Sampling Design Effects on Structure Estimates

Why Sampling Design Changes a Structure Estimate

A genetic structure analysis describes the individuals whose genotypes enter the analysis. It does not automatically describe everyone in a country, region, or clinical cohort. Geographic coverage, recruitment rules, and decisions about which ancestry profiles to include can all change the sample’s composition—and therefore change the estimated structure.

The key question is not simply whether a sample is large. It is which population the estimate is intended to describe. A carefully recruited sample of one city may characterize that city well but provide a poor summary of a geographically broad population. Conversely, a sample that includes many locations may still be unrepresentative if recruitment selects people with particular characteristics within those locations.

Definition: The target population is the set of individuals to which an analysis intends its conclusions to apply. An estimand is the particular population quantity of interest, such as the mean ancestry proportion in that target population or its distribution of genetic variation. A sample structure estimate is directly descriptive of the sampled individuals; extending it to a target population requires a sampling design that supports that interpretation.

In this tutorial, let \(q_i\) denote an individual’s estimated proportion of ancestry from one specified source, with \(0 \le q_i \le 1\). The same design principles apply to other summaries of structure, including allele frequencies, ancestry-proportion distributions, or patterns represented by genetic principal components. The numerical examples use \(q_i\) to make changes in composition visible.

Geographic Coverage and Population Composition

Genetic variation can change gradually across geography, as in a cline, or differ among locations for other historical and demographic reasons. If a sample covers only one part of a region, its estimated ancestry profile may reflect that location rather than the broader target. Even when every sampled person is genotyped accurately, restricted geographic coverage limits what the estimate represents.

Suppose a target population is divided into geographic strata \(h=1,\ldots,H\). Let \(W_h\) be the proportion of the target population in stratum \(h\), and let \(\mu_h\) be the stratum’s mean value of \(q\). The target mean is a weighted average of the stratum means:

\mu_{\text{target}} = \sum_{h=1}^{H} W_h\mu_h, \qquad \sum_{h=1}^{H} W_h=1.

An unweighted sample mean instead gives each sampled person equal influence. If the sample’s geographic proportions differ from the target’s, the sample mean need not estimate the target mean. The difference is not necessarily a flaw when the goal is to describe the sample itself; it becomes a problem when the sample estimate is interpreted as though it described a different, broader population.

Worked Example: Correcting Uneven Geographic Coverage

Hypothetical data. A target region consists of two areas. Area North contains 70% of the target population and has a mean ancestry proportion \(q\) of 0.20. Area South contains 30% and has a mean \(q\) of 0.60. A study samples 20 people from North and 80 from South. Assume the reported area means are accurate for this illustration.

The target mean is computed using the target geographic proportions:

\mu_{\text{target}} = (0.70)(0.20)+(0.30)(0.60) =0.14+0.18=0.32.

The sample mean gives the sampled proportions instead:

\bar{x}_{\text{sample}}=(0.20)(0.20)+(0.80)(0.60) =0.04+0.48=0.52.

The sample mean is 0.52, whereas the target mean is 0.32. The 0.20 difference arises here because South, the area with the larger mean ancestry proportion, is overrepresented in the sample. If the target is the full region and the area-specific estimates are suitable, weighting gives \(0.70(0.20)+0.30(0.60)=0.32\). If the intended target were instead the sampled set, 0.52 would describe that set. These are different estimands, not competing answers to the same question.

Recruitment Criteria and Selection

Geographic representation alone does not guarantee representativeness. Recruitment may depend on clinic attendance, study eligibility, language, access to a research site, or willingness to participate. If these criteria are associated with ancestry, location, or related social and demographic characteristics, the recruited sample can have a different structure from the target population.

For example, a clinic-based cohort estimates the structure of the people recruited through that clinic unless the design and analysis justify a broader interpretation. It may not estimate the structure of all residents in the catchment area. This distinction does not imply that clinic recruitment is inappropriate: it may be exactly right when the target is the clinic population. The crucial point is to state the target and avoid silently changing it.

Worked Example: A Recruitment Criterion Changes the Mean

Hypothetical data. In a target population, 25% of people meet a clinic-related recruitment criterion and have mean \(q=0.70\). The remaining 75% do not meet it and have mean \(q=0.30\). A study recruits 75% of its sample from the first group and 25% from the second. Assume the within-group means apply to the target groups and the recruited individuals are representative within each group.

The target mean, using the target group proportions, is

\mu_{\text{target}}=(0.25)(0.70)+(0.75)(0.30) =0.175+0.225=0.400.

The recruited sample’s expected mean, using its recruitment proportions, is

\bar{x}_{\text{sample}}=(0.75)(0.70)+(0.25)(0.30) =0.525+0.075=0.600.

Thus this recruitment pattern raises the sample mean from the target value of 0.400 to 0.600, a difference of 0.200 ancestry-proportion units. This is a compositional difference, not evidence that recruitment caused anyone’s ancestry proportion to change. If the target is the entire population and the group proportions are known, the weighted estimate is \(0.25(0.70)+0.75(0.30)=0.400\). If the target is people meeting the criterion, the relevant mean is 0.70.

Inclusion of Admixed Individuals

Ancestry is not always well represented by a choice between a small number of non-admixed reference groups. Admixed individuals may have ancestry proportions between those of the groups used as references, and they may be an important part of the target population. Excluding them can change the population being described and can remove meaningful variation from the analysis.

There is no universal rule that admixed individuals must always be included or always be excluded. If the estimand concerns a population that includes them, excluding them changes the target. If the scientific question is specifically about a defined subset, an exclusion may be appropriate, but the subset and the reason for restricting the analysis should be explicit. Likewise, an individual’s estimated ancestry proportion is a model-based summary, not a complete description of identity or a substitute for a population definition.

Worked Example: How Excluding Admixed Individuals Changes the Summary

Hypothetical data. A target cohort contains 40 individuals with mean \(q=0.05\), 20 with mean \(q=0.95\), and 40 admixed individuals with mean \(q=0.60\). Treat these group means as fixed for the calculation. The target cohort has 100 people.

The overall target mean is

\mu_{\text{all}}= \frac{40(0.05)+20(0.95)+40(0.60)}{100} =\frac{2+19+24}{100} =0.45.

If the admixed group is excluded, the remaining 60 individuals have mean

\mu_{\text{excluded}}= \frac{40(0.05)+20(0.95)}{40+20} =\frac{21}{60} =0.35.

Excluding the admixed individuals changes the mean by \(0.35-0.45=-0.10\) ancestry-proportion units and changes the population described from the full cohort to the non-admixed subset. Neither value is inherently preferable: 0.45 answers a question about the whole cohort, and 0.35 answers a question about the restricted subset. Reporting only the latter as though it represented all 100 people would be misleading.

A Justification for Composition Weighting

The weighting formula follows by grouping the target individuals by stratum. If stratum \(h\) makes up proportion \(W_h\) of the target and has mean \(\mu_h\), then its contribution to the overall mean is the fraction \(W_h\) times the stratum mean. Adding the contributions gives the target mean. In a finite target of \(N\) individuals, if stratum \(h\) contains \(N_h\) people, then \(W_h=N_h/N\), and:

\mu_{\text{target}} =\frac{1}{N}\sum_{i=1}^{N}q_i =\frac{1}{N}\sum_{h=1}^{H}\sum_{i\in h}q_i =\sum_{h=1}^{H}\frac{N_h}{N}\mu_h =\sum_{h=1}^{H}W_h\mu_h.

This identity explains why using the sample’s group proportions instead of the target’s proportions can change an estimate. It also clarifies a limitation: weighting addresses differences in the proportions of the specified strata, but it does not automatically remove selection differences within strata. For example, if participants recruited from one area differ systematically from nonparticipants in that same area, area weights alone do not make them representative of that area.

Planning a Sampling Design

A defensible sampling plan begins by specifying who the estimate is intended to describe and which structure summary will be reported. The plan should then cover the sources of heterogeneity relevant to that target, rather than choosing locations or eligibility criteria for convenience without considering their consequences.

Define the target and estimand.
State the population, time frame, and quantity of interest—for example, the mean ancestry proportion in residents of a defined region, or ancestry variation in a specified clinical cohort.

Map relevant sources of coverage.
Identify geographic areas and recruitment settings that may differ in genetic or demographic composition. For a clinal population, ensure sampling spans the relevant geographic range rather than relying on a single convenient location.

Make eligibility and inclusion rules explicit.
Decide whether admixed individuals and people meeting particular recruitment criteria belong to the target. Do not treat a model’s ancestry labels as an automatic sampling rule.

Record sampling information.
Retain location, recruitment pathway, eligibility, and other variables needed to describe coverage. When sampling fractions or target proportions are known, specify how they will inform population-level summaries.

Report the scope of interpretation.
Describe who was sampled, who was missed or excluded, and which population the estimate can reasonably represent. Distinguish a sample description from a claim about a wider target.

A probability-based design with known selection probabilities can support population estimates using appropriate design weights. In other settings, weights may be based on known target stratum proportions and representative stratum samples. Such adjustments require relevant sampling information and assumptions; weighting cannot recover unobserved groups or guarantee correction for unmeasured selection within groups.

Common Mistakes and Why They Matter

Assuming a large sample is automatically representative. A large, geographically concentrated sample can estimate its own composition precisely while differing from the broader target.
Interpreting the sample mean as a population mean without qualification. That interpretation needs a target definition and a sampling design or adjustment that supports it.
Using sample proportions as target weights. If the sample overrepresents a stratum, weighting by those same sample proportions preserves that overrepresentation rather than correcting it.
Excluding admixed individuals as if they were uninformative. Their inclusion or exclusion determines which population is summarized; either choice must match the estimand and be reported.
Assuming stratification solves every selection problem. Matching the target’s geographic proportions does not ensure representativeness among recruited and unrecruited people within each location.

Sampling choices matter because estimates of genetic structure are used to describe populations and to support downstream comparisons. If the sampled composition is mistaken for the target composition, differences attributed to ancestry or geography may partly reflect who was recruited. Transparent sampling descriptions let readers judge which conclusions are supported and which populations remain unrepresented.

Key takeaway: A structure estimate is meaningful only in relation to a defined target population. Geographic coverage, recruitment criteria, and inclusion of admixed individuals shape the composition being summarized; design the sample and any weighting around the intended estimand, and report the limits of representation.

Check Your Understanding

Answer each question using the distinction between the sampled individuals and the target population.

A target region is 60% eastern and 40% western. The sample is 30% eastern and 70% western. What information is needed to calculate a geographically weighted target mean of an ancestry proportion?
Explain why a clinic sample can validly describe clinic participants but not necessarily all residents in the clinic’s catchment area.
In a target population, 20% of people have mean \(q=0.80\) and 80% have mean \(q=0.25\). Calculate the target mean \(q\).
What changes in interpretation when admixed individuals are excluded from a cohort that otherwise includes them in the target population?
Why might geographic weighting fail to correct selection bias within a geographic stratum?