Causal modelling without counterfactuals and individualised effects

Overview

• Presents a new approach to causal modeling that does not rely on counterfactuals or individualized effects.
• Focuses on the distinction between causal and associational relationships in observational data.
• Introduces the concept of "causal orderings" to capture the directionality of causal effects.
• Demonstrates how this framework can be used to identify causal effects without making strong assumptions.

Plain English Explanation

This paper introduces a new way of thinking about causal modeling that doesn't rely on the concept of counterfactuals or individualized effects. The key idea is to focus on the distinction between causal and associational relationships in observational data.

Instead of trying to estimate how an individual would have responded under a different treatment, the authors propose the idea of "causal orderings" to capture the directionality of causal effects. This allows them to identify causal effects without making the same strong assumptions required by traditional approaches.

For example, let's say we want to understand the relationship between smoking and lung cancer. The traditional approach would involve estimating how an individual's risk of lung cancer would change if they had or hadn't smoked. In contrast, the new framework focuses on establishing the causal ordering - that smoking precedes and influences the development of lung cancer, rather than the other way around.

By focusing on these causal orderings, the authors demonstrate how we can draw insights about causal effects without relying on counterfactuals or making assumptions about individualized responses to treatment. This can be particularly useful in settings where those traditional assumptions don't hold, such as when there are hidden confounding factors or complex, time-varying treatments.

Technical Explanation

The paper introduces a novel approach to causal modeling that departs from the traditional Rubin Causal Model (RCM) framework, which relies on counterfactuals and individualized causal effects.

Instead, the authors propose a focus on causal orderings - the directionality of causal relationships - as the key to identifying causal effects from observational data. This allows them to sidestep the strong assumptions required by the RCM, such as the stable unit treatment value assumption (SUTVA) and the requirement of well-defined counterfactuals.

The core of their approach is to model the association structure between variables, rather than trying to estimate individualized causal effects. They use this association structure to infer the causal ordering of the variables, which then allows them to identify the direction and strength of causal effects.

This framework is demonstrated on several complex treatment scenarios, including time-varying treatments and hidden confounding. The authors show how their approach can still identify causal effects in these challenging settings, where traditional counterfactual-based methods may struggle.

Critical Analysis

The authors make a compelling case for their alternative approach to causal modeling, which avoids some of the strong assumptions required by the Rubin Causal Model. By focusing on causal orderings rather than counterfactuals, they are able to identify causal effects in a wider range of scenarios, including those with time-varying treatments and hidden confounding.

However, one potential limitation of this approach is that it may not provide the same level of individualized, granular insights as the RCM framework. The authors acknowledge that their method does not aim to estimate individualized causal effects, which could be a drawback in certain applications.

Additionally, the reliance on association structure to infer causal orderings may be vulnerable to model misspecification or the presence of unobserved common causes. The authors discuss strategies to address these challenges, but further research may be needed to fully understand the limitations and robustness of their approach.

Overall, this paper presents a novel and promising alternative to traditional causal modeling that could be particularly valuable in complex, real-world scenarios. By shifting the focus to causal orderings, the authors have developed a framework that is more flexible and accessible than the RCM, while still maintaining the ability to identify important causal relationships.

Conclusion

This paper introduces a new approach to causal modeling that moves away from the traditional reliance on counterfactuals and individualized causal effects. By focusing on causal orderings rather than counterfactuals, the authors demonstrate how we can identify causal relationships in observational data without making the strong assumptions required by the Rubin Causal Model.

This framework has the potential to be particularly useful in complex, real-world scenarios, such as those involving time-varying treatments or hidden confounding. While it may not provide the same level of individualized insights as the RCM, the authors' approach offers a more flexible and accessible way to draw causal inferences from observational data.

As the field of causal modeling continues to evolve, this paper represents an important contribution by challenging the traditional assumptions and offering a novel perspective on how we can understand the causal structure of complex systems.