Intraclass correlation: Improved modeling approaches and applications for neuroimaging

Area of Science:

• Neuroimaging
• Statistical Modeling
• Reliability Analysis

Background:

• Intraclass correlation (ICC) is a key reliability metric in neuroimaging, commonly analyzed using ANOVA.
• The conventional ANOVA framework for ICC has limitations in modeling flexibility and handling specific data circumstances.

Purpose of the Study:

• To comprehensively review ICC usage in neuroimaging.
• To address limitations of the ANOVA framework by developing and evaluating advanced ICC modeling strategies.
• To provide practical recommendations for ICC analysis and reporting in neuroimaging.

Main Methods:

• Extended the conventional ICC model to include linear mixed-effects (LME), regularized mixed-effects (RME), multilevel mixed-effects (MME), and regularized multilevel mixed-effects (RMME) approaches.
• Incorporated precision information of effect estimates and addressed negative ICC values.
• Compared model performance using theoretical considerations and a real experimental dataset.

Main Results:

• The four new modeling strategies (LME, RME, MME, RMME) offer improved flexibility, accommodate missing data, and provide fixed effect estimates.
• MME and RMME approaches provide more accurate variance component decomposition for robust ICC computation.
• The consistency version, ICC(3,1), is recommended for whole-brain ICC analysis over the absolute agreement version, ICC(2,1).

Conclusions:

• Recommends MME or RMME for ICC estimation when precision information is available, and LME or RME otherwise.
• Highlights the utility of ICC(3,1) for whole-brain analysis.
• Introduces the open-source program 3dICC for implementing these advanced ICC models.