When randomized experiments cannot be conducted in practice, propensity score (PS) techniques for matching treated and control units are frequently used for estimating causal treatment effects from observational data. Despite the popularity of PS techniques, they are not yet well studied for matching multilevel data where selection into treatment takes place among level-one units within clusters. For instance, students self-select into treatment conditions within schools (or teachers or classrooms). The authors investigate two different strategies for matching level one units (students): (i) within-cluster matching where matches are only formed within clusters (schools) and (ii) across-cluster matching where treatment and control units may be matched also across clusters. Using a simulation study, the authors show that both matching strategies are able to produce consistent estimates of the average treatment effect. However, across-cluster matching requires stronger assumptions than within-cluster matching. They also demonstrate that a lack of overlap between treated and control units within clusters cannot directly be compensated by switching to a between-cluster matching strategy. Generation of simulated target populations and a table are appended.