Automatic Grade Promotion and Student Performance:
Evidence from Brazil
Martin Foureaux Koppensteiner
Department of Economics
University of Leicester
Telephone: +44 116 252 2467
Fax: +44 116 252 2908
This paper examines the effect of automatic grade promotion on
academic achievement in 1,993 public primary schools in Brazil. A
difference-in-differences approach that exploits variation over time and
across schools in the grade promotion regime allows the identification of
the treatment effect of automatic promotion. I find a negative and
significant effect of about 7% of a standard deviation on math test
scores. I provide evidence in support of the interpretation of the
estimates as disincentive effect of automatic promotion. The findings
contribute to the understanding of retention policies by focussing on the
ex-ante effect of repetition and are important for more complete cost-
benefit considerations of grade retention.
Keywords: Grade retention, automatic promotion, incentives to learn,
primary education, Brazil.
JEL numbers: I28, I21, O15, H52
Grade retention, the practice of holding back students in the same grade for an extra
year if they fail to achieve promotion requirements – either in the form of a
performance measure or in the form of minimum attendance – is used in many
developing and in some developed countries. It is particularly widespread and
pronounced in African and Latin American countries, where repetition rates are often
as high as 30% (UNESCO 2008).
Historically grade repetition had a prominent role
in Brazil and repetition rates in Brazilian primary schools reached 24% in first grade
and 14% in fourth grade in 2005.
Retaining students has important consequences both for the individual as well as for
schools. Overall, every repeater has the same effect on school resources as enrolling
an additional student at that grade and subsequent grades and either leads to
compromising per pupil school inputs e.g. through larger class size or to a pressure on
public finances through the additional demand for teachers, classrooms, desks and
Opponents of grade repetition contend that it negatively impacts the retained
individual by stigmatizing them and harming their self-esteem, by impairing
established peer relationships and generally alienating the individual from school,
which may in turn negatively affect academic achievement and increase the
probability of dropping-out of school (Holmes 1989). Furthermore, repeating grades
delays entrance of students into the labour market which poses substantial monetary
cost on individuals over the life-cycle. In contrast, proponents argue that repetition
can improve academic achievement by exposing low performing students to
additional teaching and by allowing them to catch up on the curriculum and the
content of teaching. This is particularly important if school absence for reasons such
as illness in a given school year is the reason for retention. Grade retention may also
40 out of 43 African countries for which data is available in 2006 use grade retention (and for which
average repetition rates exceed 4% in primary school) and 18 out of 23 Latin American and Caribbean
Data available at http://stats.uis.unesco.org/unesco/ReportFolders/ReportFolders.aspx. UNESCO
Institute for Statistics, Data Centre, January 2012.
A very rough estimate of the annual cost of repetition on public finances in Brazil using average
expenditure per pupil at primary schools in 2006 of $554 (in constant 2005 US$) and 18,661,000
students enrolled at primary school and an average repetition rate over all grades of 18.7% (not
accounting for loss of students due to drop-out etc.) amounts to approximately 1.9 billion US$ (all data
from UNESCO 2008).
help to make classes more homogeneous in achievement and therefore easier to teach
by improving the match between peers in the classroom (Manacorda 2012).
There is a small but growing literature on estimating the causal effect of retention on
subsequent educational outcomes (Gomes-Neto and Hanushek 1994, Eide and
Showalter 2001, Dong 2009, Jacob and Lefgren 2004 and 2009, Manacorda 2012 and
Glick and Sahn 2010). The results are mixed, with positive as well as negative
estimates of the effect of repetition on academic achievement and school drop-out,
and the results seem to depend critically on context and age of students.
Considering these mixed empirical findings on the effect on repeaters, the use of
public resources and the undesirable consequences for public finances, the persistence
of grade retention regimes in many countries is puzzling. This is particularly the case
for developing countries where repetition rates are often very high and pressure on
public resources is large. Furthermore, repetition increases the age variation in the
classroom and repeaters may also directly lead to negative externalities on their peer
students (Manski 1993, Lavy, Paserman and Schlosser 2012).
A possible explanation for the persistence of grade retention in many countries may
be based on the deterrence effect of grade retention.
Grade retention induces students
to exert effort as it potentially inflicts substantial costs of repetition on low
performers. The ex-ante threat of retention may therefore incentivize students to study
in order to avoid being retained. This incentive effect of grade retention may have an
important effect on mean student outcomes, as it is not restricted to repeaters only,
but may create incentives for a much wider range of students. While the empirical
literature on grade retention focuses on the ex-post effect on repeaters, there exists –
to the author’s knowledge – no research on the ex-ante effect of the promotion regime
on academic outcomes of a wider set of students. This analysis examines the effect of
removing the deterrence of retention rather than estimating the effect of repetition on
repeaters. Automatic grade promotion has been introduced in Brazil on a large scale
since the early 2000’s partly to accelerate progress towards meeting the Millennium
Development Goal of universal primary education and to reducing the cost of larger
student cohorts (UNESCO 2012). I exploit credible exogenous variation in the timing
Manacorda (2012) is the first to point out such a deterrence effect of retention in the literature. A
related argument of a deterrence effect is discussed by Angrist et al. (2002) in relation to school
vouchers and by Jacob (2005) in relation to high stakes testing in the US.
of the adoption of automatic promotion for identification in a difference-in-
differences (DiD) setting.
I find that the introduction of automatic promotion significantly reduces academic
achievement measured by math test scores of fourth graders by 6.7% of a standard
deviation. Quantile DiD results show that the strongest treatment effect can be found
for the lower part of the test score distribution with considerably smaller effects in the
tails of the distribution. This is consistent with an interpretation of the estimates as a
disincentive effect of automatic promotion and the paper provides additional evidence
in support of this interpretation. There is no evidence that the results are caused by
teacher or school responses to the introduction of automatic promotion. Teachers are
no more or less likely to assign and correct their students’ homework, and class size is
unaffected by the policy introduction. Because there is only limited information on
teaching practices available it is not possible to rule out completely the possibility of
unobserved systematic teacher responses to the policy. The timing of the policy
change limits the potential for changes in the student composition of the test cohorts
and I provide strong evidence that the socio-economic composition is unaffected by
the policy and unlikely biases the estimates. There is also no evidence that the
estimates are affected by systematic changes in student mobility across schools or by
strategic test taking behaviour.
The remainder of the paper is organized as follows. Section 2 provides information on
the school system in Brazil and in the state of Minas Gerais. Section 3 presents the
data. Section 4 describes the natural experiment and outlines the assignment of
schools to treatment. Section 5 introduces the empirical strategy. The results, their
interpretation and falsification exercises are presented in section 6, and section 7
2. THE SCHOOL SYSTEM IN BRAZIL AND MINAS GERAIS
Primary school is compulsory in Brazil for children between the ages of 7 to 14 and
consists of eight years of schooling (MEC 1996).
Public schooling is free at all ages
and enrolment in primary and secondary school is open to students of all ages.
The Brazilian educational system has undergone substantial changes during the last
two decades and has achieved considerable progress in expanding access to
The school entry age has recently been lowered to 6 years and primary school has been extended to 9
education. Starting from a primary school net enrolment rate of 85% in 1991, Brazil
achieves today almost universal primary school enrolment with a net rate of 95%
(UNESCO 2008). Primary school completion and youth literacy rates have improved
notably, but the country continues to suffer from high repetition and drop-out rates.
The national conditional cash transfer programme Bolsa Família, formerly Bolsa
Escola, which is a means-tested monthly cash transfer to poor households conditional
on school enrolment and regular attendance among other conditions, plays a
significant role for the rise in school enrolment and attendance of school age children
(de Janvry, Finan and Sadoulet 2006).
This analysis focuses on the state of Minas Gerais, the second most populous state in
Brazil with an estimated population of about 19 million (IBGE 2007). Minas Gerais
contributes 10% to the Brazilian GDP and is among the most developed states in
Brazil (OECD 2005). The education system of Minas Gerais is among the most
advanced and in national performance tests students regularly perform among the top
According to state legislation, the State Secretariat of Education (SEE) has extensive
authority to plan, direct, execute, control and evaluate all educational activities in
Minas Gerais. Based on the far-reaching decentralization of education in Brazil, the
SEE transfers authority to a large extent to Regional Authorities for Education
(Superintendências Regionais de Ensino: SREs) and directly to the municipalities.
SREs and municipalities therefore play a major role in the provision of schooling and
the implementation of educational policies.
Municipal schools account for more than
half (56%) of all primary schools and state schools, that are directly under the control
of the SEE, account for 22% of all schools. Besides the public provision of education
private schools play an important role and account for the remaining 22%.
The overall repetition rate in primary schools in Brazil in 2006 was 18.7% and the total drop-out rate
for primary school 19.5% (UNESCO 2008).
The conditionalities of Bolsa Família require a minimum school attendance of 85% and extend to the
fulfilment of basic health care requirements such as vaccinations of the children and pre and postnatal
medical consultations for pregnant women. Monthly per capita income in the household cannot exceed
R$120 (US$57 in 2006) to remain eligible for the transfer. See Lindert et al. 2007 for a comprehensive
description of the programme.
The installation of FUNDEF, a federal fund established in 1996 with the aim of redistributing state
and municipal resources back to (mainly) municipalities according to student numbers contributed to
the improvement of the control of municipalities over educational decisions. See de Mello & Hoppe
(2005) for an analysis of FUNDEF.
There are also 28 federal schools in Brazil which are under the direct control of the federal
government; the single federal school in Minas Gerais has not been included in the dataset.
3. DATA DESCRIPTION
This study uses data from two sources. Information on school characteristics comes
from the annual Brazilian school census that is conducted by the National Institute for
the Study and Research on Education (INEP) under the control of the Federal
Ministry of Education (MEC). The Brazilian school census compiles data annually
from all primary and secondary schools in Brazil. The exceptionally rich data
includes information on the location and administrative dependence of schools,
physical characteristics (quantity of premises and class rooms, equipment and
teaching material), the participation in national, state and municipal school
programmes, the number of teachers and administrative staff, average class-size,
detailed information on student flows (number of students in each grade by age,
repetition, drop-out and student transfer rates) among other information. Summary
statistics for the public schools used in this analysis are presented in panel A of table
A1 in the annex.
The school census also contains the information on the grade promotion regime
adopted in each school (grade retention versus automatic promotion), which is used to
establish treatment and control groups.
The second part of the data comes from the State System of the Evaluation of Public
Education (Sistema Mineiro de Avaliação da Educação Pública: SIMAVE), which
includes the programme for the evaluation of state primary and secondary schools
(Programa de Avaliação da Educação Básica: PROEB).
Results from the
programme are used for the evaluation and design of educational policies in the state;
the results are, however, not used by the schools to evaluate individual student
performance, for example for the grade promotion decision.
The main outcome variable is student achievement in state schools in Minas Gerais
measured by math test scores in 2003 and 2006. All classes and all students in fourth
grade of each school are examined and participation of schools and students is
compulsory. The cognitive test scores are standardized to a mean of 500 and a
standard deviation of 100. In total 246,959 students have been tested in 1,993 state
schools in Minas Gerais. I use the repeated cross-section of test score data from 2003
and 2006 for this analysis. The students in the dataset have, as generally students in
public schools, a deprived socioeconomic background. Almost half (45.6%) of the
Schools under the administration of the municipality or the federal government are not included in
families with children at state schools in Minas Gerais qualify for Bolsa Família and
can be considered poor. Information on sex, date of birth, racial background and on
the socio-economic family background is also available from an adjunct
questionnaire. Unfortunately, only the 2003 wave of the socio-economic
questionnaire contains information on parental education. Chart B of table A1
presents summary statistics on these variables.
4. THE GENERAL EDUCATION ACT OF 1996: THE CASE OF A QUASI-
4.1 Policy background
The General Education Act of 1996 (Lei de Diretrizes e Bases da Educação Nacional:
LDB) paved the way for the introduction of automatic promotion policies in Brazil.
Federal Law No 9.394/1996, which came into effect in 1998, regulates the
responsibility for education between the federal, state and municipal level and
facilitated federal and state programmes to control the grade promotion regime (Pino
and Koslinski 1999). Section 3 of Art. 32 §1&2 formally distinguishes two
alternatives for educational authorities to organize student progression: besides the
conventional annual grade promotion regime the option of automatic promotion was
introduced, a system in which students progress automatically to the next grade at the
end of the school year. Between these two extremes, a mixture of both regimes was
also permitted. In the mixed regime, schools define “learning cycles” that stretch over
several – most commonly three – school years. During the initial years of the cycle
students are promoted automatically. In the final year of a cycle students that fail to
meet the minimum requirements set in the curriculum are retained. The idea behind
learning cycles is to allow students an individual studying pace (Mainardes 2010). If
students fall behind their classmates they have a longer period to catch up on the
curriculum. This particularly aims at reducing the long-run impact of negative
temporary shocks, such as school days lost to sickness or adverse family events. For
mixed regime schools that have adopted automatic promotion in learning cycles,
grade retention is not entirely eliminated, but limited to the final year of the cycles.
The LDB furthermore sets fundamental criteria on how to organize promotion under
any one regime: In every school year a minimum attendance of 75% of all school
days must be fulfilled as a general requirement for promotion, so that grade retention
is still permitted if students fall below a 75% minimum attendance.
According to the legal framework of the LDB the decision on the promotion regime
and its exact specifications is taken on the state level. Automatic promotion was
introduced at an early stage by the states of São Paulo, Minas Gerais and Paraná, to
some extent in the state of Pernambuco and by the federal resolution SE No 4, 1/98 in
all federal schools in Brazil. A recent federal resolution disallowed retention for the
first three schools years in all schools in Brazil from 2011.
In the state of Minas Gerais the new regime has been established by state resolution
No. 8.086 in 1997. It stresses the autonomy of each public school in the decision
whether to continue with the annual repetition regime or to introduce automatic
promotion. In the year 2000 1,449 out of 1,993 state schools had established
automatic promotion with two initial three-year cycles. At the beginning of the school
year 2004 the remaining 544 state schools switched to automatic promotion.
4.2 Assignment to treatment
Schools that adopted automatic promotion at the beginning of the year 2004 make up
the treatment group and schools with automatic promotion (which have adopted
automatic promotion since the year 2000) the control group. I focus on two cohorts of
fourth graders, which I call the test cohorts 2003 and 2006 for which test scores are
available. Chart 1 presents an overview of these two cohorts and the change in the
organization of promotion for the control and treatment group.
When using this division into treatment and control group for comparison a sound
understanding of the assignment process that leads to this division is essential. In the
case of state schools in Minas Gerais the 46 regional authorities for education needed
to propose a plan of implementation of automatic promotion for the schools under
their administration. The decision for early adoption of the policy was made by each
SRE in agreement with the state secretariat. The second wave was initiated by the
SEE in an attempt to introduce automatic promotion universally in all schools. As the
adoption of the policy is not randomized across schools in an experimental setting,
treatment and control schools may not be balanced in the distribution of school and
student characteristics. Although the identification strategy used in this analysis does
not rely on the distribution of covariates being balanced, it is generally reassuring to
find school and mean student characteristics of treatment and control group to be very
similar. Table A1, chart A and B present descriptive statistics of treatment and
controls schools for 2003 and 2006. T-tests (and Chi-square for categorical variable)
for the equality of means between treatment and comparison group, accounting for
clustering on the SRE level, reveal only very few small but statistically significant
differences. As sample size is partly reflected in the t-statistics, it is more useful to
look at the normalized difference
between means by
treatment status as a scale-free measure of the balancing properties of the covariates
(Imbens and Wooldridge 2009). The normalized difference is small for all covariates
and never exceeds the absolute value 0.25,
suggesting that treatment and control
schools are indeed very similar in terms of their physical school characteristics. Even
more importantly, the normalized differences of mean student characteristics, which
may indicate compositional differences of the student populations, are all very small
and are far below the suggested rule-of-thumb value of |0.25| in both years. Apart
from mean age, which differs slightly as expected,
no other variable reveals any
considerable difference at the mean. The overlap in the covariate distributions can
also be examined by looking at the distribution of the propensity score for the
treatment and control group. Figure 1 shows the propensity score for the probability
of treatment for the treatment and control group revealing substantial overlap in the
multivariate distribution of covariates and a relatively similar pattern of the
distribution of the propensity score for the treatment and control group.
In addition, I estimate a linear probability model to determine whether there are
systematic differences between schools that have adopted automatic promotion at
different points in time and to learn what observable school characteristics – if any –
determine early adoption. The results are presented in table A6. The coefficients on
the set of school characteristics are generally small and only very few are statistically
significant. When including SRE controls even fewer variables show a significant
effect and it is difficult to establish any systematic pattern.
Given the similarity of treatment and control schools with respect to school
characteristics and the student composition, it is plausible to consider the assignment
of schools to treatment and control groups as conditionally random.
This is a rule of thumb suggested in Imbens & Wooldridge 2009 to check the unconfoundedness
assumption for the use of linear regression in estimating average treatment effects.
Mean age is expected to differ as part of the treatment, which will be clarified in section 6.3.
A formal test under the null for the equality of the distribution (Kolmogorov-Smirnov) of the
propensity score is nevertheless rejected.
5. EMPIRICAL STRATEGY
To estimate the treatment effect of the policy change I use a DiD estimator exploiting
the variation in treatment status of schools over time, identifying an average treatment
effect on individuals in schools assigned to treatment. The double difference approach
is capable of removing biases resulting from permanent latent differences between
treatment and control as well as biases resulting from common trends over time. The
estimation in a regression setup allows including additional regressors on the
individual and school level to improve precision and to test for the presence of
omitted-school specific trends, in particular related to potential changes in the student
composition. Identification requires that trends in student outcomes at treated and
control schools would not be systematically different in the absence of treatment.
Under this identifying assumption, I estimate the effect of the introduction of
automatic promotion on test scores of fourth graders by the following regression
where Yist is the test score for individual i in school s at time t, ds is a school dummy
which captures school-specific time invariant effects, dt is a time dummy which
captures the common time trend of control and treatment group, dst is the
time/treatment-status interaction term containing information on the treatment status
of schools, that varies over time. γ in equation (1) is the coefficient of interest and
reflects the average treatment effect of the introduction of automatic promotion on
test scores of fourth graders. Zit is a set of covariates controlling for individual
characteristics. Xst denotes a set of exogenous covariates for class and school
characteristics, including average socioeconomic characteristics of students, detailed
the participation in federal, state and municipal educational
teacher characteristics and other.
ε is a stochastic error term.
Specifically, the covariates include initial (first grade) enrolment, number of teachers at school,
number of total staff (besides teaching staff), dummy variables describing the type of the premises
used for the school, dummies for the availability and number of teaching material (e.g. overhead
projectors, personal computers, TV and video sets etc.), the availability of computer and science labs,
school kitchen, the quality of sanitary units, number of class rooms in- and outside the school and
dummies for whether the school provides all 8 years of primary education.
These programmes include National Minimum Income Programme, Free School Lunch programme,
the provision of public school transportation, TV escola (a national education TV programme), other
educational TV programmes, computer literacy programmes, and other state and municipal school
Although non-random assignment of schools to treatment may lead to a correlation
between assignment status and outcomes, this does not violate the common trend
assumption as long as any differences that lead to the adoption of the policy are
captured by the school-fixed effects. The common trend assumption may nevertheless
be violated if selection into treatment was based on pre-treatment trends in school
characteristics that differ between treatment and control. If, for example, schools with
high performing students and low repetition rates adopt automatic promotion test
scores and treatment status are correlated for reasons other than the treatment impact
of automatic promotion. Unfortunately I do not have pre-treatment test score data to
test directly for the common trend assumption. I nevertheless can investigate whether
selection into treatment is based on pre-treatment differences in repetition rates. Table
A2 reveals that pre-intervention repetition rates (from the 1997 school census before
automatic promotion was introduced at any school) were virtually identical across
treatment and control schools, so that there is little concern for self-selection of
schools into treatment based on high or low repetition rates. The table also reports
pre-treatment class size (averaged over grades 1-4) and pre-treatment student-teacher
ratio (averaged over grades 1-4). While there is a small difference in class size of
about one student, there is virtually no difference in the student teacher ratio and the
normalized difference for both variables is well below |.25|. Classroom capacity
constraints therefore were unlikely the driving factor behind the decision for early
As I have pointed out earlier, the first wave of the policy adoption was
initiated on the SRE level, which furthermore limits the potential for individual
schools to select into treatment based on trends in test scores. The second wave was
then determined by the decision of the CEE made for all remaining schools, so that
there is virtually no scope for selection on a pre-treatment trend basis.
As the treatment regressor varies at the school level and test scores of students in the
same school are likely correlated, for example because they share the same learning
environment and/or are from the same neighbourhood, conventional standard errors
may be misleading as they do not account for the grouped error structure. The robust
This will also allow accounting for eligibility specific effects (Ashenfelter 1978). This way the
above time invariant composition assumption can be relaxed to accommodate for the case where
treatment and control group are expected to differ in covariates that may affect the outcome variable.
Lam and Marteleto (2006) show that the demographic transition in Brazil in the 1990’s had a strong
impact on student cohort sizes and enrolment rates in Brazil, but this does not seem to be relevant in
the context of this study.
standard errors reported therefore allow for clustering on the school level (Donald and
6. ESTIMATION RESULTS
6.1 Main results
The basic idea of the DiD strategy can be illustrated by a simple 2-by-2 table. Table 1
shows the levels and differences in test scores between treatment and control groups
and the changes over time. The first row reports means before treatment (year=2003),
when control schools were already under the automatic promotion regime and the
treatment schools were still under the annual grade retention regime and the mean
difference for the two groups. The entries in the first column reveal that schools that
had already adopted automatic promotion have a mean score that is 7.05% of a
standard deviation lower than schools that had not yet adopted the new regime in
2003. After the adoption of automatic promotion by schools of the treatment group
this difference almost completely disappears and students at both groups have very
similar average test scores and the difference in means is not statistically significant.
Likewise, schools in the control group have very similar mean test scores over time
with a difference that is not significantly different from zero. The lower right entry
reports the simple DiD estimate, which can be interpreted as the causal effect of
treatment under the above identifying assumptions. The adoption of automatic
promotion leads to a decrease in test scores of 6.65% of a standard deviation. Almost
the entire fraction of the DiD outcome originates from the pre-treatment difference
between control and treatment schools. After the adoption of automatic promotion in
treatment schools the difference between treated and control schools almost
This simple double difference can be amended in a regression framework following
equation (1) to improve precision of the estimates and to be able to control for
covariates and check the sensitivity of the estimates to their inclusion. Table 2
presents the estimates for different sets of controls. All specifications include school
fixed effects and year dummies. School fixed effects capture stable unobserved
characteristics of the schools and year dummies control for common trends in the test
scores that are not related to treatment. Specification (1) of table 2 includes school
controls, specification (2) controls for school and peer characteristics and
specification (3) also includes individual level covariates. The estimates in all
specifications reveal a stable negative effect of around 6% of a standard deviation and
are very precisely estimated (1% level of significance). Adding school level and peer
controls reduces the negative effect, but the reduction is relatively small. Controlling
additionally for individual characteristics delivers estimates of virtually the same size
as the simple double difference in table 1. The results reveal that the regime change
from annual grade retention to automatic promotion has a significant negative impact
on educational achievement on fourth graders in state schools in Minas Gerais. In the
next section I will discuss the interpretation of the results.
6.2 Interpretation of the results and the disincentive of automatic promotion
Table 4, column 1 reports the DiD estimates of the treatment on repetition rates for
grades 1-4, following the tested cohorts of 2003 and 2006 over grades 1-4. The
bottom entry for column 1 shows how the introduction of automatic promotion
reduces the repetition rate in fourth grade by 0.086. Prior to the policy change, about
10% of all students in treatment schools repeated fourth grade, but only about 2% did
so after the introduction of automatic promotion.
In this analysis I am interested in
understanding whether the estimated effect on test scores can be explained by the
elimination of the threat of retention for fourth grade students.
The two cohorts of
students at treatment schools face indeed very different incentives from the grade
promotion regime; while the 2003 cohort is subject to grade retention, the 2006
cohort does not face the threat of being retained.
If the estimated effect is caused by a change in study incentives to avoid being
retained, one would expect heterogeneous treatment effects along the test score
distribution. Students in the lower tail of the distribution should be more heavily
impacted by removing this incentive when compared to students in the upper tail, as
these students should be less concerned about the possibility of retention. For that
purpose I estimate equation (1) applying DiD to each quantile instead of the mean
under analogue assumptions to the standard DiD (Koenker 2004, Athey and Imbens
Repetition rates stay above zero because repetition is still possible when failing to achieve 75%
minimum school attendance.
As automatic promotion was introduced already at second grade, part of the estimated negative
effect in fourth grade test scores could be due to the disincentive students faced already in second
grade. Because of the repetition regime in third grade the effect is likely to be very small compared to
the contemporaneous effect in fourth grade.
2006, Firpo et al. 2009).
Table 4 provides the quantile DiD estimates and reveals
substantial differences in the treatment effect across the nine quantiles. The estimates
range between -9.01 and -3.92 and are more pronounced in the lower half of the
distribution, with the strongest effects centred on the fourth quantile. The effect of
automatic promotion is much smaller for the top two quantiles and not statistically
significant, yet still negative and non-negligible in magnitude. The inverse u-shaped
distribution of effects is consistent with the interpretation of the estimates as
disincentive effect of automatic grade promotion, such that the treatment effect is
largest for students left of the centre of the distribution close to the assumed grade
promotion threshold and smaller for high performing students that are unlikely to be
retained. Similarly, for students at the very bottom of the distribution the effects are
somewhat smaller with a coefficient of -7.49 but still above the mean treatment
effect. The slightly smaller effect at the bottom of the distribution could be explained
by either a different perception of the cost associated with retention or the fact that
grade retention is a possibility for these low performing students regardless of their
There is some suggestive evidence that automatic promotion indeed directly
impacts the behaviour of students and reduced their study effort. Column (1) of table
A3 in annex reports the effect of the policy introduction on the propensity of students
doing their homework.
The DiD estimates show that after the introduction of
automatic promotion fewer of the children do their homework (a decrease of 0.014);
the coefficient is only marginally significant though. Interestingly, the change in the
retention regime also changes the parents’ involvement with their children’s
homework. Column (2) of the table shows that parents are more likely to help with
their children’s homework (an increase of 0.022). This reveals that parents may well
be aware of the disincentive from automatic promotion and they may try to counteract
the potential reduction in their children’s study effort. If anything, increased parental
involvement would however bias the estimates in table 2 towards zero, rather than
explain the estimated effect.
The distribution of treatment effects in table 4 is also consistent with an explanation
based on changes in teacher incentives from automatic promotion. Teachers may for
Recent applications of quantile panel methods include Havnes and Mogstad 2010, Gamper-
Rabindran, Khan and Timmins 2010 and Lamarche 2011.
Separate estimates by socio-economic status as proxied by the number of books in the household do
not reveal heterogeneous effects along that margin (results not reported).
All of the outcome variables in table A3 are based on pupil reported behavioural responses of
themselves, their parents and their teachers and should therefore be considered more cautiously.
example focus less on students in the bottom half of the distribution if they previously
cared about them being promoted.
Information on whether teachers assign and
correct homework may shed some light on potential teacher responses to automatic
promotion. Columns (3) and (4) of table A3 report the DiD estimates for teachers
assigning and correcting homework respectively. Both coefficients are very close to
zero and not statistically significant so that there is no evidence that teachers respond
systematically to automatic grade promotion.
6.3 Changes in the student composition
For the interpretation of the estimates as disincentive effect, any possible channel of
effect of automatic promotion on outcomes – other than the disincentive effect – has
to be precluded. Most importantly, potential changes in the composition of students in
treatment and control schools over time could systematically lower test scores rather
than the changes in incentives reducing the effort of students. Because there is grade
retention in control and treatment schools in both periods at the end of third grade this
leads to a positive selection of the students entering into fourth grade in both
treatment and control schools and this mechanically limits the potential for changes in
the student composition. In the section 6.4 I will discuss the implications of this in
Table 6 reports DiD estimates for a range of mean socioeconomic variables; for each
outcome variable I have fitted a separate regression including school fixed effects and
year dummies. Only the coefficients on the mean number of fridges per household
and on mean age are statistically significant. All other indicators of the socio-
economic composition are not affected by the introduction of automatic promotion,
which is very reassuring. While the coefficient for the mean number of fridges per
household is very small and may be due to some spurious correlation, the significant
reduction in mean age by about one month is more relevant and it is important to
understand the source of this reduction in age and its consequence for the
interpretation of the result.
This reduction in age is caused by the difference in the inflow of repeaters in fourth
grade at the treatment schools before and after treatment. Whereas treatment schools
Teachers may equally worry about the lost incentive for students and target their effort on the most
affected students, so that a potential teacher response may go either way.
Section 6.8 looks separately at class size as another teaching input.
still received an inflow of repeaters from fourth grade of the previous year at the
beginning of the year 2003, there was no such inflow of repeaters in 2006, which
leads to the reduction in mean age, as repeaters are on average one year older. Table
8, column 2 shows the DiD estimate of the policy change on the net inflow of
students from first to fourth grade and from first to third grade in column (1).
Whereas the coefficient in column (1) is very small, negative and not statistically
significant, the coefficient for the net inflow of students including the inflow of
repeaters from the previous year at the beginning of fourth grade is sizeable, positive
and very precisely estimated (column (2)). Looking at the direct effect of the inflow
of repeaters on mean age of the cohort reveals that this almost exactly explains the
age effect estimated in table 6.
This means that the composition is altered due to
treatment and it is important to understand the potential bias of the compositional
change on mean achievement.
Even assuming a positive effect of repetition on educational outcomes of repeaters,
it is very plausible to assume that average performance of repeaters is still below the
mean performance of non-repeaters in the test cohort, as repeaters are selected as the
lowest performers in fourth grade in the preceding year.
How does this differential
inflow affect the outcome variable of interest? As there was an inflow of such low
performing students in 2003, but not in 2006 the results for the disincentive effect of
automatic promotion are, if anything, biased towards zero and the reported
coefficients in table 2 (chart A) need to be regarded as a lower bound of the true
effect. Unfortunately, there is no direct information in the student questionnaire on
whether and when students were retained. I can nevertheless use individual age to
single out repeaters to some extent. A regression sensitivity analysis that includes
individual age as a control variable may give an idea about the size of the bias from
the differential inflow of repeaters. Adding individual age to specification (1) leads to
an increase in the negative effect of about 20% to -7.97% of a standard deviation
compared to -6.65 % without controlling for age, reported in chart B of table 2.
Controlling for individual age in specification (2) and (3) leads to a very similar
increase of 20% of the effect to -7.33% and -6.77%, respectively.
Assuming that they are about one year older the inflow of repeaters at fourth grade leads to a
decrease of mean age of the cohort of 36 days compared to the estimated effect on mean age of 39
And a direct effect related to age, as repeaters are one year due to repeating the grade.
This is confirmed by the findings elsewhere; see Manacorda (2012) for example.
An alternative way of investigating the importance of the bias for all specifications is
to restrict the estimations to students that have never repeated by excluding all
students outside the target age range of fourth graders. Once students from the
additional inflow at fourth grade from the sample are removed, this leaves a sample
of students that have never repeated.
Chart A of table 5 reports the results for the
same specifications as in table 2, but restricts the sample to students in the target age
range for fourth graders. By restricting the sample in this way the coefficients exceed
the estimates of the original full sample in all specifications by around 30%. The
estimated effect is a further 11-16% larger compared to the estimates in chart B of
table 2. Restricting the sample to repeaters (chart B, table 5) reveals a negative effect
that is considerably smaller and no longer statistically significant. The number of
excluded students is nevertheless larger than what could be explained by excluding
fourth grade repeaters only, as removing overage students from the sample also
removes students that have repeated at third grade. As repetition is equally possible in
all schools at third grade, the additional increase in the estimates is therefore not
necessarily related to treatment. The increase rather suggests that the incentive of
grade retention may have a different impact on previous repeaters compared to
students that have never repeated a grade. The cost of repetition is likely highest for
students that have not previously repeated. In contrast the cost of being retained again
is smaller for previous repeaters, as they may already have suffered stigmatization
and have already been separated from their original peer group. The difference in
results for the restricted sample therefore may not only reflect the correction for the
differential inflow of repeaters at fourth grade, but may also more generally reflect
heterogeneous effects on repeaters and non-repeaters.
A more comprehensive analysis of the sensitivity of the estimates to the inclusion of
age controls is provided in table A4 in the annex. I present different specifications of
equation (1) with and without controlling for individual age for the full sample (chart
A) and the age restricted sample in chart B. The results support the previous findings.
Adding individual age as control (columns 4, 6 and 8) strengthens the negative effect
in the full and restricted sample for all the different specifications.
Nevertheless I cannot distinguish repeaters from students that have enrolled late at first grade. With
rather strict enforcement of the enrolment age in Minas Gerais and the incentives to parents to enrol
their children based on Bolsa Família conditions, late enrolment is nevertheless rather limited.
Besides the direct effect on the composition, there may be an indirect effect on
students of having repeaters in the class room. Repeaters may impose a negative
externality on their peers because their achievement is lower or because they may be
more disruptive in class (Lazear 2001). Lavy, Paserman and Schlosser (2012)
elaborate on the extent of ability peer effects associated with repeaters and show that
academic performance and behaviour of repeaters may be responsible for the negative
effect. By adding peer age in the DiD specification I can control for potential peer
effects from the differential inflow at fourth grade. Adding peer age as control only
moderately increases in size the coefficients in specification 1 and 2 in chart C of
table 2. Columns 2 and 5 in charts A and B of table A4 reveal that the inclusion of
peer age only has a minor effect when controlling for other peer variables and does
not strengthen the estimates of the treatment effect suggesting that there is no
noteworthy bias on the estimates. If anything, a negative peer effect of repeaters, as
suggested in the literature, would lead to an underestimation of the effect.
Conditioning on individual age and restricting the sample to non-repeaters reveals
that the differential inflow at fourth grade changes the composition of students in a
way that underestimates the true effect of automatic promotion by not taking into
account the net inflow of repeaters into fourth grade in 2003. The size of the
downward bias ranges between 20% and 30%. The estimates for the restricted sample
should nevertheless be considered with caution, as the disincentive of automatic
promotion may have differential impact on previous repeaters and non-repeaters.
6.4 Introduction of automatic promotion at second grade
Because of the introduction of automatic promotion in treatment schools at second
grade of the cohort of interest, this potentially may also have an impact on the
composition of students. Table 4 reveals how repetition rates from first to fourth
grade of the theoretic test cohorts are affected by the policy introduction. The
estimates for first and third grade show no effect of the treatment as expected. Rates
at first grade are unaffected with the policy change occurring only in the subsequent
year and rates in third grade are unaffected as the final year of the cycle remains with
grade retention for both cohorts in treatment and control group. The estimate for the
impact on the second grade reveals how the policy introduction lowers the repetition
rate by almost 12% at second grade in 2004. The potential threat to the interpretation
of the results arises from the fact that by introducing automatic promotion at second
grade for the 2006 exam cohort, this cohort may be “contaminated” by low
performers that would have been removed in the absence of treatment. The mean
repetition rate for second grade at treatment schools drops from 12.8% (2003 exam
cohort) to 3.1% (treatment cohort). Rather than looking at second grade repetition
only, in- and outflows in each grade up to the end of third grade have to be considered
when examining the effect of the adoption of automatic promotion on the student
composition. Looking at overall student flows of the test cohorts reveals that the
negative selection has largely cancelled out when the test cohort enters fourth grade
in 2006. In particular, repetition in third grade plays an important role here. The first
column of table 8 reports the effect of the policy introduction on the net flow taking
into account in- and outflows over grades 1-3. The net inflow due to the introduction
of automatic promotion is very close to zero and not statistically significant. This is
mainly based on two factors: Focussing only on treatment schools, table A5 shows
that repetition rates at third grade actually increased by about 4.3% for the 2006 exam
cohort, which filters out a substantial fraction of the low-performers already.
Furthermore, third grade repetition rates for the two cohorts have to be compared with
caution, as these may have a different impact on removing low-performers from the
previous year depending on the inflow of students into third grade at the beginning of
the year. Considering net-flows, the composition of 2003 and 2006 cohorts are
practically unaffected at the beginning of fourth grade. As mentioned earlier, the
socioeconomic composition between the cohorts (table 6) is virtually unaltered by
treatment, which supports the premise, that the policy introduction does not change
the composition of students up to fourth grade.
This is also corroborated by the fact that almost the entire fraction of the DiD result
arises from the ex-ante difference between treatment and control group in 2003,
rather than from the difference after treatment. The results for the simple difference
over time of the control schools and the difference between control and treatment
schools after treatment in 2006 are very small and not significant at conventional
As I have pointed out before, introduction of automatic promotion in second grade
may also have a separate disincentive effect already at second grade. This separate
disincentive effect is likely to be small, because of grade retention in the subsequent
grade and general decay of the effect over time. The estimated effect at fourth grade
can nevertheless be considered the composite effect of the disincentives over the two
6.5 Effect of the policy change on drop-out rates
The effect of retention on student drop-out has been studied elsewhere in the
literature (see Jacob and Lefgren 2004, 2009, Manacorda 2012). If the introduction of
automatic promotion has an effect on drop-out rates in grades prior to fourth grade,
this may change unobserved student characteristics that cannot be controlled for. I
estimate the effect of the introduction of automatic promotion on drop-out rates in a
DiD specification similar to equation (1) as (2),
using aggregated data from the school census. Column (2) of table 4 reports the
coefficients for each grade. Drop-out rates in second grade are unaffected by the
The treatment nevertheless has a small effect on drop-out rates at
third grade, by reducing the rate by half a percent. This is equivalent to a mean
reduction of 0.31 students per school/cohort and presumably negligible in its potential
impact on mean test scores.
6.6 Effect of the policy change on school transfer rates
Another potential source for compositional changes is related to student mobility
between schools. Parents that expect a negative effect of automatic promotion on
their children may for instance want to move their children to a school with grade
retention. In Minas Gerais the possibility for switching public schools is limited as
enrolment is mainly based on residence and a single public school often serves the
local neighbourhood. Given very substantial fees at private schools it is also unlikely
that parents move their children into private schools to avoid a specific grade
promotion regime. As the policy was introduced while the cohort of interest was in
second grade the incentive for parents to move their children is further reduced. To
test for any effect of the policy change on between-school mobility I estimate the
effect of the introduction of automatic promotion on student transfer rates using the
same framework as in the previous section. Columns (3) and (4) of table 4 report
point estimates for outgoing and incoming transfer rates that are close to zero and not
First grade repetition rates are also unaffected as predicted, because the policy change only takes
effect after first grade. This is a relevant observation as it shows that there are no anticipatory effects
from schools to the introduction to the policy change.
significant for any grades, so that there is no evidence that student mobility has an
impact on the student composition.
6.7 Systematic test taking behaviour
Although participation in PROEB is mandatory on the school and individual level,
some students fail to attend the test.
If the propensity to show up at the exam is
related to the capacity of the student and to the treatment status of the school, this
may bias the estimates. This might be induced by strategic behaviour of school
administrators or teachers trying to manipulate the mean test scores of their school in
the PROEB exam. If this is systematically linked to treatment status this could bias
the estimates. Notably, individually identified test results are not available to the
schools and PROEB test scores are not used by schools for the grade promotion
decisions. I use information from the official student numbers in each school from the
school census and compare these to the number of students participating in PROEB. I
estimate equation (2) using the difference between the two figures as outcome
variable. Table 7 presents the results from the regression. The coefficient is very
small (0.119 students) and not statistically significant so that there is no evidence for
systematic absence of students from the test.
6.8 Effect of the policy change on class size
There may be other teaching inputs that could be affected by the policy change; for
example a reduction in retention rates may affect class-size, which in turn may have
an impact on outcomes. There is a comprehensive literature on the effect of class size
on student performance but the overall picture about class-size effects remains rather
To rule out that the estimates are biased by an effect of the policy on class-
size I test for changes in class-size for each grade induced by the policy change for
the cohorts of interest and column (5) of table 4 reports the DiD results. There is no
evidence for an effect of the policy change on class-size in any grade, so that
estimates on test scores are unlikely biased by treatment induced class-size effects.
Even under the assumption that the introduction of automatic promotion releases
other school resources that could be allocated to fourth grade students (for which
The participation rate for the 2003 and 2006 wave of PROEB is around 95% as participation is
strictly enforced and absence is only permitted in case of illness.
See Hoxby (2000) and Angrist & Lavy (1999) for two prominent studies on class size effects.
there is no evidence in the present analysis) this would lead to underestimating the
true impact of the disincentive created by automatic promotion.
The fact that none of the above estimates (for repetition rates, drop-out rates, class-
size, transfer rates) reveal any significant effect for first grade estimates is in itself an
important falsification exercise. All these estimates are based on a placebo-treatment
as the first grade of the 2006 exam cohort was not yet affected by the policy
introduction. This also indicates that there are no anticipatory effects of the schools in
respect to the imminent introduction of automatic promotion that may affect student
outcomes at a later stage.
Existing empirical work on grade retention has to date focused on analysing the direct
effect of retention on repeaters. The focus on the ex-post effects may nevertheless
neglect an important effect of the grade retention regime that works through
incentives to study on a larger range of students than repeaters only. The introduction
of automatic promotion removes the incentives linked to the threat of retention and in
this paper I use exogenous variation in the timing of the policy adoption in public
primary schools in the Brazilian state of Minas Gerais to obtain causal estimates of
Using a DiD approach I find a negative effect of about 6,7% of a standard deviation,
significant at the 1% level. Controlling for individual age strengthens the negative
effect by about 20%, which gives an idea about the size of the bias associated with
the differential inflow of repeaters into fourth grade before and after treatment. The
estimated effect of the introduction of automatic promotion is of non-negligible size.
Considering that automatic promotion may have a negative effect in several grades,
the overall impact of the automatic promotion regime may lead to considerable loss
of academic achievement over the eight years of primary school. Quantile DiD
estimates yield an important insight into the distribution of effects. The quantile DiD
estimates reveal that a large set of students is impacted by the policy change and not
only the least performing students. The inverse u-shape of effects along the test score
distribution is consistent with an interpretation of the estimates as disincentive effect
of automatic promotion. Some further suggestive evidence on student responses
support this interpretation. Other potential channels, in particular related to changes in
the student composition, can be ruled out.
The estimation of a disincentive effect associated with automatic promotion closes a
gap in the literature on the effects of grade retention and helps to explain the
persistence of repetition regimes in many countries. Grade retention reduces internal
flow efficiency at schools and is a costly policy, but may have a positive effect on
academic achievement through a deterrence effect. Rather than focusing only on the
effect on repeaters, attempts to assess the costs and benefits of grade retention
therefore need to take into account the effects on non-repeaters as well.
This is important in the light of the universal introduction of automatic promotion in
all primary schools in Brazil that came into effect by federal legislation in 2011.
Although the Brazilian experience may not be completely transferable to other
countries – often with lower repetition rates – the findings may nevertheless be
relevant for countries facing pressures to meet the Millennium Development Goal of
universal primary education and who may regarding the introduction of automatic
promotion as a suitable way to reduce repetition rates and increase school completion
I am very grateful to Marco Manacorda for his thoughtful advice. I also thank David
Card, Jefferson Mainardes, Imran Rasul and seminar participants at University
College London, Queen Mary, the International Conference on the Economics of
Education in Zurich 2008, the Spring Meeting of Young Economists 2010, and the
Congress of the European Economic Association 2010 for valuable comments. I am
grateful to the Secretariat of Education in Minas Gerais, the Brazilian Ministry of
Education and the National Institute for Educational Studies and Research (INEP) for
the data. I thank Juliana Riani and Jorge Rondelli for assistance with obtaining the
data. I am also grateful to the Editor and two anonymous referees for their comments
which have improved the paper considerably. The usual disclaimer applies.
Angrist, J.D., Bettinger, E., Bloom, E., King, E., Kremer, M., 2002. Vouchers for
Private Schooling in Colombia: Evidence from Randomized Natural
Experiments. American Economic Review, 92(5), 1535-1558.
Angrist, J.D., Lavy, V., 1999. Using Maimonides’ Rule to Estimate the Effect of
Class Size on Scholastic Achievement. Quarterly Journal of Economics.
Ashenfelter, O., 1978. Estimating the Effect of Training Programs on Earnings.
Review of Economics and Statistics, 60(1), 47-57.
Athey, S., Imbens, G. 2006. Identification and Inference in Nonlinear Difference-in-
Differences Models. Econometrica, 74(2), 431-497.
De Janvry, A., Finan, F., Sadoulet, E., 2006. Evaluating Brazil’s Bolsa Escola
Program: Impact on Schooling and Municipal Roles. Unpublished manuscript.
De Mello, L., Hoppe, M., 2005. Education Attainment in Brazil: the Experience of
FUNDEF. OECD Economics Department Working Papers 424.
Donald, S., Lang, K., 2007. Inference with Differences-in-Differences and Other
Panel Data. Review of Economics and Statistics, 89(2), 221-233.
Dong, Y., 2009. Kept Back to Get Ahead? Kindergarten Retention and Academic
Performance. European Economic Review, 54(2), 219-236.
Eide, E., Showalter, M., 2001. The Effect of Grade Retention on Educational and
Labor Market Outcomes. Economics of Education Review, 20(6), 563-576.
Firpo, S., Fortin, M., Lemieux, T. 2009. Unconditional Quantile Regressions.
Econometrica, 77(3), 953-973.
Gamper-Rabindran, S., Khan, S., Timmins, C. 2010. The Impact of Piped Water
Provision on Infant Mortality in Brazil: A Quantile Panel Data Approach.
Journal of Development Economics, 92(2), 188-200.
Glick, P., Sahn, D., 2010. Early Academic Performance, Grade Repetition, and
School Attainment in Senegal: A Panel Data Analysis. World Bank Economic
Review, 24(1), 93-120.
Gomes-Neto, J.B., Hanushek, E.A., 1994. Causes and Consequences of Grade
Retention: Evidence from Brazil. Economic Development and Cultural Change,
Havnes, T., Mogstad, M. 2010. Is Universal Child Care Levelling the Playing Field?
Evidence from Non-Linear Difference-in-differences. IZA Discussion Paper
Holmes, T.C., 1989. Grade Level Retention Effects: A Meta-Analysis of Research
Studies. In: Flunking Grades: Research and Policies on Retention, ed. Lorrie
A. Shepard and Mary L. Smith, 16-33. London: Falmer Press.
IBGE (Brazilian Institute of Geography and Statistics), 2007. Contas Nacionais No.
21: Contas Regionais do Brasil 2002-2005. Rio de Janeiro.
Imbens, G., Wooldridge, J., 2009. Recent Developments in the Econometrics of
Program Evaluation. Journal of Economics Literature, 47(1), 5–86.
INEP (National Institute for Educational Studies and Research). 2007. IDEB - Prova
Brasil/Saeb 2007. Brasilia.
Jacob, B.A. 2005. Accountability, Incentives and Behaviour: Evidence from School
Reform in Chicago. Journal of Public Economics, 89(5-6), 761-796.
Jacob, B., Lefgren, L., 2004. Remedial Education and Student Achievement: a
Regression-Discontinuity Analysis. Review of Economics and Statistics, 86(1),
Jacob, B., Lefgren, L., 2009. The Effect of Grade Retention on High School
Completion. American Economic Journal: Applied Economics, 1(3), 33-58.
Koenker, R. 2004. Quantile Regression for Longitudinal Data. Journal of
Multivariate Analysis, 91(1), 74-89.
Lam, D. and Marteleto, L. 2006. Small Families and Large Cohorts: The Impact of
the Demographic Transition on Schooling in Brazil. In: Lloyd, C., Behrman, J.,
Stromquist, N., and Cohen, B. (eds). The Changing Transition to Adulthood in
Developing Countries. The National Academic Press, Washington.
Lamarche, C. 2011. Measuring the Incentives to Learn in Colombia Using New
Quantile Regression Approaches. Journal of Development Economics, 96(2),
Lavy, V., Paserman, D., Schlosser, A., 2012. Inside the Black Box of Ability Peer
Effects: Evidence from Variation in Low Achievers in the Classroom.
Economic Journal, 122(3), 208-237.
Lazear, E. 2001. Educational Production. Quarterly Journal of Economics, 116(3),
Lindert, K., Linder, A., Hobbs, J., de la Brière, B., 2007. The Nuts and Bolts of
Brazil’s Bolsa Família Program: Implementing Conditional Cash Transfers in a
Decentralized Context. Social Protection Discussion Paper 0709, World Bank.
Mainardes, J., 2010. Moving Away from a Graded System: a Policy Analysis of the
Primary Education Organized in Cycles in Brazil. Lampert, Saarbrücken.
Manacorda, M., 2012. The Cost of Grade Retention. Review of Economics and
Statistics, 94(2), 596-606.
Manski, C. 1993. Identification of Endogenous Social Effects: the Reflection
Problem. Review of Economic Studies, 60(3), 531-542.
MEC (Brazilian Ministry of Education), 1996. Lei de Diretrizes e Bases da Educação
National, No. 9394, Art. 24, December 1996, Brasilia.
OECD, 2005. OECD Economic Survey of Brazil 2005. Organization for Cooperation
and Development, Paris.
Pino, I.R., Koslinki, M.C., 1999. Government Programs to Eliminate Repetition,
School Dropout and Exclusion in Brazil. In: Schooling for Success: Preventing
Repetition and Dropout in Latin American Primary Schools, ed. Laura Randall
and Joan B. Anderson. New York: Sharpe.
UNESCO, 2008. EFA Global Monitoring Report. United Nations Educational,
Scientific and Cultural Organization, Paris.
UNESCO. 2012. Opportunities lost: The impact of grade repetition and early school
leaving. Global Education Digest 2012. United Nations Educational, Scientific
and Cultural Organization, Paris.
FIGURE 1: HISTOGRAM FOR THE PROPENSITY SCORE (YEAR 2003)
CHART 1: TREATMENT SCHEME
Notes: Testing takes place for all students at the end of 4th grade. The cohorts are
denoted according to the year in which they are tested through PROEB. 1 denotes
grades with grade retention, 0 denotes grades with automatic promotion.
.2 .4 .6 .8 1
Propensity score: cond. probability of treatment
TABLE 1: TEST SCORE MEANS IN TREATMENT AND CONTROL SCHOOLS
BEFORE AND AFTER THE ADOPTION IN THE TREATMENT SCHOOLS
Change in mean
Difference in mean test
Notes: Mean outcomes for treatment and control before and after treatment. Standard errors, adjusted for
clustering within SREs, are reported in parenthesis.
TABLE 2: MAIN ESTIMATION RESULTS AND SENSITIVITY TO
Dependent variable: PROEB math test scores
Observations: 244,081, number of clusters: 1,993
Chart B – adding individual age control
Chart C – adding peer age control
School fixed effects
School level controls
Peer characteristics controls
Individual characteristics controls
Notes: *** denotes significance at 1%. Robust standard errors, adjusted for clustering
within schools, are reported in parenthesis. Specification (1) contains year dummies and
school fixed effects, specification (2) additionally controls for a rich set of school
characteristics (physical characteristics of the school and the class rooms, teaching
material, teacher characteristics, participation in educational programmes etc.),
specification (3) additionally controls for peer socio-economic characteristics at the school
level and specification (4) also controls for individual characteristics.
TABLE 3: EFFECT OF THE ADOPTION OF AUTOMATIC
PROMOTION ON STUDENT FLOWS AND CLASS-SIZE
Number of schools: 1993, years 2000-2006, average cohort size: 61.24
Notes: *** denotes significance at 1%, ** denotes significance at 5%. The coefficients report the effect of introducing
automatic promotion on the dependent variables for 1st to 4th grade using data from the school census 2000-2006 following
the theoretical test cohorts. For each grade a separate regression has been fitted estimating the effect corresponding to
equation (1) as . The regression estimates are weighted by school cohort size and include
year dummies (dt) and school fixed effects (ds). Robust standard errors, adjusted for clustering within 46 SREs, are reported in
TABLE 4: QUANTILE TREATMENT EFFECTS
Dependent variable: PROEB test scores
Notes: *** denotes significance at 1%, ** significance at 5%, * significance at 10%. The coefficients report the quantile differences-in-difference
treatment effects for nine quantiles of the test score distribution. The regressions include year dummies and school fixed effects. Bootstrapped
standard errors (200 repetitions) adjusted for clustering on the school level are reported in parenthesis.
Notes: *** denotes significance at 1%. The above samples exclude students that are below the
target age range. Robust standard errors, adjusted for clustering within schools, are reported in
parenthesis. Specification (1) only includes year dummies and school fixed effects, specification
(2) additionally controls for a rich set of school characteristics (physical characteristics of the
school and the class rooms, teaching material, teacher characteristics, participation in educational
programmes etc.), specification (3) additionally controls for peer socio-economic characteristics
at the school level and specification (4) also controls for individual characteristics.
TABLE 5: ESTIMATION RESULTS FOR RESTRICTED AGE RANGES
Dependent variable: PROEB test scores
Number of clusters: 1,993
Chart A – students in target age range for 4th grade
Chart B – repeaters (outside target age range)
School fixed effects
School level controls
Peer characteristics controls
Individual characteristics controls
TABLE 6: EFFECT OF THE ADOPTION OF AUTOMATIC
PROMOTION ON THE SOCIO-ECONOMIC COMPOSITION
Robust standard error
Proportion of white students
Proportion of mixed students
Proportion of black students
Proportion if Asian students
Proportion of indigenous students
Mean age (in years)
Mean male students
HH wealth index
Washing machine mean
Notes: denotes ** significance at 5% level, *** significance at 1% level. All estimates refer
to school means or proportions at the school level. All data is taken from the socio-
economic questionnaire of PROEB. For each dependent variable the effect is estimated
separately in a regression corresponding to equation (1) as
. The regression estimates are weighted by school cohort size and include a year
dummies (dt) and school fixed effects (ds). Robust standard errors, adjusted for clustering
within SREs, are reported in parenthesis. All estimates are weighted by school cohort size.
TABLE 7: EFFECT OF THE ADOPTION OF AUTOMATIC
PROMOTION ON PARTICIPATION IN PROEB
Dependent variable: difference between official student numbers and PROEB
Notes: The coefficient reports the effect of the introduction of automatic promotion on the
difference of the number of students according to the school census and the PROEB test. The
effect is estimated by a regression corresponding to equation (1) as
. The estimates are weighted by school cohort size and include year dummies (dt) and
school fixed effects (ds). Robust standard errors adjusted for 46 clusters (on SRE level) are
reported in parenthesis.
TABLE 8: EFFECT OF POLICY ADOPTION ON STUDENT NET FLOW
Student net inflow up
to 1st – 3rd grade
Student net inflow including
Notes: *** denotes significance at 1%. The coefficients report the effect of introducing
automatic promotion on net flow (including in/outflow due to repetition using data from
the school census 2000-2006. A separate regression has been fitted estimating the effect
corresponding to equation (1) as for the two models.
Model (1) refers to net flows including 2nd and 3rd grade, model (2) refers to net flows
including 2nd, 3rd and 4th grade. The regression estimates are weighted by school cohort
size and include a year dummy (dt) and school fixed effects (ds). Robust standard errors
adjusted for 46 clusters (on SRE level) are reported in parenthesis.
TABLE A1: MEAN CHARACTERISTICS OF TREATMENT AND CONTROL GROUPS IN 2003 AND 2006
CHART A: PHYSICAL CHARACTERISTICS AND SCHOOL PROGRAMME PARTICIPATION
Table A1 cont.
Perm. class rooms
Prov. class rooms
Min inc. program
Other education TV
Classes in 1st grade
Classes in 2nd grade
Classes in 3rd grade
Classes in 4th grade
Notes: The binary variables of school characteristics and programme participation are coded 0 for not present (no participation) and 1 for present (participation). All
data is from the Brazilian school census 2003 and 2006. The p-value is reported from a test on the equality of the mean between the treatment and control groups
(independent samples). As the sample size is sufficiently large the result for using a classical t-test or taking into account the binary values and the underlying
binomial distribution deliver very similar results. As the group size and with it the variances between the groups differ, approximate t using individual sample
variances instead of the pooled variance and Welch’s approximation of the degrees of freedom have been used.
The normalized difference is computed as
, where S2 denotes the sample variance of Xi.
TABLE A1: MEAN CHARACTERISTICS OF TREATMENT AND CONTROL GROUPS IN 2003 AND 2006
CHART B: INDIVIDUAL AND FAMILY CHARACTERISTICS AT SCHOOL LEVEL
Age (in months)
% white students
% mixed students
% black students
% Asian students
% indig. students
Education ( mother)
Notes: All data is taken from the socio-economic questionnaire of PROEB 2003 and 2006. Information on educational background and literacy of parents is only available
in the 2003 questionnaire. The p-value is reported from a test on the equality of the mean between the treatment and control groups (independent samples). As the sample
size is sufficiently large the result for using a classical t-test or taking into account the binary values and the underlying binomial distribution deliver very similar results.
As the group size and with it the variances between the groups differ, approximate t using individual sample variances instead of the pooled variance and Welch’s
approximation of the degrees of freedom have been used.
The normalized difference is computed as
, where S2 denotes the sample variance of Xi.
TABLE A2: MEAN PRE-INTERVENTION SCHOOL CHARACTERISTICS IN 1997
Repetition rate 1st grade
Repetition rate 2nd grade
Repetition rate 3rd grade
Repetition rate 4th grade
Class size (grades 1-4)
Student-teacher ratio (grades 1-4)
Notes: All data is from the Brazilian school census 1997. The p-value is reported from a test on the
equality of the mean between the treatment and control groups (independent samples). As the sample
size is sufficiently large the result for using a classical t-test or alternatively taking into account the
binary values and the underlying binomial distribution deliver very similar results. As the group size and
with it the variances between the groups differ, approximate t using individual sample variances instead
of the pooled variance and Welch’s approximation of the degrees of freedom have been used.
The normalized difference is computed as
, where S2 denotes the sample variance
TABLE A3: RESPONSES OF STUDENTS, PARENTS AND TEACHERS
Notes: *** denotes significance at 1%, * denotes significance at 10%. Robust standard errors,
adjusted for clustering within schools are reported in parenthesis. The coefficients report the
effect of the introduction of automatic promotion behavioural responses of students, parents
and teachers. The effects are estimated by a regression corresponding to equation (1) as
. The binary outcome variables were constructed using
consistent information from the socio-economic questionnaire of PROEB 2003 and 2006. The
dependent variable in column(1) reports change in fraction of students always doing their
homework (mean 0.706), in column (2) the change in fraction of students always receiving
help from their parents with their homework (mean 0.652), in column (3) the change in
fraction of teachers assigning homework (mean 0.981), and in column (4) the change in
fraction of teachers always correcting homework of their students (mean 0.767).
TABLE A4: SENSITIVITY OF ESTIMATES TO INDIVIDUAL AND
PEER AGE CONTROLS FOR DIFFERENT AGE RANGES
Dependent variable: PROEB test scores
Chart A – all students
Peer and individual controls
Chart B – students in target age range for 4th grade
Peer and individual controls
Notes: *** denotes significance at 1%. All estimates include controls for school characteristics (physical
characteristics of the school and the class rooms, teaching material, teacher characteristics, participation in
educational programmes etc.) The specifications (1) include additionally controls for peer socio-economic
characteristics, specifications (2) control for per and individual characteristics, specifications (3) control for
individual characteristics. The row below specifies further controls for individual and peer age in the
TABLE A5: EFFECT OF POLICY ADOPTION ON 3RD GRADE
REPETITION RATE FOR TREATMENT SCHOOLS
3rd grade repetition
Notes: *** denotes significance at 1%. The coefficient reports the effect of introducing
automatic promotion on 3rd grade repetition rate for the cohort of interest at treatment
schools. Robust standard errors adjusted for 46 clusters (on SRE level) are reported in
TABLE A6: LINEAR PROBABILITY MODEL OF ASSIGNMENT TO TREATMENT
School provides all years of fundamental education
proper school building
building of other school
toilets outside school
toilets inside school
toilets ready for special needs
school ready for special needs
public energy supply
using 220 volt
using 110 & 220 volt
public water supply
no running water
permanent class rooms
provisory class rooms
class rooms in school
class rooms away from school
School size proxies
student enrolment 1st year
total number of staff
total number of teachers
number of teachers in classes 1-4
Minimum Income Programme
Parameters in Action
School Transport Programme
National Library Programme
State computer programme
Municipal computer programme
other state programme
other municipal programme
Free School Lunch
Free Public School Transport
Notes: * significant at 10%; ** significant at 5%; *** significant at 1%, standard errors are in parenthesis.
The data is taken from the 2003 school census. Specification (2) includes regional school administration
dummies (SRE). Most of the physical characteristics describing the schools are indicator variables on the
presence at school. Similarly, indicator variables inform about participation in education programmes. The
programme Parameters in Action is a federal programme for the professional development of teachers; FNDE
denotes a maintenance and development programme for education by the National Fund for the Development
of Education, Ouvebem is a national campaign for the importance of the sense of hearing, Reabvis is a
national campaign on visual rehabilitation, PROINFO is a federal computer literacy programme. The
coefficients reported are from two specifications of a linear model of the effect of school characteristics on the
probability for treatment. The dependent variable is an indicator that equals zero for being in treatment group
and equals 1 for being in the control group. In specification (1) only very few coefficients are significant at
the 1% and 5% level of significance, of which some disappear when including regional dummies
(specification (2). None of the coefficients of the linear model produces values outside the unit-interval and a
logit specification delivers very similar results to the linear specification diminishing doubts on the suitability
of the linear specification (not reported).