Correlation Isn’t Causation, Except When It Is

In the world at large- a clickbait, deepfake, soundbite world- nuance is everywhere on the ropes.? We seek our answers in the simplistic constraints of good/evil, left/right, yes/no, us/them, always/never.? As a result, nuance- an ineluctable element in genuine understanding- is neglected, rejected, and discarded.? The answer to almost any question worth asking involves a bracing dose of: it depends. In a world without nuance, sensible answers elude, and perhaps no questions are worth asking.? Such as: is correlation causation?

In my world- the nutrition world- much of what we know is from so-called observational studies.? We have no shortage of randomized controlled trials , let me hasten to append.? But some of what we want to know- the effects of dietary patterns on health outcomes across the entire lifespan , for instance- is just not highly conducive to randomized assignment, to say nothing of blinding and placebo control.? The ‘placebo diet’ has yet to be invented; blinding people to what they are eating is, at best, a very messy enterprise and at worst, impossible; and randomly allocating people to a diet they don’t prefer for 30 years is…well, let’s just say recruitment is apt to prove challenging.

Filling those inescapable gaps are observational studies.? Whenever one of those is published, the ranks of pundits opposed to the findings compete to be first to remind us: correlation is not causation!? In other words, just because Food A and Outcome B happen together, perhaps a lot, it does not reliably mean that Food A is causing Outcome B.? Of note, all too often those very same pundits are silent when the association sucking all of the oxygen out of a news cycle happens to be one they favor.? I applaud those who get equally rankled about correlation and causation whether the findings du jour point toward, or away from, the position they happen to own.

As for the contention itself, it can of course, in certain context, be so self-evident as to require no reminder.? People with reliable Internet access tend to have better dentition than those without- not because the Internet is good for your teeth, but because areas with fiber-optic cables and 5G cell service tend also to have dentists.? I invite you to think of even more absurdly self-evident examples.?

But then again, in other context, the contention is just as self-evidently false.? We needn’t go so far as invoking Judea Pearl, author of The Book of Why: The New Science of Cause and Effect , ?a founding father of artificial intelligence- who lays out the case that translating correlation into causal understanding is a cornerstone of natural intelligence, and essential in the construction of AI.? We may remain much homelier- and think about tossing something, an apple perhaps, into the air- and our confidence that doing so will ‘cause’ it to fall back down.? How do we really know?

Yes, it’s true, every apple ever tossed up into the air of this planet has fallen back down- but that’s just correlation, isn’t it?? Something else entirely might cause the apple to fall; solar flares, for instance, or leprechauns.? Perhaps every single time anyone tossed an apple- or for that matter, a baseball, or confetti, or a four-leafed clover - there was a solar flare, and that makes things fall.? Perhaps in the absence of those singularly timed solar flares, things tossed up don’t fall down.?

That’s the argument in ‘correlation is not causation.’? Just because A and B occur together, you cannot conclude that A caused B.

Except, of course, when you can.? Sometimes, the ‘correlation’ between A and B is off-the-charts, overwhelming.? Sometimes every instance of A is linked to an occurrence of B.? Or, if not, sometimes there is a clear mechanistic pathway (or several) linking the instance of A with the occurrence of B- as gravity links the toss of the apple with its fall.? Sometimes, there is both overwhelming consistency and clear mechanism.? And sometimes, into the bargain, no other explanation (e.g., solar flares) is remotely as plausible as the obvious correlate.

Well, then, in such circumstances, correlation is causation.? You are, I trust, fully persuaded that putting your hand into a fire will get you burned, despite the lamentable lack of randomized controlled trial data on the topic.? You likely accept that looking both ways rather than just one, or none, before crossing a busy street is causally linked to a reduced probability of being run over- although no study I have seen has randomized people to cross busy roadways with, and without, blindfolds.? You may even have presumed to admonish your offspring against running with scissors, despite the sorry want of intervention studies and meta-analyses on the mere ‘correlation’ between running with such sharp objects, and…punctured children.

There are, to my knowledge, no placebo-controlled, randomized trials proving the utility of parachutes* for those hopping out of planes at altitude.? There are no randomized trials proving the advantages of remaining inside the submarine at great ocean depths.

Let’s belabor this list no further.? Much, even most, of what we know about how to function as human beings, we know because of observation, correlation, and our ability to infer causation.? I am tempted to say that again, but you can just read it again; thank you.

Why this topic, now?? As ever, the provocation lies in a recent study , this one linking meat intake with diabetes (meat intake has been linked to other adversities of modern epidemiology in prior studies) , and various reactions to the study.? The study was observational and reported what observational studies report: a statistically significant association between exposure to A and the occurrence of B.? In this case, A was more meat consumption, and B was incident diabetes.

We were quickly reminded by those opposed to the conclusion that correlation is not causation .? But the context here belies that predictable clamor.

First, by global standards, Americans have an exceptionally high intake of meat on average, and exceptionally high rates of chronic disease to go along with that.? No matter the specific health effects of any given food, balance is good, and imbalance is bad.? Even an excess of water is potentially hazardous (causing, among other things, hyponatremia).? High meat intake in a population with high meat intake amplifies imbalance, so yes- that is bad.? So, too, the excess of saturated fat, specific amino acids, and possibly iron- among other concerns.

In addition, and importantly, every real-world question about the role of a given food or food group in the diet must contend with the critical “instead of what? ” question.? The simple fact is that once certain foods make up 100% of your daily eating, there is no room for anything else.? The more of your calories allocated to meat, then, the fewer allocated to vegetables, fruits, whole grains, nuts and seeds, and in particular legumes- the ideal nutritional alternatives to meat – all protective against diabetes (and other chronic diseases) in a variety of ways.? Americans are generally ‘deficient’ in these food groups and the important nutrients they deliver- from fiber to phytonutrients, antioxidants to potassium- and the more meat-based the diet, the greater that deficiency.?

How far-fetched is it that eating more of what we already get too much of, and consequently less of what we get too little of…amplifies imbalances, and is bad for health?? Meat in the American diet is coals to Newcastle.

So, yes, the correlation between meat intake and bad health outcomes does have causal implications- not just because of the association in a single study, but because of the consistency, context, and clear mechanisms all attached to that ‘correlation.’? Is meat intake directly responsible for increasing diabetes risk, via an inflammatory pathway, or indirectly responsible because it displaces beans, lentils, whole grains and nuts- all protective against diabetes- from the diet?? If I had to guess, I would say “yes” to this question, but in the absence of guessing- we can concede we don’t know for sure.? The causal connection is clear either way.? Perhaps an analogy will help clinch it for you: walking is good for health.? Is that because walking means more walking, or because it means less sitting?? Well, yes.?

On the topic of meat, our current intake levels are bad for us; brutal to our fellow creatures ; and devastating to the health of the planet .? As causal and clear as the first indictment is, the second and third are even more so.?

Correlation is causation when the consistency of the association extends beyond the limits of statistics into the realm of the obvious; when the mechanistic pathway is clear; when alternative explanations are implausible.? Whenever several, simple criteria are satisfied- we can infer our way to causation from mere correlation.? We do it quite reasonably and reliably, every day - and almost certainly could not function, or even survive, if we did not.? In the simpler cases, we call this super-power sense.? When there is a hint of challenge in the enterprise, we call it intelligence.

Leaving aside the many RCTs, a bounty of correlation in the context of intelligence and sense makes clear the profound implications of dietary pattern and quality for health outcomes .? Likewise, the case is clear that diets made up proportionally more of A are made up less of B.? More of what is excessive is bad; less of what is deficient is bad.? The pathways that lead to chronic, degenerative diseases are few and overlapping and well mapped: inflammation, oxidation, toxication, dysbiosis.? Take the relevant correlations, add a modicum of intelligence, and yes- meat intake is causally implicated in diabetes, as it is in much else of what ails us.

As we rush to bequeath intelligence to robots, now would be a good time to make sure we understand how the native version works.? Among its many wonders, it makes of correlation causation, reliably, and as a matter of routine.

?

-fin

*There are, alas, randomized trials showing that inadequate lifestyle interventions are inadequate, just as a study of postage-stamp-size parachutes would indicate their inadequacy.? This is yet another liability of RCTs : they can mislead us entirely when directed at a misguided question.

??

Dr. David L. Katz is a board-certified specialist in Preventive Medicine/Public Health, past president of the American College of Lifestyle Medicine , and the founder of Diet ID, Inc .?He has authored multiple editions of leading textbooks on both nutrition and epidemiology, and is the co-author, with Mark Bittman, of How to Eat .