Relationship Between Measurement Error and Unit Nonresponse in National Probability Surveys: Estimation and Evaluation of Measurement Error in the Absence of Validation Data
Peytchev, A. A., & Peytcheva, E. (2007, May). Relationship Between Measurement Error and Unit Nonresponse in National Probability Surveys: Estimation and Evaluation of Measurement Error in the Absence of Validation Data. Presented at American Association for Public Opinion Research Conference, Anaheim, CA.
Response rates in national probability surveys are falling despite higher cost of data collection from greater field efforts. The inherent threat of lower response rates is the increased potential for nonresponse bias. Some have argued that nonrespondents in a survey would be poor respondents if their cooperation is gained. If true, the higher cost per interview for cases that have already received a lot of effort could be better directed to other areas of the survey process. However, if such respondents do not exhibit more measurement error after controlling for cognitive proxies such as education and age that are can be related also to nonresponse, nonresponse bias should remain as the optimization criterion in decisions during data collection.Bringing nonresponse and measurement error combines two estimation problems in sample surveys. Nonresponse is a problem of missing survey data, while measurement error requires survey and validation data.This study uses a model for nonresponse, and proposes a model for measurement error in the absence of auxiliary information. The technique involves the simultaneous estimation of means and variances in purposefully constructed models, and provides respondent-level estimates of measurement error. These estimates provide the potential for studying linkages between survey errors, including the identification of preferable measurement conditions such as particular interviewers inducing less measurement error.The estimation methods and findings on the relationship between nonresponse and measurement error are demonstrated on two national probability surveys. Particular attention is given to the relationship between nonresponse and means (i.e., bias in point estimates) and between nonresponse and measurement error. Preliminary analysis suggests higher measurement error elicited from respondents rated as less cooperative in a previous survey administration. Implications and needed future developments are discussed.