Pooled land-use losses of local richness and abundance in PREDICTS 2016 replicated in 115 independent 2022 sources, but global coefficients predicted about 1% of within-study site variance

Question: do land-use effects on local terrestrial biodiversity from the PREDICTS 2016 release (480 sources, 666 studies) replicate in the 115 independent sources (151 studies) added in the November 2022 release? Approach: we computed site species richness and effort-corrected total abundance, then within-study log response ratios against primary vegetation. We did random-effects meta-analyses for each land use and release, with a plan written before we calculated the 2022 effects. We also measured how well the 2016 coefficients predict site differences within new studies. Results: across all non-primary land uses, richness was 16.7% lower in 2016 (k = 279; 95% CI -21.3 to -11.8%) and 13.8% lower in 2022 (k = 56; -20.4 to -6.6%; difference p = 0.49). Abundance was 9.2% lower in 2016 (k = 254) and 10.0% lower in 2022 (k = 50; p = 0.91). Twelve of 16 class-by-metric effects had the same sign. Two class differences were significant after Holm correction, both with few 2022 studies (k = 3 and 8). Between-study heterogeneity was large (tau 0.45 for richness and 0.60 for abundance, 2.4 to 6.3 times the mean effect). The 2016 prediction intervals included 98% of the 2022 study contrasts, and only 56% of new studies showed an abundance loss, as the 2016 model predicted. The 2016 coefficients explained 1.2% (95% CI -2.5 to 4.9%) of the within-study variance in log richness and -0.7% in log abundance in 2022 studies. They were better than a no-difference null in about half of the studies. Leave-one-study-out skill within 2016 was similar, while local class means of the same study explained 19 to 25%. Interactions with use intensity, taxonomic group or climate zone did not increase the skill. Random-intercept mixed models gave confidence intervals that were too narrow for new studies (8 of 16 coefficient differences had p < 0.05). Conclusion: the average effect is reproducible, but the effect in one study is not predictable from global coefficients. Limitations: few 2022 studies have a primary-vegetation reference; the releases differ in taxa and region; we used raw richness and the CC BY version 1.0 of the 2016 release.

PREDICTS · land use · local biodiversity · species richness · abundance · replication · out-of-sample prediction · meta-analysis · between-study heterogeneity · Biodiversity Intactness Index