If this is your first time reading, please check out the overview for Poor Ash’s Almanack, a free, vertically-integrated resource including a latticework of mental models, reviews/notes/analysis on books, guided learning journeys, and more.
Causation / Multicausality / Disaggregation Mental Model: Executive Summary
If you only have three minutes, this introductory section will get you up to speed on the causation / multicausality / disaggregation mental model.
Causation / multicausality / disaggregation in one quote:
I never solve the problem I am asked to solve, |because it’s| not the real, fundamental, root problem. It’s usually a symptom… it’s amazing how often people solve the problem… without bothering to question it.– Don Norman
A quick overview of causation / multicausality / disaggregation: our understanding of causation is often overly swayed by “storytelling,” as well as correlations that aren’t causations (ex. hard work is correlated with success, but does not on its own cause success independent of other factors). Consultants and other structured thinkers use a “ MECE” framework (mutually exclusive, collectively exhaustive) to segment problems into analysis.
While there are limits to this approach, it’s a starting point for evaluating situations. By inversion, aligning multiple causes can cause multiplicative “Lollapalooza” effects.
Two brief examples of causation / multicausality / disaggregation:
How do you understand heredity? The same way you eat an elephant: one bite (of peas) at a time.
Sam Kean’s “ The Violinist’s Thumb” provides an excellent overview of how Gregor Mendel used disaggregation to determine causation:
“this focus on separate, independent traits allowed Mendel to succeed where other heredity-minded horticulturists had failed.
Had Mendel tried to describe, all at once, the overall resemblance of a plant to its parents, he would have had too many traits to consider […] but by narrowing his scope to one trait at a time, Mendel could see that each trait must be controlled by a separate factor.”
Is it cheaper to give apartments to the homeless? Some evidence reviewed by Megan McArdle in “The Up Side of Down” suggests that might, counterintuitively, be the case.
The costs of the homeless to society are multicausal, and by disaggregation, the biggest line item is hospital visits – which, as David Oshinsky provides a window into in “Bellevue,” are often more for basic personal care like meals and showers than for actual medical problems.
McArdle notes that there’s thus a high opportunity cost to letting the homeless remain on the street; it might be cheaper to get them apartments. Naturally, there would be unintended consequences or n-order impacts to such a policy – McArdle explores the concept of “moral hazard” – but this sort of analysis is at least an interesting starting point for consideration.
If this sounds interesting/applicable in your life, keep reading for unexpected applications and a deeper understanding of how this interacts with other mental models in the latticework.
However, if this doesn’t sound like something you need to learn right now, no worries! There’s plenty of other content on Poor Ash’s Almanack that might suit your needs. Instead, consider checking out our learning journeys, our discussion of theinversion, margin of safety, or Bayesian reasoning mental models, or our reviews of great books like “ Polio: An American Story” , “ The Autobiography of Benjamin Franklin” , or “ Deep Survival” by Laurence Gonzales .
Correlation vs. Causation: Why Hard Work Doesn’t Cause Success
Ice cream sales and forest fires are correlated because both occur more often in the summer heat. But there is no causation; you don’t light a patch of the Montana brush on fire when you buy a pint of Haagen-Dazs.
– Nate Silver
Nate Silver’s “ The Signal and the Noise” does a great job of breaking apart correlation and causation. It’s an interesting topic, but you’re welcome to skip this section if you’re familiar with it (it’s fairly basic) and move on to some of the more intriguing parts of the model below. Silver discussed “big data” which has of course now been turbocharged with AI. There’s somewhat of a trend toward the idea that correlations are everything: i.e., the notion that we don’t need to know why the black box works… it just needs to work.
Silver is critical of this approach, exploring its drawbacks and the reasons to focus on causality (see also the Bayesian reasoning mental model.) So is Michael Mauboussin, who notes in “The Success Equation” :
Statisticians who are serious about their craft are acutely aware of the limitations of analysis. Knowing what you can know and knowing what you can’t know are both essential ingredients of deciding well.
I think most people are familiar with the idea of correlation vs. causation, so I won’t waste too many words on it: I’ll just point out a few important tenets and examples, and you can pick the rest up on your own.
One thing many people don’t know is that our tendency to associate correlation with causation is deeply embedded in our brains. Laurence Gonzales discusses this in Surviving Survival, his book on resilience and recovery in those who have been through traumatic near-death experiences. He notes:
In the brain, the cardinal rule is: future equals past; what has happened before will happen again.
Gonzales goes on to note the counterintuitive phenomenon – also mentioned by Shawn Achor in Before Happiness – then points out:
If you [imitate] a smile (say, by biting on a pencil), you gradually start to feel better… feeling good is the retrieval cue for smiling and smiling is the retrieval cue for feeling good.
Gonzales’s fascinating book discusses the “autoassociator” in our brain. So, clearly, this is something we need to be aware of, and actively counter.
A correlation between one thing and the other doesn’t tell you anything about the direction of the causal relationship. You have to ascertain that by other means. Yet people often mistake correlations for causal relationships. For example, many people hold the nonsensical view that one can do something such as “sleep too much” – that beyond a certain point, more sleep is worse for your health, productivity, and happiness.
This isn’t even remotely true. Dr. Matthew Walker explains why in his brilliant “ Why We Sleep” – my candidate for most important book of the century (no hyperbole). Walker notes that sleep is not really dose-dependent:
“No biological mechanisms that show sleep to be in any way harmful have been discovered.”
Nine hours is much better than eight, which is much better than seven, and below seven you’re in the high danger zone. Given that, why can you find a correlation between very high sleep and mortality? Walker notes that there’s a correlation vs. causation problem in mortality data associated with people who sleep a lot: many of the people who are sleeping a lot are seriously or terminally ill, and obviously we sleep more when we’re ill for a variety of reasons. Conversely, people in good health usually have jobs, hobbies, and families that take up their time; few active/healthy people have the time to sleep 10 hours per day. That doesn’t mean that the sleep is the cause of the negative health outcomes; in fact, it’s the other way around.
For example, a good friend of mine had a family member with advanced multiple sclerosis (MS). She slept about 12 hours per day. Assuming she is representative of those with her specific condition, certainly a study that included subjects like her would find a correlation between length of sleep and higher mortality, but it’s very difficult to tell a plausible story in which the sleep is the cause of her health issues – rather, it’s the other way around.
Another important example relates to what’s called a “post hoc” fallacy – short for Latin “post hoc, ergo propter hoc,” which is fancy verbiage for “after this, therefore because of this.”
Humans are very good at storytelling. As Don Norman points out in “ The Design of Everyday Things” :
[… humans’ conceptual models] are often constructed from fragmentary evidence, with only a poor understanding of what is happening, and with a kind of naive psychology that postulates causes, mechanisms, and relationships even where there are none.”
The quote was in relation to how consumer products work, but it works equally well relative to other lines of human thinking.
Here’s the second example, which is vastly counterintuitive for many people, especially thanks to the inexplicable popularity of nonsense like the “ grit” mentality. Hard work doesn’t cause success.
“But wait,” you say. “Everyone I know who is successful works hard. You can’t be successful without hard work.”
Yes and no. Success is certainly correlated with hard work; hard work is also required for many of the potential factors that cause success – like being good at a tangible skill.
But as Jordan Ellenberg points out in “ How Not To Be Wrong,” correlation isn’t transitive. That is to say: if A is correlated to B, and B is correlated to C, A isn’t necessarily correlated to C.
One of the challenges with correlations is that they don’t give you a control group. For example, if we’re trying to determine whether or not hard work causes success, we’d need to not just look at successful people and see if they all work hard, but also look at all people who work hard and see if they’re successful.
And the answer is: they’re not. A lot of low-wage jobs (construction, retail, janitorial) involve a lot of hard work. They would certainly take a lot more effort than my job, which mostly involves reading things in an air conditioned room – not lifting heavy stuff on a sweltering factory floor.
The answer here is that it’s doing things that society values highly, and doing them well, rather than hard work, that causes success. I discuss this more in the busyness vs. productivity model, but this classic quote from Stephen Covey’s “ The 7 Habits of Highly Effective People” illustrates the topic:
“envision a group cutting their way through the jungle with machetes […]
the leader is the one who climbs the tallest tree, surveys the entire situation, and yells, “wrong jungle!”
[…] as individuals, groups, and businesses, we’re often so busy cutting through the undergrowth we don’t even realize we’re in the wrong jungle.”
So how do we determine if a correlation has a causatory mechanism? We’ll investigate that.
Application / impact: just because two events are temporally associated does not mean that one caused the other.
Disaggregation / MECE: How Many Pairs of Boxer Briefs Are Sold Annually In America… And Why Don’t Teachers Get Paid A Lot?
How many pairs of boxer briefs are sold annually in America?
Yes, it’s a frivolous question, and no, you’re not allowed to go searching for answers on Google or in Hanesbrands’ annual report. It’s the type of question where the answer is less important than how you get to the answer.
How should you get to the answer? Well, let’s break it down into component pieces. Boxer-briefs are mostly only worn by males, so we should start with the number of boys and men in America – or roughly half the population, excluding a few diaper-wearing infants. (We will, for discussion purposes, exclude people who choose to go commando.)
Individuals have different preferences; these preferences might also vary by age, or even by generation due to social proof, from early childhood to adolescence to adulthood. So a reasonable starting point would be to segment the American male population into various demographics and assign percentages to determine how many males in each age group wear boxers, briefs, or boxer-briefs.
(Careful not to rely too much on the availability bias of your own preference, or that of the guys in your life.)
Third, the purchase of new underwear is likely a function of how many pairs of underwear the average guy needs, and how frequently they need to be replaced.
So we’ve broken the problem down into three constituent components – the total population, the percentage of that population that comprises our demographic, and the purchase frequency among that demographic – and if we come up with reasonable inputs for those – whether by guessing or by sourcing actual data – then we should have the right answer. A good disaggregation will be “MECE” – mutually exclusive and collectively exhaustive.
Of course, in the real world, you can probably just ask Perplexity to find you the answer – but the thought process is illustrative of how to break down a topic into manageable parts. Questions like that one, and the following, are often asked in interviews for major consulting firms like BCG, Bain, and McKinsey:
– How many gas stations are in America?
– How many golf balls could you fit on a 747?
– How many piano tuners are in Chicago?
The proper mental approach for answering these – and more complicated business “case interview” questions – are explored in Marc Cosentino’s Case in Point. I found it interesting long ago when I was considering being a consultant, and while at the time I chafed at having to answer dumb questions – “why the hell would anyone want to stuff a 747 with golf balls?” – it turns out that the line of thinking is valuable.
In fact, that last question about piano tuners is sourced from Philip Tetlock’s “ Superforecasting” , a wonderful book about forecasting. It turns out that the eponymous “superforecasters” use a very similar MECE type disaggregation process to make accurate predictions.
For example, in response to one question about how likely it was that a certain agency would detect polonium on a victim’s clothes, one “superforecaster” broke it down with various probabilistic assumptions like: would polonium be detectable today? If it is, would the murderers have had access to it?
Combining disaggregation with probabilistic thinking, Tetlock’s superforecasters wildly outperformed experts at making predictions in their own fields.
Disaggregation seems to be the style of thinking that Sam Hinkie – an ex-Bain guy – took with his famous “process” for the Philadelphia 76ers. Hinkie broke success down into its components:
– high draft picks
– lots of cap space
… then set about getting both of the above.
Disaggregation is the process Don Norman is talking about when he advocates exploring for root causes and fixing those. A similar approach was used in creating the polio vaccine: at one point, as discussed in David Oshinsky’s “ Polio: An American Story” : scientists realized that they hadn’t found polio in the blood because it didn’t manage to survive there after antibodies were generated. That didn’t mean it had never been there to begin with.
See the parallels to the polonium example? This proved to be a breakthrough. via the blood before hitting the nervous system. If you could put antibodies in the blood before polio got to the nervous system… polio would be prevented. But nobody had thought to look there.
Similar disaggregation was used by Leonard Hayflick in creating the WI-38 cell line that proved vital to future vaccine production, as discussed in Meredith Wadman’s “ The Vaccine Race.”
A final example that I find tremendously helpful – albeit deceptively simple – is Peter Thiel’s “X | Y” framework for value creation. Thiel notes in his (super-interesting, definitely worth watching” “How to Start a Startup” lecture that:
I want to suggest there’s basically a very simple formula, that if you have a valuable company two things are true.
Number one, that it creates “X” dollars of value for the world.
Number two, that you capture “Y” percent of “X.”
And the critical thing that I think people always miss in this sort of analysis is that “X” and “Y” are completely independent variables.
There’s plenty more such thinking where that came from in Thiel’s “ Zero to One” . Examples of this abound in business and life: there are some industries, like multilevel marketing, where X may be very low or zero or negative, and Y may exceed X, such that the companies are sucking value out of the system. There are other examples, like teaching and flying airplanes, where tremendous value is created, and yet very little is captured.
So if you want to make a lot of money, you not only have to create value, but also participate in a part of the ecosystem where it can be captured.
Application / impact: large problems are much easier to solve when broken down into small chunks that, in total, solve the problem.
Multicausality & The Swiss Cheese Model
One challenge when assessing causality is that given any chosen group of outputs or studied variables, there’s likely not just one cause. Going back to my previous reference to The Signal and The Noise , several parts of Dr. Jerome Groopman’s “ How Doctors Think” explore some of the challenges of making diagnoses.
For example, Groopman quotes a Dr. Light on the challenge of false positives: while MRIs are obviously wonderful and useful, the truth is that natural variation combined with aging means that if you look hard enough, you’ll find something wonky about everyone – but that doesn’t mean it’s necessarily something wrong that needs treatment or is responsible for a problem.
Groopman’s book is a wonderful exploration of multicausality, among other topics – he provides many examples of cases in which doctors tried to explain all of a patient’s symptoms with merely one diagnosis, when in fact a combination of different conditions contributed to the symptoms. Groopman provides helpful advice on how doctors – and their patients – can avoid this trap.
One of my favorite examples of this comes from chronobiologist Till Roenneberg’s “ Internal Time” . Toward the end of the book, Roenneberg discusses children suffering from a particular genetic disorder (Smith-Magenis syndrome) that was thought to cause (among other symptoms) “severe behavioral pathologies.”
The genetic disorder does cause a lot of challenges, but the way Roenneberg tells it, the behavior actually isn’t one of them (directly, anyway). The kids are actually just sleep-deprived: in some cases, the syndrome can lead to an inverted melatonin production, i.e. during the day at night.
Treating this and allowing them to sleep properly led to both them and their caretakers having a higher quality of life – but it was long overlooked because it was simply assumed that the incurable genetic disorder caused all of the associated symptoms.
This is a particularly important phenomenon in science. One of my other favorite examples of multicausality is explained in Jennifer Ackerman’s delightfully reader-friendly “The Genius of Birds” , which is a book that is as beautiful and enjoyable as it is educational. Exploring studies trying to determine “bird IQ,” Ackerman displays not only intellectual humility / scientific thinking, but also a thorough understanding of multicausality:
“It’s tricky, however. In these kinds of lab tests, all sorts of variables may affect a bird’s failure or success. The boldness or fear of an individual bird may affect its problem-solving performance.
Birds that are faster at solving tasks may not be smarter; they may just be less hesitant to engage in a new task. So a test designed to measure cognitive ability may really be measuring fearlessness.
“Unfortunately it is extremely difficult to get a ‘pure’ measure of cognitive performance that is not affected by myriad other factors,” says Neeltje Boogert [..] a bird cognition researcher at the University of St. Andrews.”
Similar things can happen when it comes to assessing canine intelligence; some dog breeds may be very smart, but not particularly motivated or inclined to follow human instructions; conversely, others (like Labradors) are often not smart, but very eager to please.
Henderson even shows how measurement itself can be gamed. As a kid, he “intentionally gave wrong answers” on a psychologist’s test because “adults didn’t seem to take my needs seriously,” so out of spite, he didn’t want to take their tests seriously.
Later, to keep expectations low, he hid evidence of ability: “I didn’t want anyone to accuse me of being smart,” so he kept his reading habit a secret. And in school he learned to pass while signaling minimal competence, figuring out “the absolute minimum effort required to pass.” Therefore, while Rob was actually highly intelligent, his grades and test scores wouldn’t reflect that.
In other words, like bird tasks that accidentally test boldness and dog tasks that reward eagerness to please, human assessments can end up measuring compliance, trust, or strategy rather than raw cognitive ability. We should be mindful of this, particularly in the age of AI, when generating polished (but wrong) output, or finding answers to test questions (without having true conceptual understanding), is easier than ever.
This sort of analysis is also helpful for understanding human psychology – Charlie Munger always talks about how many people, for example, fail to notice all the cognitive biases at play in experiments like Milgram, and this is definitely a phenomenon I’ve noticed myself.
Many authors and books and analysts take a man with a hammer approach and try to explain everything in terms of their preferred model. For example, Olds/Schwartz attribute to social connection, in “The Lonely American,” behavior that I think clearly has at least some component of learned helplessness.
Not all analysts do this, of course – one reason I particularly love Christopher Browning’s “Ordinary Men” is that he displays a thorough understanding of multicausality, carefully evaluating the contributions of different potential factors to a group of ordinary mens’ willing participation in Holocaust genocide.
One particular angle on multicausality that I find particularly fascinating is the “Swiss Cheese Model” of causality,, discussed in books like Don Norman’s “ The Design of Everyday Things,” and Megan Mcardle’s “ The Up Side of Down.” This model posits that for anything to happen, a number of stars have to align… or a bunch of slices of swiss cheese. It’s often used in context of mistakes or accidents: for example, for a specific car collision to occur, perhaps:
– the driver has to be tired or distracted, AND
– the car’s brakes have to be old, AND
– the road has to be slippery thanks to a recent oil spill
So, which of the three “caused” the car accident? Well, all of them and none of them.
Similar analyses could be conducted for other questions: why do people end up in bad relationships? Why did we not do as well we wanted on that presentation?
It’s an interesting model and there are obvious interactions with margin of safety: add a layer of Swiss Cheese, and you’re less likely to have a bad outcome.
There are a lot of practical, everyday applications of this phenomenon. Gonzales brings up the Swiss Cheese model in Surviving Survival and provides one of the single best pieces of advice I’ve read anywhere:
“… blanket a problem with overlapping solutions.”
Gonzales discusses how having “emotional reserve” helps people respond to trauma; similarly, having “cognitive reserve” helps many people escape the effects of Alzheimer’s and dementia – because they have more to work with. It goes back, in a way, to multidisciplinary rationality – we’re better when we have more than one tool to use.
Application / impact: be aware that searching for a single “cause” is often not likely to yield helpful answers, and can lead you to be a man with a hammer. An understanding of multicausality can be helpful in making more rational decisions.
Disaggregation x Dose-Dependency x Precision vs. Accuracy (x Emergence / Complexity x Inversion x Overconfidence): The Limits Of Analysis, Why Quarks Are Irrelevant, and Lollapaloozas
Like everything else, “collectively exhaustive” analysis is dose-dependent. Given the scale of the world, however extensive your analysis is, there’s more to do, and at some point you have to stop. (This is worse than ever with AI, which allows us to process vastly more information at once – going from the fire hose of the 2010s to …going from a fire hose to a tsunami. The wave hits from every direction at once, and while you can see farther than ever before, you can also be swept away by the sheer volume of “what’s possible to know.”
Let’s break it down. Even if we assume a relatively linear world, everything can be said to cause everything via the Swiss Cheese model mentioned above, kind of like the “butterfly effect.” John Lewis Gaddis presents a very thoughtful analysis of this in “ The Landscape of History,” an astonishingly thoughtful book about historians do (or should) approach their craft.
Gaddis explains that while you could go all the way back to the start of agriculture to explain Stalin’s Soviet Union (disaggregation), since in some sense the Soviet Union never could have occurred without agriculture, that doesn’t make much sense, similarly, more emphasis is placed on Truman’s decision to drop the bomb than on the men who carried it out.
So at some point, analysis fails to yield further results – precision vs. accuracy – which is why, as Gaddis notes, it’s important to analyze the right bits of data, rather than all of them.
Another problem lies in the idea of complexity / emergence: I don’t have a full mental model on it yet, but the idea is that “complex adaptive systems” like human beings, cities, or ant colonies display behaviors at one level that cannot be analyzed at another.
A classic example of this is Danish physicist Per Bak’s metaphor about a sand pile, told well by Laurence Gonzales in “ Deep Survival.” At some point, a sand pile will begin to collapse upon itself, and Gonzales observes that there’s:
“nothing in the physics of silicon dioxide that could predict the behavior of the sand pile.”
Similar problems underlie important phenomena like turbulence in pipelines and traffic on highways: given that they are highly feedback-driven processes, linear analysis like disaggregation will fail to yield a lot of useful results for modeling purposes.
One example of this comes from the aforementioned “How Doctors Think,” wherein Groopman cites a brilliant cardiologist – Dr. James Lock – who discusses a few instances in which “impeccable logic” failed to reach the right conclusions.
“My mistake was that I reasoned from first principles when there was no prior experience. I turned out to be wrong because there are variables that you can’t factor in until you actually do it.”
Lock goes on to point out that the complexity of human biology means that you can’t predict everything and often need to do real-world experimentation rather than just theoretical logic.
The takeaway here is that you have to analyze a system on a relevant scale with the relevant approach: for example, a lot of people like to throw the word “quantum” around, but it usually doesn’t mean anything. Quantum effects matter a lot if you’re making small-scale semiconductors. But they don’t do a very good job of explaining why I like chocolate, or why my shoulder hurts when I throw a football too hard.
By inversion, I believe that a lot of these phenomena come together to form what Charlie Munger calls “ lollapalooza” effects in “ Poor Charlie’s Almanack” . Basically, Munger’s proposition is that when you get a lot of mental models acting together, they act nonlinearly – so combining, say, social proof with incentives with salientfeedback is not just 3 + 3 + 3, but perhaps more like 3 x 3 x 3.
It is difficult to prove mathematically, but at least qualitatively, it seems to occur. Many of the referenced books – “Ordinary Men” on mass murder, “The Up Side of Down” on technical medical errors, or “ How Doctors Think” on cognitive medical errors – are all good to read looking for Lollapalooza-type effects.
Application / impact: we’re back to the Mauboussin quote; one damn relatedness after another: knowing the limits of analysis and understanding of causality makes us more effective at it.
Correlation vs Causality: The AI Angle
AI is a fascinating example; modern machine learning systems are built almost entirely on correlation – without understanding causation. A large language model can detect that “ice cream” often appears near “summer” and “heat” without understanding why, and yet its outputs may sound perfectly confident. It appears to reason without actually reasoning, not unlike someone who constructs elaborate justifications for their ideology but is really starting with the answer first.
In that sense, AI inherits the same storytelling bias that Don Norman and Nate Silver warned us about. Our instinct to mistake pattern for cause. When a model predicts that a person is likely to churn, or that a sentence sounds “toxic,” it is surfacing statistical shadows of past events, not the underlying mechanism that made them true. As with any black box, we can easily find ourselves trusting outputs that are right for the wrong reason. I’ve had situations personally where AI has flagged my prompts for “violating content guidelines,” because the content had words it thought were red flags (even though the context was obviously scientific). Conversely, AIs can fail to apply guardrails in obvious situations, such as a recently-publicized story in which ChatGPT provided a depressed teenager with detailed suicide instructions.
Thinking in terms of disaggregation helps us work with AI more responsibly. Instead of taking a model’s summary or score at face value, we can separate the process into smaller, knowable parts: how the data were gathered, how the prompt was framed, what examples were used, and which feedback loops are reinforcing certain responses. AI tends to work better when you give it a small task where you can verify input and output, rather than having it try to do the whole thing at once. For example, I’ve found it very helpful for structuring sections of content, summarizing content, or extending/rewording content – but if you try to get it to write a whole page at once, it doesn’t do so well. (I used AI in getting Poor Ash’s Almanack back online, but there was still a lot I had to do myself.)
Like Mendel focusing on one trait at a time, or Tetlock’s superforecasters breaking a question into smaller probabilistic steps, this approach lets us see which elements are actually causal and which are just correlated noise. It also reminds us that the right safeguards (the Swiss Cheese layers of human review, cross-model checks, and clear boundaries) matter more than any single fix.
