The evaluation of interventions such as active labor market policies or medical programs by means of a randomized controlled trial is often considered the gold standard. However, randomized experiments might face severe shortcomings especially if performed at the group level. One such problem is caused by small sample size which might prevent the experiment from developing its fundamental virtue in balancing all relevant covariates. This paper investigates the potential and limits of experimental and non-experimental approaches to the evaluation problem, in particular the use of instrumental variables, in a numerical simulation study, against the particular background of community-based interventions. In our simulations, we emphasize the trade-off between bias and precision by imposing a smaller number of communities whenever we model a randomized experiment, and by allowing for a correspondingly larger number of communities in all cases where selection into the program is not controlled completely by the analyst.