The Impact of the Bolsa Escola/Familia Conditional Cash Transfer Program

on Enrollment, Drop Out Rates and Grade Promotion in Brazil

Paul Glewwe
Department of Applied Economics
University of Minnesota
Ana Lucia Kassouf
Department of Economics
University of Sao Paulo
August, 2010

Abstract
This paper examines the impact of Brazil's Bolsa Escola (later renamed Bolsa Familia)
program on children's progress in school in Brazil. The Bolsa program, which started
in the 1990s and expanded rapidly in 2001 and 2002, provides monthly cash payments
to poor households if their children (between the ages of 6 and 15) are enrolled in
school. Using eight years of school census data (from 1998 to 2005), our estimation
method compares changes in enrollment and in dropout and grade advancement rates
across schools that adopted the Bolsa program at different times. We estimate that,
after accounting for cumulative effects, the Bolsa program has increased enrollment in
Brazil by about 5.5 percent in grades 1-4 and by about 6.5 percent in grades 5-8. We
also estimate that the program has lowered dropout rates by about 0.5 percentage points
and raised grade promotion rates by about 0.9 percentage points for children in grades
1-4, and has reduced dropout rates by about 0.4 percentage points and increased grade
promotion rates by about 0.3 percentage points for children in grades 5-8. Only about
one third of Brazils children participate in the Bolsa program, so the assumption that
these results are due only to the impact of the program on participants, with no spillover
effects onto non-participants, implies that the impact of participating in the Bolsa
program is about three times higher than these estimates. While these impacts cast a
favorable light on the program, simple calculations based on the enrollment impacts
suggest that the likely benefits in terms of increased wages may not exceed the costs of
the program.

For helpful comments and discussion, we would like to thank Marcio Bezerra, Qiuqiong
Huang, Hanan Jacoby, Edson Lopes, Sarmistha Pal, Rafael Ribas and Elisabeth Sadoulet.

I. Introduction
Many economists agree that higher levels of education increase economic growth
(Barro, 1991; Mankiw, Romer and Weil, 1992; Krueger and Lindahl, 2001; Sala-i-Martin,
2004; Hanushek and Woessmann, 2008), raising incomes and, more generally, the quality
of life. Economists support for education is matched by strong support from international
aid agencies. Two of the eight Millennium Development Goals adopted at the 2000 United
Nations Millennium Summit focus on education: first, all children should finish primary
school; and second, gender equality should prevail at all education levels.
The Millennium Goals may not be met due to constraints parents face when they
make choices for their childrens schooling. The main barriers are the direct costs (school
fees, books, uniforms, etc.) and the opportunity cost of time in school children will
have less time to work or do other activities if they spend more time studying in school
(and at home). Several countries have not only ended fees or provided free meals and
uniforms; they also pay parents if their children attend school. These are conditional cash
transfer (CCT) programs, and they have two objectives: (1) alleviate poverty today; and
(2) increase poor childrens stock of human capital to ensure their future well being. The
first objective is met when poor families receive program payments. The second is
achieved by conditioning those payments on certain behaviors, such as immunizing
young children and enrolling older children in school. Today, many developing countries
have CCT programs. The two largest are Brazils Bolsa Escola program (later renamed
Bolsa Familia) and Mexicos Progresa program (later renamed Oportunidades).
Latin America has made significant progress in education since 1980. For
example, the net primary enrolment rate rose from 70% in 1980 to 94% in 2004, and the

net secondary enrolment rate jumped from 16% to 61% (Damon and Glewwe, 2007).
But further progress is possible. For example, Brazils enrolment rates rose from 86% in
1990 to 97% in 2001 for 8-11 year old children, but among 15 year olds in 2001 it was
only 87%. Indeed, in 2001 40% (9 million) of Brazils 18 to 25 year old youths had less
than 8 years of education (PNAD, 2001). To encourage all children to complete 8 years
of school, Brazil launched the Bolsa Escola program in 2001.
Bolsa Escola (renamed Bolsa Familia in 2003) provides transfers to poor families
with school-age children if their children are enrolled in school. Several studies have
shown that CCT programs in Latin America raise students educational outcomes, but
almost all examined Mexicos Progresa or similar programs in Central America. These
studies are convincing because those CCT programs were implemented as randomized
trials. In contrast, analyses of the impact of Bolsa Escola (which was not randomized) on
education in Brazil are rare. This is unfortunate since Brazil is the largest and most
populous nation in Latin America, and Bolsa is the worlds largest CCT program. This
paper uses a school level panel data set to evaluate the impact of Bolsa Escola/Familia
on enrollment, dropping out and grade promotion at the primary and lower secondary
levels. Both school and municipio (county) level estimates are presented.
The following sections review the literature, describe the Bolsa Escola/Familia
program and the data, explain the estimation methodology, and present the results. A
final section summarizes the findings and provides suggestions for future research.

II. Literature Review

Parker, Rubalcava and Teruel (2008) review many studies that analyze the impact
of CCT programs on schooling in developing countries. Maluccio and Flores (2004)
2

estimate that Nicaraguas Red de Proteccion Social raised enrolment by 17.7 percentage
points, daily attendance by 11 percentage points, and retention rates by 6.5 percentage
points, for Nicaraguan children in grades 1 to 4. Honduras Programa de Asignacion
Familiar had positive, but smaller, impacts on daily attendance and enrolment, and a small
negative effect on dropping out, for children age 6 to 13 (Glewwe and Olinto, 2004).
Attanasio, Fitzsimons and Gomez (2005) found that Colombias Familias en Accion
increased enrolment for children age 12-17 but had no effect for 8-11 year olds. Schady
and Araujo (2008) estimated a positive impact of Ecuadors Bono de Desarrollo Humano
on enrolment. Two programs in Asia focus on girls education. Khandker, Pitt and Fuwa
(2003) found that a CCT program in Bangladesh raised 11-18 year old girls enrolment,
while Filmer and Schady (2008), estimated that the Japan Fund for Poverty Reduction
initiative in Cambodia raised secondary school girls enrolment and attendance.
Many studies examine Mexicos Progresa/Oportunidades program. Behrman,
Sengupta and Todd (2000) found that it raised 12 to 14 year old girls enrolment but did
not affect younger children; they attribute the latter to the already high enrolment rates
for younger children. They also found a significant drop in the schooling gap (difference
between actual grade attained and the grade a child would have attained had he or she
entered school at age six and never repeated) for 11-15 year olds. Schultz (2004) found a
positive effect of Progresa on girls and boys enrolment, with a larger effect for girls.
Dubois, de Janvry and Sadoulet (2004) estimate that Progresa increased childrens
probability of staying in school, and primary level grade progression and completion, but
reduced secondary grade progression. Skoufias and Parker (2001) found that Progresa
significantly increased enrolment and reduced employment among both boys and girls.

Unlike the vast research on Mexicos Progresa, Brazils Bolsa Escola/Familia is

rarely studied, perhaps because it was not implemented as a randomized trial. Bolsas first
evaluation (World Bank, 2001) focused on its operation in the Federal District (Brasilia,
Brazils capital) in 1995 and 1996. This study simply compared beneficiaries and nonbeneficiaries; Bolsa appeared to lower dropping out by 6 percentage points and raise grade
promotion by 8-10 percentage points, but had little effect on students test scores. Yet this
study has several shortcomings. First, and most important, it ignored initial differences
across beneficiaries and non-beneficiaries. Second, it included only the capital, a wealthy
area unrepresentative of Brazil. Third, Bolsa has changed since 1996, as explained below.
Bourguignon, Ferreira and Leite (2003) use a 1999 Brazilian household survey to
estimate a model of household behavior and simulate the (future) impact of Bolsa. Their
results suggest that Bolsa induces most eligible out of school youths to enroll. Yet they
caution that their results can provide only orders of magnitude for the likely effects of
transfer programs of the Bolsa Escola type. Some assumptions needed for estimation,
for example that children not in school who work away from home do not do household
work, are dubious. Thus they provide only imprecise estimates of Bolsas impact.
Cardoso and Souza (2003) and Ferro and Kassouf (2005) both estimate that Bolsa
has a large positive impact on enrollment. Yet both used cross-sectional data and did
little to control for selection into the program and omitted variable bias. Also, both used
data from before Bolsas sharp expansion, the 2000 Demographic Census and the 2001
PNAD (National Household Survey), respectively. Lastly, de Janvry, Finan and Sadoulet
(2007) find that Bolsa reduced dropping out by 8 percentage points but did not affect
repetition. Yet the analysis is limited to 5 states in Northeast Brazil, and they had many

fewer schools than do the data used in this paper, reducing the precision of the estimates.
Their data also lack some key variables, such as race; we find that estimates vary by race.
In summary, while the Bolsa Escola/Familia program is the worlds largest CCT
program (see below), there is very little research on it. The research to date suffers from
estimation problems, data that cover only a small part of Brazil, and (in most cases)
analysis of the earliest version of the program. This paper uses eight years of nationwide
data, including five years after the program greatly expanded, and implements an
estimation procedure that minimizes a wide variety of estimation problems.

III. Description of Bolsa Escola/Familia Program

The first municipios to implement Brazils Bolsa Escola were the cities of Brasilia
(the Federal District) and Campinas (in So Paulo State), in 1995. Both programs provided
cash payments to poor families with children age 6 to 15 conditional on those children enrolling in school and attending at least 85% of school days. By 1998 over 50 municipios
in 7 states (out of 27) had similar programs, but this was only 1% of Brazils municipios.
Given the programs popularity, and positive evaluations of other Latin American
CCT programs, President Fernando Henrique Cardosos government created the Federal
Bolsa Escola program in April, 2001. By the end of 2001, nearly 5 million families were
receiving payments. In October 2003, President Luis Inacio Lula da Silva merged other
programs into Bolsa Escola and renamed it Bolsa Familia. Benefits were extended to
poor families with children 0 to 5 years old or with a pregnant or breastfeeding woman,
and to all very poor families (even those without children). By 2007, over 11 million
families (about 46 million people, one fourth of Brazils population) received Bolsa payments. The government budget for the program was 7.5 billion Reais (about 4 billion
5

U.S. $) in 2006, about 0.4% of GDP. In terms of families served, this is larger than
Mexicos Progresa program, which served about 5 million families in 2006, but in terms
of program cost as a percentage of GDP both are about 0.4% of GDP (Bastagli, 2008).
To qualify for Bolsa Familia, a familys monthly per capita income must be less
than 120 Reais (about 60 U.S. $), i.e. below one-half of Brazils minimum wage. Those
with monthly incomes from 60 to 120 Reais are eligible if they have children under 16
years old or a breastfeeding or pregnant woman. These families receive 15 Reais per
month per beneficiary (either a child age 0 to 15 or a pregnant or breastfeeding woman),
up to a maximum of three (to avoid incentives to raise fertility). Families with monthly
per capita incomes below 60 Reais are classified as very poor and receive payments even
if they have no children and no pregnant or breastfeeding women. These families receive
50 Reais per month plus 15 Reais per beneficiary (up to three). 1 To receive the 15 Reais,
each child age 6 to 15 must be enrolled in school and attend at least 85% of school days,
each pregnant or breastfeeding women must obtain prenatal and postnatal health care
services, and children age 0 to 7 must have all recommended vaccinations.
To enroll in Bolsa, families fill out an application, available at their municipios
city hall, that requests information on income and household composition. That
information determines eligibility, subject to the municipios budget for Bolsa. 2 That
budget is set (and financed) by the Federal Government, based on the estimated number
of poor families in the municipio, obtained from the population census and recent PNAD
(Pesquisa Nacional por Amostra Domiclios) household surveys. Bolsa payments are

Payments were raised in 2007 and 2008. The current monthly payment to very poor families is 62 Reais.
The monthly payment per child is 20 Reais for children up to age 15 and 30 Reais for children age 16 or 17.
2
de Janvry et al. (2005) report that in almost all municipios the number of potential beneficiaries greatly
exceeded the number of beneficiaries they could fund with the budget allocated by the central government.

usually given to women, often the household head or the heads wife, via a bank card,
because studies show that women tend to use additional income to raise their families
wellbeing. Soares, Ribas and Osorio (2007) argue that this selection process may provide
benefits to ineligible families since applicants information is not carefully verified, but
Lindert et al. (2007) report that the program is well targeted to the intended beneficiaries.
Although some view Bolsa Escola/Familia as a school program, schools have no
role in funding or operating the program; Bolsa funds are transferred directly from the
Ministry of Social Development to households via bank cards, and the municipio
governments only role is to identify beneficiaries. 3 Thus schools have no incentive to
misreport information on attendance to retain some type of program funds (though
schools could misreport attendance to benefit a students family).

IV. Data Available

The main data source used in this paper is Brazils school census. Each year it is
administered to over 250,000 public and private schools, from preschools to high schools.
These schools have 53 million students and 2 million teachers. The education outcomes
in this census are enrolment, dropping out, grade promotion and repetition; there are no
data on students academic performance. The census also collects school characteristics,
including facilities (library, computer lab, internet connection, a science lab, and gymnasium) and participation in various government programs (meals, textbooks, etc.).
The school census data from 1998 to 2005 were used to create a panel of schools.
Table 1 shows the number of schools of all types in each census. We focus on schools
with grades 1-8, which covers the age range to be eligible to receive Bolsa benefits. The
3

In the first 2-3 years of operation, schools assisted by publicizing the program and accepting applications.

third column shows the number of schools with grades 1-4 or 5-8 (or both). Over time
that number steadily drops, reflecting Brazils demographic trends (lower fertility) and its
policy of closing schools in bad condition and merging small schools into larger ones.
The last column in Table 1 shows the panel data for schools with grades 1-4 and/or
5-8. Brazil had 187,514 of those schools in 1998, of which 174,153 could be matched (by
school ID codes) to the 183,475 schools in 1999 census. Non-matches reflect new schools,
schools closed or merged into larger schools, and school code errors. Each row in Table 1
shows how the panel data set is reduced as another year is added. In 2005 there are
136,114 schools with grades 1-4 or 5-8, of which 107,243 (79%) have data for all years
from 1998 to 2005. While this attrition may lead to an unrepresentative sample, Table A.1
shows no differences between the panel and cross-sectional data. Also, note that the PNAD
data indicate that only 3% of poor students enroll in private schools; the poor can rarely
afford to enroll their children in those schools, so this paper focuses on public schools.
Table 2 shows key school characteristics from 1998 to 2005 for public schools
with grades 1-4. Average total enrollment declined from 135 to 110, reflecting demographic trends and reduced repetition. The grade promotion rate is the percentage of
students who, based on academic performance, advance to the next grade; thus it is an
indicator of that performance. It increased from 68% in 1998 to 73% in 2005, which may
indicate better academic performance but also reflects a social promotion policy. The
dropout rate (fraction of enrolled students who leave school before the school year ends)
fell sharply, from 14.5% to 8.9%. Class size decreased modestly, from 27.5 to 24.5. A
final educational outcome of interest is age-grade distortion. In Brazil, students usually
start grade 1 at age 7, so they should finish grade 4 at age 11 and grade 8 at 15. A student

who finishes grade 4 at an age older than 11 has age-grade distortion. The percentage of
grade 4 students with age-grade distortion fell from 57% in 1998 to 37% in 2005.
Table 2 also has school quality indicators and Bolsa Escola/Familia participation
rates. Since 1998 the percentage of college-educated teachers increased dramatically,
from 8.4% to 26.2% in 2005. The availability of computers, printers and computer labs
also increased rapidly. The last two columns show participation in Bolsa (that is, the
percentage of schools reporting that one or more students participate in Bolsa), which is
available since 2001. In 2001 only 23.5% of schools reported student participation in the
program. This rose rapidly in 2002, to 84.7%, after which participation rose more slowly,
reaching 90.8% in 2005. The last column shows Ministry of Social Development data on
the percentage of families with children in grades 1-4 that participate in Bolsa, starting in
2001. In both 2001 and 2002, about 17% of families participated. Thereafter, the rate
steadily increased, reaching 33.8% in 2005. Unlike the school data, the sharpest jump is
from 2000 to 2001, not 2001 to 2002. This reflects the fact that the school data were
collected in March and April, while the Ministry data are from December, and (as
explained below) much of Bolsas expansion occurred in the last six months of 2001.
Table 3 shows several school characteristics in 2001, for all public schools with
grades 1-4 and separately for schools with and without Bolsa students. It shows that
simple comparisons of schools with and without Bolsa students can be misleading. The
year 2001 is used because it has the most equal split of schools with and without Bolsa
students. Enrollment and grade promotion are lower, and the dropout rate is higher, in
Bolsa schools, suggesting that the program reduced enrollment and grade promotion and
increased dropping out, which is doubtful. The obvious explanation is that Bolsa is

targeted to poor children, who have lower initial education outcomes. Table 3 also shows
that schools with Bolsa students have fewer college-educated teachers, computers,
printers and computer labs. Only students per classroom is similar for the two groups.
The data in Table 2 suggest that Bolsa raised enrollment and grade promotion,
and reduced dropping out, in schools with grades 1-4. First, average school enrollment
dropped by 4.0 students from 1999 to 2000, before Bolsas national expansion. During
Bolsas most rapid expansion, from 2000 to 2002, total enrollment dropped by only 3.1
students from 2000 to 2001 and by only 2.6 from 2001 to 2002. From 2002 to 2005, as
Bolsa expanded more slowly, and the annual drop in enrollment returned to 4 or 5
students per year. Turning to grade promotion, the rate was about 69% or 70% from
1998 to 2000. It increased to 72.4% in 2001 and 73.0% in 2002, the years of Bolsas
rapid expansion, and then stayed around 73% from 2002 to 2005 (except 2004, when it
was 71%). Finally, dropout rates were 13% or 14% from 1998 to 2000, fell to 10.6 in
2001 and 9.5 in 2002, and remained at 9-10% thereafter. Overall, all three education
outcome trends suggest improvements during the two years that Bolsa expanded most
quickly.
Next, consider public schools with grades 5-8. Unlike schools with grades 1-4,
Table 4 shows rising enrollment from 1998 to 2000, followed by a decline. Also unlike
grade 1-4 schools, grade promotion changed little, with no clear trend. Yet the dropout
rate pattern is similar to that of grade 1-4 schools; it fell from 13.6% in 1998 to 9.1% in
2005. Students per classroom is higher in schools with grades 5-8 than in those with
grades 1-4, but in both types it declines steadily over time. Finally, as in schools with
grades 1-4, the age-grade distortion rate steadily fell from 56% in 1998 to 38% in 2005.

10

The percentage of college-educated teachers in public schools with grades 5-8 is

much higher than in those with grades 1-4, and it increased gradually over time. There
was also rapid growth in the percentage of those schools with computers, printers and
computer labs, as was seen for grade 1-4 schools (though grade 5-8 schools are much
more likely to have these resources). Finally, the percentage of grade 5-8 schools with
one or more students in the Bolsa program was only 13% in 2001, but jumped to 76% in
2002, after which it slowly increased to 86% in 2005.
The last three columns of Table 3 compare public schools with grades 5-8 in 2001,
by whether they have students in the Bolsa program. As in grade 1-4 schools, schools with
Bolsa students have lower enrollment and grade promotion and more dropping out. Their
student-teacher ratio is slightly lower, but they have fewer college-educated teachers and
less computer facilities. Again, this likely reflects that Bolsa is targeted to poor children.
Finally, for schools with grades 5-8, Table 4 suggests that Bolsa boosted education
outcomes. This is hard to infer from the quadratic enrollment trend, yet grade promotion
jumped by almost two percentage points in 2001, the first year of Bolsas wide expansion.
More persuasive is the sharp decline in the dropout rate from 13.3 in 2000 to 11.1 in 2001
and 10.3 in 2002; this occurred precisely in the years when Bolsa expanded most rapidly.

V. Methodology
This section explains how Brazils school census data can be used to estimate the
impact of the Bolsa program on education outcomes. It begins by explaining what can be
estimated, and then turns to the details of the estimation method.
A. Analytical Framework. For any given education outcome, two of the most
commonly estimated program effects are: 1. Average treatment effect (ATE) over all
11

students age 6-15, both participants and non-participants; and 2. Average treatment effect
for students age 6-15 who participate in the program (ATT, average treatment effect on
the treated). 4 The 1998-2005 school census data used in this paper have two important
advantages: a very large sample, and panel data that begin before and end after the start
of the Bolsa program. Yet these data also have a serious disadvantage: for each school,
the Bolsa variable indicates only that some students participate, not the percentage of
students who participate. As explained below, this implies that one can estimate only the
average treatment effect (ATE); the average effect of Bolsa on those who participate
(ATT) cannot be estimated. Yet if there are no spillover effects from Bolsa participants
onto non-participants, one can use the proportion of students who participate, p, and an
estimate of ATE to estimate ATT: ATE = ATTp + 0(1-p), so ATT = ATE/p.
Spillover effects are likely to be minimal, for two reasons. First, enrolment among
non-eligible children is already almost 100%, so there is little scope for their enrolment to
increase; the 2001 PNAD indicates that 96.7% of children age 6-15 whose household
income makes them ineligible for Bolsa are already enrolled in school. This figure rises to
98.8% for ineligible children age 7-10, the official age range associated with grades 1-4.
Second, it is unlikely that there are noticeable effects on dropping out and grade promotion of ineligible children from bringing Bolsa students into their classrooms because our
estimated impacts on enrolment (given below) are quite modest, about 3%, and since the
average class size is 25-30 this implies only one new student per classroom. Finally, note

Another possible definition of ATE is the average treatment effect if all students had participated in the
program. In contrast, our definition is an average of ATT for those who participated and of any spillover
effects for non-participants. We prefer this definition because Bolsa is not intended to serve the entire
population, but only the poor. Moreoever, spillovers are quite possible in any education program that
serves some students and not others in the same classroom. Thus ATE as we have defined it measures the
actual overall effect of the program, which is more relevant for policy decisions than this other definition.

12

that even if spillover effects exist, estimates of ATE include them since they measure the
average program impact over both participants and non-participants.
Our identification of average treatment effects rests on the assumption that after
controlling for school and year fixed effects, state time trends (separately for schools that
implemented Bolsa in 2001 and those that did so later), time trends based on initial
school enrollment categories, and (in some regressions) observable child and school
characteristics the presence of the Bolsa program in a given school in a given year is
unlikely to be correlated with unobserved factors that determine the education outcomes
examined in this paper (enrollment, dropping out and grade promotion). This approach
can be viewed as a natural experiment; Bolsa dramatically expanded in 2001, but it did
not reach all municipios at the same time, as seen in Tables 2 and 4,
Other evidence confirms that Bolsa was implemented at different times in different
municipios. Official records show that while most municipios implemented the program
in the last six months of 2001, 78 did not implement it until sometime in 2002. 5 More
importantly, this is the official start date; registering beneficiaries took many months.
Moreover, in many municipios it did not reach all communities at the same time.
For example, in a case study of five municipios in Minas Gerais state, Bastagli (2008)
reports that the number of eligible households receiving Bolsa benefits in 2002 ranged
from 32% to 60%, but by 2003 these rates all exceeded 100%. 6 In the 2001 school census
data, 60.2% of municipios had no schools report any students in Bolsa, and 8.6% had all
schools reporting a Bolsa student; thus in the other 31.2% of municipios some schools

We thank Frederico Finan for providing us the municipio data with the date of program implementation.
Rates above 100% imply that some ineligible children received benefits; poor households whose incomes
later increased often continued to participate in Bolsa due to delays in updating of households eligibility.

13

reported students who participate while others did not, which implies that the program
did not reach all communities at the same time in those municipios. 7
Further evidence on variation is in de Janvry et al. (2005), who found considerable
heterogeneity across municipios in how Bolsa was implemented, and serious confusion
regarding its implementation. One example of the confusion found in that study, which is
based on a survey of 261 municipios in four states in Brazils Northeast, is that in 63% of
those municipios officials mistakenly thought that beneficiary selection would be done at
the Federal level. The study also shows that schools played a crucial role in registering
(potential) beneficiaries: 94% of municipios report that schools made announcements
about the program, and 85% report that registration could be done at schools. Turning to
within-municipio variation, 38% of municipios report that registration used geographic
targeting within the municipio. For these municipios, the geographic factors considered
included poor neighborhoods (62% of municipios), ease of access (42%), areas with
many schools (29%) and distance from government offices (13%).
Overall, the general confusion, combined with reliance on schools to announce the
program and register beneficiaries, budgets that were often insufficient (see footnote 2)
and in many cases geographic targeting within municipios, led to substantial variation in
when the Bolsa program became available in a given community associated with a
specific school. The next subsection explains in detail the estimation method we apply to
Brazils school census data.

An alternative explanation for municipios where some schools report having Bolsa students while others
report not having such students is that the program exists everywhere in such municipios, but in very small
schools by chance there are no participating students. Yet consider a school with as few as 20 students. If
in the community 25% of students participate in Bolsa, the probability that none of 20 randomly selected
students participate in Bolsa is 0.3%. After excluding schools with fewer than 20 students one still finds in
2001 that 30.2% of municipios have some schools with Bolsa students and others without Bolsa students.

14

B. Estimation. Let yist be an educational outcome for child i in school s at time t

for a specific set of grades. It is a function of child and household characteristics
(denoted by the vector cist), school and teacher variables (sst), and whether the Bolsa
program exists at time t in the community where school st is found (Bst): 8

yist = f(cist, sst, Bst)

(1)

A linear expansion of f( ) is a reasonable approximation given sufficient interaction terms:

yist = + cist + sst + Bst + (cistBst) + (sstBst) + ist

(1)

where is a constant and ist denotes idiosyncratic deviations from this approximation.
The cistBst interaction terms are important, for two reasons. First, the impact of
the program could vary by child characteristics. Second, not all children are eligible for
Bolsa, so ideally cist would have a binary variable indicating program participation. If so,
the corresponding to that variable would measure the (average) impact of participating
in Bolsa, and in (1) would measure spillover effects on non-participants (and on participants) in schools with students who participate. As explained below, if no participation
variable exists, does not have this interpretation, yet the interaction term cistBst still
captures heterogeneity in program impacts for different types of students.
The interaction term with school characteristics, sstBst, may also be important
since (perceived) school quality can affect parents decisions. Yet the sign of is unclear.
For example, for high quality schools the impact of Bolsa on enrolment may be higher (a
8

Community characteristics, such as child wage rates, job prospects for educated adults, and local interest
in education, could be added to equation (1). That is not done here to avoid notational clutter, and because
our data from Brazil include no community characteristics. However, it is not difficult to include such
variables in equation (1); they could be specified in the same way the school variables (sst) are specified.

15

better school, plus a transfer, persuades parents to keep a child in school) or lower (better
schools are already highly valued, so a transfer has little additional impact).
1. School level Analysis. For school level estimation, sum equation (1) over i:

yst = + cst + sst + Bst + (cstBst) + (sstBst) + st

N st

N st

N st

i =1

i =1

i =1

(2)

where yst = (1 / N st ) yist , c st = (1 / N st ) c ist , st = (1 / N st ) ist , and Nst is the number

of students in school s at time t. To ease interpretation, redefine the cst and sst variables as
deviations from their means; the only parameter affected by this normalization is .
Given this normalization of cst and sst, estimates the average treatment effect
(ATE) of the Bolsa program, and and measure how this effect varies by child and
school characteristics, respectively. To see why, consider (without loss of generality) the
case with no school variables, where pist is student is program participation and cist is
other student characteristics. With student level data, one could estimate:

yist = + ccist + Bst + (pistBst) + ccistBst + ist

(3)

(There is no pist term since pist = pistBst; students can participate only if Bolsa exists in
their community.) The (average) impact of participating in Bolsa is measured by , and
measures (the average of) any spillover effect on nonparticipants (and participants). To
interpret c, note that it is very likely that Bolsas direct effect, and any spillover effects,
vary over students; each element in c is a weighted average of both effects, with
different weights (different propensities to participate in Bolsa) for different types of
students. Next, sum (3) over all students in each school (and divide by Nst):

16

yst = + ccst + Bst + (pstBst) + ccstBst + st

(4)

where pst is the proportion of Bolsa students in school s at time t. In this school level
equation, still measures the impact of participation and still measures spillover effects.
Next, define p = pst, the participation rate conditional on program availability.
Bst =1

Let pstN be the deviation of pst from this mean: pstN = pst - p . Then (4) becomes:
yst = + ccst + Bst + [(pstN + p )Bst] + ccstBst + st

(4)

= + ccst + ( + p )Bst + (pstNBst) + ccstBst + st

Thus, when pst is replaced by its deviation from this mean, regressing yst on Bst, pstNBst
and cstBst yields a coefficient on Bst that equals + p ; this is the average treatment
effect (ATE), including any spillovers that affect both participants and non-participants.
Note that , the coefficient on pstNBst, still measures the impact on the treated (ATT).
Unfortunately, the school census data do not have the proportion of students who
participate in Bolsa (pst). Yet they do have variables that are correlated with participation,
such as race (black, mulatto and indigenous students are poorer than whites and so are
more likely to participate). Consider a linear function of student variables that determine
participation (there is no constant since all variables are normalized to have a zero mean):
pstN = pcst + stN

(5)

where stN is idiosyncratic deviations from linearity. Substituting this into (4) yields:
yst = + ccst + ( + p )Bst + pcstBst + ccstBst + stNBst + st
= + ccst + ( + p )Bst + (p + c)cstBst + stNBst + st
17

(4)

Equation (4) has two key characteristics. First, the coefficient on Bst, + p , is still the
ATE. Second, ATT, that is , is not identified; only (p + c) can be estimated, so one
cannot even assess whether = 0. Note also that since stN and st are simply idiosyncratic
deviations from linearity, they are uncorrelated with cst and Bst; one can also show that
lack of correlation between Bst and stN implies that stN is uncorrelated with stNBst.
Equation (2) simply adds schools variables to equation (4), so in (2) is the ATE
(it corresponds to + p in (4)). If data were available for all variables in cst and sst,
OLS would consistently estimate , and in equation (2). Yet many cst and sst variables
are unobserved; for example, cst may include child innate ability and parental tastes for
schooling, and sst may include teacher motivation. To see the implications for estimation,
modify (2) to separate observed and unobserved variables in cst and sst:

yst = + cst+*cst* + sst+*sst* + Bst + (cstBst)+*(cst*Bst) + (sstBst)+*(sst*Bst) + st (2)

= + cst + sst + Bst + (cstBst) + (sstBst) + [*cst* + *sst* + *(cst*Bst) + *(sst*Bst) + st]

Asterisks denote unobserved variables, and associated parameters, and variables without
asterisks now denote observed variables (their parameters have no asterisks). The second
line of (2) shows that, for estimation, all unobserved variables are in the error term.
Consistent estimation of equation (2) by OLS requires the term in brackets to be
uncorrelated with all observed variables, which is unlikely. First, the availability of Bolsa
(Bst) may reflect unobserved child (cst*) and school (sst*) characteristics. For example,
community leaders had to exert effort to implement Bolsa in their municipios, and those
leaders may affect unobserved school characteristics. Also, Bolsa was implemented

18

more quickly in communities with low education outcomes (see Table 3), outcomes that
reflect, in part, unobserved school and child characteristics.
Second, the Bolsa variable, Bst, increases over time (see Tables 2 and 4). Both
observed and unobserved child and school characteristics could also increase or decrease
over time, and the same holds for those variables interactions with Bst. Thus the composite error term in brackets in (2) could slowly increase or decrease over time, leading it to
be correlated with Bst (as well as with some elements of cst, sst, cstBst and sstBst).
To remove bias due to correlation of Bst (and other observed variables) with the
unobserved determinants of yst one could find instruments for the observed variables, but
no credible instruments are in our data. Thus we approximate all unobserved variables by
school and time fixed effects, plus state-specific and initial enrolment-specific time trends:
cst* = c,s + c,t + c,jt + c,st

(6)

sst* = s,s + s,t + s,jt + s,st

Equation (6) decomposes each variable in cst* and sst* into a (time invariant) school fixed
effect (), a (school invariant) time fixed effect (), time trends that vary over Brazils
states or over school enrolment levels in 1998 (jt, where j denotes state or a school size
category), and deviations from these fixed effects and trends (c,st and s,st). 9
Using equation (6), the terms in brackets (other than st) in equation (2) become:
*cst* + *sst* + *(cst*Bst) + *(sst*Bst)

(7)

= s + t + jt + st + *((c,s + c,t + c,jt + c,st)Bst) + *((s,s + s,t + s,jt + s,st)Bst)

= s + t + jt + st + s(B)Bst + t(B)Bst + j(B)tBst + st(B)Bst
9

Our estimates use two time trends for each state, one for schools where students began participating in
Bolsa in 2001 (early adopters) and one for schools where student participation began in 2002 or later.

19

where s = *c,s + *s,s (a school fixed effect), t = *c,t + *s,t (a time fixed effect),
j = (*c,j + *s,j), so the j terms are state-specific and initial enrollment-specific time
trends, and st = *c,st + *s,st, is an average of the deviation terms. In addition, s(B)
= *c,s + *s,s, t(B) = *c,t + *s,t, j(B) = *c,j + *s,j, and st(B) = *c,st +
*s,st. The term s(B) is an unobserved school fixed effect that turns on only when
students participate in the Bolsa program; the programs impact could vary through
interactions with (time invariant) unobserved child and school characteristics. The term
t(B) allows the time fixed effect to differ for schools with and without Bolsa students.
The j(B) term is for a time trend that is in effect only when the program is operating; it
allows the impact of Bolsa to change over time, at different rates in each state and initial
school size, due to changes in unobserved child and school characteristics that influence
that impact. Finally, st(B) is an average of the deviation terms that turn on when Bolsa
is present. Inserting (7) into (2) gives the equation estimated in this paper:

yst = cst + sst + Bst + (cstBst) + (sstBst) + s + t + jt + st

(2)

+ s(B)Bst + t(B)Bst + j(B)tBst + st(B)Bst + st

The intuition for how estimation of (2) produces consistent estimates of , and
is as follows. As explained above, the timing of Bolsas implementation across and
within municipios displays substantial variation due to a wide variety of factors. While
this suggests a natural experiment approach to estimation, much of this variation is
unlikely to be random. The appearance of the Bolsa program in school s at time t, Bst,
may be correlated with unobserved child and school characteristics, and with their
changes over time. Time invariant school characteristics, and school averages of child
characteristics, can be represented by school fixed effects. While some school and child
20

characteristics may change over time, those changes should be very gradual and so are
controlled for by adding time fixed effects plus state-specific and initial school sizespecific time trends. Thus once one conditions on these fixed effects and time trends, Bst
is unlikely to be correlated with the remaining random error terms, st, st and st(B).
One potential shortcoming of this approach is that the deviation terms in (3), which
are aggregated into st, could be correlated with observed school and child characteristics.
This would lead to biased estimates of and , but those parameters are of secondary
interest. More seriously, this implies that st(B) is correlated with the interactions of those
variables and the Bolsa variable, which would lead to biased estimates of and . This
implies that those estimates should be treated with caution, but estimation of (2) would
still produce consistent estimates of the ATE (). 10
A final estimation problem is that the school fixed effects interacted with the Bolsa
program, s(B)Bst, are, in effect, a set of school dummy variables that, when summed,
equal Bst, so , the ATE of Bolsa, is not identified. This also occurs for the year fixed
effects that interact with Bst. To resolve this, recall that cst* and sst* have a mean of zero,
so one can constrain the means of s(B) and t(B) to be zero. 11 This constraint is easy to
impose for the year fixed effects (2001 to 2005), but imposing it for the over 100,000
school fixed effects is extremely difficult. Failure to impose it is most likely to cause bias
in estimates of and . To see why, suppose there are no interaction effects in equations
(2) and (2), so = = * = * = 0. Then s(B)Bst + t(B)Bst + j(B)tBst + st(B)Bst drops
More formally, suppose that st = cst + sst + st and st(B) = cst + sst + st(B), where st and st(B)
are random noise. Then (2) would be yst = (+)cst + (+)sst + Bst + (+)(cstBst) + (+)(sstBst)
+ s + t + jt + s(B)Bst + t(B)Bst + j(B)tBst + st + st(B)Bst + st. All the terms are uncorrelated with
all of the observed variables. One cannot estimate , , or , but can still be estimated.
11
This constraint cannot be imposed if some schools never have students who participate in Bolsa, but in
fact almost 98% of the schools in our sample have Bolsa students for at least one year. A related constraint
is needed; the sum of the j(B)s is constrained to be zero to ensure that is the impact for an average school.
10

21

out of (7) and (2) and one needs to control only for s + t + jt + st, which can be
done using fixed effects and time trends. It is the estimation of interaction effects that
generates s(B)Bst, and the inability to constrain these effects to sum to zero could lead to
inconsistent estimates of observed interaction effects ( and ). This reinforces the point
of the previous paragraph that estimates of and should be treated with caution.
Finally, this approach can be made more flexible. First, state-level time trends
need not be linear; for example, is still identified if trends are quadratic. Second, the
full impact of the Bolsa program may not be felt in its first year. Enrollment, grade
promotion and dropping out in any year could also be affected by whether Bolsa operated
in previous years, since learning accumulates over time and because adding or losing
students in one year has implications for future educational outcomes. This can be
checked by adding lagged terms, denoted as Bs,t-1, Bs,t-2, etc., to equation (2).
2. Municipio Level Analysis. As explained above, because the school-level Bolsa
variable measures only the programs existence, the school data can estimate only the
average treatment effect (ATE). Fortunately, the Ministry of Social Development has
municipio level data on how many households participate in Bolsa; in principle, this
allows one to estimate ATE and the impact of Bolsa on the treated (ATT), and to
estimate spillovers as well. This is seen by adding school level variables to equation (4)
and aggregating (calculating a weighted average) over schools within each municipio to
obtain a municipio level equation:

ymt = + ccmt + smt + Bmt + (pmtBmt) + ccmtBmt + smtBmt + mt

22

(8)

where cmt, smt, pmt and mt are averages over students in municipio m at time t, 12 and Bmt is a
dummy variable indicating whether Bolsa exists in municipio m. 13
In equation (8), the ability to observe the proportion of households who participate
in Bolsa allows one to estimate both and ; estimates spillovers, estimates ATT, and
+ p is an estimate of ATE. As with the school level regressions, one should account
for the unobserved cmt and smt variables; the estimation method used for the school level
regressions, adding school and time fixed effects and time trends, can be used here.
A final point regarding the estimation of spillovers is that equation (8) assumes that
spillovers onto non-participants are the same regardless of the number of participants. If,
instead, spillovers are proportion, then Bmt in (8) is replaced by pmtBmt, but then and
are not identified; one can estimate only + , the ATT. Thus if spillover effects are proportional to participation, the municipio data cannot be used to estimate them, or even see
whether they exist; yet one can estimate the impact of the treatment on the treated (ATT).

VI. Results
This section presents school level and municipio level estimates of the impact of
Bolsa Escola/Familia on enrollment, grade promotion and dropout rates in Brazil. All
regressions have year fixed effects, school or municipio fixed effects, and state level time
trends. For school level regressions, each state has two time trends, one for schools
where Bolsa started in 2001 and another for schools where it began later. This distinction
is less clear for municipios (in early 2001, 31% of municipios had Bolsa students in some
12

In fact, cmtBmt and smtBmt only approximate the municipio sums of cistBist and sistBist; this should have
little effect on estimates of and . As explained above estimates of and must be interpreted cautiously.
13
More precisely, the municipio data are the number of households that participate in the program, so pmt is
the proportion of households that participate in the Bolsa program. This is highly correlated with, but not
exactly equal to, the proportion of students who participate. Since Bmt indicates whether Bolsa exists in
municipio m, it is set equal to one if the percentage of households who participate is greater than zero.

23

schools but not in others), so the municipio level regressions have only one time trend per
state. Finally, eight more national time trends are added, based on schools 1998 enrollment levels. This controls for Brazils policy of merging small schools to create larger
ones, which causes smaller schools enrollment to increase more rapidly.
A. School Level Regressions. Table 5 shows basic estimates of the impact of
Bolsa on (log) enrollment, dropping out and grade promotion for children in grades 1-4.
These are estimates of equation (2), excluding the interactions between the Bolsa variable
and child and school characteristics. The top panel is the simplest specification, excluding
even direct effects of child and school variables. The estimated impacts for all three
outcomes are highly significant, with the expected signs. Schools with Bolsa students
have 2.8 percent higher enrollment, a lower dropout rate (by 0.31 percentage points), and
a higher promotion rate (by 0.53 percentage points). 14 The estimated impacts are slightly
smaller, but still significant, if quadratic time trends are used (not shown in Table 5). 15
As explained in Section V, the coefficient on the Bolsa variable estimates the
ATE, i.e. the average effect over all students. Section V also pointed out that the average
effect on students who participate (ATT) equals ATE/p (p is the proportion who participate) if spillovers are small, and it suggested that they are likely to be small. About one
14
Since Bst is a binary variable indicating whether any students in a school participate in Bolsa, estimates of
the impact of Bolsa on school enrollment could be biased because it is possible that, in a given school, the
program is available but by chance no children participate in it. The larger the school, the less likely is this
possibility, which implies an upward bias in the estimated effect of the program on enrollment. Yet this bias
is unlikely to be large. Consider the smaller, grade 1-4 schools. For these schools, the 10th percentile of the
distribution of enrollment is quite small, only 17 students. Yet 30% of Brazils students age 6-15 participate
in Bolsa, so the probability that none of the 17 students in such a small school is a Bolsa student is only
0.2% (0.717). Going even lower, a 5th percentile school has only 12 students, yet this probability is still quite
small (1.4%). In fact, small schools are in remote areas that have high poverty rates and thus high participation in Bolsa, so these probabilities are likely even smaller. Even so, as a robustness check these impacts
were reestimated after dropping the smallest 5% and smallest 10% of schools. For dropping out and promotion, there was very little effect, but the impact on enrolment fell somewhat, to 0.0205 when the smallest 5%
of schools were dropped, and to 0.0185 when the smallest 10% were dropped. These smaller effects do not
necessarily imply bias; impacts may well be larger in smaller schools, which have more Bolsa students.
15
Estimates that include private schools are very similar (97% of Bolsa students are in public schools).

24

third of Brazils children participate in Bolsa, 16 so if no spillovers exist then the impacts
on participants (ATT) are about three times higher than these ATE estimates. Note also
that the estimated impact on enrollment is much larger than that of Schultz (2004) for
Mexicos Progresa program; this likely reflects Mexicos high enrolment rate (94%).
If other, unobserved changes occurred in schools near the time the Bolsa program
was implemented that affect these three education outcomes but are not fully captured by
the control variables in Table 5, these estimates are biased. To check this, consider the
first three years of data (1998 to 2000). If there are unobserved changes that are highly
correlated with the expansion of Bolsa, in some schools where Bolsa started in 2001
these unobserved changes occurred in 2000, in others they occurred in 2001 and in still
others in 2002. This implies that, using only the 1998 to 2000 data, regressing these
outcome variables on a placebo variable that is zero for all schools in 1998 and 1999
but equals one in 2000 for the schools with Bolsa students in 2001 (and otherwise equals
zero in 2000) would yield a significant impact of the placebo variable. This is done in the
second panel of Table 5; the Bolsa impacts are much smaller than those in the first panel
(an order of magnitude smaller for two of the three), and all are statistically insignificant.
This suggests that it was Bolsa itself, not some unobserved school variable correlated
with it, which caused these changes in enrolment, dropping out and grade promotion.
Another robustness check discards the 2001 data. Some school principals may not
have reported student participation in Bolsa in 2001 since it was the first year the school
census asked about it, and that years question did not use the word Bolsa. 17 (In later
years the school census specifically use the work Bolsa.) If the principals understood the
16

Brazil had 35 million children age 6-15 in 2004, of whom 11 million participated in Bolsa (2004 PNAD).
The school census asks principals which of 15 programs exist at their schools. In 2001, the box for Bolsa
was labeled minimum income program. From 2002 on, it was minimum income program/Bolsa Escola.

17

25

question better in 2002 and later years, dropping the 2001 data may yield more accurate
results. This is shown in the third panel of Table 5. The estimated impacts of Bolsa are
somewhat larger (in absolute value) than in the first panel. Perhaps school principals
made errors when filling out the 2001 census form (yet the direction of measurement
error bias is unclear since the Bolsa variable is binary and so cannot have classical
measurement error), another possibility is that Bolsa has cumulative effects; for schools
where Bolsa began in 2001, the Bolsa variable for 2002 reflects two years of operation of
the program, and if there are cumulative effects the estimated impacts will increase if the
2001 data are excluded. Indeed, Tables 7 and 10 below present clear evidence that Bolsa
does have cumulative effects for all three education outcomes.
Another possible problem is omitted variable bias. Despite using school and year
fixed effects, and state and initial school size time trends, participation in Bolsa may be
correlated with trends in school characteristics that directly affect education outcomes.
The last panel of Table 5 examines this. Adding seven school characteristics and the proportion of female pupils yields estimates of program impacts almost identical to those in
the first panel. Most of these variables are significant, with the expected signs. The main
exception is that a program to provide computers seems to worsen outcomes; while difficult to explain, note that few schools (from 3% to 9%, varying by year) have this program.
The average treatment effects (ATE) in Table 5 could vary over students, both
because the programs impact may vary over students and because the proportion of
students who participate may vary. Overall, one would expect the ATE to be larger for
disadvantaged children, since only poor families are eligible for Bolsa. This can be
checked by interacting the Bolsa variable with student characteristics. Unfortunately, the

26

school census data have only three student variables: sex, race and age. The first two give
the number of female, black, mulatto, East Asian and indigenous students (white is the
omitted category) in each school. The race data are available only for 2005, yet schools
racial composition is likely to change only very slowly. School fixed effects preclude
using race variables as regressors, but race can still be interacted with the Bolsa variable.
The age variable can be used to obtain the average age of students entering grade
1, an indirect measure of households economic status; households with few resources or
lower tastes for education tend to delay enrolling children in school. 18 This variable is
defined as the average age of grade 1 students minus the grade 1 repetition rate, which
yields average age when starting grade 1. Because Bolsa could affect this variable, all
interactions use its value in 1998, well before Bolsa began.
Table 6 presents estimates of equation (7) that include interactions between child
and school characteristics and the Bolsa variable. Overall, the school variable interactions
indicate that Bolsa has stronger effects on enrollment for better schools, amplifying
inequities in observed school quality indicators. For the child variables, Bolsa is more
effective at raising enrollment in schools with more girls. Higher delayed enrollment
in grade 1 reduces Bolsas impact on enrollment. This variable probably indicates low
income or other disadvantages (such as early childhood malnutrition or less educated
parents), so Bolsa is less effective at inducing such children to enroll in school. This
result may also reflect that children who delay enrollment are older and so have higher
opportunity costs of time in school, reducing Bolsas impact on enrollment.

18

In the 2004 PNAD, among poor families (<120 Reais per capita per month) 21.8% of grade 1 students
were 9 years old or older, while in non-poor families this figure was only 12.8%.

27

Brazils ethnic groups differ in many ways, including education outcomes. In the
2004 PNAD, the enrollment rate of white children age 6-15 is 97.3%. Rates for blacks,
mulattos and indigenous children are lower: 93.6%, 95.2% and 89.6%, respectively. The
rate for Asian children, 97.6%, is slightly above that of white children. Table 6 indicates
that Bolsa is more effective at raising the enrollment of blacks, mulattos and indigenous
children, relative to whites, so it appears to equalize enrollment by race. Surprisingly, it
also raises Asian student enrollment, relative to whites, even though their enrollment is
not below that of whites (although the rate for poor Asians, 93.5%, is below that of poor
whites, 96.0%). These impacts are large; while Bolsa raises white enrollment by about
2.6%, the increase for schools with all black students is about 13%. (About 10% of
students are black, so for all-black schools the black variable, measured as a deviation
from the mean, is 90, and the overall impact is 0.026 + 0.0011*90 = 0.125). Similarly,
the impacts on schools with all mulattos and all indigenous are about 4% and 15%,
respectively (mulattos are 50%, and indigenous students are 2%, of all students).
Finally, Bolsas impact on enrollment is smaller in relatively large schools (as
measured by enrollment in 1998). This is unsurprising since larger schools tend to be in
urban areas, where a larger percentage of children are already enrolled.
Turning to dropout rates, Bolsas impact varies little by school quality indicators.
For child characteristics, girls are less likely to drop out, but Bolsa is less effective at
reducing their dropping out, perhaps due to their already low rates. Yet it seems more
effective at keeping disadvantaged children (measured by age when starting grade 1) in
school even though, as seen above, it is less effective at inducing such children to enroll.
Bolsas negative impact on dropout rates is weaker for blacks but stronger for mulattos.

28

The 0.007 estimate for blacks implies that Bolsa slightly increases the dropout rate by 0.1
percentage points for schools with all black students (-0.524 + 0.007*90 = 0.104). Perhaps
increased enrollment of weaker students leads, in later years, to more dropping out.
Finally, consider grade promotion. Overall, Bolsas positive impact is weaker in
schools with better school quality indicators. Girls tend to outperform boys, as indicated
by their higher promotion rate, but Bolsas impact on their promotion is smaller than it is
for boys. Its positive impact on grade promotion is somewhat lower for disadvantaged
students (measured by age at enrollment), blacks, indigenous students, and Asians.
Again, Bolsas positive impact on these students enrollment may have brought weaker
students into school, who are then more likely to repeat.
The estimates thus far implicitly assume that Bolsas impact does not depend on
how long it has been in place. Yet impacts may accumulate as students are treated for
many years. Table 7 investigates cumulative effects by lagging the Bolsa variable up to
three years. For all three education outcomes, the impacts rise over time, peaking after 23 years. More specifically, the impact on enrollment is a 2.8% increase after one year, a
4.3% increase after two years, and 5.5% after three years. Assuming no spillover effects,
and recalling that about one third of students participate in Bolsa, these results indicate
that, after three years, Bolsa raises participants enrollment by 17%.
For dropping out, the Table 7 estimates indicate that Bolsa reduces dropping out
by 0.30 percentage points after one year and 0.54 points after two years. Assuming no
spillover effects, Bolsa reduces participants dropout rate by 1.6 percentage points.
Finally, Table 7 shows that Bolsa raises grade promotion by 0.5 percentage points after

29

one year and nearly 1.0 point after two years. If no spillovers exist, then over the long run
the program raises participants grade promotion rates by 3 percentage points.
The estimated impacts of Bolsa on grade 5-8 students are shown in Tables 8, 9
and 10. According to the 2004 PNAD, about one fourth of grade 5-8 students were 16 or
older, so Bolsas impacts on these students may be smaller than those for grade 1-4
students. Table 8 replicates the Table 5 regressions. The sample is only about a third as
large, since there are fewer (but larger) schools at this level of education, but it is still
very large: about 182,000 (nearly 23,000 schools over eight years). In the simplest specification (top panel), Bolsa raises enrollment by 3.2%; assuming no spillover effects, this
implies an (average) enrollment increase of 10% for participants. Bolsa also reduces
dropout rates, and raises grade promotion rates, by about 0.3 percentage points (average
over all students), thus by about 0.8 percentage points for participants. These enrolment
and dropout estimates are close to those for grades 1-4 (Table 5), but the grade promotion
impact is only half as large. This may reflect that the grade 5-8 promotion rate, at 90%,
is much higher than the grade 1-4 rate, 82%, leaving less room for improvement.
The second, third and fourth panels in Table 8 examine these results robustness.
The second panel uses the placebo Bolsa variable to check whether these estimated
impacts are due to something else. There is no evidence of bias in the first panel results,
though the dropout and grade promotion estimates are imprecise. 19 The third panel drops
the 2001 data. In two of three cases, the coefficients absolute values increase somewhat.
The exception is log enrollment; its estimate is a bit lower. In any case, the first panel

19

The coefficient estimate for enrollment in the top panel would still be highly significant in terms of the
standard error in the second panel, but this is not the case for the dropping out and promotion regressions.

30

results are not driven by the 2001 data. Finally, the last panel adds several school
variables to see if doing so affects Bolsas estimated impacts; again, they are unaffected.
Table 9 checks whether Bolsas impacts vary by school or student characteristics.
For enrollment, three school level interactions are significantly positive, and two are
significantly negative; in contrast to grades 1-4, for enrollment Bolsa does not amplify
existing inequalities in observed indicators of school quality. Yet for dropping out and
promotion it does seem to amplify some inequities.
Turning to student interactions, Bolsas positive impact on enrollment is stronger
for girls and for disadvantaged children (measured by delayed enrollment). Ethnic groups
with low enrollment (black, mulatto and indigenous) also have above average enrollment
gains. Finally, as in grades 1-4 Bolsas impact is smaller for larger schools, again perhaps
because urban schools already have high enrollment. For dropping out, Bolsas negative
impact is stronger for disadvantaged children but weaker for black, mulatto and
indigenous students, which occurred only for blacks in grades 1-4. Perhaps higher
enrollment for these groups in grades 1-4 causes more dropping out in grades 5-8. For
grade promotion, Bolsas impact does not vary by sex, but disadvantaged students seem
to benefit more. Most ethnic minorities (black, mulatto and indigenous) benefit less in
terms of their academic progress, which is consistent with a smaller program impact on
their dropping out and may again reflect weaker students.
As in grades 1-4, Bolsas full impact on grade 5-8 students may not occur in its
first year; this is examined in Table 10. As in grades 1-4, the enrollment impact rises over
time. Enrollment rises by 3.0% in the first year, but after three years it rises by 6.5%. The
impact on dropping out also increases over time, but these estimates are less precise. In

31

the first year that rate falls by about 0.3 percentage points, and after three years by 0.4 or
0.5 points. In contrast, promotion shows no cumulative effect, unlike the grade 1-4 results.
B. Municipio Level Regressions. As Section V explained, alternative estimates
can be obtained at the municipio level, and the municipio-level data on participation rates
offers the possibility of estimating spillovers. We averaged student and school variables
by municipio to create a municipio panel from 1998 to 2005. It was merged with Ministry
of Social Development data on the percentage of families receiving Bolsa transfers in each
municipio. When new municipios appear after 1998 when an existing municipio splits,
these municipios are recombined to maintain a balanced panel. Similarly, neighboring
municipios merged after 1998 into one municipio are combined for all years in the data.
Tables 11 and 12 show basic estimates of municipio regressions for grades 1-4
and 5-8, respectively. No results are shown that disaggregate the impact by school or
student characteristics; such estimates were seldom significant and displayed no clear
patterns. Lagged results are also excluded since they were insignificant or had irregular
patterns (perhaps due to high autocorrelation). In these regressions, the existence of
Bolsa variable measures spillover effects, and the percentage of families in Bolsa
variable measures the direct impact of the program on the participants.
In the top panel in Table 11, the simplest specification, a one percentage point
increase in participation in Bolsa increases grade 1-4 enrollment by 0.14 percent,
implying that enrollment among participating students rises by 14 percent. The spillover
effect is insignificant. This is close to, albeit somewhat larger than, the estimate derived
from Table 5 that the program impact on participants is 8.4% (2.8%3). 20 The top panel
of Table 11 also indicates that program participation reduces dropping out by 3.0
20

Adding the insignificant negative spillover yields an estimate 0.11 (0.14-0.03) closer to the Table 5 estimate.

32

percentage points and raises grade promotion by 3.3 points. These are also somewhat
higher than the estimates inferred from Table 5 (0.9 and 1.6 points, respectively).
The second, third and fourth panels of Table 11 present the same robustness
checks done in Table 5. Creating a placebo Bolsa variable for the year 2000 and
estimating its impact using 1998-2000 data yields insignificant, or marginally significant,
program impacts. The impact is much smaller for enrollment, and for all three education
outcomes the coefficient changes sign. Thus it seems unlikely that the first panel results
are due to an unobserved variable that is highly correlated with Bolsa. Dropping the year
2001 (third panel) has little effect on the estimated program impact, and the same holds
after adding several school variables (last panel).
For grades 5-8, all estimated impacts in Table 12 have the expected sign but only
two, those for enrollment and grade promotion, are significant. There is little evidence of
spillover effects. The enrollment estimate indicates that Bolsa raises participants enrollment by 6.4%. This is slightly below the 9.6% (3.23) effect inferred from the school
level results (Table 8). The insignificant estimate of Bolsas impact on the dropout rate
suggests that program participation reduces it by 0.7 percentage points, close to the 0.81
percentage point (0.273) effect inferred from Table 8. Finally, the sole significant effect
in Table 12, that participation in Bolsa raises grade promotion by 2.1 percentage points,
more than double the inferred impact of 0.84 (0.283) from Table 8. Overall, these
results, while imprecisely estimated, are fairly similar to those in Table 8.
Turning to other results in Table 12, the robustness checks in the third and fourth
panels reveal no problems with the estimates. This is also true for the enrollment estimate
in the second panel, which uses only the 1998-2000 data, but these data yield statistically

33

significant impacts for dropping out and grade promotion that have the opposite (and
counterintuitive) sign of the estimates in the first panel. It is unclear what generates these
results. In any case, there is no evidence that the estimated results in the first panel are
caused by an unobserved variable that is positively correlated with the program variable.

VII. Conclusion
Brazils Bolsa Escola/Familia program is the largest program in the world that
provides families monetary incentives to enroll their children in school. The impact of this
conditional cash transfer (CCT) program is difficult to estimate because, unlike other
Latin American CCT programs, it was not implemented as part of a randomized trial.
Fortunately, school census data can be used to estimate Bolsas impact given plausible
assumptions about the nature of unobserved determinants of education outcomes in Brazil.
Our school level estimates indicate that the Bolsa program increased enrollment
and grade promotion and reduced dropping out, and it appears to have equalized
enrollment by race. Including lagged effects, the program increased enrolment by about
5.5 percent in grades 1-4 and about 6.5 percent in grades 5-8, decreased dropout rates by
about 0.5 percentage points in grades 1-4 and about 0.4 percent in grades 5-8, and raised
grade promotion rates by about 0.9 percentage points in grades 1-4 and 0.3 percentage
points in grades 5-8. Assuming no spillovers onto non-participants, the impact on
participants, who constitute one third of Brazils children, is about three times as high.
Municipio level estimates are broadly similar, and they reveal no evidence of spillover
effects.
It is not particularly surprising that the Bolsa program has these effects. The real
issue for policymakers is whether its benefits exceed its costs. Simple estimates based on

34

the enrollment impacts suggest that this may not be the case. The long-run effect of Bolsa
appears to be to increase participants enrollment by about 18%. Assuming that this
leads to an 18% increase in years of schooling implies a gain of about 1.5 years for the
target population. Using the 2004 PNAD data, we estimate that an additional year of
schooling raises wages by 11%, which implies a 16-17% increase in wages among the
poorest third of the population; this amounts to an increase of about 1.5% in wages of the
whole population (see Hoffman, 2006), or perhaps about 0.8% of GDP. While this
compares favorably with the program cost of 0.4% of GDP, these costs are incurred today
while the benefits accrue over the next 40 years of the working life of beneficiaries.
Applying a 3% discount rate implies a net present value of about 0.4% of GDP, and
applying a 6% discount rate implies a net present value of 0.3% of GDP. Adding the
opportunity cost of time would reduce the benefits somewhat, as would accounting for
costs of additional teachers and school supplies. Moreover, an increase in the supply of
educated labor may reduce the returns to education. On the other hand, there may be
other benefits, such as improved health and reduced income inequality. Overall, it is
unclear that the benefits exceed the costs.
One reason why the costs may exceed the benefits is that the 18% increase in
school enrollment among participants implies that 82% of participants would have been
in school even without the program, so the 82% of the funds have no effect on
enrollment, although this income transfer could be viewed as a benefit solely on
distributional grounds. This raises the question: can Bolsa be targeted towards those
households who would not enroll their children in the absence of the program?
Answering this question is an important task for further research.

35

References
Attanasio, O., E. Fitzsimons and A. Gomez. 2005. The Impact of a conditional education
subsidy on school enrollment in Colombia. Institute for Fiscal Studies, London.
Barro, Robert. 1991. Economic Growth in a Cross-Section of Countries, Quarterly
Journal of Economics. 106(2):407-443.
Bastagli, Francesca. 2008. The Design, Implementation and Impact of Conditional Cash
Transfers Targeted on the Poor: An Analysis of Brazils Bolsa Famlia. Ph.D.
Thesis, London School of Economics.
Barros, R. Paes, M. Carvalho, S. Franco e R. Mendona. 2006. Uma Anlise das
Principais Causas da Queda Recente na Desigualdade de Renda Brasileira. Texto
para Discusso no. 1203. IPEA, Rio de Janeiro.
Behrman, Jere, and Nancy Birdsall. 1983. The Quality of Schooling: Quantity Alone is
Misleading. American Economic Review. 73(5):928-946.
Behrman, J., P. Sengupta and P. Todd. 2000. The Impact of Progresa on Achievement
Test Scores in the First Year. Intl. Food Policy Research Inst., Washington, DC.
Bourguignon, Franois, Francisco Ferreira, and Philippe Leite. 2003. Conditional Cash
Transfers, Schooling and Child Labor: Micro-Simulating Brazils Bolsa Escola
Program. World Bank Economic Review 17(2):229-254.
Bruns, B., A. Mingat and R. Rakotomalala. 2003. Achieving Universal Primary
Education by 2015: A chance for every child. The World Bank. Washington, DC.
Cardoso, E.; Souza, A.P. 2003. The impact of income transfers on child labor and school
attendance in Brazil. So Paulo: USP. http://www.econ.fea.usp.br/elianacardoso/
ECONBRAS/cardoso-souza.pdf .
Damon, Amy, and Paul Glewwe. 2007. Three Proposals to Improve Education in Latin
America and the Caribbean: Estimates of the Costs and Benefits of Each
Strategy. Department of Applied Economics, University of Minnesota.
de Janvry, A., F. Finan, E. Sadoulet, D. Nelson, K. Lindert, B. de la Brire, and P.
Lanjouw. 2005. Brazils Bolsa Escola Program: The Role of Local Governance
in Decentralized Implementation. Social Safety Nets Series. The World Bank.
de Janvry, Alain, Frederico Finan and Elisabeth Sadoulet. 2007. Local Governance and
Efficiency of Conditional Cash Transfer Programs: Bolsa Escola in Brazil. Dept.
of Agricultural and Resource Economics. University of California, Berkeley.
Dubois, P., A. de Janvry and E. Sadoulet. 2004. Effects on School Enrollment and
Performance of a Conditional Tranfers Program in Mxico. Department of
Agricultural & Resource Economics, University of California, Berkely.
36

Ferro, A., A. L. Kassouf. 2005. Avaliao do impacto dos programas Bolsa Escola sobre
o trabalho infantil no Brasil. Pesquisa e Planejamento Econmico. , v.35, p.417.
Ferro, A. R.; Kassouf, A.L. and D. Levison. 2007. The Impact of Conditional Cash
Transfer Program in Brazil on Household Work Decisions in Brazil. (mimeo).
Filmer, D. and N. Schady. 2008. Getting Girls into school: Evidence from a Scholarship
Program in Cambodia. Economic Development and Cultural Change.
56(3):581-617.
Gertler, P., H. Patrinos and M. R. Codina. 2007. Do Supply-Side-Oriented and DemandSide-Oriented Education Programs Generate Synergies? Evidence from Rural
Mexico. Paper presented at the Universidad Iberoamericana, Calidad de la
Educacion en America Latina y el Caribe. Mexico city.
Gertler, P. 2004. Do Conditional Cash Transfers Impove Child Health? Evidence from
Progresas Control Randomized Experiment. Health, Health Care and Economic
Development. Vol. 94, n. 2., p. 336.
Glewwe, Paul, Margaret Grosh, Hanan Jacoby, and Marlaine Lockheed. 1995. An
Eclectic Approach to Estimating the Determinants of Achievement in Jamaican
Primary Education. World Bank Economic Review 9 (2): 231-258.
Glewwe, Paul, and Pedro Olinto. 2004. Evaluating the Impact of Conditional Cash
Transfers on Schooling: An Experimental Analysis of Honduras PRAF
Program. Department of Applied Economics. University of Minnesota.
Hanushek, Eric, and Ludger Woessmann. 2008. The Role of Cognitive Skills in
Economic Development. Journal of Economic Literature 46(3): 607-668.
Harbison, Ralph, and Eric Hanushek. 1992. Educational Performance of the Poor:
Lessons from Rural Northeast Brazil. Oxford University Press.
Hoffman, Rodolfo. 2006. Queda na Desigualdade da Distribuio de Renda no Brasil, de
1995 a 2005, e delimitao dos relativamente ricos em 2005, in R. Paes de Barros,
M. Foguel and G. Ulyssea, eds., Desigualidade de Renda no Brasil: uma anlise
da queda reconte. IPEA: Rio de Janeiro.
Khandker, S., M. Pitt and N. Fuwa. 2003. Subsidy to promote girls secondary
education: The female stipend program in Bangladesh. Mimeo.
Krueger, Alan, and Mikael Lindahl. 2001. Education for Growth: Why and for Whom?
Journal of Economic Literature 39(4): 1101-1136.

37

Lindert, Kathy, Anja Linder, Jason Hobbs and Bndicte de la Brire. 2007. The Nuts
and Bolts of Brazils Bolsa Famlia Program: Implementing Conditional Cash
Transfers in a Decentralized Context. Social Protection Discussion Paper 0709.
The World Bank.
Maluccio, John, and Rafael Flores. 2004. Impact evaluation of a conditional cash
transfer program: The Nicaraguan Red de Proteccion Social. Discussion paper
n. 184. FCND, IFPRI, Washington, DC.
Mankiw, N. Gregory, David Romer and David Weil. 1992. A Contribution to the
Empirics of Economic Growth, Quarterly Journal of Economics. 107(2): 407-437.
Parker, Susan, Luis Rubalcava and Graciela Teruel. 2008. Evaluating Conditional
Schooling and Health Programs, in T. P. Schultz and J. Strauss, eds., Handbook of
Development Economics, Volume 4. Amsterdam: Elsevier.
PNAD. 2001. Pesquisa Nacional por Amostra de Domiclios (National Household
Survey). IBGE, Rio de Janeiro.
Sala-i-Martin, Xavier, Gernot Doppelhofer and Ronald Miller. 2004. Determinants of
Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE)
Approach, American Economic Review. 94(4): 813-835.
Schady, N. and M. Araujo. 2008. Cash Transfers, Conditions, and School Enrollment in
Ecuador. Economia. 8(2):43-70.
Schultz, T. P. 2004. School Subsidies for the Poor: Evaluating the Mexican Progresa
Poverty Program. Journal of Development Economics, 74, 199-250 p.
Schwartzman, S. 2006. Reduo da desigualdade, da pobreza e os programas de
transferncia de renda. IETS (Instituto de Estudos do Trabalho e Sociedade).
http://br.monografias.com/trabalhos/reducao-desigualdade-pobreza/reducaodesigualdade-pobreza.shtml
Skoufias, Emmanue, and Susan Parker. 2001. Conditional Cash Transfers and Their
Impact on Child Work and Schooling: Evidence from the Progresa Program in
Mexico. Economia. 2(1):45-86.
Soares, F., R. Ribas and R. Osorio. 2007. Evaluating the Impact of Brazils Bolsa Familia:
Cash Transfer Programmes in Comparative Perspective. International Poverty
Centre Evaluation Note, number 1, December.
World Bank. 2001. Brazil as Assessment of the Bolsa Escola Programs. Report n. 20208BR.

38

Table 1 Number of Schools in Brazils School Census from 1998 to 2005.

(2)

Schools with 1st to

4th and/or 5th to 8th
grade classes
(3)

School with panel

data from 1998 to
current year
(4)

267,532
266,645
261,988
264,735
256,986
253,405
248,257
248,103

187,514
183,475
181,532
177,808
172,529
169,096
143,262
136,114

187,514
174,153
166,251
157,081
148,209
141,716
116,285
107,243

School census
years

Total number of
schools

(1)
1998
1999
2000
2001
2002
2003
2004
2005

Table 2 Student and School Characteristics: Public Schools with grades 1-4, 1998-2005.
A. Student Characteristics
Grade
Number of
promotion
Drop out
students/
rate
rate
classroom
68.4
14.5
27.5
70.6
12.8
27.0
69.5
13.6
26.4
72.4
10.6
25.8
73.0
9.5
25.4
72.9
9.5
25.1
70.9
9.9
24.8
72.7
8.9
24.5

% of grade 4 students
with age-grade
distortion
56.6
53.2
50.4
47.3
43.6
40.4
37.5
37.0

Years
1998
1999
2000
2001
2002
2003
2004
2005

Total
enrollment
135.3
134.4
130.4
127.3
124.7
120.5
115.5
110.4

Years
1998
1999
2000
2001
2002
2003
2004
2005

B. Indicators of School Quality and Program Participation

Percent of
% of
% of
% of schools % of schools
teachers
schools
schools
with
with Bolsa
with
with
with
computer Escola/Familia
college
computers
printers
labs
program
8.4
8.7
8.6
1.4
9.3
11.4
10.8
1.9
9.8
13.6
12.9
2.6
11.0
16.1
15.3
3.2
23.5
12.8
18.8
17.9
4.3
84.7
16.7
21.7
20.5
5.5
88.3
20.7
21.8
21.6
12.1
90.3
26.2
23.8
21.6
8.0
90.8

39

% of
families
receiving
Bolsa
17.1
17.3
25.8
31.5
33.8

Table 3 School Characteristics in 2001: Public schools with and without Bolsa.
Schools with Grades 1-4
All
Schools
127.3
72.4
10.6
25.8
11.0
47.3
16.1
15.3
3.24

Total Enrollment
Grade Promotion
Dropping out rate
Number students per classroom
% of teachers with college
% students age-grade distortion
% of schools with computer
% of schools with printer
% of schools with computer lab

Schools with Grades 5-8

Bolsa Non-Bolsa
All
Schools Schools Schools
97.5
136.5
406.8
70.5
73.0
79.0
11.6
10.3
11.1
25.0
25.9
31.1
6.8
12.2
65.0
50.8
46.2
49.3
8.2
18.5
61.8
7.7
17.6
59.2
1.5
3.8
22.7

Bolsa Non-Bolsa
Schools Schools
334.0
418.0
78.6
79.1
12.7
10.8
29.4
31.3
51.7
67.0
55.0
48.4
46.6
64.1
44.3
61.5
15.7
23.8

The grade 1-4 sample size is between 20,548 and 20,569 for treatment schools and 66,779 and 66,843 for control
schools, except for students per classroom, which is 5,698 in treatment schools and 25,166 in control schools.
The grade 5-8 sample size varies from 3,032 to 3,035 for treatment schools and 19,732 to 19,747 for control
schools, except for students per classroom, which is 2,194 in treatment schools and 15,349 in control schools.

Table 4 Student and School Characteristics: Public Schools with grades 5-8, 1998-2005.

Years
1998
1999
2000
2001
2002
2003
2004
2005

Years
1998
1999
2000
2001
2002
2003
2004
2005

Total
enrollment
394.9
408.7
415.0
406.8
397.6
378.5
359.4
348.8

Percent of
teachers
with
college
62.4
62.9
63.8
65.0
67.9
71.1
75.6
80.3

A. Student Characteristics
Grade
Number of
promotion
Drop out
students/
rate
rate
classroom
78.2
78.2
77.3
79.0
78.7
78.4
76.7
77.4

13.6
12.7
13.3
11.1
10.3
9.9
10.2
9.1

% of grade 8
students with agegrade distortion

32.4
32.3
32.0
31.1
30.7
30.5
30.2
30.0

56.3
53.8
52.3
49.3
46.4
43.1
40.2
37.8

B. Indicators of School Quality

% of
% of
% of schools % of schools
schools
schools
with
with Bolsa
with
with
computer Escola/Familia
computers
printers
labs
program
-38.4
37.6
9.2
-48.5
46.3
16.7
-54.6
52.2
19.9
61.8
69.2
76.1
75.7
79.0

59.2
66.4
72.9
75.0
73.4

40

22.7
26.9
30.8
48.6
37.0

13.3
76.0
83.2
85.8
86.4

% of
families
receiving
Bolsa
17.1
17.2
25.8
31.3
33.7

Table 5 Program Impact on Enrollment, Dropping out and Promotion: Basic Results
(public schools with grades 1 to 4)
Variables
Basic Model (1998-2005)

Log enrollment
Coef.
S. E.

Dropping out
Coef.
S. E.

School with Bolsa Escola/Familia

.0282***

.0018 -.309 ***

Number of observations
F test

699,255
375.6 ***

698,229
350.5 ***

Promotion
Coef.
S. E.

.0582 .533 ***

.0779

698,229
182.3 ***

Basic Model (1998-2000 only)

School with Bolsa Escola/Familia
in 2001, assigned to year 2000

.00136

Number of observations
F test

262,220
71.9 ***

.0028 -.031

.136

261,845
22.3 ***

.265

.168

261,845
32.8 ***

Basic Model (dropping 2001))

School with Bolsa Escola/Familia

.0330 ***

Number of observations
F test

611,848
374.0 ***

.0025

-.582 ***

.0757

610,902
350.0 ***

.721 ***

.101

610,902
177.2 ***

Adding School/Child Variables

(1998-2005)
School with Bolsa Escola/Familia
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
Girl

.0274 ***
.0353 ***
.0491 ***
.0203 ***
.0001 ***
.0125 ***
.0042 **
-.0136 ***
.0014 ***

Number of observations
F test

699,255
353.8 ***

698,229
314.1 ***

698,229
165.3 ***

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

Control variables (all regressions)

Year fixed effects 98 05
Trend x enrollment level in 98 (08)
Trend x states x Bolsa in 2001 (27)
Trend x states x Bolsa in 2002+ (27)
School fixed effects (87,407)
Robust standard-errors.
*** significant at 1% level.
** significant at 5% level.
* significant at 10% level.

41

.0018
.0033
.0031
.0027
.0000
.0032
.0017
.0021
.0001

-.310 ***
-.094 *
-.017
.090 *
-.002 ***
-.316 ***
-.008
.174 ***
-.013 ***

.058
.056
.060
.053
.001
.108
.044
.045
.002

.530 ***
.604 ***
.269 ***
-.102
.0002
.208
-.121 **
-.403 ***
.031 ***

.0780
.0776
.0799
.0707
.0010
.1365
.0569
.0698
.0031

Table 6 - Program Impact on Enrollment, Dropping out and Promotion:

Adding Interaction Terms
(public schools with grades 1 to 4, 1998-2005)

Variables
School with Bolsa Program
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
Computer lab x Bolsa Program
Computer x Bolsa Program
Library x Bolsa Program
Teacher college x Bolsa Prog
Prog school TV x Bolsa Prog
Prog computer x Bolsa Prog
Girl
Girl x Bolsa Escola/Familia
Age-repetition 98 x Bolsa
Black 05 x Bolsa Escola/Fam.
Mulato 05 x Bolsa Escola/Fam.
Indigenous 05 x Bolsa Escola
Yellow 05 x Bolsa Escola/Fam.
(Enrollment 98/1000) x Bolsa

Log enrollment
Coef.
S. E.
0.0261 ***
0.0022
0.0194 ***
0.0045
0.0393 ***
0.0030
0.0114 ***
0.0025
-0.0001 ***
0.0000
0.0096 ***
0.0027
0.0036 *
0.0019
0.0002
0.0029
0.0184 ***
0.0049
0.0485 ***
0.0036
0.0226 ***
0.0029
0.0004 ***
0.0000
0.0002
0.0025
-0.0147 ***
0.0039
0.0004 ***
0.0001
0.0021 ***
0.0001
-0.0049 ***
0.0006
0.0011 ***
0.0001
0.0004 ***
0.0000
0.0015 ***
0.0001
0.0012 ***
0.0001
-0.155 ***
0.0069

Number of observations
F test

562,408
518.4 ***

Control variables:
Year fixed effects 98 05
Trend x enrollment in 98 (08)
Trend x states x Bolsa in 2001 (27)
Trend x states x Bolsa in 2002+ (27)
School fixed effects (87,407)

yes
yes
yes
yes
yes

Dropping out
Coef.
S. E.
-0.524 *** 0.078
-0.009
0.158
-0.204 *
0.107
-0.161 *
0.089
-0.005 *** 0.001
-0.177 *
0.097
-0.059
0.068
-0.194 *
0.102
-0.186
0.175
0.210
0.129
0.448 ***
0.104
0.003 *
0.001
0.132
0.088
0.669 ***
0.139
-0.022 *** 0.002
0.018 ***
0.003
-0.549 *** 0.023
0.007 ***
0.002
-0.005 *** 0.001
0.006 *
0.003
0.007 *
0.004
-1.801 *** 0.245

Promotion
Coef.
S. E.
0.292 ***
0.104
0.188
0.210
0.430 ***
0.142
0.274 **
0.118
0.005 ***
0.002
0.077
0.128
0.063
0.090
-0.122
0.135
0.769 ***
0.232
-0.277
0.171
-0.769 ***
0.138
-0.006 ***
0.002
-0.393 ***
0.116
-0.720 ***
0.185
0.039 ***
0.003
-0.012 ***
0.004
-0.169 ***
0.030
-0.017 ***
0.003
0.001
0.002
-0.025 ***
0.004
-0.023 ***
0.005
1.043 ***
0.325

561,789
227.3 ***

561,789
116.8 ***

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

Robust standard-errors.
*** significant at 1% level.
** significant at 5% level.
* significant at 10% level.
Notes: For all interaction terms, student and school characteristics are rescaled to have a mean of zero, so
that the coefficient on the Bolsa variable indicates the impact on average student in an average school.

42

Table 7 - Program Impact on Enrollment, Dropping out and Promotion:

Adding Program Lag Terms
(public schools with grades 1 to 4, 1998-2005)

Variables
School with Bolsa Program
School with Bolsa lagged 1 year
School with Bolsa lagged 2 year
School with Bolsa lagged 3 year
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
Girl

Log enrollment
Coef.
S. E.
0.0276 ***
0.0018
0.0153 ***
0.0019
0.0124 ***
0.0022
0.0012
0.0024
0.0348 ***
0.0033
0.0488 ***
0.0031
0.0202 ***
0.0026
0.00014 *** 0.00003
0.0128 ***
0.0032
0.0041 **
0.0017
-0.0131 ***
0.0021
0.0014 ***
0.00009

Dropping out
Coef.
S. E.
-0.299 *** 0.059
-0.245 *** 0.059
-0.023
0.065
0.046
0.077
-0.086
0.056
-0.015
0.060
0.091 *
0.053
-0.002 *** 0.001
-0.321 *** 0.108
-0.007
0.044
0.168 *** 0.045
-0.013 *** 0.002

Promotion
Coef.
S. E.
0.508 ***
0.079
0.440 ***
0.082
0.162 *
0.090
-0.196 *
0.105
0.586 ***
0.078
0.266 ***
0.080
-0.104
0.071
0.000
0.001
0.216
0.136
-0.125 **
0.057
-0.390 ***
0.070
0.031 ***
0.003

Number of observations
F test

699,255
340.4 ***

698,229
302.2 ***

698,229
159.1 ***

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

Control variables:
Trend x enrollment in 98 (08)
Year fixed effects 98 05
Trend x states x Bolsa in 2001 (27)
Trend x states x Bolsa in 2002+ (27)
School fixed effects (87,407)

yes
yes
yes
yes
yes

Robust standard-errors.
*** significant at 1% level.
** significant at 5% level.
* significant at 10% level.

43

Table 8 Program Impact on Enrollment, Dropping out and Promotion: Basic Results
(public schools with grades 5 to 8)
Variables
Basic Model (1998-2005)

Log enrollment
Coef.
S. E.

School with Bolsa Escola/Familia

.032***

Number of observations
F test

182,192
176.6 ***

Dropping out
Coef.
S. E.

.0031 -.273 ***

.075

182,007
119.1 ***

Promotion
Coef.
S. E.
.282 ***

.0925

182,007
60.9 ***

Basic Model (1998-2000 only)

School with Bolsa Escola/Familia
in 2001, assigned to year 2000

-.00004

Number of observations
F test

68,322
74.2 ***

.0061 -.157

.230

68,204
9.61 ***

-.054

.267

68,204
15.5 ***

Basic Model (dropping 2001)

School with Bolsa Escola/Familia

.0273***

.0038 -.436 ***

Number of observations
F test

129,129
155.3 ***

129,043
112.4 ***

.089

.427 ***

.111

129,043
53.2 ***

Adding School/Child Variables

(1998-2005)
School with Bolsa Escola/Familia
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
Girl

.0317 ***
.0096 ***
.0080 **
-.0079 **
.0000
-.0036
.0004
-.0144 ***
-.0014 ***

Number of observations
F test

182,191
158.5 ***

182,006
107.9 ***

182,006
56.8 ***

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

Control variables (all regressions)

Year fixed effects 98 05
Trend x enrollment level in 98 (09)
Trend x states x Bolsa in 2001 (27)
Trend x states x Bolsa in 2002+ (27)
School fixed effects (22,774)
Robust standard-errors.
*** significant at 1% level.
** significant at 5% level.
* significant at 10% level.

44

.0031
.0033
.0036
.0031
.0001
.0041
.0023
.0024
.0003

-.267 ***
-.176 **
-.060
-.181 ***
.001
-.172
-.105 *
-.122 **
-.053 ***

.075
.069
.082
.067
.001
.108
.057
.056
.006

.260 ***
.030
-.090
-.059
-.010 ***
-.341 ***
.278 ***
.169 **
.085 ***

.092
.089
.102
.084
.002
.129
.070
.071
.008

Table 9 - Program Impact on Enrollment, Dropping out and Promotion:

Adding Interaction Terms
(public schools with grades 5 to 8, 1998-2005)

Variables
School with Bolsa Program
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
Computer lab x Bolsa Program
Computer x Bolsa Program
Library x Bolsa Program
Teacher college x Bolsa Prog
Prog school TV x Bolsa Prog
Prog computer x Bolsa Prog
Girl
Girl x Bolsa Escola/Familia
Age-repetition 98 x Bolsa
Black 05 x Bolsa Escola/Fam.
Mulato 05 x Bolsa Escola/Fam.
Indigenous 05 x Bolsa Escola
Yellow 05 x Bolsa Escola/Fam.
(Enrollment 98/1000) x Bolsa
Number of observations
F test
Control variables:
Year fixed effects 98 05
Trend x enrollment in 98 (09)
Trend x states x Bolsa in 2001 (27)
Trend x states x Bolsa in 2002+ (27)
School fixed effects (22,774)

Log enrollment
Coef.
S. E.
0.0351 ***
0.0027
-0.0351 *** 0.0032
0.0042
0.0029
-0.0100 *** 0.0024
0.0003 ***
0.0000
-0.0109 *** 0.0033
0.0126 ***
0.0022
-0.0080 *** 0.0026
0.0788 ***
0.0039
0.0134 ***
0.0040
0.0103 ***
0.0031
-0.0005 *** 0.0000
-0.0301 *** 0.0030
0.0000
0.0038
-0.0038 *** 0.0002
0.0046 ***
0.0002
0.0216 ***
0.0015
0.0003 **
0.0001
0.0006 ***
0.0001
0.0008 ***
0.0002
0.0004 *
0.0002
-0.1352 *** 0.0053

Dropping out
Coef.
S. E.
-0.194 **
0.089
-0.067
0.091
0.045
0.086
-0.312 *** 0.070
0.008 *** 0.001
-0.128
0.097
0.016
0.063
-0.138 *
0.076
-0.137
0.111
-0.438 *** 0.116
0.272 *** 0.088
-0.019 *** 0.001
-0.195 **
0.086
-0.095
0.109
-0.062 *** 0.005
-0.005
0.007
-1.573 *** 0.044
0.011 *** 0.004
0.013 *** 0.002
0.040 *** 0.007
0.002
0.007
0.282 *
0.164

Promotion
Coef.
S. E.
0.364 ***
0.109
-0.202 *
0.112
-0.061
0.106
0.103
0.086
-0.013 *** 0.001
-0.368 ***
0.118
0.143 *
0.078
0.104
0.093
0.277 **
0.136
0.351 **
0.142
-0.317 *** 0.108
0.011 ***
0.002
0.157
0.106
0.188
0.133
0.101 ***
0.006
-0.001
0.008
1.093 ***
0.054
-0.030 *** 0.005
-0.025 *** 0.003
-0.053 *** 0.009
-0.015 *
0.009
-0.076
0.201

147,575
557.5 ***

147,484
120.4 ***

147,484
73.3 ***

Yes
Yes
Yes
Yes
Yes

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

Robust standard-errors.
*** significant at 1% level.
** significant at 5% level.
* significant at 10% level.

45

Table 10 - Program Impact on Enrollment, Dropping out and Promotion:

Adding Program Lag Terms
(public schools with grades 5 to 8, 1998-2005)

Variables
School with Bolsa Program
School with Bolsa lagged 1 year
School with Bolsa lagged 2 year
School with Bolsa lagged 3 year
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
Girl

Log enrollment
Coef.
S. E.
0.0297 *** 0.0029
0.0202 *** 0.0030
0.0155 *** 0.0036
-0.0029
0.0042
0.0087 *** 0.0033
0.0079 **
0.0036
-0.0078 ** 0.0031
0.0000
0.0001
-0.0037
0.0041
0.0002
0.0023
-0.0140 *** 0.0024
-0.0014 *** 0.0003

Dropping out
Coef.
S. E.
-0.264 *** 0.074
-0.111
0.072
-0.177 **
0.081
-0.076
0.097
-0.169 **
0.069
-0.059
0.082
-0.180 *** 0.067
0.001
0.001
-0.169
0.108
-0.103 *
0.057
-0.124 **
0.056
-0.053 *** 0.006

Promotion
Coef.
S. E.
0.259 ***
0.092
-0.046
0.093
-0.088
0.104
-0.068
0.126
0.034
0.089
-0.089
0.102
-0.058
0.084
-0.010 *** 0.002
-0.340 *** 0.129
0.278 ***
0.070
0.169 **
0.071
0.085 ***
0.008

Number of observations
F test

182,191
152.7 ***

182,006
104.0 ***

182,006
54.7 ***

Control variables:
Year fixed effects 98 05
Trend x enrollment in 98 (09)
Trend x states x Bolsa in 2001 (27)
Trend x states x Bolsa in 2002+ (27)
School fixed effects (22,774)

yes
yes
yes
yes
yes

yes
yes
yes
yes
yes

Robust standard-errors.
*** significant at 1% level.
** significant at 5% level.
* significant at 10% level.

46

yes
yes
yes
yes
yes

Table 11 Program Impact on Enrollment, Dropping out and Promotion: Basic Results
(municipio level regressions for public schools with grades 1 to 4)
Variables
Basic Model (1998-2005)

Log enrollment
Coef.
S. E.

Existence of Bolsa Escola/Familia

-.0298
% families w/ Bolsa Escola/Familia .00137***
Number of observations
F test

.0188
.00036

35,530
29.6 ***

Dropping out
Coef.
S. E.
-.5747
-.0302***

.5751
.0068

35,523
188.7 ***

Promotion
Coef.
S. E.
.1876
.0332 ***

.8865
.0092

35,523
66.4 ***

Basic Model (1998-2000 only)

Existence of Bolsa Escola/Familia
% families w/ Bolsa Escola/Familia
in 2001, assigned to year 2000
Number of observations
F test

-.0066

.0209

-.1913

.9194

-.7471

1.1994

-.00050

.00038

.0347**

.0166

-.0365 *

.0200

13,468
275.01 ***

13,463
62.0 ***

13,463
821.3 ***

Basic Model (dropping 2001)

Existence of Bolsa Escola/Familia
-.0237
% families w/ Bolsa Escola/Familia .00143***
Number of observations
F test

31,085
28.1 ***

.0463
.00038

-4.1529** 1.8706
.9212
-.0258*** .0071 .0245 ***
31,080
185.3 ***

2.0046
.0097

31,080
73.1 ***

Adding School/Child Variables

(1998-2005)
Existence of Bolsa Escola/Familia
% families w/ Bolsa Escola/Familia
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
GNP per capita / 1000
Girl
Number of observations
F test
Control variables (all regressions)
Year fixed effects 98 05
Trend x enrollment level in 98 (08)
Trend x states (27)
Municipio fixed effects (4,523)

-.0255
0.00136***
-0.01130
0.22366***
0.21441***
0.00030 *
0.00420
0.168 ***
-0.0124 *
0.00127**
0.00381***
35,530
36.5 ***

yes
yes
yes
yes

.0189
-.5674
.5718
.1735
0.00034 -0.0310*** 0.0069 0.0348***
0.01394 -0.2695 0.1956 1.342 ***
0.02111 0.1962 0.3184 -0.134
0.01553 0.1042 0.2178 -0.341
0.00017 -0.0025 0.0026 -0.0082**
0.01220 -0.5476** 0.2772 -0.3001
0.01133 0.0520 0.1609 -0.0321
0.00703 0.1810 * 0.1039 -0.338 *
0.00061 -0.0088 0.0092 0.0131
0.00084 0.0216 0.0148 -0.0075
35,523
288.1 ***

35,523
50.3 ***

yes
yes
yes
yes

yes
yes
yes
yes

.8835
0.0092
0.2759
0.4049
0.2838
0.0039
0.3606
0.2312
0.1941
0.0138
0.0196

Robust standard-errors. *** significant at 1% level. ** significant at 5% level. * significant at 10% level.

47

Table 12 Program Impact on Enrollment, Dropping out and Promotion: Basic Results
(municipio level regressions for public schools with grades 5 to 8)
Variables
Basic Model (1998-2005)
Existence of Bolsa Escola/Familia
% families w/ Bolsa Escola/Familia
Number of observations
F test

Log enrollment
Coef.
S. E.
.0404*
.00064*

.0211
.00038

35,654
1700.1***

Dropping out
Coef.
S. E.
.3008
-.0072

.6348
.0070

35,649
175.5 ***

Promotion
Coef.
S. E.
-.6266
.0210 **

.7247
.0087

35,649
155.2 ***

Basic Model (1998-2000 only)

Existence of Bolsa Escola/Familia
% families w/ Bolsa Escola/Familia
in 2001, assigned to year 2000
Number of observations
F test

.0098

.0243

1.0828

1.0325

-1.6433

1.1632

.00044

.00046

.0610 ***

.0187

-.0441 **

.0208

13,449
4435 ***

13,445
147.8***

13,445
449.9 ***

Basic Model (dropping 2001)

Existence of Bolsa Escola/Familia
% families w/ Bolsa Escola/Familia
Number of observations
F test

.0428
.00058

.0629
.00040

31,187
1456 ***

-.3886
-.0073

.9411
.0073

31,182
170.8 ***

.3166
.0143

1.4972
.0092

31,182
181.4 ***

Adding School/Child Variables

(1998-2005)
Existence of Bolsa Escola/Familia
% families w/ Bolsa Escola/Familia
Computer lab
Computer
Library
Teacher college
Program meal
Program school TV
Program computer
GNP per capita / 1000
Girl
Number of observations
F test
Control variables (all regressions)
Year fixed effects 98 05
Trend x enrollment level in 98 (08)
Trend x states (27)
Municipio fixed effects (4,523)

.0424**
0.00061*
-0.00289
0.0801***
0.0883***
0.00093***
-0.00448
0.0725***
-0.0171***
0.00159**
-0.00165**

.0207
0.00037
0.0094
0.0107
0.0088
0.00014
0.010
0.0076
0.0050
0.00066
0.00070

.3631
-0.0089
-0.411 **
0.0648
-0.2117
-0.0016
-0.0053
-0.0726
-0.706***
-0.0074
-0.0915***

.6228
0.0069
0.186
0.222
0.170
0.0030
0.266
0.156
0.123
0.012
0.013

-.6827
.7157
0.0227 ** 0.0086
0.2193
0.241
-0.4462
0.279
-0.1741
0.218
-0.0156*** 0.0037
-0.731 **
0.305
0.204
0.198
0.550 ***
0.159
0.0093
0.015
0.125 ***
0.016

35,654
754.1 ***

35,649
139.3 ***

35,649
94.6 ***

yes
yes
yes
yes

yes
yes
yes
yes

yes
yes
yes
yes

Robust standard-errors. *** significant at 1% level. ** significant at 5% level. * significant at 10% level.

48

Appendix
Table A.1: Variable Means for Cross-Sectional and Panel Data, Various Years

Enrolment
Promotion
Drop out
Stud/class
Age-grade
distortion
Teachers
with college
Computer
Printer
Comp. lab.
% w/ Bolsa
White
Black
Mulatto
Asian
Indigenous

1998
CrossPanel
section
135.3
120.5
68.4
70.4
14.5
14.1
27.5
27.4
56.6
54.6

2000
CrossPanel
section
130.4
120.3
69.5
71.0
13.6
13.4
26.4
26.4
50.4
48.5

2002
CrossPanel
section
124.7
123.6
73.0
74.4
13.6
13.4
25.4
25.6
43.6
41.8

2005
CrossPanel
section
110.4
109.9
72.7
72.3
8.9
9.4
24.5
24.6
37.0
37.7

8.4%

8.5%

9.8%

10.2%

12.8%

14.0%

26.2%

25.5%

8.7%
8.6%
1.4%
-

7.8%
7.7%
1.3%
-

13.6%
12.9%
2.6%
-

13.2%
12.4%
2.6%
-

18.8%
17.9%
4.3%
84.7%
-

19.4%
18.4%
4.4%
81.4%
-

23.8%
21.6%
8.0%
90.8%
36.5
9.6
50.6
1.6
1.7

23.2%
21.0%
7.9%
89.2%
32.5
10.1
53.2
1.6
2.6

49