There has been elevating research interest in social network analysis along with prosperous methodology and application studies in network science and other social sciences. Most of the studies made a great effort to explain how human behavior, as the nodal effect, and the network structure, as an environmental system, impact each other. Although the causal inference implied in these researches is naturally dynamic, most social network methods are cross-sectional. Yet, we haven't seen any attempts to incorporate longitudinal mediation analysis with social network analysis. In this dissertation, I propose a new approach to integrating the longitudinal mediation method and social network data. It can provide a better interpretation of the intervened influence between behavior and network. In this proposed method, dynamic networks are modeled as mediators and behavioral variables as either the dependent or the independent variable. The proposed longitudinal network mediation model is specified, estimated, and assessed by three simulation studies and an empirical data analysis. The proposed method achieved good estimation accuracy and computational efficiency. The constructed confidence intervals for statistical inference also showed a reasonably good coverage rate of true parameters, Type I error rates, and power under conditions of small, medium, large, and zero mediation effect. The empirical examples provided interpretation of social selection and social influence when given a plausible causality basis. A significant mediation effect of the network was found in an empirical network data set, Glasgow data. The results showed that the more similar two individual teenagers were in terms of sporting activity at Time 1, their social network would aid the assimilation of their other behaviors, which is more similar in substance use at Time 3.