Well, ideally you can observe behavior for a large group of people and tie that back to observable characteristics (i.e. "black", "poor", "urban"). Then you can

Well, ideally you can observe behavior for a large group of people and tie that back to observable characteristics (i.e. "black", "poor", "urban"). Then you can project characteristics from a new population onto unobservable behaviors (i.e. support a candidate). Pretty simple. Pretty effective. EDIT: I've actually seen "race-from-name" models. Jamaal would totally be black. Despite a lone white Jamaal, I'd still use that model.