This is addressed in the introduction. See box 4 in the flow-chart (“testable implications”).

“The listening pattern prescribed by the paths of the causal model usually results in observable patterns or dependencies in the data. [...] If the data contradict this implication, then we need to revise our model.”

"These patterns are called "testable implications" because they can be used for testing the model. These are statements like "There is no path connecting D and L," which translates to a statistical statement, "D and L are independent," that is, finding D does not change the likelihood of L."

This says nothing about testing causality, or the direction of causality. If two things are uncorrelated, then there is probably not a causal relationship between them, granted. But this is not a very novel or useful observation.

However if D and L are correlated, the test above says nothing about how to validate whether D caused L, L caused D, both were caused by a third thing (or set of things), or the correlation is just coincidence.

For a book whose entire thesis is "causality is rigorous," I expect a much more rigorous treatment of how to validate causality using more than mere correlation.

Intuitively I might guess that RCT's are the only way of rigorously establishing cause and effect. I would have been very interested if the book had confirmed or denied this intuitive conjecture of mine.

Another comment in this thread claims that you can infer causality without intervention: https://news.ycombinator.com/item?id=18884104 Perhaps this is true?

This is the kind of discussion that I wish the book had focused on. I want to probe at the line between belief and established fact, and understand what we can rigorously say given the evidence we have. I have a strong aversion to reading extended flowery descriptions of big ideas if the speaker has not rigorously shown that the model maps to the real world. Otherwise it's like listening to just-so stories.

> but note that the validity of the inference is conditional on the model being correctly specified

This strikes me as begging the question. The model is exactly what I don't trust unless it is rigorously justified, so anything conditional on the model being correctly specified I also don't trust.

It all feels like a house of cards.

You can get no causality from data alone. You always need additional assumptions.

If you can do an intervention and manipulate a variable as you wish, the assumption of its independence is warranted. A correlation with the outcome indicates a causal path (or you're being [un]lucky). Even in that case a more complex causal model is useful to get better estimates, distinguish direct and mediated effects, etc.

If you have observational data only there is not much that can be done without a causal model. Given a model, the causal effect of one variable on another may be estimated in some cases. But if your model is wrong you may conclude that there is an effect when none exists or deny the existence of a real effect.

"Beginning with simple scenarios, like whether it was rain or a lawn sprinkler that got a sidewalk wet, the authors show how causal thinking solves some of today's hardest problems, including: whether a drug cured an illness; whether an employer has discriminated against some applicants; and whether we can blame a heat wave on global warming."

In all of these "hard problems", it is the model itself that is the most contentious piece, and the most ideological. Some people have a mental model where CO2 produced by humans is causing climate change (which I agree with), and others believe that the changes can be explained by natural fluctuation. These beliefs are undoubtedly influenced by a person's biases. It's not very useful to say "once you have accepted a causal model, you can draw lots of useful inferences." Because the main point of contention is over what is causing what.

I found this statement of yours honestly more useful than anything I read in the book so far: "You can get no causality from data alone. You always need additional assumptions." The downside of this is that different people can make different assumptions, and so this implies that this kind of causal analysis can't mediate disagreements between different groups of people who see the world very differently.

Well, accepting a causal model and drawing lots of useful inferences seems better than drawing lots of misleading inferences because no attention is paid to the model (or being unable to make any inference because it's not obvious how the data can be used).

Even if people may not agree on what is the right model at least this approach makes the model explicit. And in many cases there is no reason for disagreement, but if there is no careful analysis the wrong model may be used by mistake. For example, chapter 8 has an extense discussion of the potential outcomes approach in the context of salary as a function of education and experience.