Multilevel functional principal components analysis in the case of unbalanced design and small number of subjects

Abstract

Multilevel Functional Principal Component Analysis (MFPCA) distinguishes two types of variability present in the functional data with repeated measurements - the variability between subjects and the variability within subjects. This technique is appropriate for detection of a subject-specic pattern and the differences between functional data obtained from different sessions of the same subject.

In the original article only balanced design is considered - for each subject the same number of measurements is available. On the other hand the authors claim, that the assumption of a balanced design is not restrictive and the method itself is also able to deal with unbalanced data.

However, the MFPCA algorithm faces problems when i) the number of subjects is small, ii) the variability within subjects is close to zero because their observations are almost identical for all sessions, and iii) unbalanced design is considered, where for some subjects the number of observations is significantly lower in comparison to the remaining part of the dataset. In all such cases, MFPCA is not able to properly detect subject-specific profiles and a part of the between subjects variability is mixed with within{subjects variability.

In this study we implement the algorithm which deals with the problem and we validate the approach on an artificial and real dataset.


Go back