PREDICTS 2016 land-use effects agree on average with 115 new 2022 sources, but predict near 0% of within-study abundance differences and 2% of richness differences in new studies

Question: do the land-use coefficients of the 2016 PREDICTS release, which global models such as the Biodiversity Intactness Index use, predict independent data? Approach: we fitted within-study (study fixed-effect) models of total abundance and species richness for 12 land-use and use-intensity classes to the 2016 release (444-495 studies). Then we predicted the within-study differences in the November 2022 addition (110-126 studies from 115 new sources, no shared source). We compared this transfer with leave-one-study-out prediction inside 2016, with 1,000 study-bootstrap samples. We also did a random-effects meta-analysis of study log response ratios. Results: the 2022 coefficients did not differ jointly from the 2016 coefficients (abundance p = 0.41, richness p = 0.15). They were about proportional (ratio 1.07 and 0.87). The 2016 prediction intervals contained 92% and 89% of the 2022 study effects. Transfer skill was equal to the held-out 2016 skill, but it was near zero. The 2016 coefficients explained 0.4% (95% CI -5.7 to 2.4%) of within-study abundance variance and 1.9% of richness deviance in the 2022 studies, against -0.5% and 2.4% in held-out 2016 studies. Study-specific class means explain 24-28%. The between-study SD of abundance log response ratios (median 0.69) was much larger than the mean effects (median 0.09). The means explained 2% (abundance) and 9% (richness) of the true between-study variance. Taxon-group or climatic-zone interactions decreased the held-out skill. Equal study weights gave the same results. Conclusion: the mean coefficients are reproducible across releases, but they do not predict the land-use response in one study. Limitations: fixed-effect models, not the original mixed models; the compositional-similarity component of BII is not tested; the CC BY 2016 version, not version 1.1; few 2022 studies in some classes, which gives wide CIs.

PREDICTS · land use · Biodiversity Intactness Index · out-of-sample validation · transferability · between-study heterogeneity · local abundance · species richness · meta-analysis · space-for-time