Our identification strategy requires that fluctuations in the proportion of female stu- dents within a school and within a neighborhood should not be correlated with other cohort-to-cohort changes that could affect students’ education decisions. In partic- ular, we check if changes in the proportion of female students within a school and within a neighborhood are correlated with changes in students’ observable charac- teristics. For the universe of students (N=355,808 students) the only characteristics we know are: the age of students and if a student enrolled early in school. This is the case if a student is born in the first quarter of his birth year.
However for a smaller sample of 45 schools (observations=18,670) we also know the ethnicity of students. In Table 3.4, we present some evidence that the schools in the smaller sample have no different characteristics compared to the whole population. We cannot implement the whole analysis based on this smaller sample because we need the universe of students and schools in order to construct the neighbourhoods and exploit within neighbourhood variation. We use this smaller sample of schools to check if changes in the proportion of girls are correlated with changes in students’ ethnicity and mobility rates.
Tables 3.2 and 3.3 provide evidence on the balancing tests for the whole sample and the sub-sample of the 45 schools. Table 3.2 reports the estimated coef- ficients from the OLS regression and a within school regression (school fixed effects) of students’ characteristics on the proportion of females in each school. We also re- port the estimated coefficients from a within school regression when school specific time trends are added (columns (3) and (6)). Table 3.3 reports the estimated coef- ficients from the within neighbourhood regression (neighborhood fixed effects) with 13This is more understandable when one takes into account that Greece has 227 inhabited islands, most of which are quite far from the mainland and have outdated telecommunications infrastructure (Ellinikos Organismos Tourismou (EOT), ”Greek islands”, April 2012).
(columns (3) and (6)) and without (columns (2) and (4)) adding neighbourhood linear time trends. Again the OLS estimates are reported as a point of comparison. As we notice from these two tables, the proportion of females is not related to most of the students’ characteristics, both in the OLS and the within school/ neighborhood regressions. There are some exceptions in the OLS and within school regression. In particular, the proportion of females within a school seems to be neg- atively correlated with the proportion of students with Polish and Bulgarian origin, however these correlations are reduced and become statistically insignificant when we add school linear time trends. Within neighbourhoods we find no association between the proportion of females within a neighborhood and students’ observable characteristics. All the regressions include year fixed effects. These results suggest that cohort-to-cohort changes in the proportion of female students within a school and within a neighborhood seem to be uncorrelated with changes in students’ ob- served characteristics.
We also examine whether changes in the proportion of female students within a school and within a neighborhood are related to changes in the logarithm of school enrollment. As reported in the first row of Table 3.2 there seem to be a negative association between changes in the proportion of females within a school and changes in the logarithm of school enrolment. Both, the OLS and within school regressions produce estimates negative and statistically significant at 10% . However, this correlation largely reduces and becomes insignificant when school specific time trends are added.
One could still have concerns that students might react to the unpredicted changes in gender compositions. Although students are assigned to schools based on geographical characteristics and it is not easy to switch school, one could still be worried that students might drop out from or switch to another school after being exposed to this information. For example, students who are in schools where the proportion of girls is high/low could drop out. Or transfers of students could be observed that might be correlated to the observed proportion of females in a given school. We address this concern by looking at the correlation between the proportion of female students in a school and the probability that a student drops out from or switch to another school in that year. We use the smaller sample of schools because only for these schools we have data for multiple years and we can identify students who drop out and transfers.
Our dependent variables are: a dummy variable that takes the value of one if the student drops out from school and a dummy that takes the value of one if the student is transferred to this school at the beginning of the school year. Table
3.5 reports the outcome means and the regression estimates separately for boys and girls. The first row in each panel indicates that students’ mobility from and to a school is low. Approximately 8% of boys and girls drop out from school in the twelfth grade and around 6-8% of boys and girls respectively transfer to another school at the beginning of the twelfth grade. The second row in each panel reports the regression estimates when school linear trends as well as school and time fixed effects are added. All estimates are small and statistically insignificant. Overall, changes in the proportion of females within a school seem to be uncorrelated with students’ mobility across schools and drop out rates.