Sanders, ElizabethAbe, Nathan2021-10-292021-10-292021Abe_washington_0250E_23485.pdfhttp://hdl.handle.net/1773/48008Thesis (Ph.D.)--University of Washington, 2021This dissertation represents a series of studies focused on comparing imputation approaches for single-mode networks, also known as graphs, that are missing tie information due to a variety of potential causes unrelated to network properties, such as illness or technology failure. Additionally, in social network measurement designs that ask subjects to nominate other people in the network based on free recall (rather than forced choice), missingness can arise when people outside the surveyed network are sometimes nominated. The aim of this dissertation is to understand best approaches for handling this type of missingness in terms of coefficient estimation accuracy and precision. Specifically, Study 1 compared imputation approaches for graphs with binary valued ties measured at a single time point; Study 2 investigated approaches for handling missingness in graphs with binary valued ties measured at two time points; and Study 3 focused on approaches for handling missingness in graphs with integer valued ties (e.g., counts and ratings) measured at a single time point. Monte Carlo simulations were conducted in R using statnet, with varied network sizes, densities/mean values, and missingness levels. With a focus on use of the exponential random graph family of models (ERGMs), approaches to handling missing tie data included statnet default approaches (e.g., node-wise deletion or assuming all missingness represents no tie) as well as approaches in better alignment with modern missingness handling (e.g., stochastic imputation based on mean sample values). Findings consistently showed that stochastic imputation was best for minimizing bias and maximizing precision. Last, I describe a new R package, netImp, developed to implement imputation approaches so that researchers can obtain complete data matrices prior to analysis. Limitations, policy implications, and future research directions are discussed.application/pdfen-USCC BYERGMExpponential Random Graph ModelsMissing DatanetImpSNASocial NetworksEducational evaluationStatisticsEducation - SeattleThesis