Hacker News new | ask | show | jobs
by haberman 1936 days ago
> The catch is that, whether explicitly or implicitly, you must make assumptions in the first place about the directions of causality among the variables.

That right there is the headline to me.

Compare this with the blurb in the dust jacket of the book:

> "Correlation is not causation." This mantra, espoused by scientists for more than a century, led to a virtual prohibition on causal talk. Today, that taboo is dead. The causal revolution, led by AI researcher Judea Pearl and his colleagues, has cut through years of confusion about the nature of knowledge and established the study of causality at the center of scientific inquiry.

This all but claims that these new causal tools have found a solution to the problem of "correlation is not causation." But they have done no such thing: there is no new technique offered here for establishing causation in any better or easier way than the RCT of yore.

If you get rid of the warning "correlation is not causation" and focus everyone's attention on all the exciting inferences you can make when you assume causation, I'm worried that the end result is a lot of bad science.

3 comments

> all the exciting inferences you can make when you assume causation, I'm worried that the end result is a lot of bad science.

But this, is the definition of science: you make models on top of hypothesis that you assume, you see how existing data fits your model, and then you make falsifiable predictions based on your model.

Studying data without any (causal) model of what's happening is just collecting statistically significant trivia. It is research, for sure, but that's not enough to make it science. But hey, at least you published something.

Sure, but what does Pearl bring to the table here? The idea of making falsifiable predictions long predates the "Causal Revolution." The warning of "correlation is not causation" still seems as relevant as ever.
For one it lets you avoid controlling for the wrong variables and causing e.g. spurious correlations by doing so. In fact this is one of the best examples of why a causal model is necessary, because without one you can easily end up with a correlation that doesn’t exist as is illustrated quite nicely in his book.
Do you mean that the new techniques will (1) help you prove which variables should not be controlled for, or that they will (2) help you more clearly describe your causal assumptions, so that you can more easily recognize which variables should not be controlled for according to your assumptions?

If you mean (2), I can't really disagree: explicitly specifying your causal assumptions through a DAG seems like a clarifying step in specifying a model.

If you mean (1), then I must be missing something because I'm not seeing that this set of tools can do that.

My worry is that (2) is mistaken for (1), and that writing down a causal model is conflated with proving that it is true.

For a given causal model it is (1) in my understanding.
But "for a given causal model" precisely means "given a set of statements about what causes what." Those statements must be either proved or assumed.

If they are already proved, they don't need to be proved further per (1).

If they are not already proved, then they are just assumptions and we are talking about (2).

For more on this, you might enjoy this survey of Pearl's work, and its relevance to economics research, by Guido Imbens: https://arxiv.org/abs/1907.07271.

It is long, but through. The conclusion is essentially that Pearl is over-hyped, though my summary does significant violence to the nuances of his argument.

Using «the book of Why», which is a book of popular science, not an academic book, as a reference is a bit troubling though.
A significant number of Pearl's works are cited, not just that one. (See the references section.)
I find it vaguely shocking the false statements that come up in proximity to Pearl's work, like the idea that nobody talked about causality until Pearl. Physicists talk about causality all the time, in relation to the flow of information. Causality has been a central question in econometrics since at least the 50s -- I'm sure other social sciences are the same.