A board chair sits in her office six months after a strategic decision that backfired. The company expanded into a new geography on her recommendation. Revenue came in at a third of the projection. Two senior leaders departed. The board asks the obvious question: what would have happened if we had not expanded, if we had invested those resources in deepening our existing market instead? The chair would like to answer. So would her CEO, her CFO, the board, the analysts on Wall Street. None of them can. The world in which the company did not expand does not exist. There is no data on it. There is no possibility of running the experiment. There is only this world—with its disappointing revenue and departing leaders—and the world that did not happen, about which we have only intuitions. Yet humans answer counterfactual questions all the time. We make judgments about what would have been, hold ourselves and others accountable to those judgments, use them as the basis of regret and attribution and learning. The question is whether such judgments can be made rigorous—whether we can compute counterfactuals from data and a model. Due to Judea Pearl's work, the answer is yes.

Every executive faces counterfactual questions every quarter. When a campaign underperforms, when a product launch falters, when a market entry disappoints, the question is always: what would have happened if we had made a different choice? Every patient faces them when treatment fails. Did this drug cause my illness to worsen, or would it have worsened anyway without treatment? Every plaintiff in a wrongful-death lawsuit asks it. Would the victim have survived if the defendant had acted differently? Every parent reflects on difficult conversations years later. What if I had said something different—what would that teenager have become? Counterfactuals underpin accountability, regret, attribution, and learning. We use them to judge whether we made the right decision, whether someone else acted negligently, whether we should have chosen differently. These judgments shape our actions. They are the basis of legal liability, medical ethics, and personal responsibility. If counterfactuals cannot be computed rigorously, then our judgments rest on intuition. If they can be computed, we have a foundation for evidence-based accountability.

The Ladder of Causation, introduced in Chapter 5, ranks types of causal reasoning by their evidence requirements. Rung 1 admits observational data—what we saw happen in the world. Rung 2 admits interventional data—what happens when we manipulate variables, as in a randomized trial or a controlled experiment. Rung 3 admits counterfactual data—what would have happened in a world that did not occur. Here the ladder seals. We cannot move from Rung 2 to Rung 3 by collecting more observational or interventional data, because counterfactual data cannot be collected. The world in which the company did not expand has no data. The world in which the patient received no treatment does not exist. The world in which the teenager heard different words never happened. We cannot observe what did not occur. The only path to counterfactuals is to combine the data we have with a model rich enough to predict what would have happened in worlds we did not see. That model is the structural causal model.

The SCM is more than a diagram. It is a diagram plus the functions that determine each variable from its parents and an exogenous noise term. These functions encode the mechanism—the causal machinery by which the system works. If we know the mechanism, we can answer counterfactual questions. We can ask: in this specific case, with its specific circumstances and noise terms, what would the outcome have been if the input had been different? The SCM gives us a procedure for answering. That procedure is mechanical. It does not require us to run the experiment or observe the alternative world. It takes the observed case, extracts the idiosyncratic information that distinguishes it from all other cases, modifies the model to enforce the counterfactual condition, and computes what would have resulted. The SCM's empirical credibility derives from Chapter 6: its structural assumptions imply specific patterns in observational data. If those patterns hold, the model is credible. The counterfactual computed from that model inherits that credibility. It is not a fact—it is conditional on the model being correct. But the model is testable.

Pearl's abduction-action-prediction procedure has three steps, each mechanical and each essential. The procedure takes an observation from the real world and computes a prediction about an alternative world that did not occur. Step 1, abduction, uses the observed data to infer the exogenous noise terms. The SCM says that each variable is determined by its parents and a noise term U. We have observed the variables. We can therefore recover what the noise terms must have been in order to produce the observed values. This step extracts case-specific information—the idiosyncrasies of this particular situation—that distinguishes this case from all others. It is the signature of the actual case. Step 2, action, modifies the SCM to enforce the counterfactual condition. If the question is what would have happened if X had been x′ instead of x, we replace the equation for X with the assignment X equals x′. The other equations remain unchanged. This builds the counterfactual world's model, which differs from the actual world's model only in X. Step 3, prediction, uses the modified SCM and the noise terms recovered in Step 1 to compute the value of Y. This is the counterfactual outcome.

Abduction is the first step. The SCM specifies that each variable is a function of its parents plus a noise term. In the real world, we observed X took the value x, Y took the value y, and other variables took specific values. The question is: what noise terms must have been present to produce these specific observed values? Abduction uses the SCM's equations to solve for the noise terms. This is reverse engineering. We know the output, we know the function, we solve for the noise input. The noise terms are exogenous—they are not caused by anything within the system. They represent the idiosyncratic features of this case. One campaign succeeded because the market happened to be receptive that quarter. Another campaign failed because a competitor launched a better product that week. These circumstances are captured in the noise terms. Abduction extracts them. The noise terms encode what is unique about this case—its luck, its context, its exogenous shocks. In Step 3, when we compute the counterfactual outcome, we will use these same noise terms. This ensures that the counterfactual preserves the idiosyncrasies of the actual case. We ask: given these specific circumstances, if X had been different, what would have resulted?

Action is the second step. The counterfactual question asks: what if X had been x′ instead? Action answers this by modifying the SCM. We take the original structural causal model and replace the equation for X with a direct assignment: X equals x′. This is an intervention in the logical world, not the real world. We are not saying what would happen if we actually intervened to change X in reality—that is a Rung 2 question, answerable from observational and interventional data. We are saying: suppose the equation for X no longer holds. Suppose X is simply set to x′. What would follow? The other equations remain unchanged. Y is still determined by its parents according to the original equation. Z is still determined by its parents according to the original equation. Only the determination rule for X changes. This builds the counterfactual world's model. The difference between the actual world's model and the counterfactual world's model is minimal—one variable's determination rule. Everything else is identical. This isolation of the change is essential. It ensures that any difference in Y between the actual world and the counterfactual world is caused by the change in X, all else being equal.

Prediction is the third and final step. Now we have two things: the noise terms recovered in Step 1, which encode the specific circumstances of the actual case, and the modified SCM from Step 2, which represents the counterfactual world. We feed the noise terms into the modified SCM and compute the value of Y. This is the counterfactual outcome. It is what Y would have been in this specific case, with its specific circumstances, if X had been x′ instead of x. The prediction is mechanically derived from the model and the observation. It is not a judgment or a guess. It is the logical consequence of the SCM and the abducted noise terms. The procedure is one of the most beautiful results in causal inference because it is so simple. Three steps. Each mechanical. Each testable against the structure of the SCM. The final prediction inherits empirical credibility from the model. If the SCM's assumptions are true—if the structure really captures how the world works—then the counterfactual prediction is as reliable as any causal inference can be. It may be wrong if the model is wrong. But given the model, it is the correct answer to the counterfactual question.

Let's work through a concrete example—the kind of counterfactual question the board chair faces. Suppose a simple model: a marketing campaign M drives leads L, which drive revenue R. The SCM specifies three equations. M has no parents; it is determined only by exogenous noise U_M. L is determined by M and noise U_L, with coefficient alpha. R is determined by L and noise U_R, with coefficient beta. From past data, we estimate alpha equals 0.4 and beta equals 5. We observe a specific case: the company ran a campaign at M equals 100 units of investment. This generated L equals 50 leads and R equals 200 in revenue. Abduction recovers the noise terms. We solve the equations backward. Given M equals 100 and the equation M equals U_M, we have U_M equals 100. Given L equals 50, M equals 100, and the equation L equals 0.4 M plus U_L, we solve: 50 equals 0.4 times 100 plus U_L, so U_L equals 10. Given R equals 200, L equals 50, and the equation R equals 5 L plus U_R, we solve: 200 equals 5 times 50 plus U_R, so U_R equals 50. These noise terms represent the specific circumstances of this campaign—the market responsiveness, the effectiveness context, the revenue baseline. Action modifies the model by replacing M equals U_M with a direct assignment, say M equals 150. The other equations remain: L equals 0.4 M plus U_L, R equals 5 L plus U_R. Prediction computes the counterfactual. Using the recovered noise terms: L becomes 0.4 times 150 plus 10, which equals 70. R becomes 5 times 70 plus 50, which equals 400. The counterfactual revenue is 400.

Abduction-action-prediction gives us a way to answer counterfactual questions rigorously. It transforms them from intuitions into evidence-based computations. The result is specific to the individual case. The board chair asks: what would revenue have been if we had invested differently? Not: what do campaigns of this type typically yield? Not: what is the average effect of increased investment? But: in this case, with its specific circumstances and noise, what would have happened? That is a Rung 3 question. It requires a structural causal model. If the model is correct, if the estimated parameters are accurate, if the structural assumptions hold, then the computed counterfactual is rigorous. It is not a guess. It is the logical consequence of mechanism and observation. The empirical credibility of the counterfactual is inherited from the credibility of the SCM. If the SCM's assumptions imply testable patterns in observational data—if those patterns hold—then we have reason to trust the counterfactual computed from it. This is how we move from the intuitive answers people give, based on experience and judgment, to answers that can be justified, communicated, and learned from. An executive can explain to her board: I'm not claiming certainty—we didn't run that experiment. But this is what the mechanism predicts given the actual circumstances and estimated parameters.

Counterfactuals are the hardest form of causal reasoning. They require not just observation or intervention, but a model rich enough to predict worlds we did not see. They are individual-level questions. Every other form of causal inference aims at populations and averages: what is the average effect of a treatment, the typical impact of a policy, the general relationship between two variables. Counterfactuals ask about a specific case with its specific circumstances. What would have happened in this situation, with its unique noise and context? For decades, counterfactuals seemed unanswerable. The world that did not happen leaves no traces. We cannot observe it. We cannot run the experiment. We can only speculate based on experience and intuition. Judea Pearl's abduction-action-prediction procedure changed this. It showed that counterfactuals are computable given a structural causal model. The procedure is mechanical. It requires three inputs: the structural equations, the estimated parameters, and the observed data. It produces one output: a prediction of what would have happened if the input had been different. The prediction is conditional on the model being correct. If the model is wrong, the prediction is misleading. But the model is testable. Its structural assumptions imply patterns in observational data. If those patterns hold, we have reason to trust both the model and the counterfactual it produces. This is how we move from intuition to evidence in the space of counterfactuals.