ABSTRACT Fatigue assessment in industry is often based on simple empirical models, such as the Wöhler curve or the Paris' law. In contrast, fatigue research to a great extent works with very complex models, far from the engineering practice. One explanation for this discrepancy is that the scatter in service fatigue obscures many of the subtle phenomena that can be studied in a laboratory. Here we use a statistical theory for stepwise regression to investigate the role of scatter in the choice of model complexity in fatigue. The results suggest that the amount of complexity used in different design concepts reflects the appreciated knowledge about input parameters. The analysis also points out that even qualitative knowledge about the neglected complexity may be important in order to avoid systematic errors.