关键词: 分层结构数据, 多重插补法, Gibbs算法, 固定效应模型, 随机效应模型

Abstract: Complicated sampling design are usually used in sample surveys to get multilevel survey data with hierarchical nested structures. Missing data problem is often encountered in sample surveys, however, research on imputation strategies for the multilevel structures that are often found in complex survey data is limited. In this dissertation, it tries to use Gibbs algorithm to draw imputation values for multilevel missing data, and uses fixed effect imputation model and random effect imputation model in the process of multiple imputation.Through theoretical derivation and computer simulation, under different circumstance (includingmissingness rate,intraclass correlation and cluster size, etc), it compares the result of parameter estimation from the aspects of unbiasedness and effectiveness, and also gives the selection method of imputation model.The results show that when the random effect model is used as imputation model, the estimation results are more accurate, while the fixed effect model is easy to operate. When missingness rate is small, the intraclass correlation is large and the cluster size is large, the fixed effect imputation model can be adopted, otherwise the random effect imputation model is recommended.

Key words: Multilevel structure data, Multiple imputation method, Gibbs algorithm, Fixed effect model, Random effect model