🧮  Post Hoc Power Analysis

Post Hoc Power Analysis#

Computing achieved power after a study from the observed effect — widely criticised as uninformative.

Important

✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.

What it is#

Post-hoc power analysis computes the statistical power of a test after the study is finished, plugging in the observed sample size \(n\), the observed effect size \(\delta\), and the chosen \(\alpha\). It asks: given what we actually saw, what was the probability we could have detected an effect?

Why people run it#

Usually to interpret a non-significant result (“was it real-but-missed, or genuinely null?”), to satisfy a journal asking about sensitivity, or to judge older studies in a meta-analysis. The calculation is the a-priori one with the observed effect size substituted in:

\[\text{Power} = P(\text{reject } H_0 \mid \delta_{\text{observed}}, n, \alpha).\]

Example#

If \(H_0\) is a 10% conversion rate, the treatment shows a tiny 10.2%, with 1,000 per group at \(\alpha = 0.05\), post-hoc power might be only 12% — the study was underpowered to detect so small a lift.

The tautology problem#

The deep flaw: post-hoc power is a deterministic function of the p-value, so it adds nothing. A non-significant result always yields low post-hoc power, and a significant one always high — it merely restates the test. Worse, it invites the fallacy “non-significant + low power ⇒ \(H_0\) is true,” when it only means “this study wasn’t sensitive enough.”

Report this instead#

Rather than post-hoc power, report the observed effect size (Cohen’s d, a difference in proportions, an odds ratio) and a confidence interval for the effect — these convey the strength and precision of the result without the circularity.


Theme: Statistical Inference & Power  ·  All terminology



See also

Source article Adapted (context, re-expressed) in our own words from: Post Hoc Power Analysis (insightful-data-lab.com).

Tags: purpose: reference topic: terminology level: beginner