|
Abstract:
|
We use donor imputation to construct a database that supports small area estimation. Appropriately weighted sums of observed and imputed values reproduce model-based small area estimates. We develop imputation procedures for both unit-level and area-level models, the two main classes of small area models. Each record in the imputed data set has complete data, an estimation weight, and a set of replicate weights for variance estimation. We compare imputation procedures based on area level models to those based on unit-level models through simulation. We apply the methods to the Iowa Seat-Belt Use Survey, a survey designed to produce state-level estimates of the proportions of drivers and passengers who wear a seat-belt. We develop a bivariate unit-level model to obtain county-level predictors of the proportions of belted drivers and passengers. We impute values for the number of belted/unbelted drivers and passengers onto the full population of road segments in the sampling frame. The resulting imputed dataset returns the county-level predictors based on the bivariate model.
|
