A critical issue in model-based inference for studying trait-based community assembly and a solution

Area of Science:

• Ecology
• Environmental Science
• Statistical Modeling

Background:

• Assessing trait-environment associations is challenging due to differing observation units (species traits, site environments, species-site abundances).
• Existing correlation-based methods (CWM, fourth-corner, RLQ) use a pmax test involving separate site and species resampling for statistical validity.
• Recently proposed regression-based methods using generalized linear models (GLM) with site-based resampling offer a potential alternative.

Purpose of the Study:

• To investigate the performance of GLM-based methods for trait-environment association testing, including those mimicking the pmax test.
• To evaluate the statistical validity and error rates of GLM approaches compared to traditional methods.
• To identify the underlying statistical issues with GLM-based site-resampling methods.

Main Methods:

• Simulations were conducted using models with additional random variation in species responses to the environment.
• Site-based resampling tests using GLM were evaluated for type I error rates.
• The performance of GLM-based pmax tests was compared against traditional pmax tests.
• Predictive modeling using site-based cross-validation was employed to assess trait-environment interactions.

Main Results:

• Site-based resampling tests using GLM exhibited severely inflated type I error rates (up to 90% at a 5% nominal level) when random variation was present.
• Predictive modeling frequently identified non-existent trait-environment interactions.
• The identified problem stems from ignoring random effects, not omitted variable bias.
• The GLM-based pmax test controlled type I error effectively in simpler models but showed slight inflation in complex models with missing variables.
• The fourth-corner test provided similar results to GLM-based tests for screening but with significantly less computation time.

Conclusions:

• Standard GLM-based methods with site-based resampling are unreliable for trait-environment association testing due to inflated type I errors.
• Ignoring random effects is the critical issue, leading to spurious findings.
• The GLM-based pmax test offers improved error control over simple GLM site-resampling, but caution is needed in complex scenarios.
• The fourth-corner method remains a computationally efficient and reliable option for initial screening of trait-environment associations.