Robustness of model predictions under extension

Tineke Blom, Joris M. Mooij

Abstract

Mathematical models of the real world are simplified representations of complex systems. A caveat to using mathematical models is that predicted causal effects and conditional independences may not be robust when the model is extended. This paper investigates the theoretical conditions under which qualitative predictions remain valid under model expansion and demonstrates how discrepancies between a submodel and real-world data can reveal hidden feedback loops.

Introduction

Scientific models are, by definition, simplifications. Researchers often construct a "submodel" of a complex system to derive causal predictions—such as "Variable A causes Variable B"—and subsequently test these hypotheses. However, a significant pitfall in this approach is that such predictions are frequently non-robust. For instance, extending a simple viral infection model by incorporating immune system equations can completely invert the predicted causal relationships. This lack of robustness poses a challenge for drawing broad causal conclusions from simplified models. This research addresses two primary questions: under what mathematical conditions are model predictions preserved during extension, and can the failure of these predictions be leveraged to infer unobserved system components?

The Causal Ordering Framework

To analyze model robustness, the authors employ the Causal Ordering Algorithm, a technique introduced by Herbert Simon in 1953. The methodology follows a structured pipeline:

  1. Bipartite Mapping: The system of equations is represented as a bipartite graph. One set of nodes represents endogenous variables, while the other represents the equations themselves.
  2. Perfect Matching: The algorithm identifies a "perfect matching," uniquely associating each variable with an equation that can solve for it.
  3. Graph Construction: From this matching, the authors derive a Causal Ordering Graph, which illustrates the flow of interventions, and a Markov Ordering Graph, which encodes conditional independences.
The transition from a bipartite graph to causal and Markov ordering graphs.

By comparing the graphs of a submodel against those of an extended model, researchers can determine if causal paths (ancestral relations) or d-separations (independences) are preserved.

Theoretical Insights: Robustness and Self-Regulation

A central finding of the paper is the role of self-regulating variables. A variable is considered self-regulating if it can be solved uniquely from its own derivative's equilibrium equation. The authors prove that if all variables in a system are self-regulating, the predicted presence of causal effects remains robust to model extensions.

Furthermore, the paper posits that discrepancies between model predictions and empirical data should be viewed as diagnostic tools rather than mere failures. If a submodel predicts an independence that is absent in the data, it serves as a mathematical signal that the "hidden" portion of the system contains a feedback loop involving the observed variables. This qualitative approach allows for structural assessment without requiring precise parameter values.

Application to Biological Systems

The framework was evaluated using two classic biological models: a viral infection model and a protein signaling cascade (RAS-RAF-MEK-ERK pathway).

Bipartite graph representation of a complex system, such as the viral infection model.

In the viral model, the authors demonstrate why simple representations often fail: when extended with an immune response, the immune system "takes over" the regulation of the virus, disrupting the original submodel's structure and breaking the perfect matching. In the signaling cascade example, the authors show that while a reductionist submodel predicts that an input signal affects all downstream proteins, a holistic model with feedback predicts independence at equilibrium. Observing this independence in experimental data allows researchers to infer the existence of feedback loops even when the feedback mechanism itself is unobserved.

Conclusion and Future Directions

This research provides formal sufficient conditions (Theorems 1 & 2) for the preservation of causal relations in well-posed models. However, several limitations remain. The current framework focuses on dynamical systems at equilibrium; causal relations may differ significantly during transient phases. Additionally, the theorems rely on the assumption of unique solvability and "faithfulness"—the idea that observed independences stem from system structure rather than coincidental parameter cancellations. Future work will focus on scaling these bipartite graph methods to systems with thousands of variables and exploring non-equilibrium dynamics.

Summary updated

Abstract

Mathematical models of the real world are simplified representations of complex systems. A caveat to using mathematical models is that predicted causal effects and conditional independences may not...
BibTeX
@inproceedings{blom2022robustness,
  author = {Tineke Blom and Joris M. Mooij},
  title = {{Robustness of model predictions under extension}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2022},
  url = {https://openreview.net/forum?id=AaTwe9Vx-f},
}
Figure