Causal chains are sequences of relationships

You already think in chains. You just don't trace them far enough.

When your car won't start, you don't stare at the silent engine and declare it a mystery. You think: dead battery. Then you think: I left the headlights on. Then: the door-ajar warning was broken, so I didn't notice. Three links — outcome, proximate cause, deeper cause — and suddenly the problem isn't "the car won't start" but "the warning system failed." You've just traced a causal chain.

You do this instinctively, dozens of times a day. But here is the problem: you almost always stop too early. You find the first plausible cause, feel the satisfaction of an explanation, and move on. The headlights were on. Done. Except the real leverage — the place where you could prevent this from ever happening again — is two or three links deeper in the chain. And you never get there because the first link felt like enough.

This lesson is about following causal chains to their full depth. Not because it's intellectually interesting (though it is), but because the length of the chain you can trace determines the depth of your understanding and the precision of your interventions. Short chains produce shallow fixes. Long chains reveal mechanisms — and mechanisms are what you actually need to change outcomes.

What a causal chain actually is

A causal chain is a sequence of cause-and-effect relationships where each effect becomes the cause of the next link. A leads to B, B leads to C, C leads to D. The crucial feature is that each link is a relationship — not just an event, but a connection between two things where one produces, enables, or triggers the other.

This is where causal chains connect to everything you've been building in Phase 13. You've learned that relationships are as important as entities (L-0241), that they have directionality (L-0242) and strength (L-0243). You've mapped dependency relationships (L-0244), influence relationships (L-0245), and containment relationships (L-0246). You've seen that one entity can hold multiple relationships simultaneously (L-0248) and that relationships can be indirect, propagating through intermediaries (L-0249).

A causal chain is the most consequential form of indirect relationship. When A causes B and B causes C, A has caused C — but only through the mechanism of B. Remove B, and the connection between A and C disappears. This means that to understand any outcome, you need to trace the full sequence of intermediary relationships that produced it. Skip a link, and your understanding has a hole. Find a link that doesn't actually hold, and your entire explanation collapses.

Judea Pearl, the computer scientist and philosopher whose work on causal inference earned him the Turing Award, formalized this with what he calls the "Ladder of Causation" — three ascending levels of causal reasoning [1]. The first rung is association: observing that two things tend to occur together. The second rung is intervention: understanding what happens when you actively change something. The third rung is counterfactual: reasoning about what would have happened if things had been different. Most people live on the first rung — noticing correlations and calling them causes. Tracing a causal chain forces you onto the second and third rungs, where you ask not just "do these things co-occur?" but "does this link actually produce the next one, and what would happen if it didn't?"

The anatomy of a strong causal chain

Not all causal chains are created equal. A chain is only as strong as its weakest link — and most chains people construct have at least one link that is actually an assumption dressed up as a connection.

A strong causal chain has four properties:

Each link has a mechanism. It's not enough to say "A caused B." You need to say how A caused B. What was the process, the pathway, the transmission? When John Snow traced the Broad Street cholera outbreak in 1854, he didn't just say "the pump caused the deaths." He traced the mechanism: the cesspool of a house where an infant had cholera leaked into the well that fed the pump; residents drew water from the pump; the contaminated water carried the pathogen into their bodies [2]. Every link had a physical mechanism you could examine and verify independently.

The sequence respects temporality. Causes must precede their effects. This sounds obvious, but it's violated constantly in everyday reasoning. You might say "I failed the presentation because I was anxious," but the anxiety and the failure were happening simultaneously — the actual chain is that you didn't prepare adequately (three days before), which left you uncertain about the material (one day before), which triggered anxiety (morning of), which disrupted your delivery (during). Getting the temporal sequence right often reveals links you'd otherwise miss.

Each link is individually testable. If you can't imagine a way to verify that link A actually produces link B, that link is a hypothesis, not an established relationship. This is one of the nine considerations that Sir Austin Bradford Hill articulated in 1965 for evaluating causal evidence in epidemiology: could you, in principle, run an experiment to test this specific connection [3]? You don't always need to run the experiment. But if you can't even conceive of one, the link is likely a narrative convenience rather than a genuine causal relationship.

The chain accounts for the outcome's magnitude. If a small cause produces a large effect, there's either an amplifying mechanism in the chain (which you should identify) or you're missing links. When a startup fails, "the market shifted" is rarely sufficient as a single-link explanation for an outcome as large as company collapse. The chain needs to trace through the specific ways the market shift overwhelmed specific defenses the company did or didn't have.

How practitioners trace chains in the real world

The value of causal chains isn't theoretical. It shows up in every discipline where understanding mechanisms matters more than assigning blame.

The 5 Whys. Developed within Toyota's manufacturing system, this is the simplest chain-tracing protocol: when a problem occurs, ask "Why?" five times, with each answer becoming the subject of the next question [4]. A machine stopped. Why? The fuse blew. Why? The bearing was insufficiently lubricated. Why? The lubrication pump wasn't working properly. Why? The pump shaft was worn. Why? There was no filter, so metal scrap got in. Five links, and you've moved from "the machine stopped" (a symptom) to "no filter on the lubrication system" (a root cause you can actually fix). The discipline is in refusing to stop at the first satisfying answer.

Ishikawa diagrams. Created by Japanese chemical engineer Kaoru Ishikawa in the 1940s, these "fishbone diagrams" visualize multiple possible causal chains feeding into a single outcome [4]. Instead of tracing one linear chain, you map all the potential chains simultaneously — grouping them by category (materials, methods, machines, people, environment, measurement). The result is a map of the causal landscape, not just a single path through it. This is critical because most significant outcomes are produced by multiple chains converging, not a single chain operating alone.

Epidemiological chain of infection. In infectious disease epidemiology, the "chain of infection" is an explicit causal chain model: infectious agent, reservoir, portal of exit, mode of transmission, portal of entry, susceptible host [3]. Break any single link — kill the pathogen, eliminate the reservoir, block the transmission route, protect the host — and the chain of disease is interrupted. This is why public health interventions work at so many different levels: vaccines protect the host (last link), sanitation blocks the transmission route (middle link), and vector control eliminates the reservoir (early link). Each intervention targets a specific link in the chain.

Accident investigation. Herbert William Heinrich's "domino theory" of accident causation, proposed in 1931, modeled industrial accidents as a sequence of five falling dominoes: social environment, individual fault, unsafe act, accident, injury [5]. Remove any one domino and the sequence stops. Although the theory has been criticized for overemphasizing individual behavior and underweighting systemic factors, its core insight remains foundational to modern accident investigation: every accident is the final link in a chain, and understanding the chain is the only way to prevent the next one. Modern accident investigation frameworks like James Reason's "Swiss cheese model" extend this thinking by recognizing that chains can pass through multiple layers of defense, each with its own holes.

Where causal chains break — and where your thinking breaks with them

There are systematic ways that causal chain reasoning goes wrong. Knowing these failure patterns is as important as knowing how to trace chains correctly.

Confusing correlation with causation. Two things can co-occur without one causing the other. Ice cream sales and drowning deaths both rise in summer — not because ice cream causes drowning, but because a third factor (hot weather) drives both. When you build a causal chain, every link needs to be a genuine causal relationship, not just a co-occurrence. Pearl's do-calculus provides the mathematical framework for distinguishing these, but the practical test is simpler: if you intervened to change A, would B actually change? If you're not confident, the link is suspect.

Assuming linearity. Causal chains suggest a tidy A-to-B-to-C progression, but real-world causation is messier. Multiple chains converge on the same outcome. Single causes produce multiple effects. And — as you'll explore in the next lesson on feedback loops (L-0252) — effects can circle back to influence their own causes. A causal chain is a useful simplification, but you must remember it is a simplification. The map is not the territory. The chain is not the full causal structure.

Narrative bias. Humans are storytelling creatures. Once you start building a chain, you experience a strong pull to make it coherent — to smooth over gaps, to select links that fit the narrative, to ignore links that complicate it. Daniel Kahneman documented this extensively: the mind constructs causal stories from whatever information is available and then treats those stories as if they were the only possible explanation [6]. The antidote is to explicitly ask, at each link: "Is there an alternative cause that could produce this same effect?" If yes, your chain has a branch point you need to investigate.

The infinite regress problem. Every cause has a prior cause. The 5 Whys protocol says to ask five times, but there's nothing magical about five — you could keep going indefinitely. The practical stopping rule is to stop when you reach a link where (a) you have the ability to intervene, and (b) intervening would reliably prevent the outcome. You're not looking for the philosophical first cause. You're looking for the actionable root cause — the deepest link in the chain where you can actually do something.

What children and scientists reveal about causal chains

The ability to reason in causal chains is not something you're born with — it develops, and understanding that development illuminates both the power and the limits of causal thinking.

Jean Piaget observed that children in the preoperational stage (roughly ages 2-7) exhibit what he called "transductive reasoning" — reasoning from particular to particular without tracing through a general chain [7]. A child might conclude that because the sun went down and they got tired, the sun going down caused their tiredness. They link two events in time without tracing the mechanism.

Alison Gopnik's research at UC Berkeley has substantially revised and extended Piaget's picture. Gopnik and colleagues have shown that children as young as 24 months can infer causal relationships from patterns of evidence, and by age 3-4, children construct what Gopnik calls "causal maps" — internal representations of causal structure that function like Bayesian networks [8]. Children don't just associate events. They build models of the causal chains connecting them, and they update those models when new evidence arrives.

What's remarkable is that children employ the same basic strategy that scientists use: they intervene. Where animals learn causal associations passively (this berry made me sick), human children actively manipulate their environment to test causal hypotheses. They push buttons to see what happens. They take things apart. They deliberately vary one condition while holding others constant — crude experimentation, but experimentation nonetheless. This is causal chain reasoning in its most fundamental form: if I change this link, does the downstream effect change?

The implication for your own cognitive infrastructure is direct: you already have the wiring for causal chain reasoning. You've been doing it since you were a toddler pushing blocks off a table to see if they'd fall again. The question is not whether you can trace causal chains but how far and how rigorously you do so.

Your Third Brain: what AI gets wrong about causation

If you use AI tools for analysis, understanding, or decision support — and you should — you need to know exactly where their causal reasoning breaks down.

A June 2025 study titled "Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?" found that current LLMs are capable only of what the researchers call "shallow" (level-1) causal reasoning [9]. They can recite known causal relationships embedded in their training data — smoking causes cancer, deforestation causes habitat loss — but they cannot construct novel causal chains from unfamiliar evidence. When tested on causal chains involving news events that emerged after their training data, GPT-4o's accuracy dropped from 99.1% to 69.2%. The models weren't reasoning about causation. They were pattern-matching against memorized causal narratives.

This has a precise practical implication for your cognitive infrastructure. AI is excellent at the first step of causal chain analysis: generating candidate links. Ask an LLM to brainstorm possible causes of an outcome, and you'll get a comprehensive list — often more comprehensive than what you'd generate alone. Where it fails is at evaluating those links. It cannot reliably tell you whether a proposed causal link actually holds in your specific context. It cannot distinguish a genuine mechanism from a plausible-sounding narrative. And it cannot climb Pearl's ladder from association to intervention to counterfactual.

The practical protocol: use AI to generate the widest possible set of candidate causal links. Then apply your own judgment — informed by evidence, testability, and mechanism — to evaluate which links actually hold. The chain itself must be yours. The raw materials can come from anywhere.

This is not a limitation that future models will necessarily resolve. Pearl himself has argued that causal reasoning requires a model of the world that goes beyond statistical patterns in data [1]. Until AI systems have genuine causal models — not just correlations extracted from text — they will remain powerful tools for generating hypotheses and weak tools for validating causal chains.

Protocol: The five-link trace

Here is the operational protocol for tracing causal chains in your own life and work. Use this whenever you encounter an outcome — positive or negative — that you want to understand deeply enough to reproduce or prevent.

1. Name the outcome. Write it at the bottom of a page or document. Be specific: not "the project failed" but "the project delivered three weeks late and 20% over budget."

2. Ask the first Why. What directly caused this outcome? Write it as a relationship: "X caused Y because [mechanism]." The mechanism is non-negotiable. If you can't state the mechanism, the link is a guess.

3. Chain four more links. For each cause you identify, ask again: what caused this? Continue for a total of five links. If you hit a dead end before five, branch: are there multiple causes at this level? Map them.

4. Test each link for legitimacy. For every link, ask three questions: (a) Did the cause precede the effect in time? (b) If I had removed this cause, would the effect plausibly not have occurred? (c) Can I state a specific mechanism connecting them? Any link that fails two of these three tests should be flagged as weak.

5. Identify the leverage point. Scan your chain for the deepest link where you have both the ability and the authority to intervene. This is your highest-leverage action point. A fix at link 5 prevents the entire chain from firing. A fix at link 1 only addresses the symptom.

6. Act on the leverage point. Make one concrete change at that link. Document what you changed and why. Then observe: does the downstream chain still fire? If it does, your chain model was incomplete and you need to trace again.

The discipline is not in the asking — anyone can ask "why?" once. The discipline is in asking five times, insisting on mechanisms at every link, and refusing the easy satisfaction of the first plausible explanation.

From chains to circles

You've now added a powerful tool to your relationship mapping capability: the ability to trace a sequence of causal relationships from an outcome back through its full mechanism. You can identify root causes, find leverage points, and distinguish genuine causal links from narrative conveniences.

But there's a structure that causal chains, by themselves, cannot capture. In every chain you've traced so far, causation flows in one direction — from cause to effect, from early to late, from root to outcome. The chain has a beginning and an end.

What happens when the end of the chain loops back to the beginning? When effect D influences cause A, which produces B, which produces C, which produces D again? You no longer have a chain. You have a loop — a self-reinforcing or self-correcting cycle where the outcome feeds back into its own cause.

That's the territory of the next lesson. In L-0252, you'll learn that feedback loops are circular relationships — and that when a causal chain bends back on itself, entirely new dynamics emerge that linear thinking cannot predict.

Chains explain mechanisms. Loops explain why some mechanisms accelerate and others stabilize. You need both.

Sources

[1] Pearl, J. & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books. Pearl's "Ladder of Causation" distinguishes three levels: association, intervention, and counterfactual reasoning.

[2] Snow, J. (1855). On the Mode of Communication of Cholera. Second Edition. London: John Churchill. See also: Tulchinsky, T.H. (2018). "John Snow, Cholera, the Broad Street Pump; Waterborne Diseases Then and Now." Case Studies in Public Health, 77-99.

[3] Hill, A.B. (1965). "The Environment and Disease: Association or Causation?" Proceedings of the Royal Society of Medicine, 58(5), 295-300. The nine criteria (strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, analogy) remain foundational to epidemiological causal inference.

[4] Ishikawa, K. (1976). Guide to Quality Control. Asian Productivity Organization. For the integration of Ishikawa diagrams with the 5 Whys technique, see: iSixSigma, "Root Cause Analysis: Integrating Ishikawa Diagrams and the 5 Whys."

[5] Heinrich, H.W. (1931). Industrial Accident Prevention: A Scientific Approach. McGraw-Hill. For modern critique and extension, see: Reason, J. (1990). Human Error. Cambridge University Press.

[6] Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. Particularly chapters on the "narrative fallacy" and the tendency to construct coherent causal stories from limited evidence.

[7] Piaget, J. (1930). The Child's Conception of Physical Causality. Harcourt, Brace & Company.

[8] Gopnik, A., Glymour, C., Sobel, D.M., Schulz, L.E., Kushnir, T., & Danks, D. (2004). "A Theory of Causal Learning in Children: Causal Maps and Bayes Nets." Psychological Review, 111(1), 3-32. See also: Goddu, M.K. & Gopnik, A. (2024). "The Development of Human Causal Learning and Reasoning." Nature Reviews Psychology.

[9] "Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?" (2025). arXiv:2506.21215. Found that LLMs perform shallow causal reasoning through parameter memorization rather than genuine structural causal inference.