When we conduct medical research, it is important that we design our studies well, analyze the data correctly, and communicate the results accurately.  To do these things, one inevitably has to address concepts like experimentation, observation, and—beneath it all—causality.  Many have studied causality within computer science, statistics, clinical trials, reinforcement learning, control theory, epidemiology, etc.

Today, some great points on causality were made here. I agree that causality is the default in trials (and online reinforcement learning) and that the causal frameworks, such as do-calculus and directed acyclic graphs, are useful (see my views here).

I would personally advocate, from a notational and conceptual standpoint, for merging directed acyclic graphs with a more policy-centric approach (as I describe here), which aligns with those working in control theory and reinforcement learning, and also, by nature of its focus on probability, with statisticians working in clinical trials. Also, because such a framework involves probabilities, it is a natural fit for clinicians, who become accustomed to probabilities in order to wade through the uncertainty that is medicine.

Policy-centric or not, it is good to see discussion of the concept of causality and the frameworks developed to assess and communicate it.

Taking the best from different frameworks, such as the intuitive visual qualities of directed acyclic graphs, the rigor and transparency of do-calculus and potential outcomes, the focus on study design and statistics of clinical trials, and the general action- focused and policy-focused frameworks from fields like control theory and reinforcement learning, will be paramount to moving forward.

Advances here will ultimately help us, together, make better sense of evidence and better decisions for patients.