Probing “don't know” answers: effects on survey estimates and variable relationships