LaMe: a Joint Latent Mediation Framework for High-Dimensional Systems

Gaël Poux-Médard, Kaniav Kamary, Yohann De Castro, Vivian ViallonarXiv 2026 · arXiv

TL;DR

LaMe (Latent Mediation) is a new joint probabilistic framework designed to identify how exposures influence outcomes through thousands of potential intermediate variables (mediators). By compressing high-dimensional data into a few latent "signatures," LaMe provides a scalable, robust, and interpretable way to estimate direct and indirect effects without the biases common in traditional multi-step methods.

What Problem Does This Paper Address?

In epidemiology and social science, researchers often want to know not just if a cause (exposure) leads to an effect (outcome), but how. This is mediation analysis. For example, does a specific diet (exposure) reduce heart disease (outcome) directly, or is the effect "mediated" through changes in blood cholesterol (mediator)?

Modern datasets, especially in "omics" (genomics, proteomics, etc.), present a massive challenge: we might have thousands of potential mediators. Traditional methods struggle here. Simple regression models overfit the noise. "Stepwise" approaches—where you first find patterns in the mediators and then run a regression—often fail because they might identify patterns that are irrelevant to the actual outcome. Existing joint models often assume that mediators are independent of each other once you account for the latent state, an assumption that falls apart with highly correlated biological data.

Methods

The authors introduce LaMe, a joint linear-Gaussian framework. Instead of treating every single biomarker as an independent mediator, LaMe assumes the system is driven by a small number of latent variables ($S$), which act as an information bottleneck.

Schematic comparing non-latent (left) and latent (middle) mediation systems, alongside various inference strategies (right).

Key technical innovations include:

  • Joint Probabilistic Modeling: Unlike stepwise methods, LaMe estimates all parameters simultaneously. This ensures the latent mediators are actually relevant to both the exposure and the outcome.
  • Exact Recoverability Assumption: The authors prove that as the number of manifest variables (the things we measure) grows, the latent mediators can be exactly recovered as a linear combination of those measurements. This allows for the statistical identification of "Natural Direct Effects" (NDE) and "Natural Indirect Effects" (NIE).
  • Computational Efficiency: The model can be factorized in two equivalent ways (LaMe-vX and LaMe-vM). This allows researchers to choose the version that is computationally faster based on whether they have more exposures or more measured mediators.
  • Non-Diagonal Covariance: Unlike many predecessors, LaMe doesn't assume that the "noise" in measurements is independent, making it much better suited for correlated biological data.

Datasets

The framework was evaluated using both synthetic simulations and four diverse real-world datasets:

  1. NHANES: Linking lifestyle habits to glycohemoglobin levels via 25 blood biomarkers.
  2. EPIC-CAN: Investigating how socio-demographics influence the age of cancer events through 379 proteomic features.
  3. CRIME: Analyzing how population characteristics affect crime rates through city-level contextual factors.
  4. PISA-I: Exploring the link between home environment and student math proficiency mediated by school environment factors.

Key Insights and Ideas

  • Outcome-Aware Compression: Traditional dimensionality reduction (like PCA) finds the "loudest" patterns in the mediators. LaMe finds the patterns that are causally relevant to the outcome.
  • The Power of High Dimensions: Counter-intuitively, having more manifest variables makes the latent state easier to recover. The authors show that the error in recovering the latent mediator $S$ drops as $O(J^{-1})$, where $J$ is the number of measured variables.
  • Identifiability via Bottleneck: By restricting the rank of the cross-covariance between exposures and mediators, LaMe effectively filters out spurious associations that plague standard high-dimensional regressions.

Key Results

  • Superior Parameter Recovery: In simulations, LaMe consistently outperformed baselines (like OLS, PCA+LASSO, and PLS) in retrieving the true marginal associations between variables, especially in "latent" data generation scenarios.
  • Compact Representations: In real-world applications, LaMe often achieved its best predictive performance using just a single latent dimension ($K=1$), whereas other methods required dozens of components to reach similar accuracy.
  • Scalability: The EM-based inference scales efficiently. As shown in Figure 6, the "vM" factorization is particularly effective when the number of mediators ($J$) is much larger than the number of exposures ($I$), showing a significant speed advantage over standard joint models.
Scalability of various models as a function of the number of observations (left) and the number of mediators (right).

Limitations and Future Work

While powerful, LaMe currently operates under a linear-Gaussian assumption. This means it might struggle with highly non-linear biological processes or categorical data (like "smoker" vs. "non-smoker" exposures).

The authors suggest several promising directions for future work:

  • Extending the framework to handle mixed data types (e.g., categorical exposures).
  • Incorporating non-linear relationships through kernels or neural networks.
  • Adapting the model for survival analysis with censored data, which is critical for clinical trials and long-term health studies.

Summary updated

Abstract

Mediation analysis seeks to understand how an exposure influences an outcome through intermediate variables. However, as increasingly rich datasets become available, hundreds to thousands of potential mediators make these systems difficult to model and hard to interpret. Latent-variable approaches offer a natural way to accommodate these high-dimensional settings. However, existing latent mediation methods face limitations, including overparameterization, biased inference, or limited scalability. We develop LaMe, a joint linear-Gaussian framework that summarizes a high-dimensional exposure-mediator-outcome system through a small number of latent variables. We identify natural direct and indirect effects under an explicit assumption that the manifest measurements fully capture the latent mediators, and characterize conditions under which the latent representation effectively restricts the observed cross-covariance structure. Simulations assess recovery of observed association parameters, natural effects, and manifest features associated with latent mediation pathways across a range of challenging settings. Four real-world applications demonstrate competitive predictive performance with compact latent representations, including single-dimensional representations of biomarker and proteomic measurements. Together, these results establish LaMe as a practical framework for estimating and summarizing multivariate mediation systems under explicit structural and measurement assumptions.
BibTeX
@misc{pouxmedard2026lame,
  author = {Gaël Poux-Médard and Kaniav Kamary and Yohann De Castro and Vivian Viallon},
  title = {{LaMe: a Joint Latent Mediation Framework for High-Dimensional Systems}},
  year = {2026},
  eprint = {2611.18180},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18180},
}
Figure