Online Program

Multiple Imputation for Missing Data in Longitudinal Research Synthesis: Identifying and Overcoming Assumptions in Software (303017)

Keywords: multiple imputation, longitudinal data, hierarchical data, research synthesis

In this era of increasingly available data sets, analyses that combine individual data from collections of data sets are becoming an important tool for data analysts. A commonly encountered challenge is adjusting for variables not included in all the data sets. We examine this problem as it arose in a research synthesis of longitudinal studies on the effect of pediatric traumatic brain injuries. The approach we took was to use multiple imputation to account for subject-level (non-time varying) variables not measured in all the studies.We found that implicit assumptions made by commonly used software packages were not appropriate for non-exchangeable correlation structures, which was a tenuous assumption for our data and many other scenarios. For example, there are discrepancies in how to best incorporate observation-level (time varying) variables into imputation models for missing subject-level variables. In addition, we discovered that the choice of prior distributions affected results more than we expected. We address the shortcomings of the software and how to use alternative standard procedures to overcome these issues in longitudinal analyses.