Sample Size for Comparing Two Independent Proportions

A guide to determining the sample size for studies comparing success rates, response rates, or event proportions between two groups.

Overview

This calculator determines the number of subjects needed per group to detect a specified difference between two population proportions. It is applicable to randomised controlled trials, cohort studies, and any study comparing binary outcomes (e.g., response/no response, event/no event) between two independent groups.

The calculator supports three hypothesis frameworks: equality (is there a difference?), non-inferiority/superiority (is the new treatment not worse/better by a specified margin?), and equivalence (are the two treatments sufficiently similar?).

An optional continuity correction can be applied to account for the discrete nature of proportions when using the normal approximation. Unequal allocation ratios are also supported. Both options are explained in detail in Methods & Options below.

Methods & Options: Continuity Correction and Allocation Ratio

Two optional refinements are layered on top of the basic two-proportion formula. The continuity correction adapts the formula to the discreteness of real count data; the allocation ratio adapts it to real-world recruiting constraints. Each is unpacked below.

1. Continuity Correction

The problem it fixes

The count of “successes” in each group is a whole number — you cannot have 17.3 responders. The data are discrete: they jump in unit steps. But the standard formula uses the normal approximation, which assumes a smooth, continuous bell curve. This mismatch makes the approximation slightly optimistic — it suggests the test has more power than it actually does, so the calculated sample size comes out a little too small.

What the correction does

The continuity correction (Fleiss et al.) inflates the sample size by a small amount to compensate for the discreteness, using the \( n_c \) formula given in the Formula section. The correction is largest when sample sizes are small or when the two proportions are close together, and its effect shrinks as \( n \) grows.

In the worked example above (p₀ = 0.60 vs p₁ = 0.75), the correction raises the requirement from 150 to 163 per group — about 9% more — to genuinely achieve the specified 80% power.

When to use it

Use it (the default in Statulator) when sample sizes are modest — up to a few hundred per group — or when the proportions are close together. It is the more conservative, safer choice: the study will actually achieve the power you asked for. The convention in clinical trial design is to apply it by default, because an underpowered study is a worse failure mode than a slightly overpowered one.
Skip it for very large studies, or when you need results to match software or a published protocol that did not apply the correction.

2. Unequal Allocation Ratios

What it means

By default, the formula assumes you will recruit the same number of subjects in each group (a 1:1 ratio — half treatment, half control). The allocation ratio \( r = n_2 / n_1 \) lets you plan a different split:

r = 1 — equal allocation (50/50 split)
r = 2 — twice as many subjects in Group 2 as in Group 1 (2:1)
r = 0.5 — twice as many in Group 1 as in Group 2 (1:2)

Why you might want unequal allocation

Cost or availability differences. If recruiting controls is cheap (e.g., pulling them from a registry) but treatment is expensive (a new drug, a complicated procedure), recruiting more controls per treated subject gives better statistical efficiency for the budget. A 1:2 or 1:3 ratio is common.
Ethical preference for the new treatment. In a trial of a promising therapy, investigators sometimes want more patients to receive the active treatment than the control. A 2:1 ratio (treatment:control) achieves this while keeping the trial scientifically valid.
Rare-condition pragmatics. If the treatment group must come from a small eligible population, unequal allocation may be forced on you simply because there are not enough eligible patients to fill a 1:1 design.
Existing imbalances. In observational cohort studies, group sizes often cannot be controlled — the data are what they are, and the formula must accommodate whatever ratio nature gave you.

The trade-off

Equal allocation (1:1) is the most statistically efficient: it gives the smallest total sample size for a given power. Any deviation from 1:1 increases the total. Starting from the equal-allocation per-group size \( n_{\text{equal}} \), the group sizes become:

\[ n_1 = n_{\text{equal}} \times \frac{1 + r}{2r}, \qquad n_2 = r \times n_1 \]

Worked example: suppose the 1:1 calculation gives 100 per group (200 total). With a 2:1 allocation (r = 2): n₁ = 100 × (1 + 2)/(2 × 2) = 75 and n₂ = 2 × 75 = 150, so the total rises to 225 instead of 200 — 12.5% more subjects than equal allocation, in exchange for having more subjects in Group 2.

The sweet spot: most clinical trials stay between 1:1 and 1:3. Beyond about 1:4 the returns diminish rapidly — adding more subjects to one group barely improves power, because the smaller group becomes the limiting factor.

Worked Example

Scenario: Comparing Treatment Response Rates in a Clinical Trial

A clinical researcher is planning a randomised trial comparing a new drug to the standard treatment for a skin condition. The standard treatment has a known response rate of p₀ = 0.60 (60%). The new drug is expected to achieve p₁ = 0.75 (75%). The study requires 80% power at a 5% two-sided significance level with equal allocation.

Using Statulator step-by-step:

1 Open the Sample Size Calculator for Comparing Two Independent Proportions.

2 Select Hypothesis: Equality.

3 Set Significance Level (α) to 0.05 and Power to 0.80.

4 Enter Proportion in Group 1 (p₀) as 0.60.

5 Enter Proportion in Group 2 (p₁) as 0.75.

6 Keep Allocation Ratio at 1.

7 The result should be n = 150 per group (300 total) without continuity correction, or n = 163 per group (326 total) with continuity correction (the default).

Hand calculation verification (without continuity correction):

\[ n = \frac{(z_{\alpha/2} + z_{\beta})^{2} \left[\frac{p_0(1-p_0)}{r} + p_1(1-p_1)\right]}{(p_1 - p_0)^{2}} \] \[ = \frac{(1.96 + 0.842)^{2} \times [0.60 \times 0.40 + 0.75 \times 0.25]}{(0.15)^{2}} = \frac{7.849 \times [0.24 + 0.1875]}{0.0225} = \frac{7.849 \times 0.4275}{0.0225} \approx 149.1 \]

Rounding up: n = 150 per group, matching Statulator’s output when the continuity correction is turned off.

Continuity Correction

The continuity correction compensates for using a continuous (normal) distribution to approximate a discrete (binomial) distribution. It slightly increases the sample size and is recommended for small to moderate sample sizes. The corrected sample size is:

\[ n_c = \frac{n}{4} \left(1 + \sqrt{1 + \frac{2(r+1)}{n \cdot r \cdot |p_1 - p_0|}}\right)^{2} \]

Interpretation Guide

Calculator outputs and how to interpret them
Output	Interpretation
Sample Size per Group	The number of subjects needed in each arm. With unequal allocation (r ≠ 1): Group 1 = n, Group 2 = n × r.
With/Without Continuity Correction	The corrected value is always equal to or larger than the uncorrected value. The correction is most important when the expected difference is small.
Non-Inferiority Margin	In non-inferiority designs, this is the largest acceptable amount by which the new treatment can be worse than the standard. A negative margin (e.g., −0.10) means the new treatment can be up to 10 percentage points worse and still be considered non-inferior.
Equivalence Margin	In equivalence designs, both treatments must be within ±δ of each other. The margin must be larger than the assumed true difference.

Formula

Equality (Two-Sided Test)

\[ n = \frac{(z_{\alpha/2} + z_{\beta})^{2} \left[\dfrac{p_0(1-p_0)}{r} + p_1(1-p_1)\right]}{(p_1 - p_0)^{2}} \]

Non-Inferiority / Superiority (One-Sided)

\[ n = \frac{(z_{\alpha} + z_{\beta})^{2} \left[\dfrac{p_0(1-p_0)}{r} + p_1(1-p_1)\right]}{(p_1 - p_0 - \delta_m)^{2}} \]

Equivalence (TOST)

\[ n = \frac{(z_{\alpha} + z_{\beta/2})^{2} \left[\dfrac{p_0(1-p_0)}{r} + p_1(1-p_1)\right]}{(\delta_m - |p_1 - p_0|)^{2}} \]

For equivalence testing (two one-sided tests, TOST), the type II error is split between the two one-sided tests, so the critical value is \( z_{\beta/2} \) rather than \( z_{\beta} \).

where:

\( p_0 \) = proportion in the control/reference group
\( p_1 \) = proportion in the treatment/experimental group
\( r \) = allocation ratio (Group 2 / Group 1)
\( \delta_m \) = non-inferiority, superiority, or equivalence margin
\( z_{\alpha/2}, z_{\alpha} \) = z-critical values for two-sided and one-sided tests
\( z_{\beta} \) = z-critical value for power

Continuity Correction (Fleiss et al.)

\[ n_c = \frac{n}{4} \left(1 + \sqrt{1 + \frac{2(r + 1)}{n \cdot r \cdot |p_1 - p_0|}}\right)^{2} \]

Cluster Sampling Adjustment

\[ n_{\text{cluster}} = n \times [1 + (m - 1) \cdot \rho] \]

Assumptions & Requirements

Independent groups: Subjects in the two groups are sampled independently. For paired binary data (e.g., same subjects measured twice), use the Paired Proportions calculator.
Binary outcome: Each subject has a binary outcome (success/failure, event/no event).
Normal approximation: The formula uses the normal approximation to the binomial distribution. Apply the continuity correction when the expected proportions or differences are small.
Known expected proportions: Both p₀ and p₁ must be specified. These are typically estimated from prior studies or clinical judgement.
Fixed allocation ratio: The ratio of group sizes is determined before the study begins and does not change.

Studies with More Than Two Arms

This calculator compares two groups, but many real studies have three or more arms — for example, a trial comparing control, drug 1, drug 2, and drug 3. Statulator does not yet have a dedicated multi-arm calculator; until it does, the two-group calculators can handle multi-arm designs with the recipe below. The same approach works for binary outcomes (this calculator) and continuous outcomes (the Two Independent Means calculator).

Step 1 — Be Clear About What You Are Actually Testing

This is the most important question: “comparing k groups” can mean very different things, each leading to a different sample size. Decide before recruiting:

Omnibus test — “Is there any difference among the groups?” One overall test: a chi-square test across the 2 × k table for a binary outcome, or a one-way ANOVA for a continuous outcome.
Multiple pairwise comparisons — “Does each drug differ from control? Do the drugs differ from each other?” A family of two-group tests: 6 possible pairs among 4 groups, or 3 if each drug is compared only with control.
A specific planned contrast — e.g., “Does the pooled drug group differ from control?” A single pre-specified comparison.

Step 2 — The Pairwise Recipe (Bonferroni by Hand)

1 List the pairwise comparisons you actually plan to test (e.g., each drug vs. control = 3 comparisons in a four-arm trial).

2 Bonferroni-adjust the significance level: divide α by the number of comparisons c. For α = 0.05 and 3 comparisons, enter α = 0.05/3 ≈ 0.0167 in the calculator instead of 0.05.

3 Run the two-group calculator for the smallest difference you care about detecting — the hardest comparison is the one that drives the sample size. Sizing for the largest difference would leave the harder comparisons underpowered.

4 If different pairs involve different proportions, repeat the calculation for each pair and take the maximum per-group n. Apply that n to every arm of the study.

Choosing the Comparison Structure

Each treatment vs. control only (k − 1 comparisons for k − 1 treatments): the purpose-built method is Dunnett’s test, which is more efficient but requires special tables. Bonferroni at α/(k − 1) — e.g., 0.05/3 ≈ 0.0167 for three treatments — is a usable, slightly conservative approximation.
All pairs (k(k − 1)/2 comparisons; 6 among 4 groups): the purpose-built method is Tukey’s HSD. Bonferroni at α/6 ≈ 0.008 is a conservative proxy, and the required sample is noticeably larger.
A single planned contrast (e.g., pooled drugs vs. control): no multiplicity correction is needed. Run the calculator with the contrast’s expected proportions, using the allocation ratio to reflect the pooling (e.g., r = 3 when three equally sized drug arms are pooled against one control arm).

Worked Example: Four-Arm Trial with a Binary Outcome

A four-arm trial (control, drug 1, drug 2, drug 3) with a binary response outcome. Expected response rates: control = 0.30, drug 1 = 0.40, drug 2 = 0.45, drug 3 = 0.50. Plan: compare each drug with control (3 comparisons) at overall α = 0.05 (two-sided) with 80% power.

1 Identify the smallest difference of interest: control vs. drug 1 (0.30 vs. 0.40, a 10 pp difference). This is the limiting comparison — the hardest to detect.

2 Bonferroni-adjust the significance level: α = 0.05/3 ≈ 0.0167.

3 Run this calculator with p₀ = 0.30, p₁ = 0.40, α = 0.0167, power = 0.80, allocation ratio 1. It returns approximately n ≈ 491 per group with the continuity correction (≈ 472 without).

4 Apply this n to all four arms: total ≈ 1,964 subjects.

Practical Tips

Allocation: Equal allocation across arms is usually best for omnibus tests. When each treatment is compared only with control, the control arm participates in every comparison, so some designers allocate more subjects to control — the square-root rule uses roughly \( \sqrt{k} \) times as many controls as each of the k treatment arms (about twice as many in the example above).
Dropouts: Inflate for expected loss to follow-up — with 15% expected attrition, divide the sample size by 0.85.
Multiplicity is unavoidable. Skipping the correction roughly doubles or triples the false-positive rate, and reviewers and journals will ask. Build the correction into the design, not just the analysis.
Effect-size honesty. The biggest pitfall in multi-arm sizing is not the arithmetic but optimistic effect sizes — anchor the expected proportions on pilot or published data wherever possible.

Textbook Examples

Medicine

An RCT compares a new antibiotic (expected cure rate 85%) to the standard treatment (cure rate 70%).

Inputs: p₁ = 0.85, p₂ = 0.70, α = 0.05 (two-sided), power = 80%, allocation ratio 1:1, continuity correction applied (default).
Result: n = 131 per group (262 total).
Interpretation: Enrolling 131 patients per arm gives 80% power to detect the 15 pp difference at the 5% significance level.

Education

Researchers test whether a new teaching method increases the pass rate from 60% to 75%.

Inputs: p₁ = 0.75, p₂ = 0.60, α = 0.05 (two-sided), power = 90%, continuity correction applied (default).
Result: n = 213 per group (426 total).
Interpretation: Each classroom group needs 213 students for 90% power to detect the 15 pp improvement.

Social Science

A government agency compares voter turnout between two outreach strategies: SMS reminders vs. no contact.

Inputs: p₁ = 0.55, p₂ = 0.48, α = 0.05 (two-sided), power = 80%, continuity correction applied (default).
Result: n = 825 per group (1,650 total).
Interpretation: The small expected difference (7 pp) requires a large sample. Each group needs 825 participants for adequate power.

Engineering

A factory tests whether a new assembly process reduces the defect rate from 8% to 4%.

Inputs: p₁ = 0.04, p₂ = 0.08, α = 0.05 (two-sided), power = 80%, continuity correction applied (default).
Result: n = 599 per group (1,198 total).
Interpretation: Each production batch needs 599 inspected units to detect the halving of the defect percentage with 80% power.

References

Fleiss, J. L., Levin, B., & Paik, M. C. (2003). Statistical Methods for Rates and Proportions (3rd ed.). John Wiley & Sons., Chapter 8: sample size for comparing two proportions.
Chow, S.-C., Shao, J., Wang, H., & Lokhnygina, Y. (2018). Sample Size Calculations in Clinical Research (3rd ed.). Chapman & Hall/CRC., Chapters 3–4.
Julious, S. A. (2010). Sample Sizes for Clinical Trials. Chapman & Hall/CRC., Non-inferiority and equivalence designs for proportions.
Casagrande, J. T., Pike, M. C., & Smith, P. G. (1978). An improved approximate formula for calculating sample sizes for comparing two binomial distributions. Biometrics, 34(3), 483–486.
Kish, L. (1965). Survey Sampling. John Wiley & Sons.

Back to Sample Size Calculator for Comparing Two Independent Proportions