Swarm-based systems, i.e. systems comprising multiple simple, autonomous and interacting components, have become increasingly important. With their decentralised architecture, their ability to self-organise and to exhibit complex emergent behaviour, good scalability and support for inherent fault tolerance due to a high level of redundancy, they offer characteristics which are particularly interesting for the construction of safety-critical systems. At the same time, swarms are notoriously difficult to engineer, to understand and to control. Emergent phenomena are, by definition, irreducible to the properties of the constituents which severely constrains predictability. Especially in safety -critical areas, however, a clear understanding of the future dynamics of the system is indispensable. In this paper we show how agent-based simulation in combination with statistical verification can help to understand and quantify the likelihood of emergent swarm behaviours on different observational levels. We illustrate the idea with a simple case study from the area of swarm robotics.