Background/Objectives: Dietary patterns play a crucial role in health, yet most research examines foods individually, overlooking how they interact. This approach provides an incomplete picture of how diet influences health outcomes. Network analysis (e.g., Gaussian Graphical Models, Mutual Information Networks, Mixed Graphical Models) offers a more comprehensive way to study food co-consumption by capturing complex relationships between dietary components. However, while researchers have applied various network algorithms to explore food co-consumption, inconsistencies in methodology, incorrect application of algorithms, and varying results have made interpretation challenging. To address this, the aim of this study was to review existing research and establish guiding principles for future studies. Methods: Using PRISMA-ScR criteria, a scoping review identified 18 relevant studies across different populations and health outcomes. Results: Gaussian Graphical Models were the most frequently used, often paired with regularisation techniques (e.g., graphical lasso) to improve clarity. However, several methodological challenges were identified, including the use of cross-sectional data, which limits the ability to determine cause and effect, reliance on centrality metrics in unbounded networks, and difficulties in handling non-normal data. Only a few studies addressed these limitations, such as using semiparametric extensions of Gaussian Graphical Models to manage non-normal data. Conclusions: To improve the reliability of network analysis in dietary research, this review proposes five guiding principles - model justification, design–question alignment, transparent estimation, cautious metric interpretation, and robust handling of non-normal data - that future studies can adopt. Overall, this review highlights the potential of network analysis to uncover hidden relationships between dietary components and enhance our understanding of how diet influences health. This review was preregistered https://doi.org/10.17605/OSF.IO/R5VE6.
Taylor et al. (2025) studied this question.