A Tale of Twin Dependence: A New Multivariate Regression Model and an FGLS Estimator for Analyzing Outcomes With Network Dependence

In this article, I present a new multivariate regression model for analyzing outcomes with network dependence. The model is capable to account for two types of outcome dependence including the mean dependence that allows the outcome to depend on selected features of a known dependence network and the error dependence that allows the outcome to be additionally correlated based on patterned connections in the dependence network (e.g., according to whether the ties are asymmetric, mutual, or triadic). For example, when predicting a group of students’ smoking status, the outcome can depend on the students’ positions in their friendship network and also be correlated among friends. I show that analyses ignoring the mean dependence can lead to severe bias in the estimated coefficients while analyses ignoring the error dependence can lead to inefficient inferences and failures in recognizing unmeasured social processes. I compare the new model with related models such as multilevel models, spatial regression models, and exponential random graph models and show their connections and differences. I propose a two-step, feasible generalized least squares estimator to estimate the model that is computationally fast and robust. Simulations show the validity of the new model (and the estimator) while four empirical examples demonstrate its versatility. Associated R package “fglsnet” is available for public use.

Network dependence in multi-indexed data on international trade flows

Faced with the problem that conventional multidimensional fixed effects models only focus on unobserved heterogeneity, but ignore any potential cross-sectional dependence due to network interactions, we introduce a model of trade flows between countries over time that allows for network dependence in flows, based on sociocultural connectivity structures. We show that conventional multidimensional fixed effects model specifications exhibit cross-sectional dependence between countries that should be modeled to avoid simultaneity bias. Given that the source of network interaction is unknown, we propose a panel gravity model that examines multiple network interaction structures, using Bayesian model probabilities to determine those most consistent with the sample data. This is accomplished with the use of computationally efficient Markov Chain Monte Carlo estimation methods that produce a Monte Carlo integration estimate of the log-marginal likelihood that can be used for model comparison. Application of the model to a panel of trade flows points to network spillover effects, suggesting the presence of network dependence and biased estimates from conventional trade flow specifications. The most important sources of network dependence were found to be membership in trade organizations, historical colonial ties, common currency, and spatial proximity of countries.

Network dependence testing via diffusion maps and distance-based correlations

Summary Deciphering the associations between network connectivity and nodal attributes is one of the core problems in network science. The dependency structure and high dimensionality of networks pose unique challenges to traditional dependency tests in terms of theoretical guarantees and empirical performance. We propose an approach to test network dependence via diffusion maps and distance-based correlations. We prove that the new method yields a consistent test statistic under mild distributional assumptions on the graph structure, and demonstrate that it is able to efficiently identify the most informative graph embedding with respect to the diffusion time. The methodology is illustrated on both simulated and real data.