U.S. state-level COVID-19 transmission insights from a mechanistic mobility-incidence model

SO_SCPLOWUMMARYC_SCPLOWO_ST_ABSBackgroundC_ST_ABSThroughout the COVID-19 pandemic, human mobility has played a central role in shaping disease transmission. In this study, we develop a mechanistic model to calculate disease incidence from commercially-available US mobility data over the course of 2020. We use it to study, at the US state level, the lag between infection and case report. We examine the evolution of per-contact transmission probability, and its dependence on mean air temperature. Finally, we evaluate the potential of the model to produce short-term incidence forecasts from mobility data. MethodsWe develop a mechanistic model that relates COVID-19 incidence to time series contact index (CCI) data collected by mobility data vendor Cuebiq. From this, we perform maximum-likelihood estimates of the transmission probability per CCI event. Finally, we retrospectively conduct forecasts from multiple dates in 2020 forward. FindingsAcross US states, we find a median lag of 19 days between transmission and case report. We find that the median transmission probability from May onward was about 20% lower than it was during March and April. We find a moderate, statistically significant negative correlation between mean state temperature and transmission probability, r = - .57, N = 49, p = 2 x 10-5. We conclude that for short-range forecasting, CCI data would likely have performed best overall during the first few months of the pandemic. InterpretationOur results are consistent with associations between colder temperatures and stronger COVID-19 burden reported in previous studies, and suggest that changes in the per-contact transmission probability play an important role. Our model displays good potential as a short-range (2 to 3 week) forecasting tool during the early stages of a future pandemic, before non-pharmaceutical interventions (NPIs) that modify per-contact transmission probability, principally face masks, come into widespread use. Hence, future development should also incorporate time series data of NPI use.