Functional principal component analysis with informative observation times.

Area of Science:

• Statistics
• Biostatistics
• Longitudinal Data Analysis

Background:

• Functional principal component analysis (FPCA) is crucial for understanding longitudinal data variation.
• Existing FPCA methods often assume observation times are independent of outcomes, a fragile assumption in practice.
• Real-world data frequently exhibit correlations between observation timing and outcome trajectories.

Purpose of the Study:

• To develop an FPCA method that explicitly models informative observation time processes.
• To improve the accuracy of longitudinal data analysis when observation times are outcome-dependent.
• To provide a robust framework for forecasting and model building with correlated time and outcome data.

Main Methods:

• Proposed a novel FPCA approach modeling observation times via a general counting process dependent on time-varying factors.
• Employed inverse intensity weighting for the identification of mean, covariance function, and functional principal components.
• Utilized weighted penalized splines for estimation, establishing consistency and convergence rates for the proposed estimators.

Main Results:

• Simulation studies confirmed that the proposed weighted estimators are substantially more accurate than existing methods.
• Demonstrated superior performance in scenarios with a correlation between the observation time process and the longitudinal outcome process.
• Evaluated the finite-sample performance using data from the Acute Infection and Early Disease Research Program study.

Conclusions:

• The developed FPCA method effectively handles informative observation times, offering improved accuracy.
• This approach provides a more robust tool for analyzing longitudinal data where time and outcome are intertwined.
• The findings have significant implications for forecasting and model building in various scientific fields utilizing longitudinal studies.