|
Abstract:
|
In an effort to improve U.S. Bureau of Labor Statistics (BLS) Occupational Employment Statistics (OES) survey response and quality, we developed tools and provided metrics to help managers make informed decisions such as workload distribution and to aid data collectors. Using OES collection data linked to establishment administrative files, we have developed models to estimate the likelihood of an establishment response. We expand on this work and provide additional metrics by predicting most likely modes of response/collection. Collection modes include survey form, electronic, phone call, email, fax, and other modes. These results will help managers plan soliciting methods and could help reduce collection costs. We will use various machine learning techniques applied to a multiclass prediction problem to optimize prediction accuracy. These methods include multinomial logistic regressions, lasso (and elastic net), classification trees, random forests, and boosting (AdaBoost) methods. Our predictors include various establishment characteristics such as size, age, industry, and wages, as well as paradata such as previous mode of collection and previous response.
|
