Abductive Reasoning

Thinking Tools4.1 Root Causes and the Case for Going BackAlso known as Abductive Inference, Inference to the Best Explanation

Set out by Charles Sanders Peirce, who first called it hypothesis (Deduction, Induction, and Hypothesis, 1878) and later abduction, and developed as inference to the best explanation by Gilbert Harman (1965) and Peter Lipton (1991).

Works back from a surprising fact to the explanation that would best account for it, ranking several candidates and testing the leader before acting on it.

Three explanations for a jump in sign-ups, scored before anyone acts on one

Sign-ups to a project-management app rose 60 percent in a week from Tuesday morning, while new accounts that were actually used rose only 10 percent. An explanation that cannot fit the timing is out. Score the rest on how well each fits the other facts, how often it happens, how simple it is, and how cheaply it can be tested, and change the weights: the top two are the pair to separate next, with the cheapest test that tells them apart, a look at where the new sign-ups came from. Illustrative.

Fits the timing: the jump began Tuesday at 9 a.m., the press piece ran Wednesday

MustPricing page changeBot sign-up attackPress mention

Fits the other facts: few of the new accounts were used

Weight8Pricing page change6Bot sign-up attack8Press mention4

Prior odds: how often this kind of cause produces a jump like this

Weight6Pricing page change8Bot sign-up attack4Press mention6

Simplicity: how few extra assumptions it needs

Weight4Pricing page change8Bot sign-up attack6Press mention7

Testability: how cheaply it can be checked this week

Weight5Pricing page change7Bot sign-up attack9Press mention6

Weighted total

Pricing page change163OutBot sign-up attack157OutPress mention126Out

The leading explanationNo option passes every must: rescope or find another.Pricing page changeBot sign-up attackPress mention

In order, with the runner-up second

  1. 163Pricing page changeBot sign-up attackPress mentionOut: fails a must (Fits the timing: the jump began Tuesday at 9 a.m., the press piece ran Wednesday).
  2. 157Pricing page changeBot sign-up attackPress mentionOut: fails a must (Fits the timing: the jump began Tuesday at 9 a.m., the press piece ran Wednesday).
  3. 126Pricing page changeBot sign-up attackPress mentionOut: fails a must (Fits the timing: the jump began Tuesday at 9 a.m., the press piece ran Wednesday).

An example

On a Tuesday morning, sign-ups to a project-management app jumped, and by the following Monday they were up 60 percent on the week before. The marketing team credited a mention in a trade newsletter. Growth credited the redesigned pricing page that had shipped Monday night. An engineer pointed out that accounts actually used in their first week had risen only 10 percent, and wondered aloud about bots.

The product lead asked for the three explanations to be scored before anyone acted on one. The press mention fell out first: it ran Wednesday, a day after the jump began. Of the other two, the pricing change scored higher on prior odds and simplicity, and the bot attack fit the unused accounts better. The two were close, so the cheapest test that separated them came next: an hour spent on the sign-ups' email domains and IP ranges. A third of the new accounts came from four address blocks. Both explanations were partly true, the team blocked the blocks, and the pricing page kept its credit for the rest.

Why it works

Faced with a surprise, people settle on the first explanation that fits and start acting on it. Abduction makes the step from surprise to explanation deliberate. It starts by stating exactly what is surprising, against which expectation, and then asks which explanation would make that fact unsurprising, with at least three candidates on the table, including dull ones.

Scoring the candidates on fit, prior odds, simplicity, and testability guards against the two usual errors: preferring the vivid story over the common one, and preferring the story that explains everything over the one that can be checked. The output is a ranking, never a verdict. The leader is a working hypothesis, the runner-up is its main rival, and the next move is the cheapest test that would tell them apart. That keeps a fast decision from becoming a confident mistake.

The thinking behind it

Charles Sanders Peirce Charles Sanders Peirce, a founder of pragmatism, named a third kind of inference besides deduction and induction. He first called it hypothesis, in Deduction, Induction, and Hypothesis (Popular Science Monthly, 1878), and later abduction: reasoning from a surprising fact to the explanation that would make it unsurprising. In 1878 he described it as inferring a case from a rule and a result, and in his 1903 Harvard lectures he called pragmatism the logic of abduction.

Gilbert Harman Gilbert Harman, a Princeton philosopher, named inference to the best explanation in a 1965 paper in the Philosophical Review: inferring that a hypothesis is true because it would explain the evidence better than its rivals. The paper made the term standard in discussions of abductive reasoning, and it puts the comparison with rival explanations at the center, which is why the figure scores three candidates against each other.

Peter Lipton Peter Lipton, a Cambridge philosopher of science, gave the standard book-length treatment in Inference to the Best Explanation (1991; second edition 2004). He distinguished the likeliest explanation from the loveliest, the one that would give the most understanding, and gave a contrastive account of explanation: why this fact and not some other. Both ideas show in practice when a team asks what each candidate predicts that the others do not.

How to run it

  1. State the surprise exactly. Name the fact and the expectation it broke: sign-ups up 60 percent in a week, against a steady 2 percent.
  2. List at least three explanations. Include the dull ones, such as a tracking error or a bot, alongside the flattering ones.
  3. Note what else each would predict. A real demand surge predicts more active accounts; a bot attack predicts few, from clustered addresses.
  4. Drop any that contradict a known fact. Timing, location, and segment often rule one out at once.
  5. Score the rest. Rate fit with the other facts, prior odds, simplicity, and testability, with weights the team agrees.
  6. Run the cheapest test that separates the top two. Choose the check whose result would come out differently under each.
  7. Update and repeat. Rescore with the new fact, and keep the runner-up in view until the test settles it.

Where it goes wrong

The common failure is stopping at one explanation. The first plausible story gets acted on, and evidence that fits a rival goes unnoticed. Require three candidates before anyone scores one.

A second failure is preferring the explanation that accounts for everything. A story stretched to cover every fact is often less likely than a simpler one that leaves a fact or two to noise. Weight prior odds and simplicity alongside fit. The third is treating the leader as proven. Abduction gives a working hypothesis; until a test separates it from the runner-up, plan actions that would make sense under both, as the team did when it blocked the address blocks without undoing the pricing page.

When to use something else

  • Differential Diagnosis. When the candidates are known in advance, as in a list of failure causes, and the work is ruling them out in order.
  • Bayesian Updating. When prior odds and the strength of each piece of evidence can be put in numbers.
  • Occam's Razor. When several explanations fit equally well and the question is which assumes least.

Where it sits in SOLVE eX

Go to the source

  • Deduction, Induction, and Hypothesis, Popular Science Monthly; later gathered in the Collected Papers, 1931 to 1935, 1878
  • The Inference to the Best Explanation, Philosophical Review, 1965
  • Inference to the Best Explanation, Routledge, 1991

Back to the catalog