A Hybrid Technique for the Multiple Imputation of Survey Data

Most of the background variables in MICS (Multiple Indicator Cluster Surveys) are categorical with many categories. Like many other survey data, the MICS 2014 women’s data suffers from a large number of missing values. Additionally, complex dependencies may be existent among a large number of categorical variables in such surveys. The most commonly used parametric multiple imputation (MI) approaches based on log linear models or chained Equations (MICE) become problematic in these situations and often the implemented algorithms fail. On the other hand, nonparametric MI techniques based on Bayesian latent class models worked very well if only categorical variables are considered. This article describes how chained equations MI for continuous variables can be made dependent on categorical variables which have been imputed beforehand by using latent class models. Root mean square errors (RMSEs) and coverage rates of 95% confidence intervals (CI) for generalized linear models (GLM’s) with binary response are estimated in a simulation study and a comparison is made among proposed and various existing MI methods. The proposed method outperforms the MICE algorithms in most of the cases with less computational time. The results obtained by the simulation study are supported by a real data example.