Members of a high quality, probability-based Internet panel were randomly assigned to one of two modes: (1) computer assisted telephone interview or (2) web survey. Within each mode the same series of split ballot experiments on question format were conducted. We tested the effect of unfolding opinion questions in multiple steps vs. asking a complete question in one-step, and the effect of fully verbal labelling versus end-point labelling of response categories within and between the two modes. We found small direct mode effects, but no interaction. Unfolding (two-step question) had a larger effect. When using a mixed-mode design it is advisable to avoid unfolding formats in the telephone interview and use the complete (one step) question format in both modes. Full labelling is mildly preferred above labelling endpoints only. The absence of an interaction effect is an encouraging result for mixed-mode surveys.