Stratified Analysis of 'Probabilities of Causation'

Manabu Kuroki, Zhihong CaiUAI 2006

TL;DR

This paper introduces a novel method for refining bounds on "probabilities of causation" by stratifying data based on observed covariates. This approach yields narrower, more informative bounds than previous methods, offering improved accuracy for public policy, legal determinations, and personal decision-making.

What Problem Does This Paper Address?

In many scientific and practical domains, such as epidemiology and law, it's crucial to quantify the likelihood that one event (an exposure or action) caused another (an outcome or effect). Formal methods, particularly those developed by Pearl and others, define "probabilities of causation" like the probability of necessity (PN) and probability of sufficiency (PS). While these methods provide bounds on these probabilities using experimental and observational data, these bounds can be quite wide. This paper addresses the limitation that existing methods don't fully leverage all available information, specifically covariates (e.g., age, gender) that are often measured alongside the exposure and outcome. The goal is to derive tighter bounds by incorporating this covariate information.

Methods

The core innovation of this paper is the application of stratified analysis to the problem of bounding probabilities of causation. Instead of treating all data as a single pool, the authors propose:

  1. Stratifying the data: Grouping observations based on the levels of measured covariates (S).
  2. Calculating bounds within strata: Deriving the probabilities of causation bounds for each stratum separately.
  3. Summarizing bounds: Aggregating the stratum-specific bounds to obtain overall, narrower bounds for the entire population.

This approach leverages the fact that within a specific stratum defined by covariates, the relationships between exposure, outcome, and counterfactuals might be more constrained, leading to tighter bounds. The authors use a linear programming technique and provide a simpler proof based on consistency and recursive factorizations. They also introduce a new definition for conditional probabilities of causation.

Datasets

The paper uses a real-world dataset for an illustrative example:

  • Northern Alberta Breast Cancer Registry (Newman, 2001): This dataset, originally collected to investigate the effect of receptor level on breast cancer survival, is reanalyzed to demonstrate the application of the proposed bounds. The variables include receptor level (exposure), survival indicator (outcome), and stage at diagnosis (covariate).

The paper also uses simulated data for its simulation experiments to verify the properties of the proposed bounds.

Key Insights and Ideas

  • Covariates are valuable: Even if covariates aren't confounders in the traditional sense (i.e., they don't satisfy the strong ignorability condition on their own), stratifying by them can still lead to narrower bounds on probabilities of causation.
  • Stratification improves precision: By analyzing data within more homogeneous subgroups defined by covariates, we can reduce uncertainty and obtain more precise estimates of causal probabilities.
  • Not always more is better: The paper explores the covariate selection problem, suggesting that adding more covariates doesn't always lead to better (lower variance) estimates. The relationships between covariates and their impact on the exposure and outcome are critical.
  • "No-prevention" assumption simplifies estimation: When a monotonicity assumption (often interpreted as "no-prevention" in epidemiology) is added, point estimators can be derived, further enhancing the utility of the methods.

Key Results

  • The proposed stratified analysis method yields narrower bounds for probabilities of causation (PN and PNS) compared to the existing Tian-Pearl bounds.
  • Simulation experiments show that the asymptotic variances of the proposed bounds are good approximations of the simulated variances.
  • The simulations confirm that using a single covariate (S or T) can lead to smaller variances than using both, or using a less informative covariate, depending on the specific relationships between variables.
  • In the breast cancer example, the proposed bounds for PN were (0.000, 0.778) and for PNS were (0.000, 0.168), which are narrower than the Tian-Pearl bounds (0.000, 1.000 for PN and 0.000, 0.237 for PNS), providing more specific information.

Limitations and Future Work

  • Covariate Selection: While the paper provides insights into covariate selection, it acknowledges that determining the optimal set of covariates for minimizing variance is an ongoing challenge. The paper suggests that it's not always better to use all available covariates.
  • Assumptions: The methods rely on assumptions like the SITA condition for observational data and the consistency assumption.
  • Future Work: The authors suggest that their results can be extended to more complex scenarios, such as continuous variables and more general causal structures. The improved bounds can facilitate more accurate causal effect evaluations in various fields.

Summary updated

BibTeX
@inproceedings{kuroki2006stratified,
  author = {Manabu Kuroki and Zhihong Cai},
  title = {{Stratified Analysis of 'Probabilities of Causation'}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2006},
}
Figure