PROPENSITY WEIGHTING FOR SURVEY NONRESPONSE THROUGH MACHINE LEARNING

We consider the problem of estimating the response probabilities in the context of weighting for unit nonresponse. The response probabilities may be estimated using either parametric or nonparametric methods. In practice, nonparametric methods are usually preferred because, unlike parametric methods, they protect against the misspeci cation of the nonresponse model. In this work, we conduct an extensive simulation study to compare methods for estimating the response probabilities in a nite population setting. In our study, we attempted to cover a wide range of (parametric and nonparametric) "simple" methods as well as aggregation methods like Bagging, Random Forests, Boosting. For each method, we assessed the performance of the propensity score estimator and the Hajek estimator in terms of relative bias and relative e ciency.