Handling Missing Values when Using Neighborhood Selection for Network Analysis

The handling of missing values in cross-sectional network analysis has been primarily studied in conjunction with regularization techniques. However, non-regularized alternatives, such as neighborhood selection via BIC based on node-wise multiple regression, have been shown to be viable alternatives for psychological networks. Moreover, its localized approach in model selection renders neighborhood selection particularly suitable for data sets containing missing values. In this study, we present two approaches based on multiple imputation, namely stacked and grouped MI, alongside direct and two-step EM procedures, to handle missing values. Furthermore, various approaches to calculating sample size, used for computing log-likelihood and BIC, are discussed and evaluated. A simulation study encompassing variations in network size, number of observations, and levels of missingness is conducted to investigate the performance of the missing data handling approaches and ways to determine sample size. Evaluation criteria include the recovery of the edge set, as well as bias in partial correlations and estimated node strength. The findings indicate that the two-step EM approach is the quickest method, offering adequate handling of missingness when the number of observations is very large. With a smaller ratio of the number of observations to the number of nodes and highmissingness, stacked MI emerges as the preferable option. Regarding sample size calculation, particularly in cases of high missingness, using the local number of observations per node yielded the least bias in estimated strength.