Educational systems typically contain a large pool of items (questions, problems). Using data mining techniques we can group these items into knowledge components, detect duplicated items and outliers, and identify missing items. To these ends, it is useful to analyze item similarities, which can be used as input to clustering or visualization techniques. We describe and evaluate different measures of item similarity that are based only on learners' performance data, which makes them widely applicable. We provide evaluation using both simulated data and real data from several educational systems. The results show that Pearson correlation is a suitable similarity measure and that response times are useful for improving stability of similarity measures when the scope of available data is small. [For the full proceedings, see ED596512.]