An investigation of response variance in sample surveys
The dissertation considers response variance in sample surveys in the broader context of survey quality and survey error. Following a historical review of the evolution of both the terms and the concepts a brief overview is given of earlier research in the area. The principal content of the dissertation draws on investigations carried out by the author over the last thirty years. There are three separate strands of argument, each associated with a particular approach to the analysis. First there is the descriptive (simple diagnostic) orientation of establishing the circumstances under which (or if) response variance arises, the associated issue of how it should be accommodated in analysis - primarily estimating the impact on the variance of univariate statistics - and an assessment of its likely order of magnitude. Second, there is the model-assisted orientation which attempts to decompose the effects into their constituent parts: one approach is to incorporate the correlating source (cluster or interviewer for example) as a term or terms in other models that we are estimating so that the effect is incorporated into the estimation of these models; the other is to model the response error itself -- in doing this we are trying to decompose it into its constituent parts. Third, and most radical, is to view error as information. By conceptualizing the process that generated the errors as a substantive process rather than as a set of nuisance effects we can extract from the results of the process information about both the process and the subject matter. Any particular piece of analysis may include any combination of these three approaches. The dissertation draws on special studies incorporated into a number of major sample surveys. Two principal data sets are involved. The first arises from a special investigation of response error carried out in conjunction with the World Fertility Survey; the second is the reinterview data set from the Current Population Survey carried out by the US Bureau of the Census. Four other surveys are used; an absenteeism survey in Ireland, two cross-sectional British surveys (one on Noise Annoyance, the other on Physical Handicap), and a British panel survey (the British Household Panel Survey).