins in 1988. I gathered information to ex- tend those series, from the o?cial printed reports of the Contralor?a General on the finances of the government 26 . The resulting data are quarterly frequency, and cover the 1974.1-2000.1 period. The level of disaggregation is not as detailed as for the local data, but I can distinguish current from investment spending, and two subcomponents of current spending: transfers and personnel. The de- nomination transfers, again, introduces confusions; one must note that these are current transfers, not including the transfers from the central level to the local governments, which are actually recorded as a part of investment. I therefore still regard transfers as a non-targeted type of spending. Elections occur at predetermined dates, every four years. Presidential and 26 The Contralor?a General is the same source I use for my local fiscal data. It is the entity in charge of monitoring the government?s financial statements. More details on the construction of national level fiscal variables are provided in the appendix. 151 Congressional elections are almost simultaneous (a two-month period separates them), so it is impossible to separate the e?ect of Congressional elections from that of a Presidential election. As controls, I use information on unemployment, GDP and per capita GDP. Sources, definitions, and the dates of presidential elections are listed in appendix C. Irunaregressionoftheform: y t = L X l=1 b l ?y t?l + c?x t?1 + d?elecdum t + ? t (3.7) where y t isthegrowthrateofsometypeofgovernmentspendingbetween t ? 1 and t,andx t?1 is a control equal to the growth of either unemployment, GDP, or per capita GDP, which I date in the previous period for the same reasons explained before for the local case. elecdum t is the pre-election dummy, which in this case takes the value of 1 in the two quarters prior to the elections, 0 in all others. Note that I use L lags of the dependent variable, where L is optimally chosen following the Akaike criterion. The results of this estimation are presented in Table 3.14. The table only shows the estimates for the electoral e?ect (d), and follows the same conven- tions used in all other tables. Each column corresponds to a di?erent control (unemployment, GDP, and per capita GDP for columns 1, 2, and 3 respectively). I obtain results that are broadly consistent with those observed for the lo- cal level. First, there is no significant change of total spending before elections. Thesignofthee?ect is always negative for transfers and current spending, and positive for investment. However, while the e?ect on investment is always sig- nificant, the negative e?ect on current types of spending is significant only when controlling for unemployment. 152 Dependent variable: Type Electoral effect (coefficient d ) of expenditure (1) (2) (3) TOTAL 0.068 0.149 0.148 (0.105) (0.102) (0.102) CURRENT -0.846 -1.543 -1.561 (0.103) (1.018) (1.020) PERSONNEL 0.048 0.058 0.053 (0.059) (0.061) (0.061) TRANSFERS -0.647 -0.052 -0.049 (0.105) (0.134) (0.133) INVESTMENT 0.563 0.849 0.843 (0.262) (0.275) (0.275) Notes: This table presents estimates for coefficient d in regression (3.7). Estimation is done by OLS. Standard errors in parentheses. Bold characters denote significance at 5%. Bold and italics denote significance at 10%. Each row corresponds to a different regression, where the dependent variable is a given type of government expenditure. A different control in each: unemployment rate (1), GDP (2), per capita GDP (3) Table 3.14: E?ect of elections on di?erent types of expenditure. Central Govern- ment. 153 I consider these results as indicative that the suggested pre-electoral changes in the composition of government spending occur also at the national level. One interesting extension of these results would be to examine the allocation of pre- electoral payments to the local governments (registered under the investment heading), and relate these to the level of electoral polarization that characterizes di?erent districts. At this point, however, the data on regional allocations of central government expenditures are not readily available. 3.9 Concluding remarks The goal of this paper is to o?er a more comprehensive view of electoral cycles in government spending, integrating the pre-electoral fiscal choices of incumbents with the impact of those choices on election outcomes. The picture that emerges is one where voters punish the pre-electoral deterioration of fiscal balances but reward incumbents who, before the election, increase the provision of goods most visible and valuable to voters. In terms of maximizing his probability of being re-elected, therefore, an incumbent?s optimal strategy implies simultaneously in- creasing spending on those goods favored by voters and contracting other types of spending. In the Colombian case, this is reflected in pre-electoral shifts of re- sources away from current spending and into the development of infrastructure- related projects. The evidence presented here shows that there is a logic to apparently contra- dictory pieces of previous evidence, which showed certain types of government spending growing before elections despite voters? inclination to replace incum- bents that chose high spending. It also suggests that the traditional view that incumbents have electoral incentives to run high deficits does not apply gener- 154 ally, even within the group of developing economies, which are frequently seen as the mecca of political budget cycles. On the contrary, this evidence is consistent with models that picture the political budget cycle as an electoral manipulation of the composition, rather than the size of the budget. In the same vein, it is also consistent with the argument that voters favor specific types of goods, and incumbents attempt to influence electoral results by spending on those goods. Aninterestingquestionthatisleftopenistowhatextentthegreatersuscepti- bility of specific types of spending to electoral manipulation reflects heterogeneous preferences of voters and politicians, as opposed to di?erent degrees of visibility of public goods. In simpler, although inexact, words, is the PBC more a reflec- tion of pork politics, or competence signaling? Is it perhaps even a reflection of incumbents pre-paying important campaign contributors? The answer to this question requires a di?erent kind of data and empirical strategy, possibly di?er- entiating electoral transfers to specific groups of voters or contributors, and is part of a future research agenda. 155 Appendix A Technical Appendix to Chapter 1 A.1 Proof of proposition 1 Here I use the specifics of each case (single central banker, and committee CB) to get to a unique expression. For the committee case, as discussed in section 1.4.2, E v i Pr p (? 0 = ? d | ?)=Pr p (? 0 = ? d | ?), so terms involving solely Pr p (? 0 = ? d | ?) are constants in the dove?s problem. Abstracting from these and other constants, and taking into account that Pr(? =0| v i )=1?Pr(? = ? d | v i ) and Pr(r i =0| v i )=1? Pr(r i =1| v i ), we can re-write the dove?s problem in the committeecaseas: Min {v i } Pr(? = ? d | v i ) ? ? c 2 4 ? +Pr(? = ? d | v i )? i ? c 2 2 ? ? E v i Pr p (? 0 = ? d | ? = ? d )?E v i Pr p (? 0 = ? d | ? =0) ? ?? i bPr(r 1 =1| v i )+? i ? ? c 2 4 ? (1??)Pr(r i =1| v i )?Pr(? 0 = ? d ) The last term uses the fact that, if r i =0, i?s replacement votes for ? d in the second period with probability ?. Since a dove chooses v i =0if L(v i =0)< L(v i = ? d ), proposition 1 follows. 156 For the single central banker case, the problem can be written: Min {v i } Pr(? = ? d | v i ) ? ? c 2 4 ? + ? i ? c 2 2 ? E v i Pr p (? 0 = ? d | ? = v i ) ?? i bPr(r 1 =1| v i )+? i ? ? c 2 4 ? (1??)Pr(r i =1| v i )?Pr(? 0 = ? d ) The expression in proposition 1 follows, taking into account that ?Pr(? = ? d )=1for single central banker case. A.2 Derivation of equation (1.11) Letting k be a given member of period 2 committee, the public assigns Pr(k is dove | ? = ? d )= m n ? +(1? m n )? + Pr(k is dove | ? =0)= m n ? +(1? m n )? ? so that Pr p (? 0 = ? d | ? = ? d )= n X x= n+1 2 ? n x ? " ? m? +(n?m)? + n ? x ? 1? m? +(n?m)? + n ? n?x # and a similar expression applies if ? =0,with? ? in lieu of ? + . Taking a first order approximation of this expression around ? + = ? leads to equation (1.11). 157 A.3 Explicit form of condition (1.16) Note that ?Pr(? = ? d )=Pr ? n?1 2 of others vote for ? d ? = ? n?1 n?1 2 ? (?w) n?1 2 (1??w) n?1 2 and similarly ?Pr(? 0 = ? d )= ? n?1 n?1 2 ? (?) n?1 2 (1??) n?1 2 To obtain an explicit form of the condition to choose v i =0in the committee case with government appointments, plug these expressions and equations (1.11) and (1.17) into condition (1.16). 158 Appendix B Appendix to Chapter 2 B.1 Proof of proposition 9. We first need to prove that (2.28) solves the di?erential equation (2.27). Note that Y (g h t )=e ?E ( ln? h C ) g h t ?c 0 satisfies equation (2.27) for the E(ln? h I | g h t ) E(ln? h C ) branch of (2.27) is obviously hard to solve, but we will take the view that voters rather solve an approximate, linear, form of it. We take a first order Taylor approxima- tion around Y 0 = x = ??c 0 e ?E ( ln? h C ) . This ensures that lim g??g ?E(ln? h I |g h t ) ?g h t is equal whether we approach from the left or the right. This yields (letting b? = E(?)) Y (g h t )= e E ( ln? h C ) g h t ? [K 1 ?K 2 (Y 0 ?x)] 159 where K 1 = ? 1+??Ce ?E(ln? h C ) ?? 1+??Ce ?E(ln? h C ) ? ??Ce ?E(ln? h C ) ? b???C??e ?E(ln? h C ) ? C ? 1+??Ce ?E(ln? h C ) ?? 1+??Ce ?E(ln? h C ) ? = and K 2 = ? ? b???C??e ?E(ln? h C ) ? ?C ? 1+??Ce ?E(ln? h C ) ?? 1+??Ce ?E(ln? h C ) ? . The solution to this di?erential equation takes the form: Y (g h t )=exp ? ?? ? g h t ? 2 2K 2 e E(ln? h C ) !" c 1 + K 1 K 2 Z exp ? ? ? g h t ? 2 2K 2 e E(ln? h C ) ! dg h t # (B.1) where c 1 is a constant such that Y (?g)=1. Letting K 1 K 2 = c 2 and 1 2K 2 e E ( ln? h C ) = c 3 , this is identical to (2.28) for E(ln? h I | g h t ) >E(ln? h C ). Plugging (2.28) into (2.26) we obtain (2.29). 160 Appendix C Appendix to Chapter 3 C.1 Sources and details on data C.1.1 Data for local level estimations Population, surface, distance to main markets, and the UBN indicator were pro- vided by the University of Los Andes? CEDE. State per capita GDP data are from DANE (the Colombian Bureau of Statistics). Also, GDP for ?new? states 1 is only reported since 1995. Previously, only the sum for all new states was reported. I impute pre-1995 GDP for these new states by keeping the contribu- tion of each state to total new-states-GDP constant in its 1994-1996 level. The Unsatisfied Basic Needs indicator (UBN) summarizes the fraction of households without proper housing (in terms of number of rooms and construction mate- rials), without sanitary services, with school-less children, or with a single low income for more than three people. This poverty indicator is commonly used with local-level data, because at this level income measures needed to construct 1 There is a subset of nine states that were only elevated to the state category in 1991. They were previously in a di?erent, now disappeared, category of the regional classification. These are what I call ?new? states. 161 other poverty indicators are not available. Debt in specification (3) (Table 3.3) corresponds to the sum of all deficits incurred by the city since 1984 until December of t?1.Deficit (specification (2)) is the deficit at the end of the previous year. To construct the fiscal dependence indicator, I first calculate the average share of total revenue that is represented by transfers of revenue from other levels of government. I use the average over all regional units, because the decentralization e?ect I try to account for is a process dictated by national law. Let this average fiscal dependence for year t be denoted as f t . The Fiscal Dependence index used in the regressions is calculated as: FD t =ln(f t )?ln ? 2001 X t=1984 f t T ! where T is the total number of years. The FD t index is therefore close to 0 in years of intermediate decentralization, positive in years of higher decentralization, and negative in years of lower decentralization. In the regressions, I interact FD t with the trend variable, to di?erentiate the trend e?ects related to the process of fiscal decentralization from any other trend e?ects. For the pre-1997 elections, I use electoral results recorded in the National Planning Department Databases, while for 1997 and 2000 I use o?cial results directly provided by the Registradur?a Nacional. C.1.2 Data for national level estimations National level fiscal data were taken from several issues of the Revista Informe Financiero of the Contralor?a General.Iusethefigures for ?Agreements?, which correspond to payments the government is committed to make in the period. To 162 make the investment series consistent over time, the contributions of the central level to the local governments were always included in the definition of invest- ment. Similarly, the definition of current transfers always includes ?operation contributions?. The unemployment rate series is from DANE (the National Bureau of Statis- tics). The original series has missing values for 78.2, 78.4, and 80.2, which filled using the average of adjacent quarters. GDP is only available for 1977.1-1999.4, and no unique quarterly series covers the whole period; for 1977-1995 there is a series from the National Planning Department, while DANE has been in charge of reporting quarterly GDP since 1994. Following a practice that has become standard when working with Colombian data I construct a unique series from the two by seasonally adjusting the pre-94 series to make it compatible with the DANE series, and using growth rates from one to extend the other. C.2 Matchingpoliticalmovementsandthemain parties I use information in Pach?n (2002), as well as informal consultations, to link some movements to the traditional parties. I also consider as liberal (conservative) a movement with the word ?Liberal? (?Conservative?) in its name. Table C.1 lists the matches obtained (the left column lists the name of the party as it appears in the o?cial records of election results): 163 Party or movement Mapping to larger parties Partido Liberal Colombiano Liberal Alternativa Liberal Liberal Apertura Liberal Liberal Convergencia Popular Liberal Liberalismo Ind. De Restauracion Liberal Mov. Convergencia Popular Civica Liberal Mov. Indepedinte Liberal Mil Liberal Mov.Renovador De Accion Liberal-Mor Liberal Nuevo Liberalismo Liberal Partido Conservador Colombiano Conservative Mov. Conservatismo Independiente Conservative Mov. De Integracion Regional Conservative Mov. De Participacion Popular Conservative Mov. Humbertista Conservative Mov. Nal. Conservador Conservative Mov. Unico De Ren. Conservadora Conservative Movimiento De Salvacion Nacional Conservative Movimiento Nueva Fuerza Democratic Conservative Movimiento Progresismo Democratico Conservative Movimiento Unionista Conservative Ad M-19 M-19 Union Patriotica Up Up Table C.1: Party correspondences 164 C.3 Election dates C.3.1 Local elections The first mayor elections are in March 1988. Before 1994, local elections occurred every two years, but the frequency has been extended to three years since then. Table C.2 lists years of elections. C.3.2 National elections Elections every four years at predetermined dates. Table C.3 lists the elections that occur within the period covered by the aggregate fiscal data. 165 Election delec=1 in: March 1988 1987 March 1990 1989 March 1992 1991 October 1994 1994 October 1997 1997 October 2000 2000 Table C.2: Dates of elections for Mayors Election delec=1 in: May 1976 1975.4 and 1976.1 May 1978 1977.4 and 1978.1 May 1982 1981.4 and 1982.1 May 1986 1985.4 and 1986.1 May 1990 1989.4 and 1990.1 May 1994 1993.4 and 1994.1 May 1998 1997.4 and 1998.1 May 2002 2001.4 and 2002.1 Table C.3: Dates of elections for President 166 REFERENCES Alesina, A. (1987), ?Macroeconomicpolicy ina two party systemasa repeated game?. Quarterly Journal of Economics, vol. 102. pp 651-678. Alesina, A., R. Perotti and J. Tavares (1998), ?The Political Economy of Fiscal Adjustments?. Brookings Papers on Economic Activity,1. Anderson T.W. and C. Hsiao (1982), ?Formulation and Estimation of Dy- namic Models Using Panel Data?. Journal of Econometrics, vol. 12, pp. 67-82. Arellano, M. and S. Bond (1991), ?Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations?. Review of Economic Studies, vol. 58, pp. 277-297. Backus, D. and J. Dri?ll (1985), ?Inflation and Reputation ?. American Economic Review vol. 75, pp. 530-538. Barro, R. (1986), ?Reputation in a Model of Monetary Policy with Incomplete Information?. Journal of Monetary Economics, vol. 17, pp.1-20. Barro, R. and D. Gordon (1983), ?Rules, Discretion, and Reputation in a Model of Monetary Policy?. Journal of Monetary Economics, vol. 12, pp101- 121. Blinder, A. (1998), Central Banking in Theory and Practice.MITPress. Brender, A. (2003), ?The E?ect of Fiscal Performance on Local Government Election Results in Israel: 1989-1998?. Journal of Public Economics, vol. 87, pp.2187-2205. Brender, A. and A. Drazen (2003), ?Where Does the Political Budget Cycle Really Come From??. Mimeo, Tel Aviv University. Dal B?, E. (2002), ?Committees and Supermajority Voting: Balancing Com- mitment and Flexibility ?. Mimeo, New College, University of Oxford. 167 Dixit, A. and J. Londregan (1996), ?The Determinants of Success of Special Interests in Redistributive Politics?. Journal of Politics, vol. 58, pp. 1132-1155. Drazen, A. (2000a), Political Economy in Macroeconomics. Princeton Uni- versity Press. _______ (2000b), ?The Political Business Cycle After 25 Years?. NBER Macroeconomics Annual, 2000. Ersenkal, C., M.S. Wallace, and J.T. Warner (1985), ?Chairman reappoint- ments, presidential elections and policy actions of the Federal Reserve?. Policy Sciences, vol.18, pp.211-225. Faust, J. (1996), ?Whom Can We Trust to Run the Fed? Theoretical Support for the Founders? Views?. Journal of Monetary Economics, vol. 37, pp. 267-283. Gonz?lez, M.(1999), ?Political Budget Cycles and Democracy: A Multi- Country Analysis?. Working paper, Department of Economics, Princeton Uni- versity. _______ (2001), ?Do Changes in Democracy a?ect the Political Budget Cycle? Evidence from Mexico?. Review of Development Economics,vol6. Havrilesky, T. (1996), The Pressures on American Monetary Policy.Kluwer Academic Publishers. Kneebone, R. and K. McKenzie (2001), ?Electoral and Partisan Cycles in Fiscal Policy: an Examination of Canadian Provinces?. International Tax and Public Finance, vol. 8, pp 753-774. Lohman, S. (1997), ?Partisan Control of the Money Supply and Decentralized Appointment Powers?. European Journal of Political Economy, vol.13, pp. 225- 246. Pach?n, M (2002), ?El Partido Conservador y sus Din?micas Pol?ticas?. in 168 Guti?rrez, F. Degradaci?n o Cambio: Evoluci?n del Sistema Pol?tico Colombiano. Editorial Norma. Peltzman, S. (1992), ?Voters as fiscal conservatives?. Quarterly Journal of Economics, vol. 107, pp. 327-261 Persson T. and G. Tabellini (2002), ?Do Electoral Cycles Di?er Across Polit- ical Systems??. Mimeo, Stockholm University. Rogo?, K.(1985), ?The Optimal Degree of Commitment to an Intermediate Monetary Target?. Quarterly Journal of Economics,vol. 100, pp. 1169-1190. _______ (1990), ?Equilibrium Political Budget Cycles?. American Eco- nomic Review, vol. 80, pp. 21-36. Rogo?, K. and A. Sibert (1988), ?Elections and Macroeconomic Policy Cy- cles?. Review of Economic Studies, vol. 55, pp. 1-16. Schuknecht, L. (1994), ?Political Business Cycles and Expenditure Policies in Developing Countries?. IMF Working Paper 121, 1994. Shi, M. and J. Svensson (2000), ?Political Business Cycles in Developed and Developing Countries?. The World Bank Working Paper, February. _______ (2002), ?Conditional Political Budget Cycles?. CEPR Discussion Paper #3352. Sibert, A. (1999), ?Monetary Policy Committees: Individual and Collective Reputations?. CEPR Working Paper. Tirole, J. (1996), ?A Theory of Collective Reputations (with applications to the persistence of corruption and to firm quality)?. Review of Economic Sudies, vol 63, pp. 1-22. Von Hagen, J. and S?ppel, R. (1994), ?Central Bank Constitutions for Mon- etary Unions?. CEPR Discussion Paper. 169 Waller, C. (1989), ?Monetary Policy Games and Central Bank Politics?. Jour- nal of Money, Credit and Banking, vol. 21, pp. 422-431. _______ (1992), ?A Bargaining Model of Partisan Appointments to the Central Bank?. JournalofMonetaryEconomics, vol. 29, pp. 411-428. Weingast, B., K. Shepsle and C. Johnsen (1981), ?The Political Economy of Benefits and Costs: A Neoclassical Approach to Distributive Politics?. Journal of Political Economy, vol.89, pp. 642-64. 170