Abstract
We present a new method based on Functional Data Analysis (FDA) for detecting associations between one or more scalar covariates and a longitudinal response, while correcting for other variables. Our methods exploit the temporal structure of longitudinal data in ways that are otherwise difficult with a multivariate approach. Our procedure, from an FDA perspective, is a departure from more established methods in two key aspects. First, the raw longitudinal phenotypes are assembled into functional trajectories prior to analysis. Second, we explore an association test that is not directly based on principal components. We instead focus on quantifying the reduction in $L^{2}$ variability as a means of detecting associations. Our procedure is motivated by longitudinal genome wide association studies and, in particular, the childhood asthma management program (CAMP) which explores the long term effects of daily asthma treatments. We conduct a simulation study to better understand the advantages (and/or disadvantages) of an FDA approach compared to a traditional multivariate one. We then apply our methodology to data coming from CAMP. We find a potentially new association with a SNP negatively affecting lung function. Furthermore, this SNP seems to have an interaction effect with one of the treatments.
Introduction
Traditional genetic association studies often rely on "snapshots" in time, comparing a single measurement to a patient's genotype. However, many biological processes, such as child development or disease progression, are inherently longitudinal. Standard methods for handling data over time typically involve either simplifying the data into a single scalar (e.g., total change), which discards temporal information, or using multivariate analysis (e.g., MANOVA), which often struggles with irregular visit timing and fails to account for the underlying "smoothness" of biological growth.
While Functional Data Analysis (FDA) treats these measurements as continuous curves, most existing tools rely heavily on Principal Component Analysis (PCA). PCA-based tests can be unstable and "noisy" when applied to the hundreds of thousands of markers in a typical Genome-Wide Association Study (GWAS). This paper proposes a novel FDA framework that avoids these pitfalls by utilizing a more stable $L^2$ variability test.
Methods
The proposed framework utilizes a "smoothing first, inference second" approach to handle longitudinal phenotypes.
Functional Trajectories
Raw, noisy clinical measurements are converted into smooth, continuous mathematical functions ($Y_n(t)$) using B-splines. This transformation allows the model to handle subjects with missing visits or irregular intervals by treating the phenotype as a continuous trajectory rather than discrete points.
The $L^2$ Norm Test
Instead of projecting data onto a few potentially unstable principal components, the authors measure the reduction in the sum of squared norms ($L^2$ norm). This approach essentially weights each component by its eigenvalue (the amount of variability it explains), which naturally suppresses noise from minor components. This weighting scheme ensures the method remains robust even when conducting millions of tests across the genome.
Datasets and Results
The methodology was applied to the Childhood Asthma Management Program (CAMP), a longitudinal clinical trial following 540 children over four years with 16 clinical visits each. The primary phenotype was FEV1 (forced expiratory volume), a measure of lung function, analyzed against approximately 670,000 SNPs.
Key Findings
- New Genetic Association: The study identified a significant association between SNP rs12734254 (in the gene ST6GALNAC5) and lung function ($p = 1.016 \times 10^{-8}$). This variant is associated with a decrease in lung function that worsens over time.
- Gene-by-Treatment Interaction: The SNP was found to interact with Budesonide (an inhaled steroid). While the SNP generally harms lung function, the drug appears to counter this negative effect.
- Superior Sensitivity: When the same data was analyzed using a simple scalar approach (looking only at the difference between the first and last visit), the p-value for the interaction was a non-significant 0.135. The FDA approach's ability to capture the "shape" of the response over time made the discovery possible.
Discussion
Simulation studies revealed that as the underlying biological process becomes smoother, the advantage of FDA over traditional multivariate methods grows significantly. Furthermore, the $L^2$ approach proved more stable than PCA-based methods, which often produced false signals when rare genetic variants aligned with minor, noisy principal components.
Future work will address limitations such as explicitly accounting for interpolation error during the initial smoothing step and adapting the model for non-stationary data where the correlation structure changes drastically over time. Despite these challenges, the FDA framework provides a powerful tool for uncovering the temporal dynamics of genetic influences on complex traits.
Summary updated