Probabilities of Causation in Nondeterministic Causal Models

David S. Watson, Sander BeckersarXiv 2026 · arXiv

This paper explores the "top rung" of the causal ladder—counterfactual reasoning—within a framework that doesn't assume the world is fundamentally deterministic. By relaxing the standard assumption that all uncertainty stems from unobserved factors, the authors provide a more flexible and often more informative way to calculate the probabilities of causation.

TL;DR

The authors develop a framework for bounding probabilities of causation (like the probability that a treatment was necessary for a patient's recovery) in models where mechanisms are intrinsically stochastic. By introducing "soft" versions of classical causal assumptions, they derive sharper bounds that are more practical for real-world applications like personalized medicine.

What Problem Does This Paper Address?

In classical causal inference, we often use Structural Causal Models (SCMs) to answer counterfactual questions: "Would this patient have recovered if they hadn't taken the drug?" Standard SCMs are deterministic: they assume that if you knew every single hidden factor ($U$), the outcome ($Y$) would be a fixed function of its parents.

However, this "Pearlian" determinism is a strong philosophical and mathematical commitment. In many fields, from biology to social science, it is more natural to think of mechanisms as intrinsically random. Furthermore, when we have latent confounding (unobserved variables affecting both treatment and outcome), counterfactuals are not "point-identifiable"—we can only calculate a range (bounds) for them.

Existing bounds for these probabilities are typically built for deterministic models. This paper addresses the gap by asking: How do we bound probabilities of causation when mechanisms are irreducibly stochastic, and can this nondeterministic view actually give us more useful answers?

Methods

The authors work with Probabilistic Nondeterministic Structural Causal Models (PNSCMs). Unlike standard models, these allow for stochastic outcomes even when all latent and observed variables are fixed.

To find the bounds, they propose a Generic Identification Program:

  1. Reparameterization: They break the model into three channels of latent influence: treatment propensity ($A$), and two stochastic response variables ($B$ and $C$) representing the potential outcomes.
  2. Constrained Optimization: They frame the search for bounds as a generalized nonlinear program. They optimize over probability measures to find the minimum and maximum possible values for counterfactual quantities, subject to constraints from observational and interventional data.
Modified DAG showing latent nodes A, B, and C

The authors introduce three key structural assumptions tailored for this stochastic world:

  • Margin Condition: Assumes mechanisms are "intrinsically stochastic" and never reach absolute 0 or 1 probability.
  • Soft Exogeneity: A quantitative relaxation of the "no confounding" rule, limiting how much unobserved factors can influence treatment assignment.
  • Stochastic Monotonicity: A weaker version of the classical assumption that treatment "only helps." It assumes treatment cannot reduce the probability of a positive outcome at any latent state, but it doesn't require a deterministic guarantee of success.

Datasets

The authors evaluated their approach using:

  • Adult Dataset: A semi-synthetic setup looking at the effect of education on income, treating age as a latent confounder.
  • JOBS II: Data from a randomized trial for a job-search program. They used baseline depression as a latent confounder to test their "soft exogeneity" bounds.
  • Synthetic Benchmarks: A suite of 18 different five-state PNSCMs designed to test the sensitivity of the bounds across various levels of confounding and stochasticity.

Key Insights and Ideas

  • Determinism is a Choice, Not a Necessity: You don't need to invent infinite hidden variables to explain probability; the mechanism itself can simply be random.
  • Stochasticity as Information: Counter-intuitively, assuming that a system is intrinsically stochastic (the margin condition) can lead to narrower, more informative bounds than the standard assumption-free deterministic case.
  • The Power of "Soft" Assumptions: Classical causal assumptions are often "all-or-nothing" (e.g., either there is zero confounding or the bounds are huge). The authors' "soft" parameters ($\epsilon, \delta$) allow researchers to perform sensitivity analyses, seeing how bounds change as assumptions are gradually relaxed.

Key Results

The paper demonstrates that these new bounds are significantly more useful for decision-making:

  • Significant Tightening: On the Adult dataset, the margin condition reduced the width of counterfactual intervals by roughly 80% compared to standard assumption-free bounds.
  • Breaking Deadlocks: In a case study on personalized treatment (deciding whether to treat based on the probability of benefit vs. harm), the "stochastic monotonicity" assumption lowered the upper bound on potential harm enough to justify a "treat" decision that was previously "undecided" under standard models.
Table 1: Sharp bounds on the probabilities of benefit and harm in the example, the resulting decision under rule (1) with λ = 2, and the largest harm aversion λ∗for which treatment is justified.

Limitations and Future Work

  • Parameter Sensitivity: The bounds depend on the choice of $\epsilon$ (margin) and $\delta$ (exogeneity). If these are misspecified, the resulting bounds might not contain the true value. The authors suggest using these as tools for sensitivity analysis rather than assuming "correct" values.
  • Scalability: The current work focuses on binary treatments and outcomes in simple two-node graphs.
  • Future Directions: The authors aim to extend this to multi-valued variables, situations where interventional data is unavailable, and more complex structures like Markov Decision Processes (MDPs) for sequential decision-making.

Summary updated

Abstract

We study probabilities of causation in nondeterministic causal models, where mechanisms may remain intrinsically stochastic even after all latent variables are fixed. Focusing on binary treatments and outcomes, we develop a generic identification program that represents observational and interventional data as moment constraints on latent stochastic mechanism types and computes minimal/maximal values of counterfactual quantities via constrained optimization. Within this framework, we derive sharp analytic bounds on the probabilities of causation under several assumptions tailored to nondeterministic models. Resulting identification regions are often more informative than those attainable under deterministic alternatives. We illustrate the practical relevance of these results in a case study on personalized treatment decisions and experiments on semi-synthetic datasets. Our results clarify how classical bounds relate to nondeterministic causal semantics, generalizing previous works and providing a unified route from qualitative structural restrictions to computable partial identification regions.
BibTeX
@misc{watson2026probabilities,
  author = {David S. Watson and Sander Beckers},
  title = {{Probabilities of Causation in Nondeterministic Causal Models}},
  year = {2026},
  eprint = {2611.18245},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18245},
}
Figure