Abstract
Research on the interplay between domestic animal and wildlife populations has gained momentum recently. African Swine Fever (ASF) is a highly infectious viral disease that is of particular interest due to ongoing outbreaks in Europe, where wild boar constitute a reservoir of disease spillover into domestic pigs.
It is theoretically desirable to model spread of the ASF virus through a wild boar population using a continuous-space model, where wild boar distribution is inferred from spatial mapping of the available habitat. Because continuous-space models are difficult to work with, these are typically approximated using discrete-space models, where the wild boar population is divided into patches within the landscape.
In this study, we investigate the epidemiological impact of different assumptions regarding the spatial discretisation of a wild boar population. We use a simulation study with a fixed population of wild boar within a fictional, homogeneous landscape. This landscape is discretised into patches using different methods. A deterministic compartmental Susceptible-Infected-Recovered (SIR) model is formulated to allow pairwise spread between each patch depending on a transmission matrix β, which is then fit using an ODE solver. The scenario is as follows: Two buffer zones are created with a large gap in between. An outbreak is started in one, and the time until the other buffer zone has crossed a certain prevalence threshold (τ) is recorded. This process is repeated on the same landscape, but with different discretisations of the landscape at different spatial resolutions. For each simulation, transmission between patches is controlled by a pairwise transmission matrix β.
We considered many correction factors for β that were hypothesised to re-scale overall transmission so that τ is independent of landscape discretisation choice. These included: scaling with respect to the diameter of the patch, exponential distance decay, normalising discrete distances, and normalising using an integrated distance kernel. These corrections mask the inconsistency for some scenarios, but overall τ remains highly dependent on the choice of discretisation, suggesting that the underlying epidemiological process is impacted by the spatial discretisation. We also present an alternative method of formulating β, using pairwise proximity of artificially-created high-resolution lattice points that approximates transmission in continuous space. This yields a τ that is more robust to the details of landscape discretisation, so that the overall model results are less dependent on arbitrary decisions regarding size and shape of patches.
In conclusion, we propose that the method of landscape discretisation should be regarded as a source of uncertainty that impacts disease spread, and as such should be treated as a hyper-parameter in epidemiological models involving wildlife. It is important to ensure that overall conclusions from such models are robust to landscape discretisation choice.