Hybrid multiple imputation in a large scale complex survey

Large-scale complex surveys typically contain a large number of variables measured on an even larger number of respondents. Missing data is a common problem in such surveys. Since usually most of the variables in a survey are categorical, multiple imputation requires robust methods for modelling high-dimensional categorical data distributions. This paper introduces the 3-stage Hybrid Multiple Imputation (HMI) approach, computationally efficient and easy to implement, to impute complex survey data sets that contain both continuous and categorical variables. The proposed HMI approach involves the application of sequential regression MI techniques to impute the continuous variables by using information from the categorical variables, already imputed by a non-parametric Bayesian MI approach. The proposed approach seems to be a good alternative to the existing approaches, frequently yielding lower root mean square errors, empirical standard errors and standard errors than the others. The HMI method has proven to be markedly superior to the existing MI methods in terms of computational efficiency. The authors illustrate repeated sampling properties of the hybrid approach using simulated data. The results are also illustrated by child data from the multiple indicator survey (MICS) in Punjab 2014.