Mediation analysis Notes: This figure shows the results of our mediation analysis. Panel (a) shows the decomposition of the overall treatment effect. In Panel (a), we include socio-emotional skills, parenting practices, and behaviour in class as mediators. Panel (b) shows the decomposition of socio-emotional skills. We decompose the treatment effect obtained from the unconditional outcome equation $Yis=β1PATHSs+Xis′γ+θ+εis$ shown in equation (1) in the following way: $dY/dPATHS=∑(∂Y/∂M)(∂M/∂PATHS)+R$⁠, where Y is the outcome, $PATHS$ is the treatment indicator, M is a vector of k mediators (comprised of the variables included in socio-emotional skills, parenting practices, and behaviour in class), and R is the unexplained part of the treatment effect. We estimate two additional specifications. First, we estimate the conditional outcome equation augmented with the vector M: $Yis=β2PATHSs+Misφ+Xis′γ+θ+ϵis$⁠. Second, we separately estimate the treatment effect of the intervention on each mediator $j∈k$⁠: $Misj=β3jPATHSs+Xis′γ+θ+vis$⁠. Given the longitudinal nature of our data, for all mediators we only consider measures obtained post-treatment and before the education outcome is measured. In case of multiple observations for the same mediator, we construct a summary index using the covariance weighting procedure discussed in Anderson (2008). The contribution of each mediator $j∈k$ is then computed as the ratio $φj×β3j/β1$⁠, which is shown in the colour-coded bars. The unexplained part, R, results from $R=1−∑j=1k(φj×β3j/β1)$⁠.