Statistical power

Statistical power for a one-proportion hypothesis test

Compute the probability of detecting a real effect (power) or the sample size needed for the test H₀: p = p₀. Visualize the distributions under H₀ and H₁ with their rejection regions.

Calculator

Select the mode based on what you want to compute and enter the parameters.

Result pending…
Result pending…
H₀: distribution under the null hypothesis Critical region (α) H₁: distribution under the alternative Power (1−β)

How to read the chart

The horizontal axis is the standardized Z statistic \(Z = (\hat{p} - p_0)/SE_0\). The two curves show its distribution under two scenarios:

  • Blue curve (H₀): distribution of Z if the real proportion were p₀, which follows approximately an N(0,1). The zones shaded in red are the rejection regions; their total area is α.
  • Dashed orange curve (H₁): distribution of Z if the real proportion were p₁. It is shifted according to the size of the effect. The green zone — the area under H₁ within the rejection region — is the power (1−β). The unshaded area under the orange curve within the acceptance region is β, the type II error.

The larger |p₁ − p₀| or the sample size n, the more the two curves separate and the greater the power (a wider green zone).

Statistical foundation

For the test \(H_0: p = p_0\) with the normal approximation, the statistic is:

\( Z = \dfrac{\hat{p} - p_0}{\sqrt{p_0(1-p_0)/n}} \sim N(0,\,1) \text{ under } H_0 \)

Under the alternative hypothesis \(H_1: p = p_1\), the same statistic is distributed approximately as:

\( Z \sim N\!\left(\delta,\; \sigma_1^2\right) \quad \text{where} \quad \delta = \frac{(p_1 - p_0)\sqrt{n}}{\sqrt{p_0(1-p_0)}},\quad \sigma_1 = \sqrt{\frac{p_1(1-p_1)}{p_0(1-p_0)}} \)

The parameter \(\delta\) is the non-centrality: how many standard deviations (under H₀) the statistic shifts when p = p₁. The larger \(|\delta|\), the easier it is to distinguish H₁ from H₀.

Power formulas

  • Two-tailed (\(H_a: p \neq p_0\)):

    \( 1 - \beta \;=\; 1 - \Phi\!\left(\frac{z_{\alpha/2} - \delta}{\sigma_1}\right) + \Phi\!\left(\frac{-z_{\alpha/2} - \delta}{\sigma_1}\right) \)

  • Upper one-tailed (\(H_a: p > p_0\)):

    \( 1 - \beta \;=\; 1 - \Phi\!\left(\frac{z_\alpha - \delta}{\sigma_1}\right) \)

  • Lower one-tailed (\(H_a: p < p_0\)):

    \( 1 - \beta \;=\; \Phi\!\left(\frac{-z_\alpha - \delta}{\sigma_1}\right) \)

Sample size formula

To achieve power \(1-\beta\) with level \(\alpha\) and effect \(\Delta = |p_1 - p_0|\):

\( n \;=\; \left\lceil\left(\frac{z_{\alpha^*}\sqrt{p_0(1-p_0)} + z_{1-\beta}\sqrt{p_1(1-p_1)}}{\Delta}\right)^2\right\rceil \)

where \(z_{\alpha^*} = z_{\alpha/2}\) for two-tailed tests and \(z_{\alpha^*} = z_\alpha\) for one-tailed tests.

Worked example

A company wants to check whether its satisfaction rate p₀ = 0.30 has changed after an improvement campaign, expecting it to rise to p₁ = 0.40. A two-tailed test H₀: p = 0.30 is run with α = 0.05 and a sample of n = 172 respondents.

Step 1 — Standard error under H₀:

\( SE_0 = \sqrt{p_0(1-p_0)/n} = \sqrt{0.30 \times 0.70 / 172} \approx 0.03494 \)

Step 2 — Non-centrality δ:

\( \delta = \frac{p_1 - p_0}{SE_0} = \frac{0.10}{0.03494} \approx 2.862 \)

Step 3 — Relative variance under H₁:

\( \sigma_1 = \sqrt{\frac{0.40 \times 0.60}{0.30 \times 0.70}} = \sqrt{\frac{0.24}{0.21}} \approx 1.069 \)

Step 4 — Two-tailed power (\(z_{\alpha/2} = 1.960\)):

\( 1-\beta = 1 - \Phi\!\left(\frac{1.960 - 2.862}{1.069}\right) + \Phi\!\left(\frac{-1.960 - 2.862}{1.069}\right) = 1 - \Phi(-0.844) + \Phi(-4.511) \approx 0.801 \)

With n = 172 and p₁ = 0.40, the power is ≈ 80.1%. If the real proportion has risen to 0.40, the test will correctly detect it in about 8 out of every 10 studies. The type II error is β ≈ 0.199: there is a 20% probability of not detecting the improvement even though it is real.

How to interpret the results

Power (1−β)

A power of 80% means that if the real proportion is p₁, the test will reject H₀ 80% of the time. In confirmatory studies, a power of ≥ 0.80 is recommended; in decisions where the cost of a false negative is high, ≥ 0.90.

Type II error (β)

This is the complement: β = 1 − power. With β = 0.20, one in five studies with the same design would fail to detect the effect even if it exists. Low power makes negative results uninformative: not rejecting H₀ does not mean the effect is null.

Sample size

The computed n is the minimum theoretical value under the normal approximation. In practice, apply a margin for losses: divide n by (1 − expected non-response rate). With 10% losses, plan for ⌈n / 0.90⌉ contacts.

Practical recommendations

  • If power is low (< 0.70), you need a larger sample or a bigger effect for the study to be reliable.
  • The one-tailed test has more power than the two-tailed test for the same n, but it only detects changes in one direction and must be justified a priori.
  • When p₀ or p₁ are extreme (< 0.10 or > 0.90) with a small n, consider the exact binomial test instead of the normal approximation.
  • Run a sensitivity analysis varying p₁ ±0.05 to assess the robustness of the design.

Frequently asked questions

  • What is statistical power? The probability of rejecting H₀ when the alternative hypothesis is true, that is, the probability of detecting a real effect.
  • What is type II error (β)? The probability of not detecting an effect that really exists. It is complementary to power: β = 1 − power.
  • What do the critical regions in the chart show? The red zones are the rejection regions under H₀ (total area = α). The green zones are the area of H₁ within those same regions, representing the power (1−β).
  • When should I use a two-tailed test versus a one-tailed test? Use two-tailed when you have no expected direction beforehand. The one-tailed test has more power but only detects changes in one direction.