Measuring Normative and Descriptive Biases in Language Models Using Census Data