Respondent-Related Error

Respondents and Nonresponse Error

Respondents may contribute to nonresponse bias when they refuse to participate in the study or cannot be located or contacted. Nonresponse bias varies as a function of the survey's response rate and the degree to which nonresponders differ from participants. Obtaining high response rates is generally considered an important protection from nonresponse bias. Nonetheless, nonresponse bias may be large, even with a high response rate, if those interviewed differ substantially from those who are sampled but are never contacted or those who refuse; conversely, bias may be small, even with a low response rate, if respondents are similar to noncontacts and refusers on the characteristics of interest. In longitudinal studies, nonresponse error also varies if respondents who drop out of the study systematically differ from respondents who are retained in the panel (i.e. so-called differential attrition).

Attempts to estimate the impact of different levels of response on survey estimates and nonresponse error suggest that improvements in response rates sometimes have a negligible impact on both estimates and error, but the results are unpredictable. At present there is little empirical or theoretical guidance to predict how much nonresponse or under what circumstances nonresponse will produce nonresponse error, but nonresponse can have a big impact on survey error, especially for some subgroups.

Even after they agree to participate in a survey, respondents may fail to provide data within the interview; they may do this intentionally, by refusing to answer questions or saying “don't know,” or unintentionally, by skipping questions on self-administered questionnaires. Item nonresponse occurs more often for some types of questions than others. For example, refusals are more common for questions about income and for questions that ask for sensitive or threatening information. Respondents may also provide responses that are incomplete, for instance, by not providing enough information to an open-ended question to allow their answers to be classified reliably.

...