The collection of human performance data from nuclear power plant simulators to inform Human Reliability Analysis is receiving renewed attention. Recent efforts by the community addressed the development and application of data collection protocols and the analysis of the first batches of data. The present paper highlights the need to address data variability: simulator data is collected from different plants, scenarios and crews. Within-task and within-factor variability as well as crew-to-crew variability require explicit consideration, to avoid overconfidence in the failure probability values and allow a richer and more realistic representation of the accident progression. This paper presents the use of Bayesian models, in different formulations, to capture variability. Artificial data is used to investigate different data aggregation strategies (e.g. per crew type) and the respective modelling options. A numeric example demonstrates the influence of the different approaches on the estimated HEP uncertainty distributions.