Purpose: Multiple imputation (MI) is a widely acceptable approach to missing data problems in epidemiological studies. Composite variables are often used to summarize information from multiple, correlated items. This study aims to assess and compare different MI methods for handling missing categorical composite variables.
Methods: We investigate the problem in the context of a real application: estimating the prevalence of HIV transmission category, which is a composite variable generated by applying a hierarchical algorithm to a group of binary risk source variables from a national program data set. We use simulation studies to compare and assess the performance of alternative MI strategies. These methods include the active imputation, just another variable, and the passive imputation approaches.
Results: Our study suggests that the passive imputation approach performs better than the direct imputation approach and the inclusive and general imputation model (i.e. passive imputation with interactions) performs the best. There is no need to embed the information from the variable-combining algorithm in the passive imputation modeling.
Conclusion: We recommend practitioners adopting an inclusive and general passive imputation modeling strategy.
Keywords: Composite variable; Inclusive modeling strategy; Multivariate imputation by chained equations; Passive imputation.
Published by Elsevier Inc.