Can We Infer Social Preferences from the Lab? Evidence from the Trust Game