How Do You Adjust Sample Size for Expected Dropout or Attrition?

If You Need 200 Participants at Follow-Up, How Many Should You Recruit Initially?

You are planning a semester-long experiment to investigate whether AI-generated feedback improves university students' academic writing performance. Your statistical power analysis indicates that 200 students with usable final assessment scores are needed.

However, not every student who enrolls is likely to complete the study. Some may withdraw from the course, miss the final assessment, decline further participation, or become unavailable during follow-up.

A colleague recommends recruiting an additional 20% of the required sample because you anticipate 20% attrition. Another suggests dividing the required sample by the expected retention rate.

These approaches produce different recruitment targets. Which calculation is appropriate, and what happens if students who leave the study differ systematically from those who remain?

Adjusting sample size for expected attrition is an important part of prospective research planning, particularly in longitudinal and intervention studies. Yet the calculation addresses only one aspect of the problem. Losing participants can reduce statistical information, while the reasons for their absence may also threaten the validity of the findings.

Divide the Required Usable Sample by the Expected Retention Proportion

How to Plan Recruitment When Participants May Leave Before the Study Ends

What Is Attrition in Research?

Attrition refers to the loss of participants or relevant participant information as a study progresses.

It is particularly important in longitudinal studies, randomized experiments, cohort studies, and research requiring repeated assessments.

For example, students may complete a baseline AI literacy questionnaire but fail to complete the writing assessment scheduled at the end of the semester.

Attrition may occur because participants withdraw consent, become unreachable, miss assessments, change institutions, or discontinue participation for other reasons.

However, researchers should distinguish withdrawal from the intervention, withdrawal from data collection, and missing individual outcome measurements. These circumstances may have different consequences for analysis.

Are Dropout, Attrition, and Missing Data the Same?

The terms overlap, but they are not interchangeable.

For example, a student may stop using an AI feedback platform but still complete the final writing assessment. That student has discontinued the intervention but may still contribute the primary outcome.

Conversely, a student may remain enrolled in the study but miss the final assessment.

The sample size adjustment should therefore focus on the expected availability of information required for the primary analysis rather than automatically counting every intervention discontinuation as a lost participant.

Why Does Attrition Affect Sample Size?

Sample size calculations usually depend on assumptions about the amount of usable information available for estimating or testing the primary effect.

When fewer participants provide relevant outcome data, standard errors may increase and statistical power may decrease.

For example, an experiment planned for 100 students per group may provide less information if only 70 students per group complete the primary assessment.

The consequences depend on the outcome, analysis, allocation, and missing-data pattern.

Attrition can also alter the composition of the observed sample, creating concerns beyond reduced precision.

Researchers should distinguish the numerical loss of information from the possibility that the remaining participants are systematically different.

What Is the Standard Formula for Adjusting Sample Size?

When the analysis requires a specified number of participants with usable data, a common recruitment adjustment is:

Always round upward when the calculation produces a fractional participant.

The formula is a recruitment-yield calculation, not a statistical correction for missing-data bias.

Why Is Adding 20% to the Required Sample Incorrect for 20% Attrition?

Adding 20% to the required sample and dividing by 80% retention are different calculations.

Suppose 200 usable observations are required.

Adding 20% produces 200 × 1.20 = 240 participants.

If 20% of those 240 participants are lost, the expected completed sample is only 240 × 0.80 = 192.

Dividing 200 by 0.80 produces 250 participants, with an expected completed sample of 250 × 0.80 = 200.

The difference arises because the attrition percentage applies to the initially recruited participants, not to the number researchers hope will remain.

How Does the Expected Attrition Rate Change Recruitment Requirements?

The required initial sample increases as anticipated attrition rises.

The calculations assume that every recruited participant has the same expected probability of contributing usable data.

Actual retention may differ across participant characteristics, intervention conditions, or study periods.

These values are planning targets, not guarantees that the required number will be achieved.

How Should You Estimate the Expected Attrition Rate?

Attrition assumptions should be informed by evidence relevant to the proposed study.

Useful sources may include previous studies involving similar populations, institutional follow-up records, pilot studies, and documented retention from comparable interventions.

For example, dropout in a brief online questionnaire may differ substantially from attrition in a semester-long intervention requiring repeated writing assessments.

The recruitment setting, participant burden, intervention duration, assessment schedule, and incentives may influence retention.

Researchers should avoid selecting a conventional 10% or 20% attrition assumption without considering whether it is plausible for their study.

Where evidence is uncertain, sensitivity analysis across several retention scenarios is preferable.

Should You Use the Attrition Rate from a Previous Study?

Previous research can provide useful information, but the rate should not be transferred automatically.

For example, a four-week AI learning experiment may have higher retention than a semester-long intervention involving multiple assessments.

Differences in participant motivation, recruitment procedures, institutional calendars, and follow-up methods may also matter.

Researchers should examine how attrition was defined in the earlier study.

A reported intervention discontinuation rate is not necessarily equivalent to the proportion missing the primary outcome.

The most relevant estimate concerns the type of usable information needed for the proposed primary analysis.

What If Attrition Differs Between Experimental Groups?

Group-specific attrition rates may require separate recruitment adjustments.

For example, an intensive AI feedback intervention may have a different retention rate from conventional instructor feedback.

If the study requires 100 usable observations in each group, expected retention of 75% in the intervention group and 90% in the control group produces different recruitment targets.

However, different attrition rates can create an imbalance in usable group sizes and may introduce bias if missingness is related to outcomes.

Researchers should evaluate the planned assignment ratio, expected usable allocation, and statistical analysis together.

Does Unequal Attrition Affect Statistical Power?

Yes. The precision of a between-group comparison depends on the amount of information available from both groups.

For a conventional independent-means comparison, a substantially smaller usable group may limit precision even when the total number of retained participants appears adequate.

For example, retaining 150 participants in one condition and 50 in another is not statistically equivalent to retaining 100 in each under equal-variance assumptions.

Consequently, a single overall retention percentage may conceal important group-specific information.

Researchers should consider the implications of unequal allocation for sample size when retention differs across conditions.

How Does Attrition Affect Longitudinal Studies?

Longitudinal studies may collect several observations from each participant.

For example, students may complete writing assessments at baseline, immediately after an intervention, and three months later.

Some participants may miss one assessment but return for another.

Others may discontinue participation permanently.

The amount of information lost depends on the timing and pattern of missing observations, not simply the percentage of participants with incomplete records.

Researchers should therefore calculate sample size according to the planned longitudinal model and expected missing-data structure.

See how repeated measures affect sample size requirements.

Does Every Participant Need Complete Data at Every Time Point?

Not necessarily. Some analyses require complete paired measurements, while others can incorporate partially observed longitudinal records under appropriate assumptions.

For example, a mixed-effects model may use available observations from participants who miss one follow-up assessment.

However, incomplete records generally provide different amounts of information depending on which measurements are missing.

The ability to analyze partial data does not mean that attrition has no consequences for statistical power or validity.

Researchers should not automatically apply a complete-case recruitment adjustment to a model designed to use incomplete longitudinal observations.

Simulation may be appropriate when missing-data patterns are complex.

What Is the Difference Between Missing Completely at Random, Missing at Random, and Missing Not at Random?

These concepts describe assumptions about the mechanism generating missing data.

These assumptions influence which analytical methods may be appropriate.

For example, students with poor writing outcomes may be less likely to complete the final assessment. If that tendency remains related to unobserved performance after conditioning on available data, standard MAR-based analyses may be inadequate.

The missing-data mechanism generally cannot be established conclusively from observed data alone.

Researchers should use substantive knowledge and sensitivity analysis rather than assume that missingness is harmless.

Why Can Attrition Introduce Bias?

Attrition may change the composition of participants contributing outcome data.

For example, students who struggle with AI-generated feedback may be more likely to discontinue participation or miss follow-up assessments.

If the remaining students disproportionately represent those who benefited, the observed intervention effect may be misleading.

The direction and magnitude of bias depend on the missingness mechanism and the estimand.

Importantly, low attrition does not guarantee the absence of bias, and high attrition does not necessarily imply a particular amount of bias.

The reasons for missingness matter as well as the number of missing observations.

Does Recruiting More Participants Correct Attrition Bias?

No. Increasing initial recruitment may help preserve the number of usable observations, but it does not automatically restore comparability between those who remain and those who leave.

For example, recruiting 300 students instead of 250 may produce more final assessments, but the results may still be biased if students with poor outcomes are systematically missing.

Researchers should distinguish statistical precision from the validity of the observed sample.

Retention strategies, appropriate missing-data methods, and sensitivity analyses may be necessary in addition to recruitment inflation.

Should Researchers Replace Participants Who Drop Out?

Replacing participants is not automatically equivalent to retaining the original sample.

In randomized experiments, adding new participants after observing dropout may change the recruitment process, follow-up duration, allocation, or participant composition.

Replacement procedures should not be introduced informally after examining outcomes.

Where additional recruitment is permitted, it should follow the study protocol and preserve the intended design.

Researchers should distinguish prospectively planned recruitment extensions from replacing particular participants because their data are inconvenient or incomplete.

Replacement also does not resolve the missing outcome information from the original participants.

What Is the Difference Between Attrition Adjustment and Nonresponse Adjustment?

Attrition generally concerns the loss of participants or information after enrollment or during follow-up.

Nonresponse often concerns failure to obtain participation or usable responses from initially sampled or invited units.

Both may require recruitment-yield adjustments, but they can involve different mechanisms and consequences.

For example, a study may need to invite additional students because some decline participation, then recruit enough enrolled students to account for later dropout.

These stages should be modeled separately when their probabilities and denominators differ.

See how to adjust sample size for nonresponse.

How Do You Plan for Both Nonresponse and Attrition?

When recruitment involves multiple stages, the number of invitations may be calculated using the product of expected participation and retention proportions.

The calculation assumes that the stage-specific probabilities are appropriately defined and that the product describes the expected overall yield.

It does not correct selection or attrition bias.

How Should Attrition Be Handled in Cluster-Randomized Studies?

Cluster-randomized studies may lose individual participants, entire clusters, or both.

For example, students may miss assessments while a classroom remains in the trial, or a school may withdraw entirely.

These losses have different statistical consequences.

Losing an entire cluster can substantially affect the number of independently randomized units.

A simple individual-level attrition adjustment may therefore be inadequate.

Researchers should account for the expected number of retained clusters, cluster sizes, intraclass correlation, and planned analysis.

See how clustering affects sample size.

Can Retention Strategies Reduce the Required Recruitment Target?

Potentially. Improving expected retention can reduce the number of participants who must initially enroll to achieve a given usable sample.

For example, clear assessment schedules, accessible follow-up procedures, appropriate reminders, and reduced participant burden may improve completion.

However, researchers should not assume that a retention strategy will achieve a particular rate without supporting evidence.

Retention procedures should also respect consent, privacy, and participants' right to withdraw.

Better retention can improve both efficiency and the quality of follow-up information, but its effectiveness should be monitored rather than presumed.

How Should Attrition Assumptions Be Reported?

The methodology should identify the required usable sample, anticipated attrition rate, evidence supporting that assumption, and resulting initial recruitment target.

For multi-group studies, report relevant group-specific expectations.

For longitudinal studies, explain which observations are necessary for the primary analysis and how incomplete records will be handled.

The results should distinguish participants approached, enrolled, allocated, retained, and analyzed.

Researchers should also describe missing-data methods and sensitivity analyses when attrition may influence the findings.

A transparent justification makes clear that recruitment inflation addresses expected information loss rather than guaranteeing unbiased conclusions.

Calculating Recruitment for a Semester-Long AI Feedback Experiment

The research team then examines how alternative attrition assumptions would change the target.

The values are rounded upward within each group.

The 20% scenario is appropriate only if the retention assumption is credible and the primary analysis requires the specified number of usable final outcomes.

Suppose students who experience difficulty using the AI tool are more likely to miss the final assessment. Even if 100 students remain in each group, the estimated intervention effect may be biased.

The researcher should therefore plan follow-up procedures, record reasons for missingness where appropriate, and specify a suitable missing-data analysis.

The example demonstrates that attrition adjustment is a necessary recruitment calculation in many studies, but not a complete strategy for protecting validity.

Common Misunderstandings About Attrition-Adjusted Sample Size

How to Set a Defensible Attrition-Adjusted Recruitment Target

Begin with the number of usable observations required for the primary analysis, not the number of participants you would prefer to recruit.

In the methodology section, report the original sample size requirement, expected attrition, adjusted recruitment target, and planned handling of missing data.

During the study, monitor retention using the denominators and outcome definitions specified in the protocol.

The practical objective is to preserve sufficient usable information while minimizing preventable missingness and protecting the validity of the intended inference.

Before Adjusting Sample Size for Attrition, Verify the Assumptions

Frequently Asked Questions About Dropout and Attrition Adjustments

Plan for Retention, but Do Not Confuse Extra Recruitment with Bias Correction

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.