Generalized prediction intervals for treatment effects in random-effects models.

Area of Science:

• Statistics
• Biometry

Background:

• Linear random-effects models are widely used in various scientific fields.
• Accurate prediction intervals for random effects are crucial for reliable statistical inference.
• Existing methods, such as restricted maximum likelihood (REML), may have coverage issues.

Purpose of the Study:

• To derive and evaluate generalized prediction intervals for random effects in two-way linear random-effects models.
• To compare the performance of these new intervals against established methods.
• To assess interval coverage for both balanced and unbalanced data, with and without interaction terms.

Main Methods:

• Derivation of generalized prediction intervals for random effects.
• Simulation study using data from an agricultural field experiment.
• Comparison with prediction intervals from REML, Satterthwaite, and Kenward-Roger methods.

Main Results:

• Generalized prediction intervals demonstrated coverage closer to the nominal 0.95 level compared to REML-based intervals.
• The proposed method showed improved accuracy in estimating prediction interval coverage.
• Performance was evaluated across different model complexities (with/without interaction) and data structures (balanced/unbalanced).

Conclusions:

• Generalized prediction intervals provide a more reliable alternative for random effects in linear models.
• The findings suggest an improvement over traditional REML approaches for prediction interval accuracy.
• This work contributes to more precise statistical predictions in complex experimental designs.