|
Abstract:
|
Network data typically exhibits high order dependence (as captured by k-stars etc) which causes difficulty in directly applying traditional statistical methods on network data. To capture the dependencies, the Exponential Random Graph Model is often used in network data analysis. Instead of using the ERGM model, we propose a new model – Network Model with Unspecified Higher-Order Dependence. In this model the log-odds of edges in a network is modeled as a mixed effects model with additional latent variables. Using any software for mixed effects models we first estimate the mixed effects and then perform a PCA on the residual adjacency matrix to capture higher order dependencies. We then refit the mixed model with additional random effects from the PCA. This method provides a simple way to model network data within the mixed effect model framework, allowing network data to be analyzed using pre-existing software. Further this method captures network structure in an automated ad-hoc fashion. As such data analysts will not need to pre-specify network statistics while still capturing network dependence allowing for better estimates of coefficients of covariates on edge probabilities.
|
