Generalizing Beyond Observed Donors

Generalization Begins With the Target Population

A result from a single-cell study describes the donors who contributed data directly. Extending that result to other people requires an additional argument: why should the observed donors represent the people of interest? Measuring many cells from a few donors can sharpen descriptions of those donors, but it cannot by itself make the donors more representative of a wider population.

The population to which a result applies is not determined by the sequencing platform or the number of cells. It depends on the study’s scientific question, eligibility rules, recruitment setting, and the process by which eligible people agreed to participate. A claim about “people with the condition” can be too broad if donors came only from one specialist clinic, one geographic area, or a group reached by a particular recruitment route.

The estimand must also specify what is summarized for each donor. For example, let \(Y\) denote a particular donor-level expression summary, measured in normalized expression units, and let \(\theta_T=E_T(Y)\) denote its average in target population \(T\). The expectation refers to the distribution of donors in that population. Different choices of cell type, gene, expression summary, and target population define different estimands.

This donor-level framing is consistent with the earlier material on donor-aware analysis: cells from one donor are not independent people sampled from the target population. A donor contributing 20,000 cells still contributes one donor to the question of which people the study represents. Cell sampling can affect how accurately \(Y\) is measured for that donor, but recruitment determines which donors are observed.

What Convenience Recruitment Can and Cannot Support

A convenience sample is recruited through accessibility or availability rather than by a probability design that gives eligible donors known selection probabilities. Examples include recruiting volunteers who respond to an advertisement, enrolling patients attending one clinic, or using residual specimens from a particular hospital. Convenience recruitment is often practically necessary, but it does not automatically support generalization to all eligible people.

The central issue is selection. If participation is related to the outcome of interest, the recruited donors may have different outcome values from nonparticipants. Selection can also be related to characteristics that predict the outcome, such as age, disease severity, treatment history, or location. Even when the observed sample is large, the sample mean may then describe the recruited group rather than the target population.

A useful first check is to write down the target population in operational terms: eligibility criteria, location, care setting, and relevant time period. Then compare it with the recruitment pathway. “Adults with condition X in the region” is a different target from “adults with condition X receiving care at this clinic,” even if the same sample is used. Narrowing the target to match the study can make a claim more defensible, though it does not remove selection among people within that narrower group.

When Measured Differences Can Be Addressed

Sometimes a convenience sample has a different distribution of measured characteristics from the target population, but the relevant target distribution is known. A cautious adjustment is possible if the measured characteristics capture the important sources of selection related to the outcome and if the sample includes donors across the target’s relevant groups.

For a covariate \(X\), a key condition is that, within each value or group of \(X\), recruited donors have the same mean outcome as the target donors in that group. In conditional-mean notation, this is \(E(Y\mid X,S=1)=E_T(Y\mid X)\), where \(S=1\) indicates recruitment into the study. This condition is not guaranteed by measuring \(X\): it says that, after accounting for \(X\), selection does not leave a difference in mean \(Y\). A second condition is overlap: each group represented in the target population must also be represented in the study sample. If a target group has no recruited donors, its outcome mean cannot be learned directly from the sample.

When the target proportions \(P_T(X=x)\) are known, these conditions motivate standardization: estimate the mean outcome within each group in the sample, then average those group means using the target population’s proportions rather than the sample’s proportions.

A Short Justification for Standardization

The identity behind standardization is the law of total expectation. Partition the target population by the values of \(X\). Its overall mean is the sum of each group’s mean multiplied by the proportion of target donors in that group:

Under the conditional-mean assumption, \(E_T(Y\mid X=x)=E(Y\mid X=x,S=1)\). Substituting this equality gives

The sample supplies the group means, while external information supplies the target proportions. Both are necessary. If selection remains outcome-related within a group, the sample group mean is not the target group mean, and substitution fails. If a target group is absent from the sample, its group mean is unavailable, and the displayed expression cannot be computed from the observed data alone.

When the Sample Does Not Cover the Target

A target group with no sampled donors creates a more fundamental problem than a group with a different sample proportion. Reweighting cannot recover a mean for a group that was never observed. A statistical model might produce a prediction by extrapolating from other groups, but that prediction relies on additional assumptions and is not direct evidence from sampled donors in the missing group.

Assessing the Reach of a Result

Generalization is an argument supported by study design and context, not a property automatically conferred by a small standard error or a large cell count. Before making a population claim, examine who could enter the study, who was approached, who enrolled, and which target groups are represented. If recruitment records are available, differences between enrolled donors and the eligible population can be informative, although agreement on a few measured characteristics does not rule out differences on unmeasured ones.

A useful sequence is to define the target and estimand, describe the recruitment pathway, compare available characteristics, and state the assumptions that connect the sample to the target. If adjustment is used, report the covariates, target proportions, and any target groups with weak or absent sample coverage. Sensitivity analysis can show how conclusions change under plausible differences between recruited and nonrecruited donors; it cannot establish which assumption is true.

There are several alternatives, each answering a different practical need. Probability sampling from a well-defined target frame can support population inference through its known selection design, though it may be costly or infeasible. Restricting the target to the population actually reached makes the claim narrower and clearer, but still requires care about volunteer selection. Standardization can address measured differences when its assumptions and overlap condition are plausible. Regression-based prediction can extend into unobserved covariate regions only by relying on model assumptions; it does not make a missing group empirically represented. These approaches are not interchangeable, and none turns a convenience sample into a random sample after the fact.

Common Mistakes and Practical Limits

• Equating cell count with population representation. The wrong reasoning is that thousands of cells make a few donors representative. Cell number concerns measurement within donors; representation depends on which donors were recruited.
• Generalizing from a recruitment label alone. Saying “clinic patients” does not identify all patients with the condition. Specify the clinic setting, eligibility criteria, geography, and period that define the intended population.
• Treating covariate adjustment as a cure for selection. Matching age or sex distributions cannot address selection related to unmeasured factors or selection differences that remain within those groups.
• Reweighting a group that is absent. A weight cannot create information about an unobserved target group. Any estimate for that group requires additional data or explicit extrapolation assumptions.
• Confusing a precise estimate with a broadly applicable estimate. A narrow interval can describe uncertainty under an analysis model while leaving uncertainty about whether the recruited donors represent the target population.

The practical limit is often that the full target distribution and the reasons for nonparticipation are not known. Consequently, population claims should match the evidence: describe the observed donors directly, make a narrower target claim when justified, or state the assumptions needed for a broader claim. A transparent limitation is more informative than an unqualified generalization.