Authors:Robert Ashmead (U.S. Census Bureau) (Speaker) Eric Slud (U.S. Census Bureau)
In this talk we consider the dual problems of choosing between competing small area models and validating model assumptions in an area-level model. Many classes of small area models result in an estimate that is a convex combination of the direct and the synthetic estimate for a given area. Therefore, competing models may share the same direct estimates, but give different synthetic estimates as well as relative weight on the estimates. We discuss diagnostics for model validity and goodness of fit to choose between competing models. These methods include parametric bootstrap methods and those based on cross-validation, which are not ordinarily used in small area estimation. We use the example of small area models related to the U.S. Voting Rights Act Section 203, which are used to the estimate the number of limited-English proficient and illiterate persons in certain language minority groups by jurisdiction using data from the American Community Survey.