Tree-Regularized Bayesian Latent Class Analysis for Improving Weakly Separated Dietary Pattern Subtyping in Small-Sized Subpopulations

Mengbing Li, Briana Stephenson, Zhenke WuThe Annals of Applied Statistics 2023 · arXiv

TL;DR

This paper introduces a novel "Tree-Regularized Bayesian Latent Class Analysis" (DDT-LCM) method to improve the identification of dietary patterns, especially in small subpopulations where traditional methods struggle with weakly separated patterns. By leveraging a hierarchical tree structure, the method effectively shares information between similar dietary patterns, leading to more stable and accurate estimations.

What Problem Does This Paper Address?

Dietary patterns are crucial for understanding diet-disease relationships, and Latent Class Models (LCMs) are commonly used to identify them from dietary intake data. However, LCMs often produce dietary patterns that are very similar (weakly separated), leading to numerical instability and making it hard to interpret the results. This problem is even more pronounced when analyzing small subgroups of a population, where data is limited. The authors aim to solve this by developing a method that can robustly identify distinct dietary patterns even when they are subtle and the sample size is small.

Methods

The core of the proposed method is a Tree-Regularized Bayesian Latent Class Analysis (DDT-LCM). This approach builds upon standard LCMs by introducing a Dirichlet Diffusion Tree (DDT) process as a prior. This DDT process defines a hierarchical structure (a tree) over the latent dietary patterns. Patterns that are "closer" on this tree are encouraged to be more similar through a shrinkage prior. This means that similar patterns are "pulled" towards a common ancestral pattern, effectively sharing statistical strength.

Key novelties include:

  • Tree-structured shrinkage: Instead of assuming independence between patterns, the tree structure guides how patterns are regularized.
  • Group-specific diffusion variances: The method accounts for the fact that different food groups (e.g., fruits vs. meats) might show varying degrees of separation between dietary patterns. This allows for flexible shrinkage based on food group characteristics.
  • Bayesian framework: The entire model is estimated within a Bayesian framework, allowing for uncertainty quantification and robust inference, especially crucial for small sample sizes.

The authors also detail a Metropolis-Hastings within-Gibbs sampling algorithm for posterior inference.

Datasets

The primary dataset used for application and evaluation is from the Hispanic Community Health Study/Study of Latinos (HCHS/SOL). Specifically, the authors analyze dietary intake data from a subgroup of 496 U.S. adults of South American ethnic background.

The paper also includes extensive simulation studies using synthetic data to evaluate the performance of their proposed method against various baseline models under different scenarios of class separation and sample sizes.

Key Insights and Ideas

  • Weakly separated patterns are a major challenge: The paper highlights that standard LCMs struggle significantly when dietary patterns are very similar, leading to convergence issues and unreliable results, especially in small datasets.
  • Hierarchy is key for sharing information: By imposing a tree structure on dietary patterns, the method can effectively borrow strength from similar patterns, leading to more stable estimates. This is analogous to how hierarchical models work in other fields.
  • Food group characteristics matter: Recognizing that not all food items behave the same way across dietary patterns is crucial. Allowing for group-specific shrinkage (diffusion variances) improves the model's ability to capture nuanced differences.
  • Small subpopulations need specialized tools: The proposed DDT-LCM is specifically designed to address the limitations of traditional methods when dealing with smaller, potentially more heterogeneous subgroups, which are common in epidemiological studies.

Key Results

  • Improved estimation accuracy: Simulation studies showed that DDT-LCM consistently outperformed standard Bayesian LCMs in recovering true tree structures, estimating dietary patterns (conditional probabilities), and assigning individuals to classes, especially under conditions of weak class separation.
  • Stability in small samples: The method demonstrated robustness and accuracy even with smaller sample sizes (N=100, 200, 400) where traditional methods struggled.
  • Application to HCHS/SOL data: The DDT-LCM successfully identified 6 distinct dietary patterns in the South American subpopulation of the HCHS/SOL study. The analysis revealed varying degrees of variability across different food groups, with fruit items showing the least variation between patterns.
  • Better predictive performance: The proposed method showed better predictive performance compared to a homogeneous variance version of the model, indicating the importance of group-specific diffusion variances.

Limitations and Future Work

  • Measurement error: The authors acknowledge that they did not explicitly model measurement errors in the dietary recall data, which can affect the accuracy of food item reporting. They suggest this as a potential area for future work.
  • Independence assumption within patterns: The current model assumes that food item responses are independent within a dietary pattern. Incorporating nutrition knowledge to model dependencies between food items is suggested as an extension.
  • Cross-sectional data: The HCHS/SOL analysis was cross-sectional. Extending the model to longitudinal data could provide insights into how dietary patterns evolve and impact health outcomes over time.
  • Data availability: Access to the HCHS/SOL data is restricted, limiting further in-depth analysis or replication by other researchers without specific agreements.
  • Computational cost: While the method is designed for small subpopulations, the computational cost of sampling the tree structure can still be significant, though manageable for a moderate number of classes.

Future work could explore incorporating measurement error correction, modeling inter-item dependencies within patterns, and extending the framework to longitudinal dietary data.

Summary updated

Abstract

Dietary patterns synthesize multiple related diet components, which can be used by nutrition researchers to examine diet-disease relationships. Latent class models (LCMs) have been used to derive dietary patterns from dietary intake assessment, where each class profile represents the probabilities of exposure to a set of diet components. However, LCM-derived dietary patterns can exhibit strong similarities, or weak separation, resulting in numerical and inferential instabilities that challenge scientific interpretation. This issue is exacerbated in small-sized subpopulations. To address these issues, we provide a simple solution that empowers LCMs to improve dietary pattern estimation. We develop a tree-regularized Bayesian LCM that shares statistical strength between dietary patterns to make better estimates using limited data. This is achieved via a Dirichlet diffusion tree process that specifies a prior distribution for the unknown tree over classes. Dietary patterns that share proximity to one another in the tree are shrunk towards ancestral dietary patterns a priori, with the degree of shrinkage varying across pre-specified food groups. Using dietary intake data from the Hispanic Community Health Study/Study of Latinos, we apply the proposed approach to a sample of 496 US adults of South American ethnic background to identify and compare dietary patterns.
BibTeX
@article{li2023tree,
  author = {Mengbing Li and Briana Stephenson and Zhenke Wu},
  title = {{Tree-Regularized Bayesian Latent Class Analysis for Improving Weakly Separated Dietary Pattern Subtyping in Small-Sized Subpopulations}},
  journal = {The Annals of Applied Statistics},
  year = {2023},
}
Figure