Heterogeneous treatment effects Notes: This figure shows heterogeneous treatment effects for (a) initial tracking into academic high school at age 13, (b) academic high school completion at age 20, and (c) university enrolment or graduation at age 24. The dashed line indicates the baseline treatment effects. We estimate treatment effects by family income, parental education, SES, gender, age, as well as on baseline child SBQ measures for ADHD symptoms (disruptiveness and impulsiveness), opposition and defiance, and non-aggressive conduct disorder. For family income, ADHD symptoms, opposition and defiance, and non-aggressive conduct disorder, the groups low, middle, and high are defined by the respective tertiles of the sample distribution. For parental education, low is defined as both parents not having an academic high school degree, middle is defined as one parent having an academic high school degree, and high is defined as both parents having an academic high school degree. Low SES includes families with incomes in the bottom two quintiles of the sample distribution, with a non-working father, and with both parents not having an academic high school degree or families with income in the lowest quintile of the sample distribution. Middle SES includes families with incomes in the second to fourth quintile of the sample distribution, with a working parent, and with at least one parent that has an academic high school degree or families with incomes in the third quintile of the sample distribution, with a non-working parent, and with both parents not having an academic high school degree. High SES includes families with incomes in the second to fifth quintile of the sample distribution, with a working parent, and with both parents having an academic high school degree or families with incomes in the top quintile of the sample distribution. Estimates are based on models in Table 4 that include the full set of controls. All models include strata fixed effects for the level of randomization. Each point estimate is shown with the respective 90% and 95% confidence intervals calculated based on standard errors clustered at the school level.