Imputing Missing values in Relational Event History data: A Framework for Social Network Research

Abstract

Background – Missing values are practically inevitable when it comes to data analysis and can cause loss of valuable information and introduce bias in the estimates. In social network data even a small proportion of missing values can have a substantial impact on the validity of results. The most common approach to deal with missing values is complete case analysis which does not always prevent these issues. Still, the research investigating this problem is scarce. This paper aims to address this gap by providing an overview of multiple imputation as a method of handling missing values in social network data. Methods – Relational event model analysis is performed on the fully observed relational event history dataset to produce the true estimates. Next, a simulation study is performed to introduce missingness to a fully observed dataset. Then the results of two approaches - multiple imputation and complete case analysis are compared to the results of the analysis on the fully observed dataset. Results –Multiple imputation with relational event model produced estimates with close to zero bias, high coverage rate, and low average width. However, multiple imputation produced false significant p-values. In addition, the distributions of the effect estimates were slightly skewed for all effects. Complete case analysis produced overestimated effects and standard errors but did not produce false significant p-values. Conclusion – This study made first steps in evaluating whether multiple imputation with relational event model is a valid method to address missing values in relational event history data and compared it to the more common solution – complete case analysis. The findings reveal potential benefits of multiple imputation and propose direction for future research.