From the time the Census Bureau introduced an option to identify with multiple races on its survey questionnaires, researchers within the Census Bureau have sought the best way to aggregate the possible responses into categories while preserving the information from an increasingly multiracial country. Classifying racial data into categories helps provide information to Census stakeholders so they can measure the Census Bureau’s performance in identifying and correctly enumerating each population. As planning intensified for the 2010 Census Coverage Measurement study, research staff analyzed the Matching and Correct Enumeration rates of multiracial populations, in order to model the data.

The paper details the techniques used to build models for Census Coverage data, by applying stepwise regression to the concept of CART modeling to partition the data into cells, and adding information criteria as a method of cross-validation. The paper also addresses: the specific issues inherent in modeling Dual-System Estimation data for this topic, and how they were addressed; the patterns of racial identification that were discovered; and the recommendations that were ultimately proposed.