Process Mining techniques have been successfully applied as a datadriven and domain-aware approach for improving business process performance in several organizations. Among its applications, Deviance Mining aims at uncovering the reasons why a subset of the executions of a business process deviate with respect to its expected or desirable outcomes, thus producing insights towards improving the process operation, such discoveries can be made using treatment learning techniques, which identify the sets of attributes that are most influential in the results. However, despite the fact that real-life processes are typically composed by events with non-instantaneous duration (nonatomic events), existing approaches for process mining and deviance mining in particular only deal with atomic events in their experiments. This work proposes a domain-driven method for automatically detecting deviations in processes composed by non-atomic events. The method uses the temporal dimension of non-atomic events to apply deviance mining, generating insights on how the duration and the simultaneous occurrence of events generate deviations and how these deviations affect the results of the processes. The method was successfully applied in the COVID-19 domain, to find which domain-specific sequences of nonpharmaceutical interventions mostly contributed to slowing down the rate of COVID-19 cases in countries around the world.