We analyze how different previous roles as partners or strangers in public good games affect an individual's subsequent cooperation in a partner setting. We systematically vary a group's composition from all individuals being partner over blended groups of partners and strangers to all individuals being stranger in each round. Our results show that previous group composition does not affect cooperation in the subsequent partner setting with one exception: Groups cooperate significantly less compared to all other settings, when one stranger entered the group. We further analyze this situation in-depth and find that individuals may labor under an ultimate attribution error: They feel that the newcomer is a "bad apple". The cooperativeness towards the newcomer, but also among oldtimers is disturbed in this case. We conduct additional treatments to back up this result and to show how certain information can prevent such an error.