Power and Sample Size for Paired Equivalence Test

On this page
  1. Purpose
  2. Key Capabilities
  3. Statistical Methods
  4. Input Fields & Options
  5. Output Descriptions
  6. Worked Example

Purpose

Power and Sample Size for Paired Equivalence Test plans a study in which each unit is measured twice, for example by a new and a reference method, before and after a process change, or on two fixtures, and the aim is to show the mean paired difference is small enough to ignore. You set an equivalence window for the mean difference, and the tool tells you how many pairs are needed, what power a planned number of pairs gives, or how large a true difference can be tolerated. Enter any two of sample size, difference and power, and the tool solves for the third. Use it before a Paired Equivalence Test.

Key Capabilities

  • Solves for sample size (number of pairs), power or difference: leave blank the one you want.
  • Several values at once, e.g. sample sizes “10 20 30”, give one row for each combination.
  • Any equivalence window, symmetric or not.
  • Exact power for the two one-sided tests (TOST) procedure on the paired differences.
  • Power curve of power against number of pairs for each difference, with the solved points marked.

Statistical Methods

A paired study is analysed through the differences within each pair (test − reference), so the calculation is the one-sample TOST applied to those differences. With n pairs and standard deviation of the differences σ_d, the standard error is SE = σ_d/√n with n − 1 degrees of freedom. Equivalence is claimed when the mean difference is shown to be above the lower limit L and below the upper limit U, each one-sided t-test at significance level α (the Significance level (alpha) field, default 0.05).

Power is the probability that both one-sided tests succeed when the true mean difference is δ. It is computed exactly: for a given sample standard deviation of the differences s the chance of success is Φ(−t·s/σ_d − (δ − U)/SE) − Φ(t·s/σ_d − (δ − L)/SE), with t = t₁₋α,n−1, and this is averaged over the chi-square distribution of the sample variance by numerical integration. For sample size, the number of pairs that reaches the target power is found and rounded up, and the power at that number is reported as the actual power. When solving for the difference, the tool searches between the midpoint of the limits and the upper limit.

The standard deviation to enter is that of the paired differences, not of the individual measurements. Pairing pays off when the two measurements on a unit are strongly related, because σ_d is then much smaller than the unit-to-unit spread. The difference must lie strictly between the limits; power is highest at the midpoint of the window.

Input Fields & Options

Field

Description

Sample sizes

Number of pairs. Space or comma-separated, e.g. “20 30 40”. Leave blank to solve for sample size.

Difference(s) to detect

Space or comma-separated, e.g. “0.5 0.7”. The true mean paired difference you expect; must lie strictly between the limits. Leave blank to solve for it (searches between the midpoint of the limits and the upper limit).

Power values

Space or comma-separated, e.g. “0.8 0.9”. Leave blank to solve for power.

Lower equivalence limit / Upper equivalence limit

Required. The smallest and largest acceptable mean difference; the lower limit must be less than the upper limit.

Standard deviation of the differences

Required. Assumed or historical standard deviation of the paired differences.

Significance level (alpha)

Default 0.05. Applied to each of the two one-sided tests.

Power curve

Ticked by default: draws power against sample size.

Title

Title shown on the output and the power curve.

power-and-sample-size-paired-equivalence-test-inputs


Figure 1 — Solving for power: Sample sizes = 10 20 30, Difference(s) to detect = 0.1, limits −0.5 and 0.5, Standard deviation of the differences = 0.6, alpha = 0.05.

Output Descriptions

  • Settings — what was solved for, the lower and upper equivalence limits, the standard deviation of the differences and the significance level.
  • Results — one row per combination: difference, sample size (pairs), and power when solving for power, target and actual power when solving for sample size, or difference when solving for difference.
  • Power Curve — power against sample size for each difference, with the calculated points marked.
  • PDF export via Download.

power-and-sample-size-paired-equivalence-test-output


Figure 2 — Results for the example: power 0.520, 0.886 and 0.972 for 10, 20 and 30 panels when the true bias is 0.1 µm inside a ±0.5 µm window.

Worked Example

A paint shop wants to replace its reference coating-thickness gauge with a cheaper handheld unit. Each test panel will be measured with both gauges, and the handheld gauge is acceptable if its mean bias is within ±0.5 µm. A pilot run put the standard deviation of the paired differences at 0.6 µm, and the expected bias is about 0.1 µm. The metrologist enters Sample sizes = 10 20 30, Difference(s) to detect = 0.1, Lower equivalence limit = −0.5, Upper equivalence limit = 0.5 and Standard deviation of the differences = 0.6, leaves Power values blank and keeps alpha = 0.05.

The power is 0.520 with 10 panels, 0.886 with 20 and 0.972 with 30. Ten panels would leave about a one-in-two chance of failing a gauge that is in fact acceptable. Solving for sample size instead (Power values = 0.8 0.9) gives 16 panels for 80% power and 21 for 90%. If the bias turns out to be 0.2 µm, 20 panels give a power of 0.695; with no bias at all they give 0.948.