Relating Respondent-Generated Intervals Questionnaire Design to Survey Accuracy and Response Rate
S. James Press, Judith M. Tanur
This article is concerned with relating questionnaire design to estimation accuracy and item response rate in sample surveys, in the context of the Respondent-Generated Intervals (RGI) protocol for asking questions. One RGI procedure for asking survey questions is concerned with recall of facts (How many times did you visit your doctor in the last year?). The research addresses the problem that respondents unequal memory abilities may lead to large nonsampling errors (bias). The novelty of this question protocol is asking respondents both for an answer to the recall question and also for the smallest and largest possible values they think the true answer might be. We find that a Bayesian estimator of the population mean is given by a weighted average of the basic responses, where the weights assigned to respondents estimates are larger for smaller interval lengths. We summarize four record-check surveys for which the RGI protocol has been applied. We find that interval length is related to the respondents confidence in his/her answer, and that fine-tuning the way the question is worded is directly related to the response rate, and to the accuracy of population parameter estimates. So by placing strong emphasis upon the questionnaire design we can improve the importance and usefulness of the survey.
Bayesian, brackets, questionnaires, recall, item nonresponse, surveys