We analyze the impact of a national large-scale desegregation program, targeting a greater mixing of students from different social backgrounds in middle schools, on friendship networks. We compare students in sites covered by the desegregation program (“treatment” group) with students in “matched” sites that are not covered (“control” group). We first document significant homophily with respect to socio-economic status in control schools. We then assess the effect of the program on friendship networks, finding that status homophily is higher in treated schools, which have more diverse student populations. Both baseline homophily and the increase in homophily due to the treatment reduce the effectiveness of the program in fostering more diverse friendships. We propose a novel decomposition of the treatment effect into a composition and a homophily effect, and we develop a new methodology to account for censoring in the econometrics of network formation. (work with Ghazala Azmat, Yann Bramoullé, Élise Huillery, Aristide Houndetoungan, Youssef Souidi.