[mycorrhizal]

I think you are overstating Judea Pearl's argument. He doesn't claim it is impossible to answer causal questions (on observational data) without resorting to do-calculus. What he does claim is that to answer causal (interventional or counterfactual) questions you need to define a model and reason on that model in a way that transforms the question into conditional probabilities. Do-calculus provides a framework to help you perform that reasoning, but certainly people have been using things like the backdoor criterion before the invention of do-calculus to answer interventional questions.

He might claim that any other method can be reduced to do-calculus. I'm not sure. I do believe at the core of his argument is need for an explicit model.

[jamp897]

That’s my core understanding too, the model is essential and that not only rubs statisticians the wrong way but also the current luminaries in deep learning get visibly irritated by that too. The irony is that JP was very successful with statistics having invented Bayesian Networks, and now has moved on.

Here he asks a very simple question and look at the body language from the panel: https://www.youtube.com/watch?v=mFYM9j8bGtg&t=50m47s

[nkurz]

Yes, that's a better and more accurate phrasing than I used. The issue is that the correctness of the answer depends on the whether the causal structure of the model matches the underlying generative process. Graphs and do-calculus aren't required for this, but can help to make things clearer. In a later comment on the blog post, Pearl links to a paper that describes one of his "toy" examples: http://ftp.cs.ucla.edu/pub/stat_ser/r400-reprint.pdf. Section 3 of page 584 is the beginning of the example. I'm sad to say that I'm sufficiently amateur at this that I found even the "toy" example to be at the limits of my reasoning ability, but I thought it still illustrated the argument Pearl is making.

[mycorrhizal]

I absolutely agree if your model is wrong any causal inferences you draw will be wrong. However his framework does provide mechanisms for determining if your data is inconsistent with your model (ie certain independences should or should not exist). So hopefully if your model is wrong your data will tell you so and you can change your model.