Power Analysis & Sample Size

Works out how many subjects an experimental group needs for a given power.Enter the effect size, significance level and power. A sentence for your protocol comes with the result.

⚠️ This sizes an experimental group. For how many specimens a diagnostic validation needs, see the Sample Size Calculator.

Result

Per group

26animals

Total
52
Power achieved
80.7%
With dropout
29 / 58

The sentence for your protocol

Each group was set at 26 animals. This assumes a two-group two-sided t-test at a significance level of 0.05 with 80 % power and an effect size of d = 0.8. Under those assumptions the minimum required sample size is 26 animals per group, giving an actual power of 80.7 %. Allowing for an expected dropout of 10 %, 29 animals per group are requested, 58 in total.

Copy it as it stands. Add one line saying where the effect size came from — that is the question review boards ask most.

Power curve

0%25%50%75%100%Power26Per group

Sample sizes are whole numbers, so the power achieved sits a little above the target — 80.7 %.

Where the curve flattens, more subjects buy almost no power. Stopping before that is what Reduction means in the 3Rs.

Computed exactly from the non-central F distribution, not from a normal approximation. A two-sided two-group t-test is the same calculation with one numerator degree of freedom.

If you do not know your effect size

This is where everyone stops. The three routes below are in order of how answerable they are.

1. From the difference that would matter

The only one you can answer before running anything. Give the change you want to see and how much your measurement wanders, both as a percentage of the control mean — the means cancel and only the ratio survives.

2. From a pilot or a published study

Two means, two standard deviations and two counts give Cohen's d over the pooled standard deviation.

3. The conventional labels (last resort)

Lakens (2013) restates Cohen's (1988) benchmarks as small d = 0.2, medium d = 0.5, large d = 0.8, and immediately warns that “these values are arbitrary and should not be interpreted rigidly” — to be used only where the finding is too novel to compare with anything.

How to use it

  1. Pick the test. Two groups is a t-test; three or more is a one-way ANOVA.
  2. Enter the effect size. If you do not have one, try route 1 under “if you do not know your effect size”.
  3. Check the significance level, power and expected dropout. The defaults are 0.05, 80 % and 10 %.
  4. Set the unit to whatever you are counting — animals, wells, plates. It goes into the sentence.
  5. Copy the protocol sentence into your document, and add a line saying where the effect size came from.

The formulas, and what to watch for

The sample size is the smallest one whose power reaches the target. The power itself is computed exactly from the non-central distribution.

Power

When the null hypothesis is false the test statistic follows a non-central F. Power is the chance that it lands past the critical value.

Non-centrality
λ = f² × total N
Power
P( F′(k−1, N−k, λ) > F_crit )
Two-group t-test
the same with one numerator df (|T|>c ⟺ T²>c²)

Effect size

The two-group d and the k-group f describe the same separation on different scales.

Cohen's d
(difference of means) / pooled SD
For two groups
f = d / 2
From change and CV
d = change(%) / CV(%)

Dropout

To survive dropout, divide the requirement by the fraction expected to remain and round up.

Requested
⌈ required / (1 − dropout) ⌉

What to watch for

  • ⚠️ This decides the number of subjects in an experimental group. The number of specimens for an in-vitro diagnostic validation is a different calculation, and this site's Sample Size Calculator is that one. The two share no code and no formula.
  • Sample sizes are whole numbers, so the power achieved lands a little above the target. The sentence carries the actual figure.
  • Cohen's benchmarks (0.2 / 0.5 / 0.8) are a last resort. Lakens (2013) notes that they are arbitrary and should not be interpreted rigidly.
  • The dropout allowance is a planning judgement, not something the statistics decide. Follow your institution's convention.
  • Where the power curve flattens, more subjects buy almost nothing. Stopping before that is Reduction in the 3Rs.
  • Only two-sided tests are calculated. Using a one-sided test needs its own justification in the protocol, and that judgement is not one this tool can make for you.

Frequently asked questions