Why do we say we fail to reject the null hypothesis instead of accepting it?

## The asymmetry is structural, not stylistic The wording looks like pedantry until you see where it comes from. A hypothesis test is a procedure with exactly one guarantee attached: **if the null is true, the chance it is wrongly rejected is at most `alpha`**, typically 5%. That guarantee is built into the critical value. Nothing in the construction bounds how often a *false* null survives — that depends on how big the real effect is and how much data you collected, neither of which the test controls. So the two outcomes carry very different weight: - **Reject `H0`** — backed by a designed error rate. The data would be unusual if the null were true. - **Fail to reject `H0`** — backed by nothing in particular. The data were not unusual enough under the null. That is all it says. "Accept the null" borrows the credibility of the first outcome and applies it to the second. ## Why non-rejection is so weak The deeper problem is that `H0` is a point claim while the data are compatible with a whole neighbourhood of values. Suppose a bottling line is tested against `H0: mu = 500` and the result is non-significant. The data are compatible with 500. They are usually also compatible with 499.4, with 501.2, and with everything in between — values that live in `H1` and may well matter operationally. The test simply could not tell them apart from 500 at the precision available. This gets worse the weaker the study is. A tiny sample, or a very noisy measurement, produces a wide range of compatible values and therefore fails to reject almost any null you hand it. Under that logic "accepting the null" would reward the sloppiest experiments with the strongest-sounding conclusions — the exact inversion of what evidence should do. **A non-significant result from a weak design is uninformative, not reassuring.** ## Absence of evidence is not evidence of absence The phrase is worn but it is precisely right here. "We looked and did not find it" only counts as evidence that it is not there if you looked somewhere it would have been visible. A test that could not have detected a meaningful effect even if one existed provides no information about whether one exists. This is why the honest follow-up to a non-significant result is not "so there is no effect" but "what effects can we still not rule out?" — a question about the precision of your estimate rather than about the verdict of the test. ## What you can legitimately report After a non-significant result, a defensible write-up says: 1. The data did not provide sufficient evidence against `H0` at the stated significance level. 2. Here is the estimated effect and the range of effect sizes the data remain compatible with. 3. Here is whether that range excludes the effects we would have cared about. Point 3 is where the real information lives. If the compatible range is narrow and excludes everything operationally meaningful, you have learned something substantive — "no effect larger than this" — and you learned it from the precision of the estimate, not from the test's verdict. If the range is wide and includes effects that would change decisions, the correct conclusion is "this study was not informative", and the correct next step is more data, not a claim of no effect. ## When you actually need to prove sameness Sometimes the claim you must establish genuinely is "these are not meaningfully different". The standard test cannot deliver it by failing to reject, because failing to reject is what an underpowered study does automatically. Establishing sameness requires flipping the setup so that the claim to be demonstrated is what the data must support, against a threshold of what counts as meaningfully different — a threshold that has to be stated in advance on substantive grounds. The key interview point is simply that this is a **different design**, decided before data collection, not a reinterpretation of a non-significant result after the fact. ## The language to use Acceptable: "we fail to reject `H0`", "the data are consistent with `H0`", "we found no evidence of a difference at this level", "any difference is smaller than we could detect here". Not acceptable: "we accept `H0`", "we proved there is no difference", "the groups are the same", "the effect is zero". The second list is not just loose phrasing. Each one converts a non-result into a claim the procedure never made, and once it enters a summary slide, nobody downstream can tell it apart from an actual finding. ## Interview framing A strong answer names the one-sided guarantee, explains that non-rejection is compatible with a range of non-null values, notes that weak designs fail to reject by default, and closes with what you would report instead. Candidates who stop at "you can never prove a negative" have the slogan but not the mechanism.