Journal of Development Economics 69 (2002) 103 – 128
www.elsevier.com/locate/econbase
Raising primary school enrolment in
developing countries
The relative importance of supply and demand
Sudhanshu Handa
Inter-American Development Bank, 1300 New York Avenue, Northwest, Washington, DC 20577, USA
Received 1 June 1999; accepted 1 August 2001
Abstract
Few policies are as universally accepted as raising primary school enrolment in developing
countries, but the policy levers for achieving this goal are not straightforward. This paper merges
household survey data with detailed school supply characteristics from official sources, in order to
estimate the relative impact of demand and supply side determinants of rural primary school
enrolment in Mozambique. Policy simulations based on a set of ‘plausible’ interventions show that in
rural Mozambique, building more schools or raising adult literacy will have a larger impact on
primary school enrolment rates than interventions that raise household income. When relative costs
are considered, adult literacy campaigns are nearly 10 times more cost-effective than the income
intervention and 1.5 to 2.5 times better than building more schools.
D 2002 Elsevier Science B.V. All rights reserved.
JEL classification: D1; I0; I2; O1
Keywords: Primary education; School supply; Africa
1. Introduction
Few policies, if any, are as universally accepted as that of raising primary school
enrolment in poor countries. Virtually every World Development Report published
annually by the World Bank has recognized, in one form or another, the importance of
primary schooling as an input to the social and economic progress of poor countries.1 And
E-mail address: sudhanshuh@iadb.org (S. Handa).
1
Within the overall policy goal of raising primary school enrolment, raising girls’ enrolment has received
special attention, due to the large positive externalities of female education on children and adult health, fertility,
and infant mortality.
0304-3878/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 0 4 - 3 8 7 8 ( 0 2 ) 0 0 0 5 5 - X
104
S. Handa / Journal of Development Economics 69 (2002) 103–128
within the academic literature, a host of studies has documented the market and nonmarket
return that comes from completing primary schooling, both in poor and rich countries
alike.2
However, raising primary school enrolment in developing countries is easier said than
done. The relative importance of school supply versus household demand factors remains
controversial, with serious implications for education policy.3 For example, if children’s
enrolment rates are not responsive to local school infrastructure, government interventions
aimed at increasing access to schools will have very limited impact on overall schooling
levels, thus effectively reducing the set of options available to policymakers. And even if
regional variations in schooling infrastructure can be related to household schooling
choices, as several studies have shown,4 efficient policy decisions require knowledge of
the particular dimensions of school infrastructure that matter most. This latter issue is
contentious in both developing and developed countries alike, and has been the topic of
several recent articles seeking to measure the type of schooling infrastructure (access,
quality, etc.) that makes a difference for household schooling choices.5
This study makes three main contributions to the literature on primary school enrolment
policies and school infrastructure in developing countries. First, the impact of school
characteristics on household primary school enrolment decisions are measured using a
diverse set of school ‘quality’ indicators. Aside from information on distance to the nearest
school, detailed information on school characteristics is hard to find in developing
countries, and as a result, the available published literature is small relative to that for
developed countries.6 This study thus provides an additional set of estimates with which to
assess the role of specific supply side factors in determining student outcomes. Moreover,
school characteristics are measured with the actual data that Mozambique’s Ministry of
Education uses to formulate its regional and national targets, and to develop its 5-year
plans, thus enhancing the policy relevance of the work. Second, unlike most previous
studies in this area, the interaction between school and household characteristics is
explored to see if complementarity or substitutability exists between these two sets of
factors in determining school enrolment.7 The existence of significant interactions can
provide important clues about who benefits the most from school supply interventions, and
2
For developing countries see Glewwe (1999), Handa (1999), and Lam and Duryea (1999). For developed
countries, see Rosenzweig and Schultz (1982).
3
See Simmons and Alexander (1978) for a discussion of this issue and review of the literature.
4
These studies show that community or regional fixed effects are significant determinants of household
schooling choices. For example, see Pradhan (1998) for Indonesia, Handa (1996) for Jamaica, and Alderman et al.
(1996) for Pakistan.
5
Recent studies that measure the effect of various school characteristics in developing countries include Lavy
(1996) and Glewwe et al. (1995); for developed countries, see Card and Kruger (1992), Betts (1995), and
Golhaber and Brewer (1997). The overall importance of school quality is discussed by Hedges et al. (1994),
Hanushek (1995), and Kremer (1995).
6
Recent studies that provide estimates of detailed school ‘quality’ indicators on student educational
achievement in developing countries include Glewwe and Jacoby (1994) for Ghana, Glewwe et al. (1995) for
Jamaica, and Tan et al. (1997) for the Philippines.
7
Birdsall (1985) is one of the few studies that also looks at interactions among supply and demand factors in
determining schooling outcomes.
S. Handa / Journal of Development Economics 69 (2002) 103–128
105
where programs or resources should be placed in order to benefit the most vulnerable. For
example, a significant positive interaction between household income and school access in
a particular region implies that building additional schools in this region will benefit richer
households more than poorer ones. However, policymakers may want to target regions
where the poor are more likely to benefit from the provision of schools, hence, knowledge
of specific interactions can provide useful information to help prioritize program placement.
Third, policy simulations are presented based on separate and ‘plausible’ supply and
demand side interventions, and used to evaluate which type of intervention will have the
largest impact on primary school enrolment rates. The standard policy analysis contained
in studies on school supply typically evaluates the effect on the outcome of interest (for
example, test scores or grade attainment) for a given change in a statistically significant
school supply variable (e.g. travel time to school) by multiplying the given change by the
relevant coefficient. Although statistically valid, the policy relevance of this exercise can
be enhanced significantly by recognizing that governments in the short or medium run
cannot supply 10 more books to every school, or fix every leaking roof in the country. A
typical policy intervention in the short run will involve adding more teachers in some
regions and not others, or building a few schools in a few regions. This paper estimates the
change in primary school enrolment that would come about from a set of more realistic
interventions such as building a few schools in specific regions, or targeting adult
education or income-generating programs only among the poorest households. These
simulations arguably provide a better picture of the expected benefits of the type of
interventions available to developing country governments in the short and medium term.
The data used in this paper are from Mozambique, a country that has suffered from over
25 years of armed conflict, and that is acknowledged by development experts as one of the
world’s poorest. Estimates from the rural region of this country show that both demand
and supply factors are important determinants of primary school enrolment. On the supply
side, dimensions of school quality (the number of trained teachers) and access all have
significant effects on household enrolment decisions. However, the policy simulations
show that school access on the supply side, and adult education on the demand side, are
the biggest factors in impeding primary school enrolment. When relative costs are
considered, interventions that raise adult literacy turn out to be the most efficient
alternative for raising primary school enrolment in rural Mozambique.
2. Data and description of study area
2.1. The study area
Mozambique is one of only a handful of African countries that was colonized by the
Portuguese, and by all accounts, the period of colonization was extremely repressive for
native Africans. Only a select few assimilados were allowed access to the social and
economic benefits that the colonial rulers enjoyed, and independence came in 1975 only
after a long war of independence and a change in government in Portugal. Almost
immediately after independence, the new Mozambique entered an even more brutal civil
106
S. Handa / Journal of Development Economics 69 (2002) 103–128
war, instigated by guerrillas backed by neighboring South Africa. This war of destabilization resulted in thousands of land mines being placed in rural areas in the central and
northern parts of the country. Hundreds of thousands of Mozambicans fled the countryside
for the urban centers or neighboring Malawi and Zimbabwe. The signing of the peace
treaty in Rome in 1992 essentially marked the second independence of Mozambique, but
the over 25 years of armed conflict have left a huge hole in social and economic
infrastructure that require immediate attention in order for Mozambique to realize
sustainable growth and reduce poverty. A recent study by the Ministry of Finance
(1998) estimates that 70% of the population live below the poverty line, with poverty
rates even higher in the central and northern zones that suffered most from the civil war.
2.2. The national education system
The national education system’s general education program is divided into two levels—
primary and secondary. Primary education consists of 7 years of schooling divided into
two levels, Level 1 up to Grade 5 (escola primaria do primeiro grãu or EP1) and Level 2
from Grade 6 to 7 (escola primária do segundo grãu or EP2). Secondary education
consists of 5 years also divided into two levels or cycles, first cycle secondary from Grade
8 to 10 (escola secundária geral do primeiro grãu or ESG1) and second cycle secondary
from Grade 11 to 12 (escola secundária geral do segundo grãu or ESG2).
Unlike most African countries, entrance into successively higher levels of schooling is
not based on national examinations, but on actual grades and age. Among students with
the same grades, those who are younger (and therefore either started on time or did not
repeat as often) are given priority. Access to EP1 in rural areas, and other (higher) levels
through out the country, is supply constrained. Fees do not exist in public lower primary
schools, but there is an annual matriculation fee of approximately US$5. Private EP1
school fees can range from US$150 to US$600 per year depending on ownership structure
and facilities provided.
2.3. Household data
The household data used in this paper come from the first post-war national household
survey of Mozambique undertaken in 1996/1997 by the National Statistical Institute—the
Inquérito Nacional Aos Agregados Familiares Sobre As Condicß ões de Vida (IAF). The
IAF is a multipurpose household survey that contains detailed information on consumption
expenditures, as well as modules on health (both adult and child), education, employment,
demographic composition, and a community questionnaire for rural areas describing local
infrastructure.8
The IAF is a three-staged stratified sample. Stage 1 is the 11 provinces of the country,
Stage 2 is the localidad (bairro in urban areas), and in Stage 3, households are selected
8
This data set has been used by the International Food Policy Research Institute (IFPRI) in collaboration
with the Mozambican Ministry of Finance to construct a national poverty line and to develop a poverty profile of
Mozambique (Ministry of Finance, 1998).
S. Handa / Journal of Development Economics 69 (2002) 103–128
107
Table 1
Adult literacy rates by age group (%)
18 – 65 years
All
Male
Female
Poor
66 – 99 years
Rural
Urban
Mozambique
Rural
Urban
Mozambique
32.0
52.3
15.7
31.2
71.0
85.1
57.6
61.8
40.0
59.3
23.6
36.6
29.7
42.9
17.5
28.4
69.2
78.4
60.2
59.1
37.7
50.3
26.0
34.2
from villages (or blocks in urban areas). The primary sampling unit is therefore the
localidad, and variance estimates provided in this paper account for the sample design of
the survey. The full survey covers approximately 42,000 individuals residing in 8250
households.
Tables 1 and 2 provide basic indicators of adult and child schooling calculated from the
IAF data set. Only 40% of adults aged 18 –65 can read or write; for women, the literacy
rate is even lower (24%). Note that women in rural areas have the lowest literacy rates—
16% for the 18 – 65 age group. Table 2 indicates that there are significant signs of
improvement. The net enrolment rate for primary school is 49%, and is slightly lower for
girls (45%) and rural children (44%).
2.4. School data
Information on school infrastructure in the IAF is limited to whether or not a rural
village contains a primary school. Detailed information on school characteristics has been
gathered from the Direccß ão de Planificacß ão of the Mozambique Ministry of Education
(MINED). Since 1992 MINED has administered a beginning and end-of-academic-year
questionnaire to each school in the country, soliciting information on enrolment, teachers,
teacher qualifications, pass rates, and building characteristics. This information is used by
MINED to create and keep track of its internal performance indicators. Coverage is
excellent, with over 90% of schools returning questionnaires; summaries of these data are
published in an annual report by the MINED entitled Educational Indicators (Republic of
Mozambique, various years).
Raw data from these school surveys for 1995 and 1996 were acquired from MINED
and were merged at the administrative post level with rural households from the IAF
survey.9 The analysis presented below focuses on the enrolment decision of rural children
(representing 80% of the primary school age children in Mozambique) in order to exploit
the small information on rural village level schooling availability provided in the IAF; in
all there are 634 villages in the IAF, distributed across 175 administrative posts, 112
districts, and the 10 provinces of the country (excluding the province of Maputo City).10
9
There are three levels of local administration in Mozambique: province, district, and administrative post.
The school level data are therefore aggregated to the lowest administrative unit possible.
10
In 17 cases, MINED did not have any school information for an administrative post found in the IAF. In
these cases, school information from a bordering administrative post was used.
108
S. Handa / Journal of Development Economics 69 (2002) 103–128
Table 2
Children’s current enrolment by age group (%)
7 – 11 years
All
Male
Female
Poor
12 – 17 years
Rural
Urban
Mozambique
Rural
Urban
Mozambique
43.9
49.1
39.0
41.7
70.7
73.5
68.0
63.3
49.2
53.9
44.7
45.5
43.3
51.5
33.2
42.3
63.5
65.6
61.4
54.9
48.0
54.5
40.3
44.8
2.5. Choice of school characteristics
MINED divides its educational performance indicators into three groups, measuring
coverage, quality, and efficiency of the school system, and I follow this classification
where appropriate in order for the results to be of policy relevance to the Government of
Mozambique.11 MINED has developed a set of indicators to measure each of the three
dimensions of the educational system—where possible, I used these same indicators in the
regression analysis, although there is a high degree of collinearity among the indicators,
both across and within the three dimensions of coverage, quality, and efficiency.
The basic quality indicators used by MINED are the number of trained teachers
working in the system, average class size, and the pupil – teacher ratio. The number of
trained teachers in the administrative post was used as the basic indicator of teacher
quality. However, I also find that gender of the teachers matters, and so show some results
that measure the proportion of female trained teachers in the administrative post. In
addition to teacher training, the average pupil – teacher ratio for schools in the administrative post was included. Class size is not used because many schools in Mozambique
are run on a shift system, and so smaller class sizes can be achieved by creating two shifts,
but with only a small number of additional teachers (Case and Deaton, 1999 report the
same phenomenon for South Africa).
School coverage is measured by the number of Level 1 primary schools (EP1) in the
administrative post. Given the large variations in the building structure of schools in
Mozambique, and evidence from other developing countries on the importance of building
characteristics (e.g. Glewwe and Jacoby, 1994), the number of school rooms made of
cement in the administrative post was also included. All these school supply variables are
measured at the administrative post level, so each household in the administrative post will
have the same school infrastructure characteristics. Also included is the (log of) travel time
to the nearest EP1 school to the village, taken from the IAF community questionnaire—
this controls for the very important travel time cost component of school attendance and
also allows for some village variation in school infrastructure. As in other sub-Saharan
African countries, girls’ schooling rates lag behind those of boys in Mozambique and are
thus of particular policy importance. I allow the impact of all school infrastructure
variables to differ by gender, estimating separate models for boys and girls.
11
I do not look at the impact of schooling efficiency, defined by MINED as the pass rate, since this can also
be interpreted as a school outcome indicator and not an input indicator.
S. Handa / Journal of Development Economics 69 (2002) 103–128
109
Table 3
Means for administrative post school characteristics
Mean
Coverage or access indicators
No. of EP1 schools
No. of cement rooms
EP2 school exists
Secondary school exists
21
22
0.59
0.20
Quality indicators
No. of trained teachers
No. of female trained teachers/total number of teachers
No. of female teachers/total number of teachers
No. of trained female teachers/total number of female teachers
66
0.08
0.37
0.15
Efficiency indicators
Overall pass rate
Female pass rate
Male pass rate
Portuguese pass rate
Mathematics pass rate
0.64
0.57
0.68
0.66
0.68
Data taken from survey of schools conducted by the Ministry of Education.
Table 3 provides means of the school supply variables used in the regression analysis.
These means are calculated over the 175 administrative posts found in the rural sample of
the IAF, and show that the mean number of EP1 schools is 21, with an average of one
cement room per school. Only 59% of the administrative posts have a Level 2 primary
school (EP2) school, and only 20% have a secondary school.
3. Econometric model, sample, and results
3.1. Econometric model and sample
The impact of school characteristics on household schooling decisions is measured via
reduced form demand equations for children’s schooling of the form
Si ¼ FðXc ; Xh ; Xs ; uÞ
ð1Þ
where Xc are characteristics of the individual child (age), Xh are household characteristics
that capture access to resources, differences in taste for schooling, and opportunity costs,
Xs is the vector of school infrastructural characteristics discussed above, and u is a random
error term with the usual assumptions.12 The household level variables included in the
model are the age and sex of the head, whether the head is literate, whether any adult
household member has completed Grade 7 (EP2), and whether any adult female has
completed Grade 5 (EP1). Household resources are measured with per capita daily
12
See Strauss and Thomas (1995) for a review of this methodology.
110
S. Handa / Journal of Development Economics 69 (2002) 103–128
expenditure on all goods and services including home production. This is treated as
endogenous following the recommendation of Rivers and Vuong (1988), using the cluster
median expenditure for identification.13 Also included are measures of farm assets and
production such as total land holdings, access to irrigation, agricultural commercialization,
and provincial dummy variables.
For Mozambique (and most African countries), raising primary school enrolment rates
is a priority and the focus is therefore on the analysis of school supply effects on the
primary school enrolment decisions of rural households. The sample is children of primary
school age (7– 11 years old), and the dependent variable is whether the child was currently
enrolled in school at the time of the survey. Means for these variables are presented in
Table A1 of the Appendix A.
3.2. Placement of school infrastructure
The analysis of the impact of school infrastructure on school enrolment runs the risk of
confounding cause and effect if households with a greater preference for schooling are able
to move to areas with better schooling quality. In the United States, for example,
households demonstrate preference for schooling quality through higher property prices
in districts with better schools. In Mozambique and other poor countries, allocation of
infrastructure such as school or health services may be influenced by local demand for
services. In such cases, regression estimates that do not account for endogenous program
placement will overstate the impact of school characteristics on household educational
choices. On the other hand, if policymakers purposely place programs in regions where
school outcomes are low, then standard regression estimates will lead to underestimates of
the true program effect.
Mozambique’s history of armed conflict led to destruction of physical infrastructure
including schools, roads, and health centers, and formal provision of educational centers
by the state was limited to the southern part of the country and to mostly urban zones.
During this period, very few new schools were constructed, and some were constructed
through community initiatives, which would reflect community preferences for schooling.
Since the peace accord in 1991 and the general elections of 1994, there has been a rapid
increase in the number of schools constructed in the rural areas, both due to Government
and NGO interventions. This is corroborated by the IAF community survey, which reports
that nearly half of all primary schools were built after 1992.
The education budget is distributed among the provincial directorates of education, who
allocate resources to its districts based on planning and need as articulated by the district
education directorates. In discussions with staff at the National Planning Directorate of
MINED, considerable scepticism was displayed about the ability of parents and others at
the village level to influence school placement and quality. This feeling was also expressed
by primary school teachers interviewed by the author in urban Maputo, who felt that
13
Household consumption decisions also affect leisure consumption, and are made jointly with schooling
decisions. Due to this simultaneity problem, median per capita consumption of the village is used to instrument
household consumption. This variable is highly correlated with household consumption: the simple OLS
coefficient of log (consumption) on log (cluster median consumption) is 0.88.
S. Handa / Journal of Development Economics 69 (2002) 103–128
111
parents had very little influence on how schools were run or how resources were allocated.
Presumably, this would be even more so in rural areas which are poorer and where families
are more dispersed.
As mentioned earlier, there has been considerable rehabilitation of social infrastructure
in rural Mozambique since the General Peace Accord in 1991. In the IAF sample of 634
villages, 68% report having a primary school. Of those schools that report a date of
construction (82% of cases), 42% indicate that the school was built after the signing of the
peace treaty. What determined the placement of these relatively ‘new’ schools in rural
areas of Mozambique? To evaluate the extent to which endogenous placement might bias
the estimates of program effects, the average characteristics of villages with a ‘recent’
school (i.e. built since the war) and villages without a school were compared to see if
village characteristics are sufficient to explain program placement.
I constructed program exposure time as (1997 t) where t is the date the school was
built in the village. The exposure time of villages who had a school built in the year of the
survey (1996), for example, is thus 1, while villages without a school are given an
exposure time of 0, and the resulting variable is regressed on a set of village level variables
including median village consumption expenditure, the proportion of household heads that
are literate, the proportion of households with an adult with EP2, and the proportion of
households with a female adult who has completed EP1. Since geographic location is often
an important determination of program placement, I also included the distance (km) to the
district capital, and the distance to the nearest ‘good’ road.
Ordered probit estimates of the village level determinants of program exposure are
presented in Table 4. Column 1 shows that none of the village level SES variables or the
distance variables are able to predict school placement since Mozambique’s reconstruction. Column 2 of Table 2 adds provincial dummy variables to the equation, and these
results show some significant regional variation in program placement, but the village
level characteristics remain insignificant in determining placement.
Another way that parents can influence programs is by demanding better quality.
Among recently constructed schools, are there systematic quality differences that vary by
household characteristics? To answer this question, we must go up to the administrative
post level, which is the lowest level at which school quality information can be merged
with IAF villages. Now, only those administrative posts that contain a recently (since
1991) constructed school are selected and checked to see whether school quality, measured
by the average pupil –teacher ratio and the average proportion of teachers with training,
varies according to the socioeconomic status of households in the administrative post.14
The socioeconomic variables are the same as those used earlier for the village level
analysis, and since the level of aggregation is higher, the average distance (of villages in
the administrative post) to the provincial capital was used to capture geographic targeting.
Results of this analysis are presented in Table 5, and for either measure of school
quality, the F test at the bottom of the table fails to reject the null hypothesis that the set of
14
It is important to realize that I cannot do this for all administrative posts. Administrative posts that have
had schools for many years (high exposure) will probably also have higher rates of adult literacy and primary
school completion, leading to a positive correlation between school placement or quality, and household
socioeconomic status.
112
S. Handa / Journal of Development Economics 69 (2002) 103–128
Table 4
Ordered probit estimates of years since primary school built in village
(1)
Coefficient
Median village consumption
Proportion of heads literate
Proportion of households
with female with EP1
Proportion Of households
with adult with EP2
Distance to ‘good’ road
Distance to district capital
Niassa
Cabo Delgado
Nampula
Zambezia
Tete
Manica
Sofala
Inhambane
Gaza
Log likelihood
Observations (villages)
(2)
z-statistic
Coefficient
z-statistic
0.000
0.037
0.017
1.12
0.14
0.03
0.000
0.012
0.279
0.20
0.04
0.47
0.129
0.22
0.035
0.06
0.001
0.003
0.84
0.60
0.001
0.008
0.616
1.397
0.964
0.518
0.646
1.102
0.771
0.312
1.244
374
0.50
1.59
1.44
3.03
2.44
1.32
1.55
2.67
1.74
0.75
2.85
380
266
Sample is villages that have no school or that had one built after 1991. Dependent variable is equal to 0 if village
has no school, and equal to (1997 t) if has school, where t is the year the school was built.
SES variables are jointly equal to 0. There is some indication (at the 10% significance
level) that median consumption in the administrative post is negatively correlated with the
pupil –teacher ratio (the higher the median consumption, the worse the ratio), while there
continues to be significant variation across provinces in school quality.
Another way to assess the randomness of placement rules is through a difference-indifferences approach that uses cohorts to construct control groups, as in Duflo (1999). Since
there was a major expansion in school infrastructure after 1993, children of primary school
age before 1993 faced very different school infrastructure availability relative to their
younger brothers and sisters. Children aged 14– 17 in the IAF would have been 8 –11 in
1990, just before the revitalization of school infrastructure in the country, and hence, not
subject to the ‘program’. I therefore used this cohort as a control group, and constructed a
difference-in-differences estimate by comparing the difference in school outcomes between
cohorts across administrative posts with large and small increases in schools. I compared the
1993 round of the MINED data with the 1996 round, and defined administrative posts with
‘high exposure’ as those that had more than the median number of new schools built during
this period, and those with less than the median as low exposure (the median is 5).
Table 6 provides the results of this exercise for two schooling indicators: enrolment and
grade attainment adjusted for age. For enrolment, the simple cross-sectional difference in
enrolment rates, that is, the difference among 7 –11-year olds across low- and highexposure regions is 0.099, while the difference-in-differences is 0.09, almost identical.
Moreover, the pre-program difference (the difference among the control group across
S. Handa / Journal of Development Economics 69 (2002) 103–128
113
Table 5
OLS estimates of determinants of school quality at administrative post level
Dependent variable
Proportion of heads literate
Proportion of households
with adult with EP2
Proportion of households
with female with EP1
Median consumption of
administrative post
Distance to Provincial capital
Niassa
Cabo Delgado
Nampula
Zambezia
Tete
Manica
Sofala
Inhambane
Gaza
Constant
R2
F
P value for SES variables
Observations
(administrative posts)
Proportion of teachers with training
Pupil – teacher ratio
Coefficient
z-statistic
Coefficient
0.043
0.139
0.41
0.68
11.744
13.767
1.34
0.81
0.197
1.21
21.388
1.57
0.000
0.02
0.002
1.74
0.000
0.090
0.232
0.294
0.188
0.218
0.259
0.340
0.149
0.061
0.553
0.30
2.63
0.82
102
0.61
1.19
2.77
4.09
2.69
2.79
3.56
4.05
2.07
0.77
6.77
0.034
28.619
36.078
25.574
0.879
19.038
23.375
14.642
9.368
6.096
72.452
0.60
9.17
0.17
1.46
4.54
5.16
4.27
0.15
2.92
3.85
2.10
1.56
0.91
10.64
z-statistic
Sample is rural administrative posts with at least one school built after 1991. All variables are measured at the
administrative post level, except for province dummies.
regions) is not statistically significant. The last part of Table 6 repeats the exercise using
another outcome indicator, standardized or adjusted grade attainment, defined as the
highest grade attained by the child divided by the ideal grade s/he should have attained
given age.15 In this case, the cross-sectional difference (0.0740) is over twice that of the
difference-in-differences (0.033), but even for this outcome, the pre-program difference is
not statistically significant.
In conclusion, based on the discussion with administrators in the Mozambican MINED,
as well as the results on the determinants of placement and quality of new schools in rural
areas, it appears that the program effects estimated below are unlikely to represent simply
unobserved household or community level demands for schooling.
3.3. Basic results
3.3.1. Results on school access or coverage of the educational system
Table 7a presents marginal probability estimates of the impact of school access on
EP1 enrolment by gender in rural Mozambique. Columns 1 –3 show estimates using
15
For children out of school, the IAF reports their age when they finished school.
114
S. Handa / Journal of Development Economics 69 (2002) 103–128
Table 6
Change in schooling outcomes across cohorts and by intensity of schooling construction
Schooling outcome
Enrolment
Standardized grade attainment
No increase
Large increase
No increase
Large increase
Control
(ages 14 – 17)
(N = 2290)
Treatment
(ages 7 – 11)
(N = 4119)
First difference
0.407
(0.49)
0.416
(0.49)
0.292
(0.27)
0.333
(0.29)
0.433
(0.50)
0.532
(0.50)
0.298
(0.38)
0.372
(0.39)
0.026
0.116
0.006
0.039
Standard deviation shown in parenthesis below mean. No (large) increase indicates administrative posts with less
(more) than the median number of new schools built between 1993 and 1996. Standardized grade attainment is
current grade attainment as a proportion of ideal attainment given age, and ranges from 0 to 1.
the number of schools in the administrative post, while columns 4 – 6 show estimates
using the change in the number of schools between 1993 and 1996. The number of
schools in the administrative post has a significant effect on boys but not girls
enrolment ( p value for difference in effects is 0.04), while the number of cement
rooms has a significant effect on girls’ but not boys’ enrolment ( p value for difference
is 0.08). Proximity to a school is a highly significant determinant of enrolment, the
point estimates implying that a 30-min reduction in travel time would increase
enrolment probabilities by 20 and 17 percentage points for boys and girls, respectively.
When the change in the number of schools is used instead of levels, the main
difference is the impact of schools on girls’ enrolment, which now increases and
becomes significant at 10%.
Household characteristics are also important determinants of school enrolment in rural
Mozambique, particularly for girls. In columns 1– 3, for example, all the adult schooling
variables are statistically significant, and in all cases, the marginal impact of additional
adult schooling is larger for girls than it is for boys; the existence of a female adult with
completed EP1 is especially important for girls, raising enrolment probabilities by 21
percentage points. Household access to resources (measured by per capita consumption
expenditures) also influences enrolment rates, and in this case as well, the impact is larger
for girls than it is for boys.
3.3.2. Difference-in-differences estimates of school access
The difference-in-differences analysis using older cohorts as controls can also be
applied in a multivariate context in order to provide a check on the cross-sectional
estimates presented above. Consider the following regression equation estimated over all
children ages 7 –17:
E ¼ a0 þ B1 *X1 þ B2 *ðCohortÞ þ B3 *ðProgramÞ þ B4 *ðCohort*ProgramÞ
þ ui
ð2Þ
In this framework, X1 is a vector of control variables, B2 measures the difference in
enrolment rates between older and younger cohorts in APs without any significant change
S. Handa / Journal of Development Economics 69 (2002) 103–128
115
Table 7a
Marginal impact of school access indicators on EP1 Enrolment
Log p.c. consumption
Residuala
Head literate
Adult with EP2
Female adult with EP1
Age of child in years
Log (travel time
to nearest school)
Number of cement
classrooms in AP
Number of
schools in AP
Change in number of
schools 1993 – 1996
Observations
Log likelihood
1
2
3
4
5
6
All
Boys
Girls
All
Boys
Girls
0.058
(2.27)
0.085
(2.90)
0.125
(6.02)
0.174
(4.56)
0.172
(4.15)
0.063
(11.39)
0.055
(7.77)
0.001
(2.23)
0.001
(1.53)
0.053
(1.54)
0.112
(2.79)
0.116
(3.83)
0.164
(3.33)
0.137
(2.50)
0.076
(8.69)
0.058
(7.12)
0.001
(0.91)
0.002
(2.52)
0.061
(1.97)
0.060
(1.62)
0.146
(5.51)
0.192
(3.97)
0.210
(4.29)
0.051
(6.57)
0.049
(5.88)
0.002
(2.45)
0.000
(0.27)
0.054
(2.09)
0.090
(3.01)
0.122
(5.69)
0.175
(4.52)
0.163
(3.92)
0.063
(11.28)
0.052
(7.37)
0.002
(2.95)
0.048
(1.36)
0.119
(2.86)
0.108
(3.47)
0.170
(3.42)
0.122
(2.24)
0.074
(8.40)
0.056
(6.96)
0.001
(2.10)
0.058
(1.84)
0.062
(1.60)
0.146
(5.38)
0.188
(3.84)
0.204
(4.10)
0.053
(6.67)
0.046
(5.46)
0.002
(2.54)
0.001
(2.04)
4119
2453
0.002
(1.77)
2010
1189
0.001
(1.64)
2109
1231
4290
2527
2097
1229
2193
1262
Numbers shown are marginal probabilities derived from probit estimation, with absolute z-statistics in
parenthesis. School quality variables are measured at administrative post level, except for distance to nearest
school. Constant term, provincial dummy variables, land holdings, possession of agricultural equipment, and
indicator for commercial crop production not shown. Mean of dependent variable is 0.51 and 0.40 for boys and
girls, respectively, and 0.47 for the full sample.
a
T-statistic is test for exogeneity of log p.c. expenditure.
in school infrastructure, B3 measures the pre-program difference in enrolment rates, and B4
is the double difference estimator of the impact of new schools (the ‘program’) on
enrolment. Two indicators of program exposure were used: the first measuring the actual
change in the number of schools in the administrative post between 1993 and 1996, and
the second a dummy variable indicating a high-intensity administrative post, defined as an
administrative post where the number of new schools built is above the median.16
Estimates of Eq. (2) are presented in Table 7b for the whole sample, and separately for
boys and girls. Columns 1 –3 use the actual change in schools built between 1993 and
1996 as the measure of program exposure, and the coefficient on the relevant interaction
term shown in the last line of the table indicates effects that are larger than the crosssection results, although the pattern of significance is the same. For example, the crosssection estimate of the impact of an additional school for the full sample is 0.001,
compared to 0.003 in column 1 of Table 7b. The estimates in columns 4 –6, which use the
16
This second indicator eliminates the influence of a few administrative posts reporting unreasonably large
increases in school infrastructure.
116
S. Handa / Journal of Development Economics 69 (2002) 103–128
Table 7b
Double difference estimate of impact of school access on EP1 Enrolment
Measure of
program intensity:
Change in number
of schools in
AP ( 100)
1 if large change
in schools in AP
Log (travel time to
nearest school)
Log (time) * change
in schools
1 if treated cohort
(7 – 11)
Treated cohort *
change in schools
Observations
Log likelihood
Number of new schools built
Dummy for high intensity AP
1
2
3
4
5
6
All
Boys
Girls
All
Boys
Girls
0.076
(0.73)
0.163
(1.01)
0.067
(0.54)
0.011
(0.31)
0.039
(5.03)
0.013
(1.82)
0.406
(8.53)
0.094
(2.43)
6409
3900
0.055
(1.22)
0.047
(5.33)
0.010
(1.10)
0.396
(5.00)
0.158
(3.38)
3247
2014
0.049
(0.94)
0.029
(2.77)
0.016
(1.50)
0.395
(4.88)
0.006
(0.10)
3162
1853
0.039
(5.07)
0.012
(1.70)
0.405
(8.66)
0.003
(3.68)
6409
3901
0.047
(5.36)
0.009
(0.95)
0.397
(5.31)
0.004
(2.41)
3247
2017
0.028
(2.73)
0.016
(1.51)
0.394
(4.87)
0.001
(1.23)
3162
1852
In columns 1 – 3, the program is measured by the number of new schools built in the AP between 1993 and 1996.
In columns 4 – 6, the program is measured by a dummy variable indicating whether the number of new schools
built in the AP during this period is above the median (5). Probit marginal probabilities are shown with associated
absolute z-statistics in parentheses. Control variables are the same as those in (a), plus interactions between
province and the measure of program intensity.
dummy variable for high-intensity administrative posts, also show a similar pattern of
statistical significance to the cross-sectional estimates but higher marginal effects.
3.3.3. Results on school quality
Table 8 presents probit marginal probability estimates of the impact of school quality on
EP1 enrolment by gender in rural Mozambique.17 Columns 1– 3 present the base estimates
with quality measured by the number of trained teachers and the average administrative
post level pupil –teacher ratio; the latter is insignificant but the former is positive and
significant, although the quantitative effect is small.18 Adding 10 more trained teachers
will raise the probability of enrolment by 1 percentage point.
When the total number of trained teachers is split into the proportion of male and
female trained teachers and entered as two variables (columns 4 – 6), only the proportion of
female trained teachers is significant, with the impact especially large for girls. Note that
there are very few female trained teacher in rural Mozambique (an average of 11 per
administrative post, or roughly 11% of all teachers per region). The mean of this variable is
0.08—doubling this would increase enrolment by just over 2 and 4 percentage points for
boys and girls, respectively.
17
I do not estimate a difference-in-differences model for the impact of school quality because there was no
significant change in the mean values of the quality indicators between 1993 and 1996.
18
When the proportion of teachers who are trained is used, its coefficient is positive but not significant.
S. Handa / Journal of Development Economics 69 (2002) 103–128
117
Table 8
Marginal impact of school quality indicators on EP1 Enrolment
Log (travel time
to nearest school)
Pupil – teacher ratio
No. of trained
teachers in AP
Proportion of female
teachers who
are trained
Proportion of male
teachers who
are trained
Proportion of
female teachers
Observations
Log likelihood
All
Boys
Girls
All
Boys
Girls
0.054
(7.74)
0.001
(1.08)
0.001
(3.70)
0.058
(7.02)
0.000
(0.35)
0.001
(3.81)
0.049
(5.88)
0.002
(1.30)
0.001
(2.40)
0.056
(8.10)
0.001
(1.15)
0.061
(7.40)
0.000
(0.16)
0.050
(6.15)
0.002
(1.70)
0.356
(2.85)
0.291
(1.65)
0.473
(3.43)
0.033
(0.37)
0.095
(0.78)
0.007
(0.07)
0.061
(0.24)
4290
2530
0.045
(0.13)
2097
1232
0.172
(0.63)
2193
1260
4290
2528
2097
1228
2193
1265
Numbers shown are marginal probabilities derived from probit estimation, with absolute z-statistics in
parenthesis. School quality variables are measured at administrative post level, except for travel time to nearest
school, which is for the village. The regressions also include all the control variables indicated in Table 7a. Mean
of dependent variable is 0.51 and 0.40 for boys and girls, respectively, and 0.47 for the full sample.
A recent participatory study sponsored by OXFAM (1999) in Mozambique reports that
male teachers often force students to perform chores for them such as fetching wood and
water, and that parents are reluctant to send girls to school to be taught by male teachers.
This is especially true in the heavily moslem provinces of Zambezia and Nampula, and
may explain the strong positive effect of trained female teachers reported in Table 8.
3.4. Adult literacy programs and children’s school enrolment
Table 7a reported strong effects of adult education, especially basic adult female
schooling, on the enrolment probabilities of children. Adult literacy, particularly female
literacy, is well known to be a strong determinant of children’s health, nutrition, and
schooling outcomes in developing countries, and this is true in rural Mozambique as well
(Ministry of Finance, 1998). The IAF community questionnaire provides information on
whether a sample village has had an adult literacy program, and if so, the year of the
program. Only 2.5% of the villages report having had such a program, and two-thirds of
these programs occurred since 1994. Although the incidence of villages with programs is
very small in the sample (and thus unlikely to yield statistically precise results), it is
interesting to get some initial idea of whether these programs can have an effect on
children’s schooling through their impact on adult literacy.
Table 9a shows the enrolment rates of children ages 7 –11 (treatment) and 14 –17
(control) according to whether they live in a village that ever had a literacy campaign, or
had a recent campaign (since 1994). The last line of the table indicates that the literacy rate
of household heads is higher in villages that had an adult literacy campaign relative to
118
S. Handa / Journal of Development Economics 69 (2002) 103–128
Table 9a
Village adult literacy programs and cohort enrolment rates
Control (14 – 17)
Treatment (7 – 11)
First difference
Literacy of household heads
No program
Had program
Had recent program
0.40 (0.49)
0.46 (0.50)
0.06
0.44
0.52 (0.50)
0.64 (0.48)
0.12
0.52
0.48 (0.51)
0.61 (0.49)
0.13
0.62
Recent program is one that occurred after 1993. Standard deviation in parenthesis beside proportion.
those that did not, and enrolment of children 7– 11 is similarly higher (by about 15– 18
percentage points—see row 2) in these same villages. However, the first row of Table 9a
indicates that villages with recent literacy programs had pre-program enrolment rates that
were also significantly higher (by 8 percentage points: 0.40 versus 0.48) than those with
no program. The difference-in-differences in school enrolment rates is 7 percentage points,
which is less than half the simple cross-sectional difference, but is still large.
Table 9b replicates the regression estimates in Table 7b for school access, but now
includes an indicator of whether a village had an adult literacy campaign (column 1) or a
recent campaign (column 2). Despite the very small variation in this variable, column 1
reports a large positive and significant relationship between the presence of a campaign
and children’s school enrolment probabilities. Column 2 measures only recent campaigns,
the mean of which is only about 1.6% in the sample, and the resulting coefficient estimate
is not statistically significant, although the point estimate (0.11) is of the same magnitude
as that for head’s literacy (0.12), which continues to be statistically significant.
Table 9b
Marginal impact of village adult literacy campaigns on EP1 Enrolment
(1)
(2)
dP/dX
Log p.c. consumption
Residuala
Head literate
Adult with EP2
Female adult with EP1
Age of child in years
Log (travel time to
nearest school)
Number of cement
classrooms in AP
Number of schools in AP
Village had adult
literacy program
Village had ‘recent’
adult literacy campaign
Observations
Log likelihood
z
dP/dX
z
0.056
0.084
0.123
0.178
0.174
0.065
0.053
2.22
2.91
5.99
4.66
4.15
11.72
7.77
0.058
0.080
0.124
0.178
0.171
0.065
0.055
2.23
2.72
5.86
4.64
4.08
11.80
8.05
0.001
2.21
0.001
2.10
0.001
0.200
1.59
2.59
0.001
1.61
0.114
1.17
4273
2506
4247
2491
Numbers shown are marginal probabilities derived from probit estimation, with absolute z-statistics in
parenthesis. See notes to Table 7a for explanation of variables. Recent literacy campaign means after 1993.
a
T-statistic is test for exogeneity of log p.c. expenditure.
S. Handa / Journal of Development Economics 69 (2002) 103–128
119
Table 9c
Double difference estimate of impact of literacy program on school enrolment
Village had literacy program
Treated cohort (7 – 11)
Treated cohort * literacy program
Observations
Log likelihood
(1)
(2)b
0.088 (1.01)
0.409 (8.87)
0.098 (1.15)
6654
4007
0.024 (0.30)
0.409 (8.92)
0.135 (1.11)
6609
3977
Numbers are marginal probabilities derived from probit coefficients, with absolute z statistics in parenthesis.
b
Recent literacy programs only (those after 1993).
Table 9c presents difference-in-differences estimates of the impact of literacy campaigns on school enrolment rates, based on Eq. (2) and using the two measures (ever had a
program, and had a recent program). The results in column 2 are theoretically more valid
because the control group did not have exposure to the program during their first few years
of eligibility for primary school. In column 2, the difference-in-differences point estimate
is 0.135, which is slightly higher than the cross-section estimate in Table 9b, but is not
statistically significant, probably because of the very small variation of this variable in the
sample.
3.5. Interactions with household characteristics
The influence of community infrastructure (such as school quality) may be different in
households with different characteristics. For example, the impact of a village school
may be greater for richer households if richer households are better able to take
advantage of the school. On the other hand, richer households may be able to afford to
send children to a neighboring village for schooling, in which case the impact of
constructing a school in the village may actually be greater among poorer households,
who otherwise would not have sent their children to study. The impact of community
infrastructure on household behavior may also depend on the education of adults or
parents, due to differences in preferences or access to information. In the child health
Table 10
Estimation results for school access indicators interacted with head’s literacy and household income
Interactions with:
Log (travel time to
nearest school)
No. of cement
classrooms in AP
No. of EP1
schools in APs
Interactions with head’s literacy
Interactions with household income
All (1)
Boys (2)
Girls (3)
All (4)
Boys (5)
Girls (6)
0.011
(1.04)
0.001
(1.36)
0.001
(1.07)
0.015
(1.10)
0.002
(1.54)
0.000
(0.18)
0.008
(0.60)
0.001
(0.45)
0.001
(0.95)
0.029
(3.01)
0.001
(1.01)
0.001
(0.99)
0.029
(2.12)
0.001
(0.93)
0.002
(1.10)
0.027
(2.69)
0.026
(2.31)
0.001
(0.45)
Numbers shown are marginal probabilities derived from probit coefficients of the interaction of each variable with
literacy of household head (columns 1 – 3) and household consumption (columns 4 – 6). Absolute z-statistics are in
parenthesis. Travel time is measured at the village level. Number of observations, mean of dependent variable,
and other control variables are the same as in Table 7a.
120
Interactions with:
Log (travel time
to nearest
school)
Pupil – teacher
ratio
No. of trained
teachers
Proportion of
female teachers
who are trained
Proportion of
male teachers
who are trained
Proportion of
female teachers
Interactions with head’s literacy
Interactions with household income
1
2
3
4
5
6
7
8
9
10
11
12
All
Boys
Girls
All
Boys
Girls
All
Boys
Girls
All
Boys
Girls
0.012
(1.08)
0.016
(1.13)
0.008
(0.58)
0.010
(0.89)
0.013
(0.95)
0.006
(0.44)
0.026
(2.77)
0.025
(1.91)
0.025
(2.46)
0.027
(2.78)
0.027
(2.04)
0.023
(2.26)
0.000
(0.17)
0.000
(0.63)
0.000
(0.06)
0.000
(0.51)
0.000
(0.07)
0.000
(1.10)
0.001
(1.18)
0.000
(0.42)
0.001
(0.76)
0.000
(1.78)
0.002
(1.27)
0.001
(1.90)
0.298
(1.36)
0.462
(1.55)
0.138
(0.53)
0.057
(0.36)
0.025
(0.10)
0.159
(0.82)
0.071
(0.51)
0.050
(0.26)
0.121
(0.70)
0.060
(0.41)
0.002
(0.01)
0.077
(0.48)
0.513
(2.06)
0.415
(1.02)
0.540
(1.72)
0.158
(0.70)
0.096
(0.30)
0.208
(0.75)
Numbers shown are marginal probabilities derived from probit coefficients of the interaction of each variable with literacy of household head (columns 1 – 2) and
household consumption (columns 3 – 4). Absolute z-statistics are in parenthesis. School quality variables are measured at administrative post level, except for travel time,
which is measured at the village level. Number of observations, mean of dependent variable, and other control variables are the same as in Table 8.
S. Handa / Journal of Development Economics 69 (2002) 103–128
Table 11
Estimation results of school quality indicators interacted with head’s literacy and household income
S. Handa / Journal of Development Economics 69 (2002) 103–128
121
literature, for example, the impact of mother’s education has been found to vary
significantly with community characteristics such as sewerage and sanitation conditions
(Thomas et al., 1991; Barrera, 1990).
Both household income (measured by expenditures per capita) and adult education
significantly influence schooling choices in Mozambique, and school infrastructure also
conditions these choices in rural areas. Does the impact of school infrastructure depend on
household characteristics? Are certain households more likely than others to change their
schooling decisions in response to variations in school infrastructure? These questions are
addressed by interacting the different school supply characteristics with household adult
education (measured by the literacy of the head) and household income, to see if
significant interactions indeed exist between school supply and household characteristics.
The interactions are tested sequentially, first by interacting the school supply variables
with head’s literacy, and then by interacting the same variables with household (log) per
capita consumption. Results are presented separately for two dimensions of school supply
(access and quality) in Tables 10 and 11.
Starting with school access, Table 10 presents the results of the interactions between
each school access indicator and head’s literacy (columns 1 – 3) and household
consumption (columns 4– 6). Significant interactions exist among several access indicators and household income (columns 4– 6). For both boys and girls, distance to a
primary school and household income are substitutes, hence, the positive impact of
constructing a school nearby will be greater among poorer households. Furthermore, the
positive impact of cement classrooms on girls’ enrolment is enhanced among richer
households, given by the positive and significant coefficient on the interaction term in
column 6.
Table 10 presents the results of the estimates of school quality indicators interacted
with head’s literacy (columns 1 – 6) and household consumption (columns 7 –12). For
the full sample, the impact of the proportion of female teachers depends on whether the
head is literate or not. The negative coefficient on the interaction term (column 4)
implies that these two characteristics are substitutes, and therefore, the impact of these
dimensions of school quality is significantly greater among households where the head
is not literate. The results for income in columns 7– 12 show one marginally significant
coefficient. For girls, the impact of the number of trained teachers in the AP varies with
household income; the positive coefficient in this case implying complementarity
(column 9).
4. Policy simulations
According to the Ministry of Education’s strategic plan, raising basic primary education
levels is a priority for Mozambique. In this section, the relative impact of demand side
versus supply side interventions on primary school enrolment rates in rural Mozambique is
compared. The simulations are based on the probit regressions for the determinants of
current enrolment of children aged 7– 11 years old in rural areas. The school characteristics included in the model are the number of trained teachers and the pupil – teacher ratio,
and the number of schools and cement rooms in the administrative post. All the household
122
S. Handa / Journal of Development Economics 69 (2002) 103–128
level characteristics mentioned above are included in the model, as well as the village level
variable indicating the travel time to the nearest school. Because of Mozambique’s vast
size and geographical and economic heterogeneity, the impact of the hypothetical policy
interventions to vary by province by interacting the policy variables with provincial
dummy variables was allowed.19 Systematic differences in the effect of policy interventions on boys’ and girls’ enrolment rates were not found and so estimates for the full
sample only are provided.
4.1. Supply side simulations
The supply side policy simulations consider the impact on enrolment rates of
increasing the number of schools in rural areas in Mozambique. The IAF community
questionnaire indicates that approximately 68% of rural villages have a basic primary
school, and the regression analysis shows that the distance to a school is an extremely
important determinant of children’s enrolment. The increase in EP1 enrolment that would
occur due to two separate interventions was calculated: (1) increasing the overall EP1
coverage rate to 79%, which implies building a school in 70 villages per province; (2)
increasing the overall EP1 coverage rate to 89%, which implies building schools in 140
villages per province. These rates are attained by increasing the number of schools in
each administrative post. In order to capture the impact of school characteristics (and not
just access) on enrolment, it was assumed that each school consists of three cement
rooms and comes with two trained teachers. The addition of a school in an administrative
post will reduce the average travel time to the nearest school. This indirect effect due to
changes at the administrative post level is accounted for in the policy simulations by
reducing the travel time for villages in an administrative post that receives an additional
school.20
4.2. Demand side simulations
The impact of policy interventions designed to influence demand side (or household)
characteristics is based on the same model used for the supply side simulations. Two
types of interventions are simulated, one influencing household income (or consumption) and the other influencing adult education. The income-related interventions involve
raising the per capita consumption of all households to at least the level of consumption
of the 25th percentile of the per capita consumption distribution (Mt. 2494 per person
per day in the IAF, or approximately 25 US cents); the second policy is to raise all
households to at least Mt. 3584, which is equal to median consumption in the IAF. Since
these interventions only affect poor households, they will not be evenly distributed
throughout the country. In particular, the poorer the province, the larger the share of
19
Detailed simulation results by province are not presented here, but are discussed in Handa (2000).
The average reduction in travel time was calculated by estimating an OLS regression for the relationship
between number of schools in an administrative post and village travel time. The estimated coefficient on the
variable ‘number of schools in the AP’ is used to adjust the travel time variable in the simulation.
20
S. Handa / Journal of Development Economics 69 (2002) 103–128
123
households in the bottom 25th percentile or bottom half of the per capita consumption
distribution, and thus, the larger the number of households who will be affected by the
policy. The second demand side simulation is motivated by the results presented above,
indicating that adult household education significantly conditions children’s schooling.
The impact on enrolment rates was simulated if all household heads in the bottom
quartile of the per capita expenditure distribution were literate. As in the income case,
the benefits of this intervention will not be distributed equally across provinces. While
poorer provinces have more eligible households, the policy only affects heads of
households who are not literate and so the proportion of heads who are literate also
matters.
4.3. Results of simulations
Results of the supply and demand side simulations are presented in column 1 of Table
12. These are calculated as the percentage change in overall mean enrolment with respect
to the baseline figure for predicted enrolment derived from the probit estimates without
any simulations. The policy of increasing EP1 coverage to 79% (row 1) will increase
overall enrolment by 13%, while doubling the size of the intervention would raise
enrolment by 35%. Rows 3 and 4 provide estimates of the percentage change in enrolment
due to the two income-related policy interventions described above. The overall (national)
impact is to raise enrolment rates by 2% and 4%, respectively—these effects are
significantly smaller than the estimated enrolment effects of building more schools. The
last two rows of column 1 in Table 12 present simulation results based on interventions
that raise the literacy level of heads of households in the bottom parts of the per capita
consumption distribution. The overall impact of this intervention is substantially larger
than the income intervention. Increasing literacy of heads in the bottom quartile would
increase overall enrolment by 8%; increasing literacy of heads in the bottom half of the
distribution would increase enrolment rates by 15%.
Note that if it is not literacy itself, but factors associated with literacy that lead to
increased child schooling (such as preferences or value for education), then the simulation
Table 12
Policy simulations and cost-effectiveness of supply and demand side interventions
Intervention
(1)
(2)
(3)
(4)
Benefit
Unit cost
(US$)
Total cost
(US$ million)
Effectiveness:
(3)/(1)
(US$ million)
13
35
2
4
8
15
70,000
70,000
29b
55b
30
30
49
98
24
91
14.7
27.9
3.8
2.8
12.0
22.8
1.8
1.9
a
Build 70 schools per province
Build 140 schools per province
Bring Households to 25th percentile
Bring Households to 50th percentile
Literacy to heads in bottom quartile
Literacy to heads in bottom 2 quartiles
a
b
Percentage increase in enrolment based on simulation results. See text for details.
Average per household for 1 year.
124
S. Handa / Journal of Development Economics 69 (2002) 103–128
results are likely to overestimate the benefits of adult literacy campaigns in Mozambique. I
have tried to control somewhat for these family-specific tastes or talents for education by
including the education level of other adult household members in the regression
equations. These other variables actually reduce the estimated impact of head’s literacy
by 25%.21 Moreover, one may argue that family-specific preferences for education are
likely to be smaller in a supply-constrained society such as rural Mozambique, relative to
more developed countries like the United States or Canada, where the spatial supply of
schools is relatively abundant.
4.4. Cost effectiveness of policy options
The simulations presented above provide an idea of the overall benefit of different
policy interventions without considering the cost of these same interventions. I have
gathered approximate costs for adult literacy campaigns and school construction from
NGOs working in the education sector in rural Mozambique. The cost of building a basic
three-room cement school in rural Mozambique is estimated to be US$50,000. To this
construction cost, the cost (including administrative costs) of two teachers for 10 years
was included, which, according to the pay structure for teachers, adds an additional
US$20,000 to the cost of a rural primary school. The policy simulation in row 1 of Table
12 calls for building 70 schools in each of the 10 rural provinces, at a cost of US$70,000
per school, or a total cost of US$49 million. Dividing this figure by the percentage
increase in enrolment (13%) gives approximately US$3.8 million per percentage point
increase in enrolment.
Kulima, a local NGO that has provided adult literacy campaigns in rural Mozambique, estimates a total cost per adult of US$30 for the delivery of a 1-year literacy
program in a rural village. According to the IAF, and using population weights, there are
approximately 490,000 illiterate heads of household in the bottom quartile (59% of
heads in the bottom quartile cannot read or write), and approximately 930,000 illiterate
heads in the bottom two quartiles (54% of heads are illiterate among this group).
Providing literacy for the heads in the bottom quartile at a cost of US$30 per person
leads to a total cost of US$14.7 million, which when divided by the expected percentage
increase in enrolment (8) yields US$1.8 million per percentage point increase in
enrolment.
Finally, using the (population weighted) figures for per capita household consumption
in the IAF, the total amount of transfer required to bring all households below the 25th
percentile to a per capita household consumption exactly equal to consumption in the 25th
percentile was calculated. This figure is US$24 million per year, and when divided by the
expected percentage increase in enrolment (2% – see column 1 of Table 12), yields US$12
million per unit of expected benefit.
The approximate costs associated with each intervention and the associated costeffectiveness numbers are shown in columns 2, 3, and 4 in Table 12 for each policy
21
When the variables ‘number of adults with EP2’ and ‘number of adult females with EP1’ are excluded
from the regressions, the estimated impact of ‘head literate’ increases by 25% for both boys and girls.
S. Handa / Journal of Development Economics 69 (2002) 103–128
125
simulation. These estimates clearly show that the income intervention is the least costeffective method of raising primary school enrolment in rural Mozambique. The policies
of adult literacy and improved access to schools are significantly more cost-effective
methods of raising enrolment rates, with adult literacy providing a slightly cheaper
alternative among these two options.
5. Conclusions
Raising primary school enrolment is a major development imperative, although the
interventions that can best raise enrolment are not always straightforward, and can vary
both between and within countries. Using the first national household survey of
Mozambique, coupled with detailed information on school infrastructure supplied by
the Ministry of Education, this paper evaluates the relative importance of supply and
demand side factors in determining rural primary school enrolment. Simulations based
on a set of ‘plausible’ demand and supply side interventions indicate that in rural
Mozambique, building more schools or raising adult literacy will have a larger impact
on enrolment rates than interventions that raise household income. For example, raising
the EP1 coverage rate to 79% in rural areas will increase enrolment rates by 13%, while
making household heads literate in the bottom per capita consumption quartile will raise
rural primary school enrolment by 8%. In contrast, bringing per capita consumption of
the poorest quartile up to Mt. 2494 per day will raise rural enrolment by a mere 2%.
When relative costs are considered, adult literacy campaigns become more attractive,
with cost-effectiveness ratios that are 6 – 10 times better than the income intervention,
and 1.5 –2.5 times better than building more schools. Even if we assume that half the
measured benefit of adult literacy is through unmeasured tastes for, or ability in,
acquiring human capital, adult literacy is just as cost-effective as extending coverage
through school infrastructure.
The detailed analysis of the impact of school characteristics on primary school
enrolment in rural Mozambique indicates that dimensions of school quality and access
both work to stimulate enrolment, although the effects are small and differ somewhat by
gender of child. School quality, measured by the number of trained teachers in the
administrative post, has a positive and significant impact on enrolment, but it is the
gender composition of the trained teaching staff that is even more important in
determining the household decision to send children to school. For example, the share
of female teachers who are trained is an important positive determinant of enrolment
rates. Raising this ratio from 0.08 to 0.16 in the administrative post will raise enrolment
rates by roughly 4 percentage points.
School availability also has a significant impact on enrolment rates. Reducing the travel
time to the nearest school will increase enrolment rates for both sexes by 17 – 20
percentage points, and the impact of school availability is enhanced for girls if the school
is built with cement.
Few previous studies have considered the possible interaction between school supply
indicators and household characteristics. In Mozambique, these exist particularly for girls.
In terms of policy, the most interesting of these is the positive interaction between travel
126
S. Handa / Journal of Development Economics 69 (2002) 103–128
time to a school and household income, which implies that the two factors are
substitutes—construction of a village school will increase enrolment more among poorer
households. Additionally for girls, there is a positive interaction among household income,
cement classrooms, and the number of trained teachers in the area.
Acknowledgements
Thanks to Farizana Omar, Helder Zavale, and Virgolinho Nhate for excellent
research assistance, to Manuel Rego of the Mozambique Ministry of Education for
supplying and interpreting the data, and Gaurav Datt, Dean Jolliffe, and especially Ken
Simler, for helpful comments on earlier drafts. Useful criticism was also provided by
an anonymous referee. This paper was written while the author was outposted by
IFPRI as Professor of Economics to the Eduardo Mondlane University, Maputo,
Mozambique.
Appendix A1. Summary statistics for children ages 7 – 11
Girls
Boys
Mean
SD
Mean
SD
Log daily per capita consumption
Land holdings (ha)
Have irrigation
Have agricultural equipment
Head literate
Adult in household with EP2
Adult female in household with EP2
Head female
Head’s age
Currently enrolled in school
8.154
2.564
0.046
0.046
0.476
0.078
0.109
0.187
45.508
0.406
0.61
2.41
0.21
0.21
0.50
0.27
0.31
0.39
12.80
0.49
8.143
2.590
0.046
0.042
0.454
0.077
0.109
0.184
45.146
0.498
0.64
2.70
0.21
0.20
0.50
0.27
0.31
0.39
12.71
0.50
School characteristics
Pupil – teacher ratio
No. of trained teachers
Proportion of trained female teachers
Proportion of female teachers
Pass rate
Girls’ pass rate
Boys’ pass rate
Portuguese subject pass rate
Mathematics subject pass rate
Log (travel time to nearest school)
Number of EP1 schools in administrative post
Change in number of EP1 schools
Have EP2 in AP
Have secondary school in AP
Number of observations
65.324
79.654
0.096
0.103
0.644
0.581
0.676
0.663
0.683
1.384
26.118
8.383
0.651
0.272
2293
18.07
66.47
0.10
0.12
0.05
0.07
0.05
0.05
0.05
2.11
21.08
11.35
0.48
0.44
65.157
80.414
0.094
0.102
0.643
0.580
0.676
0.661
0.682
1.317
25.945
8.301
0.657
0.273
2203
17.59
66.90
0.11
0.13
0.05
0.07
0.05
0.05
0.05
2.10
19.78
11.77
0.47
0.45
School characteristics are measured at the administrative post level, except for travel time to school.
S. Handa / Journal of Development Economics 69 (2002) 103–128
127
References
Alderman, H., Behrman, J., Ross, D., Sabot, R., 1996. Decomposing the gender gap in cognitive achievement in a
poor rural economy. Journal of Human Resources 31, 229 – 254.
Barrera, A., 1990. The role of maternal schooling and its interactions with public health programs in child health
production. Journal of Development Economics 32, 69 – 91.
Betts, J., 1995. Does school quality matter? Evidence from the national longitudinal survey of youth. Review of
Economics and Statistics 77 (2), 231 – 250 May.
Birdsall, N., 1985. Public inputs and child schooling in Brazil. Journal of Development Economics 18, 67 – 86.
Card, D., Kruger, A., 1992. Does school quality matter? Returns to education and the characteristics of public
schools in the United States. Journal of Political Economy 100 (1), 1 – 40 February.
Case, A., Deaton, A., 1999. School inputs and educational outcomes in South Africa. Quarterly Journal of
Economics 114 (3), 1047 – 1084 August.
Duflo, E., 1999. Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from
an Unusual Policy Experiment, mimeo, MIT.
Glewwe, P., 1999. Why does mother’s schooling raise child health in developing countries? Evidence from
Morocco. Journal of Human Resources 34 (1), 124 – 159 Winter.
Glewwe, P., Jacoby, H., 1994. Student achievement and schooling choice in low income countries. Journal of
Human Resources 29 (3), 843 – 864.
Glewwe, P., Grosh, M., Jacoby, H., Lockheed, M., 1995. An eclectic approach to estimating the determinants of
achievement in Jamaican primary education. World Bank Economic Review 9 (2), 231 – 258.
Golhaber, D., Brewer, D., 1997. Why don’t schools and teachers seem to matter? Assessing the impact of
unobservables on educational productivity. Journal of Human Resources 32 (3), 505 – 523 Summer.
Handa, S., 1996. The determinants of teenage schooling in Jamaica: rich vs. poor, females vs. males. Journal of
Development Studies 32, 554 – 580.
Handa, S., 1999. Maternal education and child height. Economic Development and Cultural Change 47,
421 – 439.
Handa, S., 2000. Raising Primary Schooling in a Developing Country, International Food Policy Research
Institute Discussion Paper, Washington, DC hwww.ifpri.cgiar.org/divs/fcnd/dp/dp76.htmi.
Hanushek, E., 1995. Interpreting recent research on schooling in developing countries. World Bank Research
Observer 10 (2), 227 – 246.
Hedges, L., Laine, R., Greenwald, R., 1994. Does money matter? A meta-analysis of studies of the effect of
differential school inputs on student outcomes. Educational Researcher 23 (3), 5 – 14.
Kremer, M., 1995. Research on schooling: what we know and what we don’t. World Bank Research Observer 10
(2), 247 – 254.
Lam, D., Duryea, S., 1999. Effects of schooling on fertility, labor supply, and investments in children, with
evidence from Brazil. Journal of Human Resources 34 (1), 160 – 192 Winter.
Lavy, V., 1996. School supply constraints and children’s educational outcomes in rural Ghana. Journal of
Development Economics 51, 291 – 314.
Ministry of Finance, 1998. Understanding Poverty and Well Being in Mozambique: The First National Assessment, Maputo, Government of Mozambique.
OXFAM, 1999. Access to Basic Education and Health in Mozambique, Oxfam-GB, Maputo, Mozambique.
Pradhan, M., 1998. Enrolment and delayed enrolment of secondary school age children in Indonesia. Oxford
Bulletin of Economics and Statistics 60 (4), 413 – 431.
Republic of Mozambique, Educational Indicators, Ministry of Education, Planning Directorate, Maputo, Mozambique, various years.
Rivers, D., Vuong, Q., 1988. Limited information estimators and exogeneity tests for simultaneous probit models.
Journal of Econometrics 39 (3), 347 – 366 November.
Rosenzweig, M., Schultz, P., 1983. Estimating a household production function: heterogeneity, the demand for
health inputs, and their effects on birth weight. Journal of Political Economy 91 (5), 723 – 746 October.
Simmons, J., Alexander, L., 1978. The determinants of school achievement in developing countries: a review of
the research. Economic Development and Cultural Change 26 (2), 341 – 357 Jan.
Strauss, J., Thomas, D., 1995. Human resources: empirical modeling of household and family decisions. In:
128
S. Handa / Journal of Development Economics 69 (2002) 103–128
Behrman, J., Srinivasan, T.N. (Eds.), Handbook of Development Economics, vol. 3A. North-Holland, Amsterdam, pp. 1883 – 2005.
Tan, J.-P., Lane, J., Coustere, P., 1997. Putting inputs to work in elementary schools: what can be done in the
Philippines? Economic Development and Cultural Change 45, 857 – 879.
Thomas, D., Strauss, J., Henriques, M.-H., 1991. How does mother’s education affect child height? Journal of
Human Resources 26, 183 – 211.