Selecting a separable parametric spatiotemporal covariance structure for longitudinal imaging data.

Area of Science:

• Biomedical Imaging
• Statistical Analysis
• Medical Research

Background:

• Longitudinal imaging studies provide valuable insights into anatomical changes over time.
• Analyzing such data is challenging due to inherent spatial and temporal correlations.
• Accurate statistical modeling is crucial for reliable interpretation of imaging results.

Purpose of the Study:

• To propose a linear model with a separable parametric spatiotemporal error structure for analyzing repeated imaging data.
• To evaluate the performance of information criteria in selecting appropriate correlation structures.
• To assess the impact of covariance structure misspecification on statistical inference.

Main Methods:

• Utilized a linear model incorporating separable parametric spatiotemporal correlation functions (spatial: exponential, spherical, Matérn; temporal: compound symmetric, autoregressive-1, Toeplitz, unstructured).
• Conducted a simulation study inspired by longitudinal cardiac imaging data.
• Compared information criteria for model selection and evaluated Type I and II error rates under model misspecification.

Main Results:

• Information criteria demonstrated high accuracy in selecting the correct separable parametric spatiotemporal correlation structure.
• Misspecification of the covariance structure led to inflated Type I error rates or overly conservative tests, reducing statistical power.
• Demonstrated practical application using clinical data, showing how covariance structure choice impacts fixed-effects inference.

Conclusions:

• The proposed linear model with separable parametric spatiotemporal error structures effectively analyzes longitudinal imaging data.
• Accurate selection of covariance structures is vital to avoid biased statistical inference and maintain adequate power.
• This methodology offers a robust approach for understanding anatomical changes in longitudinal studies.