Cross‐validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure

Ecography, Journal Year: 2016, Volume and Issue: 40(8), P. 913 - 929

Published: Dec. 9, 2016

Ecological data often show temporal, spatial, hierarchical (random effects), or phylogenetic structure. Modern statistical approaches are increasingly accounting for such dependencies. However, when performing cross‐validation, these structures regularly ignored, resulting in serious underestimation of predictive error. One cause the poor performance uncorrected (random) noted by modellers, dependence that persist as model residuals, violating assumption independence. Even more concerning, because overlooked, is structured also provides ample opportunity overfitting with non‐causal predictors. This problem can even if remedies autoregressive models, generalized least squares, mixed models used. Block where split strategically rather than randomly, address issues. blocking strategy must be carefully considered. Blocking space, time, random effects distance, while dependencies data, may unwittingly induce extrapolations restricting ranges combinations predictor variables available training, thus overestimating interpolation errors. On other hand, deliberate space improve error estimates extrapolation modelling goal. Here, we review ecological literature on non‐random and blocked cross‐validation approaches. We provide a series simulations case studies, which that, all instances tested, block nearly universally appropriate goal predicting to new selecting causal recommend used wherever exist dataset, no correlation structure visible fitted account correlations.

Language: Английский

---

Improving the forecast for biodiversity under climate change

Science, Journal Year: 2016, Volume and Issue: 353(6304)

Published: Sept. 8, 2016

BACKGROUND As global climate change accelerates, one of the most urgent tasks for coming decades is to develop accurate predictions about biological responses guide effective protection biodiversity. Predictive models in biology provide a means scientists project changes species and ecosystems response disturbances such as change. Most current predictive models, however, exclude important mechanisms demography, dispersal, evolution, interactions. These have been shown be mediating past present Thus, modeling efforts do not sufficiently predictions. Despite many complexities involved, biologists are rapidly developing tools that include key processes needed improve accuracy. The biggest obstacle applying these more realistic data inform them almost always missing. We suggest ways fill this growing gap between model sophistication information predict prevent damaging aspects life on Earth. ADVANCES On basis empirical theoretical evidence, we identify six commonly shape yet too often missing from models: physiology; history, phenology; interactions; evolutionary potential population differentiation; colonization, range dynamics; environmental variation. prioritize types each proxies or difficult collect. show even well-studied species, lack critical would necessary apply realistic, mechanistic models. Consequently, limitations likely override gains accuracy Given enormous challenge collecting detailed millions around world, highlight practical methods promote greatest Trait-based approaches leverage sparse make general inferences unstudied species. Targeting with high sensitivity disproportionate ecological impact can yield insights future ecosystem Adaptive schemes target while simultaneously improving OUTLOOK Strategic collections essential will allow us build generalizable our broader ability anticipate species’ other human-caused disturbances. By increasing making uncertainties explicit, deliver improved projections biodiversity under together characterizations uncertainty support informed decisions by policymakers land managers. Toward end, globally coordinated effort gaps advance climate-fueled crisis offers substantial advantages efficiency, coverage, Biologists take advantage lessons learned Intergovernmental Panel Climate Change’s development, coordination, integration projections. weather were greatly incorporating testing against station data. Biology same. need adopt meteorological approach predicting enhance mitigate services it provides humans. Emerging beginning incorporate Models used (clockwise top) evolution disease-harboring mosquitoes, environments use, physiological invasive cane toads, demographic penguins climates, climate-dependent dispersal behavior butterflies, mismatched interactions butterflies their host plants. advances, seldom necessitating new collect relevant parameterize biologically

Language: Английский

---

Global expansion and redistribution of Aedes-borne virus transmission risk with climate change

PLoS neglected tropical diseases, Journal Year: 2019, Volume and Issue: 13(3), P. e0007213 - e0007213

Published: March 28, 2019

Forecasting the impacts of climate change on Aedes-borne viruses—especially dengue, chikungunya, and Zika—is a key component public health preparedness. We apply an empirically parameterized model viral transmission by vectors Aedes aegypti Ae. albopictus, as function temperature, to predict cumulative monthly global risk in current climates, compare them with projected 2050 2080 based general circulation models (GCMs). Our results show that if mosquito range shifts track optimal temperature ranges for (21.3–34.0°C aegypti; 19.9–29.4°C albopictus), we can expect poleward virus distributions. However, differing thermal niches two produce different patterns under change. More severe scenarios larger population exposures aegypti, but not albopictus most extreme cases. Climate-driven from both mosquitoes will increase substantially, even short term, Europe. In contrast, significant reductions suitability are expected noticeably southeast Asia west Africa. Within next century, nearly billion people threatened new exposure spp. worst-case scenario. As major net losses year-round predicted project shift towards more seasonal across regions. Many other complicating factors (like limits evolution) exist, overall our indicate while lead increased viruses, increases occur at intermediate scenarios.

Language: Английский

---

Outstanding Challenges in the Transferability of Ecological Models

Trends in Ecology & Evolution, Journal Year: 2018, Volume and Issue: 33(10), P. 790 - 802

Published: Aug. 28, 2018

Predictive models are central to many scientific disciplines and vital for informing management in a rapidly changing world. However, limited understanding of the accuracy precision transferred novel conditions (their 'transferability') undermines confidence their predictions. Here, 50 experts identified priority knowledge gaps which, if filled, will most improve model transfers. These summarized into six technical fundamental challenges, which underlie combined need intensify research on determinants ecological predictability, including species traits data quality, develop best practices transferring models. Of high importance is identification widely applicable set transferability metrics, with appropriate tools quantify sources impacts prediction uncertainty under conditions.

Language: Английский

---

Iterative near-term ecological forecasting: Needs, opportunities, and challenges

Proceedings of the National Academy of Sciences, Journal Year: 2018, Volume and Issue: 115(7), P. 1424 - 1432

Published: Jan. 30, 2018

Two foundational questions about sustainability are “How ecosystems and the services they provide going to change in future?” do human decisions affect these trajectories?” Answering requires an ability forecast ecological processes. Unfortunately, most forecasts focus on centennial-scale climate responses, therefore neither meeting needs of near-term (daily decadal) environmental decision-making nor allowing comparison specific, quantitative predictions new observational data, one strongest tests scientific theory. Near-term opportunity iteratively cycle between performing analyses updating light evidence. This iterative process gaining feedback, building experience, correcting models methods is critical for improving forecasts. Iterative, forecasting will accelerate research, make it more relevant society, inform sustainable under high uncertainty adaptive management. Here, we identify immediate societal needs, opportunities, challenges forecasting. Over past decade, data volume, variety, accessibility have greatly increased, but remain interoperability, latency, quantification. Similarly, ecologists made considerable advances applying computational, informatic, statistical methods, opportunities exist forecast-specific theory, cyberinfrastructure. Effective also require changes training, culture, institutions. The need start now; time making ecology predictive here, learning by doing fastest route drive science forward.

Language: Английский

---

Predicting into unknown space? Estimating the area of applicability of spatial prediction models

Methods in Ecology and Evolution, Journal Year: 2021, Volume and Issue: 12(9), P. 1620 - 1633

Published: June 1, 2021

Predictive modelling using machine learning has become very popular for spatial mapping of the environment. Models are often applied to make predictions far beyond sampling locations where new geographic might considerably differ from training data in their environmental properties. However, areas predictor space without support problematic. Since model no knowledge about these environments, have be considered uncertain. Estimating area which a prediction can reliably is required. Here, we suggest methodology that delineates "area applicability" (AOA) define as area, cross-validation error applies. We first propose "dissimilarity index" (DI) based on minimum distance space, with predictors being weighted by respective importance model. The AOA then derived applying threshold DI calculated respect strategy used training. test ideal simulated and compare within illustrate approach case study. Our simulation study suggests at .95 quantile data. Using this threshold, comparable RMSE model, while does not apply outside AOA. This applies models trained randomly distributed data, well when clustered applied. report alongside predictions, complementary validation measures.

Language: Английский

---

A practical guide to selecting models for exploration, inference, and prediction in ecology

Ecology, Journal Year: 2021, Volume and Issue: 102(6)

Published: March 12, 2021

Abstract Selecting among competing statistical models is a core challenge in science. However, the many possible approaches and techniques for model selection, conflicting recommendations their use, can be confusing. We contend that much confusion surrounding selection results from failing to first clearly specify purpose of analysis. argue there are three distinct goals modeling ecology: data exploration, inference, prediction. Once goal articulated, an appropriate procedure easier identify. review highlight strengths weaknesses relative each goals. then present examples prediction using time series butterfly population counts. These show how approach flows naturally goal, leading different selected purposes, even with exactly same set. This illustrates best practices ecologists should serve as reminder recipes cannot substitute critical thinking or use independent test hypotheses validate predictions.

Language: Английский

---

Biodiversity scenarios neglect future land‐use changes

Global Change Biology, Journal Year: 2016, Volume and Issue: 22(7), P. 2505 - 2515

Published: March 7, 2016

Efficient management of biodiversity requires a forward-looking approach based on scenarios that explore changes under future environmental conditions. A number ecological models have been proposed over the last decades to develop these scenarios. Novel modelling approaches with strong theoretical foundation now offer possibility integrate key and evolutionary processes shape species distribution community structure. Although is affected by multiple threats, most studies addressing effects focus single threat only. We examined published during 25 years developed predict climate, land-use land-cover change projections. found mostly impacts climate largely neglect in land use cover. The emphasis has increased time reached maximum. Yet, direct destruction degradation habitats through are among significant immediate threats biodiversity. argue current state integration between system sciences leading biased estimation actual risks therefore constrains implementation policy responses decline. suggest research directions at crossroads face challenge developing interoperable plausible projections anticipate full range their potential An intergovernmental platform needed stimulate such collaborative efforts emphasize societal political relevance taking up this challenge.

Language: Английский

---

Seasonality and uncertainty in global COVID-19 growth rates

Proceedings of the National Academy of Sciences, Journal Year: 2020, Volume and Issue: 117(44), P. 27456 - 27464

Published: Oct. 13, 2020

The virus causing COVID-19 has spread rapidly worldwide and threatens millions of lives. It remains unknown, as April 2020, whether summer weather will reduce its spread, thereby alleviating strains on hospitals providing time for vaccine development. Early insights from laboratory studies research related viruses predicted that would decline with higher temperatures, humidity, ultraviolet (UV) light. Using current, fine-scaled data global reports infections, we develop a model explains 36% the variation in maximum growth rates based demography (17%) country-specific effects (19%). UV light is most strongly associated lower growth. Projections suggest that, without intervention, decrease temporarily during summer, rebound by autumn, peak next winter. Validation May June 2020 confirms generality climate signal detected. However, uncertainty high, probability weekly doubling >20% throughout absence social interventions. Consequently, aggressive interventions likely be needed despite seasonal trends.

Language: Английский

---

The priority of prediction in ecological understanding

Oikos, Journal Year: 2016, Volume and Issue: 126(1), P. 1 - 7

Published: Sept. 1, 2016

The objective of science is to understand the natural world; we argue that prediction only way demonstrate scientific understanding, implying should be a fundamental aspect all disciplines. Reproducibility an essential requirement good and arises from ability develop models make accurate predictions on new data. Ecology, however, with few exceptions, has abandoned as central focus faces its own crisis reproducibility. Models are where ecological understanding stored they source – no possible without model world. can improved in three ways: variables, functional relationships among dependent independent parameter estimates. Ecologists rarely test assess whether have made advances by identifying important elucidating relationships, or improving Without these tests it difficult know if more today than did yesterday. A commitment ecology would lead to, other things, mature (i.e. quantitative) hypotheses, prioritization modeling techniques appropriate for (e.g. using continuous variables rather categorical) and, ultimately, advancement towards general Synthesis therefore understanding. Here address how this inhibited progress explore renewed benefit ecologists. lack emphasis resulted discipline qualitative, imprecise hypotheses little concern results generalizable beyond when data were collected. allow ecologists critical questions about generalizability our making

Language: Английский

---