MerQuR+Statistics Glossary

Statistics Glossary → Mid-p value

Mid-p valueconcept

A less conservative p-value for exact tests, obtained by counting only half of the probability of the observed result.

Exact tests such as the binomial, Fisher and McNemar tests work with discrete distributions, so the p-value can only take certain values. Because the exact p puts the whole probability of the observed outcome on the 'extreme' side, the actual Type I error rate is often well below the nominal α; the test is more conservative than it needs to be. The mid-p adds half the probability of the observed outcome to the probability of more extreme outcomes: mid-p = P(more extreme) + ½·P(observed). For a two-sided test, twice the smaller one-sided mid-p is commonly used.

On average the mid-p keeps the Type I error rate closer to the nominal level and has more power than the exact p. The price is that there is no guarantee the error rate stays below α in every case; for some sample sizes it can slightly exceed it. The exact p is therefore preferred where a strict guarantee matters (for example in regulatory work), while the mid-p is a reasonable choice for method comparisons and exploratory studies.

An everyday example

A shop wants to know whether new packaging has cut the breakage rate below 10%. Only 1 of 40 boxes arrives broken. If the rate were still 10%, the chance of 1 or fewer broken boxes (the exact p) is about 0.080, and the chance of exactly 1 is about 0.066. The mid-p subtracts half of the latter, giving 0.080 - 0.033 = 0.047: the exact test misses significance at 0.05, while the mid-p reaches it.

In MerQur

It appears as the 'mid-p' checkbox in the Binomial Test and Fisher's Exact Test and as 'Extra outputs: mid-p' in the McNemar Test; in all three, half the probability of the observed outcome is removed from the exact p.

Analyses where this concept appears

Binomial TestFisher's Exact TestMcNemar Test

Related concepts

Exact test and exact p-valuep-value