Seems very dismissive and unaware of recent advances in causal inference (cf other comments on Pearl). Putting "throw the kitchen sink at it" regression a la early 2000s nutritional research (which is indeed garbage in garbage out) in the same category as mendelian randomization, DAGs, IP weighting, and G-methods is misleading. I do worry that some of these EA types dive head-first into a random smattering of google scholar searches with no subject matter expertise, find a mess of studies, then conclude "ah well, better just trust my super rational bayesian priors!" instead of talking with a current subject matter expert. Research -- even observational research -- has changed a lot since the days of "one-week observational study on a few dozen preschoolers."

A more general observation: If your conclusion after reading a bunch of studies is "wow I really don't understand the fancy math they're doing here" then usually you should do the work to understand that math before you conclude that it's all a load of crap. Not always, of course, but usually.

In my own research we are investigating how fluids cause changes in rocks that allow for mineralization of CO2 and have such problems of confounding variables (not terribly unique I suppose). One thing we note is that, well, fluid comes from the sky and goes into the ground. Thus, the deeper you go, the less fluid there is since the pathways from the sky to deep into the ground become more sparse as well as needing higher pressures to enter these regions to either overcome capillary pressures in existing fracture zones or to literally break the rock (which is highly unlikely using naturally occuring pressures from fluids from the sky). And so, literally everything in all the data sets correlates with depth in some way. But in what way? well this has many dependencies as well, did the rock that absorbed some of the fluids grow in volume because of a chemical change? are the fluid pathways currently connected? What kind of rock is absorbing the fluids? Are microbes in the fluid absorbing contents from the fluid that would otherwise be used for rock changes? and so you are left with this giant pile of data (tens of terabytes) without a clear connection between fluid and rock interactions except that there is less fluids from the sky the deeper you go into the rock. This is obvious, however it is also rather unhelpful when trying to understand the other processes that exist. Of course you might say, have you tried detrending your data? And the answer is yes and to no effect. The simple truth is that this depth dependency interacts in different ways with different systems and there is no easy way to figure out how it does for each sub-system such as the fluid rock chemistry interactions, the rock fracture mechanics, the subsequent methane and hydrogen that is produced and likely consumed by microbes, etc.

I think a lot of these kinds of studies are not really about objectively studying a phenomenon but trying to prove a predetermined point. The study is designed and adjusted until it proves what it should prove. Then it's wrapped in a nice news headline which goes away with all the details and subtleties and used for political or economic gain. Reproducing the results is not interesting and not funded. Other studies are then using these results as sources to stack the house of cards even higher. I think this does a lot of harm to science as a whole because a lot of people disregard all scientific results as a result.

Oh what fun to discover the horror of causality!

For some areas of research, truly understanding causality is essentially impossible - if well-controlled experiments are impossible and the list of possible colliders and confounders is unknowable.

The key problem is that any causal relation can be an illusion caused by some other, unobserved relation!

This means that in order to show fully valid causal effect estimates, we need to

- measure precisely

- measure all relevant variables

- actively NOT measure all harmful (i.e. falsely correlated) variables

I heartily recommend the book of why [1] by Pearl and Mackenzie for a deeper reading and the "haunted DAG" in McElreath's wonderful Statistical Rethinking.

[1] https://en.wikipedia.org/wiki/The_Book_of_Why

On a similar note, I enjoyed watching the video: https://youtu.be/mQ56uOkjccg?si=1hpwGqv2dQqLQ-ME (by Nutrition Made Simple!)

It takes a specific topic (here, health effects of red meat) and explains how each type of study can provide information, without proving anything. It helped me a lot understand the science related to nutrition, where you never have perfect studies.

Dismissing all observational study designs out of hand because they can be difficult and easy to perform poorly seems like quite the take.

I see this all the time in people’s interpretation of nutrition research, and they do exactly as this article suggests and fall back to the “intuitive option”, and go onto some woo diet that they eventually give up because they start feeling awful.

I would disagree that observational study designs should be thrown out the window or that it makes sense to, as this article seems to do, lump cross-sectional ecological data in with prospective cohort studies.

Things often “make intuitive sense” only because of these study designs. We used to get kids to smoke pipes to stave off chest infections because it made “intuitive sense” and it’s only because of observational studies that we now believe smoking causes lung cancer.

The direction of evidence from prospective cohort studies to RCTs in the field of nutrition science on intake vs intake shows a 92% agreement. If we take RCTs to be the “gold standard” of evidence that best tracks with reality, it seems a little odd that these deeply flawed observational studies that we should apparently disregard seem to do such a good job coming to the correct conclusions.

https://bmcmedicine.biomedcentral.com/articles/10.1186/s1291...

I am slowly becoming convinced that studies are in fact cargo-cultism. And there are many, many studies that confirm this.

But about causality. Long ago (old cars) I had a friend who told me that most mornings his car would not start until he opened the hood and wrapped some wires with tape (off with the old tape on with the new). Then the car would start. Every now and then it would take two wraps. Hmmm.

After he demonstrated this, I decided to try to help. I followed the wires that were wrapped. Two of them. To my surprise they were not connected at either end. This was insane, and yet his study - and my own observation - demonstrated that wrapping these two wires which were completely disconnected caused his car to start. Now there is causality for you.

Except that if you have a more complex model of cars, there is a sane explanation. Again this is an old car with a carburetor. In case you don't know this is a little bowl of gas it that provides a combustible mix of air and gas. If there is too much gas then your car won't work. The mix is controlled by a little float that controls the level of gas in the little bowl. Toilet bowls work on the same principle.

If your float is bad (or other issues) your car engine would get too much gas - be "flooded" and you have to wait until much of it evaporates. So if you flood your car engine, go and wrap some wires, it may be that your car will start right up.

So I rebuilt the carburetor and my friend never had that problem again.

The moral of the story is that I had better "model" of how cars work. But in the back of my mind I am aware that my model may be or have been just as deficient. Did you know that we are bombarded from space by an unknown type of neutrino that stops electricity from working unless there is a little pool of some liquid nearby or it is Thursday. I am going to do a study of this.

There are very good reasons to understand how frail our ability to understand causality is. And we are talking simple things here. The scientific method is about EXPERIMENTS. Yes, I did that in bold. Doing things. We have deeply complex situations we need to understand and in my opinion, studies do not help.

As with many things, just understand what you are trying to do.

If you want to predict Y and you know X, you can use data that tell you when they happen together.

If you are trying to cause (or prevent) Y, it's harder. If you can't do experiments (e.g. macroeconomics), it's borderline impossible.

"a technique called regression analysis that, as far as I can determine, cannot be explained in a simple, intuitive way (especially not in terms of how it "controls for" confounders)"

That sounds very much like a skills issue. Because it can. You call out what you consider might be confounders as independent variables (covariates). You can then use regression analysis to estimate the individual contributions from each confounder, and control for them by essentially filtering out that contribution.

Is reality harder than that? Yes. Much. The world of science isn't 9th grade math, sorry. You are not entitled to understand everything deeply with 5 minutes of mediocre effort.

> Now, a lot of these studies try to "control for" the problem I just stated - they say things like "We examined the effect of X and Y, while controlling for Z [e.g., how wealthy or educated the people/countries/whatever are]." How do they do this? The short answer is, well, hm, jeez.

You mean they don't cluster the data into sets of overlapping bins where the controlled attribute has approximately the same value and then look for the presence of an XY relationship within the bins instead of across them?

"A study using a complex mathematical technique claiming to cleanly isolate the effect of X and Y. I can't really follow what it's doing..."

This is a frustrating type of issue. Dismissing something with "I don't understand this, but I don't believe it" isn't the sort of thing I want to be doing. However, I don't have any desire to waste time trying to understand what someone has done (and did they really understand what they were doing themselves?) when it's clear that the effect isn't cleanly isolated in the data and no amount of mathematics is going to change that.

Am I the only one thinking through the reading of this: "Wait a minute... isn't this article some kind of weak X then Y also? Observation of many cases, with generalized causality concludes that he just feels like x should cause Y? hahaha. Love the article btw.

I disagree strongly with this mathematised notion of causality. Two things can be perfectly correlated at all observed points in history without necessarily being causal. There can always be some unknown variable driving change in both.

And all this before you even get to "how much of an impact of X on Y should there be before it is even close to a bottleneck that's worth actually acting on, and do we think it reaches that threshold?"

It's layers of abstraction all the way down the light cone.

The causality is always present, we just don't have the processing power to ensure with 100% certainty that all relevant factors are accounted for and all spurious factors dismissed.

> I have to say, this all was simultaneously more fascinating and less informative than I expected it would be going in.

Direct quote from the author of this post and I couldn't agree more, particulartly about the post itself.

Related. Others?

Does X cause Y? An in-depth evidence review - https://news.ycombinator.com/item?id=30613882 - March 2022 (3 comments)

Previous post: https://news.ycombinator.com/item?id=42965989

As a clinical psychologist, I find it increasingly frustrating to sift through research studies that fail to meet even the most basic standards of scientific rigor. The sheer volume of studies that claim “X is linked to Y” without properly addressing the correlation-versus-causation fallacy is staggering. It’s not just an oversight—it’s a fundamental flaw that undermines the credibility and utility of psychological research.

If a study is publicly funded, there should be a minimum requirement: it must include at least two research arms—one with an experimentally manipulated variable and a proper control condition. Furthermore, no study should be considered conclusive until its findings have been successfully replicated, demonstrating a consistent predictive effect. This isn’t an unreasonable demand; it’s the foundation of real science. Yet, in clinical psychology, spineless researchers and overly cautious annd/or power crazed ethics committees have effectively neutered most studies into passive, observational, and ultimately useless exercises in statistical storytelling.

And for the love of all that is scientific, we need to stop the obsession with p-values. Statistical significance is meaningless if it doesn’t translate into real-world impact. Instead of reporting p-values as if they prove anything on their own, researchers should prioritize effect sizes that demonstrate meaningful clinical relevance. Otherwise, we’re left with a field drowning in “statistically significant” noise—impressive on paper but useless in practice.

I like this writing style with unbound variables. Reminds me of Maya Binyam's novel "Hangman", or Kafka's novels.

so if we have a scenario where we have data points where when X ball moves white ball also moves, but we’re missing some direct evidence where they actually hit each other or not. But they correlate from the limited sample. I think this is what most correlations are like, we do not see the direct atoms causing the causation, only a probability

There is no causality, what so ever. The perceived causality is built backwards, only to make something appear sensible. Every event in this universe contributes as a cause to every other event in the universe. It's like fluid flow. Every molecule of the fluid affects the movement of every other molecule. The world evolves in a fluid motion, not through isolated causal chains.

Most important factor on results of research are personal beliefs, especially in "economics".

The question of whether X can cause Y remains open and requires further research. The article highlights the importance of thoroughly checking sources and methodology to draw clear conclusions. This is an important step towards a deeper understanding of such relationships.

Pearl's Causality is very high on my "re-read while making flashcards" list. It is depressing how hard it is to establish causality, but also inspiring how causality can be teased out of observational statistics provided one dares assume a model on which variables and correlations are meaningful.

even if you hit all the assumptions you need to make Pearl/Rubin causality work, and there is no unobserved factor to cause problems, there is still a philosophical problem.

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.

And this is even before we get into the philosophical / epistemological questions about "cause."

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...

I’ve heard Miguel Hernán’s “What If” is also excellent, but not got round to reading it.

And even if you do know there's causality (eg: the input variable X is part of software that provides some output Y), the exact nature of the causality can be too complex to analyze due to emergent and chaotic effects. It's seldom as simple as: an increase in X will result in an increase in Y

I have reflected on a good definition of causality and would be curious if anyone has thoughts or critiques of it. I am repasting part of my essay below. (https://alexpetralia.com/2023/02/25/statistics-only-gives-co...)

--

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.

That colliders and confounders have technical definitions is not known by some:

------------------ 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.

It's important to be thoughtful about research interpretation, but I'm kind of tired of kneejerk dismissal of observational studies for a couple of reasons.

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.

> We used to get kids to smoke pipes to stave off chest infections because it made “intuitive sense” and it’s only because of observational studies that we now believe smoking causes lung cancer.

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.

It is quite the take indeed; one that I posit resonates most strongly with people whose societal views tend to conflict with the available evidence.

> After he demonstrated this, I decided to try to help. I followed the wires that were wrapped. Two of them. To my surprise they were not connected at either end. This was insane, and yet his study - and my own observation - demonstrated that wrapping these two wires which were completely disconnected caused his car to start. Now there is causality for you.

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.

I liken this to the experience of playing an old school fighting game in the era before the internet. You would be mashing buttons when suddenly your character would do a power move. Then spend the rest of the day trying to figure out how to reproduce it.

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.

I think the most fascinating thing about the practice of science (and this is one of those things I wish I'd realized sooner when learning physics) is that experimental evidence often outstrips theory.

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.

> The scientific method is about EXPERIMENTS

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.

Doesn't this heavily apply to building software as well? e.g. instead of spray and pray development we should get a better understanding of the model we're working with.

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.