45 E. Funkhouser Economics of Education Review 14 1999 31–50
similar teenagers one year younger. Each of these entries is negative, large, and statistically significant.
In the next rows of the table, the birth order of the teenager within the group of teenagers aged 12–17
within the household is shown.
16
For example, 29.7 per- cent of the teenagers in the sample for rural areas had
an older sibling aged 12–17. These teenagers are 2.0 per- centage points more likely to attend school than their
older siblings. Continuing down the column, third-oldest teenagers in rural areas are 1.0 percentage points more
likely to attend school than the second-oldest and 3.0 percentage points more likely than the oldest sibling. The
data show that younger teenagers are more likely to attend school than their older school age siblings.
The next rows of Table 2 show the effect of previous education on current school attendance. These coef-
ficients show that teenagers in the sample tended to com- plete schooling in the curriculum cycles of the edu-
cational system in Costa Rica 1–6, 7–11.
17
There are large drops in the probabilities of current attendance for
those who had completed zero years of education, six years of education, and eleven years of education.
4.3. Year effects In Fig. 3, the pattern over the years of recession
showed a drop in school attendance in rural areas and little change in urban areas. The final coefficients in
Table 2 show these year effects in school attendance with controls for the other variables included. Though not
entirely attributable to regional or macroeconomic con- ditions because there may be time-varying household
characteristics not otherwise included, these coefficients provide a rough indicator of the importance of factors
outside the household in explaining school attendance.
The results indicate that there is not much variation in school attendance that is explained by changing con-
ditions over the early 1980s in urban areas, but that these factors are important in explaining the drop in school
attendance in rural areas. Though the general pattern for rural areas in the table is similar to that in the previous
figure — there is a one-time drop in attendance rates without much rebound — the timing of the change is
one-year later with controls than that without controls. In rural areas, there is a 7.1 percentage point drop in
attendance rates observed between 1981 and 1982 in Fig. 1. In contrast, between these two years there is no change
in the year effect with controls in Table 2. The drop in the year effect occurs instead between 1982 and 1983.
16
This is not necessarily the true birth order of the teenager since siblings older than 17 or siblings that do not live in the
household are not counted.
17
Note that the marginal probabilities are calculated relative to one fewer year of education.
This pattern is explained by changes in the mean values of other explanatory variables in directions associated
with lower school attendance between 1981 and 1982 and in directions associated with higher school attend-
ance between 1982 and 1982. In particular, a worsening of the predicted own wage and the changing age distri-
bution within the category of teenagers between 1981 and 1982 were the largest contributors to lower attend-
ance between these years.
Expressed in terms of changes in the standardized lat- ent variable in the attendance decision, changes in the
predicted own wage contribute 2 0.302 compared to 0.048 for the year effect between 1981 and 1982. Other
teenager characteristics, mainly age, contribute 2 0.180; household economic characteristics contribute 2 0.061;
and other household characteristics contribute 2 0.041. Between 1982 and 1983 the contribution of all of these
categories is towards higher school attendance. The year effect, though, is a reduction of 0.516 in the latent vari-
able — or 6.5 percentage points in attendance rates — between 1982 and 1983.
18
5. Extensions
In this section, I examine several extensions of the results presented in Section 3. The results of these exten-
sions are shown in Table 3 for the rural sample in Panel A and the urban sample in Panel B. Though only a subset
are reported, all of the control variables from Table 2 are included in each specification.
First, separate estimates of the determinants of school attendance are presented for males and females in col-
umns 1 and 2. Second, the possibility that the relationship between previous education on current
attendance is masking other relationships is examined with two approaches. The first is designed to counteract
the low attendance of older teenagers with low previous education not captured by age and education dummy
variables separately. In column 3, interactions for each age and years of education combination are included.
The second approach, in Column 4 is the restriction of the sample to households with only one teenager and is
designed to counteract the natural relation between second teenagers and lower levels of education. In col-
umn 5, controls for geographic fixed effects at the can- ton level are included. Finally, in column 6, controls
for selectivity in the employment decision are included in the calculation of the predicted values of the opport-
unity wage of teenagers.
The patterns in the determination of school attendance in these data are quite robust to specification and sample.
18
The lag in the macroeconomic effect suggests a role for political economy to explain some of these patterns.
46 E. Funkhouser Economics of Education Review 14 1999 31–50
Table 3 Extensions
Male Female
Age-education Single teenager
Canton fixed Heckman corr.
inter. households
effects 1
2 3
4 5
6
Panel A: Rural Predicted own
1.069 0.268 0.324 0.244
1.121 0.230 0.613 0.446
0.285 0.248 0.121 0.076
hourly wage Household
characteristics Head of
0.500 0.143 0.164 0.134
0.296 0.099 0.179 0.195
0.287 0.099 0.299 0.097
household working
Log of head 0.010 0.040
0.029 0.036 0.023 0.027
0.077 0.053 0.025 0.027
0.031 0.026 hourly income
Any other 21.055 0.138
20.414 0.129 20.709 0.095
20.890 0.195 20.662 0.095
20.711 0.093 members working
Number of non- 20.114 0.040
20.076 0.039 20.091 0.028
20.030 0.076 20.051 0.029
20.082 0.028 head members
working Log of 2nd
0.154 0.041 0.175 0.041
0.177 0.029 0.288 0.060
0.132 0.030 0.181 0.029
earner’s hourly income
Log of other 0.144 0.084
20.058 0.077 0.035 0.057
20.012 0.145 20.008 0.058
0.031 0.057 income
Other income 1.071 0.627
20.367 0.583 0.284 0.432
0.205 1.113 20.044 0.438
0.269 0.426 missing
Number of non- 0.363 0.033
0.209 0.031 0.286 0.022
0.342 0.052 0.296 0.023
0.288 0.022 workers
Presence of 20.251 0.087
20.061 0.080 20.160 0.059
0.047 0.112 20.146 0.060
20.158 0.059 children , 12
Head education 0.093 0.015
0.108 0.013 0.107 0.010
0.104 0.019 0.094 0.010
0.108 0.010 Head female
0.331 0.119 0.264 0.109
0.278 0.081 0.347 0.150
0.197 0.082 0.289 0.080
Year: 1981 data
0.408 0.144 0.324 0.134
0.488 0.108 0.360 0.215
0.288 0.110 20.107 0.198
1982 data 0.450 0.204
0.197 0.183 0.600 0.162
0.438 0.317 0.121 0.172
20.035 0.085 1983 data
20.102 0.137 20.213 0.142
0.028 0.111 20.003 0.216
20.206 0.116 20.405 0.092
1985 data 20.162 0.129
20.076 0.128 20.005 0.095
20.029 0.186 20.147 0.098
20.436 0.168 -Log likelihood
2 361.6 2 766.3
5 097.2 1 284.5
5 016.5 5 173.8
N 6 398
6 634 12 961
3 104 13 032
13 032
Panel B: Urban Predicted own
0.772 0.337 20.021 0.317
0.452 0.283 0.749 0.496
20.030 0.309 20.020 0.099
hourly wage Household
characteristics Head of
0.324 0.176 20.022 0.167
0.151 0.122 20.106 0.205
0.135 0.122 0.159 0.121
household working
Log of head 0.091 0.045
0.076 0.041 0.077 0.031
0.146 0.053 0.085 0.031
0.081 0.030 hourly income
Any other 21.156 0.185
20.611 0.180 20.829 0.129
20.519 0.227 20.849 0.129
20.841 0.128 members working
Number of non- 20.074 0.054
20.005 0.059 20.043 0.040
20.029 0.079 20.039 0.040
20.049 0.039 head members
working
47 E. Funkhouser Economics of Education Review 14 1999 31–50
Table 3 Extensions
Male Female
Age-education Single teenager
Canton fixed Heckman corr.
inter. households
effects 1
2 3
4 5
6 Log of 2nd
0.247 0.051 0.179 0.053
0.205 0.037 0.167 0.067
0.209 0.037 0.216 0.037
earner’s hourly income
Log of other 20.084 0.096
0.009 0.109 20.035 0.072
0.031 0.150 20.037 0.073
20.032 0.072 income
Other income 20.201 0.730
0.176 0.821 20.016 0.547
0.542 1.175 20.005 0.554
0.028 0.543 missing
Number of non- 0.395 0.042
0.315 0.042 0.359 0.030
0.420 0.054 0.356 0.030
0.363 0.030 workers
Presence of 20.168 0.098
20.024 0.095 20.082 0.069
20.100 0.118 20.090 0.069
20.094 0.068 children , 12
Head education 0.101 0.014
0.107 0.014 0.101 0.010
0.115 0.017 0.095 0.010
0.103 0.010 Head female
0.210 0.123 0.319 0.119
0.258 0.086 0.400 0.142
0.232 0.086 0.256 0.085
Year: 1981 data
0.158 0.173 0.058 0.174
0.142 0.128 0.297 0.241
0.014 0.131 20.081 0.267
1982 data 0.250 0.236
20.219 0.221 0.047 0.185
0.125 0.326 20.212 0.197
20.208 0.111 1983 data
0.119 0.182 20.315 0.201
20.056 0.151 0.093 0.265
20.254 0.160 20.230 0.113
1985 data 0.136 0.157
20.152 0.164 0.034 0.115
0.101 0.197 20.037 0.117
0.004 0.223 -Log likelihood
1 606.4 1 690.3
3 282.0 1 100.9
3 286.8 3 320.6
N 4 634
4 694 9 340
3 296 9 340
9 340
The main patterns are similar to those seen in Table 2. There are several findings that are common to all of these
extensions. First, the presence of working household members other than the head is negatively related to
school attendance. Second, the income of the second earner is positively related to school attendance. Third,
the number of non-working persons age 18 and above in the household is positively related to attendance. Fourth,
the education of the household head is positively related to school attendance. And fifth, teenagers in female-
headed households are more likely to be enrolled in school than other teenagers.
19
Sixth, the return to edu- cation is not significantly related to school attendance.
In addition, these extensions confirm differences between rural and urban areas seen in Table 2. First,
employment of the household head is positively related to school attendance in rural areas and has little relation-
ship in urban areas. Second, the presence of children under 12 in the household is negatively related to school
attendance in rural areas and has little relationship in urban areas. Third, hourly income of the head of the
household is not related to school attendance in rural areas and is positively related to school attendance in
urban areas.
In columns 1 and 2, separate logit estimation was conducted for the sample of males and the sample of
females. In the comparison of the two columns in each panel, the main finding is the positive relationship
between predicted own hourly wage and male school attendance while the relationship is insignificant for
females. Somewhat surprisingly, the presence of children less than 12 in the household is significantly negatively
related to school attendance for males, but not for females.
To allow for the possibility that teenagers are unlikely to return to school once they have missed a year, I allow
for separate intercepts for each completed year of school interacted with age in column 3. In this and subsequent
columns, the sample includes both males and females. The inclusion of these interactions results in little change
in the estimated coefficients for all variables. Similar results are obtained when the sample is restricted to teen-
agers that are within one year of the appropriate com- pleted years of education for the teenager’s age. As an
additional test to verify that a relationship between age and attendance based on completed years of schooling
has not been imposed, I restrict the sample to households in which only one teenager is reported in column 4.
Again, there is little change in any of the coefficients on the household variables.
20
In column 5, I allow more detailed controls for geo-
19
Characteristics of the head may reflect household compo- sition since the head does not have to be parent of the child.
20
The one exception is in the effect of children under 12 in the household in rural areas, which goes from being signifi-
cantly negative to insignificantly positive.
48 E. Funkhouser Economics of Education Review 14 1999 31–50
graphic fixed effects by including dummy variables at the level of the canton. With the inclusion of these
dummy variables, all variables should be interpreted as deviations from canton means. The main finding from
this column is unsurprising result that the variable that is calculated at the canton level predicted own wage
has less explanatory power when canton fixed effects are included. Otherwise, it is not the case that there is bias
from unobserved geographic characteristics at the can- ton level.
As a final extension, I estimate the hourly wage func- tions of Eq. 1d with a correction for self-selection using
Heckman’s two-stage approach in column 6 of each panel. With the exception of the coefficient and standard
error on this variable, there is little change in the esti- mated coefficients on the household variables once this
correction is included.
As in the previous section, the coefficients on the year dummy variables indicate the effect of macroeconomic
factors on school attendance. These coefficients do differ by specification and indicate that the overall patterns
over time mask opposing tendencies for males and females. In both urban and rural areas, the drop in year
effects between 1981 and 1982 in the raw data is seen also in the regressions for females column 2. In con-
trast, there is no drop for males between 1981 and 1982 and the drop in year effects between 1982 and 1983 seen
in the aggregate regressions of Table 2 is seen in the regressions for males column 1. In the remaining col-
umns, the pattern in rural areas is similar to that pre- viously seen in Table 2 — with a drop in attendance
between 1982 and 1983. In urban areas, columns 3–6 provide additional evidence that there was a drop in
school attendance controlling for other factors between 1981 and 1982 that had been recovered by 1985.
6. Concluding remarks