Causal Discovery for Conditionally Stationary Systems

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

Version 1 of 112 papers18 conceptsfrozen 2026-10-04

In classical causal inference, we typically assume the "rules of the game" are fixed. We represent the world as a Directed Acyclic Graph (DAG) where an arrow from $X$ to $Y$ implies a stable mechanism. However, in complex systems—from the global economy to the human body—these rules often flip. A financial policy that stimulates growth during a recession might trigger inflation during a boom; a gene that promotes cell health in youth might drive oncogenesis in old age.

When the underlying causal structure changes over time, standard algorithms fail. They attempt to find a single "average" graph that fits the entire timeline, often resulting in a tangled mess of weak or contradictory relationships. A new wave of research addresses this by modeling systems as conditionally stationary: the causal rules are not constant, but they are fixed given the current state of the system.

The Problem of Causal Averaging

The core intuition behind this research is that non-stationarity is not just noise—it is a structured shift between discrete regimes. As Valentina V. Kuskova and colleagues point out in Function-Valued Causal Influence in Nonlinear Time Series, reducing a relationship to a single importance score hides critical state-dependent behaviors like thresholds and sign changes.

If a researcher applies a static model to a system that switches between two states, the model might see a positive correlation in State A and a negative one in State B. The global average would suggest no relationship at all. By partitioning data into regimes—a technique used by Soufiane Karmouche and Jakob Runge to study Atlantic–Pacific teleconnections—scientists can uncover non-linear dependencies that are otherwise invisible.

Modeling Causal Rules as States

To solve this, researchers have proposed frameworks that treat the causal graph itself as a dynamic variable.

  1. Latent Regime Detection: In Statistical Models Coupling Allows for Complex Local Multivariate Time Series Analysis, Veronica Tozzo and Alessandro Verri couple Hidden Markov Models (HMMs) with Gaussian Graphical Models. Here, the system moves through hidden states, and each state is associated with a unique "precision matrix" (a map of conditional dependencies).
  2. State-Dependent Causal Inference (SDCI): Carles Balsells Rodas, Ruibo Tu, and Hedvig Kjellstrom formalize this in Causal discovery from conditionally stationary time-series. They model time series by learning a set of discrete causal graphs. The active graph is selected dynamically based on whether the system is in a specific regime, allowing for the mapping of shifting dynamics in fields like sports analytics and biology.
  3. Stage Variables: In health data, the "state" is often an observed stage, such as pregnancy weeks or disease severity. Bharat Srikishan and Samantha Kleinberg developed Causal Discovery with Stage Variables (CDSV), which treats these stages as meta-variables that modify the entire causal structure rather than just being nodes within it.

Meta-Causal Dynamics and Interventions

If the causal graph can change, what governs that change? Moritz Willig, Kristian Kersting, and their team introduce the concept of Meta-Causal Models (MCMs) in Systems with Switching Causal Relations: A Meta-Causal Perspective.

In an MCM, the system is a finite-state machine where each state is a causal graph. Transitions between these graphs are triggered by internal variables or external environmental shifts. This perspective is crucial for understanding how interventions work in the real world.

Traditionally, an intervention changes the value of a variable (e.g., taking a drug lowers blood pressure). But as Ranjani Srinivasan and Ilya Shpitser argue in Path Dependent Structural Equation Models, an intervention can be a "structural catalyst." It can fundamentally alter the sequence of future states and their associated causal structures. For example, a surgical intervention doesn't just change a patient's current vitals; it moves them into a new physiological regime where the "rules" of recovery are entirely different from the "rules" of active illness.

Why This Works

This direction works because it respects the modularity of mechanisms. By assuming that causal relationships are stable within a regime, we can "borrow" statistical power from different time points that share the same state.

Instead of asking "Does $X$ cause $Y$?", these models ask "Under what conditions does $X$ cause $Y$?" By answering the second question, researchers can perform more accurate counterfactual reasoning—predicting not just what will happen to a variable, but how an action might flip the entire logic of the system.

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_for_conditionally_stati_a0549c9b,
  title = {Causal Discovery for Conditionally Stationary Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-for-conditionally-stationary-systems/rdr_a0549c9b466448bf81e5971c604b1221},
  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_for_conditionally_stati_a0549c9b_v1,
  title = {Causal Discovery for Conditionally Stationary Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-for-conditionally-stationary-systems/rdr_a0549c9b466448bf81e5971c604b1221/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}