Spline-Based Models for Prediction in Survey Samples

• Activity Number: 333
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Hui Zheng*+
• Companies: Harvard Medical School
• Keywords: spline ; pps sample ; two-stage sample

Abstract

In estimation of finite population quantities from survey samples, design-based methods focus on removing the bias while model-based methods seek to improve prediction of unobserved data.

Spline based models assume flexible mean functions and reduce the potential of model-misspecification, which is often considered a limitation by design-based statisticians. We have demonstrated that model-based method can shorten the confidence (or credible) intervals without sacrificing the coverage property, hence improve the inference quality. By incorporating design information in a spline model, one can achieve design-consistency without additional bias calibration.

I review spline-based methods proposed for various sampling designs including one-stage designs to multi-stage designs with clustered outcomes. I will discuss potential applications of the same philosophy to more complex situations.