77633
Early Academic Performance, Grade Repetition,
and School Attainment in Senegal: A Panel Data
Analysis
Peter Glick and David E. Sahn
Little is known in developing country environments about how a child’s cognitive
skills manifested in the �?rst years of schooling are related to later educational success,
because the panel data needed to analyze this question have been lacking. This study
takes advantage of a unique data set from Senegal that combines test score data for
children from the second grade with information on their subsequent school pro-
gression from a follow-up survey conducted seven years later. Measures of skills from
early primary school, corrected for measurement error using multiple test observations
per child, are strongly positively associated with later school progression. A plausible
interpretation is that parents invest more in a child’s education when the returns to
doing so are higher. The results point to the need for remedial policies to target
lagging students early on to reduce early dropout. Grade repetition policies target
poorly performing students and are pervasive in Francophone Africa. Using variation
across schools in test score thresholds for promotion to identify the effects of second-
grade repetition, the analysis shows that repeating students are more likely to leave
school before completing primary school than students with similar ability who are
not held back, pointing to the need for alternative measures to improve the skills of
lagging children. JEL codes: I21, O15
Reflecting widespread recognition of the importance of human capital for indi-
vidual welfare and economic growth, an enormous amount of empirical
research in developing countries has examined the determinants of household
investments in children’s schooling.1 However, little is known about the
dynamics of schooling, in particular, the relationship of early cognitive skills to
later education outcomes. In developing countries, where households are likely
to be resource and credit constrained, decisions on how long to continue a
child’s education may depend strongly on the perceived returns to schooling
Peter Glick (corresponding author, pglick@rand.org) is a senior economist with the RAND
Corporation. David E. Sahn (david.sahn@cornell.edu) is International Professor of Economics at Cornell
University. The authors thank three anonymous referees and seminar participants at Cornell University
and RAND for helpful comments and suggestions.
1. See Glewwe and Kremer (2006) and Glick (2008) for reviews of the literature.
THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 1, pp. 93 – 120 doi:10.1093/wber/lhp023
Advance Access Publication January 25, 2010
# The Author 2010. Published by Oxford University Press on behalf of the International Bank
for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,
please e-mail: journals.permissions@oxfordjournals.org
93
94 THE WORLD BANK ECONOMIC REVIEW
and, in turn, on a child’s apparent ability. The effect of ability may be particu-
larly large among poorer families in such environments, for whom resource or
credit constraints are likely to be more binding. Knowledge of the nature and
extent of these patterns would be highly relevant for policy. In particular, a
strong connection between early achievement as manifested by test scores and
subsequent school attainment would point to the need for policies directed at
children who fall behind academically early in their school careers.
Knowledge of this relationship in developing country contexts is very sparse,
primarily because the data needed to analyze this question have been lacking.
The vast majority of studies of schooling in developing countries rely on cross-
sectional data. Dynamic analyses of school attainment, however, require panel
data containing both early measures of skills and later schooling outcomes.
Surveys with this information are rare in developing countries, especially for
Africa.
This study takes advantage of unusual panel data from Senegal that com-
bines test score data for children from the second grade with information on
their school progression from a follow-up survey conducted seven years later.
With information on school and household characteristics, the analysis is able
to control for confounding influences on school attainment in estimates of the
effects of early academic performance on later education outcomes as well as
determine whether wealth and school differences exacerbate or mitigate the
impacts of early achievement gaps among children.
The study also exploits the panel to assess the impacts of early grade rep-
etition—a pervasive policy targeting poor achievers—on later school attain-
ment. While repetition normally must be considered as jointly determined by
factors that also influence attainment, this study is able to take advantage of
the fact that, conditional on observed academic ability at the end of second
grade, having to repeat second grade is reasonably considered exogenous to
student ability or effort.
Indicators of early cognitive ability, corrected for measurement error using
multiple test observations per child, are strongly positively associated with later
school progression. If interpreted causally, this result is consistent with a model
of human capital investments under certain assumptions about returns to
schooling in the labor market. Household wealth and parental (father’s)
schooling also have (conditional on early measured ability) direct positive
effects on attainment, and there is some evidence that the impact of academic
ability on later school outcomes differs for children in poor and well-off house-
holds, as well as between girls and boys. The large impact of early test scores
on later attainment, controlling for family and school characteristics, points to
the importance of school policies that direct early remedial attention to low-
achieving children. Finally, conditional on academic ability, repeating a grade
has a negative impact on school progression, implying that the private costs
associated with stringent repetition policies exacerbate the negative effects on
attainment of poor early academic outcomes.
Peter Glick and David E. Sahn 95
I. CONCEPTUAL FRAMEWORK
A simple model of child ability and household schooling investments helps
frame the empirical analysis. The framework shows that making predictions
about the relationship of ability to schooling attainment (or about interactions
of household wealth and ability in schooling demand) is not straightforward
without speci�?c assumptions about the shape of the schooling returns func-
tions. Consider a two-period model of parental investment in the education of
two children of differing ability, A i ( ¼ 1,2), as measured, for example, by tests
early in school, and assume A 1 . A 2. For the exposition, assume that A i rep-
resents the ability endowment of the child given exogenously to the parents (as
discussed below, this cannot be assumed for the test score measures used in
this study). For a pure investment model in which parents view schooling only
as a means to future consumption through transfers from children’s period 2
wealth (W i), the lifetime utility of the parents is written as:
À  Ã�?
ð1Þ U ¼ U1 C1 ; C2 W 1 ðS1 ; A1 Þ; W2 ðS2 ; A2 Þ
where Ct is consumption in period t, and W 1 and W 2 are functions of individ-
ual child ability and the level of schooling S i. The parents maximize equation
(1) subject to the intertemporal budget constraint
C2 W 1 ðS1 ; A1 Þ W 2 ðS2 ; A2 Þ
ð2Þ C1 þ ¼ Y1 þ þ À P1 S1 À P2 S2
1þr 1þr 1þr
where P i is the unit cost of schooling for child i. The �?rst-order conditions
imply (in the absence of credit constraints) that parents invest in the education
of each child until the marginal return equals the market rate of interest:2
1 @ W 1 ðS1 ; A1 Þ 1 @ W 2 ðS2 ; A2 Þ
ð3Þ ¼ ¼ 1 þ r:
P1 @ S1 P2 @ S2
If costs are the same, the difference in the optimal levels of S 1 and S 2 will
depend on how ability affects the schooling return functions. A standard
assumption in models of human capital investment is that returns diminish
with additional schooling, which can be captured using the quadratic form
W ¼ a1S 2 a2S 2 (a1 . 0, a2 . 0). A simple way to allow returns to be influ-
enced by ability is to assume that returns to the low-ability child, relative to
those to the high-ability child, are reduced proportionately by some factor b
(0 , b , 1), so returns for the low-ability child are b(a1S 2 a2S 2). This yields
marginal return functions a1 2 2a2S for the high-ability child and ba1 2
2ba2S for the low-ability child (�?gure 1). At any level of schooling, marginal
returns are higher for the high-ability child, so for any interest rate, investments
2. For the exposition, direct schooling impacts on utility are ignored so this becomes a pure
investment model.
96 THE WORLD BANK ECONOMIC REVIEW
F I G U R E 1. Higher Marginal Returns for Higher Ability Child and Larger
Income Effects for Low-Ability Child
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
in that child’s schooling are greater. If households are credit constrained, a
similar result would be obtained, but marginal returns would be equated to the
parents’ marginal rate of substitution of future for current consumption, the
shadow interest rate (r*).
While this outcome seems intuitive, it depends on the form of the returns
functions, as pointed out by Behrman (1997). Alternative and no less arbitrary
assumptions about how ability influences human capital production could yield
a different outcome. One could flexibly represent quadratic returns to the low-
ability child as ga1S 2 da2S 2, with 0 , g , 1, 0 , d , 1, implying marginal
returns ga1 2 2da2S. Speci�?c values of these parameters would produce mar-
ginal returns that are larger for the high-ability child over most of the schooling
range but eventually become larger for the low-ability child, as depicted in
�?gure 2. Thus, at a low enough interest rate (such as r in the �?gure), schooling
investments will be greater for this child. Therefore, the relationship of total
schooling and child ability is conceptually ambiguous.3 Consumption-based con-
siderations would add more ambiguity. For example, parental preferences for
3. The likelihood that the marginal return functions cross as in �?gure 2 increases if the high-ability
child is also more productive in home or work activities. This raises the opportunity cost of schooling of
that child relative to the low-ability child and, through the increase in cost P in equation (3), shifts
down the marginal returns for the more able child.
Peter Glick and David E. Sahn 97
F I G U R E 2. Switch in Relative Marginal Returns for High- and Low-Ability
Children, and Larger Income Effects for Low-Ability Child
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
equity might lead to investing more resources in low-ability children to compen-
sate for their otherwise lower future income and welfare.
Next, consider another question of interest for the empirical analysis: how
the response to ability may be affected by the level of household income—
essentially, the sign on an interaction term of an ability measure and household
income or assets in a model estimating schooling attainment.4 In a pure invest-
ment framework with no borrowing constraints, there should be no interaction,
since investments occur until marginal returns equal the interest rate. But credit
constraints are likely, and these should impinge more on households with low
current resources. For a credit-constrained (poor) household, the rate of interest
is determined as indicated by the marginal rate of substitution of future for
current consumption. If the constraint binds, this shadow interest rate r* will
be higher than the market rate r. This is shown in �?gure 1, with schooling of
0 0
the two children under the binding constraint equal to S 1 and S 2 . The effect
of a suf�?ciently large increase in income is to relax the credit constraint so that
marginal returns are equated to r rather than r*. While in both the poor (con-
strained) and better-off household, the higher ability child receives more
schooling, this gap gets smaller as income increases, because (given the nature
of the quadratic returns functions) marginal returns are falling less steeply for
the low-ability child (see �?gure 1). Therefore, it takes more additional invest-
ment to reach equilibrium for that child at the new lower interest rate.
4. The analysis here is formally similar to Garg and Morduch’s (1998) analysis of the effects of
changes in income on girl– boy gaps in health investments.
98 THE WORLD BANK ECONOMIC REVIEW
F I G U R E 3. Higher Marginal Returns and Larger Income Effects for
High-Ability Child
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
The returns functions depicted in �?gure 2 yields this result as well. Once
again, however, different assumptions about how ability affects the returns to
schooling imply different outcomes. If, as in �?gure 2, returns to the low-ability
child are represented by the quadratic ga1S 2 da2S 2 but with a suf�?ciently
smaller absolute value of g (or larger value of d), returns could appear as in
�?gure 3. In this case, the low-ability child receives less schooling both at low
and high incomes, but the gap widens with income, because the marginal
returns fall more steeply for this child. Again, this ambiguity occurs in a pure
investment framework without resort to assumptions about parental prefer-
ences, though these would add further uncertainty to predictions. For example,
if inequality aversion were a normal good, increases in income could lead to
greater additions to investments in the low-ability child even if credit con-
straints were not operative.
Therefore, the relationship of early manifested ability and subsequent school
investments, and the role of household resources in modifying the effects of
ability on investments, are empirical questions. They can be addressed here
because of the availability of both test score measures of early ability and infor-
mation on subsequent schooling.5
5. However, a disadvantage of the data used for this study is that they do not have this information
for multiple children in the same family, so household �?xed effects cannot be used to eliminate
unmeasured household factors associated with both ability and later outcomes. The heterogeneity issue
is discussed in the next section.
Peter Glick and David E. Sahn 99
I I . D ATA AND E M P I R I CA L A P P ROAC H
The 2003 Senegal Household Education and Welfare (EBMS) survey6 was con-
ceived as a follow-up to an earlier, school-based, study, the Program on the
Analysis of Education Systems of the Conference of Francophone Ministers of
Education (PASEC). The PASEC study administered tests of written math and
French to a cohort of students beginning in second grade in 1995 and continu-
ing through subsequent years of primary school (see CONFEMEN 1999;
Michaelowa 2003). The 2003 survey attempted to re-interview these children,
who were now of middle school age (ages 14–17). Of the original 120 PASEC
clusters, 60 (28 urban and 32 rural communities) were randomly selected for
the new survey. Of the 20 PASEC children per cluster/school who were tested
in second grade in 1995–96, survey enumerators were able to �?nd, on average,
15 children in rural clusters and 17 in urban ones. It should be noted that the
sample, by construction, selected on children who made it into second grade,
which eliminates a nontrivial share of children who do not make it that far
( primarily because they never entered school).7
The present analysis combines the test score information and detailed school
and teacher information from the PASEC survey with information from the
2003 EBMS survey, which contains detailed information on household charac-
teristics and child schooling. Information was also collected on the schooling
of children who no longer reside at home. This is important since 16 percent
of the PASEC cohort was no longer living at home in 2003, and this behavior
is likely to be endogenous to schooling outcomes (for example, a young person
may live with relatives to attend secondary school, or leave school and home to
marry or �?nd work).8
In simplest form, the models estimated can be represented as follows:
ð4Þ Si ¼ a0 þ a1 A95i þ a2 Xi þ a3 Qi þ a4 AÃ Ã
95i Xi þ a5 A95i Qi þ ei ;
where Si is the later schooling measure (grade attainment reported in 2003) of
child i, A95i is ability measured in 1995–96 (test score in second grade), Xi is a
vector of individual and household characteristics, and Qi is a vector of school
inputs such as teacher background and supplies.
6. EBMS was a joint research project of Cornell University (United States), Centre de Recherche en
Economie Applique ´ e (Senegal), and Institut National de la Recherche Agronomique (France).
7. Randomly selected households with children of similar age in the same school catchment areas
were also interviewed in the EBMS survey to achieve greater representivity and a larger sample. Of this
group, 24 percent did not enter second grade, almost all of whom never enrolled. Since this analysis
focuses on the links of early academic success and later school outcomes, only the PASEC cohort, for
which early test score information is available, is used.
8. For children not living at home, information was collected on highest completed grade and
current enrollment but not on school entry and exit ages, so total years in school cannot be calculated
for these children. Therefore, the highest completed grade is used as the dependent variable in the
analysis. It will differ from total years of school because of grade repetition. The implications of
repetition are discussed in section IV.
100 THE WORLD BANK ECONOMIC REVIEW
The coef�?cients must be interpreted carefully and will depend in part on how
A95i is represented. One option is to use actual test scores from the 1995 to 1996
school year. Using the second-grade pretest (given at the start of the school year)
provides a measure of ability at the end of the child’s �?rst year of school, thus
coming close to measuring ability at the time of school entry. However, in
addition to measuring a child’s inherent ability ‘endowment,’ these pretest scores
also reflect school and home inputs into learning during �?rst grade and home
inputs during the preschool period that lead to better cognitive development;
nutrition is likely to be particularly important (Glewwe and Miguel 2008). If,
further, these unmeasured inputs also directly affect later attainment or are corre-
lated with factors that do (such as parental preferences), this will upwardly bias
the estimated effect of early achievement. Regarded as associational rather than
as strictly causal, the relation of early performance to school attainment is still
relevant for policy. Test score information is relatively easy to obtain and could
potentially be used to direct remedial action to academically lagging children.
An alternative to test scores is the ‘residual method’, in which the unex-
plained portion in a production function regression for human capital (with
the test score as the dependent variable) is taken as the individual’s genetic
ability endowment. In this approach an equation such as the following is esti-
mated:
ð5Þ TSi ¼ b0 þ b1 Xi þ b2 Q95i þ vi :
where TSi is the 1995 pretest score and the residual vi is the endowment—
ability purged (in contrast to the test score itself) of the effects of measured
household characteristics Xi and grade-speci�?c school factors Q95isuch as
teacher background and class variables (note that Q95idiffers from Qi in
equation (4), which would incorporate all years and teachers of the child at the
school). Conceptual models of parental decision-making would generally
assume that parents are able to assess the child’s ability as represented by vi
and make investment decisions based on it. For modeling behavior, then, this
may be the preferred measure. Estimation of equation (5) requires detailed
school, teacher, and household data, which the Senegal surveys have.
However, there are shortcomings with using this measure of A95i in equation
(4) as well. Since only measured school and household inputs are controlled,
the residual will pick up the effects of unmeasured household or school factors.
Moreover, these may be correlated with Si, leading to endogenity problems in
estimating the effect of “pure�? ability, just as with early test scores. A second
problem, relevant to either ability measure, is measurement error. It is easy to
see how a single test can be a noisy measure of a student’s knowledge. In par-
ticular, performance could be highly dependent on the students’ disposition on
the day of the test. If this source of error is random, it will, all things equal,
bias toward zero the estimated impacts of early test score (or similarly, endow-
ment as measured by the residual).
Peter Glick and David E. Sahn 101
Fortunately, the PASEC data permit the use of an instrumental variables
procedure to deal with this problem and correct for (nonsystematic) measurement
error. For each second-grade student there are multiple—potentially four—test
scores: French and math pretests, and French and math posttests (conducted at
the end of the school year). Say that test score TS1 is being used as a measure of
cognitive ability and that the relationship between them takes the form A95i ¼
a0 þ a1TS1i þ m1i, where mi captures the noise in the relationship. An additional
test TS2 provides another proxy for ability: A95i ¼ b0 þ b1TS2i þ m2i. If measure-
ment error is uncorrelated across these indicators (Cov(m1i, m2i) ¼ 0), their corre-
lation is due only to their common dependence on A95i. A consistent estimate of
a1 in equation (1) can then be obtained by instrumenting for one test score using
the value of the other.9 The method is applicable as well if the residual approach
is used (Pitt, Rosenzweig, and Hassan 1990); in that case, one test score residual
is used as an instrument for the other.
The data provide several options: use the French second-grade pretest to
instrument the math pretest score (and the reverse), or use the second-grade
posttests to predict pretest scores. The �?rst approach appears less desirable. Since
the two pretests were given at the same time, measurement errors are likely to be
correlated (perhaps because of the child’s disposition on that day), violating the
key assumption for the instrumental variable strategy. A second issue is that a
nontrivial number of students are missing test scores for one of the pretests. For
these reasons, the posttest average of the math and French scores (or test score
regression residual) are used to predict the average of the pretest math and
French scores (or residuals); averaging the two permits the inclusion of cases
where only one test was taken in either or both periods. The less attractive
aspect of this choice is that the posttest score instrument will incorporate the
effects of school inputs in second grade (and the posttest residual will capture
impacts of unmeasured inputs during this period). Finally, some 6 percent of
children in the sample took the posttests but not the pretests. Estimates from the
predicting equation are used to impute the pretest score for these cases.
It should be stressed that this instrumental variables strategy deals with
measurement error but not with any potential endogeneity-related bias in the
estimate of a1, which would come about if unmeasured school or home inputs
are correlated with both initial achievement and subsequent attainment.10
9. This is the multiple indicators approach (see Wooldridge 2002, for a general presentation). It is
noteworthy that even with a correlation between the error terms of the indicators (systematic
measurement error), Bound, Brown, and Mathiowetz (2001) show that the multiple indicator
instrumental variables solution will still generally be a more consistent estimator of the true parameter
of interest than will ordinary least squares estimation using the indicator with measurement error.
10. Potential instruments for early academic achievement in the data, at least for the second-grade
posttest score, include second-grade teacher variables related to quali�?cations and pedagogical practices,
provided these are truly idiosyncratic and vary across teachers/grades within a school. While some of
these factors are indeed associated with the posttest outcomes, these instruments had low power and
resulted in implausible or highly nonrobust second stage estimates of the effect of early test score on
grade attainment.
102 THE WORLD BANK ECONOMIC REVIEW
The fact that the measured relation of later success to early achievement may
be associational rather than strictly causal is not necessarily a hindrance from a
policy perspective. As already noted, knowledge of this association would
make it possible to target early on children who are otherwise likely to do
poorly later. For a behavioral interpretation, however, heterogeneity bias is
clearly a greater concern. A partial correction is to estimate school �?xed
effects, which purge estimates of bias from linear associations of school-level
unobservables with the included covariates. It is plausible that unmeasured
household preferences and child ability operate in signi�?cant part through the
choice of school (and implicitly through location since school and cluster are
interchangeable in the PASEC survey). The �?xed effects estimates will eliminate
bias from this source. They will not, obviously, account for heterogeneity
among children within schools. Therefore, it cannot be claimed that the esti-
mates of the effects of early skills represent purely causal effects.11
With respect to the functional form of equation (4), ordered probit is used
to model the determinants of years of completed schooling. The ordered probit
model treats grade attainment as the outcome of a series of ordered discrete
choices. The model is easily extended to allow for right censoring of the depen-
dent variable, an important concern since some 60 percent of the cohort is still
attending school.12
Two alternate speci�?cations are estimated, reflecting different assumptions
about unobserved (school-level) heterogeneity: random effects and (as just
described) �?xed effects. The school random effects model directly parame-
terizes the within-school correlation of errors but assumes that these are not
correlated with included regressors. Estimation of the random effects model
involves specifying the likelihoods conditional on the school-level random
effect, then integrating over all possible values of the random effect to obtain
the unconditional likelihood. This is done using Hermite integration, following
the approach suggested by Butler and Mof�?t (1982) (for further details, see
Glick and Sahn 2000). The �?xed effects speci�?cation, in contrast, does not
require the orthagonality assumption on the school-level errors; it eliminates
any linear associations of the errors with included covariates by including
dummy variables for each school. If the correlation of school-level covariates
11. It is useful to compare the approach of this study with several other analyses (Glewwe and
Jacoby 1994; Alderman and others 1996; Kingdon 1996) on cross-sectional data that estimate school
attainment or skills using contemporaneously measured ability as captured by Raven’s abstract thinking
test. This is usually presented as a measure of innate (hence preschool) ability, but as Glewwe and
Kremer (2006) note that interpretation is implausible because the test is likely to also reflect an
individual’s schooling. Therefore, these studies do not provide strictly causal estimates of the impact of
innate ability (as the present study does not), nor do they permit an understanding of how later
education success is related to ability measured at or near the start of school (which, in contrast, the
panel data in this study do permit).
12. As almost no children report a grade higher than 9 by the 2003 survey, the top category is
aggregated as grade 9 and above.
Peter Glick and David E. Sahn 103
and the error term is in fact zero, however, the �?xed effects model is less
ef�?cient than the random effects model.13
Since estimation in the �?xed effects model is based only on within-school
variation, the effects of school factors that are constant across children within
a school are not estimated. Nevertheless, the impacts of teacher characteristics
and classroom supplies can be estimated because they retain some variation
across children within the same school. Speci�?cally, the 1995–96 school
surveys included detailed information on teacher background and classroom
characteristics and supplies for the original second-grade PASEC class and for
subsequent primary grades attended by the cohort. Basic teacher and class
information was also collected on at least one class containing grade repeaters
in subsequent years (for example, a second-grade class in 1996–97, when non-
repeaters would be in the third grade). The ordered probit regressions include
the averages for these variables for each student enrolled in classes for the
school years for which complete information is available: 1995–96,1996–97,
and 1997–98 (there will be fewer than three years of data for early dropouts
or absent children).14 The paths of individual children diverge, because some
children repeat one or more grades while others progress smoothly to the next
grade (while some drop out), so there is some within-school (across child) vari-
ation in teacher characteristics and classroom supplies measures.
Information on director background and school facilities was collected only
in 1995. Because of the large number of potential covariates for classroom
supplies and facilities, principal components analysis was used for data
reduction. For classroom supplies, the variables used were presence of a dic-
tionary, ruler, compass, and eraser; for school facilities, the variables were
presence of a library, toilet, drinking water source, and canteen. The �?rst prin-
cipal component in each case was included in the school attainment models.
Finally, as the household survey did not collect consumption or income data,
factor analysis was used to construct a wealth index from information on own-
ership of durable goods (see Sahn and Stifel 2003 for details of the method).
Table 1 presents descriptive statistics on the variables used in the analysis.
13. A potentially serious concern with �?xed effects in nonlinear models is bias due to the incidental
parameters problem: the inclusion of cluster dummy variables means that as total sample size increases,
so does the number of parameters to be estimated (with a �?xed number of observations T per cluster,
increasing the sample size means increasing the number of cluster dummy variables). This renders all
the estimates inconsistent asymptotically. However, Monte Carlo evidence provided by Greene (2002)
indicates that for the values of T in the data set used here (about 15), the sample bias is small for an
ordered probit model—slightly under 10 percent—with a small downward bias in variance as well. In
view of these results, it is reasonable to use �?xed effects in this study, while recognizing the likelihood
of a small degree of bias.
14. In addition, for children who repeat second grade twice and hence are still in second grade in
the third year of the PASEC study (1997– 98), information on their teacher and class will be lacking for
that year, because data in 1997–98 were collected only for fourth and third grade classes. For these
cases, information on the child’s second-grade class in 1996–97 was used.
104 THE WORLD BANK ECONOMIC REVIEW
T A B L E 1 . PASEC Cohort: Means and Standard Deviations of Selected
Variables
First pretest score Fourth pretest score
All quartile quartile
Standard Standard Standard
Variable Mean deviation Mean deviation Mean deviation
Highest grade completed 6.03 1.65 5.28 1.45 7.07 1.51
Still in school in 2003 0.63 0.48 0.58 0.49 0.78 0.42
Standardized second-grade 0.003 0.73 2 0.895 0.20 0.971 0.33
pretest scorea
Age (years) 15.62 0.95 15.48 0.97 15.56 0.92
Girl 0.41 0.49 0.42 0.49 0.42 0.49
Asset indexb 2 0.022 0.91 2 0.293 0.81 0.204 1.00
Mother’s years of 1.53 3.06 1.07 2.46 2.30 3.79
schooling
Father’s years of schooling 2.70 4.27 2.23 3.82 3.82 5.07
Missing mother’s 0.10 0.30 0.07 0.26 0.13 0.34
schooling (percent)
Missing father’s schooling 0.17 0.37 0.13 0.34 0.18 0.38
(percent)
Rural (percent) 0.56 0.50 0.63 0.48 0.49 0.50
School director’s 12.67 8.92 9.95 8.54 14.43 9.22
experience (years)
Director has baccalaure ´ at 0.54 0.49 0.58 0.50 0.50 0.50
or higher (percent)
Teachers, average years of 11.53 7.16 9.96 7.64 12.48 6.67
experience
Teachers, share with a 0.43 0.39 0.40 0.42 0.45 0.35
baccalaure ´ at or higher
(percent)
Teachers, share female 0.24 0.36 0.20 0.34 0.32 0.38
(percent)
Classroom supplies �?rst 0.01 1.38 2 0.09 1.61 0.07 1.15
principal componentc
School facilities �?rst 2 0.07 1.42 2 0.29 1.48 0.13 1.36
principal componentc
Number of observations 834 208 209
Note: Teacher and classroom supplies variables are the child-speci�?c averages for three years
of data from the PASEC survey.
a. Average of French and math scores administered at start of second grade (1996 – 96 school
year). Predicted using second-grade posttest.
b. Constructed employing factor analysis using information on ownership of durable goods
(TV, refrigerator, bicycle, and others) and housing characteristics (including drinking water
source and toilet facilities).
c. Based on principal components analysis using classroom and school attributes described in
the text.
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 Senegal
Household Education and Welfare Survey (EBMS); see text for details.
Peter Glick and David E. Sahn 105
III . EA R LY PE R FO R M A N C E AND G R A D E AT TA I N M E N T : ES T I M AT I O N
R ES ULT S
This section briefly discusses the �?rst-stage regressions used to generate the test
score residuals as these results are of intrinsic interest. The full regression
results are shown in appendix table A-1. The dependent variables are standar-
dized scores on second-grade post- and pretests. Test scores are lower if the
student is a girl, and higher for wealthier children. For the posttest score, there
is a positive impact of mother’s but not father’s education. Pretest scores are
positively associated with the experience and education of the school director
and teachers, as well as the share of teachers that are women. The teacher vari-
ables used in the pretest model are averages over different classes for the
school, since the relevant class-speci�?c information (for most children, �?rst
grade in 1994–95) is not available. Thus, these variables capture the overall
quality of teachers. The posttest regression uses actual values for teachers and
classroom supplies for the child’s second-grade class. The teacher variables are
largely not signi�?cant in this regression, with the striking exception of a large
positive interaction of student gender (being a girl) and having a female
teacher, suggesting that girls respond better to female teachers or that female
teachers make more of an effort to encourage girls. There is also a positive
impact on the posttest score of having better equipped classrooms.
In the grade attainment-ordered probits, the estimated effects of early aca-
demic ability using either predicted test score or the predicted residual were
very similar, reflecting the high correlation between these two indicators.
Despite the presence of a number of highly signi�?cant covariates, the explained
variation in the test score regressions is low. As a result, the unexplained
portion (the regression residuals) is highly correlated with the dependent vari-
able (test score). This carries over, of course, to the measurement error-
corrected predictions—the pretest score predicted from the posttest score is
highly correlated with the pretest residual predicted from the posttest residual.
Given the similarity of results, only the results for the ordered probit attain-
ment models using the predicted test score are presented (table 2).15
Four models are presented in table 2: base models using school-level
random and �?xed effects (regressions 1 and 3 in table 2) and these models
with interactions added (regressions 2 and 4). There is a very strong, highly
signi�?cant association of grade attainment with early academic success as
measured by test score at the start of second grade. The ordered probit model
estimates the parameters of the linear index function underlying the model; to
get comparative static effects, it is necessary to calculate changes in probabil-
ities using the estimates. Calculations using the random effects estimates
15. One interpretation of the low R 2s in the test score regression is that most of the variation in
scores is due to variation in unmeasured ability. However, as suggested above, it is more likely a
combination of ability and the influence of unmeasured household or school factors.
106
T A B L E 2 . Ordered Probit Models of Grade Attainment
School-level
Variable School-level random effects �?xed effects
(1) (2) (3) (4)
Intercept 3.785 (3.26)*** 4.012 (3.16)*** 3.432 (4.43)*** 3.502 (4.36)***
Second-grade pretest scorea 0.602 (5.15)*** 0.448 (2.27)** 0.689 (8.59)*** 0.587 (7.01)***
Girl 2 0.173 (0.92) 2 0.081 (0.42) 2 0.278 (1.89)* 2 0.199 (1.43)
Asset index 0.360 (2.24)** 0.371 (2.15)** 0.407 (5.24)*** 0.414 (5.35)***
Mother’s schooling 2 0.009 (0.28) 2 0.012 (0.36) 2 0.006 (0.36) 2 0.009 (0.49)
Missing mother’s schooling 2 0.142 (0.36) 2 0.112 (0.29) 2 0.128 (0.69) 2 0.071 (0.37)
Father’s schooling 0.035 (1.38) 0.035 (1.26) 0.034 (2.34)** 0.033 (2.40)**
Missing father’s schooling 2 0.357 (1.49) 2 0.357 (1.44) 2 0.343 (1.86)* 2 0.338 (1.76)*
Pretest score  asset index 0.120 (0.98) 0.131 (1.96)**
Pretest score  girl 0.481 (1.98)** 0.364 (2.51)**
Pretest score  father schooling 0.004 (0.10) 0.002 (0.07)
THE WORLD BANK ECONOMIC REVIEW
Director’s experience 0.000 (0.01) 2 0.004 (0.18)
Director has baccalaure ´ at or higher 2 0.145 (0.45) 2 0.206 (0.67)
Teachers, average experience 2 0.016 (0.66) 0.018 (0.68) 2 0.050 (2.09)** 2 0.052 (2.22)**
Teachers, share with baccalaure ´ at or higher 0.217 (0.65) 0.182 (0.54) 0.671 (1.45) 0.624 (1.38)
Teachers, share female 2 0.640 (1.44) 2 0.652 (1.40) 2 1.485 (4.57)*** 2 1.383 (4.56)***
Girl  share teachers female 0.321 (0.90) 0.179 (0.44) 0.536 (2.32)** 0.397 (1.72)*
Classroom supplies �?rst principal component 0.041 (0.36) 0.040 (0.35) 2 0.015 (0.19) 2 0.022 (0.29)
School facilities �?rst principal component 0.002 (0.01) 0.002 (0.01) 2 0.097 (0.72) 0.000 (0.00)
Rural 2 0.256 (0.48) 2 0.314 (0.53)
Rho 0.068 (1.41) 0.070 (1.36)
Number of observations 834 834 834 834
*Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level.
Note: Numbers in parentheses are aysmptotic t-statistics. Standard errors in the �?xed effects model are adjusted for clustering at the school level. See
table 1 and text for variable de�?nitions. Models also include ethnic group dummy variables, ordered probit threshold parameters, and province dummy
variables (random effects model) or school dummy variables (�?xed effects model).
a. Predicted using second-grade posttest score.
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text for details.
Peter Glick and David E. Sahn 107
indicate that at the mean of the regressors, a one standard deviation improve-
ment in the second-grade pretest score is associated with a 22 percentage
point increase in the probability of completing the sixth grade (relative to a
mean estimated probability of 56 percent). The result for the �?xed effects
model with controls for community/school is virtually the same. These are
clearly large impacts. Although, as noted, it cannot be established how much
of this association is causal, the robustness of the estimates to controls for
school and community demonstrates that they are not simply due to an
association of high achievement and attainment with better schools or other
locality-speci�?c factors.
This very large and robust association of early academic performance with
later educational success has implications for policy. It argues strongly for pol-
icies to target children who display early cognitive disadvantages. A limitation
of the analysis, however, is that it cannot say what policies should be
implemented or at what point in the child’s life they would be most effective.
For example, if poor cognitive skills early in school reflect largely prior nutri-
tional de�?ciencies, the best time to intervene is probably well before school
age, in which case it would not be very useful to rely on early school measures.
One approach to dealing with poor primary performers—grade repetition—is
assessed later in the paper.
While there is little comparable evidence from other developing
countries, the �?ndings here accord with panel-based research in industrial-
ized countries that considers how early (at school entry) disparities in chil-
dren’s skills are related to later achievement gaps (for example, Phillips,
Crouse, and Ralph 1998; Duncan and others 2007), and how academic
performance is linked to educational attainment, dropout, and labor market
success (Robertson and Symons, 1990; Currie and Thomas 2001; Maani
and Kalb 2007).
Conditional on early test score, household wealth is a strongly signi�?cant
determinant of attainment. A one standard deviation increase in the wealth
index implies an approximately 12 percentage point increase in the probability
of completing the sixth grade. As with test score impacts, this effect is essen-
tially the same with and without controls for school �?xed effects. These wealth
impacts may reflect a number of factors, including an income effect on the
demand for education and the association of wealth with inputs such as early
childhood nutrition or school supplies that improve school performance and
progression. Among other household covariates, there is a positive impact of
father’s schooling (signi�?cant only in the �?xed effects case). This is in contrast
to the tendency in the test score regressions for mother’s schooling and not
father’s to improve test score outcomes. It is possible that educated mothers
spend more time helping children with schoolwork than educated fathers do
(leading to larger maternal education effects on scores) but that schooling dur-
ation decisions largely reflect father’s preferences, which are tied to father’s
108 THE WORLD BANK ECONOMIC REVIEW
education. In the �?xed effects model but not in the random effects model, girls
go less far in school than do boys with similar early achievement.
There are few signi�?cant coef�?cients on school covariates.16 However, it is
striking that, at least in the �?xed effects case, the share of classes taught by a
female teacher has a negative effect on attainment that is larger for boys (recall
that there is variation within schools in the share of each child’s teachers who
are women). A smaller negative effect for girls is understandable considering
the apparent positive effect of female teachers on girl’s learning, but the reason
for an overall negative effect of female teachers on attainment seems puzzling
at �?rst glance. One possibility is that (male) household decision-makers believe
incorrectly that female teachers are of lower quality.17 Perhaps more probable
in view of results discussed below is that female teachers are more likely to fail
students (especially boys), which impedes school progression.
The ordered probit models in columns 2 and 4 in table 2 add interactions of
early score with assets, gender, and paternal schooling.18 There is a positive
interaction of assets and pretest score in the �?xed effects model. As noted in
section I, even if the test score represented a measure of ability that was
exogenous to parental behavior, theoretical considerations do not lead to �?rm
predictions about the existence or direction of this interaction. In both the
random effects and �?xed effects models there is a strongly signi�?cant positive
interaction of test score and being a girl, meaning that the positive effect of test
score on subsequent school investments is larger for girls. Caution is necessary
here, however, since a higher score for girls, even as early as the end of �?rst
grade, could reflect household preferences for girls’ human capital that also
lead to higher attainment. Still, the results may indicate that the returns to
schooling are more sensitive to observed ability for girls than they are for boys.
This could happen, for example, if parents expect that a daughter with strong
academic skills will be more likely (because of higher potential pay or greater
ambition) to go on to participate in the labor market or gain full-time employ-
ment. This would increase their incentive to invest further in the girl’s school-
ing, since any earnings bene�?t from an additional year of education would be
realized over a greater number of future years of work. Since male full-time
participation is more or less universal, this effect would not occur for boys (see
Glick 2008 for discussion).
16. Also included in earlier speci�?cations was distance to the nearest lower secondary school
(collected in 2003). This was also insigni�?cant. Characteristics other than distance of lower secondary
schools may affect school continuation decisions, but information on these characteristics was not
available.
17. The share of female teachers is unlikely to be acting merely as a proxy for poor school quality:
there is no consistent pattern of correlations of female share with observed quality indicators or teacher
and director background measures.
18. In initial speci�?cations, nonlinearities in the effect of test score itself were found to be
insigni�?cant.
Peter Glick and David E. Sahn 109
T A B L E 3 . Predicted Probabilities of Completing Sixth Grade for First and
Fourth Quartiles of Second-Grade Pretest Scores
First pretest score quartile
Fourth Eliminate wealth
pretest score Eliminate and paternal
quartile wealth gapsa education gaps
Model and probability (1) (2) (3) (4)
School random effects model
Probability (grade ! 6) 0.787 0.326 0.393 0.414
Difference in probabilities for 0.461 0.395 0.374
fourth and �?rst quartile
School �?xed effects model
Probability (grade ! 6) 0.799 0.386 0.464 0.486
Difference in probabilities for 0.413 0.334 0.313
fourth and �?rst quartile
Note: Table shows predicted probabilities at test score quartile means for all variables, based
on ordered probit model estimates.
a. Column 3 shows the probability for �?rst quartile after setting wealth equal to fourth quar-
tile means. Column 4 shows the probability for �?rst quartile after setting both wealth and
paternal education equal to fourth quartile means.
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
To put the importance of early skills in perspective, the estimates were used
to predict the attainment of children in different quartiles of the pretest score
distribution (table 3). Columns 1 and 2 show the predicted probabilities of
completing at least the sixth grade, evaluated at the means of the regressors for
the fourth and �?rst quartiles. The probabilities differ substantially for the two
groups, though this also reflects differences in other factors such as wealth and
parental education. Column 3 shows the probabilities for the bottom scoring
quartile after “eliminating�? wealth differences by setting the asset index for
this group to the mean of children in the highest achievement quartile. As
expected from the association of early test performance and wealth, if house-
hold resources were the same for the two groups, the gap in probabilities
would be smaller. However, the gap remains very large, closing only from 0.46
to 0.40 in the random effects case; results are similar for the �?xed effects case.
Column 4 shows that also eliminating gaps in father’s schooling further
narrows the attainment gap, but only slightly. Controlling for differences in
school factors (in the random effects models) or cluster location (�?xed effects
models) did not substantially reduce the large differences in expected attain-
ment. Thus, the large gap in attainment between early low- and high-achieving
children reflects largely differences in early academic ability rather than associ-
ations of ability with other factors influencing attainment. This should not be
taken to mean that wealth and parental education do not contribute strongly
110 THE WORLD BANK ECONOMIC REVIEW
to attainment gaps, but rather that they appear to do so primarily through
impacts early on, as captured by achievement measures in second grade.
I V. H O W D O E S T A R G E T I N G P O O R P E R F O R M E R S FOR GRADE
RE P E T I T I O N A F F E C T AT TA I N M E N T ?
Interpreted as a causal relationship, the link between early measures of cogni-
tive ability and grade attainment indicates that school investments by parents
are strongly reinforcing with respect to ability: high-achieving children are kept
in school longer. Since the indicators of early cognitive skills may incorporate
unmeasured factors that also affect later schooling, the results should be con-
sidered suggestive of such behavior rather than conclusive evidence of it. Still,
even interpreted as an association, the �?ndings indicate clearly that students
who lag behind their peers early in primary school are at substantially higher
risk of early withdrawal from school. Thus, if education authorities are com-
mitted to ensuring greater equity in educational attainment, targeting poor
early performers for remedial help would be a logical step.
Poor performers are currently targeted for one kind of remedial treatment:
grade repetition. Repetition is particularly pervasive in Francophone Africa,
where it averages 20 percent for primary students as compared with 10 percent
in Anglophone Africa and 2 percent in Organisation for Economic
Co-operation and Development countries (Michaelowa 2003). In the
Senegalese sample, 76 percent of PASEC students report having repeated at
least one primary grade, and 40 percent report two or more repetitions.
While repetition is thought to help lagging students catch up with their
peers, the impacts of repetition on either attainment (speci�?cally primary
completion) or learning have not been established. The costs of repetition,
however, are clearly large, both for the education system (in demands on tea-
chers and classroom resources) and for families, who have to �?nance
additional time in school for children to reach a given grade. If labor
markets rewarded grade attainment—or achievement of speci�?c levels such as
primary completion—rather than solely (less easily observed) human capital,
the returns to families of additional schooling would be lower than where
there is less repetition. In terms of the model discussed in section I, the mar-
ginal returns for the low-ability (repeating) child in �?gures 1 –3 would effec-
tively shift further downward relative to those for a high-ability
(nonrepeating) child—possibly explaining in part the large negative associ-
ation of low measured ability and grade attainment.
It is of interest then to see how this “targeted�? policy affects school attain-
ment in the sample. Of course, even if there are negative impacts on attain-
ment, they may be offset by improvements in learning from repeating a grade,
so information on both outcomes is needed to assess repetition policies.
Attempts to measure the impacts of repetition must deal with the fact that
repetition is not exogenous to school attainment or dropout. Repeaters would
Peter Glick and David E. Sahn 111
be expected to be relatively poor academic performers who may well have
lower attainment than nonrepeaters, aside from any positive or negative effects
of repetition itself. In a few, mostly U.S. studies, researchers have been able to
handle endogeneity by taking advantage of natural policy experiments, such as
changes in retention policy over time or across states, or regression discontinu-
ity designs made possible by the presence of an of�?cial test score threshold for
promotion (Jacob and Lefgren 2004; Allensworth 2005; Greene and Winters
2007). While this is not the case for this study, the panel data make it possible
to control for academic achievement at the time the decision is made to repeat
a student (measured by the second-grade posttest score). Conditional on a stu-
dent’s academic performance to that date, variation in the requirement that a
student repeat the grade should reflect variation in school or teacher repetition
rules or attitudes rather than parental (or child) education preferences or
effort.19 This approach is similar to that of several analyses for developing
countries of the effects of repetition on student learning (Gomes-Neto and
Hanushek 1994; Michaelowa 2003), though unlike those studies, the present
study also estimates impacts on attainment or dropout. This is made possible
by the fact that the 2003 survey, rather than being school based, interviewed
members of the PASEC cohort whether they were still in school or not and
gathered information on the last year and grade enrolled of those who had left
school.
Because the analysis is considering the effect of a repetition occurring near
the start of a child’s education, it focuses on a measure of early attainment
rather than ultimate attainment, as in the ordered probits. It estimates the
effect of second-grade test score and second-grade repetition on the probability
that the child completes grade 4, or equivalently, drops out before grade
5. This choice of dependent variable avoids censoring problems, as very few
children (under 4 percent) who are still enrolled as of 2003 have not already
progressed beyond this level. As indicated, this analysis uses the posttest rather
than the pretest, as the posttest measures skills at the time the repetition deter-
mination is made. As before, this is corrected for nonsystematic measurement
error, now using the pretest score as an instrument.
A problem with the data is attrition from the PASEC sample. Some 20
percent of the PASEC second-grade cohort of 1995–96 is not in the testing
sample in the next year—that is, in either the third or second-grade classes
visited by the survey at the end of the 1996–97 school year. There are two
possible reasons for this: the children had already dropped out, or they were
19. It is not being assumed that a student’s academic ability ( perhaps relative to that of peers, as
discussed below) is the only factor in a teacher’s decision on whether to hold the student back; in
particular, the student’s social or emotional maturity may also be considered. However, this will not
bias the estimate of the effect of repetition unless, controlling for academic achievement, this factor
would independently lead to early dropout. There is no reason to expect a (presumably temporarily)
immature child not to continue in school if this immaturity is not causing the student to lag behind
academically.
112 THE WORLD BANK ECONOMIC REVIEW
still enrolled but were not in attendance on the day the PASEC team returned
to the school. The 2003 data indicate that the large majority of these
missing children attained third grade or higher and thus belong in the second
group.
Being missing is not random with respect to academic ability: children
who were absent for the 1997 follow-up scored signi�?cantly lower on the
prior year tests than those present, and this is the case even for children who
were still in school but merely not in attendance. Since they had lower
achievement at the end of second grade, children who were absent (for either
reason) would be more likely to be repeaters. Further, for children who were
enrolled but absent during the 1997 visit, attendance was probably poorer
on average than for nonabsent children, and factors that lead to poor attend-
ance such as higher opportunity costs or lower motivation (hence negatively
affect selection into the 1996–97 sample) might also independently result in
earlier dropout. Therefore, both early (before third grade) school leaving and
simple absence may lead to selection bias when estimating the impacts of
repetition.
It is very dif�?cult to come up with valid instruments for this selection.20
Even without this, however, the data allow a test for selectivity. For children
missing in 1996–97, information is unavailable on second-grade repetition,
but the dependent variable of interest, dropout before grade �?ve, is observed. If
there is selection on unobservables, dropout should be associated with selection
(the presence of the child in 1996–97 re-survey), conditional on test score and
other measured characteristics. In probits for early dropout using the full
1995–96 sample, there is a signi�?cant negative effect of a dummy variable for
being in the 1996–97 sample, controlling for posttest score and other covari-
ates (results available from the authors). However, this effect vanishes after
dropping the small number of children who reported leaving school right after
second grade. This �?nding indicates that selection into the 1996–97 sample is
related positively to the propensity for continuing in school, but this selection
arises from children who drop out immediately rather than those who are
simply not in attendance for the 1997 test. Of course, whatever the source, the
possibility of bias must be considered. It bears noting that selection issues in
modeling school outcomes in panels with high withdrawal or absenteeism are
hardly unique to these study data, though the ability to test for selection is
unusual.
20. The EBMS collected information on health and economic shocks to the household ( parental
illness, unemployment, negative and positive harvest shocks, and other events) that can predict ultimate
school attainment (see Glick and Sahn 2009). However, these shocks are unconvincing as instruments
in the present context: they would have to predict attendance on a given day in 1996–97 but have no
independent effects on dropout in that year or the next several years. Shock-related factors that affect
attendance (such as loss of income or demands on the child’s time) will likely also influence the decision
on whether and how long to continue schooling.
Peter Glick and David E. Sahn 113
Probit regressions ( presented in appendix table A-2) on the sample of
children present in 1996–97 con�?rm that the probability of repeating second
grade is negatively and very strongly affected by the posttest score (t-statistics
greater than 5.0), whether controlling for school �?xed effects or not. A 1 stan-
dard deviation reduction in the posttest score increases the repetition prob-
ability by about 11 percent—a large impact given the mean repetition of 14
percent.21 In the �?xed effects model, most of the school dummy variables are
highly signi�?cant and there is substantial variation in the magnitudes of the
school effects. The variation across schools in repetition probabilities con-
ditional on test performance appears to be the case at least partly because tea-
chers or schools assess students relative to their peers, as many have suggested
(Bernard, Simon, and Vianou 2005). In a speci�?cation of the non-�?xed effects
model including both individual score and mean class score, the mean score
has a positive and signi�?cant impact, meaning that a student with a given level
of ability is more likely to be held back when ranked lower within his or her
class. Repetition probabilities are also higher when the teacher is a woman.
Perhaps consistent with the bene�?cial impact of female teachers on girls’ test
scores, the positive female teacher effect on repetition (controlling for test
score) is signi�?cantly smaller if the student is a girl.
Among household variables, parental education has no impact on repetition
controlling for score, but wealth is negatively associated with repetition. Since
the wealth effect persists in the within-school model, it is evident that not all of
the variation in repetition conditional on ability is due to variation across tea-
chers or schools. It is possible that wealthier parents are able to bribe teachers
to avoid repetition or, possibly, are more willing to appeal the decisions of tea-
chers (or are more effective at it). This could pose some problems for the
identi�?cation of causal repetition effects. However, wealth is an included cov-
ariate in the dropout models, and this controls for contamination from
wealth-related behavior affecting both repetition and dropout (to effectively
capture these effects, polynomials in wealth were added to the dropout models
but were not signi�?cant). It is necessary to assume that unobserved factors not
captured by wealth are not signi�?cantly affecting both the teacher repetition
decision and later attainment outcomes conditional on skills.
The probit results for early dropout for this sample, including the indicator
for being a repeater, are presented without school �?xed effects (table 4). The
�?xed effects estimation is not feasible for this analysis, because the identi�?-
cation of repetition impacts conditional on academic performance relies on
variation across classes in teacher/school repetition practices—essentially, vari-
ation in test score thresholds for passing second grade—such that children with
similar scores but in different schools will have different probabilities of being
21. This is all the more noteworthy because teacher promotion decisions are not based on the
results of the PASEC survey tests (which were graded by survey personnel for use by the PASEC project
only) but on internal tests and other assessments of the children’s skills.
114 THE WORLD BANK ECONOMIC REVIEW
T A B L E 4 . Determinants of Dropout before Fifth Grade and 1997 Test Score
Dropouta Test scoreb
Variable Coef�?cient t-statistic Coef�?cient t-statistic
Intercept 2 1.973 2 5.961*** 0.000 2 0.001
Second-grade posttest scorec 2 0.237 2 1.837* 0.804 14.618***
Repeated second grade 0.492 2.306** 2 0.204 2 1.601
Girl 0.008 0.043 0.010 0.116
Asset index 2 0.240 2 2.475** 0.066 1.221
Mother’s schooling 2 0.032 2 1.036 0.011 1.001
Missing mother’s schooling 0.040 0.161 0.018 0.132
Father ’s schooling 2 0.001 2 0.051 0.014 1.691*
Missing father’s schooling 0.609 3.233*** 0.059 0.510
Rural 0.153 0.889 2 0.045 2 0.297
Director’s experience 2 0.001 2 0.180 0.004 0.632
Director has baccalaure ´ at or higher 2 0.067 2 0.466 0.190 1.322
Teachers, average experience 0.021 2.033** 2 0.016 2 1.833*
Teacher has baccalaure ´ at or higher 0.338 1.837* 0.097 0.553*
Teachers, share female 0.387 1.556 0.156 0.768
Girl  share teachers female 2 0.382 2 0.826 2 0.011 2 0.067
Classroom supplies principal component 2 0.023 2 0.590 0.025 0.571
School facilities principal component 2 0.031 2 0.644 0.071 1.424
Number of observationsd 664 556
*Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1
percent level.
Note: See table 1 and text for variable de�?nitions. Model also includes ethnic group dummy
variables and province dummy variables. Standard errors are adjusted for clustering at the school
level.
a. Probit model estimates. Dependent variable is leaving school before �?fth grade.
b. Ordinary least squares estimates. Dependent variable is the standardized score from the end
of the 1996– 97school year (see text for details).
c. Predicted using second-grade pretest score.
d. Sample size is lower for test score model because of missing test data.
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
held back. Considering only differences among students within a class
obviously eliminates this variation, so the repeat dummy variable would do no
more than capture a possible nonlinear impact of test score on dropout.22
The results indicate that conditional on skills as captured by posttest score
(which has the expected negative effect), repeating the second grade has a large
22. The non-�?xed effects model does not control for unobserved school or community factors that
affect both repetition and dropout. It is important to remember, however, that the regressions include
the test score, which is an ideal control to capture factors affecting schooling outcomes; conditional on
test scores and the other covariates, it seems likely that the repetition-dropout relation will not be
substantially affected by unmeasured factors. It is worth recalling as well from the attainment models
that the effects of test score itself are extraordinarily robust to adding school �?xed effects.
Peter Glick and David E. Sahn 115
and signi�?cant positive impact on the probability of early dropout. The implied
increase in the probability is about 11 percent, a large impact given the rate of
dropout before �?fth grade of 15 percent. The other estimates for the probit
model are qualitatively similar to the earlier ordered probit results, though
perhaps reflecting the relatively limited variation in the dependent variable,
there are fewer signi�?cant coef�?cients. Most notably, greater wealth reduces
the probability of leaving school before grade 5.
It can be concluded, therefore, that grade repetition as a policy for poor
achievers has detrimental impacts on attainment among children in Senegal.
The study results, subject to the caveat about selection noted earlier, suggest
that children who are made to repeat a grade are less likely to complete
primary school than children with similar academic ability who are not held
back. This is not surprising, given that repetition effectively raises the
private costs of achieving a given grade level. It may be inferred that part
of the large negative effect of poor early achievement on school attainment
operates through this pathway. It is also possible that poorly educated
parents �?nd it dif�?cult to discern by other means how much their child is
learning; such parents may take a failure to be promoted as a signal that
the returns to schooling are particularly low for their child. While there
appear to be no similar studies in other African contexts that plausibly
attempt to control for repetition endogeneity, similar negative impacts on
attainment were found for Uruguay by Manacorda (2007), who used both
regression discontinuity and a natural experiment approach.23 The �?ndings
on dropout impacts from the broader U.S. literature are mixed (Grissom
and Shepard 1989; Eide and Showalter 2001; Allensworth 2005), but the
differences in contexts between the United States and Africa are in any case
very large.
As noted, there still may be a bene�?t to repetition if children who repeat a
grade learn more by staying behind than children with similar skills who are
promoted to the next, more dif�?cult grade. The PASEC test data make it poss-
ible to address this question in a straightforward way. Students who repeated
second grade in 1996–97 were tested at the end of the school year using the
same second-grade math and French posttests taken in the previous year.
Students who were in the third grade took a different test that nevertheless had
a substantial number of questions in common with the second-grade tests.
There were 38 such anchor items on the math and French tests combined, from
23. King, Orazem, and Paterno’s (2008) study using panel data from Pakistan addresses the dropout
implications of repetition using a different methodology. Decomposing the promotion probability into
the part determined by student merit or ability and the part due to other factors, they �?nd that only
student merit has a ( positive) effect on school continuation. If the merit-based part is considered a
control for student ability, the nonmerit portion essentially captures exogenous sources of variation in
repetition probabilities much in the way the repetition indicator controlling for ability does in the
present analysis. Thus, the �?nding that nonmerit factors have no effect on continuation conditional on
academic merit is in contrast to the negative effect of repetition found in the this study.
116 THE WORLD BANK ECONOMIC REVIEW
which standardized 1997 scores were calculated for each child. The determi-
nants of this score were estimated in a linear regression including the posttest
second-grade score and the repeat dummy variable as well as the other controls
used in the dropout probit.
The results show, not surprisingly, a very large and highly signi�?cant posi-
tive association of scores in 1997 with previous year scores (see table 4).
Having repeated second grade is associated with a small (0.2 standard devi-
ation) reduction in test score relative to not repeating, though this effect is
not quite signi�?cant at the 10 percent level A negative effect of repetition is
not implausible, as children left behind may feel stigmatized or get less
encouragement to do well in school (note that the negative coef�?cient does
not mean that repeating children do not increase their skills over the sub-
sequent year, only that they gain less than those with similar initial skills
who are promoted).24
In sum, for a given level of skill (test score) at the end of second grade,
having a student repeat that grade does not lead to greater test score gains rela-
tive to being promoted and may even lead to slightly reduced improvement,
though this effect is imprecisely estimated. Similar results (showing no signi�?-
cant bene�?t) reported in Michaelowa (2003) were obtained in an analysis for
all the Francophone countries in the PASEC study employing a broadly similar
methodology. These �?ndings contrast with the widespread perception among
teachers documented in CONFEMEN (2003) that repetition leads to greater
improvement. The �?ndings suggest that there is little academic bene�?t to rep-
etition to justify the costs to schools and households.
V. C O N C L U S I O N
The estimates and simulations in this study indicate that early academic per-
formance is an important predictor of school attainment in a cohort of children
in Senegal. Conditional on measured skills at the start of second grade, gender,
father’s education, and especially household wealth also have impacts on
attainment, in expected directions. There is some evidence as well that the posi-
tive effects of early achievement are greater for wealthy children and for girls.
The strong association of early achievement and later educational attainment
points to the bene�?ts of measures directed to lagging children to avoid later dis-
advantages in skills and labor market outcomes.25 However, the one (quite
costly) policy that is currently directed at such children, grade repetition, tends
24. A negative association might also suggest that unobservable factors such as ability or motivation
are affecting both repetition and subsequent test performance. Recall, however, that as with the dropout
model, the estimation conditions on test outcomes immediately prior to the repetition decision
(second-grade posttest score), which should control for these factors.
25. Though as noted in section III, the appropriate measures and their timing will depend on the
source of the early disadvantage.
Peter Glick and David E. Sahn 117
to exacerbate the negative impacts of poor early school performance on
attainment.
Since evidence from a number of African countries indicates that there is a
premium to primary (and secondary) completion in terms of earnings or
access to formal sector employment, the �?nding that having lagging learners
repeat grades increases premature primary dropout is important for policy.
Given this �?nding as well as the evidence that being held back does not
improve learning outcomes relative to being promoted, it would likely be sig-
ni�?cantly more effective to devise other measures to improve the skills of
lagging children rather than simply having them repeat grades. One measure
might be to hire local secondary school graduates to tutor such children,
along the lines of a successful intervention evaluated in India by Banerjee
and others (2007).
FUNDING
The authors are grateful to the United States Agency for International
Development, the World Bank, the United Nations Children’s Fund, and
Institut National de la Recherche Agronomique (France) for project funding.
APPENDIX
TABLE A-1. Determinants of Second-grade Pre- and Posttest Scores
Pretest Posttest
Variable Coef�?cient t-statistic Coef�?cient t-statistic
Intercept 2 1.302 2 3.735*** 2 0.314 2 1.020
Girl 2 0.296 2 3.180*** 2 0.361 2 3.909***
Asset index 0.152 2.928*** 0.209 2.947***
Mother’s schooling 0.024 1.459 0.032 2.187**
Missing mother’s schooling 0.143 1.013 0.100 0.645
Father’s schooling 0.005 0.487 0.007 0.664
Missing father’s schooling 0.118 0.867 0.142 1.151
Rural 0.344 2.174** 0.098 0.710
Director’s experience 0.021 2.340** 0.014 1.388
Director has baccalaure ´ at or higher 0.386 3.135*** 2 0.043 2 0.309
Teacher experiencea 0.053 4.642*** 0.011 1.546
Teacher has baccalaure ´ at or highera 0.301 2.072** 0.237 1.485
Teacher femalea 0.398 1.947* 2 0.182 2 1.045
Missing teacher information in 1995– 96 1.056 3.311***
Girl  teachers share female 0.059 0.351 0.541 3.764***
Classroom supplies �?rst principal 0.009 0.215 0.094 1.996**
componenta
School facilities �?rst principal component 2 0.005 2 0.106 0.071 1.213
Adjusted R 2 0.214 0.177
(Continued )
118 THE WORLD BANK ECONOMIC REVIEW
TABLE A-1 Continued
Pretest Posttest
Variable Coef�?cient t-statistic Coef�?cient t-statistic
Number of observations 787 756
*Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1
percent level.
Note: Pretest and posttest scores are the standardized average scores on French and math tests
administered at the beginning and end of second grade, respectively. The models also include
ethnic group dummy variables and province dummy variables. Standard errors are adjusted for
clustering at the school level. The varying sample sizes reflect the number of students in the
PASEC cohort present when the tests were given. See table 1 and text for variable de�?nitions.
a. For the posttest regression, the teacher and classroom supplies variables are the values for
the child’s second-grade class (the year ending with the posttest). For the pretest regression, they
are simple averages over grades 2 – 4 in the school.
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
TABLE A-2. Determinants of Second-Grade Repetition: Probit Results
With school �?xed effects
Variable Coef�?cient t-statistic Coef�?cient t-statistic
Intercept 2 0.766 2 1.808* 2 2.094 2 4.799***
Second-grade posttest scorea 2 1.295 2 6.718*** 2 2.186 2 7.323***
Girl 0.370 2.178** 0.409 1.707*
Asset index 2 0.212 2 1.679* 2 0.469 2 2.239**
Missing mother’s schooling 2 0.209 2 0.570 0.209 0.352
Mother’s schooling 2 0.005 2 0.147 0.026 0.590
Father’s schooling 2 0.014 2 0.660 2 0.025 2 0.701
Missing father’s schooling 2 0.147 2 0.636 2 0.206 2 0.535
Rural 2 0.799 2 2.370**
Director’s experience 2 0.004 2 0.251
Director has baccalaure ´ at or higher 2 0.316 2 1.296
Teacher has baccalaure ´ at or higher 2 0.249 2 0.997
Teacher female 1.026 2.968***
Girl  female teacher 2 1.203 2 3.990*** 2 1.548 2 3.731***
Classroom supplies principal component 2 0.446 2 3.352***
School facilities principal component 2 0.146 2 1.179
Number of observations 664 664
*Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1
percent level.
Note: For �?xed effects model, school dummy variables and ethnicity dummy variables not
shown. For non-�?xed effects model, ethnicity and province dummy variables not shown.
Standard errors are adjusted for clustering at the school level. Teacher and classroom supplies
variables refer to the second-grade class.
a. Predicted using second-grade pretest score.
Source: Authors’ analysis based on data from the PASEC surveys and the 2003 EBMS; see text
for details.
Peter Glick and David E. Sahn 119
REFERENCES
Alderman, H., J.R. Behrman, D. Ross, and R. Sabot. 1996. “Decomposing the Gender Gap in
Cognitive Skills in a Poor Rural Economy.�? Journal of Human Resources 31(1):229– 54.
Allensworth, Elaine M. 2005. “Dropout Rates after High-Stakes Testing in Elementary School: A Study
of the Contradictory Effects of Chicago’s Efforts to End Social Promotion.�? Educational Evaluation
and Policy Analysis 27(4):341–64.
Banerjee, Abhijit V., Shawn Cole, Esther Duflo, and Leigh Linden. 2007. “Remedying Education:
Evidence from Two Randomized Experiments in India.�? The Quarterly Journal of Economics
122(3):1235 –64.
Behrman, J. 1997. “Intrahousehold Distribution and the Family.�? In M. Rosenzweig and O. Stark eds.,
Handbook of Population and Family Economics. Amsterdam: North-Holland.
Bernard, J.-M., O. Simon, and K. Vianou. 2005. Le redoublement: mirage de l’ecole africaine?
` mes e
Programme d’analyse des syste ´ rence des ministres de
´ ducatifs de la CONFEMEN. : La Confe
l’education.
Bound, J., C. Brown, and N. Mathiowetz. 2001. “Measurement Error in Survey Data.�?
In J.J. Heckman, and E. Leamer eds., Handbook of Econometrics. Vol 5. Amsterdam: North-Holland.
Butler, J.S., and R. Mof�?t. 1982. “A Computationally Ef�?cient Quadrature Procedure for a One-Factor
Multinomial Probit Model.�? Econometrica 50(3):761– 4.
CONFEMEN (La Confe ´ rence des ministres de l’education). 1999. “Les facteurs de l’ef�?cacite´ de
l’enseignement primaire: donne ´ sultats sur cinq pays d’Afrique et de l’Oce
´ es et re ´ an Indien.�?
CD-ROM. La Confe ´ rence des ministres de l’education, Dakar.
CONFEMEN PASEC (Le programme d’analyse des syste ` mes educatifs des etats et gouvernements de La
Confe ´ rence des ministres de l’education). 2003. Le redoublement: pratiques et conse ´ quences;
´ sultats de l’analyse longitudinale des donne
re ´ ne
´ es du Se ´ gal. Dakar: Le programme d’analyse des sys-
` mes educatifs des etats et gouvernements de La Confe
te ´ rence des ministres de l’education.
Currie, J., and D. Thomas. 2001. “Early Test Scores, School Quality, and SES: Long Run Effects on
Wages and Employment Outcomes.�? In Solomon Polachek eds., Worker Wellbeing in a Changing
Labor Market (Research in Labor Economics Vol 20). Amsterdam, New York: Elsevier JAI.
Duncan, Gregg., Amy Claessens, Aletha C. Huston, Linda S. Pagini, Mimi Engel, Holly Sexton,
and Chantelle J. Dowsett et al. 2007. “School Readiness and Later Achievement.�? Developmental
Psychology 43(6):1428–46.
Eide, Eric R., and Mark H. Showalter. 2001. “The Effect of Grade Retention on Educational and Labor
Market Outcomes.�? Economics of Education Review 20(6):563–76.
Garg, A., and J. Morduch. 1998. “Sibling Rivalry and the Gender Gap: Evidence from Child Health
Outcomes in Ghana.�? Journal of Population Economics 11(4):471–93.
Glewwe, Paul., and Hanan Jacoby. 1994. “Student Achievement and Schooling Choice in Low-Income
Countries: Evidence from Ghana.�? Journal of Human Resources 29(3):843–64.
Glewwe, P., and M. Kremer. 2006. “Schools, Teachers, and Education Outcomes in Developing
Countries.�? In E. Hanushek and F. Welch eds., Handbook on the Economics of Education.
Amsterdam: North-Holland.
Glewwe, P., and E. Miguel. 2008. “The Impact of Child Health and Nutrition on Education in Less
Developed Countries.�? In T. Paul Schultz and John Strauss eds., Handbook of Development
Economics. Amsterdam: North-Holland.
Glick, P. 2008. “What Policies Will Reduce Gender Schooling Gaps in Developing Countries: Evidence
and Interpretation.�? World Development 36(9):1623–46.
Glick, P., and D.E. Sahn. 2000. “Schooling of Girls and Boys in a West African Country: The Effects of
Parental Education, Income, and Household Structure.�? Economics of Education Review 19(1):
63– 87.
120 THE WORLD BANK ECONOMIC REVIEW
———. 2009. “Cognitive Skills among Children in Senegal: Disentangling the Roles of Schooling and
Family Background.�? Economics of Education Review 28(2):178– 88.
Gomes-Neto, J., and E. Hanushek. 1994. “Causes and Consequences of Grade Repetition: Evidence
from Brazil.�? Economic Development and Cultural Change 43(1):117– 48.
Greene, W.H. 2002. “The Behavior of the Fixed Effects Estimator in Nonlinear Models.�? New York
University, Department of Economics, Stern School of Business, New York.
Greene, Jay P., and Marcus A. Winters. 2007. “Revisiting Grade Retention: An Evaluation of Florida’s
Test Based Retention Policy.�? Education Finance and Policy 2(4):319–40.
Grissom, James B., and Lorie A. Shepard. 1989. “Repeating and Dropping Out of School.�?
In L.A. Shepard and M.L. Smith eds., Flunking Grades: Research and Policies on Retention.
London: Falmer Press.
Jacob, B., and L. Lefgren. 2004. “Remedial Education and Student Achievement: A Regression—
Discontinuity Analysis.�? Review of Economics and Statistics 86(1):226– 44.
King, E., P. Orazem, and E. Paterno. 2008. “Promotion with and without Learning: Effects on Student
Dropout.�? Iowa State University, Department of Economics, Ames, Iowa.
Kingdon, Geeta. 1996. “The Quality and Ef�?ciency of Public and Private Schools: A Case Study of
Urban India.�? Oxford Bulletin of Economics and Statistics 58(1):55– 80.
Maani, Sholeh A., and Guyonne Kalb. 2007. “Academic Performance, Childhood Economic Resources,
and the Choice to Leave School at Age 16.�? Economics of Education Review 26(3):361–74.
Manacorda, Marco. 2007. The Cost of Grade Retention. Centre for Economic Performance Discussion
Paper 878. London School of Economics and Political Science, London, U.K.
Michaelowa, K. 2003. “Determinants of Primary Education Quality: What Can We Learn from PASEC
for Francophone Sub-Saharan Africa?�? Association for the Development of Education in Africa,
Paris.
Phillips, Meredith, James Crouse, and John Ralph. 1998. “Does the Black-White Test Score Gap Widen
after Children Enter School?�? In Christopher Jencks and Meredith Phillips eds., The Black-White
Test Score Gap. Washington, DC: The Brookings Institute.
Pitt, Mark M., Mark R. Rosenzweig, and Md. Nazmul Hassan. 1990. “Productivity, Health, and
Inequality in the Intrahousehold Distribution of Food in Low-Income Countries.�? American
Economic Review 80(5):1139 –56.
Robertson, Donald, and James Symons. 1990. “The Occupational Choice of British Children.�? The
Economic Journal 100(402):828– 41.
Sahn, David E., and David Stifel. 2003. “Exploring Alternative Measures of Welfare in the Absence of
Expenditure Data.�? Review of Income and Wealth 49(4):463–89.
Wooldridge, J. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, Mass.: MIT
Press.