Imagine a classic physics demonstration: a billiard ball strikes another, causing it to roll across the table. This feels like an objective, undeniable causal event. But what if we zoom out? Was the movement caused by the transfer of kinetic energy, the player’s hand, the player’s underlying intentions, or the friction of the felt?
In data science and structural modeling, deciding where to draw the boundary around a system is known as model specification. A foundational challenge in causal inference is that causal claims are rarely absolute; instead, they are captive to the causal models we construct. Causal conclusions do not emerge purely from raw data. They are highly sensitive to subjective choices about which variables to include, how to structure equations, and how to define the granularity of the system.
The Subjective Boundary of Variable Selection
When building a structural equation model, the first step is choosing the variables. Yet, determining which variables are "natural" or "well-motivated" remains a deeply subjective process. As Joseph Y. Halpern et al. point out, automating the selection of an appropriate causal model that aligns with human intuition is an unresolved challenge. If a modeler constructs a model that is too sparse or fails to capture the correct mechanisms, the resulting causal determinations may be technically valid within that model but practically misleading in the real world (Joseph Halpern et al.).
Why does this happen? Omitting a single significant driver can introduce hidden common causes, leading an algorithm to identify entirely spurious relationships (Giorgia Di Capua et al.). This issue spans domains:
- Autonomous Agents: For an AI agent trying to build a causal world model to navigate its environment, Özgür Şimşek notes that current discovery methods lack a clear criterion for variable inclusion, leaving the agent's understanding of its environment fundamentally incomplete.
- Causal Abstraction: Even when researchers try to simplify complex, low-level models into interpretable high-level ones, the ability to maintain "causal sufficiency" depends entirely on the care with which the initial set of variables was chosen (Riccardo Massidda et al.).
Without a principled approach to selecting these variables, model construction remains an art form governed by modeler judgment rather than formal algorithms (Joseph Halpern et al.).
Modeling Relativism: Shifting Definitions of Reality
This dependency creates a form of modeling relativism, where the very properties of a system change depending on how it is represented.
Consider the definition of an "agent." Zachary Kenton et al. demonstrate that whether a system appears agentic or non-agentic is highly sensitive to the granularity and scope of the chosen environmental variables. A system might look like a goal-directed agent in a coarse-grained model, but dissolve into passive physical reactions when viewed at a finer resolution.
The same relativity applies to human concepts like blame and responsibility. David A. Lagnado et al. show that human attributions of responsibility shift dramatically depending on the specific causal model used to represent a scenario. As Joseph Halpern argues, while structural models allow us to quantify degrees of responsibility and blame, the selection of the model itself remains a subjective source of disagreement. There is no single "true" model; instead, different criteria like sufficiency versus consequentiality are chosen based on subjective psychological heuristics (James D. Park).
The Technical Challenge of Model Selection
If human intuition is subjective, can we rely on data-driven algorithms to select the right model? Algorithmic selection introduces its own layer of model-dependency.
In high-dimensional settings with many potential control variables, researchers have proposed "double-selection" methods using LASSO to intelligently filter variables (Christian Hansen et al.). However, the validity of the subsequent causal inference remains tightly bound to the specific selection algorithm used.
Furthermore, when dealing with unobserved confounders—such as in electronic health records—the accuracy of causal estimates depends heavily on the specific probabilistic architecture of the factor model chosen to infer those latent variables (Linying Zhang et al.). Even structural assumptions, such as choosing between a model that assumes constant treatment effects versus one that allows for heterogeneity, can completely alter a model's conclusions if the structure misaligns with the true data-generating process (Philipp Bach et al.).
[Raw Observational Data]
│
▼
[Subjective Choices] ──► Choice of Factor Model Architecture
│ ──► Inclusion/Exclusion of Proxies & Surrogates
▼ ──► Structural Assumptions (Heterogeneity vs. Constancy)
[Specified Causal Model]
│
▼
[Causal Inference Output] (Highly sensitive to the choices above)
This sensitivity extends to post-intervention scenarios and counterfactual frameworks. For example, in synthetic control methods, predictive accuracy degrades significantly if surrogate or proxy variables fail to capture the true underlying drivers (Jizhou Liu et al.). When moving beyond structural equations to generalized logical frameworks, Joseph Halpern et al. show that causal definitions risk producing counterintuitive results unless there is a principled way to select which specific formulas are held constant.
Bridging the Gap
Treating causal models as objective mirrors of reality leads to dangerous blind spots. In clinical settings, building prediction models without explicitly integrating causal inference can lead to harmful treatment decisions because purely predictive performance does not guarantee causal validity (Nan van Geloven et al.).
To build reliable systems, the field must treat model specification not as a solved preprocessing step, but as an active, sensitive component of the causal inference pipeline itself. This means leveraging structured relationships like secondary variables (Chris J. Oates et al.) and monitoring predictive loss during hyperparameter tuning (Philipp Bach et al.) to ground our subjective choices in empirical reality.
Key Foundational Work
Causes and Explanations: A Structural-Model Approach (2000) — contributed 4 concept elements to this direction.
Causes and Explanations: A Structural-Model Approach --- Part 1: Causes (2001) — contributed 1 concept elements to this direction.