Counts of nearby trees of the same species did not improve death predictions for small trees of 30 BCI species in 2000–2015

We tested whether counts of nearby trees of the same species improved later predictions of tree death in the Barro Colorado Island 50 ha plot, Panama. The selected data had 713,266 tree-interval records from 202,422 trees of 30 species that were common in 1985. Starting diameters were 1 to less than 5 cm. We fitted two models to 1985–2000 and froze their coefficients. Both models used species, size, counts of other species, broad location and measured follow-up length. The added model also used counts of the same species within 10 m. In 324,863 test records from 2000–2015, with 33,324 deaths, added counts raised mean log loss from 0.320583 to 0.320813, an increase of 0.0717%. The nominal 95% range from resampling spatial strips was 0.034% to 0.114%. This range was conditional on the fitted models and observed later intervals. Training log loss fell by 0.102%. Later error also rose at radii of 5 and 20 m, under two other penalty settings, and after each species was omitted from scoring. The Brier score gave the same direction. The result is a small failure of these fixed prediction rules to transfer over time. It does not reject local tree interactions or all neighborhood models. We preserve source column extracts, hashes, derived data, coefficients, predictions, figures, checks and offline reproduction code.

biodiversity · tree mortality · Barro Colorado Island · neighborhood counts · temporal prediction · density dependence · reproducible research