We present an empirical study that compares mention heads as annotated manually in four coreference datasets (for Dutch, English, Polish, and Russian) on one hand, with heads induced from dependency trees parsed automatically, on the other hand.