Multiple Imputation for Incomplete Data in Environmental Epidemiology Research

Incomplete data are a common problem in statistical analysis of environmental epidemiological research. However, many researchers still ignore this complication. We evaluate the performance of two commonly used multiple imputation (MI) methods (fully conditional specification and multivariate normal) for handling missing data and compare them to complete case analysis (CCA) method. We further discuss issues that arise when these methods are being used.

MI is a simulation-based approach to deal with incomplete data. In general, MI will perform better then ad hoc techniques such as CCA. MI is an approach which replaces the missing data with plausible values and allows for additional uncertainty due to the missing information caused by the incomplete data. To illustrate this, we use data of 944 women from the Collaborative Perinatal Project and compare estimates between these methods. The goal is to examine if each of two outcomes, birth-weight and spontaneous abortion, in the data set are associated with mothers' smoking status during pregnancy adjusting for baseline covariates in the model. Results indicate that MI is better suited for handling incomplete data and led to a significant improvement in parameter estimates compared to CCA. The two MI methods produced similar point estimates, but slightly different standard errors.