Defining the Target Population for Gene-Based Sequencing Studies

Why the Target Population Comes First

A gene-based sequencing result is not a result about “people” in general. It describes an association for a defined population, phenotype, gene, and set of variants under a particular analysis model. If the people eligible for a study differ from the people the investigators hope to describe, or if recruitment disproportionately includes some eligible people, the result may represent a narrower or different population than intended.

This distinction matters even when the sequencing and statistical analysis are flawless. A small p-value can provide evidence against a model-based null for the analyzed data, but it does not by itself establish that the same association applies to people who were ineligible, absent from the sampling frame, or unlikely to be recruited. The study’s intended population and its gene-level estimand should therefore be specified before interpreting a discovery.

Definition: The target population is the set of people about whom the study aims to make an inference, defined by relevant characteristics such as phenotype status, age, place, and calendar period. The sampling frame is the operational list or source from which potential participants can be identified. The analyzed sample is the set of recruited participants who meet the analysis requirements and have usable data.

Defining the Population in Operational Terms

A useful population definition is specific enough that another researcher could decide whether a hypothetical person belongs to it. “People with a rare disease” is usually too broad: Which diagnostic criteria count? At what age or time must the diagnosis be established? In which health system or geographic region? Are people with a particular acquired condition excluded? These choices are not administrative details; they determine the population to which the result refers.

A target definition should distinguish scientific scope from operational access. For example, investigators may want to learn about all adults with a specified diagnosis in a region during a defined period. A hospital registry may be the available sampling frame, but it may contain only diagnosed patients who received care in that hospital system. The registry is not automatically equivalent to the full target population.

Eligibility rules can narrow the target further. Rules about diagnostic certainty, age, relatedness, prior treatment, or sample quality may be justified for a particular scientific question or analysis. Their consequences should be stated plainly: if the intended inference is about adults with confirmed diagnoses who meet specified clinical criteria, then people outside those criteria are not part of that target. If the stated target remains broader, excluding them creates a gap between the target and the analyzed population.

Worked Example: Making a Target Population Explicit

Suppose a hypothetical study plans to sequence a gene in people with a particular inherited disorder. Its initial description says, “We will study patients with the disorder.” The investigators clarify that the intended population is adults aged 18–65 years, diagnosed under a named clinical definition, living in North District during 2026. They also decide that the primary analysis requires an unrelated participant and a usable sequence result.

Target population. For the stated scientific question, it is adults aged 18–65 with the specified diagnosis who live in North District during 2026. The diagnosis, age, location, and period make membership meaningful rather than vague.

Sampling frame. The investigators identify 240 potentially eligible people in a regional clinical registry. This is an operational source, not proof that every member of the target population is listed. For example, a target member treated outside the registry’s health system could be missing.

Analyzed sample. Of the 240 listed people, 18 do not meet the age or diagnostic criteria, 12 are excluded under the unrelated-participant rule, and 10 have unusable sequence data. If these groups are disjoint, the analyzed sample contains \(240-18-12-10=200\) people.

The result directly describes an analysis of those 200 participants. Its intended target is the defined eligible population, but the difference between that population and the registry frame must be considered before generalizing. In particular, unusable sequencing can reduce the analyzed sample without changing the scientific target; it does not make the missing participants cease to belong to that target.

The Gene-Level Discovery Estimand

An estimand is the precise population quantity a study aims to learn. In gene-based association analysis, it must identify more than a gene name. It depends on the target population, the phenotype and its measurement, the variant set or mask assigned to the gene, the genetic coding, and the model that defines association. Different choices can produce different estimands even when the same gene is analyzed.

For a quantitative phenotype \(Y\), let \(G_1,\ldots,G_m\) be the coded genotypes for the \(m\) variants included in a prespecified gene-level set, and let \(C\) denote the covariates in the model. A standard conditional mean model is

E(Y\mid G_1,\ldots,G_m,C)=\beta_0+\sum_{j=1}^{m}\beta_jG_j+\gamma^\mathsf{T}C.

Here, the gene-level association parameter is the vector \(\boldsymbol{\beta}_g=(\beta_1,\ldots,\beta_m)^\mathsf{T}\), defined for the target population under the stated model and coding. The global null for no conditional association of the included variants is \(H_0:\boldsymbol{\beta}_g=\mathbf{0}\). A gene-level test evaluates this joint null; it does not necessarily estimate one common effect for the gene. For a binary phenotype, a logistic model can instead define the variant coefficients as conditional log odds ratios, with the same joint-null form.

This definition is model-specific. For example, a change in the variant mask changes which coefficients are in \(\boldsymbol{\beta}_g\). A change in covariates changes the conditional association being described. If the model is only an approximation to the target population’s true outcome distribution, its coefficients should be understood as the parameters defined by that model rather than automatically as causal effects. Association alone does not establish that a variant causes the phenotype.

The test statistic and p-value are sample-based quantities; the estimand is the population parameter or hypothesis to which the analysis is directed. A significant result is evidence against the specified joint null under the test’s assumptions. It is not, by itself, a guarantee that every included variant has an effect, that the effect is clinically meaningful, or that the association applies outside the target population.

Worked Example: Stating a Gene-Level Estimand

Consider a hypothetical study of a continuous biomarker measured in adults with a defined disorder. For one gene, investigators prespecify three rare variants, code each genotype as the number of alternate alleles, and adjust for age and sex. The target is the eligible adult population in a specified region and period.

The model is

E(Y\mid G_1,G_2,G_3,\text{age},\text{sex}) =\beta_0+\beta_1G_1+\beta_2G_2+\beta_3G_3 +\gamma_1\text{age}+\gamma_2\text{sex}.

The gene-level estimand is \((\beta_1,\beta_2,\beta_3)\) for that target population and model. The discovery null is \(H_0:\beta_1=\beta_2=\beta_3=0\). If, hypothetically, the model-defined coefficients were \(0.8\), \(-0.3\), and \(0.0\) biomarker units per alternate allele, the vector would not satisfy the null, even though the third coefficient is zero. A joint gene-level test asks whether the vector is zero, not whether every single variant is individually significant.

The units of the coefficients are biomarker units per coded alternate allele, conditional on age and sex. These hypothetical values specify the meaning of the estimand; they are not a test result or evidence that the gene is associated. The population, phenotype, three-variant set, coding, and covariate adjustment are all part of the statement. Changing any one of them changes what association is being targeted.

How Recruitment Channels Shape Representation

A sampling frame can omit some target members, and recruitment channels can make participation more likely for some eligible people than for others. Recruitment through specialty clinics, patient organizations, broad community outreach, or electronic records may reach different subsets. A sample with many participants from one channel can therefore have a different composition from the target population, even if every participant meets the same written eligibility criteria.

The basic probability relationship makes this explicit. Let \(Z\) denote a characteristic such as ancestry group, and let \(S=1\) mean that a target member enters the analyzed sample. Then

P(Z=z\mid S=1) =\frac{P(S=1\mid Z=z)P(Z=z)}{P(S=1)}.

If the probability of entering the analyzed sample differs by \(Z\), the sample distribution of \(Z\) need not equal its target distribution. This does not prove that the gene-association estimate is biased: the consequence depends on how \(Z\), genotype, phenotype, and selection relate, as well as on the estimand and analysis model. It does show why sample composition and recruitment pathways are relevant to the population represented by a result.

Worked Example: A Recruitment Channel Changes Composition

Suppose the target population in a hypothetical region is 60% Group A and 40% Group B. A community clinic contributes 80 participants, 60 from Group A and 20 from Group B. A second channel contributes 20 participants, all from Group B. Assume, just for this example, that these channels together contain the complete analyzed sample of 100.

The analyzed sample has 60 participants from Group A and \(20+20=40\) from Group B. Its composition is therefore

\widehat{P}(A\mid S=1)=\frac{60}{100}=0.60,\qquad \widehat{P}(B\mid S=1)=\frac{40}{100}=0.40.

Here the sample composition matches the stated target proportions. But suppose instead that the second channel contributes 40 participants, all from Group B, while the clinic still contributes 60 from Group A and 20 from Group B. The sample then has 60 from Group A and 60 from Group B, so

\widehat{P}(A\mid S=1)=\frac{60}{120}=0.50,\qquad \widehat{P}(B\mid S=1)=\frac{60}{120}=0.50.

The sample is now 50% Group A and 50% Group B, whereas the target is 60% and 40%. The calculation establishes a composition difference, not the direction or size of any association bias. If allele frequencies or phenotype relationships differ between groups, however, the pooled sample and the target need not support the same interpretation. A researcher should report the recruitment sources and avoid presenting this sample as compositionally representative without evidence.

A Short Justification: Why Target and Sample Can Differ

The previous example used observed counts. The general implication follows directly from conditional probability. Suppose target members in group \(z\) enter the analyzed sample with probability \(q_z=P(S=1\mid Z=z)\), and the target proportion in that group is \(\pi_z=P(Z=z)\). The probability of selection is the total probability across groups:

P(S=1)=\sum_z P(S=1\mid Z=z)P(Z=z)=\sum_z q_z\pi_z.

Substituting this into Bayes’ rule gives the sample proportion

P(Z=z\mid S=1)=\frac{q_z\pi_z}{\sum_k q_k\pi_k}.

If every group has the same selection probability, \(q_z=q\), then the common factor cancels and the sample proportions equal the target proportions. If selection probabilities differ, they generally do not. Thus, the equality of sample and target composition requires either comparable selection across the groups considered or a design and analysis that appropriately accounts for the differences. Composition matching on one measured characteristic alone does not establish representativeness on all characteristics relevant to the gene-level association.

Eligibility, Generalization, and Interpretation

Eligibility rules have two distinct roles. First, they can define the scientific target: a study may intentionally ask about a narrowly specified clinical subgroup. Second, they can be practical filters imposed after a broader target has been stated. Those roles should not be conflated. If a study excludes people because their phenotype is uncertain, its result may be appropriately targeted to people with confirmed diagnoses, but it should not be described as applying automatically to everyone with suggestive symptoms.

The same care applies to recruitment channels. A result may be informative for the analyzed population and still have limited transportability to a broader population. Before extending an interpretation, ask whether the broader group has comparable phenotype definitions, variant coding and callability, relevant covariate distributions, and relationships among genotype, phenotype, and selection. These are questions about the scope of inference, not reasons to dismiss a valid result within its stated target.

For a gene-based discovery, a clear report should therefore identify the target population, the frame and recruitment sources, the eligibility and analysis exclusions, the phenotype definition, the gene’s variant set, and the association model. It should then state the gene-level null or parameter being tested. This makes it possible to distinguish a statistical discovery in the analyzed data from a claim about a different population.

Common Pitfall

Wrong reasoning: “All analyzed participants met the same eligibility criteria, so the sample represents everyone who meets those criteria.”
Correction: Shared eligibility does not ensure shared probabilities of appearing in the frame, being contacted, participating, and producing usable sequence data. Those steps can still select a subset. Describe the target and recruitment process, then limit generalization to what the design and analysis support.
Wrong reasoning: “A significant gene-level p-value means the gene has the same effect in every population.”
Correction: A p-value addresses a specified null under a particular model and analyzed data. It does not establish effect homogeneity or transportability. State the target population and model, and do not convert association evidence into a universal or causal claim.

Key takeaway: A gene-level discovery is interpretable only relative to a defined target population and a specified phenotype, variant set, coding, and model. Eligibility rules, the sampling frame, and recruitment channels determine which people contribute evidence and how far the result can reasonably be generalized.

Check Your Understanding

Answer each question by distinguishing the population a study wants to describe from the people it actually analyzes.

A study aims to describe gene association among all adults with a diagnosis in a state, but recruits only through two specialty clinics. Name the target population and one way the frame may fail to cover it.
Why does specifying a gene and a phenotype not fully define a gene-level estimand? Name at least three additional choices that matter.
If recruitment probability is equal across two groups, what does the conditional-probability formula imply about their sample and target proportions? What changes if the probabilities differ?
A study excludes participants with uncertain diagnoses. How should that exclusion affect the wording of the population to which its findings apply?
What does rejection of the joint null \(H_0:\boldsymbol{\beta}_g=\mathbf{0}\) establish, and what does it not establish about individual variants or other populations?