Sample Size Calculator

Understanding Sample Size Determination

Sample size determination is one of the most critical steps in designing any survey or experiment. Choose too small a sample, and your results will be unreliable, potentially leading to incorrect conclusions. Choose too large a sample, and you waste time, money, and resources collecting data beyond what is necessary for meaningful results. The goal is to find the minimum sample size that provides the statistical precision you need, and that is exactly what this calculator helps you determine.

I built this sample size calculator because the formulas, while not inherently difficult, involve several interrelated parameters that can be confusing without context. By showing the step-by-step calculations alongside sensitivity tables, this tool helps you understand not just the final number but how each parameter influences it. Whether you are planning a market research survey, a customer satisfaction study, an academic research project, or a quality control inspection, the underlying statistics are the same.

The Cochran Formula for Sample Size

The most widely used formula for determining sample size for proportions in survey research is Cochran's formula, developed by William Cochran in his 1977 text "Sampling Techniques." The formula calculates the minimum sample size needed to estimate a population proportion with a specified level of confidence and precision.

In this formula, n0 is the initial sample size estimate (before finite population correction), Z is the z-score corresponding to your desired confidence level, p is the expected proportion (expressed as a decimal), and e is the margin of error (also expressed as a decimal).

The z-scores for common confidence levels are 1.645 for 90% confidence, 1.96 for 95% confidence, 2.326 for 98% confidence, and 2.576 for 99% confidence. These values come from the standard normal distribution and represent the number of standard deviations from the mean that encompass the specified percentage of the distribution.

Example · Standard Survey at 95% Confidence, 5% Margin

For a typical survey with 95% confidence, 5% margin of error, and 50% expected proportion (the most conservative assumption), the calculation proceeds as follows. n0 = (1.96)^2 x 0.50 x 0.50 / (0.05)^2 = 3.8416 x 0.25 / 0.0025 = 0.9604 / 0.0025 = 384.16, which rounds up to 385. This is why you often see 385 cited as the minimum sample size for a standard survey. It does not depend on the population size, as long as the population is sufficiently large (typically above 20,000).

Finite Population Correction

When your population is finite and your sample represents a significant proportion of it (generally more than 5%), you can reduce the required sample size using the finite population correction factor.

Here, N is the total population size and n0 is the uncorrected sample size from Cochran's formula. This correction recognizes that as your sample becomes a larger fraction of the population, each additional observation provides more information about the remaining unsampled individuals.

Example · Surveying a Company of 500 Employees

Using the same parameters as above (95% confidence, 5% margin, 50% proportion), the uncorrected sample size is 385. Applying the finite population correction with N = 500, we get n = 385 / (1 + (385 - 1) / 500) = 385 / (1 + 0.768) = 385 / 1.768 = 217.8, which rounds up to 218. Instead of surveying 385 employees (which is more than 77% of the company), you only need 218, or about 44% of the workforce. The correction makes a substantial difference for small populations.

Understanding Confidence Level

The confidence level represents how certain you want to be that your sample results reflect the true population value, within your specified margin of error. A 95% confidence level is the most commonly used standard in survey research and social science. It means that if you repeated your survey 100 times using different random samples of the same size, approximately 95 of those surveys would produce results within your margin of error of the true population value.

Higher confidence levels require larger sample sizes because you are asking for greater certainty. Moving from 95% to 99% confidence increases the required sample size by approximately 75%. Moving from 95% to 90% confidence decreases it by about 30%. The choice of confidence level depends on the consequences of being wrong. Medical research and regulatory decisions typically use 95% or 99% confidence, while preliminary market research might accept 90%.

Understanding Margin of Error

The margin of error (also called the confidence interval half-width) defines the precision of your estimate. If your survey finds that 60% of respondents prefer Product A with a margin of error of plus or minus 5%, the true population preference is estimated to be between 55% and 65%.

Smaller margins of error require dramatically larger sample sizes because the relationship is quadratic. Halving the margin of error from 5% to 2.5% quadruples the required sample size (from 385 to 1,537 at 95% confidence). Reducing the margin from 5% to 1% increases the sample by a factor of 25 (from 385 to 9,604).

Choosing an appropriate margin of error requires balancing precision against practical constraints. For consumer surveys and market research, 5% is the widely accepted standard. Political polling typically requires 2 to 3% margins for competitive races. Quality control and clinical research may demand margins of 1% or less, depending on the stakes involved.

The Role of Expected Proportion

The expected proportion (p) represents your best guess at the survey result before collecting data. It affects sample size because the variance of a binomial proportion is maximized at p = 0.5 (50%). When the expected result is near 50/50, you need the largest sample. When the expected result is skewed toward 0% or 100%, you need a smaller sample because there is less variability to account for.

Using p = 0.50 is the most conservative (largest sample size) choice and is recommended when you have no prior information. If previous research or pilot studies suggest the proportion is far from 50%, using that estimate produces a smaller and more fast sample size. For example, if prior data suggests 90% of customers are satisfied, the required sample at 95% confidence and 5% margin is only 139 instead of 385.

Adjusting for Response Rate

In real-world surveys, not everyone you contact will respond. The response rate is the percentage of contacted individuals who actually complete the survey. If you need 385 completed responses and expect a 60% response rate, you need to send surveys to 385 / 0.60 = 642 people.

Response rates vary dramatically by survey method. Face-to-face interviews typically achieve 60 to 80% response rates. Telephone surveys average 10 to 30%, down significantly from historical rates due to caller screening and cell phone prevalence. Online surveys typically see 10 to 30% for unsolicited panels and 40 to 60% for targeted employee or customer surveys. Mail surveys average 20 to 40%.

Low response rates can introduce non-response bias, where the people who do not respond differ systematically from those who do. If dissatisfied customers are less likely to respond to a satisfaction survey, your results will overstate true satisfaction. This is a separate issue from sample size and requires careful survey design, follow-up protocols, and potentially non-response weighting to address.

Design Effect and Complex Sampling

Simple random sampling assumes every individual has an equal probability of selection, and each selection is independent. In practice, many surveys use more complex sampling designs such as cluster sampling (selecting groups first, then individuals within groups) or stratified sampling (dividing the population into subgroups and sampling within each).

Cluster sampling reduces data collection costs by concentrating interviews in selected locations, but it also reduces statistical efficiency because individuals within the same cluster tend to be more similar to each other than to the general population. The design effect (DEFF) quantifies this loss of efficiency. A design effect of 2.0 means you need twice as many observations with cluster sampling as you would with simple random sampling to achieve the same precision.

Typical design effects range from 1.0 (simple random sampling, no effect) to 3.0 or higher for highly clustered designs. Many large-scale household surveys such as the Demographic and Health Surveys (DHS) report design effects of 1.5 to 2.5 for most indicators. When planning a cluster-sampled survey, always account for the design effect in your sample size calculation.

Sample Size for Different Study Types

While this calculator focuses on the most common scenario (estimating a proportion), different study types have different sample size requirements.

Comparing Two Proportions

When your goal is to detect a difference between two groups (such as treatment vs. control in an experiment), the sample size depends on the expected proportions in each group, the desired power (typically 80% or 90%), and the significance level (typically 5%). Detecting smaller differences requires larger samples. For example, detecting a 5-percentage-point difference between groups (55% vs. 50%) requires approximately 1,570 per group at 80% power, while detecting a 20-point difference (70% vs. 50%) requires only about 94 per group.

Estimating a Mean

When estimating a continuous variable (like average income or mean test score), the formula uses the population standard deviation instead of a proportion.

Here, sigma is the population standard deviation and e is the desired margin of error in the same units as the variable being measured. Estimating the standard deviation typically requires pilot data or published estimates from similar studies.

Clinical Trials and Power Analysis

Clinical trials and experiments require power analysis, which determines the sample size needed to detect a specified effect size with a desired probability (power). Power analysis involves four parameters that are mathematically related, and specifying any three determines the fourth. The significance level (alpha, typically 0.05), the power (1 - beta, typically 0.80 or 0.90), the effect size (the minimum difference you want to detect), and the sample size. Specialized software like G*Power or statistical packages in R and Python are typically used for power calculations.

Common Sample Size Mistakes

Several common errors lead to inadequate or wasteful sample sizes in practice.

Using the margin of error without specifying the confidence level is one of the most frequent mistakes. A 5% margin of error at 90% confidence requires 271 responses, while the same margin at 99% confidence requires 664. The margin of error number alone is meaningless without the associated confidence level.

Ignoring non-response is another critical error. Calculating that you need 385 completed surveys and then sending out exactly 385 invitations will virtually guarantee an inadequate sample. Always inflate your target by dividing by the expected response rate.

Assuming that larger populations always need larger samples is a misconception. A survey of a city of 100,000 and a survey of a country of 330,000,000 require nearly the same sample size for the same precision. The finite population correction only matters when the population is small relative to the sample.

Confusing statistical significance with practical significance is a subtler error. With a large enough sample, even trivially small differences become statistically significant. A survey of 50,000 people might detect a statistically significant 0.5-percentage-point difference between two products, but this difference may have no practical business value. Sample size should be determined by the smallest practically meaningful effect you want to detect, not by the desire to find "significant" results.

Sample Size Tables for Quick Reference

The following table provides ready-to-use sample sizes for common scenarios, assuming 50% expected proportion (most conservative).

Applications Across Industries

Market Research

Market researchers use sample size calculations to plan consumer surveys, brand tracking studies, product concept tests, and advertising effectiveness research. A typical brand tracking study might survey 500 to 1,000 respondents per wave (monthly or quarterly) to track brand awareness and preference with a margin of error of 3 to 5 percentage points. Concept testing for new products usually targets 200 to 300 respondents per concept, balancing statistical precision against the cost of exposing proprietary information to a large number of people.

Quality Control and Manufacturing

In manufacturing quality control, sample size determines how many units from a production batch must be inspected to assess quality. Acceptance sampling plans (governed by standards like ANSI/ASQ Z1.4) specify sample sizes based on the lot size, the acceptable quality level (AQL), and the inspection level. For a lot of 10,000 units at an AQL of 1.0% under normal inspection, the standard calls for a sample of 200 units. If 5 or more defective units are found, the lot is rejected.

Political Polling

Political polls typically use sample sizes of 800 to 1,500 likely voters, producing margins of error of 2.5 to 3.5 percentage points at 95% confidence. Major national polls in the United States, such as those conducted by Gallup, Pew Research, and the major news networks, typically survey 1,000 to 1,500 adults. State-level polls often use smaller samples of 400 to 800, resulting in larger margins of error of 3.5 to 5 points.

Academic Research

Academic researchers in social sciences, education, and health sciences routinely perform sample size calculations as part of research proposals and ethics board applications. Institutional Review Boards (IRBs) require researchers to justify their sample sizes to ensure studies are not enrolling more participants than necessary (which wastes participants' time) or fewer than necessary (which wastes everyone's time by producing inconclusive results).

Healthcare and Clinical Studies

Clinical studies have particularly rigorous sample size requirements because the consequences of incorrect conclusions can directly affect patient health. Phase III clinical trials testing new drugs typically enroll hundreds to thousands of patients per treatment arm. The FDA requires adequate statistical power to detect clinically meaningful treatment effects, and underpowered studies are a common reason for regulatory rejection.

Online Survey Considerations

The growth of online surveys has changed the practical field of sample size planning. Online panels can provide large samples quickly and inexpensively, but they introduce coverage bias (not everyone is online) and self-selection bias (panel members volunteer to participate). These biases cannot be fixed by increasing sample size. A biased survey of 10,000 people is no more precise than a biased survey of 1,000.

To mitigate these issues, reputable online survey platforms use quota sampling to match the demographic profile of the target population, and some employ statistical weighting to adjust for known biases. When designing an online survey, focus on sample quality (representativeness) as much as sample quantity (size).

Practical Tips for Survey Planning

From my experience working with survey data across various contexts, here are several practical recommendations.

Always calculate sample size before collecting data, not after. Retroactive power calculations ("we had 200 responses, so our margin of error was X%") are descriptive but cannot justify the adequacy of a study that was not designed with a target precision in mind.

Plan for subgroup analysis. If you need to analyze results separately for different segments (by region, age group, gender, or customer type), each subgroup needs its own minimum sample size. A survey of 400 people that needs to compare 4 age groups is really 4 surveys of 100 people each, which is likely insufficient for any one subgroup.

Budget for oversampling. Between non-response, incomplete surveys, and data quality screening (removing responses that are too fast, show straight-lining patterns, or fail attention checks), you may lose 20 to 40% of your initial sample. Plan accordingly by inflating your target.

Consider the cost per response. If each completed interview costs $50 (not uncommon for telephone surveys with incentives), the difference between a sample of 385 and 664 is $13,950. Explicitly evaluating whether the additional precision is worth the additional cost makes for better decision-making.

Document your assumptions. Record the confidence level, margin of error, expected proportion, and population size used in your calculation, along with the rationale for each choice. This documentation is valuable for peer review, client presentations, and future replication of similar studies.

Stratified Sampling and Subgroup Analysis

When your survey needs to produce dependable estimates for specific subgroups of the population, each subgroup essentially requires its own minimum sample size. If you need to compare satisfaction levels across 4 geographic regions with a margin of error of 5% within each region, you need approximately 385 respondents per region, for a total minimum sample of 1,540. Simply surveying 385 people total and hoping for roughly equal representation across regions is insufficient because random sampling might give you 200 from one region and 50 from another.

Stratified sampling addresses this by dividing the population into homogeneous subgroups (strata) and sampling independently within each stratum. There are two main approaches. Proportional allocation assigns sample sizes to each stratum in proportion to the stratum's share of the total population. If Region A contains 40% of the population, it receives 40% of the sample. This approach is best when the goal is to estimate the overall population parameter with maximum precision.

Disproportional (or best) allocation, also called Neyman allocation, assigns larger samples to strata with greater variability. If satisfaction levels vary widely in Region B but are very consistent in Region C, more respondents should come from Region B. The formula for best allocation is n_h = n x (N_h x S_h) / sum(N_h x S_h), where N_h is the stratum size and S_h is the stratum standard deviation. This approach minimizes the overall variance of the estimate for a given total sample size.

Statistical Power and Effect Size

While this calculator focuses on sample size for estimation (determining a proportion with a given precision), many research questions require sample size for hypothesis testing (determining whether two groups differ). The concept of statistical power is central to these calculations.

Power is the probability of correctly detecting a real effect when one exists. A study with 80% power has a 20% chance of missing a real effect (a Type II error, or false negative). The conventional minimum power in biomedical and social science research is 80%, though 90% is increasingly recommended for clinical trials and confirmatory studies.

Effect size measures the magnitude of the difference you want to detect. For comparing two proportions, the effect size is simply the difference between them (p1 - p2). For comparing two means, Cohen's d = (mean1 - mean2) / pooled standard deviation. Cohen's conventions classify effect sizes as small (d = 0.2), medium (d = 0.5), and large (d = 0.8), though these labels are somewhat arbitrary and domain-specific.

The relationship between sample size and detectable effect size is quadratic. Halving the detectable effect size requires quadrupling the sample. This is why studies designed to detect small effects require large samples, and why pilot studies (which typically use small samples) are only appropriate for detecting large effects or estimating parameters for power calculations.

Bayesian Approaches to Sample Size

The Cochran formula used in this calculator is based on frequentist statistics, which is the dominant framework in applied survey research. However, Bayesian methods offer an alternative approach to sample size determination that incorporates prior information more formally.

In a Bayesian framework, you specify a prior distribution for the parameter of interest (reflecting your existing knowledge or beliefs), and the sample size is chosen to ensure that the posterior distribution (updated with the survey data) meets a specified precision criterion. For example, you might determine the sample size needed so that the 95% credible interval for the population proportion is no wider than 10 percentage points.

Bayesian sample size determination tends to produce smaller required samples when strong prior information is available, because the prior effectively contributes additional information beyond what the sample provides. When prior information is weak (a non-informative prior), Bayesian and frequentist sample sizes converge to similar values. In practice, Bayesian sample size calculations are more common in clinical trials and pharmaceutical research, where prior studies often provide substantial information about expected effect sizes.

Sequential and Adaptive Sampling

Traditional sample size calculations determine a fixed sample size before data collection begins. Sequential and adaptive designs allow the sample size to be adjusted during data collection based on the results observed so far.

In sequential sampling, data is analyzed at predetermined interim points, and the study may be stopped early if the results are already conclusive (either positively or negatively). Group sequential designs are standard in clinical trials, where it would be unethical to continue enrolling patients in a study after enough evidence has accumulated to determine treatment superiority or futility.

Adaptive designs go further by allowing modifications to the study design (including sample size) based on interim analyses. A sample size re-estimation might occur after 50% of the data is collected, using the observed variance to refine the final sample size target. This approach is particularly valuable when the expected variance is uncertain at the design stage.

These modern designs require specialized statistical methods to maintain valid Type I error rates (protecting against false positives) despite the multiple analyses. The Pocock and O'Brien-Fleming boundaries are commonly used spending functions that control the overall significance level across sequential analyses.

Finite Population Correction in Practice

The finite population correction (FPC) is one of the most misunderstood aspects of sample size calculation. Many people are surprised to learn that surveying a population of 10,000 requires almost the same sample as surveying a population of 10 million. The underlying reason is that the precision of a sample estimate depends primarily on the absolute number of observations, not on the sampling fraction.

The FPC only produces a meaningful reduction in sample size when the sampling fraction (n/N) exceeds about 5%. For a population of 1,000, the uncorrected sample of 385 represents 38.5% of the population, and the FPC reduces it to 278, a significant reduction. For a population of 100,000, the uncorrected sample of 385 represents only 0.385%, and the FPC barely changes the number (to 383).

This has a practical implication that many researchers find counterintuitive. A nationwide survey and a survey of a small company both require about the same number of respondents for the same level of precision. The difference is that the company survey will achieve a higher response rate more easily, and it may benefit from the FPC, but the base sample size is driven by the desired precision, not the population size.

Non-Probability Sampling Methods

The sample size formulas in this calculator assume probability sampling, where every member of the population has a known, non-zero chance of being selected. In practice, many surveys and studies use non-probability sampling methods where this assumption does not hold.

Convenience sampling (surveying whoever is easily accessible), voluntary response sampling (only counting those who choose to participate), snowball sampling (using existing respondents to recruit new ones), and quota sampling (targeting specific demographic profiles) are all non-probability methods. For these methods, traditional margin-of-error calculations do not strictly apply because the sampling error cannot be quantified in the same way.

Still, the sample size numbers from this calculator serve as useful guidelines even for non-probability samples. A convenience sample of 400 is generally more informative than a convenience sample of 50, even though neither has a formally calculable margin of error. Researchers using non-probability methods should report their results without formal margins of error and clearly describe the sampling method and its limitations.

Multi-Item Surveys and Multiple Comparisons

Most real surveys contain multiple questions, and the sample size needed may differ for each question depending on the expected proportion. A customer survey might ask about overall satisfaction (expected proportion near 85%), willingness to recommend (expected near 70%), and likelihood to purchase again (expected near 60%). The most conservative approach is to calculate sample size based on the question with the highest variance (closest to 50%), which ensures adequate precision for all questions.

When analyzing multiple questions from the same survey, the issue of multiple comparisons arises. If you test 20 hypotheses at the 5% significance level, you expect about one to be "statistically significant" by chance alone. The Bonferroni correction addresses this by dividing the significance level by the number of comparisons (0.05 / 20 = 0.0025 per test), but this requires larger sample sizes. For surveys with many items, the false discovery rate (FDR) approach is a less conservative alternative that controls the expected proportion of false positives among all rejected hypotheses.

Pilot Studies and Pre-Testing

Before launching a full-scale survey, conducting a pilot study with a small sample (typically 20 to 50 respondents) serves several purposes. It identifies confusing or ambiguous questions that need revision. It provides preliminary estimates of response distributions, which can refine your sample size calculation. It tests the survey logistics (delivery method, response time, technical issues). And it estimates the response rate, which is needed to determine how many invitations to send.

A pilot study should never be used alone to draw final conclusions because it is too small for dependable statistical inference. However, the information it provides dramatically improves the quality of the full study. If the pilot suggests that the expected proportion is 30% rather than the assumed 50%, the required sample size drops from 385 to 323 (at 95% confidence, 5% margin), potentially saving significant resources.

Sample Size for A/B Testing

A/B testing (split testing) in digital marketing and product development requires sample size calculations based on hypothesis testing rather than simple estimation. The typical A/B test compares a control version against one or more variants to determine whether the variant produces a statistically significant improvement.

The sample size depends on four factors. The baseline conversion rate (the current performance of the control), the minimum detectable effect (the smallest improvement you consider practically meaningful), the significance level (typically 5%, representing the false positive rate), and the power (typically 80% or 90%, representing the probability of detecting a real effect).

For a baseline conversion rate of 5% and a minimum detectable effect of 1 percentage point (detecting an increase from 5% to 6%), the required sample size per variant is approximately 14,750 at 80% power, for a total of 29,500 observations across both variants. If the minimum detectable effect is larger (say 2 percentage points, from 5% to 7%), the required sample drops to approximately 3,650 per variant. This illustrates why A/B tests for small effect sizes require substantial traffic and why it is important to define the minimum detectable effect before starting the test.

For web-based A/B tests, the "per variant" sample size translates to the number of visitors who must be exposed to each variant. If your website receives 1,000 visitors per day and you need 14,750 per variant (29,500 total), the test needs to run for approximately 30 days. Running the test for less time increases the risk of an inconclusive result or a false positive.

Weighted Sampling and Post-Stratification

When a sample does not perfectly represent the target population, statistical weighting can adjust the results. Post-stratification weights are assigned to each respondent based on how over- or under-represented their demographic group is in the sample compared to the known population distribution.

For example, if your survey sample is 60% female and the target population is 51% female, each female respondent would receive a weight of 0.51/0.60 = 0.85, and each male respondent would receive a weight of 0.49/0.40 = 1.225. The weighted results better reflect the population distribution. However, weighting increases the effective variance of the estimates, meaning the effective sample size is smaller than the actual sample size. The design effect due to unequal weighting can be approximated as 1 + CV^2, where CV is the coefficient of variation of the weights. Highly variable weights (indicating a severely unrepresentative sample) can substantially increase the effective margin of error.

For this reason, sample size calculations should anticipate the need for weighting by building in a buffer. If you expect a design effect of 1.3 due to weighting, multiply your calculated sample size by 1.3. This ensures that even after weighting, the effective sample size provides the desired precision.

Reporting Sample Size in Research Publications

Academic journals and professional organizations have specific requirements for how sample size decisions should be reported. The CONSORT statement (for clinical trials), STROBE statement (for observational studies), and CHERRIES checklist (for online surveys) all require transparent reporting of sample size calculations.

A well-written sample size justification includes the formula used, the values for each parameter (confidence level, margin of error, expected proportion, population size), the rationale for choosing those values (citing pilot data or prior studies where applicable), any adjustments for design effect, clustering, or non-response, and the final target sample size. This transparency allows reviewers and readers to evaluate whether the study was adequately powered and to replicate the calculation if needed.

A common deficiency in research reporting is stating the sample size without explaining how it was determined. Statements like "we surveyed 200 participants" without justification leave reviewers unable to assess whether 200 is adequate. Even worse is the practice of performing the sample size calculation retroactively to justify an already-collected sample. Best practice is to determine and document the required sample size during the study design phase, before any data is collected.