Multivariate Functional Mixed Model and Application
10.11783/j.issn.1002-3674.2025.04.003
- VernacularTitle:多元函数型混合模型及其应用
- Author:
Yufei MA
1
;
Chunxia LI
;
Tao ZHANG
Author Information
1. 山东大学齐鲁医学院公共卫生学院 250012
- Publication Type:Journal Article
- Keywords:
Longitudinal data;
Multivariate functional data;
Principal component;
Interpretability;
Colorectal cancer
- From:
Chinese Journal of Health Statistics
2025;42(4):491-495
- CountryChina
- Language:Chinese
-
Abstract:
Objective To introduce the fundamental principles of the multivariate functional mixed model(MFMM)and provide a methodological basis for interpreting the principal components derived from dimension reduction of multivariate functional data.Methods The MFMM models the mean function,auto-covariance function,and cross-covariance function within functions non-parametrically,clearly separating the shared latent processes common to all indicators from the specific latent processes unique to each indicator.In the case study,data from colorectal cancer patients at the Yunnan Cancer Hospital are used to extract the common and specific principal components of longitudinal carcinoembryonic antigen(CEA)and carbohydrate antigen 125(CA125)within 12 months post-surgery using MFMM,coupled with interpretation through a random survival forest prediction model.Results The MFMM identified 2 shared principal components for CEA and CA125,5 specific components for CEA,and 5 specific components for CA125.The random survival forest prediction model constructed based on these extracted components showed higher prediction accuracy than the baseline model.Furthermore,variable importance analysis revealed that the first specific component of CEA was the most important predictive variable after clinical baseline variables,while the importance of CA125 specific components was lower than that of CEA,consistent with the current consensus on colorectal cancer.Conclusion MFMM effectively manages the correlations among multiple longitudinal indicators and their changes over time,capturing the shared and specific processes of multiple longitudinal measurement indicators as they evolve.This enhances the model's interpretability and offers theoretical and practical advantages.
