Should Underrepresented Groups Be Oversampled When Subgroup Inference Is Important?

Should Researchers Recruit More Participants From Underrepresented Groups to Support Subgroup Comparisons?

A research team plans to evaluate an educational intervention across several student populations. One group represents only 10% of the eligible population, but the researchers want to determine whether the intervention works differently for that group.

If participants are recruited in proportion to their population distribution, relatively few members of the smaller group may enter the study. Their treatment-effect estimates could be imprecise, and the analysis may be unable to detect differences that matter.

Should the researchers deliberately recruit a larger proportion of participants from that group? Would doing so improve the scientific and equity-related value of the study, or would it compromise representativeness?

The answer depends on the intended inference. Oversampling can improve information about a population that would otherwise be poorly represented, but it also changes the sample composition and may require careful design and analysis.

Oversampling Is Often Appropriate When Important Subgroup Questions Would Otherwise Remain Unanswered

How Oversampling Can Strengthen Subgroup Inference and When It May Introduce New Problems

What Does Oversampling Mean in Research?

Oversampling means deliberately selecting or recruiting a greater proportion of participants from a particular population than would ordinarily occur under a specified reference sampling strategy.

For example, if a population contains 10% members of a particular group, a proportionate sample of 1,000 participants would be expected to contain approximately 100 members of that group. A disproportionate sampling design might instead recruit 300 members to obtain more precise subgroup estimates.

Oversampling does not necessarily mean recruiting more participants overall. Researchers may redistribute a fixed sample across groups or increase total enrollment to obtain additional information about a particular population.

These approaches have different implications for statistical precision, recruitment costs, and the intended population-level analysis.

Why Might Proportional Recruitment Be Inadequate for Subgroup Analysis?

Proportional recruitment can provide a sample composition that resembles the target population, assuming an appropriate probability-sampling process and sufficiently representative participation.

However, a sample that reflects population proportions may contain too few participants from smaller groups to support precise subgroup-specific estimates.

Suppose an intervention study includes 500 participants and a particular group constitutes 8% of the population. Proportional recruitment would yield approximately 40 participants from that group.

If the study compares intervention and control conditions, those 40 participants may be divided into roughly 20 per treatment arm. The resulting estimates may have wide confidence intervals, particularly when outcomes are variable or events are uncommon.

Thus, a sample can be reasonably representative of population composition while remaining poorly suited to answering an important subgroup question.

When Is Oversampling Scientifically Justified?

Oversampling is most defensible when a clearly defined research objective requires adequate information about a population that would otherwise contribute relatively few observations.

Several circumstances may support this decision.

• Confirmatory effect modification: Researchers intend to determine whether intervention effects differ between a smaller group and another population.
• Subgroup-specific estimation: The study requires sufficiently precise estimates within a particular population, even when no formal interaction hypothesis is planned.
• Safety assessment: Important adverse outcomes may need to be examined in populations for whom the intervention could have different consequences.
• Equity-related questions: Researchers seek evidence about populations that have historically been insufficiently represented in relevant studies.
• Measurement or implementation research: The study must determine whether instruments, procedures, or interventions function appropriately across populations.

These reasons should be evaluated separately. Improving representation, estimating a subgroup effect, and detecting an interaction are different objectives with different sample size requirements.

Is Oversampling the Same as Stratified Sampling?

No. Stratified sampling divides a population into defined strata and selects participants according to a sampling strategy within each stratum.

Oversampling may be implemented through disproportionate stratified sampling, in which a particular stratum receives a larger sampling fraction than its population share.

However, stratification does not necessarily involve oversampling. Proportionate stratified sampling may preserve the population distribution.

Oversampling can also occur through targeted recruitment in nonprobability samples, although such approaches introduce different considerations for representativeness and selection bias.

How Does Oversampling Improve Statistical Precision?

Subgroup-specific estimates become more precise as the amount of relevant information increases, assuming the additional observations are informative and the statistical model is appropriate.

For a simple comparison of two independent subgroup means with equal variances, the standard error of their difference depends on the sample size in each group.

Under these simplified assumptions, the more balanced allocation provides a more precise estimate of the difference between group means.

This formula concerns a direct comparison of means, not a treatment-by-subgroup interaction. An interaction involves comparing treatment effects across groups and generally requires information from treatment-by-subgroup combinations.

Nevertheless, the illustration shows why increasing representation of a small group can improve the precision of comparisons involving that group.

How Does Oversampling Affect Power for Treatment-by-Subgroup Interactions?

When the scientific question concerns whether treatment effects differ between groups, the relevant statistical target is an interaction.

Suppose a randomized study includes a group representing only 10% of the eligible population. Under proportional recruitment, relatively few participants from that group may be allocated to each treatment condition.

Oversampling can increase the number of observations in those treatment-by-subgroup cells and improve the precision of the interaction estimate.

However, the benefit depends on the subgroup distribution, outcome variability, treatment allocation, and other design characteristics.

Researchers should calculate statistical power for the intended subgroup interaction rather than assuming that a particular oversampling percentage will be sufficient.

For rare-event outcomes, the anticipated number of events may remain limiting even after additional recruitment.

Should Researchers Aim for Equal Numbers in Every Subgroup?

Not automatically. Equal subgroup allocation can be statistically efficient for certain comparisons under simplified assumptions, but it is not a universal requirement.

The optimal allocation depends on the study's objectives, outcome variability, recruitment costs, subgroup prevalence, and the importance of pooled versus subgroup-specific estimates.

For example, a study primarily intended to estimate a population-average outcome may benefit from a different allocation than a study designed to compare two subgroup treatment effects with similar precision.

Recruiting equal numbers from groups with very different population sizes may also require substantial effort and could affect the feasibility of the project.

Researchers should select the allocation that serves the intended inference rather than treating numerical equality as an automatic methodological ideal.

Does Oversampling Compromise Representativeness?

Oversampling changes the sample's composition relative to the reference population, but it does not necessarily invalidate population-level inference.

In probability-sampling designs, unequal inclusion probabilities can often be addressed using appropriate sampling weights and design-based estimation procedures.

For example, if one population group is deliberately sampled at a higher rate, each sampled member of that group may represent fewer individuals in the target population than a participant from a less intensively sampled group.

Correctly constructed weights can help recover population-level estimates under the relevant sampling assumptions.

However, weighting does not automatically eliminate nonresponse bias, coverage error, measurement problems, or unmeasured selection mechanisms.

In convenience samples or targeted recruitment designs without known inclusion probabilities, population representativeness cannot simply be restored by assigning weights based on demographic proportions.

When Are Sampling Weights Necessary?

The need for weighting depends on the estimand, sampling design, and analytical method.

If the objective is to estimate a population mean or prevalence from a disproportionate probability sample, design weights may be necessary to account for unequal selection probabilities.

If the objective is a subgroup-specific estimate, the relevant analysis may require a different treatment of weights, depending on the sampling process and model assumptions.

For randomized trials, oversampling changes the distribution of participant characteristics in the enrolled sample. Randomization can still support internally valid treatment comparisons within the trial, provided its implementation and analysis are appropriate.

However, the trial's unweighted average treatment effect may represent the enrolled sample rather than the broader target population when treatment effects vary across the oversampled characteristic.

Transporting or generalizing treatment effects to a target population may therefore require additional assumptions, external population information, or suitable weighting and standardization methods.

Researchers should distinguish correcting a probability-sampling design from transporting a trial result to a different population. These are related but not identical analytical tasks.

Can Oversampling Introduce Selection Bias?

Yes, particularly when targeted recruitment changes not only the proportion of a group in the sample but also which members of that group participate.

For example, recruiting underrepresented students primarily through highly engaged campus organizations may disproportionately include students with stronger institutional connections.

Those participants may not represent the broader population of students who share the demographic characteristic.

Oversampling can therefore increase sample size while leaving important selection problems unresolved.

Researchers should examine recruitment pathways, eligibility criteria, participation barriers, and nonresponse patterns. Where feasible, probability-based sampling or carefully designed recruitment procedures can strengthen the interpretation of subgroup estimates.

Does Oversampling Automatically Improve Research Equity?

No. Increasing the number of participants from an underrepresented population may improve the availability of evidence, but equity also concerns how research questions are selected, how participants are recruited, and whether the findings address issues important to the population.

Researchers should consider whether recruitment procedures are accessible, whether participation burdens are reasonable, and whether relevant communities have opportunities to inform the study's priorities.

Oversampling should not be treated as a purely numerical exercise in which representation is improved without considering the conditions of participation.

Equity-oriented recruitment may also require attention to language, accessibility, compensation, scheduling, trust, and institutional barriers.

The appropriate approach depends on the research context and the populations involved.

Should Every Demographic Group Be Oversampled?

No. The decision should follow the scientific and ethical objectives of the study.

Age, sex, gender, ethnicity, disability, socioeconomic background, and other characteristics may be relevant to representation or effect modification, but their presence in a dataset does not automatically justify separate confirmatory comparisons.

Researchers should first determine whether demographic subgroup analysis is scientifically appropriate.

If subgroup-specific inference is important, the recruitment plan should be aligned with the required precision and the characteristics of the target population.

Oversampling multiple intersecting populations may introduce additional complexity, particularly when some combinations are rare or difficult to recruit.

How Should Oversampling Be Planned Before Recruitment?

Researchers should begin by specifying the intended subgroup estimands and the precision or power required to support them.

Next, they should estimate subgroup prevalence, anticipated recruitment rates, outcome variability, and relevant event frequencies.

The proposed allocation should be evaluated using design-specific sample size calculations or simulations.

Recruitment feasibility should also be considered. Some populations may be difficult to reach through standard procedures, and additional recruitment effort may not translate proportionately into completed participation.

The statistical analysis plan should explain how the oversampling design will be reflected in estimation, variance calculation, and any population-level generalization.

Where subgroup hypotheses are confirmatory, the plan should also specify the interaction tests and relevant multiplicity strategy.

When Might a Separate Study Be Better Than Oversampling?

Oversampling within an existing study may be insufficient when the subgroup requires different measurements, intervention adaptations, recruitment procedures, or contextual considerations.

For example, a study designed around university students in conventional online courses may not adequately address the experiences of working adults in a substantially different learning environment merely by recruiting more of them.

If the subgroup question requires a distinct design or substantially different implementation conditions, a separate focused investigation may provide more defensible evidence.

Researchers should distinguish a shortage of subgroup observations from a mismatch between the original study design and the subgroup research question.

How Oversampling a Smaller Population Can Improve Interaction Precision

For the simplified difference-in-differences interaction, the standard error is calculated from the four treatment-by-subgroup cell sizes.

Under the stated assumptions, oversampling reduces the interaction standard error by approximately 39% without increasing the total number of participants.

This improvement does not establish that the study has adequate power for a particular interaction. The researchers must still specify the smallest meaningful difference, significance criterion, and desired power.

The example demonstrates how oversampling can improve interaction precision when a smaller population would otherwise contribute relatively little information.

However, the resulting sample contains 40% working students even though the target population contains approximately 10%. Any estimate intended to represent the broader population must account appropriately for that difference and the underlying sampling process.

Common Misconceptions About Oversampling and Subgroup Inference

How to Decide Whether Oversampling Is Worth the Additional Recruitment Effort

Begin by identifying the intended subgroup inference. If the objective is merely to describe participant composition, proportional sampling may be adequate. If the objective is to estimate an important subgroup effect or detect an interaction, the smaller group's information requirements deserve explicit consideration.

Next, evaluate the subgroup's expected prevalence, recruitment feasibility, outcome variability, and the sample size needed for the intended precision or power.

The decision should also consider whether the subgroup hypothesis is scientifically justified. A more balanced sample does not establish that the proposed interaction is meaningful or plausible.

When planning recruitment, document the target population, sampling frame, subgroup definitions, intended allocation, and analytical consequences of disproportionate enrollment.

Researchers should be particularly careful when oversampling demographic populations. Group categories must be appropriately defined, and recruitment should avoid reinforcing unsupported assumptions about homogeneous experiences or responses.

The practical objective is to obtain sufficient information for a question that matters while preserving the validity and transparency of the study's conclusions.

Before Oversampling an Underrepresented Group, Check the Design Requirements

Frequently Asked Questions About Oversampling Underrepresented Populations

Oversample When Better Subgroup Evidence Justifies the Design

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.