A local tree death model gave no clear decrease in log loss compared with past group rates at Barro Colorado Island, 2005–2015

We tested whether recent local tree deaths improve prediction of deaths in the next census interval in one 50 ha Panama forest plot. Rules fitted before 2005 used past species and size rates, broad location and older local death records. The added term used deaths in 20 m cells from the most recent completed interval. The later tests had 415,488 tree records and 47,546 deaths in 2005–2010 and 2010–2015. The added term reduced mean binary log loss by 0.119% against the adjusted reference. The nominal 95% spatial resampling range was 0.083% to 0.154%. Both fitted rules overestimated the mean later death fraction. A comparison added after the main results used simple past species and size death rates, without logistic adjustment or spatial terms. Against that rule, the local model gave a log-loss decrease of 0.012%, with a range of −0.158% to +0.178%. Its Brier loss was higher by 0.163%. No fitted rules or predictions were changed for this comparison. Thus, the small planned gain did not establish a clear decrease in loss from simple past rates. The test is retrospective; earlier descriptive results for these intervals were known. Spatial ranges hold predictions fixed and omit training and future-time uncertainty. The result does not identify a cause of death or apply to other forests. The package has source extracts, provenance, all code, fitted rules, probabilities, resampling draws, figures and reproduction checks.

tree mortality · forest dynamics · chronological validation · spatial prediction · Barro Colorado Island · reference models · open data