A candidate explanation for the persistence of heterogeneous behavior in a sequential social dilemma played many times is the existence of heterogeneous preferences. Preferences-dependent conjectures about opponents’ behavior are an additional source of heterogeneity. By behaving differently, different preference types acquire different information. Thus, when observing only outcomes of own past interactions heterogeneous and possibly wrong conjectures about opponents’ strategies may endogenously arise and persist. In a Centipede game experiment played for forty rounds, we manipulate the type of ex post information and the method of play. We find that, when the game is played in its reduced normal form and subjects have only access to personal statistics, heterogeneity of behavior across preference types persists in the long run. In this case, behavior resembles a self-confirming equilibrium: selfish subjects take at earlier nodes due to an unjustified lack of trust. When subjects have also access to public statistics, heterogeneity disappears: selfish subjects tend to pass more often and play moves towards Bayes Nash equilibrium.