Does X cause Y? An in-depth evidence review (2021)

EA types spend a lot of time talking with subject matter experts, see e.g. https://www.givewell.org/international/technical/programs/vi...

[0] edit: Hansen was awarded the Nobel Memorial Prize in Economics in 2013 for GMM, not that that means it can't fail, but clearly a lot of people have found it useful.

E.g. I certainly found myself agreeing with his points about observational studies, and there are plenty of real-world examples you can point to where experts have been lead astray by these kinds of studies (e.g. alcohol consumption recommendations, egg/cholesterol recommendations, etc.)

But when he talked about his reservations re "the wheat" studies, they seemed really weak to me and semi-bizarre:

1. Regarding "The paper doesn't make it easy to replicate its analysis." I mean, no shit Sherlock? The whole point is that it would be prohibitively expensive or unethical to carry out these real experiments, so we rely on these "natural" experiments to reach better conclusions.

2. "There was other weird stuff going on (e.g., changes in census data collection methods), during the strange historical event, so it's a little hard to generalize." First, this seems kind of hand-wavy (not all natural experiments have this issue), but second and more importantly, of course it's hard to "generalize" these kinds of experiments because their value in the first place is that they're trying to tease out one specific variable at a specific point in time.

3. The third bullet point just seemed like it could be summarized as "news flash, academics like to endlessly argue about shit."

I think the fundamental problem when looking for "does X cause Y", is that in the real world these are complex systems: lots of other things cause Y too (or can reduce its chances), so you're only ever able to make some statistical statement, e.g. X makes Y Z% more likely, on average. But even then, suppose there is some thing that could make Y Z% more likely among some specific sub-population, but make it some percent less likely in another sub-population (not an exact analogy but my understanding is that most people don't really need to worry about cholesterol in eggs, but a sub-population of people is very reactive to dietary cholesterol).

Basically, it feels like the author is looking for some definitive, unambiguous "does X cause Y", but that's not really how complex systems work.

While this is true, putting the onus on the reader to understand a lot of advanced math makes it easy to avoid scrutiny by increasing the complexity of your math such that the only people who are ever going to be able to critique you are the intersection between PhD-level mathematicians and experts in the field your paper actually pertains to. Anyone can just say whatever they want and assure you that they must be right because they know more math than everyone else who's interested in that problem.

Instead of, "Understand the math before you conclude that it's all a load of crap," I would say, if it's an unreasonable level of complexity for the particular problem and you can't find a large body of other papers doing something similar with the same problem, just ignore it.

So then you are hands in the air, trying to get others internalize, a discussion culture where you actually can as normal thing claim or are by default assumed to be probably superficial.

Even if someone knows lots about foundational things like philosophy & what they want etc., statements have a certain feel to it regardless of mathematical truthness, and at some point you will do the value benefit calculation that you might be superficial but you should learn about others topics first because you know you can for less time gain more information that is valuable from them.

---

Pearl's causality as far as I see is best modelled with cdr car + NBG as embedded agency foundations + computability in the mix + signal theory (time as discrete evolution from one state in the chain to another; signal theory is relevant when you have multiple agents or the environment & PARTS of you run at different clocks or something more complex) i.e. part of formal embedded agency. Doesn't feel too meaningful without it except epistemologically (close what can we know type questions) where causality might be a good lens, especially to filter inquiries.

One of the first time I got interested in reading medical studies was when I saw a bunch of headlines announcing that a randomized controlled trial had proved that echinacea was ineffective for treating respiratory problems. This surprised me, because I'd always been a dogmatic drinker of echinacea tea whenever I had a cold, and had thought that it helped. But then again, I come from a culture of damn dirty hippies, so I was open to being wrong about it. Rather than rely on the headlines, I decided to dig up the study itself.

Here's what the study actually found: that rubbing an echinacea-infused ointment on your wrists has no effect on respiratory health.

Er... yeah, no shit, Sherlock. Literally nobody uses echinacea that way. You've just falsified a total straw-man of a hypothesis, and based on the number of headlines generated off the back of this, I think it's reasonable to presume there was some kind of funded apparatus for disseminating that bogus result.

Ever since then, I've learned not to trust the headlines when it comes to trials, reserving judgment until I've looked at the methodology. When I do, a lot come up short.

--

Can we nevertheless extract causality from correlation?

I would argue that, theoretically, we cannot. Practically speaking, however, we frequently settle for “very, very convincing correlations” as indicative of causation. A correlation may be persuasively described as causation if three conditions are met:

Completeness: The association itself (R²) is 100%. When we observe X, we always observe Y.

No bias: The association between X and Y is not affected by a third, omitted variable, Z.

Temporality: X temporally precedes Y.

it all assumes you can divide the world cleanly into variables that can be the nodes of your DAG. The philosopher Nancy Cartwright talks about this a lot, but it’s also a practical problem.

You can make the argument, from correlative data, that bridges and train tracks cause truck accidents. And more importantly, if you act like they do when designing roadways, you actually will decrease truck accidents. But it's a common-sense-odd meaning of causality to claim a stationary object is acting upon a mobile object...

------------------ Confounders ------------------

A variable that affects both the exposure and the outcome. It is a common cause of both variables.

Role: Confounders can create a spurious association between the exposure and outcome if not properly controlled for. They are typically addressed by controlling for them in statistical models, such as regression analysis, to reduce bias and estimate the true causal effect.

Example: Age is a common confounder in many studies because it can affect both the exposure (e.g., smoking) and the outcome (e.g., lung cancer).

------------------ Colliders ------------------

A variable that is causally influenced by two or more other variables. In graphical models, it is represented as a node where the arrowheads from these variables "collide."

Role: Colliders do not inherently create an association between the variables that influence them. However, conditioning on a collider (e.g., through stratification or regression) can introduce a non-causal association between these variables, leading to collider bias.

Example: If both smoking and lung cancer affect quality of life, quality of life is a collider. Conditioning on quality of life could create a biased association between smoking and lung cancer.

------------------ Differences ------------------

Direction of Causality: Confounders cause both the exposure and the outcome, while colliders are caused by both the exposure and the outcome.

Statistical Handling: Confounders should be controlled for to reduce bias, whereas controlling for colliders can introduce bias.

Graphical Representation: In Directed Acyclic Graphs (DAGs), confounders have arrows pointing away from them to both the exposure and outcome, while colliders have arrows pointing towards them from both the exposure and outcome.

------------------ Managing ------------------

Directed Acyclic Graphs (DAGs): These are useful tools for identifying and distinguishing between confounders and colliders. They help in understanding the causal structure of the variables involved.

Statistical Methods: For confounders, methods like regression analysis are effective for controlling their effects. For colliders, avoiding conditioning on them is crucial to prevent collider bias.

This is an interesting example, because I don’t know of any studies (although there probably are some, if only old ones) specifically about whether smoking pipes staves off lung infections, but the “intuitive sense” answer has changed because of adjacent evidence. And in this case, it’s not the lung cancer evidence that makes it intuitively unlikely that pipe smoking would be helpful, but a broader understanding of what causes lung infections, and what tobacco smoke contains and doesn’t contain.

First, experiments have their own varieties of horrors. Many are small N, with selective data reporting, and lack external validity — that is, the thing you really want to randomize is difficult or impossible to randomize, so researchers randomize something else as a proxy that's not at all the same. Other times there's complex effects that distort the interpretation of the casual pathway implied by the experiment.

Second, sometimes it's important to show that any association exists. There are cases where it's pretty clear an association is non-existent, based on observational data and covariate analysis. You just don't hear about those because people stop talking about them because of the null effects. So there's a kind of survivorship bias in the way results are discussed, especially in the popular literature.

It's easy to handwave about limitations of studies, it's much harder to create studies that provide evidence, for logical, practical, and ethical reasons. Why you'd want less information about an important phenomenon isn't clear to me.

You didn't show causality, though. You never randomized anything. His study and your observation was purely observational. At no point did you open the hood, get ready to wrap the wires, and flip a coin to decide whether to wrap the wires or do a placebo wrapping somewhere else.

Had you done that, you would have found, per your ultimate explanation, that the wrapping made no causal difference: you did the procedure, and either way, the car turned on. Hence, there is no causality for you.

IMO the model of that story is the S at the end of experiments is more than just repeating the same things. Fixing the carburetor was the second and vastly more informative experiment, but your friend could have tried various alternatives to doing exactly what he was doing which would then uncover the time component.

Science digs into problems, so the most important part of meta analysis, which is often ignored, is asking if the question even makes sense in a generic context. Just as crucial is narrowing down the effect size which may be dwarfed by random noise in some experiments etc.

If my parser gets null's when it should be non-null then I first need to find where they could potentially even come from, aka get a better understanding of the model I'm working with.

If you could reproduce it, it would usually be intermittent. Eventually you would learn “when I X then my character will Y, but only sometimes.

This is due to the real command being a subset or being a slight variation of what you thought was correct that you accidentally do sometimes.

Even when it’s ephemeral and seemingly random I still find these things valuable. It’s better to be able to reproduce it sometime instead of never. Answering the question “is doing this better than random?” (P95) can help you throw away a bad hypothesis. Most people don’t realize that when they are providing evidence for causality they are competing with random. If they had instead done jumping jacks or said a prayer to the engine gods X times, then the correlation between the wires and the engine might suddenly seem much weaker.

Once you have one hypothesis you can test it against others and I believe that’s powerful. Provided it’s done systemically and with at least a mild understanding of probability and error. Also a hypothesis without a theory first scientific. Why did your friend wrap the wires to begin with?

It’s okay to act in random until we find some effect, but then we also need to take the time to roll back (as you did) to ask “WHY did this happen?” In which case you can begin the process with a fresh hypothesis.

I feel when we are taught the scientific method in elementary school it doesn’t stick for most of us, even engineers. Especially non-engineer folks. It seems at first blush like some truisms strung together, but that simplicity hides very powerful capabilities and subtle edge cases.

There are all manner of observable, reproducible behaviors in nature that we barely have an explanation of. Those things remain observable and reproducible whether we can tell a tidy story about why they happen.

In a very meaningful sense, the local healer applying poultices formulated from generations of experimentation is using science much as the medical doctor is (assuming, of course, they're taking notes, passing on the discoveries, and the results are reproducible). The doctor having tied their results to the "germ theory of medicine" vs. the local healer having tied theirs to "the Earth Mother's energies impregnate the wound" is an irrelevant distinction until (and unless) a need comes along to unify the theory to some other observable outcomes.

I don't actually know how the method you suggest compares in the limit of finer bins. It's possible it might only achieve similar results?

What worth is a result with p<0.01 when the 10 previous articles with negative results were never actually written?

If you define everything to be "inside", then causality disappears because intervention disappears.

Baby boom as solar panels sales skyrocket.

Causality is a largely orthogonal problem to frequentist/bayesian - it makes everything harder, not just one of those!