Comparing students by a matching analysis – on early school leaving in Dutch cities

Abstract

In case of regional discretionary on the implementation of policy measures, central governments often consider differences in outcomes as an indication that one policy was more effective than another policy. If uniform incentives are provided to motivate regional policy makers, these incentives can be discouraging when the underlying populations differ. Empirically, this study compares early school leaving between the four largest Dutch cities. It shows that considering regional differences as performance measures can be dangerous if differences in population characteristics are not properly taken into account. Methodologically, this study contrasts the use of a traditional probit model with a more advanced iterative matching procedure.

Keywords:

• uniform incentives
• early school leaving
• comparative
• matching analysis

JEL Classification::

• C14
• C61
• C23
• I21

Acknowledgements

We are grateful to two anonymous referees, Sofie Cabus, Wim Groot, Henriette Maassen van den Brink, and seminar participants at NICIS Institute and the Dutch Ministry of Education (OCW) for their constructive remarks. The authors acknowledge financial support of NICIS. The usual caveat applies.

Notes

¹ Dropout is also a major issue in other continents. For instance, US president Obama mentioned in his inauguration speech: ‘Every American will need to get more than a high school diploma. And dropping out of high school is no longer an option’.

² As the examples below indicate, the implications of this article are not restricted to the educational sector, but can straightforwardly be extended to other sectors. To facilitate the use of the iterative matching procedure, the Stata code is available upon request.

³ Following the official Dutch and European definition, we define an early school leaver (or dropout) as a person younger than 23 who leaves education without a higher secondary education degree.

⁴ The number of dropouts is more easily observed at post-secondary level. Therefore, a significant part of older literature is focussing on the dropout of university students as data are readily available by university professors.

⁵ The body of this article includes Amsterdam and Rotterdam as main focus. These cities are the largest two cities of the Netherlands, which arguably received most attention from policy makers. In Appendix, we extend the data to two other large Dutch cities: Utrecht and The Hague.

⁶ Robustness tests pointed out that replacing the more detailed ethnicity variable by a migrant dummy variable (i.e. native or migrant) does not significantly change the results.

⁷ We extensively analysed alternative probit specifications as robustness tests (e.g. including birth year of the student, detailed information on ethnicity or excluding variables). However, the difference between Amsterdam and Rotterdam proved not to be an artefact of the model specification but consistently remained significant under all specifications.

⁸ Kernel estimators or matching estimators based on a propensity score are not necessarily inferior to Mahalanobis matching. Each matching method has its own advantages and disadvantages, and for an elaborate description of the available matching methods we refer to Cameron and Trivedi (2005). As a robustness check, we matched students based on a conditional probability of living in Rotterdam and based on a kernel function and we found that the results and conclusions were similar to results with Mahalanobis matching. We matched students from Rotterdam to one, five and 10 students from Amsterdam using the propensity score, and we matched students on the basis of caliper and kernel matching. When we match on the propensity score, we match on the conditional probability, p(x), that a student lives in Rotterdam given x. The matching set is then A_i (p(x)) = {p_j min _j ‖p_i  − P_j ‖}. Caliper matching is essentially a propensity score matching estimator where we impose that p_i  − p_j  < ε. For ε we take the values 0.05 and 0.01. When we performed Kernel matching, we used an Epanechnikov kernel funtion with 0.6 as bandwidth and the weight that defines the Kernel matching estimator is then w_{i ,  j}  = . The outcomes of the alternative matching models are available upon request.

⁹ Remark that in the matching analysis, we experimented with various robustness test. For example, we included control variables such as age, detailed information on ethnicity and family status (divorced or married parents). The results did not significantly change.