Studies on missing network data have largely focused on the impact of missing data on network structure rather than inference from a statistical model. In particular, there has very little research on the impact of missing data when fitting latent variable network models, so we examined the impact of common missing data mechanisms and subsequent missingness treatment and imputation methods when working with latent variable network models, focusing on the latent space model (Hoff et al., 2002). By removing the common definitions of missingness, our simulation study found large differences in inference, parameter, and network feature recovery based on the missingness mechanism and treatment method. In addition, we induced missingness using a real-world dataset and explored how treatment methods impacted subsequent inference and network recovery. We found that missingness based on a node covariate that also predicted network ties was the most problematic form of missingness and that complete case analysis and Bayesian estimation generally worked as well or better than other methods. • Missing data in latent space models has been studied in depth. • We explored the effects of missing data mechanisms and the treatment method. • Bayesian estimation had the lowest regression coefficient bias. • Regression imputation and multiple imputation produced worse results than expected. • We recommend multiple treatment methods when estimating models with missingness.
Sweet et al. (Mon,) studied this question.