Inference from Complex Sample Surveys: Past Controversies, Current Orthodoxies, Future Paradigms

• Activity Number: 322
• Type: Other
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: ASA
• Author(s): Roderick J. Little*+
• Companies: University of Michigan
• Keywords: Survey Inference ; Calibrated Bayes ; Design-Based Inference ; Superpopulation Models

Abstract

Recent U.S. Census Bureau director Bob Groves argues that government statistics is at a crossroads -- probability samples are increasingly hard and expensive to collect, and combining survey information with data from alternative sources such as the internet or administrative records is increasingly desirable. This situation has revived interest in the inferential basis of survey inference. Survey sampling is perhaps unique in being the only area of current statistical activity where inferences are primarily based on the randomization distribution rather than on statistical models for the survey outcomes. It is an area where the age-old debate between randomization-based and model-based inference is most sharply drawn. I describe the three main inferential approaches to survey sampling -- design-based or randomization inference, superpopulation modeling, and Bayesian inference, and compare them for some basic survey designs. I review the tasty historical debates and controversies surrounding these alternative paradigms, describe and critique current survey inference orthodoxy, and contrast it with calibrated Bayes, which I see as an attractive alternative paradigm for the future.