On Borrowing Information Over Time in Small-Area Estimation

• Activity Number: 114
• Type: Invited
• Date/Time: Monday, August 10, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): William R. Bell* and Carolina Franco
• Companies: U.S. Census Bureau and U.S. Census Bureau
• Keywords: linear mixed model ; best linear unbiased prediction ; random walk ; bivariate model

Abstract

We examine alternative approaches to borrowing information over time in small area estimation with the goal of improving on estimates from the model of Fay and Herriot (1979) when this model is applied to only current data. We focus on the case of a moderate to large number of areas and a small number of time points, and consider two situations distinguished as models with strong covariates versus models with weak or no covariates. Alternatives considered include autoregressive and random walk dependence structures, as well as a bivariate model applied to current estimates and an average of past estimates. Theoretical calculations indicating how much improvement might be expected out of borrowing information from past data for the alternative models are compared to results from empirical examples.