This paper is based on research about solving Bayesian problems, with instructions presented in natural frequencies or conditional probabilities (Gigerenzer et al. 1995; Zhu & Gigerenzer, 2006). In addition, it is attempted to use the theory of knowledge spaces to generate and validate a structure of the underlying skills needed to solve such problems, using the competence-performance approach (Korossy, 1996). According to previous research, Bayesian problems represented in percentages are less likely to be solved than those represented in natural frequencies. The conducted study shows similar results. The analysis of the established competence-performance-model shows several infractions of the model. The model is not superior in explaining the participants answer patterns to a linear modelling based solely on task difficulty.