lglasso: Longitudinal Graphical Lasso

Jie Zhou, Jiang Gui, Weston D. Viles, Anne G. HoenOpenalex Unknown 2022 · DOI · OpenAlex

Abstract

The lglasso package provides a robust framework for the $L_1$ penalized maximum likelihood estimation of precision matrices and correlation parameters in high-dimensional settings. Specifically designed for correlated observations, it accommodates longitudinal data with dampening correlation structures and clustered data with uniform correlation. By addressing the complexities of irregularly spaced observations, lglasso enables the identification of sparse interaction networks, such as those found in microbial 16S rRNA sequence data.

TL;DR

The lglasso R package extends the standard Graphical Lasso to handle non-independent longitudinal data. It allows researchers to accurately map high-dimensional interaction networks by accounting for temporal decay and irregular sampling intervals, ensuring robust results even when observations are correlated within subjects.

The Challenge: High-Dimensional Correlated Data

In fields ranging from genomics to ecology, understanding how variables like species or genes interact is critical. The Graphical Lasso (glasso) is a staple for estimating sparse precision matrices to reveal these networks. However, a core assumption of glasso is that observations are independent.

In longitudinal studies, this assumption fails because measurements from the same subject are naturally correlated over time. Furthermore, real-world data is often "messy," featuring irregularly spaced time points or clustered observations. Ignoring these dependencies leads to biased network estimates. lglasso addresses this by integrating correlation modeling directly into the network estimation process, specifically targeting scenarios where the number of variables ($p$) exceeds the number of observations.

Methodology

The package implements an $L_1$ penalized maximum likelihood estimation approach. Its innovation lies in the simultaneous estimation of the network (precision matrix) and the temporal correlation structure.

  • Correlation Modeling: It utilizes a dampening correlation function, typically $\exp(-\tau |t_i - t_j|^\text{type})$, to model how relationships decay over time ($t$).
  • Heterogeneity vs. Homogeneity: Users can choose between heterogeneous models, where each subject has a unique correlation parameter ($\tau_i$), or homogeneous models, where a global parameter is shared.
  • The EM Algorithm: For heterogeneous models, an Expectation-Maximization (EM) algorithm iteratively refines the precision matrix and the distribution of correlation parameters.
  • Model Selection: The Extended Bayesian Information Criterion (EBIC) is used to select optimal tuning parameters for the Lasso penalty.
Key parameters used in the lglasso estimation functions, including the tuning parameter (rho) and correlation type.

Built-in Datasets

To demonstrate its utility, the package includes sample_data, a 100-by-22 matrix derived from a longitudinal prospective project. It tracks 20 microbes across 13 clusters (subjects) over various time points. This dataset exemplifies the high-dimensional, sparse, and correlated nature of microbiome data that lglasso was built to analyze.

Core Insights

  • Beyond Independence: By treating the correlation parameter as a variable to be estimated alongside the network, the model becomes resilient to the realities of clinical and biological sampling.
  • Handling Irregularity: Unlike methods requiring binned or perfectly aligned time points, lglasso models correlation as a function of the actual time distance between samples.
  • Biological Relevance: While mathematically general, the package is optimized for 16S rRNA sequencing data, where identifying microbial interactions in dynamic environments is a primary goal.

Key Results and Outputs

The package provides several functions to extract statistical and biological insights:

  • Precision Matrix ($\Omega$): The primary output representing the inferred interaction network.
  • Correlation Parameters ($\tau$ and $\alpha$): Estimates of the rate at which correlations decay within subjects.
  • Likelihood and Testing: Functions like lli_homo and logdensity support rigorous statistical testing and model comparison via EBIC.
The package calculates internal components like the quasi-covariance matrix for individual subjects to refine global estimates.

In practice, lglasso successfully estimates sub-networks and calculates total interaction edges while accounting for different correlation types (e.g., "abs" for absolute distance or "sqr" for squared distance), providing a more nuanced view of variable interactions than traditional methods.

Summary updated

Abstract

For high-dimensional correlated observations, this package carries out the L_1 penalized maximum likelihood estimation of the precision matrix (network) and the correlation parameters. The correlated data can be longitudinal data (may be irregularly spaced) with dampening correlation or clustered data with uniform correlation. For the details of the algorithms, please see the paper Jie Zhou et al. Identifying Microbial Interaction Networks Based on Irregularly Spaced Longitudinal 16S rRNA sequence data &lt;<a href="https://doi.org/10.1101%2F2021.11.26.470159" target="_top">doi:10.1101/2021.11.26.470159</a>&gt;.
BibTeX
@inproceedings{zhou2022lglasso,
  author = {Jie Zhou and Jiang Gui and Weston D. Viles and Anne G. Hoen},
  title = {{lglasso: Longitudinal Graphical Lasso}},
  booktitle = {Openalex Unknown},
  year = {2022},
  doi = {10.32614/cran.package.lglasso},
}
Figure