Abstract:
|
In today's survey environment, we use various types and numbers of variables for survey sample design, as well as survey data collection and adjustment. Adaptive survey designs seek to use paradata to customize survey protocols based on various scenarios that can, in part, be driven by estimated behavior modeled using available frame and paradata. Nonresponse adjustments are often restricted to including a limited number of frame variables, but in some sampling scenarios such as ABS sampling, more geographic and demographic variables can be appended to samples, thus enlarging the number of possible predictors for nonresponse adjustment models. At the design stage, several of these geographic variables also may be appended to sampling frames, allowing for a more diverse set of variables to be included in the construction of sampling strata, PSUs, and overall sampling strategy. In this roundtable we will discuss various ways machine learning methods have and can be used throughout the survey data-collection process and survey weighting adjustment phases, including random forests, cluster analysis, and k-nearest neighbors approaches.
|