Despite the popularity of propensity score (PS) techniques they are not yet well studied for matching multilevel data where selection into treatment takes place among level-one units within clusters. This paper suggests a PS matching strategy that tries to avoid the disadvantages of within- and across-cluster matching. The idea is to first identify groups of clusters that are homogenous with respect to the selection model, and then to estimate the PS and treatment effect within each of the homogeneous group. The purpose of the study is to demonstrate that across-cluster matching within homogenous groups of clusters is less prone to bias than within-cluster matching and complete across-cluster matching (across the whole population of clusters). The results indicate that a matching strategy that first groups the data into homogeneous groups of clusters and then estimates the treatment effects via across-cluster matching within each of the groups can outperform within-cluster matching and across-cluster matching without any grouping information. Tables and figures are appended.