Inference to the Best Explanation

What it is

There are three different things a reader can be handed in support of a causal claim, and confusing them is the commonest failure in popular science writing.

The first is a correlation. It sizes an association: how much of the variation in one thing goes with variation in another. It says nothing about order or direction.

The second is an experiment or quasi-experiment. It fixes the order of events, by assignment or by exploiting a change whose timing was decided elsewhere, and so can license a causal claim about the population, exposure and outcome it covered.

The third is an inference to the best explanation. It does neither. It surveys the candidate causes of a pattern, argues each of them away, and nominates the survivor. No coefficient is produced and no counterfactual is observed. What is produced is an elimination.

This is not a cheat, and the page would be useless if it were read as one. A great deal of real science works this way, especially where the phenomenon is historical and cannot be rerun. Continental drift, the asteroid account of the end-Cretaceous extinction and the identification of a disease vector before the mechanism is known have all rested, at some stage, on there being no other plausible story. Labelling an argument as this kind of argument is a classification, not a demotion.

But it has a characteristic weakness and it is sharp. The argument is only as strong as the completeness of the list of alternatives, and as strong as the rigour with which each was dismissed. A short list makes the conclusion look inevitable and makes the argument weak, and nothing inside the argument signals which has happened. An elimination that missed a candidate delivers a confident wrong answer with no internal warning.

So the right question to put to an argument of this shape is never "what is the effect size", because it does not have one. It is: what else was on the list, and how was each one ruled out? And a secondary question that does a surprising amount of work: who assembled the list?

There is a stronger version of the form, worth recognising because it is better practice. Instead of simply asserting that nothing else fits, the arguer specifies what a rival cause would have to look like, in enough detail that someone could go and find one. That converts an elimination into a standing invitation, which is closer to falsifiable than the weak version gets.

In effect

Research: the synchrony argument, named as such

The clearest instance in this publication's archive is Jonathan Haidt's international argument in The Anxious Generation. At page 44: "The only plausible theory I have found that can explain the international decline in teen mental health is the sudden and massive change in the technology that teens were using to connect with each other." His figures cover Canada, the United Kingdom, Australia, the Nordic countries and a 37-country analysis of school loneliness and alienation.

The form is explicit in his own sentence. The only plausible theory I have found. It is not a coefficient and it is not a design, and the book never claims it is either. This matters because the international material is often summarised as though it were the book's strongest evidence, when it is in fact the book's strongest argument, which is a different kind of thing. His own hedge sits on the same page: "Data is less abundant in non-Western nations, and the patterns there are less clear."

He also produces the stronger version of the form, at page 290: "Until someone finds a chemical that was released in the early 2010s into the drinking water or food supply of North America, Europe, Australia, and New Zealand, a chemical that affects adolescent girls most, and that has little effect on the mental health of people over 30, the Great Rewiring is the leading theory." That is rhetoric, and it is also a specification of what a competing cause would have to satisfy, which is more than most theses offer. Note the phrase he lands on, twice on that page, in his own words: the leading theory.

What a reader should do with this is neither dismiss it nor upgrade it. The pattern is real and the elimination is argued seriously. Whether the rivals have been eliminated is the live question, and it is live in part because the main rival in this field, that distress came first and changed how young people use these platforms, is itself a candidate the available designs cannot cleanly rule out.

Everyday life: the appliance tripping the breaker

A household works out which appliance is tripping the fuse by unplugging each one in turn. Nobody measures a current, nobody consults a wiring diagram. The conclusion arrives when the list is exhausted and one candidate is left standing.

It works, and the reason it works is instructive: the list of appliances in a kitchen is short, closed and visible, so an exhaustive elimination is genuinely available. That is exactly what a social question lacks. Miss one appliance, perhaps because it is plugged in behind a cupboard, and the method hands back a wrong answer with complete confidence, and the household rewires something that was never faulty. The failure mode is identical in a kitchen and in a literature, and the only defence is the completeness of the list.

Commerce and product design: attributing a bad quarter

Sales fall. The team rules out seasonality, price changes, stock-outs and the competitor's launch, and concludes that the website redesign did it. No experiment was run and no counterfactual exists, and yet the slide gets presented with the confidence of a measured result.

The inference may well be right. The honest version of the slide says "the best remaining explanation" rather than "the cause", and the useful question in the room is which candidates were never on the list: a change in the marketing mix nobody mentioned, a shift in a channel partner's algorithm, a cohort of customers reaching the end of a contract. Note also who built the list, since the team that shipped the redesign is rarely the one that puts it at the top.

A/B testing exists precisely because this argument form, however conscientiously conducted, is not a substitute for a comparison. Organisations that run proper experiments are not smarter than those that reason by elimination. They have simply noticed that elimination arguments are unfalsifiable from the inside.

What it does not say

An inference to the best explanation does not establish a cause in the sense an experiment does, and it carries no effect size, so it cannot be put into a meta-analysis, cannot be pooled with a coefficient, and cannot be reported as "studies show".

Classifying an argument this way says nothing about whether it is sound. The international-synchrony argument discussed above is argued seriously and is disputed by named researchers with published numbers, and this page takes no view on which way it comes out. What it insists on is that it be quoted at the strength its author gives it, which is his own phrase, the leading theory, rather than at the strength a book cover asserts.

It is also not interchangeable with a tally. An elimination argument that genuinely engages rival explanations is doing more work than a count of studies pointing one way, and the two are often confused because neither reports a magnitude. See vote counting for the weaker of the pair.

And one thing must be said plainly about the dispute this page is drawn from, because it is routinely mis-told. Both the author of the synchrony argument and the author of the most prominent published criticism of it accept that adolescent mental health has deteriorated, and both want the platforms reformed. Candice Odgers, reviewing the book in Nature, writes that "we have a generation in crisis and in desperate need of the best of what science and evidence-based solutions can offer", in the same review in which she holds that the rewiring claim "is not supported by science" and that "such serious claims require serious evidence". The disagreement is about the direction of the longitudinal associations, the mechanism, and two specific policies. It is a methodological dispute conducted in journals, and it should not be read as a fight.

Two further evidence claims in that review are high-value and are not cited here, because the copy this publication read was an excerpted course copy whose numbered reference list was absent, so the works behind them cannot be named without guessing.

Who developed it

No single originator is recorded here. The form is a standing one in the philosophy of science rather than a proposal by a single researcher, and this publication verified no attribution for the phrase at source on this pass. Rather than print a plausible name and date that were not checked, the vault records the origin as unknown.

---