Comparison of regression methods for phenology.

Area of Science:

• Ecology
• Biometeorology
• Statistical Modeling

Background:

• Investigating relationships between biological records and weather data is crucial.
• Existing methods include mechanistic models (e.g., thermal time) and linear regression.
• Stepwise regression is common in phenology but can lead to information loss due to data aggregation.

Purpose of the Study:

• To compare the performance of stepwise regression, penalised signal regression (PSR), and fusion methods for analyzing phenological data.
• To evaluate these methods using simulations from mechanistic spring warming and sequential models.
• To identify the most robust and effective regression technique for phenological studies.

Main Methods:

• Simulations were conducted using two mechanistic models: spring warming and sequential.
• Three regression methods were compared: stepwise regression, penalised signal regression (PSR), and fusion (a sparse PSR variant).
• Performance was evaluated based on the accuracy and robustness of predictions using temperature days as covariates.

Main Results:

• PSR and fusion outperformed stepwise regression for the spring warming model when appropriate temperature covariates were selected.
• PSR demonstrated the best performance for the sequential model.
• Fusion showed robustness to a large number of redundant temperature covariates, unlike PSR whose performance declined.

Conclusions:

• PSR and fusion are valuable alternatives to stepwise regression in phenological studies.
• Using PSR and fusion in tandem, and varying the number of covariates, is recommended for optimal results.
• Fusion offers robustness against redundant covariates, making it a reliable choice for complex datasets.