Constructing comprehensive operational models of intended system behaviour is a complex and costly task. Consequently, practitioners have adopted techniques that support partial behaviour description and focus on elaborating these descriptions iteratively. Scenario-based specifications, for example, are incrementally elaborated to cover system behaviour that is of interest. However, how should partial behavioural models described by different stakeholders with different viewpoints be composed? How should partial models of component instances of the same type be put together?In this thesis, we use model merging based on observational refinement as a general solution to these questions, where merging consistent models is a process that results in a minimal common refinement. We prove several mathematical characterizations of merging and consistency, study algebraic properties of the merge operator, and give new and improved algorithms related to constructing merge. Finally, we present a case study that illustrates the utility of our results.