Title: Donor imputation for multivariate missing data
Authors: Audrey-Anne Vallee - Universite Laval (Canada) [presenting]
Yves Tille - University of Neuchatel (Switzerland)
Esther Eustache - University of Neuchatel (Switzerland)
Abstract: Swiss cheese nonresponse, also known as non-monotone nonresponse, occurs when every variable of a survey contains missing values without a particular pattern. The estimators of the parameters of interest can be considerably affected by the missing values which introduce a bias and an increase in the variability. To reduce the effects of nonresponse, the missing values are usually imputed. When several variables of a dataset need to be imputed, it may be difficult to preserve the distributions and the relations between the variables. Balanced $K$-nearest neighbor imputation is extended to treat Swiss cheese nonresponse. The method uses random imputations by donors, and it is constructed to meet the following requirements. First, a nonrespondent should be imputed by neighboring donors. Next, all missing values of a nonrespondent should be imputed by the same donor. Last, the donors are selected in order to satisfy some balancing constraints allowing to decrease the variance of the estimators. To meet all the requirements, a matrix of imputation probabilities is constructed using calibration techniques. The donors are then selected with these imputation probabilities and balanced sampling methods.