What Is Inference to the Best Explanation? Reasoning Backwards From Evidence

Deduction moves from general premises to a conclusion that must be true if they are, and it adds no new content, since the conclusion was contained in the premises. Induction generalises from observed instances to unobserved ones, concluding that what has held will hold, and its conclusion goes beyond the premises and can be wrong. This third pattern, sometimes called abduction following Charles Sanders Peirce, works in the opposite direction from deduction, starting with a surprising observation and reasoning back to what would have produced it. Finding the ground wet, concluding that it rained, is the schoolroom example, and the reasoning is plainly fallible since a burst pipe or a street cleaner would produce the same result. What makes it more than a fallacy is the comparative element, since the conclusion is not simply that a hypothesis would explain the evidence but that it explains it better than the available alternatives, which requires the alternatives to be considered.

The criteria used to compare hypotheses are widely agreed in outline and hard to make precise:

• •Explanatory scope, meaning how much of the evidence it accounts for rather than leaving unexplained
• •Simplicity, preferring the hypothesis that requires fewer independent assumptions, which is the working content of Occam's razor
• •Consistency with established knowledge, since a hypothesis requiring the abandonment of well-supported theory carries a heavy cost
• •Testability, since a hypothesis making further checkable predictions is worth more than one that merely accommodates what is known
• •Absence of ad hoc modification, meaning that the hypothesis was not patched specifically to fit the awkward evidence
• •Unification, meaning that it connects things previously thought unrelated, which is a strong mark in the history of successful theories

The pattern faces a serious challenge frequently called the bad lot argument. Selecting the best available explanation only supports believing it if the true explanation is among those considered, and there is no guarantee of that, so the procedure may reliably select the best of a bad set. The history of science supplies support for the worry, since theories that were overwhelmingly the best available explanation of their evidence have been replaced, which is the basis of the pessimistic induction against scientific realism. A second objection concerns whether explanatory virtues track truth at all, since simplicity and unification are aesthetic or pragmatic preferences and their connection to how the world actually is requires an argument rather than an assumption. Defenders reply that the same worry applies to every form of ampliative reasoning, that the criteria have a good track record, and that a Bayesian reconstruction can recover much of the practice as a form of probabilistic updating, which is a substantial and unfinished project.

Recognising the pattern makes a range of activities legible as the same operation. Medical diagnosis generates candidate conditions that would produce the symptoms and discriminates between them by testing, and the discipline of considering alternatives is exactly what checklists and differential diagnosis enforce. Criminal investigation reconstructs events that would account for the evidence, and the standard failure mode is fixing on one account early and interpreting everything subsequently in its light, which is documented in wrongful conviction cases. Historical reasoning infers past events from surviving traces. Palaeontology and geology reconstruct processes from their products. Fault diagnosis in engineering does the same. In every case the strength of the conclusion depends on how thoroughly the alternatives were canvassed, which is why the most useful practical discipline derived from this is not a rule of inference but a habit, namely asking what else would produce exactly this evidence.

The reasoning runs backwards from evidence to whatever would have produced it, and it becomes more than guessing only when rival explanations are actually considered, since the conclusion is comparative. Scope, simplicity, consistency and testability are the standard criteria. The serious objection is that choosing the best available explanation helps only if the true one was among those on offer, which the history of replaced theories makes a live worry.