Experimental Design
Target schools
Our target population consists of Finnish-speaking middle schools with high dropout rates located in 16 municipalities (Espoo, Helsinki, Hyvinkää, Hämeenlinna, Jyväskylä, Järvenpää, Kirkkonummi, Kotka, Lahti, Lappeenranta, Oulu, Porvoo, Salo, Tampere, Turku, Vantaa). We exclude speciality schools such as Steiner schools, schools exclusive targeted to disabled students and schools that do not have at least two parallel classes for grades 7-9. In total, our target population consisted of 92 schools that had 480 regular classes and approximately 8,400 ninth grade students.
School-level randomization
We used randomized block design to divide schools into treatment and control groups. We created stratas using school-level dropout rates constructed with data from Statistics Finland's Sijoittumispalvelu. "Dropouts'' were defined as individuals who had not obtained a secondary degree and were not enrolled in any school four years after graduation. Within each municipality, we ranked the schools according to their earlier dropout rates and divided them into bins of two schools. The randomization for the first two rounds were conducted simultaneously. For these rounds, we also randomized the bins into those treated during the 2016--2017 and 2017--2018 academic years. The motivation for this step was to ensure that we do not create correlation between the year the school was treated and school "quality''. From each bin, we then randomized one school to become a treatment school, while the other one would become a control school. The school-level randomization of the remaining two rounds were conducted during the spring term of the academic year preceding the intervention.
Within-school randomization
Once the schools were recruited to participate, we randomized classes in the treatment schools into intensive and light treatment groups (see above). In the control schools, we randomized classes into a group for which we conducted the same survey as in the treatment schools and into "pure control", which was not contacted at all.