Increasing spatial approximation complexity can degrade prediction quality in distribution models

Spatial and spatiotemporal models are increasingly critical for understanding species distributions, tracking population change, and informing conservation decisions. As biological processes are influenced by increasing external pressures, including human disturbance or environmental change, accurate model predictions become essential for adaptive management. However, the reliability of spatial predictions depends on often-overlooked modelling choices, including the spatial resolution used to approximate underlying processes. Using long term monitoring data from a large-scale groundfish survey in the California Current ecosystem, we investigated how spatial model complexity affects the quality of ecological predictions and derived indices used for management. We fit spatial and spatiotemporal models of ocean temperature and fish biomass density for 27 commercially important species using varying levels of spatial resolution. We evaluated both in-sample and out-of-sample prediction, and effects on area-weighted biomass indices. Counter to common assumptions, increasing spatial approximation resolution did not universally improve predictions. Our case studies demonstrate that for many datasets, out-of-sample prediction quality peaked at intermediate spatial resolutions and declined at the finest scales. Through simulation testing, we found this pattern was strongest when spatial patterning had a small range and high spatial variance, and observation error was low. For most species, spatial resolution had a minimal effect on biomass trend estimates used in management, but for several commercially important rockfish species, resolution choices substantially affected both the scale and uncertainty of population indices. Our findings demonstrate that spatial model specification can substantially affect ecological inference, with direct implications for management and conservation planning. We provide practical guidance for ecologists on selecting appropriate spatial complexity through cross-validation. When out-of-sample prediction is a focus, appropriate approximation complexity should improve both parameter estimation accuracy and derived quantities.