Abstract:
|
The use of response models to group or stratify members of a target survey population into homogeneous response groups, informed by auxiliary data characterizing the sample units, is a well-established practice to reduce bias and improve reliability of survey estimates. More recently, adaptive survey design methods have extended the use of propensity models to inform tailored design changes. This paper explores the use of various Internet response models to stratify the American Community Survey (ACS) sample frame to enable a tailored initial assignment of the mail and Internet self-response modes. These models include traditional logistic regression as well as some of the more recent machine learning techniques - decision trees, random forests, boosting, support vector machines, and K-nearest neighbors. To inform the models, we augment the ACS sampling frame using administrative records. Using data from the April 2011 ACS Internet Test, we establish that offering mail or a choice of mail and Internet self-response modes are viable options for members of the low Internet response stratum.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.