A0307
Title: Fast heavy-tail count models for automated probabilistic computing and pandemic preparedness
Authors: Oliver Ratmann - Imperial College (United Kingdom) [presenting]
Abstract: The aim is to present new heavy-tail models for discrete count data, including the Zipf distribution and the beta-negative-binomial distribution (BNB) that we contributed for fast inference to Stans C++ Math library, also enabling GPU library support. The Zipf and BNB distributions are, respectively, one and three-parameter distributions to flexibly model scale-free and/or heavy-tail behaviour of a large class of outcomes in generalised linear regression frameworks. These models are implemented and applied to infer how the structure of human contact networks evolved during the COVID-19 pandemic using one of the largest ever, individual-level longitudinal contact survey studies comprising n 50,000 participants in Germany from early 2020 to the end of 2021. The C++ implementations in Stans Math library allow tracking individual-level heterogeneity in contact behavior by age, gender, sub-national geographical areas and attitudes over four pandemic phases. It characterizes which sub-populations high-degree hubs re-emerged fastest and how their relative contribution to all contacts changed over time. At the network level, it is shown how time trends in compounding individual-level behavior are associated with phase transitions in how effectively pandemic spread can be disrupted by estimating population-level scale-free network properties and predicting the existence of giant network components.
