Make the Model Do Something

An analogy can make a system feel familiar, but the next question is less comfortable: what actually makes what happen? A story may describe events in the right order and still leave the cause untouched. A label may name a pattern without explaining it. To understand an outcome, trace the influence that produces it—and treat that trace as part of a model to test, not a conclusion in itself.

Consider a familiar summertime puzzle, used here as a causal illustration rather than as evidence from a named dataset: in some places, ice-cream sales and drownings rise together. The numbers move in tandem. Does ice cream cause drownings? The correlation alone cannot tell us. It shows that two things vary together, not that one produces the other.

Look for the missing link. Hot weather encourages people to buy ice cream. It also sends more people to beaches and pools, where more swimming creates more opportunities for drowning. The heat helps explain both increases. The causal sketch is not ice cream → drowning, but hot weather → more ice cream sales and hot weather → more swimming → more drownings.

That difference matters because a description can masquerade as an explanation. “Drownings rise when ice-cream sales rise” reports a relationship. “Summer brings heat, more swimming, and greater exposure to water” begins to explain why the relationship appears. And “summer” is not a magical cause by itself: it is a convenient label for conditions—temperature, daylight, school schedules, habits—that may influence what people do.

When you meet a claim about cause, try drawing the chain in plain language. What changes first? Through what intermediate steps could it affect the outcome? What other conditions must be present? In the summer example, heat does not create drownings without people entering water; swimming does not make every swimmer drown. The outcome depends on a chain of influences and circumstances, not a single switch.

This is also why sequence is not enough. If one event happens before another, it may be a cause, but “before” does not prove “because of.” A rooster crows before sunrise; its crowing does not bring up the sun. To explain, you need a mechanism: some account of how an influence travels through the system and changes the result.

The useful question is not merely, “What happened alongside this?” It is, “What would have to be true for this to produce that?” If your answer jumps straight from a broad label to an outcome, slow down and fill in the links. Then test whether those links support reasoning beyond this illustration. Understanding often lives in those links—in the ordinary, sometimes hidden conditions that make an observed result follow.

Once you have sketched the links between a cause and an outcome, you can inspect them by changing the story. Hold everything else as steady as you can, alter one important feature, and ask: What would happen now?

This is a counterfactual question—a question about what might have happened if one part of the situation had been different. It is not a request to fantasize freely. It is a way to put pressure on an explanation. If the model is any good, it should tell you which changes matter, through what route, and with what limits.

Imagine a town where a severe flu outbreak leads the school to close for a week. A simple explanation might be: “Closing the school reduced infections.” But how? Suppose the proposed chain is that students have fewer close encounters, so there are fewer opportunities for the virus to pass from person to person. Now change one thing: the school remains open, but students stop eating together in crowded rooms. Under that model, reducing crowded shared meals should reduce some opportunities for transmission even while the school stays open. That is a prediction derived from the proposed mechanism; imagining it does not supply evidence that it happened.

Change a different thing. Keep the school closed, but imagine that students gather in large groups elsewhere. The model predicts less benefit from reducing encounters, because the relevant opportunities for transmission may have moved rather than vanished. This counterfactual clarifies the proposed mechanism—fewer opportunities for transmission—but it does not establish that a school closure caused any observed change in infections. Real outcomes can reflect other factors as well. A 2021 systematic review of observational studies of COVID-19 school closures found varied results and noted that other interventions introduced around the same time made the effects of closures difficult to isolate ^[1].

The thought experiment makes the hidden “because” inspectable. To find out whether that explanation fits the world, we need observations that could challenge it. Thinking through a model tests its implications; comparing those implications with evidence tests its fit.

The one-variable habit

A useful counterfactual begins with a deliberate edit: increase, remove, or reverse one feature while leaving the rest of the imagined scene alone. Then trace the consequences through the model.

Suppose a shop raises the price of a popular product. If you are thinking about demand, ask what might happen if the price rose further. Some customers may buy less; others may switch to alternatives. If the product has no close substitute, the drop in purchases may be small. The prediction depends on more than the word price: it depends on how customers respond and what choices they have.

Now remove the price change from the story. If you still expect purchases to fall for the same reason, then your explanation may be relying on something else—perhaps a seasonal shift or a shortage. The comparison exposes which variable is doing the work.

This is why “What if?” can be more revealing than “What is?” A description gives you a snapshot. A counterfactual makes the model move. It shows whether the parts you named are connected in the way you think they are.

Not every imagined change has a clean answer. If you ask what would happen if gravity suddenly reversed, you have changed so much that ordinary assumptions about the world collapse. A useful test changes one feature while keeping the surrounding rules intact. You are not trying to invent any possible world; you are checking the logic of this one.

When a prediction surprises you, follow the path. Which link did you assume? Which condition did you forget? Perhaps the effect is delayed, blocked, or outweighed by another cause. A miss can show where the model needs another variable or a more careful boundary.

The practical question is simple: If I changed this one thing, what else should change—and why? If you can trace the answer, you are beginning to reason with the model rather than stand outside it, repeating its description.

Return to the extension cord at the start of the book. The loop model explained why current flows, but not why the wire warms. We need one more relationship: a wire has resistance, and electrical energy is converted into heat as current passes through it. For the same material and temperature, a longer wire has more resistance; a thicker wire has less. At the same current, more resistance means more heating. For a wire whose resistance stays approximately constant, doubling the current produces four times the heating power. ^[2]

Now the opening questions have answers. At the same current, a shorter cord of the same material and thickness produces less heat overall; increasing the current makes heating rise sharply. Notice the condition we kept repeating: at the same current. Changing resistance can itself change current, depending on the rest of the circuit. The improvement in understanding lies in knowing both the relationship and what must be held steady to use it.

An observed result that differs from a prediction is a signpost: a relationship, an assumption, or even the observation may need checking. The next step is to follow it in two directions: forward, by asking what the model implies, and backward, by rebuilding the model from its essential parts. These are not separate tricks; they test whether the model supports reasoning beyond the example that first made it plausible.

Consider a simple electric circuit: a battery connected in a closed loop to a bulb. You may know the usual description—current flows, and the bulb lights—without yet being able to use it for a prediction. Try one forward step. What happens if you add a second bulb in the same loop?

For a simplified comparison, assume identical incandescent bulbs connected in series and a battery that maintains the same voltage as the load changes. Adding a second bulb increases the total resistance, so the current decreases; each bulb dissipates less power and glows less brightly. The result depends on those assumptions: it is not a prediction for every kind of bulb or battery.

Now work backward from that result. The bulbs share one series path, so adding one changes the resistance of the whole circuit and reduces the current through each. That is enough of the circuit model for this particular prediction; the arrangement matters. If the bulbs are connected in parallel, each has its own branch across the battery, and the prediction is different. A real battery’s behavior under load and a bulb’s electrical characteristics can also affect the outcome.

Prediction and reconstruction catch different weaknesses. A prediction asks whether the model can travel beyond the example you were given. Reconstruction asks whether you can assemble the explanation without relying on its polished wording. One tests the model’s reach; the other tests whether you know what holds it together.

When a prediction fails, ask what broke in the chain. Was a mechanism missing? Did you quietly assume a condition that wasn’t present? Did the model work only within a limited range? In the circuit example, a prediction made for bulbs in series fails if the actual setup is parallel—not because electricity has become mysterious, but because the model was applied outside its stated arrangement. The miss marks the edge of the model and gives you a more focused next question.

There is a further test: could a different explanation predict the same result? Suppose the café sells fewer coffees after raising its price. Perhaps customers are responding to the price; perhaps nearby roadworks have reduced passing trade. Both explanations fit the fall in sales. They point toward different observations, however. The roadworks account suggests fewer people passing the door; the price account suggests that, among people who still visit, more choose cheaper drinks or leave without buying. Neither observation settles everything, and both causes may operate. But now you know what evidence would help distinguish them.

Before checking, state what you expect and what finding would make you reconsider. Otherwise it is easy to reshape the story after every result and mistake that flexibility for understanding. With uncertain outcomes, one miss need not disprove a model: a 30 percent chance of rain allows both rain and no rain. The evidence must match the kind of prediction being made.

A useful explanation should let you derive a consequence, trace it back through the parts that produced it, and identify evidence that could count against it. A model can feel perfectly clear and still be wrong about the world.