Resumen 

We study how group composition in an existing training program for Chilean low-skilled adult women may impact its effectiveness. We experimentally separated the target population into groups according to their propensity to work and tested whether offering the program to homogenous groups is more efficient than implementing the program with heterogeneous participants. Because we kept group assignment secret we can estimate the pure effect of group composition, with less of a potentially planned response by the instructors. Our evidence suggests, then, that tracking per-se does not pay off for everyone, but there are benefits from having more diverse peers. In particular we find that non-household heads (relatively less experienced) placed in high-labor attachment groups benefits the most. We argue that these results may come from learning from peers, as network formation is unaffected.