Clustering of functional data prone to complex heteroscedastic measurement error

TL;DR

This paper develops and evaluates clustering methods for functional data affected by complex heteroscedastic measurement errors, demonstrating improved accuracy through error adjustment and flexible application across different clustering approaches.

Contribution

It introduces two-stage mixed-effects model-based methods to adjust for measurement errors before clustering functional data, addressing a gap in handling error-prone measurements.

Findings

Measurement error adjustment improves clustering accuracy.

The methods are robust across different initializations.

Application to real datasets demonstrates practical utility.

Abstract

Several factors make clustering of functional data challenging, including the infinite-dimensional space to which observations belong and the lack of a defined probability density function for the functional random variable. To overcome these barriers, researchers either assume that observations belong to a finite-dimensional space spanned by basis functions or apply nonparametric smoothing methods to the functions prior to clustering. Although extensive literature describes clustering methods for functional data, few studies have explored the clustering of measurement error--prone function-valued data. In this work, we consider clustering methods for functional data prone to complex, heteroscedastic measurement errors. Two stage-based methods using mixed-effects models are first applied to adjust for measurement error bias, followed by cluster analysis of the measurement error--adjusted curves. Through simulations, we investigate how varying sample size, the magnitude of measurement error, and the presence of complex heteroscedastic measurement errors influence the cluster analysis of functional data. Our results indicate that failing to account for measurement errors and the correlation structures associated with frequently collected functional data reduces the accuracy of identifying the true latent groups or clusters. The method consistently produces better results regardless of the initial clustering values used. Moreover, it is flexible and can be applied to various clustering approaches, based on the specific distribution of the data. The developed methods are applied to two data sets: a school-based study of energy expenditure among elementary school-aged children in Texas and data from the National Health and Nutrition Examination Survey on participants' physical activity monitored by wearable devices at frequent intervals.