Spoken dialog systems (SDSs) have to respond adequately in many different situations to a multitude of different, partly misrecognized user inputs. Thus, user simulation is a valuable means to test such systems during design time. Although the user models used for the simulation are often incomplete and not always accurate, the simulated data contain much of the information found in a user test (Engelbrecht, 2012). Thus, next to reducing the effort to adapt the models to new systems, an interesting research question is how to analyze large amounts of generated data efficiently. This paper contributes to two types of analysis, namely design error detection, and prediction of perceived system quality.