B0415
Title: A nonparametric copula-based method for the imputation of dependent data
Authors: F Marta L Di Lascio - Free University of Bozen-Bolzano (Italy)
Aurora Gatto - Free University of Bozen-Bolzano (Italy) [presenting]
Abstract: Imputation methods are useful in all situations where missing values occur and limiting the analysis to complete cases prevents proper inference. The copula function has been only partially explored in the context of imputation, where the algorithm called CoImp is the main proposal in the literature. Although CoImp allows the imputation of multivariate missing data of any pattern and the dependence structure underlying the data is preserved, it has some drawbacks, such as the computational burden and the limited types of copula models used. The potential of a fully nonparametric copula-based approach is explored for the imputation of data exhibiting complex multivariate dependence structures. The proposed method is based on the empirical copula, which is highly flexible and distribution-free. The main idea is to numerically reconstruct the conditional empirical copula of missing data given the observed data through the numerical version of the Monte Carlo inverse transform method. The performance of the proposal has been investigated on simulated data and compared with the CoImp algorithm. Finally, the method has been successfully applied to a real data set concerning plant protection products used in agriculture.
