A power and sample size calculator for cluster randomised trials

clustertrial.app calculates statistical power, sample size, standard errors and minimum detectable effects for cluster randomised trials, including stepped wedge, parallel, parallel with baseline, cross-over and staircase designs. Rather than selecting a design from a fixed list, you draw the allocation schedule directly on a grid of randomised sequences by time periods, so incomplete designs, implementation periods and irregular rollouts can all be costed accurately.

Power depends on how a trial will be analysed, so the calculator supports ten approaches to inference: model-based mixed model inference with normal or t reference distributions, Satterthwaite and Kenward–Roger small-sample corrections, generalised estimating equations with independence or exchangeable working correlation and robust sandwich variances, and classical design effect calculations for comparison. Continuous, binary and count outcomes are supported, with conditional and marginal estimands distinguished explicitly.

• Four correlation structures: exchangeable, nested exchangeable with a cluster autocorrelation, exponential decay, and exponential function with a lengthscale.
• Cross-sectional, closed cohort and open cohort sampling, with individual autocorrelation and a replacement rate.
• Optimal allocation weights for cells and sequences, rendered directly onto the design grid.
• Up to three designs open at once for side-by-side comparison of power, degrees of freedom, standard error and minimum detectable effect.
• Power curves and power surfaces exportable as PNG or SVG, and design diagrams exportable as PNG.
• Sessions saved to and restored from JSON, so a calculation can be revisited when assumptions change.
• Warnings when the selected estimator is materially more optimistic than the alternatives, and when specified correlation parameters cannot be realised in the fitted model.
• A verification bundle containing every matrix in the computation plus an R script that independently reproduces the standard error.

Everything runs locally in your browser through a WebAssembly build of the statistical engine. No design, parameter or result is transmitted anywhere, and there is nothing to install and no account to create.

• Getting started

  A step-by-step walkthrough from an empty grid to an exported bundle.

• Stepped wedge sample size

  The model, the parameters that matter, and a worked example.

• Choosing an estimator

  What each method assumes, and when it will flatter your design.

• Correlation structures

  ICC, CAC and IAC, and why binary outcomes constrain what is achievable.