|
Abstract:
|
We consider surveys that are conducted in a hierarchical fashion, where a country is partitioned into census divisions, states are chosen randomly from these divisions, then counties are randomly selected from each chosen state, and finally survey units are selected from each chosen county. To estimate a county-level parameter, we propose a three-fold sub-subarea level model with a division-level random effect and a state-level random effect. To obtain a parsimonious model, we proposal two easy-to-implement transformation-based methods to select covariate variables. One transformation is model-parameter dependent and the other transformation is model-parameter free. The effectiveness of the two variable selection methods will be assessed by a simulation study.
|
