Centipede games represent a classic example of a strategic situation, where the equilibrium prediction is at odds with human behavior. This study is explicitly designed to discriminate among the proposed explanations for initial responses in Centipede games. Using multiple and different Centipede games, our approach determines endogenously whether one or more explanations are empirically relevant. We find that non-equilibrium behavior is too heterogenous to be explained by one unique model. However, most non-equilibrium choices can be fully explained by level-k thinking and quantal response equilibrium, roughly in equal proportions. Preference-based models play negligible role in explaining the non-equilibrium play.