Abstract
In the present work, I adopt the cognitive hierarchy approach to analyze the centipede game. To this end, I present and study an extensive-form version of Camerer et al.'s (2004) original normal-form model. The resulting predictions are evaluated empirically using laboratory data borrowed from a previously published experiment. The paper features two main contributions. First, it presents a parsimonious model that can, in principle, be generalized to any two-person extensive-form game of perfect information. Secondly, it demonstrates that in the centipede game the cognitive hierarchy approach leads to predictions which are not fully backwardly inductive and that can help to explain some key feature of the experimental data.