Improved abundance prediction from presence-absence data

Many ecological surveys record only the presence or absence of species in the cells of a rectangular grid. Ecologists have investigated methods for using these data to predict the total abundance of a species from the number of grid cells in which the species is present. Our aim is to improve such predictions by taking account of the spatial pattern of occupied cells, in addition to the number of occupied cells. We extend existing prediction models to include a spatial clustering variable. The extended models can be viewed as combining two macroecological regularities, the abundance-occupancy regularity and a spatial clustering regularity. The models are estimated using data from five tropical forest censuses, including three Panamanian censuses (4, 6 and 50 ha), one Costa Rican census (16 ha) and one Puerto Rican census (16 ha). A serpentine grassland census (8 x 8 m) from northern California is also studied. Taking account of the spatial clustering of occupied cells improves abundance prediction from presence-absence data, reducing the mean square error of log-predictions by roughly 54% relative to a benchmark Poisson predictor and by roughly 34% relative to current prediction methods. The results have high statistical significance.