Relying on the private sector: The income distribution and public investments in the poor

Journal of Development Economics 107 (2014) 320–342

Contents lists available at ScienceDirect

Journal of Development Economics

j ourna l homepage: www.e lsev ie r .com/ locate /devec

Relying on the private sector: The income distribution and publicinvestments in the poor

Katrina Kosec ⁎Stanford University, Graduate School of Business, 655 Knight Way, Stanford, CA 94305, USA

⁎ 2033 K Street NW,Washington, DC 20006, USA. Tel.:467 4439.

E-mail address: [email protected].

0304-3878/$ – see front matter © 2013 Elsevier B.V. All rhttp://dx.doi.org/10.1016/j.jdeveco.2013.12.006

a b s t r a c t
a r t i c l e i n f o
Article history:Received 25 October 2011Received in revised form 6 December 2013Accepted 14 December 2013

JEL classification:D78H42H75I24O15

Keywords:Fiscal federalismLocal government expendituresPolitical economyEducation policyMunicipalities

What drives governments with similar revenues to provide very different amounts of goods with private sectorsubstitutes? Education is a prime example. I use exogenous shocks to Brazilian municipalities' revenue during1995–2008 generated by non-linearities in federal transfer laws to demonstrate two things. First, municipalitieswith higher income inequality or highermedian income allocate less of a revenue shock to education and are lesslikely to expand public school enrollment. They are more likely to invest in public infrastructure that is broadlyenjoyed, like parks and roads, or to save the shock. Second, I find no evidence that the quality of public educationsuffers as a result. If anything, unequal and high-income areas aremore likely to improve public school inputs andtest scores following a revenue shock, given their heavy use of private education. I further provide evidence thatan increase in public sector revenue lowers private school enrollment.

© 2013 Elsevier B.V. All rights reserved.

1. Introduction

What drives governments with similar revenues to publicly providevery different amounts of goods with private sector substitutes? Educa-tion and health care are prime examples; in almost all countries, thepublic and the private sectors simultaneously provide versions ofeach. Because access to them may have profound impacts on growthand poverty, it is important to understand what factors lead govern-ments to provide more of the public version versus relying more onthe private version. This is especially so since the poor frequently cannotafford a private version. I show that in areas with higher incomeinequality or higher median income, governments allocate less of anexogenous revenue shock to goods with private substitutes (specifical-ly, education) andmore to goodswithout private substitutes (like parksand roads), which aremore broadly enjoyed. Unequal and high-incomeareas are also more likely to save a revenue shock, running a budgetsurplus rather than raising consumption in the short run. Importantly,however, I find no evidence that the quality of public education suffersdue to this lower propensity to invest in it. If anything, unequal andhigh-income areas are more likely to use a revenue shock to improve

+1 202 421 3393; fax: +1 202

ights reserved.

public school inputs and standardized test scores, since their heavyuse of private education ensures that public education resources arespread less thinly.

My results are consistent with the political economy models ofBarzel (1973), Besley and Coate (1991), Epple and Romano (1996),Glomm and Ravikumar (1998), and de la Croix and Doepke (2009).They hypothesize that areas with more income inequality have morepeople who consume private sector versions of publicly-providedgoods. Because they are consuming private versions, they vote for lowspending on the public sector counterparts. Under majority voting, thegovernment thus provides less of goods with private substitutes. Thiscontrasts somewhat with Meltzer and Richard (1981, 1983), Alesinaand Rodrik (1994), and Persson and Tabellini (1994), who predict thatthe size of public expenditures increaseswith inequality, asmean incomerises relative to the income of themedian (decisive) voter. I partly recon-cile these models by showing empirically that the type of public in-vestment matters. Government investment in education, whichhas private substitutes and predominantly benefits the poor, be-haves according to the first set of political economy models. Govern-ment investment in broadly-enjoyed public goods that lack privatesector substitutes (like parks, roads, and other infrastructure) behavesaccording to the Meltzer–Richard model. My results are also consis-tent with work modeling control over public policy increasing in in-come. As inequality grows, the bottom half of the income distribution

http://crossmark.crossref.org/dialog/?doi=10.1016/j.jdeveco.2013.12.006&domain=pdf

http://dx.doi.org/10.1016/j.jdeveco.2013.12.006

mailto:[email protected]

http://dx.doi.org/10.1016/j.jdeveco.2013.12.006

http://www.sciencedirect.com/science/journal/03043878

3 A related literature analyzes the effects of the distribution of income on redistributionmore generally. Casamatta et al. (2000) describes how a positive level of social security ispolitically sustainable since retirees andworkers withmediumwages form amajority co-alition supporting it. Bellettini and Ceroni (2007) describe how the poor (liquidity

321K. Kosec / Journal of Development Economics 107 (2014) 320–342

becomes less able tomake the government invest in publicly-providedgoods that the poor use. Prominent studies in this vein include Pande(2003), Foster and Rosenzweig (2003), Keefer and Khemani (2005),Bardhan and Mookherjee (2006), Banerjee and Somanathan (2007),Araujo et al. (2008), and Karabarbounis (2011).

This paper's primary contribution is an empirical analysis of how anarea's income distribution affects how its government allocates revenuebetween goods with and without private substitutes. It also examinesthe public service quality implications of this differential propensity to in-vest. I focus on education in Brazil, which has abundant private substitutes.The strength of my analysis stems from exploiting a change in Brazil'sschool finance law which generates exogenous variation in revenue.

I focus on education spending in Brazil for three main reasons. First,Brazil has over 5,000municipalities with enormous discretion over howmuch to invest in education. Thismakes it an ideal setting to understandhow the local income distribution affects public investment. Second,Brazilian municipalities differ greatly in their income distributions buthave similar institutions and constraints that make them comparable.Finally, education is the single largest item in the municipal budget(accounting for about 30% of revenue) and is the chief municipally-provided service with abundant private substitutes.1 This allows me tocleanly capture tradeoffs between investment in goodswith vs. withoutprivate substitutes by focusing on education.

The analysis faces a significant challenge to identification. Unob-served factors that affect revenue levels may also influence citizens'preferences over goods. For example, if highly-educated individualsplace a relatively high value on education and have more taxableincome, revenues may be higher precisely where public education ismost valued. Endogenous household sorting is likely to exacerbate iden-tification problems. I circumvent these problems by exploiting a 1998change in Brazil's education finance law that generates exogenous var-iation in municipalities' revenue. Specifically, I form a simulated instru-mental variable that encapsulates the credibly exogenous variation inrevenue generated by the law, but which excludes variation due to mu-nicipalities' own actions. I instrument for actual (endogenous) revenuewith the simulated instrument, thereby testing how different munici-palities spend an exogenous shock to revenue, all else equal.

Briefly, the law—“Fund for the Maintenance and Development ofFundamental Education and Valorization of Teaching” (FUNDEF)—forced each of Brazil's 26 states to gather 15% of each of its municipali-ties' revenue in a state fund. Eachmunicipality in the state then receiveda share of the state's fund equal to its share of public school students inthe state (with students at different grade levels weighted differently indifferent years). Because redistribution took place only within states(and also because of various nonlinearities in the rules), similar munic-ipalities experienced very different changes in their finances. I form asimulated instrument which is the predicted revenue of municipality iin year t. I predict revenue using time variation and discontinuities inthe law's parameters applied to municipalities' pre-law (1997) schoolenrollment and finances. Thus, only the law-induced changes—not mu-nicipalities' responses—are incorporated in the simulated instrument.

The main results of the paper are as follows. An exogenous revenueincrease boostsmunicipal spending and raises public school enrollmentrates. However, more unequal and higher-income municipalities aresignificantly less likely to expand education spending and public schoolenrollment. The funds they do not spendon education aremore likely toend up in public infrastructure like parks and roads, which lack privatesubstitutes. They are also more likely to save a revenue shock ratherthan raising short-run consumption.2 Examining the implications of

1 State and federal governments are the main providers of other goods with privatesubstitutes.

2 A budget surplus can also increase access to credit and lower future taxes (Ball andMankiw, 1995). In Brazil, municipal taxes come primarily from two sources: Taxes on Ser-vices (ISS) and Real Estate Taxes (IPTU). The first is a head tax paid by self-employed indi-viduals and enterprise owners. The second is a property tax. Both tax rates vary bymunicipality.

these differential spending patterns for public service quality, I findthat unequal and high-income areas are actually more likely to use arevenue shock to raise the average quality of public education. Whilethey invest less in education, their more modest increases in publicschool enrollment more than offset this, leading to greater overallimprovements in school quality. Thus, there is no evidence that thislower propensity to invest in education harms the poor. I further pro-vide evidence that an exogenous increase in public sector revenuelowers enrollment in private primary schools.

The remainder of the paper is organized as follows. In Section 2, I putthe paper into context by describing contrasting predictions in thetheory literature. In Section 3, I describe education finance and thepolitical system in Brazil, including the 1998 FUNDEF law and subsequentrevisions to it. In Section 4, I outline the empirical strategy and data. InSection 5, I present the empirical results. Section 6 concludes.

2. Theoretical context

A large body of theory literature considers the impact of the distribu-tion of income on the demand for publicly-provided goods and incomeredistribution more generally. This literature is somewhat divided in itspredictions. One strand—described in the seminal work of Meltzer andRichard (1981, 1983), and expanded on by Alesina and Rodrik (1994)and Persson and Tabellini (1994)—proposes that under majority rule,an increase in mean income relative to the income of the medianvoter increases total public spending. An increase in inequality of thistype raises public spending by lowering the median (decisive) voter'stax price of raising revenue.

In contrast, a second and still burgeoning strand of the literaturefocuses on the existence of private sector substitutes for publicly-provided goods. Papers in this vein include Barzel (1973), Besley andCoate (1991), Epple and Romano (1996), Glomm and Ravikumar(1998), de la Croix and Doepke (2009), and Gutiérrez and Tanaka(2009). Their models use a variety of approaches but follow a similarlogic. If education quality is a normal good, the existence of private ed-ucation leads parents past a threshold income to enroll children in pri-vate school. This is the income at which the cost of private school isjust offset by the utility gain from higher-quality education. In societieswithmore unequal income orwith highermedian income, more peopleare past this threshold and use private schools. They thus oppose spend-ing on public education, leading to less public education under majorityvoting.3 This is in contrast to government provision of goods like infra-structure, which lack private substitutes.4 Related to this literature,Suárez Serrato and Wingender (2011) model a positive valuation ofgovernment services that is larger for unskilled workers. Interestingly,De la Croix and Doepke (2009) model inequality raising public educa-tion quality, since greater use of private education lowers the numberof children that must be educated publicly.

Existing empirical work has not definitively supported orrejected the Meltzer–Richard model. A number of cross-countrystudies suggest that inequality is associated with less public spend-ing (Lindert, 1994, 1996; Perotti, 1996; Benabou, 1996; Rodríguez,2004; Moene and Wallerstein, 2001; Schwabish, Smeeding, andOsberg, 2006). Several historical studies also find that unequal

constrained) and the richmay forma coalition to push for low taxes,meaning that theme-dian income voter is not pivotal.

4 While Stiglitz (1974) showed that preferences are not single-peakedwhenprivate ed-ucation is available, several approaches address this and obtain existenceof amajority vot-ing equilibrium: imposing a single crossing property in a median voter model (Epple andRomano, 1996; Gutiérrez and Tanaka, 2009), identifying the decisive voter (Barzel, 1973;Glomm and Ravikumar, 1998), and using a probabilistic voting model (de la Croix andDoepke, 2009).

322 K. Kosec / Journal of Development Economics 107 (2014) 320–342

communities tend to raise less revenue locally and provide lowerlevels of publicly-provided goods (Goldin and Katz, 1999;Ramcharan, 2010; Galor, Moav and Vollrath, 2009; Zolt, 2009). Incontrast, several studies of the U.S. find that inequality is associatedwith higher public spending and more progressive tax codes(Husted and Kenny, 1997; Chernick, 2005; Schwabish, 2008;Corcoran and Evans, 2010; Boustan et al., 2013). Identification prob-lems complicate the study of the effects of the distribution of incomeon public expenditures, especially given endogenous householdsorting (Tiebout, 1956).

Meltzer and Richard (1981) assume that the absolute tax burdenincreases with income while the benefits of government activity aremore equally shared. When a good is predominately consumed bypoorer individuals—as is education, in the Brazilian context—the bene-fits of this expenditure are less equally shared than in the case of abroadly-used good like public infrastructure. Thus, we might expectthe distribution of income to differentially affect public investment ineducation vs. in infrastructure. This motivates the empirical analysis ofthis paper.

A related political economy literature models the distribution ofincome affecting policy by affecting the distribution of political influ-ence. Theoretical models in this literature give the rich control overpolicy that is roughly proportionate to their share of income as op-posed to their share of votes (Benabou, 1996; Besley and Burgess,2002; Grossman and Helpman, 1994). They can offer larger cam-paign contributions and bribes, and may have lower costs of politicalactivity due to personal connections with politicians. However, therich also have less demand for redistribution. Accordingly, onemight expect inequality or higher income to reduce redistributionand the production of publicly-provided goods that mostly benefitthe poor.

Several empirical studies evaluate these models. Foster andRosenzweig (2003) find that in democratic villages of rural India,an increase in the share of landless increases road investment(which the landless want) and decreases irrigation investment(which they do not want) more than in non-democratic villages.5

Mansuri and Rao (2004) demonstrate, in a review of World Bankcommunity-driven development projects, that most projects aredominated by local elites and do not target the poor. Bardhan andMookherjee (2006) find that Indian villages with high land inequal-ity and large incidence of low-caste status and poverty receive lessmoney from the central government. Within villages, these factorsaffect the allocation of public goods and the propensity to select pro-jects that generate employment for the poor.6 And Karabarbounis(2011), using panel data from OECD countries, finds that when theincome of a group of citizens increases, aggregate redistributive policiestilt towards its most preferred policies. This literature provides an addi-tional political economy mechanism through which higher inequalityand incomemay reduce the size of government. Of course, this literaturedoes not examine the role the existence of private sector substitutes playsin public investment decisions.

In summary, the theoretical literature points to two separate channelsthrough which the distribution of income might impact revenue alloca-tion. Specifically, it may affect the number of people who support publiceducation (a “collective choice” channel), or it may affect the relativepower and influence of poor individuals, who support public education(a “political power” channel). The bulk of recent literature on both chan-nels would suggest that the propensity to allocate revenue to public

5 Several studies find similar results. Bates (1973, 1976, 1981) shows that some ethnicgroups in post-independence Africa increased their access to public funds by increasingtheir political salience. Varshney (1995), Jaffrelot (2003), Pande (2003), Chandra (2004),and Banerjee and Somanathan (2007) find similar results for low caste groups in India.

6 Keefer andKhemani (2005) describe howpoliticalmarket imperfections like informa-tion gaps, social fragmentation, identity-based voting, and credibility problems may drivesuch findings.

education is decreasing in both the level of income and income inequality.This paper aims to empirically test whether this is the case. One maysuspect that an increase in median incomemost strongly affects revenueallocation through the collective choice channel, by increasing the num-ber of rich people (who oppose public education expenditure). Converse-ly, one might suspect that an increase in income inequality (holdingmedian income constant) most strongly affects revenue allocationthrough the political power channel; the preferences of the medianvoter are preserved, but the rich are now (relatively) richer and thusmore able to influence policy. While the paper cannot empirically distin-guish which channel is most important in transmitting the effects of me-dian income on revenue allocation, nor which channel is most importantin transmitting the effects of inequality, each of their independent effectsmay come more from one channel than the other.

3. Background

3.1. Municipal governments and their finances

Brazil has over 5000municipal governments with remarkable po-litical autonomy. They are led by an elected mayor and an electedcity council. The mayor is the legislative, budgetary, and administra-tive leader, and the city council is largely responsible for approvingthe mayor's policies (Couto and Abrucio, 1995; Wampler, 2007).

Federal, state, and municipal governments all publicly providegoods. Municipalities primarily engage in providing pre-primary andprimary education, within-city transportation, and public infrastruc-ture. The municipal revenue base is comprised of state and federal gov-ernment transfers, municipal tax collection, and miscellaneous revenue(e.g., earnings from industrial, agricultural, andmineral activities).Mostmunicipalities dependheavily on transfers for revenue. In 1997, theme-dianmunicipality got 88% of revenue from transfers. Further, 80%ofmu-nicipalities got at least 75% of revenue from transfers. See Appendix Afor more details.

3.2. Education in Brazil

Education is one of the few publicly-provided goods with privatesubstitutes that municipal governments supply. They provide twolevels: pre-primary (ages 0–6) and primary (ages 7–15), with higherlevels handled by state and federal governments. Primary educationmust be available to all children and attendance is compulsory. Pre-primary is optional for governments and for parents; it takes place increches for children aged 0–3 and in pre-schools for children aged4–6. Municipalities must spend at least 25% of revenue on education,though this requirement is non-binding for the vast majority ofmunicipalities; as of 1997, nearly 2/3 of municipalities were spendingabove the minimum on education. Municipal governments providemost public primary education (83% of students in 2007), but states runsome primary schools.

Fig. 1 shows the share of children in private school in 2000 by quin-tile of per capita household income and by child age. For children aged0–3 (creche aged), 4–6 (pre-school aged) and 7–15 (primary schoolaged), enrollment rates in private school increase monotonically withincomequintile. At thefirst quintile of income, theprivate creche enroll-ment rate is b1%, the private pre-school enrollment rate is 4%, and theprivate primary school enrollment rate is 2%. However, at the fifthquintile, the private creche enrollment rate is 14%, the private pre-school enrollment rate is 43%, and the private primary school enrollmentrate is 33%.

Enrollment in pre-primary education is of particular interest forthree reasons. First, public and private versions are both widely avail-able. Not only are private creches and pre-schools abundant in Brazil,but also a family that cares for a young child at home uses a private ver-sion. Second, Brazilian municipalities have almost full discretion overhow much to invest in pre-primary. Third, pre-primary education may

Fig. 1. Enrollment rate in private schools in 2000 by age and income quintile. Sources:Author's calculations based on data from 2000 Brazilian Census, 11.7% sample (IBGE).


have very high returns; Currie and Thomas (1995), Currie (2001),Carneiro and Heckman (2003), Cunha et al. (2006), Engle et al. (2007,2011), Heckman and Masterov (2007), Berlinski et al. (2009), andChetty et al. (2011) describe substantial benefits from pre-primary,especially for poor children whose parents cannot privately provideadequate substitutes. Patrinos (2007) estimates a pre-school benefit-cost ratio of 2.0 for Brazil—larger than that of most public industrialand agricultural investments.

Pre-primary education in Brazil has expanded rapidly but unevenlyrecently. During 1995–2008, the national enrollment expanded from8% to 18% among 0–3 year olds (creche aged) and from 48% to 80%among 4–6 year olds (pre-school aged). Yet enrollment rates vary alot across states; in 2008, the creche enrollment rate ranged from 6%to 32%, and the pre-school enrollment rate ranged from 57% to 91%.7

3.3. The FUNDEF/B education finance reforms

Brazil has a history of within- and across-municipality incomeinequality.8 This has led to high variance in per capitamunicipal revenue.In 1998, the federal government passed an education finance reform topartially equalize funding across municipalities: “Fund for the Mainte-nance and Development of Fundamental Education and Valorization ofTeaching” (FUNDEF). Federal policymakers worried that disparities inbasic education investment threatened national progress. Rich statesopposed nation-wide redistribution, and a compromise of within-stateredistribution was reached.

FUNDEF obligated each of Brazil's 26 states to gather 15% of eachmunicipality's revenue, and 15% of state government revenue, in astate fund. Each municipality then received a share of the fund equalto its share of total public primary school students in the state. Enroll-ment data come from the Ministry of Education's annual Census ofSchools.9 The 15% payment to the state fund did count toward the 25%of revenue minimum expenditure on education, but 10% more was re-quired. All fund receipts had to go to education. If receipts did notreach a federal minimum per primary school child, the federal govern-ment would ‘top off’ the fund, bringing it to this level.10 In 2007, the

7 Averages are based on parental self reports from PNAD (1995–2008). For a discussionof heterogenous trends in access to pre-primary education inBrazil during 1996–2009, seeEvans and Kosec (2012).

8 In 2007, GDP per capita was 13,400 (constant 2005 R$). (The average 2005 exchangerate was 0.4 USD/Reais). It ranged from 4,300 R$ in Piauí to 37,700 R$ in the Federal Dis-trict (IBGE, 2007).

9 State and municipal departments of education must provide data to receive publicfunds.10 In 1998, Bahia, Ceará, Maranhão, Pará, Pernambuco, and Piauí had their funds toppedoff. The per child minimum changes annually. It began varying with age in 2000 and ur-banization status in 2005.

policy continued under “Fund for the Development of Basic EducationandAppreciation of the Teaching Profession” (FUNDEB),which gradual-ly increased the fraction of revenue paid to the fund and added otherlevels of education to the redistribution algorithm. See Appendix A formore detail.

Municipalities benefitting from the reform looked poor compared toothers in their state (even if they were rich in the national sense).Table 1 summarizes the frequency of high, moderate, and low netFUNDEF receipts per child in 1998 by high, moderate, and low averageincome per capita. Many rich municipalities gained while many poorones lost funds. In 1998, the average municipality got back 1.1 timeswhat it paid in. At the 25th percentile of net receipts, municipalitiesgot back 1/3 of the amount paid in. At the 75th percentile, they gotback 1.6 times the amount paid in. As municipalities pay in 15% of rev-enue, this implies a loss of 10% of revenue at the 25th percentile, and again of 9% of revenue at the 75th percentile. These are large effectsthat should lead municipalities to re-optimize spending.

4. Empirical strategy

The roles of inequality and median income in mediating how arevenue shock is spent can be described by the following equation:

log yitð Þ ¼ γlog ritð Þ þ δ log ritð Þ � gið Þ þ θ log ritð Þ � log dið Þð Þ þ ηXitþ αi þ βt þ uit ð1Þ

where i indexes municipalities and t indexes years. yit is a public expen-diture outcome such as education expenditure per capita, infrastructureexpenditure per capita, the per capita budget surplus, or measures ofschool enrollment or quality. gi is the Gini coefficient and di is medianper capita income. αi are municipality fixed effects and βt are yearfixed effects. Xit is a vector of control variables for municipality i inyear t, described in Section 4.4. I estimate the model using panel dataon Brazil's over 5000 municipalities during 1995–2008. The data comefrom several sources, matched at the municipality level. They aresummarized in Table 3 and described below.

By including an interaction of revenue with median income in theregression, I uncover that part of the effect of inequality that is not driv-en by its correlation with municipal income. In essence, δmeasures theeffect of a median-preserving spread of per capita income on how reve-nue is spent. A is amedian-preserving spread of B if and only if it has thesame median, and the tails of A are uniformly larger than those of B.Malamud and Trojani (2009) describe the relevance of median-preserving spreads inmacroeconomic inequality applications. Similarly,θ uncovers that part of the effect of median income that is not driven byits correlation with inequality.

Municipalities may be on different time trends according to theirvalues of the pre-reform (1997) variables used to compute the excludedinstruments. I address this concern by including the interaction of a lin-ear time trend with a vector of the 1997 values of per capita revenue,per capita transfer revenue, the public pre-primary school enrollmentrate, the public primary school enrollment rate, and the share of prima-ry school students in state or federally-run schools. This allows for het-erogeneous secular trends in municipalities with different pre-reformcharacteristics.

4.1. Identification

There are two identification problems likely to affect this publicinvestment analysis. The first is the potential for omitted variable bias.The second is the potential for observed per capita revenue to be endog-enous due to reverse causality.

Observed per capita revenue is the result of factors affecting both thesupply of public funds and the population's demand for funds. I want torely solely on variation in revenue that comes from the supply side. Fac-tors which affect demand for revenue may have a direct effect on

Table 2Means of variables describing quality of educational inputs, which suggest that private education is of higher quality than public.

Panel A: pre-primary schools Private Public Difference

(1) (2) (3)

Students per teacher 8.8 14.3 5.5(0.105) (0.132) (0.134)***

Fraction of teachers with some post-secondary education 0.394 0.430 0.036(0.001) (0.001) (0.001)***

Fraction of institutions with own school building 0.979 0.949 −0.030(0.001) (0.002) (0.002)***

Fraction of institutions with electricity 0.997 0.903 −0.095(0.001) (0.001) (0.002)***

Fraction of institutions with indoor bathroom 0.960 0.799 −0.161(0.001) (0.001) (0.003)***

Fraction of institutions with library 0.623 0.197 −0.427(0.003) (0.001) (0.003)***

Fraction of institutions with computer 0.849 0.370 −0.479(0.002) (0.002) (0.003)***

Panel B: primary schools Private Public Difference

(1) (2) (3)

Students per teacher 8.0 30.4 22.4(0.118) (0.445) (0.416)***

Fraction of teachers with some post-secondary education 0.801 0.800 −0.001(0.001) (0.001) (0.001)

Fraction of institutions with own school building 0.950 0.930 −0.020(0.001) (0.001) (0.002)***

Fraction of institutions with electricity 0.999 0.871 −0.129(0.001) (0.001) (0.002)***

Fraction of institutions with indoor bathroom 0.959 0.763 −0.197(0.001) (0.001) (0.003)***

Fraction of institutions with library 0.799 0.334 −0.465(0.003) (0.001) (0.003)***

Fraction of institutions with computer 0.920 0.447 −0.473(0.002) (0.001) (0.004)***

Notes: Standard errors are in parentheses. ***indicates p b 0.01.Sources: Author's calculations based on school- and teacher-level data from the 2008 Census of Schools.

Table 1Number of municipalities, by tercile of income Per capita and tercile of net per-child receipts (net per-child benefits) from FUNDEF education finance reform, 1998.

Lowest receipts Middle receipts Highest receipts

Poorest 51 372 710Middle 664 695 369Richest 638 286 273Total 1353 1353 1352

Notes: The total number of municipalities in the table is smaller than the national total, due to missing data on per capita municipal income or per-capita FUNDEF receipts data.Sources: Author's calculations based on data from Tesouro Nacional (2011) and IBGE (2007).


education and other policy choices. If they do, this would bias estimatesof the effects of revenue. There are several possible sources of omittedvariable bias (I detail three below). First, having an older populationmay mean more working adults per capita and more tax revenue. Butif it also means fewer children and less need for public education, thenordinary least-squares (OLS) estimates of the effects of revenue on edu-cation spending would be (likely downward) biased. The magnitude ofthe bias might shrink as measures of population-by-age were addedEq. (1), but it would be impossible to eliminate all bias. Second, mayorswith high discount rates or who are corrupt may generate higher reve-nue to maximize private gains.11 However, short-sightedness may alsolower investment in education, which has mostly long-term payoffs.

11 de Janvry et al. (2010) suggest such a possibility. Theyfind that second termmayors inBrazilian municipalities, who are not eligible for reelection, have less transparent policiesand are less likely to reduce school drop-out rates using federal funds designated for thispurpose. Ferraz and Finan (2009a) show that first termmayorsmisappropriate 27% fewerresources than second term mayors, which they also link to electoral incentives.

This, too, would likely downward-bias the estimates. Finally, municipal-ities with higher revenue may pay higher wages to public officials.Ferraz and Finan (2009b) show that, in Brazil, this leads to more politi-cal competition, and to the election of better-educated, more experi-enced, more productive candidates. If these qualities increase publiceducation investment, this would likely upward-bias estimates of theeffects of revenue on such investment.

Municipal revenue per capita can also be a response to educationpolicy—a classic problem of reverse causality. Municipalities that investheavily in public education may need to levy more taxes to pay for it.One response to these identification problems is a set of valid instru-ments. These instruments should affect revenue and its interactionswith inequality and median per capita income, but should be uncorre-lated with factors affecting the demand for revenue.

4.2. Instruments for per capita revenue

To address threats to identification, I exploit the federal FUNDEF/Beducation finance reforms described in Section 3.3, which generated

Table 3Summary statistics.

Variable Mean Std. dev.

Revenue per capita, 100 s 2005 Reais 9.84 5.92Simulated revenue per capita, 100 s, 2005 Reais 6.12 2.98Population, 100,000 s 0.32 1.92Population aged 0–6, 100,000 s 0.04 0.21Population aged 7–15, 100,000 s 0.06 0.3Gross municipal product (GMP) per capita, 10,000 s of 2005 Reais 0.68 0.58Gini coefficient (year 2000, computed by IBGE using full Census data) 0.39 0.03Median income per capita in 2000, 10,000 s of 2005 Reais 0.23 0.13Predicted Gini coefficient (computed by author w/Census 11.7% sample) 0.53 0.08Predicted median income per capita, 10,000 s of 2005 Reais 0.25 0.12Education spending per capita, 100 s 2005 Reais 2.67 1.42Infrastructure and urban development spending per capita, 100 s 2005 Reais 0.88 0.82Budget surplus per capita, 100 s 2005 Reais 0.53 1.01Municipal public pre-primary school students per age 0–6 population 0.25 0.14Municipal public primary school students per age 7–15 population 0.58 0.29Private pre-primary school students per age 0–6 population 0.03 0.05Private pre-primary school students per age 0–6 population in 1997 0.02 0.03Private primary school students per age 7–15 population 0.02 0.03Private primary school students per age 7–15 population in 1997 0.02 0.03Municipal public pre-primary school infrastructure quality index 0 1.73Municipal public primary school infrastructure quality index 0 1.81Private pre-primary school infrastructure quality index 0 1.78Private primary school infrastructure quality index 0 1.91Share of municipal public pre-primary schools w/ own school building 0.82 0.33Share of municipal public pre-primary schools w/ electricity 0.82 0.34Share of municipal public pre-primary schools w/ an indoor bathroom 0.76 0.36Share of municipal public pre-primary schools w/ a library 0.17 0.28Share of municipal public pre-primary schools w/ a computer 0.22 0.35Share of municipal public primary schools w/ own school building 0.81 0.34Share of municipal public primary schools w/ electricity 0.7 0.37Share of municipal public primary schools w/ an indoor bathroom 0.66 0.38Share of municipal public primary schools w/ a library 0.21 0.31Share of municipal public primary schools w/ a computer 0.23 0.35Share of private pre-primary schools w/ own school building 0.75 0.38Share of private pre-primary schools w/ electricity 0.89 0.31Share of private pre-primary schools w/ an indoor bathroom 0.85 0.33Share of private pre-primary schools w/ a library 0.49 0.4Share of private pre-primary schools w/ a computer 0.55 0.42Share of private primary schools w/ own school building 0.78 0.38Share of private primary schools w/ electricity 0.86 0.34Share of private primary schools w/ an indoor bathroom 0.83 0.36Share of private primary schools w/ a library 0.63 0.42Share of private primary schools w/ a computer 0.64 0.43Municipal public pre-primary students per teacher 19.43 7.96Municipal public primary students per teacher 26.17 12.87Private pre-primary students per teacher 13.66 7.18Private primary students per teacher 11.59 6.87Share of municipal public pre-primary teachers w/ post-secondary education 0.22 0.27Share of municipal public primary teachers w/ post-secondary education 0.31 0.31Share of private pre-primary teachers w/ post-secondary education 0.24 0.28Share of private primary teachers w/ post-secondary education 0.45 0.34Prova Brasil 4th grade Portuguese test score 171.92 18.93Prova Brasil 4th grade math test score 188.78 23.35Fraction of population that lived in urban areas in 2000 0.59 0.23Labor force participation rate (share of pop. economically active) in 2000 0.55 0.09Racial fractionalization index in 2000 0.38 0.12

Notes: Data are summarized over 1995–2008 over all municipalities for which data are available, except where a specific year is indicated. (N = 59,424).Sources: Author's calculations based on data from IBGE, Ministry of Education, Tesouro Nacional, IPEA, MEC, TSE, and BNDES.


exogenous variation in municipalities' revenue. These laws redistributedrevenue within states according to an algorithm handed down by thefederal government, and thus exogenous to any one municipality's in-vestment decisions. The rules changed slightly each year, but alwaysidentify two things: how much each municipality has to pay into itsstate's education fund (always a set fraction of its revenue), and howmuch of the fund is given to each municipality (equal to its share oftotal public school students in the state, with students at differentgrade levels weighted differently each year). Given municipal publicschool enrollment and revenue data for all municipalities, the algorithmdelivers a new (i.e. adjusted) revenue to each municipality. It is lower

for net losers and higher for net winners from the reform. Each year,there is a slightly different algorithm.

These reforms almost certainly induced endogenous revenue andenrollment responses by municipalities and parents (Gordon, 2004;Hoxby, 2001). A good instrument should encapsulate the credibly exog-enous variation in revenue generated by the law, but exclude variationdue to municipalities' own actions. To do this, for each year I simulatethe revenue each municipality would have if the current year's algo-rithm were applied to 1997 (pre-reform) enrollment and revenuedata. I then instrument for actual (endogenous) net revenue withthe simulated instrument, which allows me to test how different


municipalities differentially spend an exogenous shock to revenue,all else equal. I instrument for the interaction of per capita revenuewith covariates using the interaction of simulated per capita reve-nue with those covariates.12 See Appendix B for further details onconstruction of the instruments.

The first stage equations state that per capita revenue and its inter-actions with the Gini coefficient and per capita median income are afunction of simulated per capita revenue, sit and its interactions withinequality and per capita median income:

log ritð Þ ¼ λ1log sitð Þ þ π1 log sitð Þ � gið Þ þ ϕ1 log sitð Þ � log dið Þð Þþ μ1Xit þ ρ1;i þ κ1;t þ ϵ1;it

ð2Þ

log ritð Þ � gi ¼ λ2log sitð Þ þ π2 log sitð Þ � gið Þ þ ϕ2 log sitð Þ � log dið Þð Þþ μ2Xit þ ρ2;i þ κ2;t þ ϵ2;it

ð3Þ

log ritð Þ � log dið Þ ¼ λ3log sitð Þ þ π3 log sitð Þ � gið Þ þ ϕ3 log sitð Þ � log dið Þð Þþ μ3Xit þ ρ3;i þ κ3;t þ ϵ3;it :

ð4Þ

ρi aremunicipality fixed effects and κt are yearfixed effects. In Section 5,I demonstrate that these instruments satisfy the inclusion restriction:they are correlated with observed per capita revenue andwith its inter-actions with inequality and per capita median income. The exclusionrestriction should hold since the instruments depend only on pre-reform municipality characteristics (captured by fixed effects) andtime-varying federal government rules, which are exogenous to anyparticular municipality's investments.

4.3. Data and variable measurement

4.3.1. Inequality and median incomeI take two approaches to measure gi. First, I take the Gini coefficient

at one point in time and assume that it broadly characterizes the aver-age level of inequality in a municipality over 1995–2008. Specifically,gi is the Gini coefficient computed by the Brazilian Institute of Geogra-phy and Statistics (IBGE, 2010) using 2000 Census data, and it only ap-pears in interaction in Eq. (1). This seems appropriate as the nationalGini coefficient was 0.59 in 1960—the first year the Census collectedhousehold income data—and also 0.59 in 2000 (Skidmore, 2004). Itcan also be defended on practical grounds given that the only Censusduring the sample period was in 2000, and only Census data permit cal-culation of a municipality-specific Gini.13 Second, I predict a time-varying municipal Gini, as in Boustan et al. (2013). If national trendsmask changes in the distribution of income in individual municipalities,such an approach is important.14 Using Integrated Public UseMicrodataSeries (IPUMS) Census data, I find that the correlation coefficient be-tween the 1991 and 2000 municipal Gini coefficients is only 0.16—sug-gesting changes over time. Barros et al. (2009) also note that inequalitydeclined steadily since the 2000 Census, from a Gini of 0.59 in 2001 to0.55 in 2007 as per estimates from the National Household Sample Sur-vey (PNAD). I thus use both approaches (the 2000 Gini and a time-vary-ing, predicted Gini), showing that they yield substantially similarresults.

I predict each municipality's Gini coefficient for each yearduring 1995–2008 as follows. I first take actual per capita incomeobservations on the over 5 million households from the 2000 Cen-sus and compute the percentile of the national income distribution

12 In a similar vein, Martínez-Fritscher et al. (2010) use commodity price shocks in Brazilduring 1889–1930 as an instrument for state capacity to spend on publicly-providedgoods. They find that education expenditures increased during this period in responseto positive price shocks.13 No other survey indicates municipality of residence for individuals in all Brazilianmunicipalities.14 For example, if the incomes of the poor are rising more rapidly than those of the rich,then poor areas will see greater increases in the share of revenue from local taxes. AndGadenne (2012) shows that revenue from taxes (rather than transfers) is relatively morelikely to go into education.

(1st, 2nd, …, and 99th) to which each household belongs. Second, Itake household-level per capita income data from PNAD for eachyear during 1995–2008; while these data do not identify thehousehold's municipality of residence, they provide useful infor-mation on national patterns of income growth. I use them to com-pute the annual income growth rate for each percentile of theincome distribution (1st, 2nd, …, and 99th) during 1995–2008.Third, I apply these national income patterns to estimate whatwas each of the 5 million 2000 Census households' per capita in-come in each year during 1995–1999 and 2001–2008. For example, dur-ing 2001–2007, the incomes of the poorest decile grew by 7%while thoseof the richest decile grew by only 1.1% (Barros et al., 2009).15 Accord-ingly, I would predict that the incomes of poor households (as of the2000 Census) would growth more over 2001–2008 than would theincomes of richer households. Finally, armedwith predicted incomesof each household in each year, I compute a predicted Gini (andalso a predicted median per capita household income) for each mu-nicipality i in each year t. I similarly measure di (median income) intwo ways: the actual 2000 Census median per capita income, and atime-varying, predicted median per capita income.

By design, changes in the predicted Gini and predictedmedian in-come cannot be influenced by endogenous responses to revenuelevels (including household sorting); they capture expected changesin the municipal income distribution driven purely by nationaltrends, such as changes in the return to skill or other aspects of thelabor market. When using time-varying, predicted values, I estimatea modified variant of Eq. (1) that also includes the Gini and medianper capita income in their non-interacted forms:

log yitð Þ ¼ γplog ritð Þ þ δp log ritð Þ � gitð Þ þ θp log ritð Þ � log ditð Þð Þþ ψgit þωlog ditð Þ þ σXit þ υi þ χt þ εit :

ð5Þ

Results using the actual vs. predicted distribution of incomenaturally have different interpretations. When I use the actualincome distribution (and interpret δ and θ from Eq. (1)), I testwhether municipalities characterized by relatively high inequalityor high median income in 2000 spend exogenous shocks to revenueduring 1995–2008 differently. When I use the predicted incomedistribution (and interpret δp and θp from Eq. (5)), I test whethermu-nicipalities that one expects to become more unequal or to reach ahighermedian income during 1995–2008 (due to prevailing nationalincome patterns) spend an exogenous shock to revenue differentlythan do municipalities in which one does not anticipate such large,exogenously-driven increases.

4.3.2. Public expenditure and revenuesI measure education expenditure, infrastructure expenditure, and

the budget surplus in Reais per capita spent by municipality i in year t(in 100 s of constant, 2005 $R).16 I similarly measure revenue in netReais per capita (net of FUNDEF/B receipts) in municipality i in year t(in 100 s of constant, 2005 $R) using data from Tesouro Nacional(2011). As a nominal amount of currency has more purchasing powerin some areas—possibly in poorer or more rural areas—than in others,I convert all measures of expenditure, revenue, and income to constant,2005 São Paulo currency using a set of spatial incomedeflators for Brazilcomputed by Ferreira et al. (2003).

4.3.3. School enrollment levelsThe Ministry of Education's annual Census of Schools (Censo

Escolar) provides detailed school-level data on the number of students

15 Rather than computing national income growth rates by decile as in Barros et al.(2009), I compute them by percentile—thus computing 100 different rates rather than 10.16 These jointly account for about half of municipal revenue. Additional revenue goes toadministrative costs (e.g., paying public employees and operating the judiciary), public as-sistance (welfare), and a number of smaller uses which make it difficult to identify thebeneficiaries of the spending.


enrolled by grade level and type of school (public or private) for eachyear during 1995–2008. Combining these data with population by agedata, I also compute the share of 0–6 year olds and 7–15 year olds en-rolled in school (public and private).17

4.3.4. Public education qualityThe Census of Schools also provides detailed information on teachers

and school infrastructure in eachmunicipality i in each year t. I computeseveral measures of the average quality of educational inputs: studentsper teacher (i.e. class size), the fraction of teachers with at least somepost-secondary education,18 and the fraction of schools with a designat-ed school building,19 electricity, an indoor bathroom, a library, and acomputer. Table 2 summarizes these variables for public vs. privatepre-primary and primary schools in 2008.20

Private schools outperform public on all metrics but the educationlevel of pre-primary teachers (public teachers are 3.6 percentage pointsmore likely than private to have some post-secondary education).Public and private primary school teachers are equally educated. Onall other dimensions, private schools outperform public schools, andthe differences are statistically significant at the 0.01 level. Publicschools have 5.5 more pre-primary students and 22.4 more primarystudents per teacher. Indoor bathrooms are present in 80% of publicpre-primary schools and 76% of public primary schools, but in 96% ofeach of their private counterparts. A library is present in 20% of publicpre-primary schools and 33% of public primary schools, but in 62% ofprivate pre-primary schools and 80% of private primary schools. Givenhigh correlations among the infrastructure quality variables, I use prin-cipal components analysis to combine them into an index and use thefirst principal component. For public pre-primary (primary) schools,the first principal component explains 60% (67%) of the variation inthe infrastructure quality variables, with an eigenvalue of 3.01 (3.33).

The Ministry of Education also provides municipality-level averagemath andPortuguese test scores of 4th grade students inmunicipal publicprimary schools for 2005, 2007, and 2009. These come from the ProvaBrasil National Assessment of Educational Achievement. Assuming edu-cational investments (inputs) have a lagged impact on education outputs,I measure test scores at time t + 1 and thus merge these data onto reve-nues data from 2004, 2006, and 2008. These data are only available forpublic schools.

4.4. Demographic, political, and economic trends

One potential concern is possible correlation between simulatedrevenue and demographic, political, or economic trends that affectpublic investment. I thus include total population, age 0–6 population,and age 7–15 population (all in 100,000 s) in Xit—the vector of controlvariables for municipality i. I also include gross municipal product(GMP) per capita (in 10,000 s of 2005 R$), as this affects the tax base.

As a robustness check, I also show specifications controlling forthree time-varying political factors, measured using municipal elec-tions data from the Superior Electoral Court (TSE) for 1996, 2000,2004, and 2008. They include three features of the last mayoral elec-tion: the mayor's vote share, the number of political parties compet-ing, and an HHI of between-party competition.21 These mightinfluence the quality of policies or the level of corruption. As many

17 Population by age data from IBGE are available for 1991, 1996, 2000, and 2007. I com-pute annual estimates by linear interpolation.18 This includes any post-high school education, whether or not a degree was obtained.19 Some schools report using a teacher's house, church, community center, business,barn, or shed to carry out classes.20 Infrastructure data come from an average across all schools. Teacher education datacome from an average across all teachers. Students per teacher data come from an averageacross municipalities.21 I compute the standard Herfindahl–Hirschman Index typically used to measure com-petition among firms in an industry. A market is a municipality, a firm is a political party,and its market share is its share of votes during the last election.

municipalities are rural, I also take into account shifts in the worldprices of three important agricultural products: coffee, cocoa, andbananas. I multiply the current year world price of each (from the In-ternational Monetary Fund's International Financial Statistics) bytotal hectares of it in the municipality in 1994 (from the Institute ofApplied Economics Research). This allows the severity of a priceshock for a municipality to be determined by the intensity withwhich it grows the commodity.

4.5. Controlling for another government program: PMAT

In 1998—the same year FUNDEF was launched—the Brazilian Devel-opment Bank (BNDES) initiated a new program targeting municipalgovernments: the Tax Administration Modernization Program (PMAT).PMATwas available to allmunicipalities, providing applicantswith subsi-dized loans to invest inmodernizing their tax administration. About 6% ofBrazil's municipalities (331 in total, covering about 40% of the Brazilianpopulation) elected to take out a PMAT loan at some point during1998–2008. Exploiting exogenous variation in the time between applica-tion to the program and receipt of funds, Gadenne (2012) estimates thatan increase in local tax revenues (which PMAT delivers) boosts invest-ment in municipal public education more than does an equivalentincrease in transfer revenue.

Onemayworry that time-variation in the rules governing FUNDEF issomehow correlated with receipts from PMAT. As a robustness check, Iestimate specifications that take into account the effect of PMAT, toensure that such a correlation does not drivemy results. I try: a) control-ling for a dummy for participation in PMAT by municipality i in year t;b) controlling for the PMAT loan amount per capita in municipality iin year t;22 and c) omitting the roughly 6% of municipalities that had aPMAT program sometime during 1998–2008. None of these has anappreciable impact on my results. Data on participation in PMAT arethose collected by Gadenne (2012) through interviews with BNDES.

5. Results

5.1. First stage results

Table 4 presents estimates of the first-stage regressions. Actualmunicipal revenue is robustly positively correlated with simulatedrevenue. Column (1) indicates that a 10% increase in simulated revenueis associated with a 4.6% increase in revenue.23 The F-statistic on theexcluded instrument is 1566. In these and all specifications, I cluster stan-dard errors at the municipality level since revenue and public invest-ments vary at this level.

Column (2) explicitly allows municipalities to convert revenueshocks from FUNDEF/B into revenue differently, according to theiryear 2000 levels of inequality and median income. An increase in simu-lated revenue increases revenue more in higher-income municipalities(significant at the 0.01 level). However, there is little evidence that theeffects of simulated revenue vary with the 2000 Gini. Column (3) esti-mates an alternative specification, interacting simulated revenue withthe time-varying, predicted Gini and median income values (describedin Section 4.3.1). I find that municipalities predicted to become richerover the sample period becomemore likely to convert simulated revenueinto revenue. Further, municipalities predicted to become more unequalover the sample period become less likely to convert simulated revenueinto revenue. The F-statistic on the joint significance of the excluded

22 Since a PMAT program is six years long, I take the total amount loaned divided by six tobe the annual loan amount during each of the six years of the loan (in all other years, it is 0).23 Several factors can explainwhy this coefficient is not equal to 1.0. First, municipalitiescan reduce collections of tax revenue in response to an influx of money from FUNDEF/B.Second, the panel is rather long; changes in the political and economic landscape of a mu-nicipality over time may drive the coefficient up or down.

25 A budget surplus increases the supply of loans available to private borrowers, pushing

Table 4IV First stage results, showing the effect of simulated revenue from the law change on actual revenue.

Dependent variable: Log (municipal per capita revenue) Revenue × Gini Revenue × medianincome

Revenue × predictedGini

Revenue × predictedmedian income

(1) (2) (3) (4) (5) (6) (7)

Log (simulated revenue) 0.459*** 0.839*** 0.784*** 0.515*** −0.226*** −0.044 −0.251***(0.012) (0.087) (0.029) (0.179) (0.036) (0.044) (0.016)

Log (simulated revenue) × Gini 0.145 −0.457 1.513***(0.206) (0.422) (0.087)

Log (simulated revenue) × log (median income) 0.213*** 0.830*** 0.091***(0.013) (0.027) (0.005)

Log (simulated revenue) × predicted Gini −0.132*** 0.522*** 1.274***(0.019) (0.034) (0.013)

Log (simulated revenue) × log (predicted median inc) 0.140*** 0.791*** 0.093***(0.014) (0.022) (0.007)

Log (population) −0.654*** −0.635*** −0.647*** 1.064*** −0.245*** 0.978*** −0.333***(0.017) (0.017) (0.017) (0.034) (0.007) (0.031) (0.009)

Log (0–6 population) 0.038* 0.053*** 0.046** −0.155*** 0.031*** −0.137*** 0.028***(0.020) (0.020) (0.020) (0.040) (0.008) (0.035) (0.011)

Log (7–15 population) 0.281*** 0.260*** 0.278*** −0.341*** 0.096*** −0.342*** 0.145***(0.023) (0.023) (0.023) (0.047) (0.009) (0.040) (0.012)

Log (GMP per capita) 0.079*** 0.081*** 0.080*** −0.174*** 0.032*** −0.148*** 0.043***(0.006) (0.006) (0.006) (0.011) (0.002) (0.009) (0.003)

Observations 51,212 51,159 51,172 51,159 51,159 51,172 51,172R-squared 0.848 0.850 0.849 0.836 0.848 0.799 0.836Number of municipalities 3919 3915 3918 3915 3915 3918 3918F Stat, excluded instruments 1565.72 665.3 584.66 1321.01 735.83 1579.81 3690.32

Notes: An observation is a municipality - year. Robust standard errors are in parentheses, and clustered at the municipality level. All specifications include municipality and year fixedeffects, as well as a linear time trend interacted with the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enrollment rate),log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ***indicates p b .01; **indicates p b .05; *indicates p b .10. Populationis measured in 100,000 s. The mean Gini coefficient in 2000 is 0.39, the mean value of median income per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23,and mean revenue per capita (in 100 s 2005 Reais) is 9.84. The mean predicted Gini coefficient during 1995–2008 is 0.53 and the mean predicted median income per capita during1995–2008 is 0.25 (measured in 10,000 s of 2005 Reais).Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


instruments is over 580 in each first stage specification, leaving noconcerns of weak instruments.

5.2. Effects of a revenue shock on public investment

Table 5 presents OLS and IV estimates of the effect of revenue on var-ious measures of municipal investment, with and without populationand GMP controls. These include education and infrastructure spendingper capita, the per capita budget surplus, the enrollment rate in publicpre-primary schools (municipal public pre-primary students per popu-lation of 0–6 year olds), and the enrollment rate in public primaryschools (municipal public primary students per population of 7–15 yearolds). In every specification—unsurprisingly—revenue is positively corre-lated with the investment measure. Further, adding controls has littleimpact on the point estimates.

The IV estimates suggest that a 10% increase in revenue induced bythe FUNDEF/B reforms is associated with an 11% increase in educationspending per capita (Column 2) and a 5% increase in infrastructurespending per capita (Column 4). It is also associatedwith a 12% increasein the per capita budget surplus (Column 6). The increase in educationspending does impact public school enrollment rates, though its effectsare muchmore dramatic in the area of pre-primary education than pri-mary education. A 10% increase in revenue is associated with a 5% in-crease in the public pre-primary enrollment rate (significant at the0.01 level), but with a meager 0.09% increase in the public primary en-rollment rate (a result which is further statistically insignificant at con-ventional levels). This latter result suggests that municipal publicprimary education is not a level for which enrollment levels areconstrained by revenue. This may be in part due to the fact that publicprimary education must be provided to any child whose parentsdemand it, while municipalities have nearly complete discretion inproviding public pre-primary education.24

24 While age 4–6 education is now a legal requirement, it was not so during the sampleperiod.

The OLS and IV estimates are sufficiently different to suggest bias.For education spending and pre-primary enrollment, the OLS estimatesare smaller in magnitude than the IV estimates, consistent with thepreviously-described channels of downward bias. For example, ifrevenue is higher precisely where there are fewer children, then OLSmight downward-bias the effects of revenue on education spendingand pre-primary enrollment. In the case of infrastructure spendingand the budget surplus, the OLS estimates are larger in magnitudethan the IV estimates. This might be due to reverse causality, wherebymore public infrastructure and a larger budget surplus tend to supportthe business climate and thereby raise revenues.25

5.3. The distribution of income and government spending

The levels of inequality and median income in a municipalitymoderate the effects of a revenue shock on government spending,as shown in Table 6. Panel A estimates the coefficients on revenueand its interactions with inequality and median income fromEq. (1). Panel B interacts revenue with terciles of inequality andmedian income, thus allowing the effects of revenue to vary byinequality and income categories.

More unequal and higher-income municipalities are less likely tospend a revenue shock on education, as seen in Panel A. This is truewhether we use the 2000 levels of inequality and median income(strategy 1) or use time-varying, predicted values based on nationalpatterns of income growth (strategy 2)—seen in Columns (1) and(2), respectively. The interaction between revenue and the Gini isalways negative and significant at the 0.01 level, as is the interactionbetween revenue and median income. This suggests two things:a) more unequal and richer places (as measured in year 2000) are

down the interest rate (Ball and Mankiw, 1995). Further, DeLong and Summers (1991)suggest that the accumulation of capital stimulates technological change and increaseseconomy-wide productivity.

Table 5OLS and IV results, showing the effects of revenue on several measures of government investment.

Dependent Variable: Log (education spendingper capita)

Log (infrastructurespending per capita)

Log (budget surplus percapita)

Log (public pre-primaryschool students per age0-6 pop.)

Log (public primaryschool students per age7-15 pop.)

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Panel A: OLS results

Log (revenue) 0.695*** 0.655*** 1.112*** 1.132*** 3.105*** 3.462*** 0.065*** 0.083*** 0.117*** 0.198***(0.010) (0.011) (0.034) (0.038) (0.055) (0.059) (0.015) (0.015) (0.012) (0.013)

Log (population) −0.212*** 0.244*** 1.654*** 0.346*** 0.477***(0.020) (0.086) (0.113) (0.040) (0.031)

Log (0–6 population) 0.083*** −0.452*** −0.532*** −0.329*** 0.161***(0.023) (0.101) (0.132) (0.050) (0.047)

Log (7–15 population) 0.123*** 0.443*** −0.172 −0.203*** −0.386***(0.026) (0.114) (0.150) (0.054) (0.049)

Log (GMP per capita) 0.035*** 0.074*** −0.257*** 0.048*** 0.008(0.006) (0.027) (0.037) (0.013) (0.011)

Observations 50,714 50,714 46,257 46,257 33,788 33,788 49,456 49,456 48,877 48,877R-squared 0.784 0.786 0.157 0.158 0.258 0.269 0.563 0.566 0.531 0.538Number of municipalities 3919 3919 3918 3918 3918 3918 3919 3919 3919 3919

Panel B: IV results

Log (revenue) 1.067*** 1.131*** 0.396** 0.450*** 1.160*** 1.245*** 0.360*** 0.469*** 0.021 0.089(0.025) (0.033) (0.167) (0.174) (0.196) (0.239) (0.056) (0.068) (0.047) (0.058)

Log (population) 0.148*** −0.227 0.041 0.634*** 0.396***(0.032) (0.153) (0.201) (0.060) (0.049)

Log (0–6 population) 0.023 −0.383*** −0.262* −0.375*** 0.175***(0.025) (0.104) (0.140) (0.051) (0.047)

Log (7–15 population) 0.019 0.594*** 0.335** −0.290*** −0.362***(0.027) (0.125) (0.163) (0.057) (0.051)

Log (GMP per capita) −0.008 0.149*** −0.041 0.012 0.019(0.007) (0.033) (0.043) (0.015) (0.013)

Observations 50,713 50,713 46,256 46,256 33,777 33,777 49,455 49,455 48,867 48,867R-squared 0.760 0.753 0.137 0.143 0.201 0.206 0.555 0.556 0.530 0.537Number of municipalities 3918 3918 3917 3917 3907 3907 3918 3918 3909 3909

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. Public school refers tomunicipal public schools, run by the local government. All specifications includemunicipality and yearfixed effects, aswellas a linear time trend interacted with the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enrollment rate), log(public primaryenrollment rate), and log(share of primary school students in state or federally-run schools). Population is measured in 100,000 s and mean per capita income is measured in 10,000 s of2005 R$. The mean revenue per capita (in 100 s 2005 Reais) is 9.84.Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


less likely to invest an exogenous revenue shock in education, andb) places which we expect to become more unequal and richer dueto shifting national income patterns—such as changes in the returnto skill and in labor market institutions—become increasingly lesslikely to invest a revenue shock in education.

From Column (1), given a 10% increase in revenue, very equal andvery unequal municipalities invest the shock very differently in educa-tion. If we set income at its sample median and allow the 2000 Gini toincrease from the 10th percentile (0.35) to the 90th (0.44), then themunicipality will go from increasing education spending by 8.6% toincreasing education spending by only 8.0% (a 0.06/0.86 = 7% decreasein the elasticity).26 Similarly, very poor and very rich municipalitiesbehave differently. If we set the 2000 Gini at its sample median andallow 2000 median income to increase from the 10th percentile (895$R) to the 90th percentile (4144 $R), then the municipality will go fromincreasing education spending by 10.2% to increasing education spendingby only 7.0% (a 0.32/1.02 = 31% decrease in the elasticity).27 Column (2)shows similar results using the predicted distribution of income.

Higher-income municipalities are much more likely than low-income municipalities to spend a revenue shock on infrastructure, alsoshown in Panel A. This can be seen in Columns (3) and (4), which usestrategy 1 and strategy 2 for measuring the distribution of income,

26 8.6% = 0.774 + (log(0.218) × −0.206) + (0.35 × −0.646), and 8.0% = 0.774 +(log(0.218) × −0.206) + (0.44 × −0.646).27 10.2% = 0.774 + (log(0.089) × −0.206) + (0.39 × −0.646), and 7.0% = 0.774 +(log(0.414) × −0.206) + (0.39 × −0.646).

respectively. The interaction between revenue and median income isalways positive and significant at the 0.01 level. From Column (3),given a 10% increase in revenue, very equal and very unequal munici-palities behave similarly in their infrastructure spending. However,very poor and very rich municipalities behave differently. If we set the2000 Gini at its sample median and allow 2000 median income to in-crease from the 10th percentile to the 90th, then the municipality willgo from increasing infrastructure spending by only 8.3% to increasinginfrastructure spending by 12.3% (a 0.40/0.83 = 48% increase in theelasticity).

Finally, unequal and higher-income municipalities are both muchmore likely to save a revenue shock by running a budget surplus—seen in Panel A, Columns (5) and (6) (using strategies 1 and 2, respec-tively). The interaction between revenue and the Gini is always positive,and it is statistically significantwhen using strategy 1 (at the 0.05 level).The interaction between revenue andmedian income is always positiveand significant at the 0.01 level. From Column (5), given a 10% increasein revenue, very equal and very unequal municipalities behave different-ly. If we set income at its sample median and allow the 2000 Gini to in-crease from the 10th percentile to the 90th, then the municipality willgo from increasing its budget surplus by 23.6% to increasing it by 26.0%(a 0.24/2.36 = 10% increase in the elasticity). Similarly, very poor andvery rich municipalities behave differently. If we set the 2000 Gini at itssample median and allow 2000 median income to increase from the10th percentile to the 90th, then themunicipality will go from increasingthe budget surplus by 19.1% to increasing it by 28.7% (a 0.95/1.91 = 50%increase in the elasticity). From these findings, we see that while higher

Table 6IV Results, showing how inequality and median income moderate the effect of revenue on government spending.

Dependent variable: Log (education spendingper capita)

Log (infrastructure spendingper capita)

Log (budget surplusper capita)

(1) (2) (3) (4) (5) (6)

Panel A: interactions of revenue with inequality and median income

Log (revenue) 0.774*** 0.748*** 1.540*** 1.370*** 2.382*** 2.395***(0.081) (0.054) (0.379) (0.234) (0.548) (0.391)

Log (revenue) × Gini −0.646*** −0.196 2.647**(0.165) (0.744) (1.091)

Log (revenue) × log (median income) −0.206*** 0.262*** 0.622***(0.013) (0.055) (0.084)

Log (revenue) × predicted Gini −0.220*** −0.066 0.614(0.060) (0.253) (0.422)

Log (revenue) × log (predicted median income) −0.186*** 0.227*** 0.455***(0.015) (0.064) (0.103)

R-squared 0.782 0.778 0.158 0.158 0.255 0.243Kleibergen-Paap rk Wald F statistic 640.41 645.04 336.81 350.37 235.65 234.60

Panel B: interactions of revenue with terciles of inequality and median income

Distribution of income used: Actual(year 2000)

Predicted(time-varying)

Actual(year 2000)


Actual(year 2000)


Log (revenue) 0.880*** 0.973*** 1.058*** 0.811*** 2.338*** 1.924***(0.031) (0.031) (0.142) (0.148) (0.223) (0.236)

Log (revenue) × Gini bottom third 0.006 0.013 0.026 −0.026 −0.186** −0.104(0.011) (0.009) (0.055) (0.041) (0.085) (0.066)

Log (revenue) × Gini top third −0.027** −0.027*** 0.020 −0.007 0.002 −0.024(0.012) (0.008) (0.054) (0.032) (0.085) (0.055)

Log (revenue) × Median income bottom third 0.158*** 0.127*** −0.242*** −0.116** −0.365*** −0.374***(0.013) (0.012) (0.054) (0.047) (0.085) (0.081)

Log (revenue) × Median income top third −0.096*** −0.025** 0.148** 0.063 0.529*** 0.050(0.016) (0.011) (0.067) (0.042) (0.102) (0.072)

R-squared 0.779 0.771 0.159 0.156 0.252 0.233Observations 50,660 50,674 46,218 46,223 33,745 33,758Number of municipalities 3914 3917 3913 3915 3903 3906Kleibergen-Paap rk Wald F statistic 398.52 346.96 225.47 171.34 183.51 136.31

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ThemeanGini coefficient in 2000 is 0.39, themean value ofmedianincome per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84. The mean predicted Gini coefficient during1995–2008 is 0.53 and the mean predicted median income per capita during 1995–2008 is 0.25 (measured in 10,000 s of 2005 Reais).Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


median income and higher income inequality tend to reduce the amountof a revenue shock allocated to education, an important alternate use ofthat revenue is the generation of a budget surplus (which can helpincrease access to credit and lower future taxes).

In Panel B, I interact tercile dummies of inequality and medianincomewith revenue to see whether it is particularly high or particularlylow inequality and income that drive the results. A few results are inter-esting and worth noting. Column (1) shows that when a municipality inthemiddle third on both inequality andmedian income is given a 10% in-crease in revenue, it spends 8.8% more on education. If this municipalitywere instead in the bottom third on inequality, it would spend similarly.It is being in the top third of inequality that has an especially large, nega-tive effect on education spending. That is, the effect of inequality on edu-cation investment comesmainly fromvery unequal areas being less likelyto invest a shock in education. Column (3) shows that when amunicipal-ity in the middle third on both inequality and median income is given a10% increase in revenue, it spends 10.6% more on infrastructure. Bothbeing in the bottom and being in the top third on median income signif-icantly affect infrastructure spending, though the negative effect of beingin the bottom third is larger inmagnitude than the positive effect of beingin the top third. That is, the effect of median income on infrastructureinvestment comesmainly from very poor areas being less likely to investa shock in infrastructure. Finally, Column (5) shows that when a munici-pality in the middle third on both inequality andmedian income is given

a 10% increase in revenue, it increases its budget surplus by 23.4%. If thismunicipality were instead in the top third on inequality, it would spendsimilarly. However, being in the bottom third on inequality has a strong,negative impact on the propensity to save a revenue shock. That is, the ef-fect of inequality on the budget surplus seems to comemainly from veryequal areas being less likely to save.

The results strongly suggest that municipalities at different levelsof inequality and median income spend the same revenue shock dif-ferently. Unequal and higher-income municipalities are significantlyless likely to invest a shock in education. They are more likely to invest itin infrastructure or a budget surplus. That inequality leads to less invest-ment in education fails to support the predictions of Meltzer and Richard(1981). However, it is in keeping with a class of models predicting thatthe existence of private substitutes means inequality reduces publiceducation investment. That higher income leads to less investment ineducation supports the model of Suárez Serrato and Wingender (2011),underwhich poor and unskilledworkers value such government servicesmore than richer, skilledworkers. Interestingly, thefindings on educationdo not hold for public infrastructure investment, which is broadlyenjoyed by most citizens and does not target the poor. Unequal areasare no less likely to spend a revenue shock on infrastructure, and high-income areas are much more likely to spend it on infrastructure. Thisprovides evidence that the distribution of income affects not only thesize of government, but also the composition (or type) of government

Table 7IV Results, showing how inequality and median income moderate the effect of revenue on school enrollment.

Dependent Variable: Log (public pre-primaryschool students per age0–6 pop.)

Log (public primaryschool students per age7–15 pop.)

Log (private pre-primary studentsper age 0–6 pop.)

Log (private primarystudents per age7–15 pop.)

(1) (2) (3) (4) (5) (6) (7) (8)

Panel A: interactions with inequality and median income

Log (revenue) −0.028 −0.590*** −0.577*** −0.514*** 0.019 −0.487 −0.514*** 0.540(0.164) (0.110) (0.159) (0.107) (0.180) (0.430) (0.157) (0.470)

Log (revenue) × Gini −0.737** −0.841*** 1.228 −0.051(0.324) (0.270) (0.886) (0.999)

Log (revenue) × log (median income) −0.257*** −0.325*** 0.001 0.324***(0.026) (0.024) (0.072) (0.063)

Log (revenue) × predicted Gini −0.031 −0.284***(0.118) (0.085)

Log (revenue) × log (predicted income) −0.390*** −0.284***(0.032) (0.029)

R-squared 0.561 0.559 0.494 0.514 0.244 0.245 0.084 0.104F Stat, excluded instruments 612.54 610.72 579.18 577.77 581.37 280.24 580.71 298.23Observations 49,410 49,427 48,822 48,839 27,414 27,386 25,081 25,053

Panel B: interactions with terciles of inequality and median income

Actual(year 2000)


Actual(year 2000)


Actual(year 2000)

Actual(year 2000)

Log (revenue) 0.079 0.106* −0.361*** −0.182*** −0.142 −0.128(0.066) (0.063) (0.069) (0.064) (0.156) (0.135)

Log (revenue) × Gini bottom third 0.049** 0.015 0.021 0.026* −0.091 −0.144(0.025) (0.019) (0.017) (0.014) (0.082) (0.088)

Log (revenue) × Gini top third −0.009 −0.015 −0.021 −0.034*** 0.027 −0.091*(0.024) (0.014) (0.019) (0.012) (0.059) (0.054)

Log (revenue) × Median income bottom third 0.228*** 0.222*** 0.308*** 0.216*** 0.169*** −0.078(0.025) (0.022) (0.021) (0.018) (0.065) (0.061)

Log (revenue) × Median income top third −0.102*** −0.055*** −0.022 −0.000 0.146* 0.317***(0.032) (0.020) (0.035) (0.019) (0.074) (0.066)

R-squared 0.562 0.564 0.513 0.528 0.247 0.104Observations 49,410 49,427 48,822 48,839 27,386 25,053Number of municipalities 3914 3916 3905 3907 2768 2629Kleibergen-Paap rk Wald F statistic 376.44 321.85 358.45 307.74 175.31 183.57

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ThemeanGini coefficient in2000 is 0.39, themean value ofmedianincome per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84. The mean predicted Gini coefficient during1995–2008 is 0.53 and the mean predicted median income per capita during 1995–2008 is 0.25 (measured in 10,000 s of 2005 Reais).Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


spending. Unequal and high-income areas tend to invest more heavily ingoods that broadly benefit all citizens rather than those that target thepoor.

Ball and Mankiw (1995) provide insight into why unequal andhigher-income areas aremore likely to save a revenue shock as a budgetsurplus. Equal and low-incomemunicipalitiesmay have greater popularpressure for raising consumption in the short run given the decreasingmarginal utility of consumption. Individuals in such municipalitiesmay also have a higher discount rate, or expect to catch-up and becomericher in the future—thuswanting to consume today and pay the tax billlater.28 More unequal and high-income municipalities, on the otherhand, may wish to improve the business climate by running a budgetsurplus.

5.4. School enrollment rates

Municipalities increasing education spending may increase publicschool enrollment—either by attracting children not yet in school (more

28 Ball and Mankiw (1995) note: “Because of technological progress, the income andconsumption of a typical individual in the economy rises over time. Because budget defi-cits shift taxes forward in time, they benefit relatively poor current taxpayers at the ex-pense of relatively rich future taxpayers.”

likely among 0–6 year oldswho are not legally required to attend school)or by drawing children from private schools to public. Table 5 providedevidence of an expansion in enrollment due to a revenue shock that isparticularly large and statistically significant in the case of pre-primary(age 0–6) education. In this section, I consider how inequality andmedianincomemoderate the effects of a revenue shock on school enrollment. Inthe next section, I turn to the average quality of educational inputs andoutcomes following enrollment expansions.

Unequal and high-income municipalities are less likely than theirmore equal and low-income counterparts to raise public school enroll-ment following a revenue shock, as shown in Table 7. This is true forboth public pre-primary education (Columns 1 and 2) and public primaryeducation (Columns 3 and 4), and results using the 2000 levels ofinequality and median income are similar to those using time-varying,predicted values. From Panel A, the interaction between revenue andthe Gini is negative and significant at the 0.05 level in the case of publicpre-primary education and at the 0.01 level in the case of primary ed-ucation. The interaction between revenue and median income isnegative and significant at the 0.01 level in all specifications. Whilea revenue shock tends to expand public school enrollment(Table 5), the impact of the shock on enrollment varies significantlyaccording to the levels of median income and inequality when weflexibly allow its impact to do so (Table 7).

Table 8IV Results, showing how inequality and median income moderate the effect of revenue on the quality of education inputs and outputs.

Dependent Variable: Infrastructure quality index(5 components)

Share of teachers w/somepost-secondary education

Log (students per teacher) Log (Prova Brasil 4thgrade test score)

(1) (2) (3) (4) (5) (6) (7) (8)

Panel A: public schools

Pre-primary Primary Pre-primary Primary Pre-primary Primary Portuguese Math

Log (revenue) −0.424* −1.132*** −0.108 −0.372*** −0.709*** −1.361*** 0.081 −0.265(0.249) (0.376) (0.086) (0.096) (0.182) (0.180) (0.151) (0.187)

Log (revenue) × Gini 1.896*** 2.248*** 0.816*** 1.435*** 0.343 1.519*** 0.109 0.435**(0.508) (0.739) (0.174) (0.207) (0.378) (0.370) (0.139) (0.170)

Log (revenue) × log (median income) 0.313*** 0.352*** 0.120*** 0.129*** −0.542*** −0.618*** 0.033*** 0.066***(0.038) (0.062) (0.013) (0.013) (0.027) (0.029) (0.009) (0.011)

R-squared 0.899 0.876 0.379 0.651 0.315 0.249 0.285 0.456Observations 49,393 41,866 45,485 44,876 45,401 44,778 11,152 11,152Number of municipalities 3914 3904 3913 3904 3913 3904 3981 3981

Panel B: private schools

Pre-primary Primary Pre-primary Primary Pre-primary Primary

Log (revenue) 0.078 −0.033 0.127 −0.257* −0.742*** 0.334(0.362) (0.388) (0.130) (0.149) (0.259) (0.297)

Log (revenue) × Gini 0.193 0.386 0.227 0.905*** 0.643 −1.871***(0.747) (0.817) (0.255) (0.303) (0.527) (0.627)

Log (revenue) × log (median income) −0.070 0.068 0.087*** 0.033 −0.462*** −0.315***(0.055) (0.055) (0.020) (0.022) (0.043) (0.043)

R-squared 0.894 0.936 0.212 0.405 0.264 0.551Observations 26,605 20,408 25,124 22,837 25,017 22,604Number of municipalities 2538 2214 2659 2472 2650 2464

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). Pre-primary education is age 0–6 education, and primary educationis age 7–15 education.Measures of the quality of educational inputs (Columns 1–6) aremeasured at time t. Test scores (Columns 7–8) aremeasured at time t + 1. Themean Gini coefficient in2000 is 0.39, the mean value of median income per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84.Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.

29 Recall that the index is the first principal component of a principal components anal-ysis using five variables: the share of municipal schools with their own building, electric-ity, indoor toilet, library, and computer. Table C.1 presents results for the five individualinfrastructure quality variables.


To interpret the magnitudes of the effects, it is again helpful tocompare a municipality in the middle third on both inequality andincome with others, as shown in Panel B. Column (1) shows that ifone doubled its revenue, this municipality would increase its publicpre-primary enrollment rate by 7.9%. If this municipality were insteadin the top third on inequality, it would behave similarly. However, if itwere instead in the bottom third on inequality, it would increase pre-primary enrollment by a substantially larger 12.8% (7.9 + 4.9)—a dif-ference statistically significant at the 0.05 level. Similarly, if this munic-ipality in the middle third of both inequality and income (whichexpands pre-primary enrollment by 7.9% following a doubling of reve-nues) were instead in the bottom third on income, it would increaseits pre-primary enrollment rate by a substantially larger 30.8%(7.9 + 22.8) in response to the shock.Were it in the top third on income,it would actual reduce its public pre-primary enrollment rate by 2.1%(7.9–10.0) in response to the shock. Column (2) shows similar resultsusing the predicted distribution of income. These differences by incomeare statistically significant at the 0.01 level. Results for primary education(Columns 3 and 4) similarly suggest that the distribution of incomeimportantlymoderates the effect of a revenue shock on public school en-rollment, with municipalities with a Gini coefficient and median incomein the bottom third being most likely to expand primary enrollment.

One might expect changes in public education investments to nega-tively impact enrollment in private schools. From Panel A, Columns (5)and (7), we see that a revenue increase has no significant impact on en-rollment in private pre-primary. However, a 10% increase in revenue isassociated with a 5.1% decrease in the private primary school enrollmentrate. Column (8), Panels A and B show that this reduction in privateschooling in response to more public sector revenue is much larger inmagnitude in municipalities with lower median income. The effects donot vary consistently with the level of inequality.

5.5. Quality of education inputs and outputs

Municipalities increasing education spending may increase ordecrease the quality of educational inputs and outputs, depending onwhether enrollment expansions outpace the additional investments.We have already seen that higher inequality and median income tendto raise the amount of a revenue shock invested in education, as wellas raise the public school enrollment rate; a natural question is whatthese effects imply for educational quality. Conceivably, they mayraise or lower the quality of factors including infrastructure, teachers,and learning. In this section, I consider how inequality and medianincomemoderate the effects of a revenue shock on educational quality.I measure educational quality using both educational inputs and ameasure of educational outputs: standardized test scores. I considerboth public and private school quality, at the pre-primary and primarylevels. Public school quality should be directly affected by the invest-ment decisions of local governments. Private school quality may beindirectly affected through changes in demand for private schools.

Despite their lower propensity to invest a revenue shock in educationor expand public school enrollment, more unequal and higher-incomemunicipalities aremore likely to raise the average quality of public educa-tion following a revenue shock, as shown in Table 8. Panel A presents re-sults for public schools on three indicators of educational input qualityand two indicators of output quality. Panel B presents the same resultsfor private schools. Indicators of input quality include the school infra-structure quality index described in Section 4.3.4,29 the share of teachers

Table 9IV Results, showing how the pre-reform supply of private education moderates the effects of a revenue shock.

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

Dependent variable: log (education spending per capita)

Log (revenue) 0.906*** 0.833*** 0.854*** 0.891***(0.122) (0.125) (0.123) (0.122)

Log (revenue) × Gini −1.022*** −0.743*** −0.726*** −0.696***(0.237) (0.257) (0.259) (0.255)

Log (revenue) × log (median income) −0.223*** −0.215*** −0.212*** −0.213***(0.020) (0.020) (0.020) (0.020)

Log (revenue) × log (No. private primary schools in 1997) −0.021***(0.008)

Log (revenue) × log (No. private primary classrooms in 1997) −0.016***(0.006)

Log (revenue) × log (No. private primary students in 1997) −0.020***(0.006)

Log(population) −0.068* −0.066* −0.065* −0.063(0.040) (0.040) (0.040) (0.040)

Log(0–6 population) −0.023 −0.008 −0.011 −0.007(0.044) (0.044) (0.044) (0.044)

Log(7–15 population) 0.127*** 0.122*** 0.123*** 0.123***(0.047) (0.047) (0.047) (0.047)

Log(GMP per capita) −0.001 −0.002 −0.002 −0.003(0.010) (0.010) (0.010) (0.010)

Observations 22,453 22,453 22,453 22,453R-squared 0.803 0.804 0.804 0.804Number of municipalities 1732 1732 1732 1732Kleibergen-Paap rk Wald F statistic 344.00 257.16 256.55 256.61

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ThemeanGini coefficient in2000 is 0.39, themean value ofmedianincome per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84.Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


with at least some post-secondary education, and students per teacher,all measured in period t. Indicators of the quality of educational outputsare the municipality's public school average 4th grade Portuguese andmath test scores in year t + 1 (not available for private schools).

For public schools, the interaction between revenue and the Gini ispositive and statistically significant at the 0.01 level in all specificationsexcept pre-primary students per teacher and 4th grade Portuguese testscores, where it is insignificant at conventional levels. More unequal mu-nicipalities are more likely to improve pre-primary and primary schoolinfrastructure quality and also more likely to employ pre-primary andprimary teachers with at least some post-secondary education. They arealsomore likely to raise 4th grademath test scores (though nomore like-ly to raise 4th grade Portuguese test scores). One caveat is that more un-equal municipalities are more likely to raise primary school class sizes(though not pre-primary class sizes). Overall, however, inequalitymakes a municipality more likely to use a revenue shock to increase theaverage quality of educational inputs, and this further translates into agreater likelihood of raisingmath test scores.30 Thus, we find no evidencethat a lower propensity of unequal areas to invest additional revenuein publicly-provided goods like education compromises the qualityof such goods; if anything, unequal areas are more likely to use arevenue shock to raise the average quality of public education.

The interaction between revenue and median income is positiveand statistically significant at the 0.01 level in all specificationsexcept students per teacher, where it is negative and statisticallysignificant at the 0.01 level. Higher-incomemunicipalities are significant-ly more likely to improve education along all of the input and outputmeasures of quality. At both the pre-primary and primary schoollevels, they are more likely to improve school infrastructurequality, more likely to employ teachers with at least some post-secondary education, more likely to reduce class sizes, and more

30 Math test scoresmaybe inherently easier to raise through changes in education inputsthan language test scores; this is consistent, for example, with Muralidharan andSundararaman (2011).

likely to raise 4th grade Portuguese and math test scores. Resultsinstead using a predicted Gini and median income are similar, andavailable upon request.

Panel B of Table 8 presents evidence on the impact of an exoge-nous public sector revenue shock on the average quality of privateschool inputs. Recall from Table 7 that a 10% increase in municipalrevenue does not affect enrollment in public pre-primary school,but leads to a 5.1% decrease in private primary school enrollment.Now, we see that the revenue increase has no impact on privateschool infrastructure (either at the pre- primary or primary level).Fixed private school assets remain unchanged—perhaps becausethey are fixed. However, we do observe changes in the supply andeducation levels of private school teachers—mostly in more un-equal and high-income municipalities. In more unequal municipal-ities, a revenue increase is far more likely to increase the share ofprivate school teachers with some post-secondary education and to re-duce the number of private primary school students per teacher (thoughsimilar results are not found for private pre-primary). In higher-incomemunicipalities, a revenue increase is far more likely to increase theshare of private pre-primary teachers with some post-secondary educa-tion, and to reduce the number of private school students per teacher inboth pre-primary and primary schools. Given relatively greater improve-ments in public school quality in such municipalities, this is consistentwith private schools being pressured to increase their quality along flex-ible dimensions like teacher inputs.

5.6. Mechanisms analysis: initial supply of private schools

Thus far, we've seen evidence that municipalities with higherincome inequality are less likely to allocate a revenue shock to educa-tion. This is consistent with a class of theoretical models predictingthat the existence of private substitutes for education makes unequalmunicipalities relatively less likely to invest in public education. Asinequality increases, the share of the population sending children to

32 Since municipalities may have changed their revenue levels and school enrollment in


public schools declines, leading to less demand for public educationunder majority voting. However, a natural question is whether thislower propensity to invest in education is in part due mechanically tothe large initial supply of private schools in unequal municipalities. In1997—a year before the FUNDEF/B reform—theGini coefficientwas pos-itively correlated with a number of indicators of the supply of privateprimary education: the number of schools (correlation coefficient0.17), the number of classrooms (0.14), and the number of students(0.16). While the supply of private schools may contract in responseto greater public education investment, it may be slow to do so in theshort-run. Parents may hesitate to pull children out of their currentprivate schools, and private schools may thus be slow to reduce theircapacity or close.

I explore the extent to which the supply of private schooling drivesthe results on inequality by including an interaction between revenueand the 1997 (i.e. initial) supply of private education in Eq. (1). Doingso provides insight into roughly what share of the impact of the Giniwas due to its correlation with the private schooling supply. Table 9presents these results.31

Column (1) is the baseline specification (Eq. (1)). Columns (2)–(4)each add to the baseline specification a control for logged revenueinteracted with a logged 1997 measure of the supply of private primaryschooling. Column (2) uses the number of schools, Column (3) uses thenumber of classrooms, and Column (4) uses the number of students. Inall three specifications, the interaction between revenues and the 1997measure of private education supply is negative and statistically signifi-cant at the 0.01 level. In municipalities with a larger initial supply ofprivate schooling, a revenue shock is less likely to be spent on education.However, what is particularly interesting is how the coefficient onrevenue interacted with the Gini (δ) changes once these new controlsare added. Compared with Column (1), δ is 27% smaller in Column (2),29% smaller in Column (3), and 32% smaller in Column (4). Thisprovides suggestive evidence that an increase in revenue available tolocal governments interacts with the availability of private schooling toaffect government choices regarding education spending. Some 30% ofthe effect of inequality on education spending may be due to the greaterinitial supply of private schools inmore unequalmunicipalities. However,the remaining 70% of the effect may be due to other factors—such as alower demand for public education and less political influence of the poor.

5.7. Robustness

5.7.1. Controlling for PMAT and political and economic trendsIf time variation in simulated revenue is somehow correlated with

other government programs or with political and economic trends thatimpact public investment, the estimates would be biased due to omittedvariables. (Sections 4.4 and 4.5 describe these potential concerns in de-tail). Table C.2 accordingly examines the sensitivity of the results to anumber of alternative specifications. I first take into account the effectsof PMAT (a federal program initiated in 1998 which provides subsidizedloans to municipalities to improve their tax administration) in severalways. Next, I try controlling for a vector of time-varying political and eco-nomic controls: the mayor's vote share, the number of political partiescompeting for mayor, an HHI of between-party competition for mayor,and the current year world prices of coffee, cocoa, and bananas—eachweighted by the 1994 hectares of each crop in the municipality. Column(1) presents the baseline results for the effects of revenue and its interac-tionswith inequality andmedian income on all of themain public expen-diture and enrollment variables (education spending, infrastructurespending, budget surplus, pre-primary enrollment, and primary enroll-ment). Column (2) omits the 6% of municipalities that had a PMAT pro-gram sometime during 1998–2008; Column (3) adds a dummy forparticipation in PMAT by municipality i in year t; Column (4) controls

31 The sample size is now somewhat smaller, reflecting the fact thatmanymunicipalitieshave no private education at any point during the sample period.

for the PMAT loan amount per capita in municipality i in year t; and Col-umn (5) controls for both the PMAT loan amount per capita and a vectorof time-varying political and economic controls (described in Section 4.4).

As Table 2 shows, the addition of controls for PMAT and for politicaland economic shocks has virtually no effect on the coefficientsinterpreted in this analysis, or their standard errors. That is, the effectsof the revenue shocks seem to be due to exogenous changes in therules governing the FUNDEF/B program and not somehow due to theircorrelations with these time-varying variables.

5.7.2. Heterogeneous expenditure constraintsStarting in 1998, FUNDEF provided municipalities with revenue

shocks that had to be spent on education; this was true for all munic-ipalities—unequal and equal, rich and poor. However, municipalitiesalready accustomed to spending beyond the minimum on education(nearly 2/3 as of 1997) could simply reduce their voluntary (abovethe minimum, 25% of revenue) education spending in response to awindfall. Doing so effectively allowed them to use some of the wind-fall on any spending priority (not just education). For municipalities:a) with a higher share of revenue going to education in 1997, or b)with asmaller expectedwindfall from FUNDEF, these new fundswere especiallyfungible.

An example is helpful. Suppose amunicipality receives a 5% increasein net revenue in 1998. This money must go into education. If themunicipality spent at least 30% of revenue on education in 1997 (5%more than it was required to!), then it could simply reduce this voluntary(above minimal) education spending from 30% in 1997 to 25% in 1998.With the 5% revenue shock plus this 25%, themunicipality would contin-ue to spend what it did on education in 1997, and yet would also be ableto put 5% more revenue into any other spending priority. If the samemunicipality received any smaller revenue shock, it would also be ableto invest the full shock into any spending priority of its choosing. If thismunicipality received a shock larger than 5%, however, it would need tospend some part of the shock on education. For example, with a 10%revenue shock, it would have to spend half of the shock on education inorder to continue to spend 30% on education (though it would have anadditional 5% of revenue to put into another spending priority). However,a municipality with 1997 education spending of 35% or higher would beable to spend the full 10% revenue shock on any spending priority by sim-ply allowing the new revenue to crowd out all voluntary educationspending.

This implies that FUNDEF/B receipts are more fungible (lessconstrained in how they are to be used) for two types of municipalities:a) those spending far beyond the minimum on education in 1997, andb) those receiving small amounts from FUNDEF. Education investmentis effectively more expensive for a less-constrained municipality, asthere are more potential uses of the money. That is, there is a price ef-fect. While less-constrained municipalities include equal and unequal,high and low income municipalities, a potential concern is that theintensity of the price effect is correlated with inequality or medianincome.

To show that such a correlation between the distribution of incomeand the intensity of the price effect is not driving the results, I estimateEq. (1) for sub-samples of municipalities that we expect to be amongthe least constrained to spend a revenue shock on education: a) thosein the top third of 1997 education spending as a share of revenue, andb) those in the bottom third of net FUNDEF receipts as a share ofrevenue.32 In these municipalities, the price effects should be muted,allowing us tomore clearly observe the effects of inequality andmedianincome on spending.

response to the reform, here I simulatewhat 1997 FUNDEF receiptswould be (as a share of1997 revenue) if FUNDEF were already in place in 1997 (using the 1998 rules forredistribution).


Table C.3 presents these results. The dependent variable is loggededucation spending per capita, and I first use actual 2000 inequalityand median income, and then use the time-varying, predicted mea-sures of inequality and median income. I find that both inequalityand median income still have a strong and statistically significantnegative impact on the propensity to invest revenue in education.If anything, the interaction between revenue and inequality is nowlarger in magnitude—whether we use the 2000 or the predictedmeasures of the distribution of income. The effects of median incomeinteracted with revenue are quite similar to the baseline estimates. Acorrelation between the distribution of income and how constraineda municipality is to spend revenue on education does not appear tobe driving the results.

5.7.3. Correlates of inequality and incomeI treat inequality andmedian income as broad concepts summarizing

important dimensions of society. However, my findings may be less in-teresting if they are simply pickingup the effect of somevariable correlat-ed with them that also conditions how revenue is spent. For example, itmay be more expensive and complicated to provide public education inmore urbanized places or in thosewith extensive racial and social hetero-geneity. To explore this possibility, I individually add three interactionswith revenue to the model: the year 2000 fraction of the municipalitythat was urban, the labor force participation rate, and an index of racialfractionalization.33 If the effects of inequality andmedian income inmod-erating a revenue shock were to vanish when including these controls,then these effects would seem rather mechanical rather than due to themechanisms hypothesized in the theoretical models presented.

As Appendix Table C.4 shows, inclusion of these interactions withrevenue does not significantly affect themagnitude of themain findingsor their statistical significance. More heavily urbanized municipalitiesare less likely to invest a revenue shock in education or to expand publicprimary enrollment. Municipalities with greater labor force participa-tion are more likely to expand public pre-primary and primary enroll-ment rates. Finally, municipalities with greater racial fractionalizationare more likely to expand public pre-primary enrollment. That the ef-fects of inequality and median income are similar despite the inclusionof these controls is consistent with the political economy mechanismsproposed. That is, inequality and median income seem to be capturingdifferences in the policy preferences of citizens, or in the influencewielded by relatively poor citizens (whomost support public education).

6. Conclusions

A long-standing debate in the theory literature concerns how the dis-tribution of income affects the size of government and its propensity toinvest in publicly-provided goods. The models of Barzel (1973), Besleyand Coate (1991), Epple and Romano (1996), Glomm and Ravikumar(1998), and de la Croix and Doepke (2009) suggest that greater inequal-ity leads to less redistribution, while Meltzer and Richard (1981, 1983),Alesina and Rodrik (1994) and Persson and Tabellini (1994) predictthat it leads to more redistribution.

This paper partly reconciles these models by focusing on the type ofpublic sector investment. I use credibly exogenous shocks to Brazilianmu-nicipalities' revenue during 1995–2008 generated by non-linearities infederal transfer laws to examine the effects of the distribution of incomeon goods with vs. without private sector substitutes. This reveals an inter-esting distinction between the two types of goods. Municipalities withhigher income inequality or higher median income allocate less of a reve-nue shock to education (a publicly-provided good that predominantlybenefits the poor) and are less likely to expand public school enrollment.

33 Urbanization is as classified by IBGE for the 2000 Census. Labor force participation isthe fraction of the adult population that is economically active. Racial fractionalization isan HHI computed using data on the population in each of four categories: black or brown,white, Asian, and indigenous.

However, they are more likely to invest in public infrastructure, likeparks and roads, which is enjoyed by everyone. They are also more likelyto save a revenue shock, running a budget surplus. Yet I find no evidencethat the quality of public education suffers due to this lower propensityto invest. If anything, unequal and high-income areas are more likely touse a revenue shock to improve public school inputs and test scores.

These findings have at least three policy implications. First, revenuetransfers that target the poorest local governments and those with thelowest income inequality are likely to be most effective in expandingpublic education investment and public school enrollment. Policymakersin those municipalities already have a relatively strong desire to spendtheir next R$ on these uses, and national policymakers should take thisinto account. In the Brazilian case, a nation-wide education finance equal-ization reformwould boost education spending in uniformly poormunic-ipalities better than a within-state distribution like FUNDEF/B, as there issubstantial cross-state revenue inequality. Second, the existence of avibrant private education sector is important. By absorbing many richerstudents, it can ensure that limited education resources are not spreadtoo thinly, and quality remains high—especially in highly unequal andhigh-income areas that are relatively less likely to invest a revenueshock in education. Finally, the steady decline in inequality taking placein Brazil has real implications for public investments—as do changes ininequality in other parts of the world. In particular, one can expect localgovernments to become increasingly more likely to invest a revenueshock from FUNDEF/B into education and to expand public schoolenrollment. One could imagine that this would make it increasinglyless necessary to mandate minimum expenditures on education.

Acknowledgments

This research was supported by a National Science FoundationGraduate Research Fellowship and a Stanford Institute for EconomicPolicy Research Dissertation Fellowship. I am incredibly grateful to mydissertation committee for excellent advising: Caroline Hoxby, SeemaJayachandran, Saumitra Jha, and Romain Wacziarg. I thank MichaelBoskin, David Evans, Fernando Ferreira, Maria Fitzpatrick, Claudio Ferraz,Alex Frankel, David Laitin, Darío Maldonado, Karthik Muralidharan,Kathryn Shaw, Ken Shotts, Fabián Valencia, Emiliana Vegas, ScottWallsten, Joel Wiles, and two anonymous referees for helpful comments.I also thank Érica Amorim, Leonardo Bursztyn, Flavio Cunha, FernandaEstevan, Pedro Maciel, Guilherme Lichand, David Plank, Joseph Sands,Pedro Silva, and Brian Wampler for data and useful discussions aboutlocal government policy in Brazil. Remaining errors are my own.

Appendix A. Institutional detail: public education in Brazil, FUNDEF,and FUNDEB

In 1985, Brazil began a process of re-democratization following mili-tary dictatorship. The constitution of 1988 decentralized substantialpower to state and municipal governments. As of 2000, Brazil had 5507municipalities. Municipal governments are led by an elected mayor andan elected city council comprised of a minimum of nine and a maximumof 55 councilmen, according to population size. Local elections are com-petitive and held every four years, and voting is compulsory for thoseover 17. Everyone 16 and older can vote. The mayoral candidate withthe largest vote share is elected (with runoff elections in municipalitieswith populations over 200,000 if no candidate wins at least 50% of thevote). City council seats are filled by a system of open-list proportionalrepresentation. Legislative, budgetary, and administrative authority areconcentrated in the mayor's office, and the city council is largely respon-sible for approving mayoral directives (Couto and Abrucio, 1995;Wampler, 2007).

The federal government has placed much emphasis on pre-primaryeducation over the last two decades. The 1988 constitution formally rec-ognized age 0–6 education as the right of every child and the responsibil-ity of municipal governments. The 1996 National Education Guidelines


and Frameworks Law emphasized the educational value of pre-primaryeducation by integrating crèches (age 0–3) and preschools (age 0–6,and 0–5 starting in 2007) as the domain of theMinistry of Education. Na-tional enthusiasm for pre-primary education has been spurred on by analarming rate of failure out of first grade: 29.3% as of 2001 (UNESCO,2006). This compares unfavorably even with most Sub-Saharan Africancountries, which have the world's highest regional average failure rate.Additionally, Brazil performs far below peers with similar per capitaGDP on Program for International Student Assessment (PISA) tests.With a lot of studies emerging saying that early childhood educationcan equip children with the cognitive and non-cognitive skills requiredfor success in primary school and beyond, the Brazilian federal govern-ment has been keen to capitalize on this. Souza (2005), Minister of Edu-cation during 1995–2002, describes the federal government'smotivations for greater education investment andmonitoring of per-formance indicators.

Despite this rhetoric, however, the federal government has largely re-cused itself frompre-primary and primary education policy. Part of this isan artifact of the 1988 Constitution, which gave considerable policy au-tonomy to municipal governments. As Plank (1996) notes: “TheConstitution's concession of independence to municipalities significantlyreduces the power both of the Ministry of Education and state govern-ments,while greatly expanding opportunities for administrative and pol-icy innovation.” Political pressure for municipal autonomy had beenbuilding during the 20 years of military rule preceding the 1988 Consti-tution. About 54% of the congressmen drafting the 1988 Constitutionhad been previously involved in local politics as mayors or city council-men, and they wanted to financially weaken the federal government.

The federal government has interfered in municipal education policyby imposing an education spending floor: at least 25% of municipalrevenue from any source—local or transfer revenue. The federalgovernment has also passed several education finance reform laws.In 1998, the government implemented the “Fund for the Develop-ment of Elementary Education and Teachers” (Fundo deManutençãoe Desenvolvimento do Ensino Fundamental e de Valorização doMagistério), or FUNDEF. This policy was expanded in the 2007“Fund for the Development and Maintenance of Basic Education”(Fundo de Manutenção e Desenvolvimento da Educação Básica), orFUNDEB.

FUNDEF obligated each of Brazil's state andmunicipal governments topay 15% of each of four intergovernmental transfers into a state-level ed-ucation fund. These four transfer receipts were typically the largest foursources of transfer revenue for municipal governments, and included:transfers from the municipal participation fund (FPM), the tax on goodsand services (ICMS), the tax on industrializedproducts (IPI), and transfersunder the complementary law. In Portuguese, these are the Fundo deCompensacao dos Estados Exportadores (IPI), the Imposto SobreCirculacao de Mercadorias e Servicos (ICMS), the Parte do Fundo deParticipacao dos Municipios (FPM), and the Lei Complementar (LC 87/96) transfer. The federal government sets IPI tax rates and the state gov-ernment sets ICMS tax rates subject to federal minima andmaxima. FPMis the “Revenue Sharing Fund of the Municipalities.” Of the federalgovernment's proceeds from the collection of taxes on income and earn-ings and on industrialized products, 22.5% is given to Brazil's municipali-ties as an FPM transfer. Usually, these four transfers form the majority ofthe municipal revenue base. They are in some sense local taxes; the stateand federal governments set the rates and bases, and collect the taxes, butthey then transfer the revenue tomunicipalities on the basis of itsmunic-ipality of origin. These payments into the school fund did count towardthe 25% of revenue that must be spent on education, but municipalitiesalso had to spend 25% of all other revenue (from sources other thanthese four transfers), and an additional 10% of these four transfers, on ed-ucation. For more details, see Gordon and Vegas (2005), Estevan (2007),and UNESCO (2007).

FUNDEF generated 27 education funds: one for each state, and onefor the Federal District. Fund receipts were redistributed back to

municipalities within each state so that each government's share ofthe pie was equal to its share of the state's total enrolled primary schoolchildren. The state government itself also paid into the fund by thesame set of rules, and also claimed a share of the pie equal to its shareof students (although state governments are generally less involved inprimary education, and almost entirely uninvolved in pre-primary educa-tion). All fund receipts had to be spent on education. Additionally, 60% offund receipts had to be spent exclusively on teacher compensation. Eachmunicipality was also required to establish a council that would overseeexpenditures (Sands, 2008). If gross fund receipts alonewere not enoughto achieve a federal minimum investment level per primary school child(set annually), the federal government would ‘top off’ the fund to ensureminimal primary school expenditure per child. Importantly, this federal‘top-off’ ensured that primary education was already universal and ofminimal required quality before the municipality spent the 10% ormore of additional revenue mandated by the 1988 Constitution's25% spending floor on education. As UNESCO (2007) notes, the in-tention was that states would spend this additional (non-schoolfund) money on secondary education, and municipalities wouldspend it on pre-primary education and other education-relatedexpenditures.

On December 23, 1997, President Fernando Henrique Cardoso andMinister of Education Paulo Renato Souza announced the size of thefederal top-off policy that would be put into effect for 1998 (andwould be revised periodically). FUNDEF was implemented at the startof the new school year in 1998, with municipal receipts from the fundbased on the 1997 Census of Schools (Censo Escolar) enrollments data.

In 2007, FUNDEB increased the fraction of transfer revenue to becontributed to the state education fund: from 15% in 2006 to 16.67% in2007, 18.33% in 2008, and 20% in 2009 onward. Also, while FUNDEFonly considered primary school students in assigning capitation grants,FUNDEB gradually expanded this to include pre-primary, primary, andsecondary school students.

Appendix B. Construction of simulated instruments

The constructionof the simulated revenue variable follows techniquesstandard in the public finance literature. The exact redistribution algo-rithms defined by the FUNDEF (1998) and FUNDEB (2007) education fi-nance reform laws provide me with an exogenous source of variation inmunicipal per capita revenue during 1998–2008. This variation is exoge-nous since these federal lawswere enactedwithout regard for any specif-ic municipality or state's revenue or education policies.

The FUNDEF law defined how revenue would be redistributed with-in states in a very careful algorithm, detailed in Ministério da Educação(2006). For 1998–1999, having data on the following variables allowsone to determine how each municipality will fare under FUNDEF: totalrevenue of eachmunicipality in the state, total revenue of the state gov-ernment itself, total children enrolled in municipal primary schools ineach municipality in the state, total children enrolled in state primaryschools in the state, and the annually-set federal ‘minimumexpenditureper primary school student’ amount. Given these variables alone, onecan determine exactly which municipalities will be net winners,which will be net losers, and by how much.

During 2000–2004, the algorithm began to make the federal top-offamount vary based on the primary school level of students. Children ingrades 1–4 of primary school were weighted differently than children ingrades 5–8. During 2005–2006, the algorithm began to make federaltop-off depend not only on the primary school level, but also on urbaniza-tion status. Thus, onewould additionally need toknowhowmany childrenwere enrolled in grades 1–4of urbanprimary schools, howmanywere en-rolled in grades 1–4 of rural primary schools, howmany were enrolled ingrades 5–8 of urban primary schools, and how many were enrolled ingrade 5–8 of rural primary schools. Beginning in 2007, under FUNDEB,the algorithm was complicated even more, to depend (in addition to allthese previous factors) on the number of children enrolled in urban vs.

Table C.1IV Results, showing how inequality and median income moderate the effect of revenue on school infrastructure quality.

Dependent variable is share of schools with: Own building Electricity Indoor bathroom Library Computer

(1) (2) (3) (4) (5)

Panel A: public pre-primary schools

Log (revenue) 0.006 −0.267*** −0.211*** 0.064 0.233**(0.043) (0.053) (0.066) (0.073) (0.091)

Log (revenue) × Gini −0.087 0.490*** 0.709*** −0.011 0.273(0.093) (0.117) (0.148) (0.125) (0.177)

Log (revenue) × log (median income) −0.005 −0.031*** 0.045*** 0.062*** 0.255***(0.007) (0.008) (0.010) (0.011) (0.014)

Observations 49,393 49,393 49,393 49,393 49,393

Panel B: public primary schools

Log (revenue) −0.062 −0.691*** −0.479*** 0.077 0.371***(0.042) (0.092) (0.096) (0.103) (0.109)

Log (revenue) × Gini 0.077 1.156*** 0.920*** −0.214 −0.309(0.090) (0.190) (0.209) (0.178) (0.204)

Log (revenue) × log (median income) −0.001 −0.043*** 0.028* 0.095*** 0.235***(0.008) (0.014) (0.015) (0.017) (0.017)

Observations 41,866 41,866 41,866 41,866 41,866

Panel C: private pre-primary schools

Log (revenue) 0.105 0.009 −0.020 0.097 −0.132(0.112) (0.026) (0.063) (0.176) (0.154)

Log (revenue) × Gini −0.034 −0.030 0.087 −0.078 0.241(0.240) (0.055) (0.140) (0.362) (0.322)

Log (revenue) × log (median income) 0.015 −0.003 −0.000 −0.031 −0.057**(0.018) (0.005) (0.011) (0.026) (0.024)

Observations 26,605 26,605 26,605 26,605 26,605

Panel D: private primary schools

Log (revenue) 0.100 −0.033 0.037 −0.019 −0.147(0.109) (0.036) (0.066) (0.185) (0.182)

Log (revenue) × Gini −0.138 0.043 −0.110 0.197 0.477(0.247) (0.077) (0.144) (0.394) (0.386)

Log (revenue) × log (median income) 0.004 −0.014** −0.008 0.048* 0.050*(0.016) (0.006) (0.010) (0.027) (0.028)

Observations 20,408 20,408 20,408 20,408 20,408

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ThemeanGini coefficient in2000 is 0.39, themean value ofmedianincome per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84.Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.

Appendix C. Appendix tables


Table C.2IV Results, showing robustness of the results to controls for the PMAT program and time-varying political and economic shocks.

Baseline Exclude PMAT municipalities Dummy for PMAT PMAT R$ per capita control PMAT R$, polit. & econ. controls

(1) (2) (3) (4) (5)

Dependent variable: Log (education spending per capita)

Log (revenue) 0.774*** 0.809*** 0.773*** 0.775*** 0.784***(0.081) (0.084) (0.081) (0.081) (0.082)

Log (revenue) × Gini −0.646*** −0.687*** −0.643*** −0.650*** −0.663***(0.165) (0.170) (0.165) (0.165) (0.167)

Log (revenue) × log (median income) −0.206*** −0.201*** −0.206*** −0.205*** −0.203***(0.013) (0.014) (0.013) (0.013) (0.013)

Observations 50,660 47,040 50,660 50,546 50,321

Dependent variable: Log (infrastructure spending per capita)

Log (revenue) 1.540*** 1.628*** 1.543*** 1.550*** 1.515***(0.379) (0.404) (0.379) (0.379) (0.383)

Log (revenue) × Gini −0.196 −0.404 −0.208 −0.206 −0.156(0.744) (0.789) (0.744) (0.745) (0.754)

Log (revenue) × log (median income) 0.262*** 0.273*** 0.260*** 0.261*** 0.257***(0.055) (0.058) (0.055) (0.055) (0.056)

Observations 46,218 42,885 46,218 46,104 45,917

Dependent variable: Log (budget surplus per capita)

Log (revenue) 2.382*** 2.402*** 2.390*** 2.389*** 2.456***(0.548) (0.573) (0.547) (0.549) (0.554)

Log (revenue) × Gini 2.647** 2.279** 2.612** 2.609** 2.493**(1.091) (1.124) (1.089) (1.091) (1.106)

Log (revenue) × log (median income) 0.622*** 0.614*** 0.617*** 0.611*** 0.610***(0.084) (0.089) (0.085) (0.085) (0.086)

Observations 33,745 31,284 33,745 33,657 33,494

Dependent variable: log (public pre-primary enrollment)

Log (revenue) −0.028 −0.035 −0.027 −0.028 −0.041(0.164) (0.174) (0.164) (0.164) (0.164)

Log (revenue) × Gini −0.737** −0.780** −0.741** −0.742** −0.705**(0.324) (0.341) (0.324) (0.325) (0.326)


Observations 49,410 45,923 49,410 49,298 49,074

Dependent variable: log (pubic primary enrollment)

Log (revenue) −0.577*** −0.623*** −0.569*** −0.575*** −0.566***(0.159) (0.163) (0.158) (0.159) (0.160)

Log (revenue) × Gini −0.841*** −0.864*** −0.865*** −0.852*** −0.911***(0.270) (0.281) (0.270) (0.270) (0.272)


Observations 48,822 45,393 48,822 48,710 48,486

Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). PMAT is a federal subsidized loan program aimed at improvingmunicipal tax collection. PMAT R$ per capita is positive for the years a municipality had a PMAT program, and is 0 otherwise. Political controls include the mayor's vote share, the numberof political parties competing formayor, and anHHI of between-party competition formayor. Economic controls include the current yearworld pricemultiplied by the 1994number of hectaresfor coffee, cocoa, and bananas. The mean Gini coefficient in 2000 is 0.39, the mean value of median income per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, andmean revenue per capita (in 100 s 2005 Reais) is 9.84.Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


Table C.3IV Results, showing results on education spending for a sub-sample of particularly unconstrained municipalities.

All municipalities Municipalities in top third of 1997 educationspending as a share of revenue

Municipalities in bottom third of FUNDEFreceipts as a share of revenue

(1) (2) (3)


Panel A: 2000 inequality and income

Log (revenue) 0.774*** 0.939*** 1.033***(0.081) (0.110) (0.207)

Log (revenue) × Gini −0.646*** −1.080*** −0.936*(0.165) (0.212) (0.490)

Log (revenue) × log (median income) −0.206*** −0.203*** −0.160***(0.013) (0.021) (0.038)

Observations 50,660 17,207 14,989R-squared 0.782 0.824 0.673Number of municipalities 3914 1346 1134Kleibergen-Paap rk Wald F statistic 640.41 219.41 157.81

Panel B: predicted inequality and income

Log (revenue) 0.748*** 0.696*** 0.976***(0.054) (0.087) (0.107)

Log (revenue) × predicted Gini −0.220*** −0.244*** −0.303**(0.060) (0.085) (0.143)

Log (revenue) × log (predicted income) −0.186*** −0.198*** −0.082*(0.015) (0.026) (0.046)

Observations 50,674 17,207 14,991R-squared 0.778 0.823 0.672Number of municipalities 3917 1346 1135Kleibergen-Paap rk Wald F statistic 645.04 212.46 172.59

Notes: An observation is a municipality-year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates p b .10. All specifications includemunicipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ThemeanGini coefficient in2000 is 0.39, themean value ofmedianincome per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84. The mean predicted Gini coefficient during1995–2008 is 0.53 and the mean predicted median income per capita during 1995–2008 is 0.25 (measured in 10,000 s of 2005 Reais).Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


Table C.4IV Results, showing how the effects of inequality and median income vary when including other socioeconomic variables interacted with revenue.

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


Log (revenue) 0.778*** 0.698*** 0.782*** 0.748***(0.081) (0.084) (0.082) (0.090)

Log (revenue) × Gini −0.654*** −0.458*** −0.651*** −0.612***(0.165) (0.173) (0.166) (0.173)


Log (revenue) × log (% urban in 2000) −0.046***(0.016)

Log (revenue) × log (labor force participation rate in 2000) 0.014(0.041)

Log (revenue) × log (racial fractionalization index in 2000) −0.016(0.014)

Observations 50,449 50,449 50,449 50,449Kleibergen-Paap rk Wald F statistic 641.34 480.73 481.20 422.11

Dependent variable: log (public pre-primary enrollment rate)

Log (revenue) −0.023 −0.080 0.032 −0.133(0.165) (0.168) (0.165) (0.179)

Log (revenue) × Gini −0.750** −0.611* −0.705** −0.593*(0.325) (0.335) (0.326) (0.339)


Log (revenue) × log (% urban in 2000) −0.033(0.028)

Log (revenue) × log (labor force participation rate in 2000) 0.217**(0.087)

Log (revenue) × log (racial fractionalization index in 2000) −0.056*(0.029)


Dependent variable: log (public primary enrollment rate)

Log (revenue) −0.575*** −0.665*** −0.523*** −0.550***(0.159) (0.164) (0.159) (0.166)

Log (revenue) × Gini −0.842*** −0.620** −0.801*** −0.877***(0.271) (0.282) (0.271) (0.276)


Log (revenue) × log (% urban in 2000) −0.053**(0.021)

Log (revenue) × log (labor force participation rate in 2000) 0.204***(0.060)

Log (revenue) × log (racial fractionalization index in 2000) 0.012(0.026)


Notes: An observation is amunicipality - year. I instrument for revenue using simulated revenue. Robust standard errors are in parentheses, and clustered at themunicipality level. ***indicatesp b .01; **indicates p b .05; *indicates pc.10. All specifications include municipality and year fixed effects, controls for population, age 0–6 population, and age 7–15 population, a control forgrossmunicipal product (GMP), and a linear time trend interactedwith the pre-reform, 1997 levels of log(per capita revenue), log(per capita transfer revenue), log(public pre-primary enroll-ment rate), log(public primary enrollment rate), and log(share of primary school students in state or federally-run schools). ThemeanGini coefficient in 2000 is 0.39, themean value ofmedianincome per capita (as of the 2000 Census, measured in 10,000 s of 2005 Reais) is 0.23, and mean revenue per capita (in 100 s 2005 Reais) is 9.84.Sources: Author's calculations based on data from IBGE, Ministry of Education, and Tesouro Nacional.


rural creches, preschools, secondary schools, and adult education (EJA).Fortunately, the annual Census of Schools contains all of the required infor-mation, as it is the basis for the distribution of FUNDEF/B funds. Ministérioda Educação (2008) describes the changes initiated in 2007.

While the laws induced endogenousmunicipal responses, I can simu-late (predict) each municipality's total revenue by using the exact algo-rithms of the laws in each year, but using pre-reform (1997) data onrevenue and enrollments. To better predict how enrollments will expandover the sample period without introducing endogenous information, Iexploit the nation-wide rate at which municipal governments took overthe state's role in primary education during 1998–2008. Primary educa-tion is usually run by municipal governments, but the state governmentis involved in some municipalities more than in others. I wanted toaccount for the fact that some municipalities had a greater potential tosecure higher revenue through take-over than did others. The federal

rate of takeover is exogenous to pre-primary education policy in anygiven municipality. However, by taking this into account and thusallowing simulated revenue to grow more in municipalities with heavyinitial state and federal involvement in primary education in the pre-period, I can better predict revenue without drawing on any onemunicipality's endogenous response to the reforms.

References

Alesina, Alberto, Rodrik, Dani, 1994. Distributive politics and economic growth. Q. J. Econ.109 (2), 465–490.

Araujo, M. Caridad, Ferreira, Francisco H.G., Lanjouw, Peter, Ozler, Berk, 2008. Decentral-ization and economic development. J. Public Econ. 92, 1022–1046.

Ball, Laurence, Mankiw, N. Gregory, 1995. What do budget deficits do? NBER WorkingPaper No. 5263.

Banerjee, Abhijit, Somanathan, Rohini, 2007. The political economy of public goods: someevidence from India. J. Dev. Econ. 82, 287–314.

http://refhub.elsevier.com/S0304-3878(13)00178-8/rf0005









Bardhan, Pranab, Mookherjee, Dilip, 2006. Pro-poor targeting and accountability of localgovernments in West Bengal. J. Dev. Econ. 79, 303–327.

Barros, Ricardo, de Carvalho, Mirela, Franco, Samuel, Mendonça, Rosane, 2009. Markets,the state and the dynamics of inequality: Brazil's case study. Working Paper.

Barzel, Yoram, 1973. Private schools and public school finance. J. Polit. Econ. 81 (1),174–186.

Bates, Robert, 1973. Ethnicity in contemporary Africa. Monograph, Program of EasternAfrican Studies, XIV. Maxwell School of Citizenship and Public Affairs: SyracuseUniversity.

Bates, Robert, 1976. Rural Responses to Industrialization: A Study of a Village in Zambia.Yale University Press, New Haven.

Bates, Robert, 1981. States and markets in tropical Africa: the political basis of agriculturalpolicy. Series on Social Choice and Political EconomyUniversity of California Press,Berkeley.

Bellettini, Giorgio, Ceroni, Carlotta Berti, 2007. Income distribution, borrowing constraintsand redistributive policies. Eur. Econ. Rev. 51 (3), 625–645.

Benabou, Roland, 1996. Heterogeneity, stratification, and growth: macroeconomicimplications of community structure and school finance. Am. Econ. Rev. 86 (3),584–609.

Berlinski, Samuel, Galiani, Sebastian, Gertler, Paul, 2009. The effect of pre-primary educationon primary school performance. J. Public Econ. 93 (1–2), 219–234.

Besley, Timothy, Burgess, Robin, 2002. The political economy of governmentresponsiveness: theory and evidence from India. Q. J. Econ. 117 (4), 1415–1451.

Besley, Timothy, Coate, Stephen, 1991. Public provision of private goods and the redistribu-tion of income. Am. Econ. Rev. 81 (4), 979–984.

Boustan, Leah, Ferreira, Fernando, Winkler, Hernan, Zolt, Eric M., 2013. The effectof rising income inequality on taxation and public expenditures: evidencefrom US municipalities and school districts, 1970–2000. Rev. Econ. Stat. 95 (4),1291–1302.

Carneiro, Pedro Manuel, Heckman, James, 2003. Human capital policy. In: Heckman,James, Krueger, Anne (Eds.), Inequality in America: What Role for Human CapitalPolicies? MIT Press, Boston.

Casamatta, Georges, Cremer, Helmuth, Pestieau, Pierre, 2000. The political economy ofsocial security. Scand. J. Econ. 102, 503–522.

Chandra, Kanchan, 2004. Why Ethnic Parties Succeed: Patronage and Ethnic Head Countsin India. Cambridge University Press, Cambridge.

Chetty, Raj, Friedman, John N., Hilger, Nathaniel, Saez, Emmanuel, WhitmoreSchanzenbach, Diane, Yagan, Danny, 2011. How does your kindergarten class-room affect your earnings? Evidence from project Star. Q. J. Econ. 126 (4),1593–1660.

Chernick, Howard, 2005. On the determinants of subnational tax progressivity in the U.S.,Nat. Tax J. 58 (1), 93–112.

Corcoran, Sean, Evans, William N., 2010. Income inequality, the median voter, and thesupport for public education. NBER Working Paper 16097.

Couto, Cláudio, Abrucio, Fernando Luiz, 1995. Governando a Cidade? A Força e a Fraquezada Câmara Municipal. São Paulo Perspect. 9, 57–65.

Cunha, F., Heckman, J., Lochner, L., Masterov, D., 2006. Interpreting the evidence on lifecycle skill formation. In: Hanushek, E., Welch, F. (Eds.), Handbook of the Economicsof Education. Elsevier, Amsterdam.

Currie, Janet, 2001. Early childhood education programs. J. Econ. Perspect. 15, 213–238.Currie, Janet, Thomas, Duncan, 1995. Does head start make a difference? Am. Econ. Rev.

85, 341–364.de Janvry, Alain, Finan, Frederico, Sadoulet, Elisabeth, 2010. Local electoral incentives and

decentralized program performance. NBER Working Paper No. 16635.de la Croix, David, Doepke, Matthias, 2009. To segregate or to integrate: education politics

and democracy. Rev. Econ. Stud. 76, 597–628.DeLong, J. Bradford, Summers, Lawrence H., 1991. Equipment investment and economic

growth. Q. J. Econ. 106, 445–502.Engle, Patrice L., Black, Maureen M., Behrman, Jere R., Cabral, Meena, de Mello, Paul J.,

Gertler, Lydia Kapiriri, Martorell, Reynaldo, Young, Mary Eming, 2007. Child develop-ment in developing countries. Lancet 369, 229–242.

Engle, Patrice L., Fernald, Lia C.H., Alderman, Harold, Behrman, Jere, O'Gara, Chloe,Yousafzai, Aisha, Cabral de Mello, Meena, Hidrobo, Melissa, Ulkuer, Nurper, Ertem,Ilgi, Iltus, Selim, 2011. Strategies for reducing inequalities and improving develop-mental outcomes for young children in low-income and middle-income countries.Lancet 378, 1339–1353.

Epple, Dennis, Romano, Richard, 1996. Ends against the middle: determining publicservice provision when there are private alternatives. J. Public Econ. 62, 297–325.

Estevan, Fernanda, 2007. The quality of public education and private school enrollment:an assessment using Brazilian data. Working Paper.

Evans, David, Kosec, Katrina, 2012. Early child education: making programs work forBrazil's most important generation. World Bank Study.

Ferraz, Claudio, Finan, Frederico, 2009a. Electoral accountability and corruption: evidencefrom the audits of local governments. NBER Working Paper 14937.

Ferraz, Claudio, Finan, Frederico, 2009b. Motivating politicians: the impacts of monetaryincentives on quality and performance. NBER Working Paper 14906.

Ferreira, Francisco H.G., Lanjouw, Peter, Neri, Marcelo, 2003. A new poverty profile forBrazil using PPV, PNAD and census data. Rev. Bras. Econ. 57 (1), 59–92.

Foster, Andrew, Rosenzweig, Mark, 2003. Democratization, decentralization and thedistribution of local public goods in a poor rural economy. BREAD Working PaperNo. 010.

Gadenne, Lucie, 2012. Taxme, but spendwisely: public finance and government account-ability. Working Paper.

Galor, Oded, Moav, Omer, Vollrath, Dietrich, 2009. Inequality in landownership, the emer-gence of human-capital promoting institutions, and the great divergence. Rev. Econ.Stud. 76 (1), 143–179.

Glomm, Gerhard, Ravikumar, B., 1998. Opting out of publicly provided services: a major-ity voting result. Soc. Choice Welf. 15, 187–199.

Goldin, Claudia, Katz, Lawrence F., 1999. The Returns to skill in the United States acrossthe twentieth century. NBER Working Paper No. 7126.

Gordon, Nora, 2004. Do federal grants boost school spending? Evidence from TitleI. J. Public Econ. 88 (9–10), 1771–1792.

Gordon, Nora, Vegas, Emiliana, 2005. Educational finance, equalization, spending, teacherquality, and student outcomes: The case of Brazil's FUNDEF. In: Vegas, Emiliana (Ed.), In-centives to Improve Teaching: Lessons From Latin America. The World Bank,Washington, D.C.

Grossman, Gene, Helpman, Elhanan, 1994. Protection for sale. Am. Econ. Rev. 84 (4),833–850.

Gutiérrez, Catalina, Tanaka, Ryuichi, 2009. Inequality and education decisions in develop-ing countries. J. Econ. Inequal. 7 (1), 55–81.

Heckman, James, Masterov, Dimitriy V., 2007. The productivity argument for investing inyoung children. Working Paper.

Hoxby, CarolineM., 2001. All school finance equalizations are not created equal. Q. J. Econ.116 (4), 1289-1231.

Husted, T.A., Kenny, L.W., 1997. The effect of the expansion of the voting franchise on thesize of government. J. Polit. Econ. 105 (1), 54–82.

Instituto Brasileiro de Geografia e Estatística (IBGE), 2007. 2007 Regional AccountsReport. http://www.ibge.gov.br/english/estatistica/economia/contasregionais.

Instituto Brasileiro de Geografia e Estatística (IBGE), 2010. Cities@. http://www.ibge.gov.br/cidadesat/.

Jaffrelot, Christophe, 2003. India's Silent Revolution: The Rise of the Lower Castes in NorthIndia. Columbia University Press, New York.

Karabarbounis, Loukas, 2011. One dollar, one vote. Econ. J. 121 (553), 621–651.Keefer, Philip, Khemani, Stuti, 2005. Democracy, public expenditures, and the poor:

understanding political incentives for providing public services. World Bank Res.Obs. 20 (1), 1–27.

Lindert, Peter H., 1994. What limits social spending. Explor. Econ. Hist. 31 (1), 1–37.Lindert, Peter H., 1996. The rise of social spending, 1880–1930. Explor. Econ. Hist. 33 (1),

1–34.Malamud, Semyon, Trojani, Fabio, 2009. Variance, risk and median preserving spreads.

Swiss Finance Institute Research Paper No. 09–13.Mansuri, Ghazala, Rao, Vijayendra, 2004. Community-based and -driven development: a

critical review. World Bank Res. Obs. 19 (1), 1–39.Martínez-Fritscher, André, Musacchio, Aldo, Viarengo, Martina, 2010. The Great Leap

Forward: The Political Economy of Education in Brazil, 1889–1930.Meltzer, A., Richard, S., 1981. A rational theory of the size of government. J. Polit. Econ. 89

(5), 914–927.Meltzer, A., Richard, S., 1983. Tests of a rational theory of the size of government. Public

Choice 41, 403–418.Ministério da Educação, 2006. Legislação. http://www.fnde.gov.br/index.php/fundef-

legislacao.Ministério da Educação, 2008. FUNDEB: Manual de Orientação. http://www.fnde.gov.br/

index.php/fundeb-publicacoes.Moene, Ove Karl, Wallerstein, Michael, 2001. Inequality, Social Insurance, and Redistribu-

tion. Am. Polit. Sci. Rev. 95 (4), 859–879.Muralidharan, Karthik, Sundararaman, Venkatesh, 2011. Teacher performance pay:

experimental evidence from India. J. Polit. Econ. 119 (1), 39–77.Pande, Rohini, 2003. Can mandated political representation provide disadvantaged

minorities policy influence? Theory and evidence from India. Am. Econ. Rev. 93 (4),1132–1151.

Patrinos, Harry A., 2007. The living conditions of children. World Bank Policy ResearchWorking Paper No. 4251.

Perotti, Roberto, 1996. Growth, income distribution, and democracy: what the data say.J. Econ. Growth 1 (2), 149–187.

Persson, Torsten, Tabellini, Guido, 1994. Is inequality harmful for growth? Am.Econ. Rev. 84 (3), 600–621. http://www.jstor.org/discover/10.2307/2118070?uid=2&uid=4&sid=21103258199933.

Plank, David N., 1996. The Means of Our Salvation: Public Education in Brazil, 1930–1995.Westview Press, Boulder.

Ramcharan, Rodney, 2010. Inequality and redistribution: evidence from U.S. Counties andStates, 1890–1930. Rev. Econ. Stat. 92 (4), 729–744.

Rodríguez, Francisco, 2004. Inequality, redistribution, and rent-seeking. Econ. Polit. 16(3), 287–320.

Sands, Joseph C., 2008. Education funding in a recentralizing democracy: a cautionary taleof four Brazilian Cities 50 (3), 93–120.

Schwabish, Jonathan A., 2008. The effects of earnings inequality on state social spendingin the United States. Public Financ. Rev. 36 (5), 588–613.

Schwabish, Jonathan, Smeeding, T.M., Osberg, Lars, 2006. Income distribution and socialexpenditures: a cross-national perspective. In: Papadimitrious, Dimitri B.,Northampton, Mass, Elgar, Edward (Eds.), The Distributional Effects of GovernmentSpending and Taxation.

Skidmore, Thomas, 2004. Brazil's persistent income inequality: lessons from history. Lat.Am. Polit. Soc. 46 (2), 133–150.

Stiglitz, Joseph, 1974. The demand for education in public and private school systems.J. Public Econ. 3 (4), 349–385.

Souza, Celina, 2005. Brazil's Tax System: the dilemmas of policy reform. FOCAL PolicyPaper 05–02.

Suárez Serrato, Juan Carlos, Wingender, Philippe, 2011. Estimating the incidence of gov-ernment spending. Working Paper.

Tesouro Nacional, 2011. https://www.tesouro.fazenda.gov.br/ (last accessed October 2011).Tiebout, Charles M., 1956. A pure theory of local expenditures. J. Polit. Econ. 64 (5),

416–424.







































































































http://www.ibge.gov.br/english/estatistica/economia/contasregionais

http://www.ibge.gov.br/cidadesat/

http://www.ibge.gov.br/cidadesat/




















http://www.fnde.gov.br/index.php/fundef-legislacao

http://www.fnde.gov.br/index.php/fundef-legislacao

http://www.fnde.gov.br/index.php/fundeb-publicacoes

http://www.fnde.gov.br/index.php/fundeb-publicacoes












http://www.jstor.org/discover/10.2307/2118070?uid=2&uid=4&sid=21103258199933

http://www.jstor.org/discover/10.2307/2118070?uid=2&uid=4&sid=21103258199933























https://www.tesouro.fazenda.gov.br/




United Nations Educational, Scientific, and Cultural Organization (UNESCO), 2006. Strongfoundations: early childhood care and education. EFA Global Monitoring Report.UNESCO Publishing, Paris.

United Nations Educational, Scientific, and Cultural Organization (UNESCO), 2007. Policyreview report: early childhood care and education in Brazil. UNESCO EducationSector, Early Childhood and Family Policy Series No. 13–2007.

Varshney, Ashutosh, 1995. Democracy, Development, and the Countryside: Urban–ruralStruggles in India. Cambridge University Press, New York.

Wampler, Brian, 2007. Participatory Budgeting in Brazil: Contestation, Cooperation andAccountability. Pennsylvania State University Press, Pennsylvania.

Zolt, E.M., 2009. Inequality, collective action, and taxing and spending patterns of stateand local governments. Tax Law Rev. 62 (4), 445–539.













Documents

Relying on the private sector: The income distribution and public investments in the poor