Developed by Robert Stalnaker (A Theory of Conditionals, 1968), David Lewis (Counterfactuals, 1973), and Judea Pearl (Causality, 2000).
Tests whether something caused an outcome by changing that one factor in an imagined version of events and tracing what would have followed.
A lost account, and three candidate causes tested one at a time
A food distributor lost its largest restaurant account in March. Pick one cause to undo and how to undo it; the line under the rows traces what would have followed with everything else held as it was, and says whether that cause mattered. Illustrative.
Undo one cause
How to undo it
- Pick a cause to undo and how to undo it.Both January deliveries arrive on time. The buyer still leaves in February, her successor still puts every supplier out to tender as a matter of policy, and the rival still bids 6% lower. The account is most likely still lost: the delays did not decide it.No price rise. The tender still happens, and the rival's bid is 2% lower instead of 6%. The gap is small enough that the service record might have carried it: a contributing cause, and not enough on its own.The buyer stays. She renewed twice before despite late boxes and price rises, and nobody calls a tender. The account is most likely kept: this is the cause that mattered.Too big a change. A new account team, a new price list, and a new carrier together describe a different company, and the outcome says nothing about any one cause. Undo one thing and hold the rest fixed.
- Both January deliveries arrive on time. The buyer still leaves in February, her successor still puts every supplier out to tender as a matter of policy, and the rival still bids 6% lower. The account is most likely still lost: the delays did not decide it. (cause: The late deliveries, how: Smallest change only)
- No price rise. The tender still happens, and the rival's bid is 2% lower instead of 6%. The gap is small enough that the service record might have carried it: a contributing cause, and not enough on its own. (cause: The December price rise, how: Smallest change only)
- The buyer stays. She renewed twice before despite late boxes and price rises, and nobody calls a tender. The account is most likely kept: this is the cause that mattered. (cause: The buyer leaving, how: Smallest change only)
- Too big a change. A new account team, a new price list, and a new carrier together describe a different company, and the outcome says nothing about any one cause. Undo one thing and hold the rest fixed. (how: Rewrite the whole year)
An example
A food distributor lost its largest restaurant account in March, and the first review blamed the operations team for two late deliveries in January. The sales director asked for one more pass before anyone acted on it, testing each candidate cause on its own.
The review team took the three things that had happened: the late deliveries, a 4 percent price rise in December, and the departure in February of the buyer who had run the account for six years. For each, they imagined the smallest change that removed it, kept everything else as it was, and traced forward. On-time deliveries changed nothing, since the new head of purchasing put every supplier out to tender as a matter of policy and the rival bid 6 percent lower. Without the price rise the gap would have been 2 percent, which helped and was not enough. If the buyer had stayed, no tender would have been called. The fix was multiple contacts at every top account, and the operations team kept its bonus.
Why it works
After a bad outcome, people blame whatever went wrong nearest to it or most visibly. Asking what would have happened without one factor turns that hunch into a test: if the outcome follows anyway, the factor did not decide it, however bad it looked.
The discipline is in the size of the change. Remove one cause by the smallest edit that does it, and hold fixed everything that did not flow from it; a sweeping rewrite of events produces a different world that says nothing about the cause in question. Tracing forward step by step exposes the other forces that were already in motion, and trying nearby versions of the change shows how firm the conclusion is.
The thinking behind it
Robert Stalnaker Robert Stalnaker, a philosopher at MIT, gave conditionals a possible-worlds semantics in A Theory of Conditionals (1968, in Studies in Logical Theory, edited by Nicholas Rescher): "if A, then B" is true when B holds in the world most like ours in which A holds. Finding that nearest world is the smallest-change rule that runs through this method. With David Lewis's account it forms the Stalnaker-Lewis theory of counterfactuals.
David Lewis David Lewis, a Princeton philosopher, analyzed statements of the form "if it had been, it would have been" in Counterfactuals (1973). On his account a counterfactual is true when its consequent holds in the possible worlds most similar to ours in which its antecedent holds. Judging similarity, what must change and what may be held fixed, is the work a careful counterfactual does.
Judea Pearl Judea Pearl, a UCLA computer scientist and recipient of the 2011 Turing Award, developed Bayesian networks and a formal theory of causal inference, set out in Causality: Models, Reasoning, and Inference (2000). It uses directed acyclic graphs to draw causal structure, the do-operator to represent an intervention, and structural models to compute counterfactuals, which turned the philosophers' question into something that can be calculated from a model.
How to run it
- Name the outcome. State exactly what happened and when, such as the account lost in March.
- List the candidate causes. Write down every factor someone has blamed, including the ones that feel obvious.
- Undo one cause by the smallest change. Describe the minimal edit that removes it, such as two deliveries that arrive on time.
- Hold everything else fixed. Keep every event that did not flow from that cause exactly as it was.
- Trace forward step by step. Ask at each step what follows from the change, using what the people involved actually did elsewhere.
- Compare the two outcomes. If the outcome happens anyway, the cause did not decide it; if it vanishes, it did.
- Try nearby versions. Vary the change a little and see whether the conclusion survives.
Where it goes wrong
The change is too big. Imagining a different team, a different market, and a different strategy all at once tests nothing in particular. If the counterfactual needs more than one sentence, it is probably several counterfactuals.
The trace runs on hope. People imagine the alternative world going smoothly and conclude that the cause mattered. Ground each step in what the people involved actually did in similar situations.
One cause gets all the credit. Several causes can each have been necessary, or none alone sufficient. Test each in turn and report which ones the outcome depended on, and by how much.
When to use something else
- Difference-in-Differences. When comparable units exist that did and did not get the change, so the counterfactual can be estimated from data.
- Causal Diagram. When the causes affect one another and the structure needs drawing before any one of them is tested.
- Mill's Methods. When many cases are on hand and the cause can be found by comparing where the outcome did and did not occur.
Where it sits in SOLVE eX
Go to the source
- A Theory of Conditionals, Studies in Logical Theory, 1968
- Counterfactuals, 1973
- Causality: Models, Reasoning, and Inference, 2000