Predicting abundance from presence-absence data

My research, joint with colleagues, studies simple, tractable, broadly applicable models of the spatial distributions of species, with applications to large plant datasets.  In this presentation, we study the problem of predicting the abundance of a species in an area from data on its presence or absence in the cells of a grid.  Such prediction is important because abundance is important to conservation and restoration, and because presence-absence data are common and easily obtained.  In the standard prediction approach, the number of occupied cells on a grid is the sole prediction variable.  Our hypothesis is that a second variable – a measure of the spatial clustering of the occupied cells – will improve prediction.  We explore several clustering metrics, pretest them using simulation data, and then test them with real data from six plant censuses.   An adjacent-cell autocorrelation works best.  It improves prediction with high statistical significance, reducing the mean square error of log-predictions on the order of 60%.