Simulating (Multiple) Imputation in Relational Event History (REH) Data: Missingness in Time, Sender, and/or Receiver.

Abstract

Relational event history (REH) data is a specific type of dynamic network data containing the time-stamped interactions between nodes in a network. REH data is characterized by its high resolution compared to regular network data and is increasingly available due to technological advancements. Therefore, it can potentially be crucial in investigating complex social and behavioral phenomena. The state-of-the-art method to analyze REH data is through a relational event model (REM) and managing missing data within this context is crucial as it can significantly impact the validity of results. While (multiple) imputation methods are well-regarded for their reliability, they remain underexplored within the realm of dynamic social networks, particularly in REH data. This study aims to bridge this gap by focusing on REH data to improve the robustness of REM analyses involving missing data in social network research. By simulation and imputation of missing data in part of the Apollo 13 mission data, this study compares REM analyses of imputed data to their true and complete case analysis counterparts. Multiple imputation was employed for missingness in sender and receiver nodes, while time values were interpolated in several ways. Bias, coverage, and confidence interval width are evaluated in reciprocity, in-degree sender, out-degree receiver, and the same location statistics. Results revealed biases in the estimated statistics. Imputed analyses showed reduced absolute and relative bias, but incorrect statistical significance compared to complete case analysis. Multiple imputation improved effect size estimation compared to complete case analysis, suggesting its potential in REH data. This finding also highlights the need for refined methods specifically tailored to imputing time data, ensuring more accurate and reliable analyses in the study of dynamic social networks.