Correlation vs Causation Mental… | Faster Than Normal
Mathematics & Probability
Correlation vs Causation
The critical distinction between two variables that move together and one actually causing the other — confusing the two is among the most common reasoning errors.
Model #0086Category: Mathematics & ProbabilityDepth to apply:
Two things can move together without one causing the other. That sentence contains more decision-making power than most MBA curricula.
Correlation measures the statistical relationship between two variables — when one moves, the other tends to move in a predictable direction. Causation means one variable actually produces a change in the other through a identifiable mechanism. The gap between these two concepts is where fortunes are made and lost, where policies succeed or catastrophically backfire, and where the difference between a rigorous thinker and a confident fool becomes measurable.
The human brain is a causation-manufacturing machine. Evolution wired it that way. A rustle in the grass correlated with a predator attack often enough that our ancestors who inferred causation — and ran — survived to reproduce. The ones who paused to design a controlled experiment did not. The bias served its purpose on the savanna. In a world of complex systems, multivariate data, and trillion-dollar decisions, it's a liability that scales with the stakes.
The formal study of correlation began with Francis Galton in the 1880s. Galton — Charles Darwin's half-cousin, a polymath who also invented the weather map and fingerprint classification — noticed that tall parents tended to have children who were tall, but less extremely so. He called this "regression toward mediocrity" (now regression to the mean) and developed the concept of correlation to quantify the relationship. His protégé Karl Pearson formalised the mathematics in 1896 with the Pearson correlation coefficient, still the most widely used measure of linear association. Pearson was meticulous about the math and almost completely uncritical about its interpretation — a pattern that has persisted for 130 years.
The problem isn't the statistic. The problem is what people do with it. A correlation coefficient of 0.85 between advertising spend and revenue looks like proof that advertising drives sales. It might be. Or revenue growth might drive advertising budgets — companies spend more on marketing when they have more cash. Or a third variable — say, a booming economy — might independently drive both. Or the correlation might be entirely coincidental, an artefact of small samples and the human appetite for pattern.
Tyler Vigen's "Spurious Correlations" project catalogued hundreds of these accidents: the divorce rate in Maine correlates with per capita margarine consumption (r = 0.99). Nicolas Cage film appearances correlate with swimming pool drownings (r = 0.67). These are absurd — and that's the point. With enough variables and enough time periods, you will find correlations between anything and everything. The universe of possible variable pairs is effectively infinite. The subset that share a causal mechanism is vanishingly small.
The philosophical foundation runs deeper than statistics. David Hume argued in 1739 that causation itself is unobservable — we never see one billiard ball cause another to move; we see one event consistently followed by another and infer causation from the regularity. Hume's "problem of induction" remains unsolved in philosophy. In practice, science bypassed the philosophical impasse by developing methods — randomised controlled trials, instrumental variables, natural experiments, Bradford Hill's criteria — that don't prove causation with logical certainty but establish it with sufficient confidence to act.
Austin Bradford Hill articulated the most influential framework in 1965, proposing nine criteria for evaluating whether an observed association is likely causal: strength of association, consistency across studies, specificity, temporality (the cause must precede the effect), biological gradient (dose-response), plausibility, coherence with existing knowledge, experimental evidence, and analogy. No single criterion is sufficient. No single criterion is necessary. The framework is a structured way of asking: how many independent reasons do we have to believe this correlation reflects a real mechanism?
The practical implications are immediate and consequential. Every metric dashboard in every company displays correlations. Revenue correlates with headcount. Customer satisfaction correlates with feature count. Employee engagement correlates with free lunch quality. The question that separates useful analysis from expensive noise is always the same: is this relationship causal, and if so, in which direction? Get the direction wrong and you hire people to generate revenue that was actually driven by product-market fit. You build features to increase satisfaction that was actually driven by customer success onboarding. You invest in perks to boost engagement that was actually driven by meaningful work.
Consider the most expensive correlation-causation confusion in modern financial history. In the years preceding the 2008 crisis, credit rating agencies relied on models that assumed housing prices in different geographic regions were largely uncorrelated — because historically, they had been. National diversification, the models implied, caused portfolio risk reduction. The correlation structure was real in the historical data. The causal claim — that geographic diversity mechanically reduced default risk — was wrong. The underlying cause of the apparent decorrelation was that no national-level housing shock had occurred in the sample period. When subprime lending practices created systemic risk across all regions simultaneously, the decorrelation vanished. AAA-rated tranches built on the assumption of causal independence lost 60–80% of their value. The total cost exceeded $2 trillion in losses and triggered the deepest recession since the 1930s. The models were sophisticated. The causal reasoning was absent.
The cost of confusing correlation with causation compounds silently. Each misattributed cause generates a misallocated resource. Each misallocated resource produces a result that's slightly worse than expected. The shortfall gets explained by another spurious correlation, which generates another misallocation. The cycle is self-reinforcing, invisible from the inside, and often lethal to organisations that mistake data-richness for understanding.
Section 2
How to See It
The confusion between correlation and causation is so pervasive that the hard part isn't finding examples — it's finding domains where the error isn't present. Once you train yourself to ask "but is it causal?" after every claimed relationship, you'll find that most confident assertions about what drives what are untested correlations dressed in the language of mechanism:
Investing
You're seeing Correlation vs Causation when an analyst argues that a company's stock price rose 30% following a new product launch, therefore the product launch drove the stock price. The S&P 500 might have risen 25% over the same period. The sector might have been reclassified by index funds. A short squeeze might have coincided with the announcement. The temporal correlation is real. The causal claim requires controlling for every alternative explanation — and in public markets, that's nearly impossible. Jim Simons built Renaissance Technologies on the discipline of distinguishing genuine statistical signals from the thousands of spurious correlations that surface in any sufficiently large dataset.
Technology
You're seeing Correlation vs Causation when a growth team reports that users who complete onboarding within 24 hours retain at 3x the rate of those who don't, and concludes that faster onboarding causes better retention. The direction might be reversed: highly motivated users — who would have retained anyway — happen to complete onboarding faster because they're already engaged. Or a confound might explain both: users acquired through a specific high-intent channel complete onboarding faster and retain better, with the channel driving both outcomes. Facebook's early "7 friends in 10 days" metric faced exactly this challenge — the correlation was robust, but establishing that friend connections caused retention rather than reflecting pre-existing engagement required years of experimentation.
Medicine
You're seeing Correlation vs Causation when observational studies show that people who take vitamin supplements live longer than those who don't, and the supplement industry cites this as evidence of efficacy. The confound is socioeconomic: people who buy supplements tend to be wealthier, better educated, more health-conscious, and have better access to medical care. When randomised controlled trials — which neutralise confounders by random assignment — tested the same supplements, most showed no benefit and some showed harm. The Women's Health Initiative, a landmark $625 million RCT launched in 1991, found that hormone replacement therapy, previously "supported" by decades of observational correlations showing reduced heart disease, actually increased cardiovascular risk. The observational correlation had been real. The causal conclusion was lethal.
Personal life
You're seeing Correlation vs Causation when someone notices that their most productive weeks coincide with early morning wake-ups and concludes that waking early causes productivity. The relationship might run in reverse: high-energy periods produce both early waking and high output. Or a third factor — reduced social obligations, a compelling project, better sleep quality — might drive both. The "4 AM club" productivity literature is built almost entirely on survivor-bias-laden correlations mistaken for causal mechanisms. The people writing those books were productive before they started waking at 4 AM. The alarm clock gets credit for the discipline.
Section 3
How to Use It
Decision filter
"Before accepting any claimed relationship between two variables, ask three questions: Is there a plausible mechanism? Could the causation run in the opposite direction? Is there a third variable that could explain both? If you can't eliminate at least two of the three alternatives, you don't have a causal claim — you have a hypothesis."
As a founder
Every metric your team celebrates is a correlation until proven otherwise. Monthly active users correlate with revenue. NPS correlates with retention. Feature velocity correlates with growth. The founder's job is to identify which relationships are causal and which are coincidental — because resources allocated to non-causal relationships are resources burned.
The operational discipline: run experiments wherever possible. A/B tests are miniature randomised controlled trials. They isolate the effect of one variable by holding everything else constant. When experiments aren't feasible, use natural experiments — a pricing change in one market but not another, a feature rolled out to one cohort first. Jeff Bezos built Amazon's culture around "input metrics" — controllable actions believed to cause desired outcomes — and insisted on testing the causal link rather than assuming it. The distinction between an input metric and a vanity metric is the distinction between correlation and causation applied to business operations.
As an investor
The most expensive error in investing is mistaking a correlation for a competitive advantage. A company's revenue grew 40% in a year when its industry grew 38%. The company's management presents the 40% as evidence of superior execution. The incremental 2% might be noise. The 38% was the tide. Confusing the two leads to paying a premium multiple for an asset that's riding a macro wave — and discovering the error only when the wave recedes.
Charlie Munger's insistence on understanding the mechanism behind a business's success — not just the statistical track record — is fundamentally a demand for causal evidence. Why does this company win customers? Not "customers correlate with marketing spend" but "this specific feature solves a specific problem that no competitor addresses, which produces switching costs that compound." The mechanism is the causal claim. The correlation is just the symptom. Every durable investment thesis rests on a causal mechanism that survives when the macro environment changes.
As a decision-maker
When presented with data supporting a recommendation, ask the presenter to explain the mechanism — the specific pathway through which X produces Y. If they can only point to a correlation ("companies that do X tend to see Y"), push for the next level: why? What's the chain of events? Where could it break?
The most dangerous presentations are the ones with beautiful scatter plots and high R-squared values that explain nothing about causation. A perfect correlation between two variables tells you they move together. It tells you nothing about what happens when you intervene on one of them. Intervening on a correlated-but-not-causal variable doesn't just fail to produce the expected result — it consumes resources and creates false confidence that the problem is being addressed, while the actual cause continues unchecked.
Common misapplication: Demanding proof of causation before acting on any correlation. In practice, perfect causal evidence is rare, slow, and expensive. Bradford Hill himself warned against requiring proof: "All scientific work is incomplete. That does not confer upon us a freedom to ignore the knowledge we already have, or to postpone the action that it appears to demand." The goal isn't to reject every correlation — it's to hold correlational evidence with appropriate uncertainty, invest in establishing causation where the stakes justify the cost, and avoid building strategy on relationships you've never tested.
A second trap: assuming that because a specific correlation is spurious, the entire dataset is uninformative. Correlations are genuinely useful as starting points for causal investigation. They tell you where to look. They just can't tell you what you'll find.
Section 4
The Mechanism
Section 5
Founders & Leaders in Action
The correlation-causation distinction isn't theoretical for the people below. Each built a career on the ability to identify genuine causal mechanisms in domains saturated with spurious correlations — and each paid meaningful tuition, in dollars or in reputation, for the times they got the distinction wrong before building the systems that got it right.
The pattern across these cases is consistent: the competitive advantage wasn't in seeing more data. It was in asking better questions about the data everyone could see. The question "is this causal?" is the single highest-return-on-investment analytical habit available to any decision-maker. These leaders proved it across mathematics, investing, technology, and physics.
Jim SimonsFounder, Renaissance Technologies, 1988–2020
The central challenge of quantitative finance is that financial data generates an almost infinite supply of correlations, and the vast majority are spurious. Run enough regressions on enough variable pairs across enough time periods and you will find patterns everywhere — patterns that look robust in backtests and disintegrate the moment real capital is deployed. The industry term is "overfitting." The conceptual error is mistaking correlation for causation.
Simons — a former NSA codebreaker who had worked on pattern recognition in classified signals intelligence — understood this problem at a level most financial practitioners did not. Renaissance Technologies' Medallion Fund, which averaged roughly 66% gross annual returns over three decades, was built on a rigorous process for separating signal from noise. The firm hired physicists, mathematicians, and computational linguists rather than finance professionals precisely because the core skill was statistical discrimination, not market intuition.
The discipline was specific. Renaissance reportedly tested patterns against multiple independent datasets and time periods before allocating capital to any strategy. They demanded that patterns be persistent (present across different market regimes), non-obvious (not already captured by known factors), and statistically significant by standards far more stringent than the p < 0.05 threshold common in academic finance. A pattern that appeared in backtested data but lacked a plausible generating mechanism was treated with extreme scepticism.
Robert Mercer, co-CEO from 2010 to 2017, reportedly told new employees that for every genuine signal Renaissance identified, the team had rejected hundreds of equally compelling-looking correlations that turned out to be noise. The Fund's sustained performance — in a domain where most quantitative strategies decay within years — suggests the filtering process worked. The lesson isn't that correlations are useless in finance. It's that the ratio of spurious to genuine correlations is so high that without institutional discipline, you'll mistake the noise for the signal every time.
Charlie MungerVice Chairman, Berkshire Hathaway, 1978–2023
Munger's investing framework rested on a demand that would have embarrassed most Wall Street analysts: explain the mechanism. Not the correlation, not the regression output, not the historical pattern — the causal chain through which a business generates and sustains profits.
When Munger evaluated Coca-Cola in 1988, he didn't start with the stock price correlation to earnings growth. He started with the psychology of brand preference, the thermodynamics of distribution logistics, the economic structure of concentrate pricing, and the regulatory moat of a century-old trademark. Each element was a causal claim, testable against evidence. The brand produces repeat purchase behaviour. The distribution system produces retail availability. The concentrate model produces gross margins. The trademark produces legal protection. The stock price was a downstream effect of these mechanisms, not a cause of anything.
The contrast with correlation-based investing is instructive. Quantitative screens that select stocks based on low P/E ratios, momentum, or earnings surprises capture correlations that have historically predicted returns. Munger didn't reject these correlations — he acknowledged that cheap stocks, on average, outperform expensive ones. But he insisted that the average tells you nothing about the specific stock in front of you. A stock might be cheap because the market is wrong (causal opportunity) or because the business is deteriorating (causal trap). Only the mechanism distinguishes the two.
"I never allow myself to have an opinion on anything that I don't know the other side's argument better than they do," Munger told a Berkshire shareholder meeting. That's a causal-reasoning demand dressed as intellectual humility. Understanding the other side's argument requires understanding the mechanism they believe in — and then testing whether that mechanism or yours better explains the evidence.
Bezos built Amazon's operating system around a distinction that maps directly onto correlation versus causation: input metrics versus output metrics. Output metrics — revenue, profit, stock price — are results. They correlate with countless variables. Input metrics — selection breadth, delivery speed, page load time, price competitiveness — are controllable actions that Amazon believed caused the desired outputs through specific mechanisms.
The distinction mattered operationally. A team optimising for an output metric can always find a correlation to claim credit for results that were actually driven by something else — seasonal demand, a competitor's stumble, a marketing campaign from a different team. A team optimising for an input metric has to demonstrate the causal mechanism: we reduced page load time by 100 milliseconds, which produced a measurable increase in conversion rate, which generated incremental revenue. The mechanism is specified. The claim is testable.
Bezos's annual shareholder letters repeatedly emphasised this framework. The 2005 letter distinguished between "math-based decisions" where causal relationships could be quantified and "judgment-based decisions" where they could not — and insisted that the company maximise the former category. Amazon's internal culture of "working backwards" from the customer press release is itself a causal-reasoning exercise: start with the desired effect (customer delight), then specify the mechanism that would produce it, then build the mechanism. The approach is the opposite of pattern-matching to historical correlations and assuming they'll persist.
Feynman's 1974 commencement address on "Cargo Cult Science" is the most precise articulation of the correlation-causation error applied to institutional practice. The metaphor came from Melanesian islanders who, after World War II, built replica control towers and runways from bamboo, mimicking the forms they'd observed at military airfields — believing that the physical structures had caused the planes to land. The structures correlated with plane arrivals during the war. The structures did not cause them.
Feynman used the metaphor to diagnose a failure mode in science itself: researchers who replicate the superficial forms of rigorous methodology — lab equipment, statistical significance, peer-reviewed publication — without the underlying causal discipline. "They follow all the apparent precepts and forms of scientific investigation, but they're missing something essential." The missing element was precisely the willingness to test causal claims rather than accepting correlational evidence.
His most devastating example: educational research that correlated teaching methods with student outcomes without controlling for student selection, teacher quality, socioeconomic background, or measurement validity. The field produced thousands of studies, hundreds of policy recommendations, and no reliable causal knowledge about what actually made students learn. The correlations were everywhere. The causation was nowhere — because nobody had designed the experiments that could distinguish the two.
Feynman applied the same logic to the Challenger disaster investigation in 1986. NASA management had observed that O-rings showed damage on some flights but no catastrophic failure had occurred — and concluded the damage was acceptable. The correlation between O-ring damage and non-catastrophic outcomes was real across the available sample. The causal mechanism — that O-ring resilience degraded with temperature, and that sufficiently cold conditions would produce catastrophic failure — required engineering analysis that management declined to perform. They had data. They had correlations. They didn't have causation. Seven astronauts died.
Thorp's career is a masterclass in distinguishing genuine causal relationships from statistical noise — first in gambling, then in financial markets. His 1962 book Beat the Dealer identified a causal mechanism in blackjack: the composition of the remaining deck causally determines the player's expected value, and tracking dealt cards updates that expectation in real time. This wasn't a correlation between card counting and winning. It was a mathematical proof that the information in the dealt cards causes a shift in edge from house to player under specific distributional conditions. The casinos didn't ban Thorp because he got lucky. They banned him because his causal model was correct.
When Thorp moved to financial markets with Princeton Newport Partners in 1969, the discipline transferred directly. Options markets in the 1970s were priced using rules of thumb and historical correlations between volatility and price. Thorp developed a theoretically grounded options pricing model — independently of and slightly before Black and Scholes — that specified the causal relationship between an option's value and its underlying variables: stock price, time to expiration, volatility, and the risk-free rate. When the market price diverged from the model price, Thorp didn't merely observe a correlation between mispricing and profit. He understood the causal mechanism — arbitrage convergence — that would close the gap.
The distinction mattered in risk management. Correlation-based traders blow up when correlations shift. Mechanism-based traders like Thorp sized positions based on the strength of the causal relationship, not the historical pattern. Princeton Newport Partners delivered 19% annualised returns over nearly two decades with minimal drawdowns — a record built on causal understanding, not pattern-matching.
Section 6
Visual Explanation
Section 7
Connected Models
Correlation versus causation doesn't operate in isolation — it intersects with models that either sharpen causal reasoning, exploit the confusion, or represent the downstream consequences of getting the relationship wrong. Some models reinforce the discipline of causal thinking by providing complementary frameworks. Others create productive tension by revealing how cognitive biases convert correlations into false causal beliefs. And some represent the natural next step once genuine causal understanding is established.
Understanding the web of connections tells you where the error hides and which tools can surface it:
Reinforces
Bayes' Theorem
Bayes' Theorem provides the mathematical machinery for updating causal beliefs as new evidence arrives. When you observe a correlation and want to know whether it's causal, you're implicitly asking a Bayesian question: given this correlation (evidence), how much should I update my prior belief that A causes B? The likelihood ratio depends on how surprising the correlation would be if the relationship were causal versus if it were spurious. Strong, dose-dependent, temporally consistent correlations carry high likelihood ratios for causation. Weak, inconsistent, or easily confounded correlations carry low ones. Bayesian updating quantifies what Bradford Hill's criteria approximate qualitatively.
Reinforces
Scientific Method
The scientific method is the institutionalised process for converting correlations into causal knowledge — or rejecting them. Hypothesis formation identifies a candidate causal relationship. Experimental design controls for confounders. Replication tests whether the relationship persists across different conditions. Peer review subjects the causal claim to adversarial scrutiny. Each step addresses a specific failure mode of naive correlational reasoning: confounding, reverse causation, sample bias, researcher degrees of freedom. The correlation-causation distinction is the central problem the scientific method was designed to solve.
Tension
[Narrative](/mental-models/narrative) Fallacy
The narrative fallacy — Nassim Taleb's term for the compulsion to construct coherent stories from random data — is the primary vehicle through which correlations get promoted to causal status without evidence. A compelling story ("the CEO's new strategy drove the turnaround") feels like a causal explanation, but it's often a post-hoc narrative draped over a correlation. The narrative provides the mechanism that the data lacks — but the mechanism is fictional, constructed to fit the pattern rather than tested against it. Every time a business book attributes a company's success to a specific practice without controlling for survivorship bias and confounding variables, the narrative fallacy is converting correlation into mythology.
Section 8
One Key Quote
"We have no other notion of cause and effect, but that of certain objects, which have been always conjoined together. We cannot penetrate into the reason of the conjunction."
— David Hume, A Treatise of Human Nature (1739)
Section 9
Analyst's Take
Faster Than Normal — Editorial View
This is one of the few mental models where the concept is simple, universally taught, and almost universally ignored in practice. Every statistics student learns that correlation doesn't imply causation. Then they spend the rest of their career acting as if it does.
The reason is structural, not intellectual. Causal evidence is expensive. Randomised experiments take time, money, and organisational commitment. Observational correlations are cheap — they're sitting in your database right now, available to anyone with SQL access and a hypothesis. The incentive gradient points sharply toward correlational analysis dressed in causal language. A dashboard that says "users who do X retain at 3x" is actionable and impressive in a slide deck. A dashboard that says "we don't know if X causes retention" is accurate and career-limiting.
The most destructive form of the error isn't in academic research or public policy — it's in business metrics. Every company I've seen operates on a set of assumed causal relationships between inputs and outputs that have never been experimentally validated. Marketing spend "drives" leads. Feature releases "drive" engagement. Hiring "drives" throughput. These relationships are almost always correlational, frequently confounded, and occasionally pointing in the wrong causal direction entirely. The company that actually tests its assumed causal relationships — through A/B experiments, holdout groups, or natural experiments — discovers that roughly half of them are weaker than assumed and some are nonexistent.
The finding isn't demoralising. It's liberating. If half your assumed causal relationships are wrong, then half your resource allocation is at least partially wasted. Identifying which half frees enormous resources for redeployment to the relationships that actually hold. Jeff Bezos understood this — Amazon's culture of experimentation is fundamentally a culture of causal verification. The company runs thousands of experiments per year not because it loves statistical rigour for its own sake, but because every validated causal relationship translates directly into more efficient capital allocation.
The most subtle version of the error is confusing a proxy with a mechanism. College degrees correlate with professional success. Companies that screen for degrees believe they're screening for competence. But the degree might be a proxy for family wealth, social networks, conscientiousness, or signalling ability — none of which require the specific knowledge the degree supposedly certifies. Peter Thiel's fellowship programme, which paid students to drop out of college and start companies, was a direct test of the causal claim. The results — multiple billion-dollar companies founded by fellows — didn't disprove that college causes success. They demonstrated that the correlation between degrees and success was substantially confounded by selection effects that had nothing to do with the education itself.
Section 10
Test Yourself
The ability to spot the correlation-causation error in the wild is a skill that improves with deliberate practice. The error rarely presents itself in obvious form — nobody argues that Nicolas Cage films cause drowning deaths. In practice, the correlations that get mistaken for causation are plausible, supported by data, and endorsed by credible people. The test of the model is whether you can distinguish the error when it's dressed in authority.
These scenarios are designed to test not just whether you can identify the error, but whether you can distinguish between situations where the error is present and situations where genuine causal reasoning is at work.
Is this mental model at work here?
Scenario 1
A health study finds that people who drink moderate amounts of red wine have lower rates of heart disease than non-drinkers. A wine industry trade group cites the study as evidence that red wine prevents heart disease and recommends daily consumption.
Scenario 2
A SaaS company runs an A/B test where 50,000 randomly assigned users see a simplified onboarding flow and 50,000 see the original. The simplified flow produces a 12% increase in 30-day retention with p < 0.001. The company rolls out the simplified flow to all users.
Scenario 3
A management consultant presents data showing that companies with diverse executive teams have 35% higher profitability than companies with homogeneous teams. She recommends that a client prioritise executive diversity as a profitability strategy.
Scenario 4
A city installs speed cameras on a stretch of highway where fatal accidents have spiked. In the following year, fatal accidents drop 40%. The transportation department credits the cameras and installs them citywide.
Section 11
Top Resources
The essential reading spans philosophy, statistics, economics, and applied decision-making. The field has produced some of the clearest writing in all of science — partly because the stakes of getting the distinction wrong are so visible across medicine, policy, and finance. Start with Pearl for the formal framework, Kahneman for the psychology, and Angrist & Pischke for the applied toolkit.
Pearl's framework for causal inference — directed acyclic graphs, the do-calculus, and the "ladder of causation" — is the most rigorous modern treatment of the correlation-causation distinction. The book makes the mathematics accessible without sacrificing precision. Essential for anyone who works with data and wants to understand the formal boundary between correlation and causation.
Kahneman's chapters on regression to the mean (Chapter 17) and the illusion of understanding (Chapter 19) explain the psychological machinery that drives causal inference from correlational data. His treatment of System 1's automatic causal storytelling is the clearest account of why the error is so persistent and so difficult to override through awareness alone.
The practical handbook for extracting causal estimates from observational data. Angrist and Pischke's treatment of instrumental variables, regression discontinuity, and difference-in-differences is the toolkit behind the credibility revolution in economics. Technical but indispensable for anyone who needs to make causal claims from non-experimental data.
Vigen's collection of absurd but statistically robust correlations — per capita cheese consumption and death by bedsheet entanglement, for instance — is the most visceral demonstration of why correlation alone proves nothing. The project is entertaining and pedagogically powerful: if you can find a 0.95 correlation between random variables, the statistic itself carries zero causal information.
The foundational framework for evaluating whether an observed association is likely causal. Hill's nine criteria — strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, and analogy — remain the most widely used checklist in epidemiology and have broad applicability to any domain where causal claims are drawn from observational evidence. Short, clearly written, and as relevant today as when published.
Correlation vs Causation — Why two variables moving together tells you nothing about the mechanism connecting them
Tension
Confirmation Bias
Confirmation bias selectively surfaces correlations that support existing beliefs and suppresses those that contradict them — then treats the surviving correlations as causal evidence. An investor who believes a sector is overvalued will notice every correlation between current conditions and past bubbles while ignoring correlations between current conditions and sustained growth periods. The bias doesn't just distort which correlations you see. It distorts which causal stories you construct from them. The combination produces confident causal claims built on cherry-picked associations — the most dangerous form of the correlation-causation error because it feels rigorous.
Leads-to
Second-Order Thinking
Understanding true causal relationships — rather than surface correlations — is the prerequisite for second-order thinking. If you know that A actually causes B (not just correlates with it), you can reason about what B causes in turn, and what those downstream effects produce. If you're working from a spurious correlation, every second-order inference inherits the original error and compounds it. Causal reasoning is the foundation that makes downstream inference reliable. Without it, second-order thinking becomes second-order speculation — elaborate chains of reasoning built on a false premise.
Leads-to
Map vs. Territory
Confusing correlation with causation is one of the most common ways that maps diverge from territory. Your mental model of how the world works — your map — is built from observed relationships. When those relationships are causal, the map predicts accurately: intervene on A and B changes as expected. When the relationships are merely correlational, the map looks accurate in normal conditions but fails catastrophically when you act on it. The subprime mortgage crisis demonstrated this at civilisational scale: risk models built on historical correlations between housing prices and default rates produced maps that matched the territory beautifully — until the underlying causal structure shifted, and the map became worthless overnight.
The operational discipline I'd recommend to any decision-maker: maintain a list of the ten most important assumed causal relationships in your business. For each one, write down the specific mechanism through which you believe the cause produces the effect. Then write down the two most plausible alternative explanations — reverse causation and confounding. If you can't design a test that would distinguish between your causal claim and the alternatives, you're operating on faith. Faith is fine for personal beliefs. It's expensive in resource allocation.
One final observation that most treatments of this topic miss: the correlation-causation distinction is itself subject to diminishing returns. At some point, demanding more causal evidence becomes a form of analysis paralysis. The pharmaceutical industry's gold-standard RCT process takes 10–15 years and $1–2 billion per drug, partly because the causal evidence threshold is set extremely high. That's appropriate when the stakes include patient safety. It's inappropriate when the stakes are a landing page redesign. The skill isn't applying the same rigour everywhere — it's calibrating the level of causal evidence you demand to the reversibility and magnitude of the decision you're making. Irreversible, high-stakes decisions demand strong causal evidence. Reversible, low-stakes decisions can proceed on correlational evidence with a plan to validate causation later. Bezos's framework of Type 1 versus Type 2 decisions maps neatly onto this: Type 1 decisions (irreversible) demand causal understanding. Type 2 decisions (reversible) can tolerate correlational hypotheses.
The AI and machine learning revolution has made this worse, not better. Modern ML systems are correlation engines of extraordinary power. A neural network trained on customer data will find every correlation in the dataset — including the spurious ones — and weight them all into its predictions. The model works beautifully on the training data and often on new data drawn from the same distribution. But intervene based on the model's correlational outputs — change a feature it identified as predictive — and you may find the prediction breaks, because the feature was correlated with the outcome but didn't cause it. The most sophisticated prediction systems in history are also the most vulnerable to the correlation-causation error at scale. The organisations that understand this limitation will use ML for prediction and experiments for causation. The ones that don't will deploy models that optimise for the wrong variables with extraordinary precision.
The mental model isn't "never act on correlations." It's "know when you're acting on a correlation, size your commitment accordingly, and build the feedback loop to discover whether the causation holds before you've bet the company on it."
