Using a model-aided sampling paradigm instead of a traditional sampling paradigm in a nationally representative establishment survey

Abstract

We compare traditional survey inference, which is based on probability sample selection and weighting, with a model-based approach based on sampling quotas and model-based weighting. Compared with the traditional approach, the model-based approach more efficiently controls subgroup sample sizes when a large number of rare subgroups are studied. Using data from a national survey of US businesses, we simulated a model-based paradigm and compared estimates with those under the traditional paradigm. In this study, the findings suggest that the model-based approach offers advantages over the traditional sampling approach; however, a hybrid approach capturing the advantages of both paradigms proved best.