Monotonicity and Stability in Causal Model Refinement

12 concepts6 papers

When modeling a complex system, we rarely simulate every atom. Instead, we choose a level of granularity that balances accuracy and simplicity. A coarse-grained model is computationally efficient but omits details; a fine-grained model captures the nuances but quickly becomes unwieldy.

Ideally, moving between these levels of abstraction should be seamless. However, without strict mathematical constraints, changing a model's granularity causes structural breakdowns. Causal claims can flip-flop indefinitely as detail is added, and searching for the "right" level of abstraction can become computationally impossible.

To solve this, a shared research direction focuses on constraining the relationships between models at different scales. By enforcing properties like monotonicity and structural normality, we can guarantee that causal models remain both stable and computationally tractable as they are refined.


The Pathology of Unconstrained Refinement

Why do causal models break when we change their resolution? As Halpern et al. note, causal definitions are highly sensitive to the richness of the underlying structural model. If a model is too abstract and omits intermediate mechanisms, formal definitions can fail to distinguish between genuine causes and incidental events.

However, blindly adding detail makes things worse. As explored by Halpern et al. and Halpern et al., unconstrained model refinement leads to an infinite toggle:

[Coarse Model]  ---> Event A is a CAUSE of B
       ↓ (Add variable X)
[Medium Model]  ---> Event A is NOT A CAUSE of B
       ↓ (Add variable Y)
[Fine Model]    ---> Event A is a CAUSE of B ... (and so on)

Without constraints, an event can alternate between being classified as a cause and a non-cause indefinitely as new variables are introduced. This instability undermines the reliability of any causal explanation.

A parallel issue arises in computation. If we want an automated system to find the simplest adequate model of a physical device, the search space across arbitrary abstractions is massive. Nayak demonstrates that finding the simplest explanation for a system's behavior is inherently NP-hard. Similarly, when optimizing causal decisions across complex networks, Huang et al. show that navigating internal functional dependencies pushes the computational complexity to an intractable $NP^{PP}$-complete.


Taming the Hierarchy via Monotonicity

To resolve the computational bottleneck, we must restrict how simple and complex models relate to one another. The key insight from Nayak and Nayak et al. is the concept of a causal approximation.

Under this framework, model simplifications are constrained such that a simpler model's effects must be a strict subset of a more complex model's effects. This introduces a monotonic relationship between model complexity and explanatory power:

$$\text{Effects}(\text{Model}_{\text{Simple}}) \subseteq \text{Effects}(\text{Model}_{\text{Complex}})$$

Because of this subset relationship, adding detail can only refine or add to existing causal links, never remove them. This monotonicity guarantees that if a simple model fails to explain a phenomenon, any even simpler model is guaranteed to fail as well. By pruning the search space this way, the task of finding the optimal model collapses from an NP-hard problem into a polynomial-time task, enabling efficient, greedy-style search algorithms.


Stabilizing Claims with Normality

While causal approximations ensure monotonicity in physical system behaviors, we still need a way to prevent the semantic flip-flopping of actual causality. To achieve this, Halpern et al. introduce conservative extensions—a formal way to expand a causal model into a more detailed version without altering the truth values of the original variables under counterfactual interventions.

To fully halt the infinite toggle, the authors integrate considerations of "normality" (typical system behavior). When new variables are added to a model, they are constrained to follow established structural equations in the actual, real-world context.

Under this condition of normality, causal judgments stabilize. Crucially, once a factor is identified as a non-cause in a simpler model, adding further structural details will never retroactively turn it into a cause.


Simplification via Exogenous Selection

A final, complementary strategy for managing the complexity of causal hierarchies is to restrict where in the model structure we intervene.

When selecting target units for optimal treatment, calculating effects through layers of endogenous variables is what triggers the $NP^{PP}$-complete complexity. Huang et al. prove that if the selection of target variables is restricted strictly to the exogenous (root) nodes of a causal model, the complexity drops to $NP$-complete.

[Exogenous Roots]   <-- Restricting selection here bypasses internal complexity
      ↓      ↓
[Endogenous Layer]  
      ↓      ↓
   [Outcome]

Because exogenous variables are independent of the model's internal functional dependencies, targeting these inherent, non-descendant traits simplifies the underlying search space. Just like causal approximations and conservative extensions, this constraint leverages the structural hierarchy of the model to make causal inference practically executable.

Key Foundational Work

Appropriate Causal Models and Stability of Causation (2014) — contributed 4 concept elements to this direction.
Causal Approximations (1992) — contributed 3 concept elements to this direction.
Automated Model Selection Using Context-Dependent Behaviors (1992) — contributed 2 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_approximation_hierarchy_a7153d98,
  title = {causal approximation hierarchy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-approximation-hierarchy/rdr_a7153d983cea4105a8669d7d4b15201e},
  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_approximation_hierarchy_a7153d98_v1,
  title = {causal approximation hierarchy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-approximation-hierarchy/rdr_a7153d983cea4105a8669d7d4b15201e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake