Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Democratzaton and clentelsm: Why are young democraces badly governed? Phlp Keefer Development Research Group The World Bank pkeefer@worldbank.org The comments of Scott Gehlbach and George Clarke and partcpants at a semnar at the Unversty of Calforna, San Dego, are gratefully acknowledged. Ths paper and ts conclusons are entrely those of the author and not those of the World Bank or ts drectors. Abstract: Ths paper dentfes and eplans systematc performance dfferences between younger and older democraces: younger democraces are more corrupt; ehbt less rule of law, lower levels of bureaucratc qualty and lower secondary school enrollment; and spend more on publc nvestment and government workers. One eplanaton for ths s that poltcans n young democraces are less credble. Keefer and Vlacu (004) argue that the nablty of poltcal compettors to make credble promses to ctzens leads them to underprovde publc goods, overprovde transfers to narrow groups of voters, and engage n ecessve rent-seekng. A varety of tests suggest that ths s the only theory that eplans the performance of young democraces. The effect of democratc age remans large even after controllng for the possbltes that voters are less well-nformed n young democraces, that young democraces have systematcally dfferent poltcal and electoral nsttutons, or that young democraces ehbt more polarzed socetes World Bank Polcy Research Workng Paper 3594, May 005 The Polcy Research Workng Paper Seres dssemnates the fndngs of work n progress to encourage the echange of deas about development ssues. An objectve of the seres s to get the fndngs out quckly, even f the presentatons are less than fully polshed. The papers carry the names of the authors and should be cted accordngly. The fndngs, nterpretatons, and conclusons epressed n ths paper are entrely those of the authors. They do not necessarly represent the vew of the World Bank, ts Eecutve Drectors, or the countres they represent. Polcy Research Workng Papers are avalable onlne at http://econ.worldbank.org.. WPS3594
Democratzaton and clentelsm: Why are young democraces badly governed? Phlp Keefer Many democraces fall short of many autocraces n the provson of publc servces or the protecton of economc or human rghts. Recent contrbutons to the lterature on democracy underlne how puzzlng ths s. Acemoglu, et al. (00) and others, for eample, argue that unversal suffrage, compettve electons and restrants on the eecutve branch should have a profound effect on the securty of property rghts. Engerman and Sokoloff (00) argue that the epanson of the franchse should open the way to wder access to educaton. The evdence, though, ndcates that the mere fact of electons and poltcal checks and balances seems to guarantee lttle n terms of government performance, just as a large lterature fnds lttle or no relatonshp between democracy and economc growth. A natural eplanaton for uneven democratc performance s the etreme heterogenety of democraces. Whch precse aspect of heterogenety that matters most for government performance s stll not clear, however. Indeed, even defnng the dmensons of heterogenety has proven dffcult. For eample, classc contrbutons from an older lterature (e.g., Huntngton 97) argue that democraces dffer sgnfcantly n the etent to whch they are poltcally nsttutonalzed. Ths term ressts easy operatonalzaton. The argument and evdence n ths paper pont to two key characterstcs of democraces that eplan performance dspartes among democraces: ther age and the ablty of poltcal compettors to make credble promses to voters. The fndngs below are the frst to demonstrate that young democraces perform systematcally dfferently than older democraces across a range of polcy areas, from corrupton controls to government
ownershp of newspapers to secondary school enrollment. The mportant queston, however, s why ths s so. The analyss below supports a new eplanaton: the relatve nablty of poltcans n younger democraces to make credble promses to voters pror to electons eplans the precse dfferences n polcy choces of poltcans n younger and older democraces. Keefer and Vlacu (004) argue that poltcal compettors who are unable to make credble promses to most voters wll, upon takng offce, underprovde publc goods, overspend on transfers to narrow groups, and engage n sgnfcant rent-seekng. That s, the behavor of such poltcans can be characterzed as hghly clentelst. However, poltcans n older democraces have had greater opportunty than ther counterparts n younger democraces precsely to buld up polcy reputatons across a wde swath of the electorate. To the etent that credblty and age of democracy are related, though, younger democraces should ehbt the same behavor under-provson of publc goods, over-spendng on transfers to narrow groups of ctzens and hgh rent-seekng predcted to emerge n countres where poltcans are credble to only lmted numbers of voters. In fact, the tests below, across seven polcy dmensons, reveal that younger democraces ehbt precsely ths behavor. Ths fndng s robust to numerous specfcatons, ncludng poltcal and electoral nsttutons, levels of voter nformaton, ethnc fractonalzaton and the potental conflatng nfluence of democratcally-elected leaders who ehbt long tenures. Moreover, evdence presented below suggests that one common measure of voter nformaton, newspaper crculaton, may n fact reflect n large measure the underlyng credblty of poltcal promses to ctzens. Ths suggests a potental alternatve or complementary nterpretaton of studes that have lnked hgher newspaper crculaton to such outcomes as low corrupton (e.g., Adserà, et al. 003).
3 The credblty of pre-electoral promses and government performance The etreme heterogenety of democratc performance s easly demonstrated. The rule of law n half of all countres ehbtng ether checks and balances or compettve electons n the 990s was the same or worse as n the medan country lackng ether one or the other. Comparsons usng corrupton ndcators yeld smlar results, as do comparsons usng objectve ndcators of performance. More than 0 percent of countres ehbtng strong checks and balances and compettve electons ehbt worse gross secondary school enrollment than the medan country that does not. Lndert (003) documents ths as well, showng that full democraces are more lkely than elte democraces to provde prmary educaton. Baum and Lake (003) fnd no relatonshp between democracy and female secondary school enrollment. The sources of ths vared performance reman a contnung puzzle of growng mportance, snce the number of countres holdng compettve electons has doubled, from 53 to 0 between 985 and 000, and the number ehbtng some checks and balances has rsen from 6 to. What eplans the heterogenety of democratc performance? Keefer and Vlacu (004) pont to the credblty of pre-electoral poltcal promses. Most models assume ether that all pre-electoral promses are credble to all voters, or none are (see Persson and Tabelln 000 for a revew). There are, however, scarcely any democraces n whch these polar cases apply. Keefer and Vlacu (004) argue that when poltcal compettors are less able to make credble promses to voters regardng publc good provson and economc performance generally, they attempt ether to buld ther credblty among smaller groups of The rule of law measure s from Poltcal Rsk Servces Internatonal Country Rsk Gude and the measures of checks and balances and compettve electons from the Database on Poltcal Insttutons. These are dscussed below. Based on World Development Indcators PPP adjusted ncome per capta, and usng the Database on Poltcal Insttutons varables Eecutve Inde of Electoral Competton (EIEC) and Legslatve Inde of Electoral Competton (LIEC), and Tenure of System (tensys) whch are eplaned n more detal below.
4 voters or to rely on patrons who, n turn, can make credble promses to clents. Ether strategy nfluences the polcy makng ncentves of poltcans. Specfcally, n low-credblty countres where poltcans are relant on patrons or have succeeded n establshng credble relatonshps wth a few voters poltcans focus on government polces that beneft small, targeted groups of voters, gve shorter shrft to publc good provson (polces that beneft voters as a whole) and are more free to engage n rent-seekng. The effort of non-credble poltcans to establsh credblty wth small groups of voters mrrors the well-documented relatonshp n the lterature between patrons and clents. Scholars who have closely studed patron-clent relatonshps n the feld have found that they are personalzed, on-gong and recprocal characterstcs suffcent for reputatonal equlbra to est n a non-cooperatve game. For eample, Scott (97, 9) characterzes patron-clent relatons n Southeast Asa as ones n whch an ndvdual of hgher socoeconomc status (patron) uses hs own nfluence and resources to provde protecton or benefts, or both, for a person of lower status (clent) who, for hs part, recprocates by offerng general support and assstance, ncludng personal servces, to the patron. The analyss here eplans why clentelsm at the socal and local levels enters poltcs at hgher levels: the credblty of patrons wth respect to clents makes patrons valuable alles of poltcal compettors. However, snce patrons only value publc polces that transfer benefts to ther clents, poltcal relance on patrons also leads non-credble poltcans to underprovde publc goods. 3 The mplcatons for young democraces are mmedate. Poltcal compettors n younger democraces have had less chance to develop polcy reputatons wth voters and 3 Patron preferences here are related to a generc problem n poltcal economy. When poltcans have narrow consttuences, t s easer for consttuents to gve credt to poltcans for publc servces that the consttuents and no one else receves than for publc servces that beneft all ctzens and that many poltcans could have had a role n provdng (e.g., Mayhew 974).
5 ther poltcal partes are lkely to be less well-establshed as vehcles for conveyng credble polcy stances. As a consequence, they should be more susceptble to relance on patrons as a means to establsh credble bonds wth voters. Ths mples that the polcy choces of young democraces should resemble most closely those predcted by the foregong arguments: lower levels of publc good provson, hgh levels of prvate, targeted goods, and hgh rent-seekng. It s ths hypothess that s the subject of the emprcal analyss below. Qualtatve evdence for the pervasveness of clentelsm n young democraces The concluson that poltcal compettors n young democraces are less credble, more relant on patrons, and more lkely to focus publc polcy on transfers and rent -seekng than broad publc good provson fnds substantal mplct support n the case study lterature. Varous contrbutors n Malloy and Selgson (987), lookng at countres eperencng the transton from authortaran to democratc government, repeatedly note the relance of new poltcal compettors on narrow benefts to targeted consttuences. Conaghan, for eample, characterzes the partes of the young Ecuadoran democracy as fundamentally clentelst (p. 57), and Rosenberg descrbes poltcal decson makng n young Central Amercan democraces as personalzed and based on vertcal patronage networks (p. 97). The democratc regme that succeeded the authortaran government of Getulo Vargas n post-world War II Brazl was tself soon replaced n 964 by the mltary. One of the mltary s purported ams n replacng ths young democracy was to create the condtons for the ntroducton of a clean democracy, one n whch the ctzenry were free of clentelst tes to poltcal bosses and where rural voters were not controlled by country bosses (Duncan Baretta and Markoff, 53). Sayar (977) wrtes that n the early years of Turksh democracy n the 940s, party strateges for peasant moblzaton were based largely on the recrutment of notables nto
6 party ranks who were then entrusted wth the task of provdng ready vote banks...ths strategy met a favourable resopnse from the notables snce assumng the leadershp post of a party s local unt meant that a notable could (a) gan addtonal status and prestge vs-à-vs rval notables, (b) secure new sources of outsde support for members of hs facton, and (c) mantan and mprove hs economc standng through party tes. (p. 07). These notables were the at the heads of etended clentelst networks. Sayar notes the mportance to partes of provd ng ndvdualzed assstance: frst, n navgatng the bureaucracy (whch are relayed to local party leaders or deputes, 08) and, second, n the provson of publc nvestment for rural development projects (08). Polcy varables that capture dfferences between young and old democraces More systematc emprcal evdence also supports the clam that poltcans n younger democraces prefer transfers to narrow groups of voters, engage n greater rentseekng and underprovde publc goods. Seven varables are used here to represent these dfferent dmensons of government polcy. A common measure of corrupton captures rent-seekng tendences; measures of bureaucratc qualty, the rule of law, secondary school enrollment and government ownershp of newspapers are employed to track government ncentves to provde publc goods or goods that are dffcult to target to partcular groups of voters; and publc nvestment and the central government wage bll are used to capture polces that are easy for poltcans to target to specfc voters at the epense of voters generally. Each of these s eplaned n greater detal below. Measurng rent-seekng: Corrupton Rent-seekng the dverson of economc resources to the prvate requrements of poltcal decson makers s a key measure of government ncentves to satsfy broad socal nterests. Measures of corrupton are the most-used proes for rent -seekng. One that has
7 been used n a number of artcles (e.g., Knack and Keefer 995) and that, unquely, has ample country and year coverage, s the corrupton ndcator from Poltcal Rsk Servce s Internatonal Country Rsk Gude. Ths s a subjectve measure of the etent to whch brbes are a sgnfcant determnant of government decson makng. The measure has broad country coverage, and s avalable snce 986. All of the Poltcal Rsk Servce varables are scaled so that hgher values are better. Hence, hgher values of the corrupton varable sgnfy reductons n corrupton. Measurng publc or non-targeted good provson: The rule of law and the qualty of bureaucracy The rule of law s easly seen as a consequence of government ncentves to respond to broad or narrow nterests n socety. 4 When the rule of law prevals, the umbrella of secure property, contractual and other rghts etends over all ctzens. The benefts of the rule of law faster growth, for eample smlarly etend to all ctzens. However, when specal nterests (ncludng poltcans themselves) can use the power of government to abrogate the government s oblgatons to average ctzens for eample, to protect the property or contractual rghts of average ctzens the rule of law s weak. 5 The costs of weak rule of law are borne by ctzens n general. Government protecton of property and contract rghts can therefore be seen as a polcy that has publc good characterstcs and s not easly targeted to partcular groups of ctzens. The rule of law measure from Poltcal Rsk Servce s Internatonal Country Rsk Gude s a common measure of the etent of 4 The lterature (North and Wengast 989, Acemoglu, et al. 00, Clague, et al. 996) more typcally argues that secure property and contractual rghts are the product of poltcal checks and balances that create nsttutonal road blocks to government efforts to eproprate. Emprcally, however, poltcal checks and balances are weak predctors of the property rghts measures used here. The dvergent ncentves of poltcans to pursue broad publc nterests, the core problem n the analyss here, eplans why checks and balances may not always be a strong guarantee of the securty of property rghts. 5 Ths defnton of rule of law s not the only one. However, the specfcs of the defnton ft squarely n any defnton, and are tracked by the Poltcal Rsk Servces rule of law ndcator.
8 government credblty and ts wllngness to respect property rghts of ndvduals (see, e.g., Knack and Keefer 995 and Acemoglu, et al. 00, Clague, et al. 996). Hgh bureaucratc qualty s another government polcy wth publc good characterstcs. Lower bureaucratc qualty mples that the qualty of publc servces offered generally by government to the average ctzen s low, and that the average ctzen confronts more obstacles to dong busness wth government. At the same tme, those obstacles can always be elmnated for favored consttuents of the government. Hence, bureaucratc qualty s a useful ndcator of the tradeoff that governments make between the pursut of general and prvate nterests. It can, agan, be measured usng the eponymous varable from Poltcal Rsk Servce s Internatonal Country Rsk Gude. Measurng publc or non-targeted good provson: secondary school enrollment These measures of publc or non-targeted good provson are subjectve. The best objectve measure would be one that quantfed government spendng on publc or nontargeted goods. Unfortunately, even when we can dentfy categores of government spendng that appear to be untargeted or dedcated to publc good provson, the actual destnaton of these funds s often targeted. For eample, educaton spendng can be amed at rasng achevement for all chldren (through currculum reforms, testng, spendng on hgh qualty teachers); or t can be targeted (by buldng new schools n some areas, but not n others). Cross-country budget data on educaton do not dstngush whch. To crcumvent ths dffculty, an outcome ndcator, secondary school enrollment, s used below to appromate the etent to whch the publc component of educaton spendng predomnates. 6 If poltcans care relatvely more about poltcal targetng than they do about 6 Prmary school enrollment could, n prncple, also be used, but most countres send most of ther chldren to prmary school The coeffcent of varaton n the sample used here s 7.6 percent for prmary school enrollment, and 50.8 percent for secondary school enrollment.
9 provdng the publc good of qualty educaton to all chldren, the overall qualty of schoolng should suffer. For eample, they mght locate schools or hre teachers based on crtera other than the educatonal payoff to those decsons. As qualty falls, famles should demonstrate ncreasng reluctance to ncur the fnancal and opportun ty costs of sendng ther chldren to secondary school and secondary school enrollment should fall. Confdence n the valdty of secondary school enrollment as a proy for publc good provson should be hgh f t s the case that the relatonshp between (targetable) spendng and learnng outcomes s low, f the relatonshp between (non-targetable) educaton reforms and learnng s hgh, and f learnng outcomes are correlated wth famly decson to enroll chldren. The estng evdence s generally supportve of these relatonshps. A large fracton of educaton spendng goes to teachers. The evdence suggests lttle clear relatonshp between teacher spendng and learnng qualty, however. Frst, teacher absenteesm s a notorous problem n developng country educaton systems and has been strongly assocated wth student absenteesm (see Ehrenberg, et al. 99 on the Unted States; Carlson 000 on Chle, and Harber 989 on Ngera). Student absenteesm s obvously an ndcaton of low learnng. It s also partcularly relevant for the analyss here, where student enrollment s the varable of nterest. Second, teacher qualty seems more closely related to turnover of teachers and the prevalence of frst year teachers (educaton management polces that are dffcult to target) than to teacher compensaton (Hanushek, et al. 005). In sum, teacher compensaton can be hgh, but f t s meant only to beneft teachers as a poltcal consttuency and f there s lttle ncentve to mprove educaton outcomes, compensaton does not lead to hgher teacher attendance or qualty. Addtonal evdence underlnes the postve effects on learnng of (relatvely cheap) management reforms, whch do not target resources to specfc narrow consttuences.
0 Lookng at four Francophone countres, Mchaelowa (00) fnds that a vst by a school nspector n the past year ncreased scores. On the other hand, teacher absenteesm substantally reduces student performance (math/french) for ffth graders, as dd whether a teacher was a cvl servant or a unon member (both of whch are ndcators of the degree of teachers ablty to act collectvely n ther prvate nterests). Moreover, these negatve nfluences on learnng more than offset the postve effects on learnng assocated wth the presence of tetbooks or the lteracy of a chld s parents. 7 Even when the relatonshp between addtonal physcal nputs and enrollment or educatonal achevement has been shown to be postve, the magntude of the effect s small (see, e.g., Mchaelowa 00 and Glewwe and Jacoby 994 on blackboards n four Francophone Afrcan countres and Ghana, respectvely; and Harbson and Hanushek for a broad revew). Glewwe and Jacoby (994) only fnd a sgnfcant effect of the physcal condton of classrooms on student testng, but condton s related to mantenance ependtures, whch are precsely the types of government spendng that are dffcult for poltcans to target (t s dffcult for a poltcan to get credt for the repar of a door or wndow compared to the constructon of a new school). Taken together, ths evdence shows that a focus on qualty or on systems that nfluence outcomes across schools, nether of whch s easy for poltcans to target to partcu lar consttuences, are crtcal to learnng, but fnancal nputs nto educaton are less clearly assocated wth learnng. However, s low learnng, the unobservable publc good n cross-country analyss, assocated wth low enrollment, the varable that can be observed? Here there s less evdence, but t supports the contenton that low learnng envronments 7 Ballou (996) also concludes that a good academc record does lttle to boost an applcant s chances to be hred as a teacher n the Unted States and argues that ths s because admnstrator ncentves are too weakly lnked to classroom performance and student achevement to persuade them to place weght on a canddate s academc success, even when ths s assocated w th better classroom performance.
drve down enrollment. One pece of evdence s reported above: teacher and student absenteesm move together. Bommer and Lambert (000) fnd, n addton, that students enroll earler n Tanzanan schools where school qualty s hgher. In the estmatons below, therefore, gross secondary school enrollment, from World Development Indcators, s taken as an ndcator of ncentves to provde publc rather than prvate and targetable goods. Measurng publc or non-targeted good provson: government ownershp of newspapers The fnal ndcator of publc good provson s related to government polcy towards ctzen nformaton. Snce ctzen nformaton s a publc good, government polces that epand or restrct ctzen access to nformaton should be nfluenced by the same forces that drve publc good provson more generally ncludng the credblty of poltcal promses. One such polcy s the market share of government-owned newspapers as a fracton of the market share of the top fve newspapers n a country, collected by Djankov, et al. (003). Where the market share of government newspapers s hgh, ctzens have less access to a dversty of nformaton sources. Elected governments can more easly deny ctzens ths publc good f publc goods are not electorally relevant, as when poltcal compettors cannot make credble promses to a large fracton of ctzens. It may of cours e also be true that elected governments wth no ntenton of surrenderng power mght also foreclose ctzen access to nformaton. Such a government would pass from the category of compettvely elected, eventually, when t suppressed an electon or won an electon wth a large super-majorty, to the non-compettvely elected category, creatng an assocaton, by defnton, between government ownershp and the contnuous years of electons. In ths case, however, the hghest government share of the newspaper market should be found where elected leaders have the longest tenure. There s, n fact, a strongly postve relatonshp between the tenure of elected eecutves and the
market share of government-owned newspapers. However, evdence reported below shows that the negatve effect of contnuous years of compettve electons on government market share s robust to controls for leader tenure. Government-owned newspapers mght also be seen as a way to ncrease the flow of nformaton to ctzens, for eample f they are heavly subsdzed and dstrbuted below cost. There seem to be few eamples where government newspapers are set up and managed for the purpose of provdng unbased nformaton about, for eample, the performance of elected offcals. However, to the etent that ths s true n some cases, estmates below wll be based towards rejectng the hypothess that poltcal credblty reduces government restrctons on ctzen nformaton. Measurng targeted government spendng: Publc nvestment and the government wage bll Government employment and publc nvestment or nfrastructure spendng are tradtonal ways for governments to target benefts to partcular consttuences. Ths s straghtforward to see n the case of patronage n government employment, whch benefts job holders and ther famles. The government wage bll as a fracton of GDP, taken (lke publc nvestment) from Government Fnancal Statstcs, tracks government spendng on jobs and therefore offers a measure of government ncentves to channel spendng to targeted consttuences. The fracton of socety that benefts from a gven publc nvestment s more varable than jobs spendng. Lke educaton spendng, publc nvestment has both publc good and targetable, prvate components. The key characterstc of publc nvestment s not that ts benefts are entrely prvate and targeted, but that relatve to other categores of publc spendng, poltcans can target a hgher fracton of publc nvestment undert aken every year. Most educaton spendng, for eample, s on teachers and other admnstratve staff. On the
3 margn, poltcans can nfluence whch teachers are hred or where they are posted, but most teachers and postngs are fed from year to year. Publc nvestment spendng, on the other hand, dsproportonately conssts of new constructon that benefts partcular consttuences. Gven the characterstcs of publc nvestment, poltcans nterested n provdng narrowly targeted servces to consttuents should therefore be dsproportonately nterested n ncreasng publc nvestment. In fact, as the phrase pork barrel spendng recalls, poltcal preferences for targetable publc nvestment spendng are well-known. Publc nvestment spendng as a fracton of GDP s therefore used here as a measure of government ncentves to target publc spendng to specfc consttuences. Eplanng dfferences between young and old democraces The key concern here s to assess whether the age of democracy eplans varaton n polcy choces across democraces. Numerous alternatve hypotheses regardng polcy varaton must also be taken nto account, however. Ths secton descrbes the age of democracy varable and the other controls that enter nto the econometrc specfcatons analyzed below. Measurng the age of democracy The puzzle eplored n ths paper s why the nsttuton of compettve electons, a defnng characterstc of democraces, yelds sharply dfferent outcomes across countres. In constructng a measure of the age of democracy, therefore, the core democracy varable s based on the seven-pont Legslatve and Eecutve Indces of Electoral Competton (LIEC and EIEC) from the Database on Poltcal Insttutons (Beck, et al. 00). These varables are objectve and narrow, ecludng, unlke other measures of democracy, ether other nsttutonal features of countres (the presence of checks and balances) or outcomes
4 (whether governments are constraned by formal nsttutons, or whether electons have led to the replacement of an ncumbent by a challenger). The hghest score on each of these two ndces (seven) s assgned to countres n whch multple partes compete n legslatve and eecutve electons and no party receves more than 75 percent of the vote. In all of the dscusson here, the contnuous years of compettve electons are measured as the number of consecutve years n whch a country has the hghest score on both ndces. From 975 000, the perod under study here, the medan number of years of contnuous electons s eleven and the mean 7 among countres that have ever held compettve electons. 8 Other control varables Other geographc and demographc features of countres are lkely to have a drect effect on the knds of polces that ctzens demand or that have the hghest payoff for ctzens. Country land area and populaton capture the demand for publc nfrastructure, snce the demand for roads or the costs of provdng educaton or ensurng the rule of law are surely dfferent n large, sparsely populated than n small, densely populated countres. Countres wth a larger fracton of young people should ehbt a greater demand for educaton. More young people may also ncrease the pressure on poltcans to provde jobs and rase the costs of mantanng the rule of law, n vew of the well-known assocaton between youth and crme. There may be economes (or dseconomes) n the applcaton of polces across a larger populaton; hence, total populaton s taken nto account. 8 DPI begns n 975. To establsh the number of contnuous years of democracy n 974, the age of democracy varable from Clague, et al. (996) s employed. Ther methodology truncates the age of the oldest democraces, so the mamum contnuous years of electons n the year 000 n the analyss below s therefore 70, attaned by 9 countres.
5 Fnally, the demand for, and the costs of provdng, nfrastructure, educaton or even the rule of law and ntegrty n government may dffer substantally across rural and urban areas. These are all taken from World Development Indcators. The estmates of the determnants of gross secondary school enrollment have addtonal controls to solate the publc good aspect of school enrollment. Gross secondary school enrollment s a product of how much governments spend on educaton and of how they spend the money. The goods and servces that governments buy may or may not be publc goods. However, the way governments spend money and the oversght they gve to spendng outcomes are more lke publc goods and less targetable to specfc consttuences. The educaton regressons therefore control for government spendng on educaton n order to solate the mpact of years of contnuous electons on the publc good aspects of school polcy that nfluence enrollment. In addton, gross prmary school enrollment s controlled for, reflectng the smple fact that secondary school enrollments must drop when prmary school enrollments are low; at the same tme, t controls for the possblty that government spendng on educaton mght be hgh, but dedcated dsproportonately to prmary schools. One mght argue that government ependtures should be ncluded as a control varable drectly, snce nearly all of the polcy varables n the estmates of Tables and 3 are related to government ependture. In fact, government ependtures as a fracton of GDP are assocated wth sgnfcantly lower corrupton and hgher publc nvestment and school enrollment. Government ependture s nevertheless ecluded from the core specfcatons because ts ncluson reduces sample szes by appromately 5 percent. Results, though, are robust to ts ncluson. Wherever the electons varable s sgnfcant n Tables and 3 below, t remans sgnfcant when controllng for government ependture as a fracton of GDP.
6 In eamnng the robustness of the results below, a number of other potental poltcal characterstcs of countres are also eamned, ncludng poltcal and electoral nsttutons, ethnc heterogenety, and voter nformaton. Table presents a summary of all of the varables used here. Note that each contnuous epsode of compettve electons consttutes an observaton. All varables are therefore averaged over the duraton of the epsode. Table : Summary statstcs Varable N Mean medan std dev. Rule of law 0 6.87 6.67.48 Share of government-owned newspapers/ market share of top fve newspapers 75.6 0.3 Corrupton n government 00 5.89 5.5 Bureaucratc qualty 0 5.86 5.39.6 Gross seco ndary school enrollment 6.885 65.3 3.87 Publc nvestment/gdp 86 0.039 0.03 0.07 Gov t. wages/gdp 9 0.06 0.055.035 Average contnuous years of compettve electons 37 3.35 5.5 7.65 Ethnolngustc polarzaton 8 0.56 0.038 0.7 Newpaper crculaton per 000 nhabtants 8 9.66 7.99 3. Majortaran () or non-majortaran 37 0.38 0 0.47 Presdental (), Sem-presdental (), or Parlamentary (0) 37.096.69308 0.95 Percent populaton young 33 0.336 0.353 0. Total populaton (0 mllons) 35 33600000 8330978 04000000 Percent populaton rural 34 0.487 0.4975 0. Land (mllons km ) 33 85394 43000 9043 Total government ependtures/gdp 9 0.86 0.8 0. Prmary school enrollment 3 99.47 00.7 7.59 Total educaton ependtures/gdp 5 0.034 0.03 0.0 N.B. Observatons are epsodes of contnuous compettve electons. Varables are the sum of the
7 yearly observatons dvded by the number of years the epsode lasts n the sample (a mamum of 6 years, snce the data run from 975 000). Estmaton strategy The deal test of the hypothess that the credblty of pre-electoral poltcal promses nfluences polcy would be: () Polcy choce = ß 0 + ß (credblty of poltcal compettors )+ X B + e. where ndees the democratc epsode under observaton. Two problems make ths approach dffcult and are the focus of ths secton. Frst, credblty s unobservable. Second, credblty s lkely to be a functon of some of the elements n X, gvng rse to multcollnearty. The prevous dscusson suggests a close relatonshp between the number of contnuous years of compettve electons and credblty, suggestng the electons varable as a natural proy for unobserved credblty. In partcular, because credblty can emerge from reputaton, and reputaton s a possble product of repeated eperence wth poltcans, one mght epect a postve assocaton between the contnuous years of compettve electons and the ablty of poltcans to make credble pre-electoral promses. Gven ths, we can epress credblty as a functon of the electons varable, () Credblty =α 0 +α (contnuous years of compettve electons) + u.. Substtutng () nto () yelds, (3) Polcy choce = ß 0 + ß (α 0 +α (contnuous years of compettve electons) + u.)+ X B + e. Attenuaton bas and multcollnearty It s always the case that the use of a proy varable ntroduces attenuaton bas n the estmate of the coeffcent of the proed varable. That s, because the years of contnuous electons captures the effects of credblty mperfectly, wth error u, the estmate of ß s based downwards. The attenuaton bas naturally ncreases the less close s the connecton
8 between credblty and the electons varable (the greater s σ u ). It s less often observed, however, that attenuaton bas worsens n the presence of multcollnearty. Ths s a partcu lar ssue here because of strong assocatons between credblty and some of the elements n X. To see ths effect, assume that there s only one element n X =, set credblty=,,,,, =. Then the estmate of suppress constant terms, and let ( ) Q β s gven by b = αq β σ u [,, (. ) (, ) ] +, (, ) αq ( ). Recall that the correlaton coeffcent between and credblty= s equal to,,,,, and note that t s nversely related to Q. It s evdent, then, that the greater s the correlaton between the two credblty and other determnants of polcy, the larger s u αq and the greater s attenuaton bas (whch s zero when σ u s zero). 9 The polcy determnant may nfluence credblty drectly or, through ts nfluence on the contnuous years of compettve electons, ndrectly. Effects are smlar n ether case, but the second case s easest to demonstrate. Assume, as s plausble, that the number of contnuous years of compet tve electons s determned by a process lke σ z = γ w + γ, + ν, where w s an eogenous and unobserved varable unrelated to polcy outcomes that nfluences the electons varable. But then the process determnng 9 See Anne for the full dervaton.
9 credblty, equaton (), can be rewrtten as, = α ( γ w + γ, + ν ) + u. The correlaton between =credblty and s gven by αγ. That s, the more that nfluences the contnuous years of electons, the greater s the downward bas n the estmaton of the effect of credblty on polcy outcomes when usng the electons proy. Three of the control varables ncluded below n X are lkely also to nfluence credblty: the fracton of the populaton that s rural, the fracton that s young, and ncome per capta. Not only do they have drect effects on polcy, they also are lkely to have drect and ndrect effects (through the contnuous years of compettve electons) on credblty. Rural voters are more solated and strateges to mprove poltcal credblty wth respect to these voters, such as personal contact wth the voters, are more costly per voter. Ths drectly nhbts the development of credblty. Younger voters have less eperence wth poltcal actors. Eperence s key to the formaton of reputaton and credblty, so t s more costly for poltcans to convey credble promses to them. Income per capta s also lkely to have strong effects on credblty. Poltcans can nvest resources n makng ther appeals to ndvdual consttuents credble, promsng prvate transfers, jobs or assstance n tmes of dstress; or they can nvest n makng ther appeals to large groups credble, buldng the credblty of ther promses to provde publc goods that wll beneft large groups n the most cost-effectve way. The poorer s a country, the more lkely s the frst strategy to be the most cost-effectve way to acheve poltcal support (as Dt and Londregan 996 observe). There s substantal evdence for ndrect effects. The years of contnuous electons s negatvely correlated wth the percent of the populaton that s rural (ρ = -0.5) and the percent that s young (ρ =-.50). It s postvely correlated wth ncome
0 per capta (ρ =.7). 0 To the etent that rural, young or poorer populatons are less wellserved by democraces, they are also less lkely to be more passve n the face of efforts to overthrow democracy. For eample, f rural voters are more lkely to be the captves of patrons, they would be less lkely to beneft from poltcal competton under democracy and less opposed to threats to democracy. To address attenuaton bas and multcollnearty, two strateges are employed. The frst s to use nstrumental varables. These are a well-known sol uton to attenuaton bas due to measurement error (e.g., f credblty were drectly measured, but wth error), where the nstrument s correlated wth true credblty and uncorrelated wth the error term n the observed credblty measure. In the case of proy varables, as here, the vald nstrument must be correlated wth contnuous years of electons, but not wth the error term u. Multcollnearty of the type suggested here dsrupts the effcacy of the IV approach, however. The predcted value of the contnuous years of electons, from the frst stage of the IV estmaton, s a functon of the nstruments for the contnuous years of electons and the controls X from the second stage of the equaton. However, credblty and the contnuous years of electons are partally determned by X. Consequently, the nstrument wll only be vald (sgnfcant n the frst stage regresson of the nstrument on contnuous years of electons) f t has substantal eplanatory power above and beyond that of the varables n X. To the etent that these varables are sgnfcant determnants of the electons varable or of credblty, the nstruments wll fal. In fact, as we mght epect, the 0 The connecton between poverty or agrcultural domnance and democratc nstablty s well-establshed (Przeworsk, et al. 000, Bo and Stokes 003, Bo 00). These relatonshps are typcally eplaned n terms of class conflct, the margnal returns to captal n poor countres and wealth redstrbuton (see Bo and Stokes for a summary). Unlke ths eplanaton, the argument here takes nto account that key aspects of poltcal competton (e.g., whether voters are nformed or not, or poltcans are credble or not) vary systematcally among democraces and that class-based partes and conflct over wealth dstrbuton are rare n most young and poor democraces.
nstruments that are wdely-used n the cross-country nsttutons lterature are rarely sgnfcant n the frst stage regresson n the presence of the varables n X. The second strategy used here s therefore to nvestgate the effects of droppng the collnear varables. The usual crteron to judge the reasonableness of ths step s whether the resultng ncrease n precson s large relatve to the ntroducton of an unknown bas due to the omsson of these varables. Here, because the collnear varables create a sgnfcant source of downward bas n the OLS regressons, a stronger case than usual can be made that the tradeoff s worthwhle. In the estmates below, therefore, estmates of () are therefore undertaken wth and wthout those varables n X for whch there s a strong a pror case that they nfluence both credblty and polcy choce. The change n coeffcent values after droppng these varables s consstent wth the a pror argument that the presence of the varables X eacerbate the attenuaton bas n the OLS regressons and spurously undermne the valdty of nstruments n the IV regressons. Endogenety The IV estmatons below address both measurement error and possble endogenety concerns. These concerns are easy to eaggerate, however. One mght argue, for eample, that causalty s reversed and that polcy choce tself drves regme duraton. That s, governments that choose bad polces trgger greater opposton than those that do not. Ths argument alone s not enough to establsh that reverse causaton drves the results we observe. Frst, bad polcy should lead ctzens to prefer challengers to ncumbents, but not necessarly to prefer autocracy to democracy. One needs, n addton, an argument that suggests why bad polcy should lead to regme change. The credblty arguments here provde precsely such an argument: the less credble are poltcal compettors n a democracy, the less acountable they are and the more ndfferent ctzens should be between
elected and unelected poltcal leaders. But one cannot use ths argument to assert that the results below n Table are spurous snce ths argument and the credblty arguments advanced earler to nterpret Table are entrely consstent wth each other. A dfferent model of reverse causaton mght assert that an omtted varable s a determnant of polcy, though not of regme duraton, but t s polcy that determnes the number of contnuous years of compettve electons. What mght that omtted varable be, however? Polcy s the product of government ncentves; the latent credblty varable s one that eplans the pattern of polces observed across democraces. Agan, then, to the etent that credblty drves polcy choce, t s dffcult to argue that reverse causaton eplans the results observed n equaton (). The lterature suggests other possble latent nfluences. All of these are eamned below, however, and found not to eplan the dfference between democraces of dfferent ages. Fnally, n any case, the IV strategy used here also addresses potental endogenety concerns. Results Publc nvestment, the wage bll of the central government, the rule of law, and bureaucratc qualty are all appromately one standard devaton hgher and secondary school enrollment one standard devaton lower n countres where compettve electons have been held for fewer than the medan number of contnuous years. The market share of government-owned newspapers s 3 percent of the market share of the top fve newspapers n countres wth fewer than the medan years of contnuous compettve electons and close to zero n countres wth more years. These bvarate assocatons emerge strongly n the regressons n Table. For each of the seven polcy varables, Table presents two sets of estmates of equaton (). The frst s a smple regresson controllng for the land area of the country and total populaton
3 and, n the equaton for secondary school enrollment, controllng as well for educaton spendng and prmary school enrollment. The second set of estmates controls, n addton, for the three varables lkely to have strong effects on government credblty: the percent of the populaton that s young or rural, and ncome per capta. The pattern of electon coeffcents across the dfferent polces s eactly consstent wth the credblty eplanaton: rent-seekng falls (recallng that the corrupton s worst when the corrupton varable s lowest) and government provson of publc goods (the rule of law, bureaucratc qu alty, the absence of government newspapers and secondary school enrollment) rses the more contnuous years of compettve electons that countres have eperenced. In contrast, government provson of targeted goods through publc nvestment or publc employment falls. The earler argument suggests that the ntroducton of the varables percent populaton that s young or rural and ncome per capta would lead to a drop n the estmate of the electons coeffcent. The second regresson n each of the regresson pars, demonstrates that ths s the case: n the presence of these three varables, the estmates of the electons coeffcent, though stll hghly sgnfcant, drop substantally n magntude. The magntude of the electon or credblty effect s n most cases large. For eample, n the more parsmonous specfcatons n Table, a one standard devaton ncrease n the contnuous years of compettve electons (appromately 0 years) s assocated wth a reducton n corrupton of more than one-half a standard devaton, ncreases the rule of law by one-thrd of a standard devaton and bureaucratc qualty by If the three control varables are determnants of poltcal credblty, ther sgns should be consstent wth the credblty arguments made earler. In all those cases where the estmated coeffcents of the varables are sgnfcant, ther sgns are consstent wth the credblty hypothess: corrupton s lower, publc good provson s hgher, and fewer targeted goods are provded the greater s ncome; corrupton and the government wage bll are hgher and some publc good provson s lower the younger s the populaton.
4 more than one-half a standard devaton, and secondary school enrollment by one-fourth of a standard devaton. Clague, et al. (996) and Tresman (000) fnd that measures of the age of democracy are postvely related to the rule of law and corrupton but do not eplan the fndng. The results here and below etend ther fndngs n three ways. Frst, the results here reflect tests of specfc eplanatons of why age of democracy should matter for poltcal ncentves to mantan the rule of law or to seek rents years of electon s not smply an addtonal control varable. Second, the estmates here are based on an objectve measure of democracy that emphaszes the specfc nsttuton of votng rather than broader characterzatons of democracy. Fnally, the results here are found to be robust to a number of alternatve eplanatons.
Table : Effect of the persstence of compettve electons (OLS) Dependent varable: Contnuous years of compettve electons Corrupton n Government.088 (.00).05 (.00) Rule of law.083 (.00).0 (.03) Bureaucratc qualty.0 (.00).03 (.00) Market Share, Gov t-owned newspapers -.005 (.00).00 (.07).93 (.00) Gross secondary school enrollment.4 (.0) Central gov t. wage bll/gdp -.0005 (.00) -.0005 (.0) Publc nvestment/gdp -.0003 (.00) -.00006 (.59) Total populaton (0 mllons) -.4 (.003) -.3 (.0) -.7 (.0) -. (. 08). (.8). (.00) -.34 (.0) -.49 (.00) -3.5 (.00) -.64 (. 38) -.006 (.00) -.007 (.00) -.0003 (.00) -.0005 (.00) Land (mllon km ) -.0 (.8) -.4 (.8).9 (.39).06 (.9).3 (.5).3 (.6) -.03 (.65) -.0 (.9) -.9 (.86) -5.6 (. 3) -.03 (.00) -.03 (.03) -.00 (.00) -.000 (.9) GDP/capta (real, PPP-adjusted, thousands) Percent populaton young Percent populaton rural.083 (.6) -7. (.0). (.4).5 (.00) -.9 (.00).5 (.3).8 (.00) -. (.45).7 (.4).005 (.47).7 (.003).5 (.64).84 (.06) -9.4 (.00) -.89 (.4).0 (.3).3 (.0).05 (.3).00 (.9).0 (.54).03 (.05) Prmary school enrollment Total educaton ependtures/gdp R.44 (.03) -9.6 (.53).6 (.004).57.66.43.7.55.73.5.40.40.86..30.7.3 99 97 00 98 00 98 74 74 07 07 90 90 86 85 N N.B. p-values n parentheses. Each observaton s the average across a country-regme, begnnng n the frst year a country has compettve electons (or 975) and endng n the last year that a country has compettve electons (or 000). Robust standard errors are reported, clustered by country so that multple country-regmes from the same country are not treated as ndependent observatons. All regressons nclude a constant (not reported). Instruments n the IV regresson marked wth * do not satsfy the ecluson assumpton. 39.5 (.)
Table 3 reports the IV estmates of the regressons n Table usng two nstruments, lattude and Brtsh colonal hertage. These estmates are meant to address both possble endogenety ssues and measurement error due to the use of the electons proy varable. Wth respect to endogenety, nether nstrument s plausbly rel ated to polcy choces from 975 to 000, the years over whch data are avalable for the estmatons. Wth respect to measurement error, nether seems plausbly related to the measurement error u that emerges from usng the electons varable as a proy for credblty. At the same tme, lattude (or geographc locaton more generally) and colonal eperence are lkely to capture unobserved factors that mght nfluence whether countres become democratc and how long they are able to reman democratc. 3 Recallng the earler dscusson, the strong relatonshp between years of contnuous electons, credblty and the controls for the two populaton varables and ncome per capta mean that no nstruments sgnfcantly predct the electons varable n the presence of these controls. Table 3 results are based on the more parsmonous specfcatons n Table. In every case, the coeffcent estmate of the electons varable s larger n the IV estmates n Table 3 than n the correspondng estmates n Table. The use of nstruments therefore seems to reduce attenuaton bas due to measurement error, suggestng that measurement error downward bas rather than endogenety-drven upward bas s the man concern arsng from the OLS regressons of Table. 3 These are two of several varables that are standard n the lterature lnkng cross country outcomes to nsttutonal varables. Others (such as settler mortalty, fracton of Englsh speakers, years snce ndependence) fal as nstruments because they are nsgnfcant determnants of the contnuous years of compettve electons.