Hierarchical or multilevel linear models are widely used for longitudinal or cross-sectional data on students nested in classes and schools, and are particularly important for estimating treatment effects in cluster-randomized trials, multi-site trials, and meta-analyses. The models can allow for variation in treatment effects, as well as examination of the reasons for treatment effect variation. In this paper the authors propose a method that pulls the group-level standard deviation estimate off the boundary while producing estimates that are consistent with the data. The idea is to specify a weakly informative prior distribution for the standard deviation and to maximize the resulting posterior distribution, a method that can also be viewed as penalized maximum likelihood estimation.