Introduction

Abstract

The basis for unbiased inference from relatively small observed samples to largely unobserved populations is probability sampling. Such samples, in turn, provide the foundation for studies that inform national, regional, and local economic, social, and health policies, as well as marketing and political decisions (e.g., National Research Council 2006). Underlying this inferential process, however, is the assumption that all elements designated for the sample are actually observed or measured. In recent years, increasing violation of this assumption because of an inability to contact sample members or because of their unwillingness to participate has led to increasing attention to survey nonresponse. “Nonrespondents” are those missing from a probability sample; the largest components of nonresponse, and therefore those of greatest concern to survey methodologists, are noncontact and refusal. These components have different causes and may, therefore, have different consequences for survey estimates (Groves and Couper 1998; Groves, this issue). Concern about survey nonresponse is of course not new. Smith (2002, pp. 27–28) notes that “early research extends back to the emergence of polling in the 1930s and has been a regular feature in statistical and social science journals since the 1940s. An analysis of JSTOR statistical journals dates the first nonresponse article from 1945 and the Public Opinion Quarterly index’s earliest reference is from 1948. The index of Public Opinion Quarterly contains 125 articles on this topic; a full-text search of journals covered in JSTOR finds the following number of articles, by subject area, that included the word ‘nonresponse’: political science—62, economics—87, sociology—146, and statistics—431. . . . The vast majority of these articles focus on the main issues in the field: reducing nonresponse, measuring nonresponse bias, and compensating for nonresponse by imputation and/or weighting.” Statisticians have been more concerned with ways of adjusting for the bias introduced by nonresponse; social scientists and survey methodologists have tended to focus on understanding and reducing nonresponse itself.