Model composition plays a key role in many tasks in model-centric software development, e.g., evolving UML diagrams to add new features or reconciling models developed in parallel by different software development teams. However, based on our experience in previous empirical studies, one of the main impairments for the widespread adoption of composition techniques is the lack of empirical knowledge about their effects on developers' effort. This problem applies to both existing categories of model composition techniques, i.e., specification-based (e.g., Epsilon) and heuristic-based techniques (e.g., IBM RSA). This paper, therefore, reports on a controlled experiment that investigates the effort of (1) applying both categories of model composition techniques and (2) detecting and resolving inconsistencies in the output composed models. We evaluate the techniques in 144 evolution scenarios, where 2,304 compositions of elements of UML class diagrams were produced. The main results suggest that (1) the employed heuristic-based techniques require less effort to produce the intended model than the chosen specification-based technique, (2) there is no significant difference in the correctness of the output composed models generated by these techniques, and (3) the use of manual heuristics for model composition outperforms their automated counterparts.