Balanced imputation for swiss cheese nonresponse

Swiss cheese nonresponse or non-monotone nonresponse occurs when all the variables of a survey can contain missing values without a particular pattern. Imputation of missing values allows to reduce the bias and the variability due to nonresponse. It is difficult to preserve the distributions and the relations between the variables when imputing in the swiss cheese nonresponse case. In this presentation, balanced K-nearest neighbor imputation Hasler and Tillé (2016) is extended to treat swiss cheese nonresponse. It is a donor imputation method that is random and constructed to meet some requirements. First, a nonrespondent can be imputed by donors which are close to him. The distances are calculated with the observed values. Next, all the missing values of a nonrespondent are imputed by the same donor. Last, the donors are chosen so that if the observed values of the nonrespondents were imputed, the estimated totals would be the same as the one calculated with the observed values only. To meet all the requirements, a matrix of imputation probabilities is constructed with calibration techniques. The donors are selected with these imputation probabilities and balanced sampling methods. The advantages and the properties of the method are investigated in a simulation study.

, Workshop on statistical data editing de la Commission Économique des Nations Unies pour l'Europe, Neuchâtel