Model selection in occupancy models: Inference versus prediction.

Area of Science:

• Ecology
• Ecological Modeling
• Statistical Ecology

Background:

• Occupancy models are crucial for understanding species occurrence patterns.
• Model selection often involves choosing between various occupancy and detection covariates.
• Information-theoretic approaches, like Akaike's Information Criterion (AIC), are widely used for model selection.

Purpose of the Study:

• To investigate the impact of collider bias on occupancy models.
• To evaluate the performance of AIC and Bayesian Information Criterion (BIC) in model selection under collider bias.
• To differentiate the effects of collider bias on parameter estimation versus prediction accuracy.

Main Methods:

• Utilized simulation studies to examine collider bias (M-bias) in occupancy and detection processes.
• Assessed model selection using AIC and BIC with simulated data.
• Compared parameter estimates and prediction accuracy across different model scenarios.

Main Results:

• Collider bias in the occupancy process led to inaccurate focal covariate effect estimates but improved prediction accuracy.
• Collider bias in the detection process did not affect focal estimates and slightly improved prediction accuracy.
• AIC and BIC selected models with better predictive performance but potentially biased parameter estimates.

Conclusions:

• Information criteria can be used for covariate selection in occupancy models if prediction is the primary goal.
• Caution is advised when using information criteria for inference on environmental effects on occupancy due to potential bias.
• Detection covariates can generally be selected using information criteria irrespective of the modeling objective.