Does modeling causal relationships improve the accuracy of predicting lactation milk yields?

Area of Science:

• Animal Science
• Statistical Modeling
• Machine Learning

Background:

• Accurate estimation of lactation milk yields is crucial for dairy herd management and health monitoring.
• Various statistical models exist, differing in their assumptions about relationships between variables like milk yield and health conditions.

Purpose of the Study:

• To compare the accuracy of correlational and causal models for estimating lactation milk yields.
• To evaluate model performance under different conditions, including the presence of mastitis.
• To assess the impact of model complexity and biological assumptions on prediction accuracy.

Main Methods:

• Simulated lactation data using Wood lactation curves as a benchmark.
• Compared three correlational models: best prediction, linear regression, and feed-forward neural networks.
• Compared two causal models: recursive structural equation model and recurrent neural networks.
• Assessed model accuracy based on prediction performance, particularly in scenarios with and without mastitis.

Main Results:

• Individual Wood lactation curves offered excellent parametric interpretation of lactation dynamics.
• Best prediction models excelled in the absence of mastitis but were suboptimal when mastitis was present and not accounted for.
• Recurrent neural networks demonstrated the highest accuracy when mastitis was present.
• Causal models, while aiding in inferring causality, faced challenges in precisely capturing complex biological relationships, and misspecification led to accuracy loss.

Conclusions:

• Modeling causal relationships does not inherently guarantee improved prediction accuracy for milk yields.
• A parsimonious model, balancing complexity and accuracy, is often preferred in practice.
• The choice of statistical model and the accurate accounting for covariates affecting milk yields are critical for reliable estimations.