Abstract:
|
Sample attrition is a common problem in longitudinal studies. This study examines how paradata, particularly observations documented by interviewers in the course of data collection, can improve the accuracy of prediction of attrition from one round of data collection to the next. We modeled round-to-round attrition using questionnaire data and paradata from rounds of data collection in the 2016 Medicare Current Beneficiary Survey (MCBS), a survey of a nationally representative sample of the Medicare population, conducted by the Centers for Medicare & Medicaid Services (CMS) through a contract with NORC at the University of Chicago. We show that attrition can be effectively predicted with a model incorporating covariates such as number of contact attempts, length of previous interview, and respondent demographics. The predictive validity of the model was improved by adding predictors related to interviewer observations from the previous survey round. We describe the process of transforming text from interviewer observations into model predictors and explore how other longitudinal studies could use such data.
|