B0516
Title: Variance estimation for survey estimators based on statistical learning procedures
Authors: Mehdi Dagdoug - McGill University (Canada) [presenting]
David Haziza - University of Ottawa (Canada)
Abstract: Predictive models are widely used in survey sampling. Common examples include the model-based framework, the model-assisted framework, as well as the treatment of nonresponse with both imputation and reweighting. The last two decades have witnessed an increasing attention of applied and theoretical statisticians towards statistical learning; a field dedicated to the study and development of predictive models. More recently, survey statisticians started studying the use of statistical learning procedures in a survey framework. Statistical learning brings a new set of highly flexible tools for survey researchers, as well as new challenges; variance estimation is one of them. It is shown that traditional variance estimators often do not perform well when applied to survey estimators built from complex statistical learning procedures. The reason for these ill behaviours is investigated and explained. Alternative variance estimators will be suggested, and their performances will be discussed through theoretical and empirical results.
