Research sampling: recruit the people your decision depends on
Define a research population and sampling frame. Compare probability, quota and synthetic designs and work through a B2B coverage problem.
Sampling starts with what you need to count. Imagine you run a B2B software company and want to understand renewal decisions. In this illustrative case, one account has one renewal event, but several people can explain it. Your customer relationship management system, or CRM, contains administrators; that contact list may miss other decision makers or eligible accounts.
Complete the audience brief before choosing a number
Before choosing a sample size, decide whether you need administrator experiences, linked committee perspectives or account renewal outcomes. Finance and security teams may influence the same renewal.
One renewal event, several linked perspectives
Record level | Details |
|---|---|
Account renewal event | Evidence: One eligible event in the stated period What is counted: One account outcome |
Administrator | Evidence: Workflow experience linked to account What is counted: One person perspective |
Finance / technical roles | Evidence: Relevant approval requirements linked to account What is counted: Additional perspectives, not additional renewals |
CRM frame gap | Evidence: Missing or stale eligible contacts What is counted: Access problem to repair |
The table shows why several role interviews should not become several renewal events. Keep each perspective linked to the account it explains.
A random selection of administrators is still a sample of administrators in that frame. It is not automatically a sample of all buying-committee members. The team must decide whether the question concerns administrator experience, account outcomes or committee decisions, then obtain evidence appropriate to that unit.
Interviewing several people from the same company adds perspectives but does not create the same number of independent company decisions. Preserve organization identifiers and account-level relationships in the analysis.
A consumer frame has different omissions
A DTC email list may exclude marketplace customers, guest orders, non-subscribers and people who considered the category but bought elsewhere. If the decision concerns existing identifiable customers, say so. If it concerns acquisition, add a suitable route to prospects and competing-option buyers. Randomizing names on the email list cannot repair the list's scope.
Population: accounts facing renewal in a stated study period. Eligibility: define eligible accounts and the relevant renewal event, and require recent, direct involvement in its evaluation, including renewed, declined or delayed decisions where the question needs them. Frame: the current CRM account/contact list, with stale or missing contacts identified.
Recruitment: find the administrator and additional roles relevant to each account; state how contacts will be verified and verify what decision each person made. Analysis: keep role-to-account identifiers so perspectives can be linked without multiplying outcomes, and keep person-level experience separate from the account’s renewal event. Deliverable: show the requirements that change the offer and which accounts or roles remain missing.
For DTC acquisition, the missing experience is different: an email list of customers misses people who chose another option. A suitable existing-customer frame can answer a retention question while being inadequate for acquisition. The decision decides the repair.
Choose the evidence design and its inference together
Evidence design | Details |
|---|---|
Human probability sample | Selection and information to disclose: Random selection from a defined contact/selection frame with known nonzero inclusion probabilities. Report population, frame, selection, response and analysis design. Precision limit: Sampling error depends on the design; other errors remain. |
Human nonprobability sample | Selection and information to disclose: Opt-in, convenience, purposive or quota recruitment. Report recruitment, eligibility, quotas, missing groups and inferential assumptions. Precision limit: Conventional random-sample precision does not automatically apply. |
Synthetic population | Selection and information to disclose: Generated or modeled units from grounding and construction choices. Report sources, coverage, model, scenario settings and validation. Precision limit: Generated N is not an independent human sample size. |
The AAPOR disclosure standards distinguish probability and nonprobability recruitment and call for explicit assumptions when reporting precision from nonprobability research. An online survey can use either human sampling design; the screen does not determine it.
“Online” is a mode, not a probability design. A frame and selection procedure establish how human records enter. Opt-in or quota recruitment has different assumptions from known random inclusion. A modeled population requires source and construction disclosure rather than a human sampling margin derived from its profile count.
Choose the recruitment procedure after access is understood
These examples are design choices, not guarantees of representativeness.
Method and business use | Details |
|---|---|
Simple random: estimate one customer-list proportion | How to select the units: Randomly select eligible accounts from a complete defined list. Practical limit: Missing accounts and nonresponse still matter. |
Stratified: include new and repeat accounts in a probability design | How to select the units: Divide the frame into defined groups, or strata, then randomly select within each. Practical limit: Analysis must reflect selection and stratum sizes. |
Cluster: study customers across sampled outlets | How to select the units: Randomly select groups such as stores, then the specified units within them. Practical limit: Within-store similarity affects precision. |
Systematic: select from an eligible account register | How to select the units: Use a random start, then every kth unit in a suitable ordered frame. Practical limit: Periodic ordering can distort selection. |
Purposive: interview abandoners, continuers and alternative users | How to select the units: Deliberately select relevant or contrasting experiences. Practical limit: Supports depth and contrast, not a random-sample margin. |
Quota: include purchase occasions in opt-in concept work | How to select the units: Recruit until predefined group counts are filled. Practical limit: Quotas do not supply known inclusion probabilities. |
Convenience: pilot wording with reachable category buyers | How to select the units: Include readily accessible eligible units. Practical limit: Findings may reflect who was easy to reach. |
Synthetic construction is a separate evidence design. A generated group labeled a quota is still a modeled population, with grounding and shared-model assumptions to disclose.
Stratification can ensure planned groups enter a probability selection when the frame supports it. Purposive selection can deliberately investigate an important contrasting experience. Quotas can fill planned characteristics in a nonprobability comparison. Those are useful functions, but they do not make the designs interchangeable.
Count roles, events and outcomes separately
Decide whether the report concerns administrator experiences, linked committee perspectives or account renewal events. A person-level distribution cannot become an account outcome by changing the chart title. Reconcile multiple role accounts and count the event once.
Keep disagreement within the account: the administrator may favor a feature while finance objects to cost. That conflict can explain a purchase requirement; averaging it can hide the decision.
Repair the frame when experiences are missing, linkage when perspectives are being miscounted and response volume when a suitable design has too few valid answers. A larger N from the wrong contact list does not make the original inference fit.
Set size around the analysis the team needs
Quotas can ensure selected characteristics or experiences are present in a nonprobability study. For example, a concept exploration might include new and repeat category buyers so the team can examine both perspectives. Quotas alone do not establish known inclusion probabilities or remove unmeasured selection differences.
For qualitative research, purposive selection can investigate varied experiences and contrary cases. Its purpose is explanatory coverage, not an implied population percentage. Describe the selection logic in the report.
For a simple random human sample estimating one proportion, agree the target confidence level, anticipated proportion and desired precision before setting the completed-answer requirement. Subgroup comparisons, clustered units, weighting, rare populations, experiments and nonresponse may require other planning. A whole-study total does not guarantee usable bases for every segment.
For synthetic work, plan scenario diversity, grounding and independent validation. A human margin-of-error formula does not establish precision for generated responses.
Use our research bias guide to examine missing groups and our B2B research guide for committee roles. Discuss a modeled audience with Sapien, including what the grounding supports and what the decision still requires.
Plan the subgroup or outcome bases, not only a whole-study total. An experiment, cluster design, rare audience or weighted estimate may need different planning from one simple proportion. Ask the analyst to explain the target quantity and design before accepting a sample-size claim.
The usable sampling deliverable is an audience definition, contact/selection route, expected gaps, analysis unit and inference. With those fixed, size can answer a concrete requirement instead of supplying a reassuring round number.
