The morning routine that never worked
A friend of mine spent two years perfecting his morning routine. Cold shower, bulletproof coffee, ten minutes of breathwork, a specific Spotify playlist, journaling with a fountain pen. He had data: on days he completed the full ritual, his self-reported productivity score averaged 8.2 out of 10. On days he skipped any element, it dropped to 5.7. The correlation was undeniable. He wrote a blog post about it. He told everyone at dinner parties.
Then he got a puppy. The puppy destroyed the routine. He was stumbling out of bed at odd hours, skipping the cold shower, drinking whatever coffee was fastest, journaling in fragments between walks. And his productivity stayed at 8.2.
It took him three weeks to figure out what had actually changed two years earlier, when the "routine" began working: he had switched from a reactive morning — checking email and Slack first thing — to a proactive morning where he decided what to work on before anyone else's agenda reached him. The cold shower was not the variable. The breathwork was not the variable. The variable was cognitive sovereignty over the first ninety minutes of the day, and the routine had been a vehicle for it, not the cause of it. The puppy, by eliminating the phone-checking habit entirely (because the puppy needed immediate attention), preserved the actual causal factor while stripping away the correlated rituals.
He had spent two years optimizing the wrong variables because he confused what happened together with what caused what.
This is the single most common epistemic error in personal pattern recognition. You identified a pattern in L-0101 through L-0108. You named it, journaled it, tracked it. Now comes the hard question: do you actually understand why it works?
Your brain is a causal story machine
The human brain did not evolve to be a careful statistician. It evolved to survive in environments where fast causal inference — even wrong causal inference — was better than no inference at all. The rustle in the grass might be wind or it might be a predator. The ancestor who assumed causation and ran survived more often than the ancestor who waited for a controlled experiment. This is the deep evolutionary origin of what logicians call the post hoc ergo propter hoc fallacy: after this, therefore because of this.
Daniel Kahneman identified the mechanism precisely. In Thinking, Fast and Slow (2011), he described how System 1 — the fast, automatic, associative mode of cognition — generates causal stories continuously and effortlessly. You see two events in sequence and your brain fills in an arrow of causation between them before you have time to evaluate whether the arrow is warranted. "I gave that presentation and then got promoted" becomes "that presentation caused my promotion." "I started meditating and then my anxiety decreased" becomes "meditation cured my anxiety." The stories feel true because they are neurologically compelling — the brain generates them with the same machinery it uses to perceive physical causation, like one billiard ball striking another.
Nassim Taleb named this tendency the "narrative fallacy" in The Black Swan (2007). He argued that humans are constitutionally unable to look at sequences of facts without weaving causal explanations into them. The problem is not that we sometimes construct false causal stories. The problem is that we do it automatically, constantly, and without noticing — and then we treat the stories as observations rather than interpretations. You separated fact from story in L-0091. This lesson asks you to apply that same discipline specifically to causal claims.
The illusion runs deep. Leonid Rozenblit and Frank Keil demonstrated in 2002 what they called the "illusion of explanatory depth" — people consistently overestimate how well they understand causal mechanisms. Ask someone to rate their understanding of how a zipper works, then ask them to actually explain it step by step, and their confidence rating drops sharply. The same effect operates in personal patterns: you believe you understand why your good days are good, why your relationships follow certain dynamics, why you procrastinate on certain tasks. But when you are forced to articulate the actual causal mechanism — not just the correlation — the explanation often falls apart.
Three ways correlation masquerades as causation
Understanding the general principle is not enough. You need to recognize the specific structures that make spurious correlations feel like genuine causes in your own life.
The confounding variable
This is the most common structure. Two things appear connected because a third, unobserved variable is driving both. Tyler Vigen's Spurious Correlations project (2015) famously demonstrated this with absurd examples: the per capita consumption of mozzarella cheese correlates with the number of civil engineering doctorates awarded (r = 0.96). The divorce rate in Maine correlates with per capita margarine consumption. These are funny because the absurdity is obvious. But the same structure operates in your personal data in ways that are not obvious at all.
You notice that you sleep better on days you exercise. You conclude: exercise causes better sleep. But both might be caused by a third variable — your stress level. On low-stress days, you have the energy to exercise and the cognitive calm to sleep well. On high-stress days, you skip the gym and lie awake. The exercise and the sleep are both downstream effects of the same cause. If you force yourself to exercise on a high-stress day, you may find it does not fix the sleep — because exercise was never the operative variable.
The classic example in statistics is the strong correlation between children's shoe sizes and their reading ability. Bigger shoes do not make children read better. Both variables are driven by age: older children have larger feet and also read more proficiently. In your personal patterns, the "age" equivalent is often something structural — your schedule, your energy level, your social environment, your hormonal cycle — that produces multiple correlated effects without any of them causing the others.
The reverse causation
Sometimes the arrow points the other direction. You notice that you check your phone more on days when you feel anxious, and you conclude that phone use causes anxiety. But it may be that anxiety causes phone use — you reach for the device as a coping mechanism when the anxiety is already present. The intervention that follows from each interpretation is completely different. If the phone causes anxiety, you should restrict phone use. If anxiety causes phone use, you should address the anxiety directly — and phone restriction will just redirect the coping behavior to something else.
Reverse causation is particularly tricky in personal patterns because introspection is temporally imprecise. You reconstruct the sequence of events after the fact, and the reconstruction is shaped by your existing causal theory. If you believe phones cause anxiety, you will remember the phone use as preceding the anxious feeling, even if the anxious feeling was already building before you picked up the device.
The coincidental pattern
With enough data points, some patterns will appear by pure chance. If you track twenty variables daily for three months, some pairs of variables will correlate at impressive levels simply because of random variation. This is the multiple comparisons problem, and it does not just afflict scientific researchers — it afflicts anyone who journals extensively and then scans their journal for patterns. The more variables you track, the more spurious correlations you will find. Vigen's entire project is built on this principle: mine enough datasets and you will find that U.S. spending on science, space, and technology correlates with suicides by hanging, suffocation, and strangulation (r = 0.99). The correlation is real. The causation is nonexistent.
In your personal data, this shows up as the pattern that works for three weeks and then stops working — because it was never a pattern in the first place, just a coincidence that your tracking system happened to capture.
The ladder of causation: three levels of knowing
Judea Pearl, the computer scientist and philosopher whose work on causal inference earned him the Turing Award, proposed a framework in The Book of Why (2018) that transforms how we think about the relationship between correlation and causation. He called it the ladder of causation, and it has three rungs.
Rung 1: Association. This is the level of observation. You see that two things happen together. "People who exercise tend to sleep better." "On days I meditate, I am more focused." All you have is co-occurrence. Every pattern you identified in lessons L-0101 through L-0108 lives on this rung. Association is where pattern recognition begins, but it is not where understanding lives.
Rung 2: Intervention. This is the level of doing. You change one variable and observe what happens to the other. "If I start exercising, will my sleep improve?" This is fundamentally different from association, because intervention controls for confounders. If you randomly assign yourself to exercise on some days and rest on others, keeping everything else constant, and sleep improves only on exercise days, you have climbed from association to intervention. The gold standard in science — the randomized controlled trial — lives on this rung.
Rung 3: Counterfactual. This is the level of imagining. "Would my sleep have improved even if I had not exercised?" Counterfactual reasoning requires a mental model of how the world works — a causal model — not just data. It is the highest form of causal reasoning, and it is what separates genuine understanding from sophisticated correlation tracking.
Pearl's critical insight is that most of what we call "data analysis" — including most machine learning, most personal analytics, and most journaling-based pattern recognition — operates exclusively on Rung 1. You cannot climb from Rung 1 to Rung 2 by collecting more data. You climb by intervening — by changing something and observing the result. And you cannot climb from Rung 2 to Rung 3 by running more experiments. You climb by building a causal model that explains the mechanism, not just the outcome.
For personal epistemology, this means: your pattern journal (L-0108) gives you Rung 1 data. Valuable, but not causal. To get causal understanding, you need to run personal experiments — change one variable at a time and observe the effect. And to get deep understanding, you need to build a model of why the causal relationship holds, so you can predict when it will and will not apply.
Testing your causal claims: a personal protocol
Austin Bradford Hill, the British epidemiologist who helped establish the link between smoking and lung cancer, proposed in 1965 a set of nine viewpoints for evaluating whether an observed association is likely to be causal. He explicitly warned against treating them as a checklist — they are considerations, not criteria. But adapted for personal pattern analysis, five of them are immediately useful.
Strength. How strong is the association? A pattern where your productivity is 3x higher on meditation days is more suggestive of causation than one where it is 5% higher. Weak associations are more likely to be explained by confounders.
Consistency. Does the pattern hold across different contexts? If meditation improves your focus at home, at the office, and while traveling, the causal claim is stronger than if it only works in one setting. Confounders tend to be context-specific; genuine causes tend to be context-independent.
Temporality. Does the proposed cause reliably precede the effect? This is the one criterion Hill considered nearly indispensable. If you sometimes feel focused before meditating (perhaps the meditation happens because you already feel calm enough to sit still), the causal arrow may be reversed.
Experiment. When you deliberately change the proposed cause, does the effect change? This is Pearl's Rung 2. Remove the meditation for a week. Add it back. Remove it again. Track the outcome. This is the single most powerful tool for personal causal inference, and almost nobody uses it because it requires temporarily giving up something you believe is working.
Coherence. Does the causal claim make mechanistic sense? Can you articulate how the cause produces the effect? "Meditation reduces cortisol, which reduces attentional interference, which improves sustained focus" is a coherent mechanism. "I just feel better when I meditate" is an observation, not a mechanism. Coherence does not prove causation, but the absence of any plausible mechanism should make you suspicious.
The protocol in practice: pick one causal belief about your own patterns. Subject it to all five viewpoints. If it survives — if the association is strong, consistent across contexts, temporally ordered, robust to experimental manipulation, and mechanistically coherent — you have a causal claim worth building on. If it fails any of these tests, you have a correlation that needs further investigation before you optimize around it.
Correlation, causation, and your Third Brain
Here is where AI becomes a genuinely useful tool for causal reasoning — and where its limitations matter most.
A large language model can help you with Rung 1 analysis at a scale your working memory cannot match. Feed it your pattern journal and ask it to identify every correlation in the dataset. It will find associations you missed — including ones between variables you did not think to compare. It can flag potential confounders: "You noted that both your sleep quality and your exercise frequency vary with your work schedule. Have you controlled for workload when evaluating whether exercise improves sleep?" This is valuable. This is what AI does well.
But AI cannot climb the ladder for you. It cannot run your personal experiments. It cannot tell you whether exercise causes your good sleep or merely correlates with it, because that question requires intervention — changing a variable in the real world and observing the result. Machine learning systems, including the most sophisticated deep learning models, operate almost entirely on Pearl's Rung 1. They find associations in data. They are extraordinarily good at finding associations in data. But as Pearl himself argues, no amount of associational data, however large, can establish causation without either experimental intervention or strong structural assumptions about how the variables relate.
The practical implication: use AI to generate hypotheses, not conclusions. When your AI tool says "your data suggests that morning sunlight exposure correlates with better afternoon focus," treat that as a Rung 1 observation that needs Rung 2 testing. Design the experiment: get morning sunlight on some days, skip it on others, keep everything else constant, track the outcome. The AI found the signal in the noise. You have to test whether the signal is causal.
The partnership works because each side compensates for the other's limitation. You cannot scan a hundred journal entries for non-obvious correlations — the AI can. The AI cannot intervene in your life to test whether a correlation is causal — you can. Neither alone produces causal understanding. Together, you move from pattern recognition to pattern explanation.
Simpson's paradox: when aggregation lies
There is one more structure you need to recognize, because it is genuinely counterintuitive and it shows up in personal data more often than people realize.
Simpson's paradox occurs when a trend that appears in aggregated data reverses or disappears when the data is split into subgroups. The most famous example comes from a 1973 study of graduate admissions at UC Berkeley. The aggregate data showed that men were admitted at a significantly higher rate than women, suggesting gender discrimination. But when the data was broken down by department, women were admitted at equal or higher rates in nearly every department. The paradox arose because women applied disproportionately to more competitive departments with lower overall admission rates.
In personal patterns, Simpson's paradox might look like this: you track your energy levels and notice that on average, you have more energy on days you drink coffee. But when you split the data by sleep quality, a different picture emerges. On well-rested days, coffee makes no difference. On poorly-rested days, coffee gives a temporary boost followed by a crash that leaves you with less total energy. The aggregate data says coffee helps. The disaggregated data says coffee helps only in a specific context and hurts in another. The aggregate correlation is real but misleading.
The lesson: when you find a pattern in your data, always ask whether it holds across subgroups. Does the pattern hold on weekdays and weekends? During high-stress and low-stress periods? In social and solitary contexts? If a pattern reverses in a subgroup, you have found a Simpson's paradox, and the aggregate pattern is not the real story.
The bridge to second-order patterns
This lesson is not an argument against pattern recognition. It is an argument for rigorous pattern recognition — the kind that distinguishes between what happens together and what causes what, between Rung 1 associations and Rung 2 causal relationships, between aggregate correlations and subgroup-specific effects.
The patterns you identified in L-0101 through L-0108 are real observations. They are data. But data without causal structure is a collection of co-occurrences, not understanding. The transition from pattern recognition to pattern explanation is what separates someone who notices that things repeat from someone who understands why they repeat.
In L-0110, you will begin looking for second-order patterns — patterns in how your patterns behave. This requires causal reasoning, not just correlational tracking. A second-order pattern is not "these two patterns tend to co-occur." It is "this pattern reliably produces that pattern, and here is the mechanism." You cannot do second-order pattern analysis without the causal hygiene this lesson teaches, because second-order claims are causal claims by nature.
Every time you catch yourself saying "X causes Y" about your own life, pause. Ask: have I tested this at the intervention level, or am I operating on association alone? The answer will determine whether your pattern recognition is building genuine understanding or an elaborate, fragile edifice of coincidences that feel like insights.
Sources:
- Pearl, J. & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books.
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
- Taleb, N.N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House.
- Vigen, T. (2015). Spurious Correlations. Hachette Books.
- Hill, A.B. (1965). "The Environment and Disease: Association or Causation?" Proceedings of the Royal Society of Medicine, 58(5), 295-300.
- Rozenblit, L. & Keil, F. (2002). "The Misunderstood Limits of Folk Science: An Illusion of Explanatory Depth." Cognitive Science, 26(5), 521-562.
- Bickel, P.J., Hammel, E.A., & O'Connell, J.W. (1975). "Sex Bias in Graduate Admissions: Data from Berkeley." Science, 187(4175), 398-404.