Innovations in Statistical Methodology for Complex Surveys

Iowa State University, Ames IA

Investigators

Abstract

This research project will modernize survey sampling tools for inference from survey data that can be used for complex problems such as survey integration or small area estimation. Increasingly, data sources have characteristics that traditional sampling tools are not readily designed to handle. The project will address important inference problems, such as model selection and hypothesis testing, based on data from complex samples and will extend the methods from parametric models to more flexible nonparametric models, such as quantile regression. The new methodology will have broad applicability to surveys in diverse disciplines, including agriculture, health, and demographics. Products of this research will include resources to equip practicing survey statisticians with tools to better meet the demands of policy makers and the public. Software and metadata produced by this project will be made publicly available. Graduate students will receive education and training, with an emphasis on the relationships between design-based and model-based inference. This research project will develop innovative methods that will enable survey statisticians to exploit modern data structures and models in a statistically defensible way. The project will focus on three areas: (1) the use of inverse sampling and reweighting to obtain valid inferences from nonprobability samples, (2) methods for prediction and analytic inference under complex sample designs, and (3) hierarchical modeling strategies that are feasible to implement with large, diverse data sources. Methods to obtain approximately unbiased inferences from non-probability samples are highly relevant because of increases in nonresponse and use of non-survey data, such as administrative sources and satellite information. This examination of inference under complex sample designs will further research on hypothesis testing and the use of semiparametric models, particularly for situations in which the sample design is informative for the specified model. The investigators will further develop a hierarchical modeling approach that aggregates estimates obtained for sub-divisions of a large data source. The investigators have vetted this procedure using non-survey data and will apply the approach in a large-scale survey context.