Estimation of measurement bias using a model prediction approach

Abstract

It is well-known in the survey literature that when responses are obtained from respondents in sample surveys, the actual observed values of measured characteristics may differ markedly from the true values of the characteristics. Evidence of these socalled measurement errors in surveys has been collected in a number of ways. For example, the recorded response may be checked for accuracy against administrative records or legal documents within which the true (or at least a more accurate) value of the characteristic is contained. An alternative means relies on revised reports from respondents via reinterviews. In a reinterview, a respondent is recontacted for the purpose of conducting a second interview regarding the same characteristics measured in the first interview. Rather than simply repeating the original questions in the interview, there may be extensive probes designed to elicit a more accurate response, or the respondent may
be instructed to consult written records for the "book values" of the characteristics. For some reinterview surveys, descrepancies between the first and second interviews are reconciled with the respondem umil the
interviewer is satisfied that a correct answer has been obtained. Forsman and Schreiner (1991) provide an overview of the literature for these types of
reinterviews. Other means of checking the accuracy of survey responses include" a) comparing the survey statistics (i.e., means, totals, proportions, etc.) to statistics from external sources that are more accurate;
b) using experimental designs to estimate the effects on survey estimates of interviewers and other survey personnel; and c) checking the results within the same survey for intemal consistency.