A new approach for modeling generalization gradients: a case for hierarchical models

Area of Science:

• Psychology
• Statistics
• Behavioral Science

Background:

• Repeated measures analysis-of-variance (rANOVA) is the standard statistical method for generalization research.
• rANOVA imposes restrictions, including limitations on including continuous independent variables and assumptions like sphericity.
• Violations of sphericity can increase Type I errors, particularly in typical sample sizes used in generalization studies.

Purpose of the Study:

• To advocate for the use of hierarchical models in analyzing generalization gradients.
• To highlight the advantages of hierarchical models over rANOVA in generalization research.
• To demonstrate the practical benefits and statistical superiority of hierarchical models.

Main Methods:

• A simulation study was conducted to compare hierarchical models with rANOVA.
• The study evaluated the performance of hierarchical models concerning continuous independent variables and sphericity assumptions.
• A worked example of a hierarchical model was presented for practical interpretation.

Main Results:

• Hierarchical models demonstrated superiority over rANOVA in the simulation study.
• Mauchly's sphericity test was found to be inefficient for typical sample sizes in generalization research.
• Violations of sphericity were confirmed to elevate the risk of Type I errors.

Conclusions:

• Hierarchical models provide a more flexible and robust statistical framework for generalization research.
• The adoption of hierarchical models can overcome key limitations of rANOVA, leading to more accurate analyses.
• Researchers are encouraged to utilize hierarchical models for a deeper understanding of generalization gradients.