Optimal Stratification Based on Nonparametric Regression Methods

• Activity Number: 278
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Enrico Fabrizi*+ and Carlo Trivisano
• Companies: Università di Bologna and Università di Bologna
• Keywords: optimal stratification ; nonparametric regression ; finite population

Abstract

In the design of surveys, stratification is introduced for a number of different purposes, among which the gain in efficiency for design based estimators of population parameters. A stratification providing the most efficient stratified mean estimator is therefore called "optimal." A landmark in the literature about optimal stratification is the Dalenius method. We propose to generalize and extend the Dalenius method to situations in which the target variable (or a proxy) is available for a sample and a large set of covariates is known for each unit in the population. This situation may arise in repeated surveys or when stratifying first-level units in multistage designs. We propose to replace the linear regression which form the kernel of the Dalenius method with more flexible non parametric regression techniques such as MARS, generalized additive models, and regression trees. All evaluations and comparisons are carried out with reference to a large simulation exercise in which different hypotheses on the target variable, the covariates, and the relation between them are considered. Some real-life example is also considered.