Functional principal component analysis forsparse censored data

Area of Science:

• Statistics
• Functional Data Analysis
• Biostatistics

Background:

• Functional principal component analysis (FPCA) is crucial for analyzing functional data.
• Existing FPCA methods fail with censored functional observations, leading to biased results.
• Censoring occurs when measurement instruments limit recorded data to a specific range.

Purpose of the Study:

• To extend FPCA to handle noisy, sparse, and censored functional data.
• To develop bias-corrected estimators for mean, covariance, and FPCA scores.
• To enable accurate functional data analysis in the presence of instrument-induced censoring.

Main Methods:

• Utilized local loglikelihood maximization to estimate mean and covariance surfaces.
• Developed a procedure for positive semidefinite covariance surface estimation.
• Constructed a predictor for FPCA scores conditional on censored data.
• Applied the method to a generalized functional linear model.

Main Results:

• The proposed method provides smooth estimates of latent functional processes.
• Achieved a positive semidefinite covariance surface without eigenvalue adjustments.
• Demonstrated improved predictive performance and reduced bias in simulations.
• Established convergence rates for the new estimators.

Conclusions:

• The novel FPCA framework effectively accommodates censored functional data.
• The method offers a significant improvement over existing techniques for censored data.
• Successfully applied to classify eating disorder diagnoses using censored blood glucose data in type 1 diabetes patients.