Multivariate Bayesian clustering using covariate‐informed components with application to boreal vegetation sensitivity

• Published in: Biometrics Vol. 78; no. 4; pp. 1427 - 1440
• Main Authors: Scharf, Henry R., Raiho, Ann M., Pugh, Sierra, Roland, Carl A., Swanson, David K., Stehn, Sarah E., Hooten, Mevin B.
• Format: Journal Article
• Language: English
• Published: United States Blackwell Publishing Ltd 01.12.2022
• Subjects: Alaska; Alaska national parks; Bayes Theorem; Bayesian analysis; Climate Change; Clustering; clustering regression coefficients; community ecology; Ecosystem; joint distribution models; Labels; Mathematical models; Multivariate analysis; Plants; Robustness; Sensitivity; Smoothness; Statistical models; Vegetation; Alaska; Alaska national parks; clustering regression coefficients; joint distribution models; community ecology; climate change
• ISSN: 0006-341X; 1541-0420
• DOI: 10.1111/biom.13507

Climate change is impacting both the distribution and abundance of vegetation, especially in far northern latitudes. The effects of climate change are different for every plant assemblage and vary heterogeneously in both space and time. Small changes in climate could result in large vegetation responses in sensitive assemblages but weak responses in robust assemblages. But, patterns and mechanisms of sensitivity and robustness are not yet well understood, largely due to a lack of long‐term measurements of climate and vegetation. Fortunately, observations are sometimes available across a broad spatial extent. We develop a novel statistical model for a multivariate response based on unknown cluster‐specific effects and covariances, where cluster labels correspond to sensitivity and robustness. Our approach utilizes a prototype model for cluster membership that offers flexibility while enforcing smoothness in cluster probabilities across sites with similar characteristics. We demonstrate our approach with an application to vegetation abundance in Alaska, USA, in which we leverage the broad spatial extent of the study area as a proxy for unrecorded historical observations. In the context of the application, our approach yields interpretable site‐level cluster labels associated with assemblage‐level sensitivity and robustness without requiring strong a priori assumptions about the drivers of climate sensitivity.