Is Inaccuracy on Factual Survey Items Item-Specific or Respondent-Specific?

The assumption that reporting errors are uncorrelated across survey items was tested using data from the 1949 Denver Community Survey. Respondent reports to 14 questions in that study were later validated with official records. Inaccuracy was found to be item-specific for questions about seven generally unrelated subjects. By contrast, for seven questions on electoral behavior, all of which were significantly associated, the assumption of uncorrelated errors was clearly violated: respondents inaccurate on one of the seven were disproportionately inaccurate on each of the other six. This held for groups defined by sex, age, education, and political interest. The correlated error term, combined with a tendency for levels of inaccuracy to be greater for those who expressed higher political interest, substantially increased the size of the relationship between voting and political interest. The implications of these results for survey research are discussed.
The author is Director of the Detroit Area Study in the Sociology Department, at the University of Michigan. This paper was begun while he was at the University of North Carolina's Institute for Research in Social Science. Patricia Rector rendered excellent assistance in setting up the computer work, and George Rabinowitz provided valuable help with part of the analysis. Duane Alwin, Frank Andrews, James House, and Howard Schuman offered good criticisms of an earlier draft. [ABSTRACT FROM AUTHOR]