Two things every field trial needs before it starts and rarely has in one place: an answer to “how small an effect could we detect with this sample?”, allowing for clustering, attrition, imperfect take-up and the number of outcomes; and a randomisation that is stratified, seeded, reproducible and documented, with a balance table produced at the moment of the draw. Both run in the browser. The baseline you upload for randomising never leaves it.
Outcome
Enter 1 to work in standard-deviation units.
A baseline measure of the outcome typically gives R² of 0.3 to 0.6; its square root is what matters, so a weak baseline buys little.
Design
Real-world adjustments
Bonferroni: α is divided by the number of outcomes. Holm and Romero–Wolf are less conservative but cannot be planned for in closed form.
Minimum detectable effect
Formula
Following Bloom (1995) and Duflo, Glennerster and Kremer (2007): MDE = (t1−α/2 + t1−β) · √[1 / (P(1−P))] · √[σ²(1−R²) / N] for individual randomisation, and the same with N = J·m and the design effect √[1 + (m−1)ρ] for cluster randomisation, with degrees of freedom N−2 or J−2 for the t quantiles. Attrition scales N by (1 − attrition) before the calculation, and imperfect compliance divides the intention-to-treat MDE by the difference in take-up to give the detectable effect on compliers (Wald/LATE). For a binary outcome σ² = p0(1 − p0). Multiple outcomes replace α by α/k. The required sample for a target effect inverts the same expression by bisection.
Sensitivity
Baseline
or
The demo is a simulated 720-household baseline in 36 villages with the covariates a typical survey has. It is labelled as such.