Causal Discovery in Heterogeneous Populations

18 concepts14 papers1 proposal

Traditional causal discovery algorithms operate under a comforting assumption: all data points in a dataset are independent and identically distributed (i.i.d.) from a single underlying Directed Acyclic Graph (DAG). But real-world data is rarely uniform. In medicine, economics, or neuroscience, a population is typically a mixture of distinct subgroups—such as patients who respond to a treatment versus those who do not, or individuals with varying genetic backgrounds.

When we apply standard causal discovery methods to such heterogeneous data, the algorithms do not just degrade gracefully; they fail catastrophically. Forcing a single graph onto mixed data yields inaccurate "compromise structures" that fail to represent any real subpopulation, as noted by Björkman et al.. Even worse, Partovi Nia et al. demonstrate that aggregating data from diverse subgroups can completely reverse the inferred causal directions, leading to fundamentally incorrect conclusions.

Why Aggregation Distorts Causality: Emergent Edges

To understand why population heterogeneity breaks traditional causal inference, consider what happens statistically when you mix two different causal mechanisms. Suppose that in Subpopulation A, $X$ causes $Y$ ($X \rightarrow Y$). In Subpopulation B, $X$ and $Y$ are completely independent, but both have significantly higher baseline means due to an unobserved environmental factor. If you pool the data from both groups together, a strong statistical dependency between $X$ and $Y$ will appear purely because of the group-level differences.

Varıcı et al. formalize this phenomenon as emergent edges: statistical dependencies that do not exist in any single underlying causal model but materialize purely as an artifact of mixing. When a standard algorithm encounters these emergent edges, it misinterprets them as direct causal links. The fundamental challenge of causal discovery in heterogeneous populations is to separate the true, subgroup-specific causal mechanisms from these deceptive mixture artifacts.

Disentangling Mixtures via Joint Clustering and Structure Learning

The most direct way to solve this "causal cocktail" problem is to learn the subpopulation labels and their respective causal graphs simultaneously. If we can cluster the data points by their underlying data-generating mechanisms, we can infer the correct DAG for each group.

One approach, introduced by Hu et al., leverages a mixture of Additive Noise Models (ANMs). Because the true causal direction in an ANM leaves a distinct footprint in the independence of the residuals, the authors show that one can alternate between assigning data points to clusters and testing causal directions within those clusters.

Expanding on this setup, Mameche et al. developed Causal Mixture Models (CMM), which jointly discover the causal graph and identify multiple independent latent factors that define the subpopulations. Instead of treating heterogeneity as noise to be averaged out, CMM treats the latent subpopulation assignment as an explicit variable within the broader causal system.

What if we do not know how many subgroups exist? Saeed et al. proved that even when group labels are completely unknown, standard constraint-based algorithms applied to the mixed distribution will still recover a "union" of the underlying causal structures. Crucially, the authors show that it is possible to mathematically isolate exactly which variables differ across the mixture components, providing a diagnostic roadmap for where the population splits.

From Discrete Mixtures to Continuous Personalization

Not all heterogeneity can be neatly carved into discrete clusters. Often, causal mechanisms vary continuously from person to person based on individual characteristics, or "moderators."

To capture this fluid variation, Huang et al. propose a framework that splits causal mechanisms into shared templates and specific individual variations. This allows a model to leverage the statistical power of the entire dataset while still tailoring the causal links to individual subjects.

Taking this a step further into classical statistical frameworks, Xie et al. introduced DAG-MM, which incorporates mixed-effects structural equation models. By modeling causal effects as a combination of fixed effects (governed by observed covariates) and random effects (capturing unobserved individual heterogeneity), they allow every single subject to have a unique, personalized DAG.

Similarly, Watson et al. developed Heterogeneous Causal Graphs (HCGs). In an HCG, the presence or strength of a causal edge is explicitly conditioned on individual-level moderators. By introducing a debiasing step into score-based structural learning, the authors show how to map personalized causal paths, enabling precise estimation of individual-level treatment effects even through complex, multi-stage mediating pathways.

Key Foundational Work

Causal Inference and Mechanism Clustering of A Mixture of Additive Noise Models (2018) — contributed 1 concept elements to this direction.
Specific and Shared Causal Relation Modeling and Mechanism-Based Clustering (2019) — contributed 1 concept elements to this direction.
Learning Subject-Specific Directed Acyclic Graphs With Mixed Effects Structural Equation Models From Observational Data (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_heterogeneity_f914397c,
  title = {causal heterogeneity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-heterogeneity/rdr_f914397cc00c4b51a148069b56d1d5a7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_heterogeneity_f914397c_v1,
  title = {causal heterogeneity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-heterogeneity/rdr_f914397cc00c4b51a148069b56d1d5a7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake