Missing data is a phenomenon that cannot be ignored in network analysis, especially due to the complex nature of network data and the plethora of models in this field. This dissertation studies the effect of missing two-mode tie data on coefficient estimates of the influence mode in two-mode network analysis. A new imputation method based on the log odds of attending events within- vs. outside- cluster is proposed. The new imputation method is compared with the multiple imputation method under the missing at random mechanism. Network data are simulated based on different parameter values, including the network density, number of actors, number of events, and the odds ratio (i.e., clustering effect). Fifty-four unique network settings are examined, and 2000 replicates are generated for each unique setting. The multiple imputation method performs the best in terms of bias, empirical standard error, and root mean square error, partly because the missing data generation mechanism favors the multiple imputation method. The proposed imputation method performs well when there are medium to strong clustering effect in the network.
Read
