In this paper we estimate the causal effect of a training program for disadvantaged youths on their long-run labor market outcomes. Individuals receive lottery offers to participate in the program, but are allowed to choose when to leave the program and to participate in alternative programs. We consider a multistage decision setting, where individuals sequentially select which program to participate in at every stage. The standard IV estimator using initial random assignment as instrumental variable identifies a weighted average of the effects of the treatment for subgroups of individuals differing in terms of potential duration of participation and choice of alternative programs. We estimate a sequential choice model that allows us to estimate the effect of the treatment for these different subgroups separately. We use the estimated model to investigate the dynamic complementarity between different training programs and explore program targeting to improve the cost-effectiveness relative to the existing program.