How Does Outcome Prevalence Affect Sample Size?

Why Can a Rare Outcome Require More Participants Even When Its Prevalence Is Low?

You are designing a university survey about students' use of generative AI. One research question concerns the proportion of students who use AI tools weekly. Previous institutional surveys suggest that this behavior may be relatively common.

Another question concerns the proportion of students who have experienced a serious academic integrity investigation involving generative AI. This outcome is expected to be uncommon.

Should both outcomes require the same sample size?

A colleague suggests that rare outcomes should require fewer participants because their proportions are closer to zero. Another argues that rare events require substantially larger samples because very few participants will experience them.

Both observations can be correct under different statistical objectives. Estimating a population proportion with a fixed absolute margin of error differs from comparing rare-event probabilities or developing a regression model for an uncommon outcome.

The methodological question is how the anticipated prevalence of an outcome influences the amount of information obtained from each observation and, consequently, the sample size needed for the intended analysis.

Prevalence Affects Sample Size Differently Depending on What You Want to Estimate or Test

How Prevalence Changes Statistical Information and Sample Size Requirements

What Is Outcome Prevalence?

Prevalence refers to the proportion of a defined population that has a particular characteristic or condition at a specified point or during a specified period.

For example, a study might estimate the proportion of university students who used generative AI for academic writing during the previous semester.

The definition must specify the population, outcome, and reference period.

Prevalence should not automatically be used as a synonym for incidence. Incidence concerns new events or cases occurring over a defined period among those at risk.

In sample size planning, researchers sometimes use prevalence or event probability more broadly to describe the anticipated proportion of observations with a binary outcome.

However, the distinction between prevalence, cumulative risk, and event rate remains important when interpreting results.

Why Does Prevalence Affect the Variance of a Binary Outcome?

A binary outcome takes two possible values, commonly coded as zero and one.

If the probability of an event is p, the variance of a Bernoulli outcome is:

The variance is largest at p = 0.50 and decreases as the probability approaches zero or one.

This mathematical relationship helps explain conventional sample size planning for estimating a single proportion.

Why Is 50% Prevalence Commonly Used When the True Proportion Is Unknown?

For estimating a population proportion under simple random sampling, a commonly used initial approximation is:

Because p(1 − p) is maximized at 0.50, using p = 0.50 gives the largest initial sample requirement under this formula when the other assumptions are fixed.

This is why 50% is often used as a conservative planning assumption when no credible prevalence estimate exists.

However, the approach is conservative specifically for this absolute-precision calculation. It does not establish an appropriate sample size for every rare-event analysis or comparison.

Does Lower Prevalence Always Mean a Smaller Sample?

No. The answer depends on the inferential objective.

When estimating a single proportion with the same absolute margin of error, a proportion near zero generally requires fewer observations under the conventional approximation than one near 0.50.

However, if researchers want a margin of error that is small relative to the prevalence itself, rare outcomes can require very large samples.

For example, estimating a prevalence of 1% within five percentage points would provide little useful precision because the target margin exceeds the prevalence.

Estimating the same prevalence within half a percentage point requires substantially more information.

Researchers should therefore distinguish absolute precision from relative precision.

What Is the Difference Between Absolute and Relative Precision?

Absolute precision describes uncertainty in percentage points or the original proportion scale.

Relative precision describes uncertainty in relation to the size of the parameter being estimated.

Suppose the anticipated prevalence is 2%.

An absolute margin of one percentage point corresponds to a relative margin of 50% of the anticipated prevalence.

If researchers require a relative margin of 20%, the corresponding absolute margin is 0.4 percentage points.

The latter target is substantially more demanding and will generally require a much larger sample.

Why Can Rare Outcomes Require Large Samples?

Rare outcomes produce relatively few observed events in a sample.

For example, if an event probability is 1%, a sample of 200 participants would contain approximately two events on average.

That expected count may be inadequate for stable comparisons or regression analyses.

Increasing the sample to 2,000 participants would produce approximately 20 events on average under the same event probability.

However, the actual number of events remains random and may differ from its expectation.

Researchers should distinguish the total number of participants from the amount of information provided by observed events.

How Do You Estimate the Expected Number of Events?

For a simple setting with n independent observations and common event probability p, the expected number of events is:

These are expectations, not guaranteed event counts.

The number of events required for a particular analysis depends on the statistical model, effect of interest, predictor structure, and estimation objectives.

How Does Prevalence Affect the Sample Size for Comparing Two Proportions?

Comparing two proportions requires assumptions about the event probability in each group and the effect researchers intend to detect.

For example, a study might compare the prevalence of weekly AI use between students receiving different forms of instructional guidance.

A comparison between probabilities of 40% and 50% has different planning requirements from a comparison between 1% and 2%.

Although the second comparison involves a larger relative difference, its absolute difference is only one percentage point.

The sample size requirement depends on the chosen effect measure, significance level, desired power, and allocation ratio.

Researchers should not assume that a rare outcome automatically makes a comparison easier or harder without specifying the relevant alternative.

Does the Choice of Risk Difference, Risk Ratio, or Odds Ratio Matter?

Yes. Different effect measures express different aspects of the relationship between outcome probabilities.

A risk difference describes the absolute difference between two probabilities.

A risk ratio compares their ratio, while an odds ratio compares the corresponding odds.

For example, increasing an event probability from 1% to 2% represents an absolute increase of one percentage point and a risk ratio of two.

The same risk ratio applied to a much more common outcome corresponds to a different absolute difference.

Researchers should therefore specify the effect measure and baseline probability when planning comparisons of binary outcomes.

Effect size should be defined in terms relevant to the scientific decision, not merely selected for computational convenience.

How Does Outcome Prevalence Affect Logistic Regression?

Logistic regression models the relationship between predictors and the probability of a binary outcome.

When the outcome is rare, a dataset may contain relatively few events even if the total sample is large.

Limited event information can produce unstable coefficient estimates, wide confidence intervals, and difficulties with model fitting.

Complete or quasi-complete separation may occur when predictors distinguish event and nonevent observations too strongly for ordinary maximum likelihood estimation.

However, the adequacy of a logistic regression sample cannot be determined solely by the event count.

The number of predictors, their distributions, anticipated associations, outcome prevalence, and modeling objectives also matter.

Is the Ten Events per Variable Rule Sufficient?

No. The commonly cited rule of ten events per variable is a historical heuristic, not a universal sample size requirement.

Its adequacy varies according to model complexity, effect sizes, predictor distributions, outcome prevalence, and the purpose of the analysis.

Prediction model development may require substantially different information from a narrowly specified explanatory regression.

Modern approaches, including those developed by Riley and colleagues, consider multiple criteria for determining minimum sample size in prediction model development.

Researchers should avoid treating an arbitrary events-per-variable threshold as proof of model adequacy.

How Does Prevalence Affect Predictive Modeling?

Prediction models require sufficient information about both events and nonevents to estimate relationships and evaluate predictive performance.

For example, a model predicting a rare academic integrity investigation may contain thousands of students but relatively few positive outcomes.

This can complicate parameter estimation, calibration assessment, and validation.

Model complexity and anticipated predictive performance influence the required sample.

Researchers should also distinguish discrimination from calibration. A model may separate higher-risk and lower-risk observations while providing poorly calibrated predicted probabilities.

Sample size planning should reflect the intended prediction model and evaluation strategy.

Can Oversampling Rare Outcomes Reduce the Required Recruitment Effort?

In some designs, deliberately selecting more event cases can improve the efficiency of studying rare outcomes.

For example, a case-control study may recruit students with a documented outcome and a comparison group without it.

However, this design does not directly reproduce the outcome prevalence in the source population.

Estimating absolute risks or prevalence may require additional information and appropriate analytical methods.

Similarly, oversampling rare groups in a probability survey may require sampling weights for population estimation.

Researchers should distinguish improving the availability of relevant cases from obtaining a representative population prevalence estimate.

Does Prevalence Affect Sample Size in Clustered Studies?

Yes, but the relationship is more complicated than in an independent Bernoulli model.

For example, students within the same classroom may share experiences that influence their probability of using generative AI.

Within-cluster correlation can reduce the effective amount of independent information.

The number of clusters, cluster sizes, event prevalence, and intraclass correlation may all affect precision and power.

Rare events may also be unevenly distributed across clusters.

Researchers should use design-specific calculations rather than apply an independent-observation formula without adjustment.

See how clustering affects sample size.

What Happens When Prevalence Is Very Close to Zero or One?

When prevalence approaches either boundary, conventional normal approximations for confidence intervals may perform poorly, particularly when the expected number of events or nonevents is small.

For example, observing no events among 100 participants does not establish that the population event probability is exactly zero.

Alternative interval procedures, including exact binomial or score-based methods, may be more appropriate.

Sample size planning should use methods consistent with the anticipated event frequency and intended precision.

Researchers should not rely uncritically on symmetric normal intervals when the outcome is extremely rare.

Can You Use a Pilot Study to Estimate Prevalence?

Yes, but estimates from small pilots may be unstable.

Suppose a pilot survey of 40 students observes one occurrence of a rare outcome.

The observed prevalence is 2.5%, but that estimate has considerable uncertainty.

Using it as though it were the exact population prevalence may lead to inappropriate sample size planning.

Researchers should examine relevant external evidence, uncertainty intervals, and plausible prevalence scenarios.

For uncommon outcomes, a small pilot may provide little information about the underlying event probability.

What If No Previous Prevalence Estimate Exists?

The appropriate response depends on the research objective.

For a standard single-proportion precision calculation, p = 0.50 provides a conservative initial assumption under the conventional formula.

However, that assumption may be unsuitable for planning a rare-event regression or a comparison involving a plausibly uncommon outcome.

Researchers may examine plausible prevalence ranges, consult relevant administrative records, conduct preliminary work, or use simulation.

Uncertainty should be represented explicitly rather than concealed through one unsupported estimate.

See estimating sample size without a formal calculation for broader planning alternatives.

Does a Larger Sample Solve Rare-Outcome Selection Bias?

No. Sample size and selection bias are different concerns.

For example, a voluntary survey about academic integrity may disproportionately attract students with strong opinions or particular experiences.

Recruiting more volunteers may increase the number of observations without eliminating systematic differences between respondents and the target population.

Probability sampling, appropriate weighting, careful measurement, and transparent assumptions may be necessary for defensible population estimation.

Researchers should avoid reporting a conventional probability-based margin of error for a convenience sample solely because it meets a numerical target.

How Should Prevalence Assumptions Be Reported?

The methodology should identify the anticipated prevalence or event probability used in planning and explain its source.

For proportion estimation, report the desired precision, confidence level, sampling design, and relevant adjustments.

For comparisons, report the anticipated probabilities in each group, effect measure, significance level, power, and allocation.

For regression or prediction, describe the expected event frequency, model complexity, and sample size criteria used.

If prevalence is uncertain, report sensitivity analyses across plausible values.

The justification should demonstrate why the selected assumptions are relevant to the actual research objective.

Why Common and Rare Outcomes Require Different Planning Strategies

These values use n₀ = 1.96²p(1 − p) / 0.05², rounded upward. They are initial normal-approximation values before finite population, complex-design, and nonresponse adjustments.

The 5% prevalence scenario produces a smaller calculated sample because its binomial variance is lower.

However, a margin of five percentage points is very wide relative to a prevalence of 5%. The resulting estimate would provide limited relative precision, and the normal approximation may be unreliable with few expected events.

Suppose the researcher instead wants to estimate the 5% prevalence within one percentage point.

The corresponding initial calculation becomes 1.96² × 0.05 × 0.95 / 0.01² = 1,824.76, requiring approximately 1,825 observations.

This larger requirement reflects the more demanding precision target.

If the researcher instead intends to model an outcome with 1% prevalence, the relevant planning problem may concern the expected number of events and model stability rather than a simple margin-of-error formula.

The example illustrates why prevalence cannot be interpreted independently of the statistical objective.

Common Misunderstandings About Prevalence and Sample Size

How to Use Outcome Prevalence in Sample Size Planning

Begin by specifying whether the research aims to estimate a proportion, compare event probabilities, or develop a model involving a binary outcome.

In the methodology section, identify the outcome probability used for planning and explain why it is appropriate.

Distinguish the number of participants from the expected number of events and avoid using one prevalence-based formula for unrelated analytical objectives.

Prevalence determines how information is distributed across binary outcomes, but the required sample depends on what the study intends to learn from that information.

Before Calculating Sample Size for a Binary Outcome, Check the Prevalence Assumptions

Frequently Asked Questions About Outcome Prevalence and Sample Size

Prevalence Matters Because It Changes Both Variance and Available Event Information

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.