How to plan a study 01 Pick the test that matches the design. Two independent means is tTwo; paired measurements is tPaired with d<sub>z</sub>; proportions are propTwo with h. 02 Choose what to solve for. Power, N or effect — the other two must be supplied. 03 Enter the effect in the test's own units. d = 0.5 is half a standard deviation; h = 0.541 is an arcsine-transformed proportion difference. The panel converts between conventions (d ↔ r, η² ↔ f, proportions ↔ h) so the effect can be stated in the form your source uses. 04 Set α, tails and any correction. Two-tailed α = 0.05 is the default; a Bonferroni or Šidák correction across k comparisons is applied to the effective α before the power computation. 05 Read the distribution card. Power, N (split into n₁ and n₂), critical value, df and the noncentrality parameter δ are all reported, which is what makes the number auditable.
Worked readouts
Design · input
Readout
Two-sample t, d = 0.5, α = 0.05, 80 % power, two-tailed
N = 128 (64 + 64), power 80.146 % , critical t 1.978971 , df 126 , δ 2.82843
The same design, one-tailed
N = 101 (51 + 50), power 80.241 % , critical t 1.660391 , df 99
The same design at N = 100
power 69.689 % (δ 2.5 , critical t 1.984467 , df 98 )
Solve for effect at N = 100
minimum detectable d = 0.56588
Target 90 % power
N = 171 (86 + 85), δ 3.26912
d = 0.2 instead of 0.5
N = 787 (394 + 393)
Bonferroni across 2 comparisons
effective α 0.025 , N = 155 (78 + 77)
Two proportions, h = 0.541
N = 108 (54 + 54)
A/B: 120/1000 vs 150/1000
z 1.96305 , p 0.04964 , significant; post-hoc power at observed rates 50.203 % ; N for 80 % power at those rates 2036 per group
Why the answer comes from a noncentral distribution Under the null hypothesis the two-sample t statistic follows a central t distribution with 126 degrees of freedom, and the test rejects when |t| exceeds 1.978971 . Under the alternative it follows a noncentral t with the same df and a noncentrality δ = d·√(n₁n₂/(n₁+n₂)) = 2.82843 at d = 0.5 and 64 per arm. Power is the probability that a draw from that shifted distribution lands beyond the critical value — 80.146 % . Nothing about the computation is a rule of thumb: it integrates the noncentral density, and the panel reports δ and the critical value so the result can be reproduced.
The same machinery inverts in the other two directions. To find N, the solver increases n until the noncentral probability crosses the target; the jump from 128 to 171 observations for 90 % power is the price of the extra assurance. To find the minimum detectable effect, it searches the effect that gives exactly the target power at fixed N; at N = 100 that is d = 0.56588 , so the design cannot reliably see the d = 0.5 effect it was sized for. The one-tailed row shows the leverage of a directional hypothesis: the same 80 % power needs 101 observations instead of 128, and a Bonferroni correction across two comparisons raises N to 155 .
Limits and edges
Power is a property of the design, not the data. The numbers describe the probability of a future experiment under an assumed effect; once the data are in, power is not a statistic you can observe. The A/B view's post-hoc figure is a diagnostic, not evidence.
Every answer inherits the assumed effect and distribution. d = 0.5 from a pilot is an estimate with its own uncertainty; the two-proportion test assumes the arcsine approximation, and exact binomial power exists for small single-proportion studies where discreteness matters. What it does not do. It does not run the hypothesis test on your data, fit a model or adjust for clustering and missingness. Test real samples in Hypothesis Tests , summarise them in Descriptive , and monitor a running process with the SPC panel .
The observed-power trap The A/B example is the trap in one table: 120/1000 against 150/1000 gives z = 1.96305 and p = 0.04964 , significant at α = 0.05, yet the post-hoc power at those observed rates is only 50.203 % — a coin flip. The result is not wrong; it is fragile. Quoting post-hoc power as if it validated the finding confuses the probability of the data under the null with the probability of a future study. The honest move is to size the next study: at those rates, 80 % power needs 2036 per group.
The second trap is the units of the effect. d = 0.5 and h = 0.541 are both "medium" by their conventions but are not interchangeable; applying a conversion to the wrong test changes the required N. The third is discreteness: for the exact binomial example with n = 30, p₀ = 0.25 and p₁ = 0.5, the nominal α = 0.05 admits only the integer rejection region X ≤ 2 or X ≥ 13, whose actual two-sided size is α = P(X ≤ 2 or X ≥ 13 | p₀) = 0.032190 , and whose power is the same region's probability under p₁ = 0.5, 81.920 % . When n is small, read the actual α, not the one you typed.
Where it earns its place Sizing a two-arm experiment For a half-standard-deviation effect, two-tailed at 80 % power, the answer is 128 observations; doubling the assurance to 90 % costs 171 , and correcting for two comparisons costs 155 . Each number comes with its δ and critical value so a reviewer can check the arithmetic.
Reading an underpowered result correctly A pilot with 50 per arm and d = 0.5 has power 69.7 % , so a non-significant result is not evidence of no effect — it is the expected outcome in three of ten runs. "Solve for effect" turns that into a concrete statement: the pilot can only reliably detect d = 0.566 and above.
Privacy All power calculations, curves and A/B results are computed in your browser; nothing is uploaded.
References
Wikipedia, Power of a test , en.wikipedia.org (访问日期:2026-10-07)— power, β and the design interpretation.
Wikipedia, Noncentral t-distribution , en.wikipedia.org (访问日期:2026-10-07)— the shifted distribution behind the t-test power.
Wikipedia, Sample size determination , en.wikipedia.org (访问日期:2026-10-07)— inverting power for N and effect size.
Wikipedia, Bonferroni correction , en.wikipedia.org (访问日期:2026-10-07)— the multiplicity correction applied to α.
G*Power (Faul et al., 2007), psychologie.hhu.de (访问日期:2026-10-07)— the reference tool for noncentral-distribution power analysis.
Calculators in this hub
Hand-picked tools, one click away. The mini versions compute live and carry your values into the full calculator.
Sources & review
Reviewed by CalcX Editorial Team
Updated 2026-10-07