Causal Discovery from Nonstationary Distribution Shifts

This direction is merged into another direction: non-stationary causal discovery. Its last version stays readable and citable here.

Version 1 of 130 papers43 conceptsfrozen 2026-10-04

In classical statistics, non-stationarity is a nuisance. When the statistical properties of a dataset change over time or across domains—like a stock market regime shift or a sensor drifting—the standard toolkit suggests we should "fix" it by detrending, differencing, or filtering the data.

However, a growing body of research suggests that these shifts are not noise to be removed, but a "blessing" for causal inference. By treating distribution shifts as natural experiments, researchers can distinguish between mere correlation and true causation in ways that are impossible with static, stationary data.

The Intuition: Distribution Shifts as Virtual Interventions

To understand why change is helpful, consider the fundamental problem of causal discovery: if two variables $X$ and $Y$ are correlated, does $X$ cause $Y$, does $Y$ cause $X$, or does a hidden variable $Z$ cause both? In a stationary setting, $X$ and $Y$ are often symmetric, making the direction of influence unidentifiable.

As Kun Zhang, Biwei Huang, and Bernhard Schölkopf argue in Causal Discovery from Nonstationary/Heterogeneous Data, non-stationarity breaks this symmetry. If $X$ causes $Y$, the mechanism that generates $Y$ from $X$ (the conditional distribution $P(Y|X)$) is typically an independent physical process. If the environment changes, the distribution of the cause $P(X)$ might shift, but the mechanism $P(Y|X)$ often remains invariant.

If we observe that $P(X)$ changes while $P(Y|X)$ stays the same, we have strong evidence that $X \to Y$. If we assumed $Y \to X$, we would likely see that both $P(Y)$ and $P(X|Y)$ change simultaneously, which is statistically less likely under the principle of independent mechanisms. This approach, formalized in the CD-NOD framework, transforms distributional shifts into "virtual interventions" that reveal the underlying causal skeleton and orientation.

Invariance as a Causal Fingerprint

This principle of looking for what doesn't change is known as Invariant Causal Prediction (ICP). Niklas Pfister, Peter Bühlmann, and Jonas Peters extended this to time-series in Invariant Causal Prediction for Sequential Data.

The core idea is that a causal relationship should be robust across different environments or time segments. By testing which sets of predictors yield a model that remains valid even as the background distribution shifts, researchers can identify the true causal parents of a variable. This is particularly powerful for sequential data, where non-stationarity allows for the detection of instantaneous causal effects that traditional methods like Granger causality—which relies strictly on time delays—often miss.

Recovering the Hidden Gears of the World

The utility of non-stationarity extends beyond observed variables into Causal Representation Learning (CRL). In CRL, the goal is to recover latent causal factors (like "gravity" or "position") from raw, high-dimensional data (like video pixels).

In Learning Temporally Causal Latent Processes from General Temporal Data, Weiran Yao and colleagues introduced LEAP, a framework that proves latent causal structures are mathematically identifiable if the transition rules change over time. They treat these shifts as "soft interventions." If you train a model on a pendulum swinging on Earth and then on Mars, the change in gravity provides the necessary variability to distinguish the latent variable "gravity" from other factors that might appear identical in a single environment.

Similarly, Philippe Brouillard and Dhanya Sridhar demonstrate in Causal Representation Learning in Temporal Data via Single-Parent Decoding that structured temporal dynamics allow models to uniquely recover underlying causal factors—such as El Niño climate patterns from global pressure maps—without needing manual labels or experimental interventions.

From Theory to Practice: Brains and Climate

The shift from viewing non-stationarity as a problem to viewing it as a tool has led to breakthroughs in complex domains:

  1. Neuroscience: Traditional brain connectivity models struggle with the "drift" inherent in EEG and fMRI data. The INCAMA framework, developed by Sangyoon Bae and others, treats these neural drifts as natural experiments to resolve driver-follower relationships in the brain, as detailed in Recovering Whole-Brain Causal Connectivity under Indirect Observation.
  2. Climate Science: By identifying unobserved "contexts" or regimes (like shifting weather patterns), researchers like Jakob Runge can discover how causal links themselves evolve over time, providing a more granular view of planetary systems in Context-Specific Causal Graph Discovery.

By embracing the "messiness" of a changing world, these methods provide a rigorous way to move from observing what is happening to understanding why it is happening. Non-stationarity, once the enemy of the statistician, has become the primary signal for the causal scientist.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_discovery_from_nonstationary_dist_b13cb646,
  title = {Causal Discovery from Nonstationary Distribution Shifts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-from-nonstationary-distribution-shifts/rdr_b13cb6467c7d4d4389290e10406f319e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_discovery_from_nonstationary_dist_b13cb646_v1,
  title = {Causal Discovery from Nonstationary Distribution Shifts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-from-nonstationary-distribution-shifts/rdr_b13cb6467c7d4d4389290e10406f319e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}