
Imputation is commonly used to compensate for item nonresponse. With a suitable imputation model and method, the bias due to nonresponse can be reduced relative to using only the observed data. In this dissertation, methods are developed for making inferences using a data set where imputation has been used to replace missing data;Variance estimation methods based on a pseudo data set used only for variance estimation are developed. Standard complete data variance estimators applied to the pseudo data set lead to consistent variance estimators for linear estimators under various imputation methods;Two modelling approaches, the cell mean model and the cell response-probability model, for hot deck imputation are discussed. Imputation methods are developed for each of the approaches and a fully efficient fractional imputation procedure is introduced. A replication variance estimation method for fully efficient fractional imputation is proposed and shown to be consistent for both of the approaches. The variance estimator is consistent for domain means and for functions of vector variables. Empirical comparisons with existing methods are made through simulation studies. The proposed variance estimator generally has smaller variances than methods suggested in the literature;A consistent variance estimation method for multi-phase sample estimators is proposed. The proposed method can be used for a reweighted expansion estimator as well as for a double expansion estimator. The proposed method is applied to the Accuracy and Coverage Evaluation survey in 2000 US Census of Population and Housing.
Available at: http://works.bepress.com/jae-kwang-kim/8/