Sample Size for Proportion Calculator

Hamza Gujjar Reviewed by Hamza Gujjar, Algorithm Architect & Software Developer Last updated: August 17, 2026
📚 Confused about how this is calculated? Read the full The Margin of Error, and the Errors It Doesn't Cover →

Related Statistics Calculators

Sample Size for Mean Calculator Power Calculator Probability of Success Calculator Bayes Theorem Calculator Mann-Whitney U Test Calculator Wilcoxon Signed-Rank Calculator Kruskal-Wallis Calculator McNemar Calculator Cronbach's Alpha Calculator Z-Test Calculator Paired T-Test Calculator Welch's T-Test Calculator

View all 54 Statistics calculators →

Sample Size Determination for Binomial Proportions: Cochran's Formula

In polling, market research, epidemiological prevalence surveys, and digital A/B testing, estimating an unknown population proportion (such as voter support, conversion rate, or disease prevalence) with a specified level of precision requires Cochran's Sample Size Formula.

The required sample size depends on three fundamental parameters: the Confidence Level (Z), the Margin of Error (E), and the Estimated Baseline Proportion (p).

Mathematical Formulation of Sample Size for Proportions

1. Cochran's Formula for Infinite / Large Populations:
n0 = [ Zα/22 × p(1 − p) ] / E2

Where:
• Zα/2: Standard normal critical value corresponding to desired confidence level (1.960 for 95% confidence, 2.576 for 99% confidence, 1.645 for 90% confidence).
• p: Estimated true population proportion (0.0 to 1.0). When completely unknown, set p = 0.50 (Maximum Variance Principle) to guarantee a conservative, safe sample size.
• E: Maximum acceptable Margin of Error (half-width of confidence interval, e.g., ±0.03 for ±3%).

2. Finite Population Correction (FPC):
When sampling without replacement from a finite population of size N where the sample exceeds 5% of the population (n0 / N > 0.05):
n = n0 / [ 1 + (n0 − 1) / N ]

Required Sample Size Matrix: Infinite vs Finite Populations (95% Confidence)

Population Size (N) ±1% Margin of Error ±2% Margin of Error ±3% Margin of Error ±5% Margin of Error
100 99 96 92 80
500 475 414 341 217
1,000 906 706 516 278
10,000 4,899 1,936 964 370
100,000 8,762 2,345 1,056 383
Infinite (∞) 9,604 2,401 1,067 384

Step-by-Step Practical Calculation: National Presidential Polling

A polling organization wants to estimate national candidate approval with a 95% confidence level (Z = 1.96) and a ±3.0% margin of error (E = 0.03). Baseline support is unknown, so maximum variance p = 0.50 is used:

  • Step 1: Calculate Base Sample Size (n0):
    n0 = [ (1.96)2 × 0.50 × (1 − 0.50) ] / (0.03)2
    n0 = [ 3.8416 × 0.25 ] / 0.0009 = 0.9604 / 0.0009 = 1,067.11 → 1,068 Respondents.
  • Step 2: Account for Finite Registered Voter Population (N = 50,000 in a specific district):
    n = 1,067.11 / [ 1 + (1,067.11 − 1) / 50,000 ] = 1,067.11 / [ 1 + 0.02132 ] = 1,067.11 / 1.02132 = 1,045 Respondents.
  • Step 3: Adjust for Expected 20% Survey Response Rate:
    Required Invitations = 1,045 / 0.20 = 5,225 Poll Invitations Dispatched.

Frequently Asked Questions About Sample Size for Proportions

Why does p = 0.50 yield the maximum required sample size?

The mathematical variance of a Bernoulli trial is p(1 − p). This quadratic parabola peaks at p = 0.50 (0.50 × 0.50 = 0.25). If true proportion is skewed (e.g., rare defect p = 0.05), p(1 − p) = 0.0475, requiring far fewer observations for the same absolute margin of error.

Why does population size matter very little once N > 100,000?

As population size N approaches infinity, the finite correction term (n − 1)/N approaches zero, causing the required sample size to plateau at the infinite threshold (e.g., n = 1,068 for ±3% at 95% confidence whether the population is 1 million or 1 billion).

What is the Design Effect (DEFF) in cluster sampling?

When sampling clusters (e.g., classrooms, hospital clinics) rather than independent individuals, observations within clusters correlate (intra-cluster correlation ICC). The required sample size must be multiplied by DEFF = 1 + (m − 1)ICC, where m is average cluster size.

How does halving the margin of error affect sample size?

Because margin of error E is squared in the denominator (E2), cutting the margin of error in half (e.g., from ±4% to ±2%) requires 4 times the sample size. Reducing it from ±3% to ±1% requires 9 times the sample size.

What is the difference between Wald confidence intervals and Clopper-Pearson intervals?

Cochran's formula is based on the normal approximation (Wald interval). When sample sizes are small or proportions are near 0 or 1 (np < 5), the normal approximation fails, requiring exact binomial Clopper-Pearson or Wilson Score intervals.

A/B Testing Sample Sizing: Minimum Detectable Lift (MDL) for Conversion Rates

In digital product experimentation and online conversion rate optimization (CRO), teams compare two proportions (Control conversion rate p1 vs Variant conversion rate p2):

Two-Sample Proportion Sample Size Formula (Equal Allocation):
nper variant = [ Zα/2√(2p̄(1 − p̄)) + Zβ√(p1(1 − p1) + p2(1 − p2)) ]2 / (p1 − p2)2
where p̄ = (p1 + p2) / 2 is the pooled baseline conversion rate.

Relative Minimum Detectable Lift (δrel): If baseline conversion is p1 = 3.0% and the business requires detecting a 10% relative lift (δrel = 0.10 → p2 = 3.30%, absolute difference Δ = 0.003):
nper variant ≈ 173,000 Users per Variant (346,000 Total Visitors) for 95% Confidence and 80% Power.

A/B Test Sample Size Matrix per Variant (80% Power, α = 0.05 Two-Tailed)

Baseline Conversion Rate (p1) +5% Relative Lift +10% Relative Lift +20% Relative Lift +30% Relative Lift
1.0% (0.010) 2,480,000 per variant 615,000 per variant 151,000 per variant 66,000 per variant
3.0% (0.030) 800,000 per variant 198,000 per variant 48,500 per variant 21,200 per variant
5.0% (0.050) 465,000 per variant 115,000 per variant 28,200 per variant 12,300 per variant
10.0% (0.100) 215,000 per variant 53,000 per variant 13,000 per variant 5,600 per variant
25.0% (0.250) 65,000 per variant 16,000 per variant 3,900 per variant 1,700 per variant

Stratified Sampling: Neyman Optimal Allocation

When sampling heterogeneous sub-populations (e.g., polling across distinct geographical regions or customer spend tiers), Neyman Optimal Allocation minimizes estimation variance for a fixed sample budget:

Neyman Stratum Sample Size Formula:
nh = n × [ Nh × √(ph(1 − ph)) ] / [ ∑ (Ni × √(pi(1 − pi))) ]
Stratum sample size scales directly with stratum population weight (Nh) and internal stratum binomial variance.

Sample Sizing for Unequal Variant Allocations (k = n2 / n1)

In live production digital A/B tests, routing 50% of traffic to a radical, unproven design can pose unacceptable business revenue risk. Teams frequently assign 80% of traffic to Control and 20% to Variant (allocation ratio k = n2 / n1 = 0.25 / 0.75 = 0.333):

Unequal Allocation Sample Size Formula:
n1 (Control) = [ (Zα/2 + Zβ)2 × (p1(1 − p1) + p2(1 − p2)/k) ] / (p1 − p2)2
n2 (Variant) = k × n1

The Statistical Penalty: Unequal allocation increases the total combined sample size required for identical power. An 80/20 split requires 56% more total visitors than a balanced 50/50 split.

Sequential Testing and Multi-Armed Bandit (MAB) Sample Sizing

Modern tech companies deploy Multi-Armed Bandits (Thompson Sampling) and continuous sequential testing algorithms (mSPRT) to dynamically allocate traffic to winning variations while minimizing user exposure to inferior designs:

Experimentation Framework Sample Size Determination Traffic Allocation Model Primary Business Objective
Fixed-Horizon A/B Testing Calculated strictly a-priori (e.g., 50,000 users per arm). Fixed 50/50 split throughout test duration. Rigorous causal scientific hypothesis testing.
Sequential Testing (mSPRT) Variable; continuous monitoring with always-valid p-values. Fixed split; stops early when boundary is crossed. Accelerating decision velocity without false positives.
Multi-Armed Bandits (MAB) No fixed sample size; runs continuously. Dynamic Bayesian routing to highest-performing arm. Maximizing immediate cumulative revenue / clicks.

The 10-Point Survey and Polling Sample Size Protocol

  1. Establish the Margin of Error Tolerances: Determine whether the research goal mandates ±1%, ±3%, or ±5% precision.
  2. Default to p = 0.50 When Baseline is Unknown: Assume maximum binomial variance to avoid underpowered survey designs.
  3. Apply Finite Population Correction for Small Groups: Use FPC whenever sample size exceeds 5% of total organization/membership size.
  4. Calculate Stratum Sizes with Neyman Allocation: Over-sample high-variance subgroups to optimize overall survey accuracy.
  5. Factor in Multi-Stage Cluster Design Effects: Multiply required sample size by DEFF when sampling geographical clusters or institutions.
  6. Forecast Realistic Response Rates: If phone response rates are 5%, dispatch 20× invitations to achieve target completed interviews.
  7. Check Subgroup Analysis Sample Depth: Ensure individual demographic sub-cells (e.g., age 18–29) contain at least n ≥ 100 respondents.
  8. Avoid Early Stopping in Fixed-Horizon Tests: Never declare a winner early in standard A/B tests; data peeking severely inflates Type I errors.
  9. Weight Sample Post-Stratification: Apply rake weighting to adjust completed sample demographics to census benchmarks.
  10. Document Non-Sampling Error Reserves: Recognize that question wording and sampling frame coverage errors often exceed statistical sampling error.

Detailed Proportion Sample Size FAQs

How does cluster sampling inflate required sample sizes?

In cluster sampling (e.g., surveying all employees across 10 company branches), individuals within the same branch think alike (positive ICC). The Design Effect (DEFF = 1 + (m − 1)ICC) typically requires doubling or tripling total sample size to match simple random sampling.

Why do national polls sample only 1,000 people for a country of 330 million?

By Cochran's formula, 1,067 randomly sampled individuals achieve a ±3% margin of error at 95% confidence regardless of whether the population is 1 million or 1 billion, because the finite population correction term is negligible in large populations.

What is the minimum sample size for testing rare proportions (e.g., 0.1% defect rate)?

For rare events, normal approximations fail. The Rule of Three states that if zero events are observed in n trials, the 95% upper confidence bound for the true rate is 3/n. To detect a 0.1% defect rate with high precision requires tens of thousands of samples.

How does non-response bias affect sample size calculations?

Non-response bias cannot be fixed by merely increasing sample size. If non-respondents differ systematically from respondents, collecting 10,000 biased responses simply produces a more precisely incorrect estimate.

Can a sample size calculation be based on a one-sided margin of error?

Yes. In regulatory compliance where only the upper limit matters (e.g., confirming contamination proportion is < 1%), one-sided critical values (Z = 1.645 for 95% confidence) reduce required sample size by ~30%.

What is the difference between statistical precision and accuracy in polling?

Precision (Margin of Error) reflects random sampling variability governed by sample size. Accuracy reflects freedom from systematic bias, non-response distortion, and leading survey questions.

Exact Binomial Sizing for Rare Events: Wilson Score and Clopper-Pearson Intervals

When monitoring ultra-low defect proportions (e.g., aerospace avionics failures, rare adverse vaccine events where p < 0.01), the normal approximation used in Cochran's formula breaks down completely because np < 5:

1. Wilson Score Interval Inversion:
p̃ = [ X + Z2/2 ] / [ n + Z2 ]
Center and bounds remain strictly within the valid [0, 1] probability range without negative lower limits.

2. The Rule of Three for Zero-Failure Sampling:
If an auditor tests n components and finds exactly 0 defects, the 95% one-sided upper confidence limit for the true defect proportion is:
pupper ≤ 3 / n
Example: To prove with 95% confidence that an automotive airbag failure rate is below 0.01% (0.0001), testing must observe 0 defects across at least n = 3 / 0.0001 = 30,000 Airbags Tested.

Case Study: High-Reliability Semiconductor Yield Audit

A foundry quality control director wants to confirm that a microprocessor fabrication line maintains a defect rate of no more than p = 2.0% (0.020) with a ±0.5% margin of error (E = 0.005) at 99% Confidence (Z = 2.576):

Step-by-Step Resolution:
1. Calculate Infinite Sample Size (n0):
n0 = [ (2.576)2 × 0.020 × (1 − 0.020) ] / (0.005)2
n0 = [ 6.6358 × 0.0196 ] / 0.000025 = 0.13006 / 0.000025 = 5,202.4 → 5,203 Microchips.

2. Finite Lot Correction (Production Batch N = 20,000 Wafers):
n = 5,202.4 / [ 1 + (5,202.4 − 1) / 20,000 ] = 5,202.4 / [ 1 + 0.26007 ] = 5,202.4 / 1.26007 = 4,129 Microchips Inspected.
Outcome: Finite population correction reduces required destructive testing by 1,074 chips (−20.6%), saving substantial foundry testing expenditure.

Additional Proportion Sample Size FAQs

How does cluster sampling affect polling precision?

Cluster sampling increases standard error by the square root of the Design Effect (√DEFF). A poll with DEFF = 2.0 has an effective sample size equal to half its nominal sample size (neff = n / 2).

What is the minimum sample size for a valid Chi-Square Goodness-of-Fit test?

Cochran's guideline requires that at least 80% of categories have expected frequencies of 5 or more (Ei ≥ 5), and no category has an expected frequency less than 1.

How does sample size change when comparing 3 or more conversion rates?

Comparing $ proportions requires a Chi-Square contingency test or Marascuilo procedure, sizing the experiment based on Cohen's w effect size (w = √∑((pobs − pexp)2 / pexp)) with degrees of freedom df = k − 1.

Why do political polls often report ±3.1% margin of error?

A sample size of exactly n = 1,000 respondents at 95% confidence (Z = 1.96) and p = 0.50 produces a margin of error of E = 1.96 × √(0.25 / 1,000) = 1.96 × 0.01581 = 0.0310 (±3.1%).

Historical Foundation: Survey Sampling Theory from Neyman to Cochran

The statistical science of determining optimal sample sizes for proportions evolved from foundational mathematical milestones in survey sampling:

  • Jerzy Neyman (1934): Published his groundbreaking paper on representative sampling, introducing stratified sampling and the mathematical proof of Optimal Allocation, which proved that stratified random sampling is strictly superior to purposive selection.
  • William G. Cochran (1953): Published the landmark textbook Sampling Techniques, formulating the definitive closed-form equations for margin of error, maximum variance principles ( = 0.50$), and the Finite Population Correction (FPC) that govern contemporary polling and auditing.
  • Leslie Kish (1965): Formalized complex multi-stage cluster sampling and introduced the Design Effect (DEFF), establishing the exact variance inflation multiplier required when clustering geographic areas or institutions.

Multi-Arm Multi-Stage (MAMS) Platform Trial Sample Sizing

In modern adaptive platform trials testing multiple experimental therapies simultaneously against a shared control group (e.g., oncology master protocols):

Shared Control Efficiency Formula:
Control Group Allocation n0 = narm × √K
where K is the number of active treatment arms.
Efficiency Gain: Evaluating 4 treatments against a shared control requires 37.5% fewer total patients than running 4 independent pairwise 2-arm trials.

Proportion Sample Size Troubleshooting and Diagnostics Matrix

Sampling Diagnostic Issue Underlying Cause Statistical Risk Remediation Protocol
Required Sample Exceeds Total Population (n > N) Demanding extreme margin of error (±0.5%) without applying FPC. Wasting budget; mathematically impossible sample target. Apply Finite Population Correction: n = n0 / [ 1 + (n0 − 1)/N ].
Underpowered A/B Test (Traffic Deficit) Overly optimistic Minimum Detectable Lift assumption (+25% relative lift). Test inconclusive after weeks; fails to detect genuine +5% revenue lift. Re-evaluate baseline conversion rate; extend testing window or combine variants.
Zero Events Observed in Safety Audit Rare defect proportion violates normal approximation (np < 5). Standard Wald confidence intervals yield negative lower bounds. Use Rule of Three (pupper ≤ 3/n) or exact Clopper-Pearson binomial intervals.
Severe Polling Non-Response Bias (Response Rate < 3%) Systematic demographic skew among individuals answering telephone polls. Collecting larger sample simply magnifies biased estimate precision. Deploy multi-mode sampling (SMS + Web) and apply iterative post-stratification rake weighting.

Glossary of Survey Sampling and Proportion Terminology

Margin of Error (E):
The radius (half-width) of a confidence interval expressing the maximum expected random sampling error at a specified confidence level.
Confidence Level (1 − α):
The long-run percentage of repeated independent samples in which the calculated confidence interval encompasses the true population proportion.
Maximum Variance Principle:
The practice of setting baseline p = 0.50 in Cochran's formula to guarantee the largest, most conservative required sample size.
Finite Population Correction (FPC):
A mathematical reduction factor applied to standard error when sample size exceeds 5% of a finite, known population.
Design Effect (DEFF):
The ratio of the variance of a complex cluster sample estimate to the variance of a simple random sample of identical size.
Intraclass Correlation Coefficient (ICC):
A measure of the degree of homogeneity or similarity among individuals grouped within the same sampling cluster.
Stratified Random Sampling:
A sampling method dividing the population into mutually exclusive subgroups (strata) and sampling independently from each stratum.
Rake Weighting (Iterative Proportional Fitting):
A mathematical algorithm adjusting survey respondent weights so sample marginal totals match known census demographic distributions.

Sample Size Allocation in Financial Auditing and Fraud Detection

In external financial statement audits and corporate compliance testing, certified public accountants (CPAs) use Monetary Unit Sampling (MUS) and attribute sampling to test for accounting control deviations:

AICPA Attribute Sampling Size Formula:
n = [ Reliability Factor (R) ] / [ Tolerable Deviation Rate (TDR) − Expected Population Deviation Rate (EPDR) ]
where Reliability Factor R is determined from Poisson risk tables (e.g., R = 3.00 for 95% confidence with 0 expected errors).
Auditing Application: If tolerable deviation rate is 5% and expected error is 0%, an auditor must test n = 3.00 / 0.05 = 60 Invoices to conclude with 95% confidence that internal billing controls are operating effectively.

Executive Summary: Guidelines for Survey and Proportion Sizing

When designing polling surveys, market research, and binomial experiments:

  • Set Conservative Baseline Proportions: Default to p = 0.50 unless reliable prior empirical data exists, ensuring maximum variance protection.
  • Apply Finite Population Correction: Adjust sample size whenever surveying small corporate departments, schools, or specialized professional organizations.
  • Account for Cluster Design Effects: Inflate sample targets by DEFF when sampling geographically clustered populations.
  • Budget for Non-Response Attrition: Calculate dispatch targets based on expected survey completion rates to ensure final completed quotas are met.

Step-by-Step Calculation Protocol: Sizing Binomial Surveys and Polls

Follow this 6-step computational protocol to determine exact sample size requirements for any binomial proportion study:

  1. Step 1 — Establish Desired Confidence Level: Select standard normal critical value Z (e.g., Z = 1.960 for 95% confidence, Z = 2.576 for 99% confidence, Z = 1.645 for 90% confidence).
  2. Step 2 — Set Acceptable Margin of Error (E): Define maximum tolerable half-width of the confidence interval in decimal form (e.g., E = 0.03 for ±3.0%).
  3. Step 3 — Estimate Expected Baseline Proportion (p): Use prior research or pilot data; if completely unknown, set p = 0.50 for maximum conservative variance protection.
  4. Step 4 — Calculate Base Cochran Sample Size (n0): Evaluate n0 = [ Z2 × p(1 − p) ] / E2.
  5. Step 5 — Apply Finite Population Correction if Applicable: If total population size N is known and n0 / N > 0.05, adjust via n = n0 / [ 1 + (n0 − 1)/N ].
  6. Step 6 — Inflate for Expected Non-Response: Divide completed sample size target by anticipated response rate: Ndispatch = n / Response Rate.