Statistical Inference on Tree Swallow Migrations, Using Random Forests
share
Species migratory patterns have typically been studied through individual observations and historical records. In this work, we adopt a data driven approach to modelling the presence of the North American Tree Swallow, Tachycineta bicolor, throughout the Eastern United States, using data collected through the eBird project at Cornell University's Lab of Ornithology. Preliminary models suggest a qualitatively different pattern in Tree Swallow occurrence between the years of 2008 to 2009 and 2010 to 2013. We implement a global hypothesis test based on the functional predictions of Random Forests (RFs) to evaluate whether this effect is significant or not. In order to better understand the effect of climate change, we also conduct a test evaluating the effect of daily maximum temperature anomaly in predicting tree swallow occurrence. We implement a local test using the asymptotic normality of the predictions of a modified RF, which relies on subsampled trees. This test is conducted at 6 locations in space throughout the northeastern U.S. Finally, we present visual evidence that maximum temperature is affecting the predictions of RF models via a heat map of the differences in RF predictions. We also demonstrate that there is a spatial pattern in the effect using Moran's I statistic.
READ FULL TEXT