Comparing risk of chronic wasting disease occurrence using Bayesian hierarchical spatial models and different surveillance types

•R-INLA software suitable for modeling diseases across large spatial extents.•Chronic wasting disease occurrence lower in roadkill compared to hunter harvest deer.•Modeling single surveillance type from disease investigations can bias model results.•Combining multiple surveillance types into models provides more robust estimates.•Spatial patterns may provide insight into disease ecology and epidemiology. Spatial modeling of wildlife diseases can be used to describe patterns of disease risk, understand biological mechanisms of disease occurrence, and for spatial prediction. Risk of wildlife disease occurrence in relation to environmental variables is often modeled and predicted using Markov chain Monte Carlo (MCMC) methods, which are unsuitable for large datasets and those covering large spatial extents. Integrated nested Laplace approximation (INLA) and INLA using the stochastic partial differential equation (INLA-SPDE) approach have become popular alternatives to MCMC for Bayesian inference because of their fast computational time and ability to process large datasets. Studies investigating risk of disease occurrence in wildlife, to our knowledge, have not yet compared Bayesian hierarchical spatial models over large spatial extents using real world data. Using chronic wasting disease (CWD) surveillance data from white-tailed deer (Odocoileus virginianus) collected in Pennsylvania, United States, as a case study, we first demonstrate how parameter estimates compare among MCMC, INLA, and INLA-SPDE modeling frameworks. We then model CWD (detected/non-detected) using INLA-SPDE over a much larger spatial extent than has been conducted previously for this disease to determine how surveillance type (e.g., hunter harvest, roadkill, or all surveillance) influences model parameters and predicted risk of CWD occurrence at locations not sampled. Fixed effects considered in the models included deer age and sex, elevation, slope, distance to streams, percent clay, and proportion of two habitat classes (forest and open) known to influence deer movements. We found INLA to produce comparable estimates to MCMC and permit modeling large datasets covering expansive spatial extents much faster and more efficiently than MCMC. We identified potential biases in surveillance types, indicating the value of including all surveillance in models rather than only a single type. Comparing modeling tools available for mapping diseases of wildlife in relation to ecological variables at l