Estimate whether a labeling experiment is large enough to detect an accuracy improvement. Compare baseline and target accuracy with the sample per group to see approximate two-sided power, absolute lift, relative error reduction, and a suggested sample target.
The calculation uses equal group sizes and a two-sided normal approximation for two independent proportions.
Use the main result together with the supporting bounds, counts, or capacity figures. The estimate is only as reliable as the input definitions, sampling process, and operating assumptions.
Independent random samples are assumed. Paired reviews, repeated items, class imbalance, adjudication rules, and multiple metrics require a tailored analysis.
Comparing 88% with 92% using 500 observations per group produces a 4.0-percentage-point effect. The calculator converts that effect and its standard errors into approximate power.
Power is the approximate probability that the specified two-sided test detects the entered difference when that difference is real.
Small differences are harder to separate from random sampling variation, so more observations are required.
It is the number of observations per group. The displayed total is twice that value for two equal groups.
No. It describes a long-run detection probability under the assumptions, not a guarantee for one experiment.
Not directly. Paired or repeated measurements require a method that accounts for within-item correlation.
| Input | Meaning |
|---|---|
| Baseline rate | Expected control proportion |
| Target rate | Smallest effect of interest |
| Sample per group | Independent observations in each arm |