Dynamical Correlation for Multivariate Longitudinal Data

Joel A. Dubin, Hans‐Georg MüllerJournal of the American Statistical Association 2005 · DOI

TL;DR

This paper introduces "dynamical correlation," a novel nonparametric method to quantify the relationships between multiple time-varying measurements (longitudinal data). It offers a more robust and interpretable alternative to existing methods, particularly for complex datasets with irregular sampling, and is demonstrated on a study of blood protein levels.

What Problem Does This Paper Address?

Analyzing data collected over time from the same subjects (longitudinal data) is crucial in many fields, from medicine to economics. When we have multiple such measurements for each subject (multivariate longitudinal data), understanding how these different time series relate to each other becomes a significant challenge. Existing methods often struggle with irregular sampling times, require restrictive assumptions, or are difficult to interpret. This paper aims to fill this gap by providing a flexible, nonparametric method to capture the "dynamical" dependencies – how different measurements co-vary over time, including their rates of change and time shifts.

Methods

The core of the paper is the introduction of "dynamical correlation." This method treats each time series as a realization of a random process and defines correlation based on the "angle" between standardized versions of these curves in a Hilbert space. Key aspects include:

  • Standardization: Curves are standardized to have zero mean and unit variance, making them comparable.
  • Functional Inner Product: A generalized inner product is used to define relationships between functions.
  • Extension to Derivatives and Lags: The method is extended to capture correlations between the rates of change (derivatives) of the curves and to account for time shifts (lags) between them.
  • Estimation: A practical estimation procedure is proposed, involving pre-smoothing of discrete data and averaging individual subject correlations.
  • Bootstrap for Intervals: A two-stage bootstrap procedure is used to construct confidence intervals for the estimated correlations, accounting for smoothing uncertainty.

Datasets

The methods are illustrated using:

  • Simulated Data: Used to assess the sensitivity of the dynamical correlation estimate to bandwidth choices and to compare it with functional canonical correlation.
  • Longitudinal Protein Data: Data from a study of 35 hemodialysis patients, measuring five acute-phase blood proteins over time (12-18 measurements per subject). This dataset is used to demonstrate the practical application of the method.

Key Insights and Ideas

  • Dynamical vs. Static Correlation: Dynamical correlation captures how variables move together over time, offering a richer understanding than static (cross-sectional) correlation, which only looks at a single point in time.
  • Robustness to Irregular Sampling: The method is designed to handle irregular and non-matching observation times, a common issue in real-world longitudinal studies.
  • Interpretability: Dynamical correlation provides a more intuitive interpretation, akin to the cosine of an angle between functions, making it easier to understand the nature of the relationship.
  • Breakdown of Functional Canonical Correlation: The paper highlights that functional canonical correlation, a related but different method, can break down and become uninterpretable with a moderate to large number of repeated measurements, a problem dynamical correlation avoids.
  • Biomedical Relevance: The application to blood proteins reveals biologically plausible correlations, such as negative correlations between negative acute-phase proteins (NAPP's) and positive acute-phase proteins (APP's).

Key Results

  • Sensitivity Analysis: The dynamical correlation estimate is shown to be robust to the choice of smoothing bandwidth, except at extreme values.
  • Comparison with Functional Canonical Correlation: Dynamical correlation performs reliably even with a large number of repeated measurements, whereas functional canonical correlation breaks down, yielding uninterpretable results.
  • Protein Data Analysis: The estimated dynamical correlation matrix supports the hypothesis of negative correlations between NAPP's and APP's. The first dynamical principal component explains 80% of the longitudinal variability, suggesting a linear contrast between these protein groups.
  • Lagged Correlations: Analysis of lags reveals that changes in albumin appear to be anticipated by changes in C-reactive protein, a biologically plausible finding.

Limitations and Future Work

  • Number of Measurements: While robust, the method generally requires a sufficient number of repeated measurements per subject (caution advised for fewer than 10, and generally not recommended for 3 or 4).
  • Bandwidth Selection: While the estimate is not highly sensitive to bandwidth, an automatic bandwidth selection procedure would be desirable.
  • Future Directions: Extensions include investigating correlations over specific time intervals, incorporating covariates, and developing automatic bandwidth selection.

Summary updated

Abstract

Nonparametric methodology for longitudinal data analysis is becoming increasingly popular. The analysis of multivariate longitudinal data, where data on several time courses are recorded for each subject, has received considerably less attention, despite its importance for practical data analysis. In particular, there is a need for measures and estimates to capture dependency between the components of vector-valued longitudinal data. We propose and analyze a simple and effective nonparametric method to quantify the covariation of components of multivariate longitudinal observations, which are viewed as realizations of a random process. This includes the notion of a correlation between derivatives and time-shifted versions. The concept of dynamical correlation is based on a scalar product obtained from pairs of standardized smoothed curves. The proposed method can be used when observation times are irregular and not matching between subjects or between responses within a subject. For higher-dimensional data, one may construct a dynamical correlation matrix that then serves as a starting point for standard multivariate analysis techniques, such as principal components. We iliustrate our methods via simulations as well as with data on five acute-phase blood proteins measured longitudinally from a study of hemodialysis patients.
BibTeX
@article{dubin2005dynamical,
  author = {Joel A. Dubin and Hans‐Georg Müller},
  title = {{Dynamical Correlation for Multivariate Longitudinal Data}},
  journal = {Journal of the American Statistical Association},
  year = {2005},
  doi = {10.1198/016214504000001989},
}
Figure