Characterizing the impact of incorporating spatially aggregated human mobility data into infectious disease models

Models of infectious disease dynamics should align the spatial scale of mobility data to the scale of travel relevant to infer disease introduction events and subsequent local transmission. Despite this, the biases of spatially aggregating mobility data, which are more commonly available, on model inferences are rarely explored. Here, we examine the sensitivity of infectious disease modeling results to different spatial scales of human mobility by integrating multiscale mobility data from Sri Lanka into SEIR metapopulation models. Aggregated mobility data were obtained from mobile phone records at three increasingly coarser spatial scales to simulate epidemic spread. We found that travel was not evenly distributed amongst nested spatial units when data were disaggregated, with rural subunits exhibiting more external travel than urban subunits. In simulations of non-specific disease transmission, these different scales of mobility aggregation yielded wide variations in estimates of epidemic size and spatial invasion timing. However, modeled differences depended on disease characteristics such as transmissibility and exogenous factors like seeding location urbanicity. Our results carry implications for infectious disease modeling best practices and public health response, particularly the policy decisions made from model inferences that were or were not informed by the relevant spatial scale of mobility data such as intervention timing, risk communication, and resource allocation. Significance StatementInfectious disease models serve an important role in disease forecasting and outbreak response. As data on human mobility have become increasingly detailed and widely used, the question of spatial scale is paramount to effectively approximating disease dynamics. Our work finds significant discrepancies in estimations of spatial invasion timing and epidemic magnitude simply by changing the scale of mobility data integrated into a transmission model. This could have consequences for estimation of key parameters obtained from such models, the interpretation and communication of infection risk, and ultimately the public health response to a disease threat. We also highlight situations where large model discrepancies are unlikely, such as for highly infectious pathogens and where disease is initially seeded into an urban location.