Functional Generalized Additive Models.

Area of Science:

• Statistics
• Biostatistics
• Functional Data Analysis

Background:

• Association studies require robust models for scalar responses and functional predictors.
• Existing methods like principal component-based models may not fully capture functional covariate information.
• Generalized additive models (GAMs) provide a flexible framework but need extension for functional data.

Purpose of the Study:

• To introduce the functional generalized additive model (FGAM) as a novel regression approach for association studies.
• To extend generalized additive models by directly incorporating functional predictors.
• To provide a flexible and powerful tool for analyzing scalar responses with functional covariates.

Main Methods:

• The functional generalized additive model (FGAM) models the link-transformed mean response using an integral of an unknown regression function with respect to the functional covariate.
• Estimation of the regression function is performed using tensor-product B-splines with roughness penalties.
• A pointwise quantile transformation is applied to the functional predictor to ensure stable B-spline estimation.

Main Results:

• The FGAM was evaluated using simulated data, demonstrating its effectiveness.
• Predictive performance was compared favorably against existing scalar-on-function regression alternatives.
• The model's utility was illustrated through applications in brain tractography, analyzing diffusion tensor imaging data for disease status and cognitive test scores.

Conclusions:

• The functional generalized additive model (FGAM) offers a significant advancement for association studies involving functional predictors.
• The direct incorporation of functional covariates provides a more natural and potentially more powerful analysis than component-based approaches.
• FGAM is a valuable tool for diverse applications, including neuroimaging, where complex functional data is prevalent.