Skip to content

Halving every two-sided p-value

In the symmetric normal case, half the two-sided p-value is the smaller tail; it need not be the planned one-sided p-value.

On this page

Locate the lost direction

The shortcut discards whether the test statistic is positive or negative. An upper alternative looks for positive evidence, while a lower alternative looks for negative evidence. The same absolute statistic can therefore imply a small or large one-sided probability.

Use a concrete counterexample

For z = −2, the two-sided normal p-value is about 0.0455. Half is about 0.02275, but the upper-sided p-value is about 0.97725. The shortcut would turn evidence pointing away from an increase into apparent evidence for it.

Reconstruct the intended tail

Repair checklist
StepCheck
AlternativeWas it upper, lower or two-sided before the result?
StatisticWhat is its signed value?
DistributionIs the null reference symmetric and continuous?
TailWhich event is counted by the intended procedure?

Avoid extending the shortcut to every procedure

Doubling a smaller tail is a particular symmetric convention. Do not apply it automatically to asymmetric distributions or discrete tests without reading the method definition. Preserve the originally planned analysis in the report.

Further references