Intelligent behavior of adaptive educational systems is based on student models. Most research in student modeling focuses on student learning (acquisition of skills). We focus on prior knowledge, which gets much less attention in modeling and yet can be highly varied and have important consequences for the use of educational systems. We describe several models for prior knowledge estimation--the Elo rating system, its Bayesian extension, a hierarchical model, and a networked model (multivariate Elo). We evaluate their performance on data from application for learning geography, which is a typical case with highly varied prior knowledge. The result show that the basic Elo rating system provides good prediction accuracy. More complex models do improve predictions, but only slightly and their main purpose is in additional information about students and a domain. [For complete proceedings, see ED560503.]