Understanding Deeply · 08 · Fictional data

One study’s p-value. Many studies’ power.

Two invented learning routines, A and B, produce a continuous task score. Higher is better. Test H₀: their population mean scores are equal. This simplified experiment uses independent normal outcomes and a known common standard deviation.

Try a story:

Before seeing the data

Now inspect one study

Start with an illustrative observed gap, or draw simulated group means.

Changing δ alone leaves this observed gap and its p-value unchanged. In a real study δ is unknown. Changing n or σ here explores a hypothetical study with the same observed gap.

The four possible outcomes of a test

The columns are different hypothetical truths. α and β condition on different columns; they are not the probabilities that a particular conclusion is wrong.

Test decisionH₀ is true: δ = 0
Reject H₀
Do not reject H₀
New random studies; the design stays fixed.

Run the experiment to compare theoretical probabilities with simulated frequencies.

What if you try many tests?

One pre-specified test at α = 5% is different from looking for a small p-value among many tests. This panel assumes all null hypotheses are true and the tests are independent.

The mathematical bridgeSample size → standard error → rejection boundary → power → observed p-value.
Interpretation, assumptions and sourcesA p-value is not the probability that the null is true.

What each quantity means

α is the long-run chance of rejection when H₀ is true, under the test’s assumptions. β is the chance of not rejecting at a specified nonzero effect. Power is 1 − β for that effect and design. A two-sided p-value measures how often H₀ would produce a mean difference at least as far from zero as the one observed, under the model. These quantities condition on different things.

What the result does not say

A p-value does not give the probability that H₀ is true, the probability the result occurred “by chance,” the size of an effect or its practical importance. Not rejecting H₀ does not establish no effect or equivalence. Statistical significance does not repair confounding or measurement bias. See the bias experiments.

Planning and reporting

Power here is calculated for a specified effect before interpreting study results. Replacing that effect with the observed estimate to calculate “observed power” is not a substitute for reporting the estimate and confidence interval. The displayed interval has confidence level 1 − α, matching the two-sided test. In repeated use under the model, that interval procedure covers the fixed population difference at the stated rate; it is not a posterior probability for this particular interval.

A deliberately simple model

Both fictional groups have independent, normally distributed scores and the same known population SD. B’s population mean is 50 and A’s is 50 + δ. The app draws sample means directly from their exact normal sampling distributions. It does not manufacture individual records. In real analyses, an estimated SD generally calls for a suitable t procedure or another model, and clustering, missingness and study design may require different methods.

Multiple testing

For independent true nulls, the chance of at least one false positive is 1 − (1 − α)ᵏ. Bonferroni uses α/k per test to control the family-wise error at at most α, including with dependent tests. It usually reduces power for each individual test. This toy panel does not model repeated peeking or optional stopping; those also require appropriate design and analysis.

All scores and settings are invented for teaching. Numerical normal probabilities use an approximation; the app labels very small probabilities as bounds. No real study or product is represented.