I think this is a really thoughtful post. I'd encourage us (readers in general) not to trip over the specific illustrations that are being used to show a general point—namely, that when humans attempt to identify cause-effect relationships we're nearly always performing some kind of selection on top of the "raw" data available to us (which is, itself, nearly always a mere sub-set of all the "raw" data that could exist on that subject).

This observation, if we take it seriously, will encourage us (humans in general trying to make sense of the world) to be humble about the conclusions we reach and mildly skeptical about the conclusions others reach, even when those conclusions feel really incontrovertible.

It's not causal explanations you need to avoid, it's junk! And while I'm glad that the article criticises one terrible theory promoted by Scott Alexander, I'm not impressed by their alternative LSD theory, or their many bold, unsubstantiated claims about what affected what (we do not, for instance, know that political assassinations affected civil rights law in both directions).

Causal graphs are only circular if you are sloppy with time! We do not know if LSD in 1968 caused hippies in 1969, but we can rule out that hippies in 1969 caused LSD in 1968. And it may seem trivial, but it's not. That things don't happen before their causes is the main tool we have to break endless debates about what caused what, but both the article writer, Scott A. and the author he reviews seem to thrive a little too well in the world of muddy explanations.

I liked the article - I'm reading a management book that uses system diagrams to describe business processes and it separates resources from rates and qualities. The article muddies that with its diagrams but still gets the point across.

An example from the book are the resources of recurring customers and staff. If you add marketing, you'll gain customers, and those customers will feedback and bring in more customers by word of mouth if your service is high quality (positive feedback), but you'll also lose customers with poor qualify service and that can be negative feedback. These feedbacks aren't always flowing in only one way. If marketing leads to a surge in customers, and you don't hire more staff quality will sink, and staff might leave due to being overworked . This feeds back into having less staff. The book grabs hold of the causality, you have to understand how your actions will affect the entire system and plan for it.

In terms of determining causality, it’s true that major events aren’t ever simply driven by a single cause, but philosophy has studied this to deep depths and that’s where I think the article leaves the discussion too early. The arrow of time is not discussed and neither are necessary sufficient or contributory causes. LSD and Harvard Admissions may be contributory to the events of the 1960s, but not sufficient or necessary to explain assassinations.

What is the difference between causal projection and just not having enough data to determine the structure? Is it just that someone has picked sides and thus tests a slightly different graph?

Also I've noticed that many substack articles are doubled up. I read the article, and then it repeats itself. What's up with that?

How much of our propensity to make causal statements depends on the language we use?

In English, we structure sentences as either a question, answer, or command of causality. We have "independent clauses" but the moment a clause gains utility is when dependence is added to it.

The only methods we have to express noncausal ideas are additive. Like an expansion pack for the English language, we just slap on some extra grammar and jargon, then call it good. It's a lot of messy boilerplate that we have to use every time we need abstraction. It's no wonder that we have ended up with doctor-speak and programmer jargon.

This has me wondering: what if we did have first-class noncausal language? What would that even look like? How would we structure clauses? Is there even a way to do it, or is what I am proposing simply what one finds searching the antonym of causal itself: redundant, frivolous, irrelevant, or meaningless?

Made me think of Markov chains, and the PageRank algorithm. We should construct causal graphs, then give each of them 1/n weight and iterate over the casual weight splitting out, until the Markov chain reaches a steady state. Then you know what is important causally and what isnt.

"Causal Explanations Considered Harmful" seems to clearly want to invoke Dijkstra's "GOTO statements considered harmful", which played an important role in changing programming style. But do we really want to see causal explanations to go the way of the "goto"? Good luck getting your car fixed, and god help you if you are ill. I'm more comfortable with the essay "superficial analyses considered harmful".

I view the article as aspirational, and a lot of the comments here seem to fetishize a trinity of time and certitude and causality.

I appreciate attempts to explore diagramming, and so I appreciate the article. I am also informed by past exposure to Markov chains.

I am aware of Prigogine, and so I accept that reasoning about a dynamic system relies on its dynamic state. A simple example would be:

  opening the gates causes water to be released from the dam

But that's predicated on there being water captured behind the dam. Which requires rainfall. It also requires a dam.

A different example would be:

  adding additional propellant makes it go faster

This is reasonably straightforward construction for a projectile, less so for a jet engine or a piston engine which dynamically reaches the functional state which we wish to reason about.

As for LSD, how did the author miss Tim Leary? Referring to a Leary personality diagnosis radar chart I can't find an obvious mapping in the dynamic of post vs comments. Best WAG is that neither is expressing dominance (both are seeking order and guidance), with the article more in the region of acceptance and the comments more in the range of suspicion and hostility.

About David Brooks explanations:

https://statmodeling.stat.columbia.edu/2015/06/16/the-david-...

This is why we do randomized experiments, just FYI.

Obviously, there are many situations where we can't do them (e.g., randomizing people to smoking / not smoking). In those cases, the challenges the article discusses become relevant.

The reason why we know having no father figure leads to a higher chance of committing violent crime is not some kind of guesswork of causality, it is because it is the best predictor.

Dishonesty is considered harmful.

I'll add that causal relationships should be multi-dimensional: the Vietnam vet and LSD-fueled hippy having their own truth can be modeled as subgroups that add arrows in their dimensions to the graph.

The rabbit hole goes deeper when you introduce the concepts of blame and responsibility. People all too often fully assign blame to one piece of the giant network of causes that result in an outcome.

Another "considered harmful" post. Please stop.

Ok, sure, when David Brooks, or the Freakonomics guys, or the Guns, Germs, Steel guy, claim they've nailed the causality of something, that's often facile gibberish.

All this does nothing to undercut the validity of causality itself. I hope that's obvious, despite the clickbait title of the OP.

Aren't "Considered Harmful" statements themselves a causal explanation?