Lewis's Counterfactuals and Causal Overdetermination — Epoche B2
Fragile Events Will Not Save the Counterfactual Theory The wrong problem, and a repair its author refused David Lewis's counterfactual analysis of causation is usually introduced with a difficulty attached [1] . Two riflemen fire at once, either shot would have killed, the animal dies. Neither shot satisfies the counterfactual test, since without it the animal would have died anyway. So the theory says neither is a cause, and that seems wrong. Two things about this standard presentation are worth disputing, and the second matters more than the first. Symmetric overdetermination of this kind was not a problem Lewis felt obliged to solve [2] , and he said so: our intuitions about such cases are unsettled, so a theory may deliver an odd verdict without being refuted. The case that does the damage is preemption , and specifically late preemption. And the repair usually offered on Lewis's behalf — describe the effect finely enough and dependence reappears — is one Lewis rejected [3] , for a reason internal to his own project. This essay argues that the fragility repair cannot be rehabilitated, and not because it is inelegant. It relocates the causal verdict onto a parameter that has nothing to do with causation, and there are cases where it has nothing to work on at all. The conclusion is not that Lewis's theory is beyond help [4] ; it is that help had to come from somewhere else, which is why he replaced the analysis in 2000. A specialist can deny this, and the best denial is set out below. The 1973 analysis, stated properly A counterfactual conditional says what would have been the case had things been otherwise. Lewis's semantics for these [5] , developed in Counterfactuals (1973 [6] ), evaluates them over possible worlds: "if C had not occurred, E would not have occurred" is true just in case, among the worlds where C does not occur, those most similar to the actual world are worlds where E does not occur either. What "most similar" comes to is not settled in that book; the system of weights that fixes it arrives later, in 'Counterfactual Dependence and Time's Arrow' (1979). In the paper 'Causation' of the same year Lewis defines causal dependence with this apparatus. Where C and E are distinct actual events, E depends causally on C just in case, had C not occurred, E would not have occurred. Then comes the step the textbook presentation usually drops, and without which nothing that follows makes sense. Lewis does not identify causation with dependence. Causation is the ancestral of dependence: C causes E when there is a chain of actual events, each depending on the one before, running from C to E. Dependence is sufficient but not necessary. He takes that step for a reason. Consider early preemption. An assassin fires; a backup assassin, watching, would have fired had the first not. The victim dies. The victim's death does not depend on the first shot, because the backup would have acted. But there is an intermediate — the bullet in flight, halfway to its target — which depends on the first shot, and on which the death depends. The chain restores the right verdict. That is the whole point of the ancestral. Where the chain fails Late preemption defeats it. Suzy and Billy each throw a rock at a bottle. Suzy's throw is a fraction earlier; her rock shatters the bottle; Billy's rock flies through the space where the bottle was. Billy's throw is preempted, and preempted by the effect itself rather than by anything on Suzy's causal path. Now look for the chain. Take any intermediate on Suzy's process — her rock at the halfway point, say. The shattering does not depend on it, because had that intermediate been absent, Billy's rock would have arrived and the bottle would have shattered. Every link fails the same test, for the same reason. The ancestral has nothing to build with. Notice how this differs from the two riflemen. There, the two processes are symmetrical, and it is genuinely unclear what to say; Lewis was entitled to shrug. Here, everyone agrees that Suzy's throw caused the shattering and Billy's did not. The theory delivers no verdict at all, and there is a determinate verdict to be got wrong. The fragility repair The obvious response is to describe the effect more finely. The shattering that actually happened occurred at a particular instant, with a particular pattern of fragments, from a particular direction of impact. Had Suzy not thrown, the bottle would have shattered — but a moment later, differently arranged, struck from the other side. So that shattering, the one that actually occurred, does depend on Suzy's throw after all, and the chain is restored. Everything here turns on how fragile events are: how much of an event's actual character belongs to its essence, such that a differently timed or differently arranged occurrence would be a different event rather than the same event otherwise. Lewis saw the move and would not take it. His objection was that fragility over-generates. If an effect's essence includes its exact time, then anything that shifted the effect by a fraction of a second becomes a cause of it. The waiter who delayed the meal is thereby a cause of everything that happened afterwards, in exactly the way the murderer is a cause of the death. Loosen event essences enough to block that, and the preemption cases stop being handled; tighten them enough to handle preemption, and the theory calls almost everything a cause of almost everything. Why no setting of the dial works It is tempting to read this as a calibration problem — as though there must be some principled degree of fragility, discoverable by a better theory of events, at which the two demands are met. The trouble is deeper, and it can be put as a structural point. On the fragility repair, whether C caused E depends on whether the alternative process would have produced a numerically distinct event. That is settled by facts about event essences: by how much variation in time and manner an event can tolerat