tergmLite-package {tergmLite}

Fast Simulation of Simple Temporal Exponential Random Graph Models

Description

Package:	tergmLite
Type:	Package
Version:	2.6.1
Date:	2022-07-20
License:	GPL-3
LazyLoad:	yes

Details

The statistical framework of temporal exponential random graph models (TERGMs) provides a rigorous, flexible approach to estimating generative models for dynamic networks and simulating from them for the purposes of modeling infectious disease transmission dynamics. TERGMs are used within the EpiModel software package to do just that. While estimation of these models is relatively fast, the resimulation of them using the tools of the tergm package is computationally burdensome, requiring hours to days to iteratively resimulate networks with co-evolving demographic and epidemiological dynamics. The primary reason for the computational burden is the use of the network class of object (designed within the package of the same name); these objects have tremendous flexibility in the types of networks they represent but at the expense of object size. Continually reading and writing larger-than-necessary data objects has the effect of slowing the iterative dynamic simulations.

The tergmLite package reduces that computational burden by representing networks less flexibly, but much more efficiently. For epidemic models, the only types of networks that we typically estimate and simulate from are undirected, binary edge networks with no missing data (as it is simulated). Furthermore, the network history (edges or node attributes) does not need to be stored for research-level applications in which summary epidemiological statistics (e.g., disease prevalence, incidence, and variations on those) at the population-level are the standard output metrics for epidemic models. Therefore, the network may be stored as a cross-sectional edgelist, which is a two-column matrix of current edges between one node (in column one) and another node (in column two). Attributes of the edges that are called within ERGMs may be stored separately in vector format, as they are in EpiModel. With this approach, the simulation time is sped up by a factor of 25-50 fold, depending on the specific research application.