Causal Discovery in Heterogeneous and Mixed Data

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 296 papers150 conceptsfrozen 2026-10-04

Causal discovery aims to move beyond simple correlations to uncover the "why" behind data—the directed mechanisms where one variable actually influences another. However, most classical algorithms assume a "clean" world where every variable is of the same type, usually all continuous (Gaussian) or all discrete (multinomial).

In reality, data is messy and heterogeneous. A medical dataset might combine continuous gene expression levels with discrete genetic mutations and categorical patient outcomes. A manufacturing log might mix sensor readings (continuous) with batch IDs (categorical) and error counts (discrete). When we force these diverse data types into a single homogeneous framework, we lose information, introduce bias, and often fail to identify the true causal structure.

The Failure of Homogeneity Assumptions

Traditional causal discovery often relies on two extremes: discretizing everything or assuming everything is continuous. As Chao Li and Shohei Shimizu point out in Combining Linear Non-Gaussian Acyclic Model with Logistic Regression Model for Estimating Causal Structure from Mixed Continuous and Discrete Data, "lossy" discretization destroys the density of continuous measurements, while treating categorical labels as numbers creates mathematically nonsensical relationships.

When we apply a single-graph method to a mixed population, we often end up with an "inaccurate compromise structure" that represents no one. Zachris Björkman and colleagues argue in Incorporating Expert Knowledge into Bayesian Causal Discovery of Mixtures of Directed Acyclic Graphs that effective discovery requires models capable of identifying multiple distinct causal mechanisms simultaneously to avoid averaging out critical differences.

Tailoring Tests to the Variable Pair

One intuitive solution is to stop looking for a "one size fits all" statistical test and instead use a library of tests tailored to the specific variables being compared. Michail Tsagris demonstrates this in Constraint-based causal discovery with mixed data by using Likelihood Ratio Tests (LRT) derived from different regression models:

  • Linear regression for continuous targets.
  • Multinomial logistic regression for categorical targets.
  • Ordered logit models for ordinal data.

By selecting the appropriate model for each pair, researchers can maintain a unified framework for causal discovery without forcing variables into unnatural distributions. Similarly, Andrew J. Sedgewick and Joseph D. Ramsey show in Mixed Graphical Models for Causal Analysis of Multi-modal Variables that estimating an undirected graph first to prune the search space, followed by these specialized tests, significantly improves both speed and precision.

Information as a Universal Currency

If statistical tests are too specific, information theory offers a more universal language. The core idea is that the true causal direction is the one that allows for the "simplest" description of the data.

Alexander Marx and Jilles Vreeken leverage the Minimum Description Length (MDL) principle in algorithms like CLIMB and GLOBE (Discovering Fully Oriented Causal Networks). By treating causal discovery as a data compression problem, they can identify the directed structure that best explains the data across diverse types. This approach is echoed by Sahar Behzadi in the ITGH framework (ITGH: Information-Theoretic Granger Causal Inference on Heterogeneous Data), which uses Generalized Linear Models (GLMs) to handle Poisson, Bernoulli, and Gamma distributions within a single compression-based metric.

The Latent Bridge: Copulas and Ranks

Another way to handle mixed data is to assume that discrete variables are actually "shadows" of hidden continuous variables. This is the logic behind Gaussian Copula models. Ruifei Cui and Tom Heskes use this in Learning the Causal Structure of Copula Models with Latent Variables to bridge the gap between ordinal survey responses and continuous measurements.

By mapping discrete data into a latent continuous space, researchers can use robust rank-based measures—like Spearman’s $\rho$—to perform causal discovery. Naftali Harris showed in PC algorithm for Gaussian copula graphical models that these rank-based methods maintain high-dimensional consistency for a broad class of distributions with almost no computational penalty.

Proving Identifiability in the Wild

A major hurdle in causal discovery is "identifiability"—mathematically proving that we can distinguish $A \to B$ from $B \to A$. In mixed settings, this becomes even more complex.

Recent work has expanded the boundaries of what is identifiable:

These advancements ensure that when we find a causal link between a continuous predictor and a discrete outcome, the direction is grounded in rigorous mathematical proof rather than just statistical correlation.

Papers in this version

Showing 60 of 96 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 71 papers
  2. Version 1 · 2026-10-04 · 96 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_mixed_type_causal_discovery_a043f50c,
  title = {mixed-type causal discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-type-causal-discovery/rdr_a043f50c5a6b453d8c99338504bc04a2},
  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_discovery_in_heterogeneous_and_mi_a043f50c_v1,
  title = {Causal Discovery in Heterogeneous and Mixed Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-in-heterogeneous-and-mixed-data/rdr_a043f50c5a6b453d8c99338504bc04a2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}