Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Biostatistics / Sample size and power

Clusters, non-inferiority, attrition and writing the section

14 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Apply the design effect for cluster-randomised and cluster-sampled studies

  • Explain non-inferiority and equivalence designs and calculate the sample size

  • Inflate sample size for loss to follow-up and write a clear sample size paragraph

Cluster designs and the design effect

When you randomise or sample groups (villages, schools, clinics), people in the same group resemble each other. The intracluster correlation coefficient (ICC, ρ) measures this: the share of total variation that lies between clusters. Typical values: 0.01 to 0.05 for many outcomes; higher for infectious disease and behaviour.

Observations within a cluster carry less independent information than the same number of independent people. The design effect (DEFF) says how much to inflate the sample:

DEFF = 1 + (m - 1)ρ

where m is the average cluster size. Multiply the sample size for an individually randomised trial by DEFF.

Worked example

An individually randomised trial would need 300 children per arm. In a village-randomised trial with m = 30 children per village and ρ = 0.05:

DEFF = 1 + 29 × 0.05 = 2.45, so you need 300 × 2.45 = 735 children per arm, which is 25 villages per arm (735 / 30 = 24.5).

Even a modest ICC of 0.05 more than doubles the sample size. Adding more clusters is far more efficient than adding more people per cluster: with ρ = 0.05, going from 30 to 60 per cluster raises DEFF to 3.95, so you gain little. The upper limit of efficiency is 1/ρ people per cluster.

0246110203050100Average cluster size (m)Design effectDesign effect (ICC = 0.05)
The design effect rises linearly with cluster size, so large clusters waste participants.

Analyse clusters as clusters

A cluster trial analysed as if every person were independent gives falsely narrow intervals and too many 'significant' results. Use cluster-level summaries, mixed-effects models or GEE. Few clusters (under about 15 per arm) need small-sample corrections.

Non-inferiority and equivalence

A superiority trial asks 'is the new treatment better?'. Often the new treatment is cheaper, safer or simpler and only needs to be not unacceptably worse.

  • Non-inferiority: the new treatment is not worse than the standard by more than a pre-specified margin (δ).
  • Equivalence: not worse and not better than δ (two-sided).

The hypotheses are reversed: H0 says the new treatment is worse by at least δ; rejecting it shows non-inferiority. It is tested by checking whether the lower limit of the CI for the difference lies above −δ.

StudyEstimate (95% CI)A: non-inferior (and not superior)-1 (-6–4)B: inconclusive-4 (-12–4)C: non-inferior and superior7 (2–12)D: inferior-12 (-17–-7)-10010Difference in cure rate (percentage points); non-inferiority margin = −10
Non-inferiority is shown when the entire CI lies to the right of the margin (here −10). Cases A and C pass; B and D do not.

Worked example: a simplified regimen

Standard regimen cures 80% of patients. A shorter regimen is acceptable if it is not worse than 10 percentage points lower (δ = 0.10). One-sided α = 0.025, 80% power, expecting equal 80% cure:

n = (2 p(1-p) (z_(1-α) + z_(1-β))^2)/(\delta^2) = (2 × 0.16 × 7.85)/(0.01) = 251.2, so 252 per arm.

Choose δ by clinical reasoning and historical evidence of the standard treatment's benefit over placebo (preserving a fraction of that effect). A margin chosen to make the trial feasible is not a justification. Analyse by both intention-to-treat and per-protocol, because ITT biases toward 'no difference', making non-inferiority easier to claim (chapter 10).

Inflating for losses

Some participants will drop out, withdraw or be missed. If you expect a fraction L lost, divide by 1 - L:

n_enrol = (n_analysis)/(1 - L)

With 294 needed per arm and 15% loss: 294 / 0.85 = 345.9, so 346 per arm. Never add 15% (294 × 1.15 = 338), because that undershoots.

Other adjustments

  • Interim analyses with stopping rules slightly raise the needed sample (group sequential designs, chapter 10).
  • Multiple primary outcomes or arms need α adjusted or the largest n.
  • Unequal allocation (2:1) increases total n slightly but may improve recruitment or safety data.
  • Crossover designs need fewer participants but need washout and stable disease.

A model sample-size paragraph

'The primary outcome is clinical cure at 28 days. Assuming cure of 70% with standard care and detecting an improvement to 80% (an absolute difference of 10 percentage points that the investigators judge the minimum worth adopting), with a two-sided α of 0.05 and 80% power, 294 participants per arm are required (588 total). Allowing for 15% loss to follow-up, 346 per arm (692) will be enrolled. Calculations used the normal approximation with pooled variance (software: R, package pwr).'

Use the same checklist every time

Outcome and comparison, α, power, effect and its justification, variability and its source, design effect and loss adjustment, final numbers per arm and in total, software or formula.

Summary

  • Cluster designs need DEFF = 1 + (m − 1)ρ times the individually randomised sample; more clusters beat bigger clusters.

  • Non-inferiority tests whether the CI lies above −δ; the margin must be justified clinically.

  • Inflate for dropouts by dividing by (1 − loss), not by adding a percentage.

  • Write the calculation transparently, with sources and a sensitivity analysis.

Check your understanding

Question 1
1

With m = 20 and ICC = 0.05, the design effect is:

Question 2
2

For a cluster trial, the cheapest way to gain power is typically to:

Question 3
3

In a non-inferiority trial, non-inferiority is shown if:

Question 4
4

Required analysis n = 200 and 20% loss expected. Enrol:

Question 5
5

Which statement about ITT in non-inferiority trials is true?

Question 6
6

A study reports 'n = 100 per arm is enough' without any justification. The reviewer should:

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Calculating sample size for means, proportions and eventsIntention-to-treat, per-protocol and missing data