Why Sampling Design Matters for GWAS
A genome-wide association study needs more than genotype data: it needs participants whose phenotypes can be classified and compared under a clearly defined sampling plan. That plan determines whether researchers begin with people at risk of an outcome, recruit people according to outcome status, or use an existing biobank resource. It also affects how cases accumulate, what the comparison group represents, and which conclusions the sample can support.
For common-variant discovery, a design is useful when it can provide enough well-characterized cases and appropriate controls, with reliable phenotype information and usable genetic data. No sampling design guarantees discovery. A large sample with poorly defined outcomes or unsuitable controls can be less informative than a smaller, carefully assembled study.
Three Designs, Three Recruitment Logics
The distinction between a cohort and a case-control study is the recruitment logic, not simply whether records were collected in the past. A study can use existing records and still have cohort logic if participants were selected independently of the outcome and their outcomes are identified over a period. Conversely, a dataset assembled by selecting known cases and controls has case-control logic even if some source records were collected prospectively.
In a prospective cohort, researchers define eligibility and enroll participants before the outcomes of interest occur. They then use follow-up, clinical records, registries, or other planned measurements to determine who develops the outcome. Recruitment can be broad and population-based, or restricted to a particular eligible group. The cohort’s case count depends on the number enrolled, the rate at which the outcome occurs during follow-up, completeness of outcome observation, and whether participants have usable genetic data.
In a case-control study, researchers identify cases and select controls from the population that could have become cases under the study’s eligibility rules. Researchers can recruit many cases without having to enroll and follow a much larger group in which most participants may never develop the outcome. However, controls must be appropriate to the cases: controls drawn from a different source population may create a misleading comparison. Recruitment based on outcome status also means the sample’s case proportion is set by design rather than reflecting the outcome’s frequency in the population.
A biobank commonly combines broad participant recruitment with stored samples and linked health information. Depending on its recruitment and follow-up, it may function as a prospective cohort, an existing cohort with retrospective outcome identification, or a resource from which researchers select cases and controls. Its scale and breadth can help researchers assemble many participants, but case availability still depends on who joined, what outcomes are recorded, how complete those records are, and which participants have genetic data suitable for the analysis.
Outcome Observation and Case Availability
A cohort’s advantage is that outcomes can be observed after recruitment under a defined follow-up plan. This supports counting incident cases during that period, provided the outcome is measured consistently and follow-up is adequate. A cohort does not automatically yield many cases: for an uncommon outcome, even a large cohort may produce a modest number. Loss to follow-up and incomplete records can further reduce the number whose outcomes are known.
A case-control study can deliberately recruit a target number of cases and controls. This can be efficient when cases are uncommon or take a long time to develop. The number recruited is not, by itself, evidence that the sample represents the population’s outcome frequency. The case proportion in the study reflects recruitment choices, and estimating population risks or prevalence requires information beyond the case-control sample fraction.
In a biobank, phenotypes may come from participant reports, clinical records, or linked registries. These sources can make it feasible to search large existing collections for cases and controls, but recorded status is not necessarily complete status. Some participants may not have enough follow-up or data to classify the outcome. In addition, the people who enrolled in a biobank may differ from the population the researchers wish to describe. Prior work on cohort eligibility and target-population alignment is relevant: the intended population and the actual recruitment frame should be made explicit.
The count that matters for a genetic association analysis is not always the total number enrolled. It is the number with a usable outcome classification and genetic data, after applying eligibility rules and any required quality checks. For case-control comparisons, the usable numbers of cases and controls both matter. A large control group does not replace missing cases, and a high case count does not repair controls sampled from an unsuitable population.
Worked Example: Projecting Cases in a Prospective Cohort
A hypothetical study enrolls 80,000 eligible participants who do not have the outcome at baseline. The design anticipates that 3% will develop the outcome during five years of follow-up. Assume for this planning calculation that outcome ascertainment is complete and that 85% of the expected cases will also have usable genetic data. How many usable cases are projected?
First calculate the expected number of incident cases under the stated five-year rate:
Then apply the assumed 85% usable-genotype fraction:
The projection is therefore 2,040 usable cases. This is a planning estimate, not a guaranteed count. It depends on the assumed outcome rate, complete ascertainment, and the stated genetic-data yield. If follow-up is incomplete or the actual outcome rate differs, the number will change. The cohort also contains many participants who do not become cases; those participants may contribute as controls if the study’s analysis and sampling rules make them appropriate.
Comparing the Designs for Common-Variant Discovery
| Feature | Prospective cohort | Case-control study | Biobank resource |
|---|---|---|---|
| Recruitment | Enroll without selecting on the outcome; follow participants | Recruit by outcome status; select controls from the cases’ source population | Depends on how the resource recruited and retained participants |
| Outcome observation | Often planned follow-up, with incident outcomes observed after enrollment | Outcome status is used to identify cases and controls; observation may use existing or new data | May rely on follow-up, participant reports, records, or linked data |
| Case count | Accumulates from the enrolled cohort over follow-up | Can be set by recruiting a target number of cases | Depends on resource size, outcome records, and participant eligibility |
| Particular strength | Outcome occurrence can be observed within the enrolled cohort | Can assemble cases efficiently when cases are uncommon | Can provide scale and existing samples or linked data |
| Key limitation | Uncommon outcomes may yield few cases, even in a large cohort | Sample case fraction does not represent population frequency by design | Enrollment, records, and available data may limit who is represented or classifiable |
For common-variant discovery, all three can be suitable. A large cohort can be valuable when it accumulates sufficient cases and has reliable outcomes. A case-control design can be practical when direct case recruitment is more feasible than long follow-up of a very large population. A biobank can support either approach if its participants, phenotype data, and genetic data meet the study’s needs. The relevant design question is not which label is universally best, but whether the recruitment and outcome data yield a valid, sufficiently informative comparison for the intended association analysis.
Why Case-Control Sampling Can Preserve an Association
Outcome-based sampling changes how many cases and controls appear in the study. Under specific conditions, however, it need not change the genotype odds ratio being estimated. Consider a binary outcome \(D\), with \(D=1\) for cases and \(D=0\) for controls, and a genetic dosage \(G\) coded as 0, 1, or 2 effect alleles. Suppose the source population follows the logistic model
Here, \(\exp(\beta_1)\) is the source-population odds ratio for the outcome per one-unit increase in dosage, under this model. Let \(S=1\) indicate selection into the study. Assume selection depends on outcome status but, within each outcome group, does not depend on genotype: \(P(S=1\mid D,G)=s_D\). Also assume cases and controls come from the same defined source population and that the model is appropriate for the association being estimated.
By conditional probability and the selection assumption, the case-to-control odds among selected participants are the population odds multiplied by the case-to-control selection-probability ratio.
The sampled odds equal \(\frac{s_1}{s_0}\exp(\beta_0+\beta_1G)\). Taking logs adds \(\log(s_1/s_0)\) to the intercept, while leaving the coefficient of \(G\) equal to \(\beta_1\).
The sampled case fraction and logistic intercept may differ from their source-population values, but the dosage coefficient—and thus its odds ratio—remains the same under these assumptions.
In equations, the justification is:
This result explains why case-control sampling can be useful for estimating genetic odds-ratio associations. It does not mean the sample case fraction estimates population prevalence, nor does it automatically establish that controls are valid. If selection within case or control groups depends on genotype, or controls do not represent the cases’ source population, the argument does not apply as stated. The result also does not turn an odds ratio into a risk ratio.
Worked Example: Interpreting a Case-Control Recruitment Plan
A hypothetical study identifies 4,000 eligible cases in a health system and recruits 70% of them. It then recruits 1,400 controls from the population that produced those cases. Assume, for the purpose of this example, that recruitment within each outcome group is unrelated to genotype. How many participants are enrolled, and what does the resulting case fraction mean?
The number of recruited cases is:
Adding the 1,400 recruited controls gives:
The fraction of the enrolled sample that consists of cases is:
or about 66.7%. This is a consequence of the recruitment plan, not an estimate that 66.7% of the source population has the outcome. Under the selection assumption and the stated source-population condition, the genotype odds-ratio association can still be estimated from the case-control comparison. The sample fraction itself cannot be used to infer the population outcome frequency.
Worked Example: Finding Cases in a Hypothetical Biobank
A hypothetical biobank contains 150,000 participants. Usable outcome information is available for 88% of them. Among participants with usable information, 1.5% meet the study’s initial case classification. Assume 90% of these classified cases also have usable genetic data. How many genetically usable cases are available under these assumptions?
First find the number of participants with usable outcome information:
The number classified as cases is:
Finally, apply the assumed genetic-data yield:
The estimate is 1,782 usable cases, before any additional exclusions required by the analysis. The calculation shows why the total biobank enrollment is not the same as the effective case count. The resource could be analyzed as a cohort or used to select cases and controls, depending on its recruitment and the analysis plan. The calculation alone does not show that the outcome records are accurate or that the sampled participants represent a particular target population.
Common Mistakes and Design Checks
- Treating “biobank” as a sampling design. A biobank describes a resource. Determine whether the analysis uses cohort follow-up, outcome-based case-control sampling, or another defined design.
- Interpreting a case-control sample fraction as population frequency. The study may have recruited cases and controls in chosen numbers. Report the recruitment design rather than treating the sample fraction as prevalence.
- Selecting controls only because they are convenient. Controls should represent the source population from which the cases arose, under the study’s eligibility and sampling rules. Otherwise, the genetic comparison can reflect selection differences as well as outcome association.
- Equating enrollment with analyzable sample size. Outcome classification, eligibility, and usable genetic data can each reduce the available case and control counts.
- Assuming large scale solves phenotype limitations. A very large resource cannot compensate for systematically unreliable or incomplete outcome information. Recruitment scale and outcome quality must be assessed together.
A practical design review should therefore state who was eligible, how participants entered the study, how case and control status was determined, how many participants have usable outcomes and genetic data, and what population the comparison is intended to represent. These descriptions make clear whether the design is well suited to the planned common-variant association analysis and prevent sample composition from being mistaken for population frequency.
Check Your Understanding
Answer each question using the recruitment logic and assumptions described above.
- What feature distinguishes cohort recruitment from case-control recruitment?
- A case-control study enrolls 500 cases and 1,000 controls. Can the case fraction of one-third be interpreted as the outcome frequency in the source population? Explain.
- A biobank has 200,000 participants, but only a subset has usable outcome and genetic data. Why is the total enrollment not the relevant case count for association analysis?
- State the key selection condition under which outcome-based sampling leaves the genotype coefficient in the logistic model unchanged.
- Why must controls be sampled from the population that gave rise to the cases?