03 · What You Need to Know
How Recruiting Individuals Within Groups Shapes Research Design
What Does Recruiting Individuals Within Selected Groups Mean?
This arrangement occurs when researchers first identify or select collective entities and subsequently recruit individuals associated with them.
Groups might include schools, classrooms, hospitals, households, companies, neighborhoods, or other defined populations.
For example, researchers may select schools and then recruit teachers, select hospitals and then recruit patients, or select communities and then recruit households.
The design may involve formal probability sampling, purposive selection, convenience recruitment, or a combination of procedures at different stages.
Consequently, the term group-based recruitment describes the structure of participant access rather than establishing a particular sampling method.
How Is Group Selection Different From Individual Recruitment?
Group selection determines which institutions or collective entities become eligible recruitment settings. Individual recruitment determines which members of those settings are approached and ultimately participate.
These stages may use different procedures.
For example, schools might be randomly selected from a regional register, but teachers within those schools might distribute survey invitations only to classes they teach.
Alternatively, schools might be selected conveniently while students within each participating school are randomly sampled from enrollment records.
The second stage does not automatically correct selection limitations introduced at the first stage.
Researchers should distinguish the unit of recruitment from the unit of analysis and describe how both levels contribute to the final sample.
What Are the Main Design Implications?
| Design issue |
Why it matters |
What to examine |
| Group selection |
Selected groups may differ from groups outside the study. |
Sampling frame, selection procedure, coverage, and institutional nonparticipation |
| Individual selection |
Participation within groups may be uneven or systematic. |
Eligibility, invitations, response rates, and nonresponse patterns |
| Statistical dependence |
Individuals within groups may share outcome-related influences. |
Clustering, intracluster correlation, and model assumptions |
| Unequal group sizes |
Groups may contribute different amounts of information. |
Selection probabilities, weighting, and cluster-size variation |
| Sample size |
Individual counts do not necessarily represent independent group replication. |
Number of groups, individuals per group, and desired precision |
| Generalizability |
The recruited population may not represent the intended target population. |
Group coverage, population differences, and limits on inference |
These issues interact. For instance, selecting only technologically advanced universities can restrict population coverage while also reducing variation in institutional technology resources.
Is This Always Cluster Sampling?
No. Cluster sampling is a sampling design in which groups are selected as sampling units, usually as part of a probability-based procedure.
Recruiting individuals within selected groups may resemble cluster or multistage sampling, but the classification depends on how the groups and individuals were selected.
Suppose a researcher randomly selects 20 schools from a complete regional frame and then randomly selects 30 students within each school. This is a two-stage probability sampling design, assuming the selection procedures and inclusion probabilities are appropriately specified.
By contrast, inviting ten conveniently accessible schools and allowing students to volunteer is a group-based convenience recruitment strategy.
Both produce clustered observations, but they do not provide equivalent foundations for population inference.
Researchers should avoid calling a sample probability-based merely because it contains multiple schools or because individuals were randomly selected within conveniently chosen institutions.
How Do Inclusion Probabilities Work in Multistage Sampling?
In a simple two-stage probability design, an individual's inclusion probability depends on both the probability that their group is selected and the probability that they are selected within that group.
If these stages operate according to the specified conditional selection process, the overall inclusion probability is:
For example, if a school has a 0.20 probability of selection and an eligible student has a 0.10 conditional probability of selection within that school, the student's overall selection probability is 0.02.
The corresponding base sampling weight would be 1 / 0.02 = 50, before any justified nonresponse or calibration adjustments.
These calculations are meaningful when the probabilities arise from a documented sampling design. They should not be manufactured retrospectively for an uncontrolled convenience sample.
What If Some Groups Are More Likely to Participate?
Institutional participation can introduce selection bias when willingness or ability to join the study relates to the characteristics being investigated.
For example, schools with stronger administrative capacity may be more likely to participate in a study requiring extensive data collection.
If administrative capacity also relates to student outcomes, participating schools may differ systematically from nonparticipating schools.
Similarly, hospitals with established research offices may be more likely to participate than facilities with fewer resources.
Researchers should compare participating and nonparticipating groups using available characteristics such as location, size, institutional type, and relevant contextual indicators.
Observed similarity cannot prove that selection bias is absent, but substantial differences can reveal limitations requiring attention.
What If Individual Participation Also Varies Within Groups?
Even when groups are selected appropriately, individual recruitment can introduce further selection.
Teachers may distribute invitations unevenly, administrators may restrict access to particular classes, or students with stronger academic engagement may be more likely to complete questionnaires.
The resulting sample may therefore differ from the eligible population within each participating institution.
Researchers should document the number eligible, invited, and participating at each recruitment stage whenever feasible.
When individual response probabilities differ systematically, appropriate nonresponse adjustments may help under defensible assumptions and with adequate auxiliary information.
However, weighting does not automatically remove bias associated with unmeasured determinants of participation.
Why Does Recruiting Within Groups Create Statistical Clustering?
Individuals from the same group may share experiences or conditions that influence their outcomes.
Students within a school may experience common institutional policies, while patients within a hospital may receive care under shared clinical protocols.
This can create intracluster correlation, meaning observations within a group are statistically related.
Ignoring such dependence may produce misleading standard errors or confidence intervals.
The researcher should assess when clustering needs to be accounted for in research design, considering the estimand and how the sample was obtained.
Potential approaches include survey design-based estimation, cluster-robust standard errors, multilevel models, or other methods appropriate to the data structure.
How Does Intracluster Correlation Affect Precision?
When observations within groups are positively correlated, adding individuals from an existing group may provide less new information for certain estimands than recruiting individuals from additional independent groups.
For a simple equal-size cluster design, the following approximation is commonly used:
The approximation assumes equal group sizes and an exchangeable correlation structure. More complex sampling and analytical designs may require different calculations.
Should You Recruit More Groups or More Individuals per Group?
The answer depends on the research question, recruitment costs, intracluster correlation, and desired precision.
For group-level comparisons, additional independent groups are often particularly valuable because they increase information about variation between groups.
For individual-level descriptive questions, additional individuals within groups may also improve precision, although gains can diminish when clustering is strong.
For example, recruiting 500 students from two universities provides a different information structure from recruiting 500 students across 25 universities.
Neither design is automatically superior for every purpose. However, the two-university design provides much less institutional coverage and may be inadequate for estimating university-level relationships.
The distinction between participant counts and cluster counts should inform sample size planning before recruitment begins.
What If Group Sizes Are Unequal?
Selected groups may differ substantially in their number of eligible members and recruited participants.
For example, one university may contribute 400 respondents while another contributes 40.
Unequal group sizes can affect precision, sampling weights, and the population quantity being estimated.
If every individual receives equal weight, larger groups may contribute more strongly to the analysis. If every group receives equal weight, the estimate may instead represent an average across groups.
These approaches can answer different questions.
Researchers should specify whether their estimand concerns an average individual, an average institution, or another population quantity.
When group size is associated with outcomes, additional attention may be needed to avoid inappropriate weighting or informative cluster-size effects.
Can Group Selection Affect the Relationship Being Estimated?
Yes. Group selection can affect both population representation and the variation available for analysis.
Suppose a study of educational technology adoption recruits only universities with well-developed digital infrastructure.
Because these institutions have relatively similar infrastructure, the sample may provide limited information about the relationship between institutional infrastructure and student outcomes across the broader university sector.
Selection may also distort associations if institutional participation depends on variables related to both exposure and outcome.
Researchers should distinguish selection-related bias from reduced variation or limited transportability. These are related but not identical problems.
How Do Group-Based Recruitment and Multilevel Analysis Relate?
Multilevel models can represent outcomes measured on individuals nested within groups.
For example, a researcher may examine whether individual student motivation predicts achievement while allowing average achievement to differ across schools.
Such a model may distinguish individual-level relationships from contextual differences between schools.
However, multilevel modeling does not automatically make conveniently selected schools representative of a wider population.
Nor does it guarantee valid inference when the number of groups is very small.
The model should correspond to the research question, group structure, selection process, and assumptions about unobserved group characteristics.
What If Only a Few Groups Are Available?
Studies with few participating groups may have limited information for estimating group-level variation and relationships.
For example, 1,200 students from four schools provide many individual measurements but only four school-level cases.
Conventional cluster-robust inference may be unreliable with very few clusters, and complex multilevel models may depend heavily on assumptions that are difficult to assess.
Researchers may need to recruit more groups, simplify the analysis, use appropriately justified small-sample methods, or narrow the scope of their claims.
Statistical sophistication cannot substitute for independent groups that were never included.
What About Ethical and Administrative Considerations?
Group-based recruitment often requires cooperation from institutional gatekeepers.
Schools may approve access to students, hospitals may authorize access to eligible patients, and organizations may facilitate employee recruitment.
Institutional permission does not automatically replace individual informed consent where consent is required.
Researchers should also consider whether gatekeepers might pressure individuals, restrict invitations, or gain inappropriate access to individual responses.
Recruitment procedures should protect voluntary participation, confidentiality, and equitable access according to applicable ethical requirements.
In cluster-randomized research, recruitment after treatment allocation requires particular care because knowledge of allocation can influence who enters the study.
How Should the Design Be Reported?
Methods sections should identify the number of groups eligible, selected, approached, and participating, along with the number of individuals eligible, invited, and included within those groups.
Explain the selection procedure at each stage and whether probability-based sampling was used.
Report group sizes, relevant response rates, sampling weights, clustering adjustments, and limitations on generalizability.
If different recruitment stages involved different eligibility rules or response mechanisms, describe them separately.
STROBE provides relevant reporting principles for observational studies, while the CONSORT extension for cluster-randomized trials addresses additional concerns involving cluster allocation and recruitment.
