A low-incidence study is one where a small proportion of the general population qualifies: a rare medical condition, a niche category purchaser, a specific job function, a small ethnic or linguistic group within a market. The methodological problem is not finding the audience. It is reaching a usable base size for that group without distorting the study around it or spending disproportionately to get there.
This page covers fielding. For the separate engineering problem of getting a restrictive screener through survey testing, where simulation defaults stop simulated respondents terminating before they reach the rest of the questionnaire, see Simulating responses for restrictive screeners.
Step 1: Estimate incidence
Estimated incidence drives cost, timeline, and whether the study is achievable at all, so it should be measured rather than assumed.
Feasibility takes the target definition, an estimated incidence rate, and an expected duration, and returns achievability and cost against that specification. Run it on the real target rather than a simplified version, and run it on the alternatives too: a definition that is unreachable at 2 percent incidence may be workable at 6 percent with a marginally broader screen, and that trade-off is better made before fielding than after.
Feasibility is available in the platform and through the public API and MCP integration. See MCP and OpenAPI integration.
Step 2: Decide whether to combine related audiences
A combined design can improve effective incidence when several related audiences or items can share one questionnaire and sample frame.
Separate studies screen independently for each audience. A combined study instead admits the union of those audiences and uses routing to determine which questions or items each respondent sees.
This changes the incidence calculation, but it also creates design constraints: the questionnaire must keep respondent burden manageable, preserve a usable base for each item, and apply consistent screening and assignment logic.
The multi-brand case
Brand lift is the clearest example. Consider fifty brands.
Fifty separate studies each screen for the audience of one brand. Incidence is calculated separately, and differences in fielding window or sample frame can reduce comparability between brands.
A combined study screens for the union of the fifty brand audiences. Respondents answer only for the brands assigned to them, and every brand is measured from a common sample frame during the same fielding window. Effective incidence can therefore be much higher than it is for any one brand considered alone.
This can provide a significant cost advantage over other methods.
How it is implemented
- Survey Blocks define where item-specific questions appear without hard-coding them. The template controls screening, looping, and overall flow, while each brand or concept carries its own tailored follow-ups, editable visually rather than in code.
- Looping with per-respondent selection means no respondent answers about all fifty. Each sees a manageable subset, and the base per brand accumulates across respondents rather than across studies.
- Stimulus topic in lift reporting runs the analysis separately for each brand from the single dataset, so one study produces fifty independent reads. See Lift measurement methodology.
- Metadata tagging carries item attributes into reporting, so brands can be netted by category, tier, or client without restructuring the data.
Design checks
Base size per item is the constraint, not total sample. Two thousand respondents each seeing five of fifty brands yields roughly two hundred per brand. Size the study against the precision you need per item, not against the headline n.
Randomize item order. Position effects are real in long item lists, and an unrandomized order bakes them into the comparison the study exists to make.
Interview length still binds. A questionnaire covering many brands can increase respondent burden and break-off. Limit how many items each respondent sees and check the resulting base per item.
Screening logic must stay clean. A combined study needs a screener that admits the union of the audiences, with per-item logic deciding who is asked about what. Admitting everyone and sorting it out in analysis wastes interview time and muddies the base definitions.
Step 3: Design quotas with minimums, not just percentages
Quota lines can be expressed two ways, and the distinction is what protects a small group.
- A percentage of total completes target is used for weighting.
- A minimum number of respondents is used for boosting.
For each quota line, the platform calculates the count implied by the percentage and compares it to the configured minimum, and uses whichever is larger. This is what prevents a small but analytically important group from being sized out of significance by its own natural incidence.
A worked example. On a 500-complete study, a quota line at 5 percent implies 25 respondents. Set a minimum of 50 on that line and 50 is used instead. Once lines with binding minimums are locked in, the remaining completes are redistributed across the other lines in proportion to their percentages, normalized to what is left. On a 1,000-complete study with lines at 50, 30, and 20 percent, where the 20 percent line carries a minimum of 300, the result is 438, 262, and 300 rather than 500, 300, and 200.
Range quotas, with both a minimum and a maximum, are available where a group needs a floor for analysis and a ceiling to avoid over-representation. Full detail is in Boosting sample size for specific audiences.
Step 4: Use a boost source to protect the topline
Raising a subgroup base inside the main sample changes the composition of the main sample. A boost source avoids that.
A boost is a separate respondent source with its own targeting definition, configured to weight back to the primary source. The additional interviews raise the achieved base for the group, improving the precision of subgroup analysis, while the weighting scheme holds the overall distribution to the primary source's targets.
In practice: on a US general population study, a boost source targeting African American respondents at 300 additional interviews, weighted to the "US Genpop" primary source, substantially improves significance testing within that group without shifting the topline result. Setup is covered in How to set up respondent sources.
It works because it separates two questions that get conflated: how many of this group we need in order to say something reliable about them, and what the study as a whole should look like.
Step 5: Choose the fielding strategy
The fielding strategy determines when the survey closes and whether a respondent is accepted when their cell is already full. At normal incidence the choice is minor. At low incidence it dominates cost and timeline.
- Strict enforces quotas exactly, terminating any respondent who does not qualify for an open line. Cleanest composition, and the most expensive at low incidence, because hard-to-find respondents are turned away once their cell fills while easy cells are still being worked.
- Guaranteed quota prioritizes filling every quota line. Appropriate when the subgroup base is the point of the study.
- Guaranteed respondents prioritizes hitting the total target. Appropriate when overall n matters more than exact cell composition, with residual imbalance corrected by weighting.
The trade-off and the mechanics are documented in Quota management methodology. Where a looser strategy is used, the platform estimates the respondents required to hit quotas so the implication is visible before fielding.
Step 6: Go beyond online panel where panel is thin
Rare audiences are precisely the case where a single sourcing mode fails, because panel incidence for a rare group is a function of who that panel happened to recruit.
Options that change the reachable population rather than working the same pool harder:
- First-party lists. If the audience can be identified in a CRM, upload the list and generate unique links, up to one million. This changes the limiting factor from incidence to response rate.
- Interactive SMS, text-to-web, and AI voice. Reaches respondents who are not on panel at all, and works from voter files and texting-house lists.
- Call center. Interviewer administration for audiences where an assisted interview materially improves participation.
- In-ad surveys. Reaches the audience inside the media environment, which can be more efficient than screening for it cold.
- Several respondent sources. Footprints in a rare audience vary widely between sample providers. Fielding across several and comparing live incidence per source finds the ones that actually have the group.
See Respondent sourcing methodology.
Step 7: Watch incidence per source in field
Estimated incidence is a planning input. Live incidence measures what is happening during fielding, and the difference affects the number of entrants required to reach the target base.
The summary view reports live incidence per source, broken into completed, in progress, terminated, and poor quality. Compare sources against each other rather than against the estimate: a supplier materially below the others on a rare audience usually does not have that audience, and is better paused than widened. Allow at least 50 to 100 entries on a source before acting, since the figure moves during the first few hundred completes.
Termination detail by demographic in Field reports shows where in the screener the audience is being lost, which frequently reveals a screener problem rather than an incidence problem.
Step 8: Make the screener efficient
At low incidence most respondents entering the survey will terminate, so screener order affects how many questions non-qualifying respondents answer.
Order screening questions so the most restrictive criterion is asked first. Use Standard screeners for common market definitions so they remain consistent across studies. Avoid stacking multiple rare criteria into a single definition without checking the compound incidence, since two independent 10 percent criteria produce a 1 percent audience.
Step 9: Report the base alongside the result
A boosted subgroup is still a small subgroup, and the reporting should say so.
- Base sizes are configurable per question and should be shown, not suppressed, on subgroup analysis.
- Weighted margin of error is reported on all results, and boosting increases weight heterogeneity, which is reflected in the effective sample size adjustment described in Weighting methodology.
- Significance testing flags results that clear the configured threshold and filters those that do not, so a subgroup difference that a small base cannot support is not presented as a finding. See Understanding Stat Testing in MX8 Labs Reports.
A low-incidence result is precise enough to support some claims and not others. Make that boundary visible in the reporting rather than leaving a reader to assume the subgroup carries the same weight as the topline.