71 P. Glick, D.E. Sahn Economics of Education Review 19 2000 63–87
school exit decisions in this way allows in a limited way for a dynamic analysis of schooling transitions using
cross-section data. For example, a key demographic vari- able on the right-hand side is the number of siblings
under 5, i.e. the number of brothers and sisters born and surviving and remaining in the household during the last
5 years. The analysis thus shows how school status is affected by “fertility shocks”, that is, additions to the
family of young children treated as exogenous during the period.
The determinants of leaving school are estimated using random effects probit on the samples of boys and
girls aged 15–18 who, using the method just described, are inferred to have been enrolled at the start of the per-
iod. Note that because of the way the dependent variable is constructed, we are effectively analyzing behavior
since age 10 for children currently 15, since 11 for those now 16, and so on. Since leaving school earlier than age
10 is rare, restricting the sample to children 15 and older is appropriate.
24
4. Data and descriptive statistics
The data used in this study are taken from a survey of 1725 households conducted in Conakry in 1990. The
survey contains detailed information on a wide range of socioeconomic factors such as education, labor force
activity and earnings, assets, and health. Since the house- hold survey does not contain information on the charac-
teristics of schools, the focus in this study is on the effects of household and individual factors on schooling.
Our sample for analysis consists of boys and girls aged 10–18 living with at least one parent. Means and stan-
dard deviations of the explanatory variables for boys and girls are given in Table 1.
We restrict the sample to children residing with a par- ent for two reasons. First, one of our primary concerns
is the impact of parental education on child schooling,
24
Since the sample consists of individuals inferred to have been in school 5 years earlier, the results must be regarded as
being conditional on prior enrollment. Two other selection- related concerns should be noted. First, those who have recently
left school and have also left the household are not recorded in the survey. Second, years since leaving school will, in general,
be underestimated for children who started school late or repeated more grades than the typical child in their last grade,
since these children would have been attending their last grade at a higher age than the median for the grade; the reverse is true
for early starters and those who repeated less than the average. Overestimation of years since departure will cause some in the
second group to be incorrectly dropped from the sample as non- recent i.e. not within the last 5 years school leavers, which
might impart a bias to the estimates on the sample excluding them.
and the survey does not record the educational attain- ment of parents who do not live with their children or
of children who do not live with their parents. Second, for older children who no longer live with their parents,
the characteristics of their present household, such as income and demographic composition, are not likely to
be the relevant ones for determining the education out- comes of interest and may, in fact, be the outcomes of
prior schooling decisions: consider, for example, the case of a young man quitting school, going to work, and start-
ing his own family. Given the cumulative nature of schooling decisions, the circumstances of the household
in which the child was raised are more relevant, so for this reason as well we focus on children still living in
the household of the parent or parents.
Although the choice of sample is necessary for these reasons, it involves a significant sample reduction. A siz-
able percentage of children do not live with either parent and this ratio rises with age. For example, among all
children in the sample age 13–18, 31 of girls and 44 of boys are living away from their parents. The data thus
suggest that child fostering is an important phenomenon in Guinea, as elsewhere in West Africa Ainsworth,
1992. In addition to fostering, older teenage children may move out of the households in which they were
raised in order to marry, work, or attend school; early marriage of girls in particular is evident from our data.
This raises the issue of selection bias in our estimates of the determinants of schooling, since households in which
children are living with their parents may differ from other households in terms of unmeasured preferences or
propensities for school or children who leave may be different from children in the same household who
stay.
25
The Conakry survey does not record the numbers
25
The potential biases involved in restricting estimation of schooling functions to the sample of children living with a par-
ent can be illustrated with the following elaboration of the Heckman selectivity model Heckman, 1979; see also Pitt
1997 for a more detailed exposition of a model similar to the following. Assume that equations for the level of schooling S
and selection into the sample of children observed to be living at home H, an index function such that H 0 implies selection
into the sample take the following linear forms:
S
i
5 x
i
b
1
1 ub
2
1 e
i
5 x
i
b
1
1 e
i
H
i
5 w
i
g
1
1 ug
2
1 u
i
5 w
i
g
1
1 m
i
where x
i
and w
i
are vectors of exogenous variables, u is an unobserved to the researcher heterogeneity factor representing
household parental preferences for schooling we focus on household rather than individual heterogeneity in this example,
and e
i
and u
i
are independent, normally distributed random error terms. Although e
i
and u
i
are independent, e
i
and m
i
are not, because of the presence of u in both terms: thus we have
corre
i
,m
i
5 b
2
g
2
s
2 u
. To find the effect of a change in an exogenous variable x
k
appearing in both equations on schooling
72 P. Glick, D.E. Sahn Economics of Education Review 19 2000 63–87
Table 1 Boys and girls ages 10–18: variable means and standard deviations
Girls Boys
Variable Mean
Standard Mean
Standard deviation
deviation Years of schooling
3.714 2.933
4.595 2.708
Currently in school 1 5 yes 0.608
0.488 0.798
0.402 WITHDRAW 5 1 if left school in the last 5 years
a
0.317 0.467
0.175 0.379
Age years 13.679
2.568 13.633
2.500 Mother’s years of schooling
2.574 4.465
2.086 3.970
Father’s years of schooling 3.149
5.119 3.287
4.982 Mother missing 1 5 yes
0.101 0.301
0.167 0.373
Father missing 1 5 yes 0.154
0.361 0.116
0.320 Log expenditure per adult
10.508 0.581
10.494 0.583
Siblings , 5 0.701
0.804 0.600
0.762 Brothers 5–12
0.646 0.764
0.648 0.779
Sisters 5–12 0.521
0.677 0.475
0.649 Brothers 13–20
0.535 0.724
0.554 0.719
Sisters 13–20 0.466
0.657 0.448
0.689 Other children , 5
0.850 1.169
0.948 1.247
Other children 5–12 1.192
1.472 1.316
1.504 Other boys 13–20
0.612 0.952
0.602 0.934
Other girls 13–20 0.569
0.927 0.680
1.022 Men 21–64
1.995 1.499
1.890 1.392
Women 21–64 2.201
1.362 2.169
1.347 Men 64
0.102 0.303
0.114 0.317
Women 64 0.089
0.319 0.085
0.290 Ethnicity excluded 5 Soussou
Fulani 0.195
0.396 0.253
0.435 Malinke
0.193 0.395
0.191 0.394
Other ethnic 0.064
0.245 0.065
0.246
a
For ages 15–18 and enrolled 5 years prior to survey see text.
conditional on being in the sample of children living at home, substitute for corre
i
,m
i
in the standard expression for a regression conditional on sample selection Heckman, 1979
and take the derivative with respect to x
k
: ∂
E S
i
|H
i
5 1
∂ x
k
5 b
k
2 g
k
A
i
b
2
g
2
s
2 u
where A
i
5 l
2 i
1 w
i
g
1
l
i
, l
i
is the inverse Mills ratio evaluated at w
i
g
1
and is positive, as is A
i
as long as w
i
g
1
0 as we would expect for positive selection into the home sample. The second
term on the right-hand side represents the bias in the estimate of b
k
, the effect of the change in x
k
on schooling, arising from sample selection. The direction of bias will depend on the signs
of g
2
and b
2
, the effects of schooling preference on sample selection and schooling, respectively, as well as on the sign of
g
k
, the coefficient of x
k
in the selection equation. Since b
2
is positive, for a variable that increases the probability of selection
into the with-parent sample g
k
0, the estimate of b
k
will be biased downward if g
2
is also positive, that is, if parents who conditional on observed factors prefer educated children are
also more likely to keep them at home.
of children of parents in sample households who do not reside at home, so we do not know the exact extent of
fostering-out and marrying-out of the estimating sample. However, it is likely to be far less than implied by the
high ratios just cited of children not living with parents to children living with parents. This is because most of
the children in the former group come from non-Conakry non-survey households; that is, they are fostered into
households in our sample from other, typically rural areas rather than from other households within Conakry.
We infer this from the fact that the majority of children in the sample who do not live with parents are listed as
recent migrants to Conakry the opposite is true for chil- dren with parents. This is in agreement with evidence
for the region that child fostering is more prevalent from rural to urban households than among households in
urban areas especially the same urban area Ainsworth, 1992. In addition, although early marriage for girls is
not uncommon in Guinea 21 of girls aged 15–18 in the survey are married, most teenage girls in Conakry
who are married are also migrants, and most of these
73 P. Glick, D.E. Sahn Economics of Education Review 19 2000 63–87
apparently came to Conakry after getting married, or to get married.
26
Thus girls who marry young and no longer live with their parents for the most part have not married
out of Conakry households. These rural to urban patterns of fostering, marriage,
and migration in the data suggest that children and ado- lescents flow into households in Conakry from which
our estimating sample is drawn far more than they flow out. Consequently, households or more accurately,
parents in Conakry are “missing” fewer children as a result of out-fostering or early marriage than implied by
the means for all children, which include children from non-sample rural households. This alleviates to some
extent concerns over a large-scale selective elimination of observations from sample households.
27
On the other hand, selective rural–urban migration would imply that
our randomly selected sample of households in Conakry, even with all children counted, is not representative of
Guinea households overall. Given this possibility, it should be kept in mind that inferences regarding house-
hold determinants of schooling are conditional on resi- dence in Conakry.
28
26
Among girls aged 14–18 in the Conakry sample who are married and living away from their parents, 81 were born
outside of Conakry and the vast majority of these had arrived in Conakry in the previous 5 years. Their average age at the
time of migration is high 14.2 years, suggesting that they were married when they migrated or came to get married rather than
simply having migrated with their parents at a young age in which case they would indeed have selected out of the sample
of Conakry-resident households. Thus while there are many married girls in the sample under 18 not living at home, they
appear by and large not to have come from sample households. As further evidence of this, we note that the ratio of boys to
girls living at home is stable rather than increasing with age at least until age 19, even though boys unlike girls do not marry
young or enter the labor force early rates of marriage and par- ticipation are very low for males under 20: if girls were marry-
ing and moving out of sample households in large numbers, we would expect an increasing male–female ratio with age.
27
Survey households containing parents of children aged 10– 18, all of whom are living away from home, obviously would
be excluded entirely from the analysis, but if as we suspect, the portion of children from Conakry households not living at home
is relatively small, there will be few such cases.
28
To the extent that, in spite of the considerations noted in the text, there is a selectivity bias problem from absent children,
it is possible to make inferences about the likely direction of the bias for certain regressors, provided we can make reasonable
assumptions regarding the signs of the key parameters in the analytical framework presented in note 25. We expect that par-
ents in Conakry who have preferences for educated children are more likely to keep them at home i.e. g
2
0. As noted in the text, fostering out children for schooling or other pur-
poses appears to be more important from rural to urban house- holds than from urban Conakry households, and it is likely that
very few families in our sample can afford to send children to
We now examine some relevant descriptive statistics for the sample. Table 2 shows enrollment status of boys
and girls by age. As would be expected, Conakry has substantially higher enrollment rates than the national
largely rural averages noted in the introduction, but a large gender gap remains. Boys’ enrollments are consist-
ently higher than girls’ and the gap widens with age, reflecting girls’ earlier withdrawal from school. This is
especially apparent after age 15: for example, 73 of boys aged 16–17 are in school compared to only 49
of girls. We should note that the relatively high
Table 2 Current enrollment rates by age and sex in Conakry
Age Males
Females 6–7
0.58 0.45
8–9 0.83
0.72 10–11
0.84 0.71
12–13 0.86
0.63 14–15
0.80 0.63
16–17 0.73
0.49 18–20
0.63 0.48
20–21 0.52
0.33
other countries for schooling. At least among older children, therefore, leaving home early is associated primarily with mar-
riage for girls or work or apprenticeship for boys and thus also with leaving school. Thus selection of children out of sam-
ple should be negatively correlated with unobserved preferences or propensities for schooling, implying g
2
0. Since b
2
is also positive, the effect of a variable on schooling will be under-
over-estimated if that variable increases decreases selection into the sample of children staying at home. Parental schooling
is likely to increase the desire for child schooling and for this reason will also make it more likely that the child does not
leave home early. Thus the positive effect of parental education on schooling for a random sample of children will be under-
stated by estimation on the at-home sample when there is selec- tivity. With regard to household income, it is likely that g
k
and b
k
are again both positive, the latter through an income effect on schooling demand and the former because households with
more resources have less need to marry off daughters or encour- age sons to leave and support themselves on the other hand,
it is possible that poor households will retain sons to contribute to household income through work in a family business. Thus
the effect of income on schooling demand will also be under- estimated. Finally, with respect to the effect of younger siblings
on girls’ schooling discussed in detail below, b
k
will be nega- tive since having more young children raises the demand for a
girl’s time in household work. If parents are also inclined to keep girls at home when there is a greater need for their house-
work services, g
k
will be positive, so that selectivity results in a downward bias, i.e. an overstatement of the absolute value
of the reduction in girl’s education from an additional young sibling.
74 P. Glick, D.E. Sahn Economics of Education Review 19 2000 63–87
enrollment rate for boys in this age group does not mean that most boys let alone girls in Conakry progress to
upper secondary school: all but 4 of enrolled boys aged 16–17 and a similar proportion of enrolled girls are still
attending either lower secondary school grades 7–10 or primary school, a reflection of late starting age or high
grade repetition.
29
Table 3 shows years of schooling or grade attainment of children 10–18 by sex and education level of the
mother and father. We should point out that average schooling levels of parents and especially of mothers is
very low: about 75 of the mothers in the sample and 65 of the fathers have less than a primary education,
and almost 70 of mothers and 60 of fathers have no schooling at all. In looking at these descriptive results,
it should be kept in mind that most of these children 80 of boys and 61 of girls are currently attending
school and thus will likely have an ultimate grade greater
Table 3 Boys and girls aged 10–18: mean years of schooling by sex and
level of mother’s and father’s education Boys
Girls Mother’s education
a
Less than primary 4.37
3.28 Completed primary
5.14 4.66
Completed secondary or higher 5.54
5.47 Father’s education
b
Less than primary 4.31
3.05 Completed primary
4.81 4.48
Completed secondary or higher 5.55
5.54
a
On sample with mother schooling data.
b
On sample with father schooling data.
29
Enrollments for children who do not live with a parent show a similar pattern of gender differences but are substan-
tially lower for both boys and girls. Of course, some older chil- dren in this group have finished their schooling and left home,
but younger children not living with a parent i.e. fostered children also have lower enrollments. Multivariate analysis on
the samples of all boys and all girls under 15 results available from the authors indicates an independent negative effect of
being fostered on the probability of enrollment controlling for household income and other factors. This does not prove, how-
ever, that the practice of fostering is detrimental to child school- ing; in fact, anthropological evidence suggests that fostering is
often used by parents to secure an education for their children see Ainsworth, 1992 and references therein. As noted, living
without a parent is correlated with having been born outside of Conakry, which for most such children would mean a rural area
where schooling opportunities are relatively limited. Fostered children’s access to schooling in Conakry may thus be superior
to what it would have been in the place of origin, even if it is inferior to that of native born children or those living with
a parent.
than their current grade. Because of this, the association of parental education and completed schooling will be
understated by the cross tabulations in the table, which shows the relation of parents’ education to current
grade.
30
This is a sample censoring problem and is explicitly addressed in the ordered probit model of grade
attainment discussed in the previous section. Even with censoring, the table indicates a substantial positive
impact of parental education on children’s schooling, especially for girls. Girls whose mothers have less than
a completed primary education have, on average, 3.3 years of school compared with about 4.4 years for
boys, but girls with mothers who have a completed sec- ondary education have about 5.5 years of schooling,
similar to boys. For father’s education the pattern is simi- lar. These figures thus suggest that improvements in the
education of parents are associated with greater increases in the schooling of daughters than of sons, an issue
addressed in the estimations discussed below.
5. Empirical results