Power, by simulation.

Statistical power is the chance a study actually detects an effect that is really there. For simple tests you can get it from a formula. For the models researchers really use, you often cannot. This tool takes a different route: it builds data where the effect is real, runs the analysis, and checks how often the result comes out significant. Do that hundreds of times and the share of significant results is your power. Move the sliders and watch it change.

Design & assumptions

Effect size (Cohen’s d) iHow big the gap between the two groups is, in standard-deviation units. As a rough guide: 0.2 is small, 0.5 is medium, 0.8 is large.0.50

Sample size (per group) iHow many participants you plan to recruit in each group (or each cell in the 2×2 design). Type a number or drag the slider.

Significance level (α) iThe false-positive rate you are willing to tolerate: how often you would call a result significant when nothing is really there. .05 is the usual default.

Simulations iHow many datasets to simulate. More gives a steadier, less jumpy estimate but runs a little slower. 500 is a good balance.

…%

Main effect A

Main effect B

Interaction

iThe line shows how power grows as you add participants. Where it crosses the 80% mark is a common target for the sample size you need. The vertical line marks your current sample size.

Each square is one simulated study · coloured = it detected the effect

Methods write-up and R code

The paragraph updates with your settings, so you can paste it straight into a preregistration or a methods section. The download gives you runnable R code that reproduces this exact analysis, so anyone can check or adapt it.

How this works

Power formulas exist for simple tests, but they break down for the models applied research really uses, like logistic regression with correlated predictors, mixed models, and interactions. So instead of a formula, this tool simulates. It builds a dataset in which the effect you specified is real, runs the analysis, and checks whether the result is significant. Repeat that a few hundred times and the share of significant results is the estimated power. The same approach scales from this two-group example up to the multi-predictor logistic models I use in practice. It just needs a richer step for generating the data.