Skip to content
Field Notes

The research side of permaculture.studio.

Climate & Earth5 min read

Global biodiversity models get the average right and the place wrong

A systematic test finds global biodiversity models track average land-use effects but predict poorly at local scales, and worse when extended to new places.

What the research shows

  • A systematic evaluation found that global biodiversity models using study attributes as random effects had low predictive accuracy. Meta-analysis
  • Replacing study random effects with biome, realm and taxonomic parameters improved predictions, but results remained modest in the contexts sampled. Meta-analysis
  • Predictive performance stayed low when models were extended to other contexts, which the authors relate to shifts in environmental conditions and biodiversity responses. Meta-analysis
  • The authors attribute the constraint to current data availability, including geographic and taxonomic biases. Meta-analysis
  • Land-use effects estimated by these models were in line with previous research, so the failure is in prediction rather than in the direction of the estimated impact. Meta-analysispartially supported

Each of these was checked against the abstracts of the papers it cites. See the receipts.

Global biodiversity models get the average right and the place wrong

A systematic test of global biodiversity models finds what its authors call a challenging gap between effect-size inference and local prediction [3]. The models can say that land use harms biodiversity on average, in line with earlier research, but they are poor at saying what is happening at a particular place. Mixed models that treat study attributes as random effects — the workhorse of biodiversity meta-analysis and of many global indicators — showed low predictive accuracy, and the paper describes the cause as structural rather than incidental [3].

That distinction matters: this is not a finding that land-use impacts on biodiversity are unknown or that the underlying ecology is wrong. The study reports land-use effects consistent with earlier work [3]. The failure is where a global model is asked to say something specific about a specific location.

What was actually tested

The authors evaluated alpha-diversity and beta-diversity models against representative global data [3]. Alpha diversity is the variety within a site; beta diversity is the change in composition between sites. Both are quantities that conservation planning tends to want at a resolution finer than the globe.

The abstract does not report numeric accuracy metrics, sample sizes, or which datasets were used beyond the description "representative global data" [3]. So the magnitude of the failure cannot be judged from the abstract alone. What can be judged is the pattern: low accuracy for the standard modelling approach, modest improvement for a better-specified one, and low performance again when either is pushed outside the conditions it was fitted on.

Why averaging breaks prediction

The mechanism the authors identify is worth following, because it explains why a model can be simultaneously well-calibrated on average and useless locally.

Mixed models in this literature typically include study attributes as random effects. A random effect for each study absorbs the between-study variation — differences in site, method, taxon, region, sampling design — into a per-study offset. That is a reasonable way to estimate a pooled effect size. It is a poor way to predict, because the prediction then leans on the fixed effects, which are highly averaged across all those studies [3]. The model has, in effect, learned the mean of a very heterogeneous set of places and called it a coefficient.

The authors tested a fix. They replaced the study random effects with explicit biome, realm and taxonomic parameters, so that the model is forced to describe how responses differ across ecological and geographic settings rather than absorbing that variation into a nuisance term. This improved performance, but the paper is careful about the size of the gain: results were "improved but still modest" in the contexts that were sampled [3].

Then they pushed further, extending predictions to contexts not represented in the fitting data. Performance there remained low. The authors connect this to distribution shifts — both in environmental conditions and in how biodiversity responds to them [3]. This is the familiar problem of a model asked to extrapolate: the further the target sits from the training distribution, the less the fitted relationship means.

The tension the paper names

The framing the authors land on is a tension between two things that are both improving and are not improving together. High-resolution predictors — land-use layers, climate surfaces, remotely sensed covariates — are increasingly available at fine grain. The granularity at which biodiversity responses can be reliably estimated is not keeping pace [3]. You can describe a place in detail without being able to say what its biodiversity will do.

That gap has a geographic shape. The paper notes that countries with extensive data can build higher-fidelity national indicators, while data-poor regions need accelerated and systematic data collection to support localized, actionable insight [3]. The places where a global model is most likely to be the only tool available are also the places where it is least likely to be validated.

What the evidence does and does not support

Supported: that a common class of global biodiversity model has low predictive accuracy in this evaluation, that the reliance on highly averaged fixed effects is the stated driver, that an ecologically structured alternative does better but not dramatically so, and that transfer to new contexts is weak [3].

Not supported by this abstract: any specific accuracy figure, any claim about which regions or taxa performed worst, or any statement that a particular named indicator or policy tool is invalid. The paper's own conclusion is not that models should be abandoned — it states that models are essential for informed conservation efforts while their applicability is constrained by current data availability [3].

Also worth flagging: this is a single study, published in 2026, and the abstract describes its conclusions at a general level. It is a systematic test, which is more than the field previously had, but one paper is one paper.

Where the evidence runs out

The abstract gives no numbers. No R², no error metrics, no breakdown by region or taxon, no list of the datasets behind "representative global data" [3]. A reader cannot use it to estimate how wrong a given map is, only that the class of model tested underperforms at local prediction.

It is also unclear from the abstract how much of the improvement from ecologically structured models would survive in a genuinely new region, since the paper reports those gains as modest in sampled contexts and low when extrapolating [3]. The boundary between "modest" and "usable" is not drawn.

And the central constraint the authors identify — data availability with geographic and taxonomic biases — is stated as a diagnosis, not tested as an intervention. The claim that accelerated systematic data collection would raise local predictive accuracy is a reasonable inference from the analysis, but the abstract does not report a test of it [3].

The honest summary: the models are not broken at what they were built to do, which is estimate average effects. They are being asked to do something else, at a scale the data do not yet support. The paper is a clear statement of that limit, and a clear statement of what would have to change. It is not a measurement of how far off any particular map is.

The study is published in Nature Ecology and Evolution: 10.1038/s41559-026-03205-z [3].

References

  1. [1]Lynx reintroduction from a bottlenecked source exposes a lethal genetic defectJanna Niehaus, Michelle Imlau, Saskia Keller, Iris Marti, Christine Breitenmoser, Anna Letko · 2026 · Nature Ecology & Evolutiondoi:10.1038/s41559-026-03204-0
  2. [2]Species-rich and genomically diverse: Comparative genomics reveals how fusions, fissions, and sex chromosomes have shaped beetle evolutionDwayne G. Tally, Charlie Pittman, Ryan R. Bracewell · 2026 · PLOS Geneticsdoi:10.1371/journal.pgen.1012371
  3. [3]Global biodiversity models exhibit low predictive accuracy on local scalesJakob Nyström, Jeffrey R. Smith, Lisa Mandle, Andrew Gonzalez, Thomas B. Schön, Tobias Andermann · 2026 · Nature Ecology & Evolutiondoi:10.1038/s41559-026-03205-z
  4. [4]Unravelling groundwater's role in soil-plant-atmosphere continuum: integrated ecohydrological modelling approach using STEMMUS-SCOPE and MODFLOW 6Mostafa Gomaa Daoud, Fakhereh Alidoost, Yijian Zeng, Bart Schilperoort, Christiaan Van der Tol, Maciek W. Lubczynski · 2026 · Hydrology and Earth System Sciencesdoi:10.5194/hess-30-7031-2026
  5. [5]Multifunctional Bioactive Compounds from Medicinal Plants: Antioxidant, Antimicrobial and Corrosion Inhibition PerspectivesM. Jisha, Vinayan P. Nair · 2026 · Journal of Bio- and Tribo-Corrosiondoi:10.1007/s40735-026-01282-y
  6. [6]Examining urban expansion and spatial morphological evolution in Shillong city using geospatial techniquesDebojyoti Dhali, Sanjit Paul, Sarfaraz Alam, Jason Jyrwa · 2026 · Discover Geosciencedoi:10.1007/s44288-026-00803-z
  7. [7]Towards sustainable weed management: the role of artificial intelligence, sensing, and roboticsAshish Chaudhary, Swas Kaushal, Het Samir Desai, Navjot Singh, Muskan Jakhar, Devanshi Het Desai · 2026 · Frontiers in Agronomydoi:10.3389/fagro.2026.1898506
  8. [8]Neighborhood vertical-profile similarity and aboveground carbon variation in subalpine dwarf bambooHsiao-Lung Pan, Cho-ying Huang · 2026 · Ecological Indicatorsdoi:10.1016/j.ecolind.2026.115823
  9. [9]A comprehensive benchmark and dual-encoder baseline for terraced field extraction in complex mountainous regionsPei Tian, Yujie Sun, Yan Chen, Yuanzhe Cai, Ruoxuan Deng, Zheng Zhang · 2026 · Frontiers in Ecology and Evolutiondoi:10.3389/fevo.2026.2011654
  10. [10]Spatial matching between urban green-space services and older adults’ potential health-support needs: a community-scale assessmentWenjia Bai, Xinkai Liao, Mingyu Chen, Xinyi Li, Xin Zheng, Aomei Guan · 2026 · Frontiers in Public Healthdoi:10.3389/fpubh.2026.1991406
  11. [11]Soil organic matter quality in agrivoltaic systems: contrasting responses of SOM pools and microbial carbon utilizationMartin Brtnicky, Jiri Holatko, Petra Prochazkova, Jan Vopravil, David Hájek, Oldrich Latal · 2026 · Frontiers in Soil Sciencedoi:10.3389/fsoil.2026.1993729
  12. [12]Quantifying Total Shrimp Pond Coverage Using Remote Sensing and a Machine Learning Algorithm: A Case Study in the Ecuadorian CoastRomina Cifuentes, Jorge Vulgarin, Daniel Ochoa, Stanislaus Sonnenholzner · 2026 · New Zealand Journal of Marine and Freshwater Researchdoi:10.1002/nzm2.70091
  13. [13]Artificial intelligence virtual patientsYanhong Pei, Xiaoya Li, Xin Yin, Yinglong Zhang, Ping Song, Yuanqiao Liang · 2027 · EngMedicinedoi:10.1016/j.engmed.2026.100171
  14. [14]Comparative assessment of bacterial diversity in small water bodies of the Eastern Pamir mountains using amplicon-based and metagenomic approachesŁukasz Łach, Agnieszka Rudak, Anna Karnkowska, Michał Karlicki, Metody Hollender, Valentina Smacchia · 2026 · Scientific Reportsdoi:10.1038/s41598-026-79844-w

All 14 references resolved against Crossref on Nov 15, 2026. Full verification record.

Get the findings, not the noise

A short email when something worth knowing clears the verification gate. No takes. Unsubscribe in one click.