Why Sampling Units Matter
A single-cell data set can contain hundreds of thousands of measured cells and still represent only a small number of people. The count of rows in a cell-level matrix is therefore not, by itself, a count of independent biological replicates. To understand what a study can support, identify the objects sampled at each stage: often donors, then specimens from those donors, then cells recovered from each specimen.
The distinction matters because the intended population and the sampling process determine what the observed data can reasonably represent. A study that recruits donors from one clinic and sequences cells from one tissue type does not automatically represent all people, all tissues, or all cells in the body. A careful design description names both the target population and the feasible sampling frame: the population the investigators intend to learn about, and the concrete set of eligible units from which they can actually obtain observations.
Units at Different Levels
If a condition is assigned to a whole donor, the donor is the experimental unit for that assignment. If the study observes naturally occurring disease status rather than assigning it, “experimental unit” may not be the most informative label; the donor is still commonly the independent biological sampling unit. A specimen, such as a biopsy, can be a sampling and measurement unit within a donor. A cell is usually an observational unit for cell-level measurements and may be sampled from a specimen, but cells from the same donor are not generally independent biological replicates.
The appropriate unit depends on the question and design, not on which data table has the most rows. For example, if a study compares disease groups using one specimen per donor, donors provide the between-person replication. If it measures multiple tissues per donor, specimens are repeated observations nested within donors. If thousands of cells are captured from each specimen, those cells provide detailed information about that specimen, not thousands of independently recruited donors.
| Level | Possible unit | What it represents | Common limitation |
|---|---|---|---|
| People | Donor | Between-person biological variation | A narrow donor pool limits population coverage |
| Material within a person | Specimen, tissue, or biopsy | Variation across sampled sites or materials | Several specimens may share donor-level characteristics |
| Measured objects within material | Captured cell | Cell-level composition or expression in the assayed material | Cells from one specimen share biological and technical context |
Target Population and Feasible Sampling Frame
A target population should specify the unit and relevant scope. “People with condition A” is incomplete if eligibility depends on age, care setting, geography, or time period. “Cells” is also incomplete if the intended cells are from a particular tissue, cell type, or physiological state. The study protocol should make clear which units and characteristics define eligibility, even when the final analysis uses measurements at a different level.
A practical single-cell sampling process often has multiple frames. A donor roster may include people eligible for recruitment; a specimen frame may include collected biopsies that contain the tissue of interest; and a cell frame may consist of the cells that survive dissociation, capture, sequencing, and quality control. Each stage can narrow or alter the population represented. The cells observed at the end are not necessarily a probability sample of all cells in the target tissue: recovery can depend on cell type, tissue handling, viability, and other factors.
It is useful to distinguish who is in scope from who could be reached. The target might be adults with a particular disease across a region, while the feasible donor frame consists of eligible patients attending two specialist clinics during a recruitment period. If those clinics serve a different mix of patients from the region, the frame may not cover the target population well. Naming the mismatch does not solve it, but it makes the limits of generalization visible.
Worked Example: Finding the Independent Replication Level
Suppose a hypothetical study collects one specimen from each of 6 donors and captures 200 cells per specimen. Disease status differs between donors; it is not randomly assigned to individual cells. How many biological sampling units support a comparison between disease groups?
There are \(6\) donors, \(6\) specimens, and \(6 \times 200 = 1{,}200\) captured cells. The 1,200 cells are observational units for cell-level measurements. But the disease status is a donor-level characteristic, and cells from the same donor share that status and biological background. Thus, the independent biological replication for a between-donor disease comparison is at most 6 donors, not 1,200 cells.
For clarity, suppose 3 donors are in each group. The group sizes for independent biological replication are 3 donors and 3 donors. Reporting 600 cells per group does not change those donor counts. Increasing the number of captured cells per specimen may improve the description of each specimen, but it does not add new donors or independent disease assignments.
This count assumes the six donors are distinct and the group labels are correctly recorded at donor level. If two specimens came from one donor, they would not be two independent donors. The study would need to preserve the donor linkage rather than treat those specimens as unrelated people.
Why Cells from One Donor Are Not Independent Replicates
A short variance argument shows why cell counts can exaggerate precision when within-donor dependence is ignored. Consider a simplified measurement \(Y_{ij}\) for cell \(j\) from donor \(i\), with \(D\) independent donors and \(m\) cells per donor:
Here, \(u_i\) represents donor-to-donor variation shared by that donor’s cells, and \(e_{ij}\) represents cell-level variation. Assume for this illustration that donor effects are independent across donors, cell-level errors are independent across cells conditional on donor, and the two components are independent. The mean for donor \(i\) is \(\bar{Y}_i = \mu + u_i + \bar{e}_i\), where \(\operatorname{Var}(\bar{e}_i)=\sigma_e^2/m\). Therefore:
The donor component is divided by \(D\), not by the total number of cells \(Dm\). If all \(Dm\) cells were incorrectly treated as independent measurements with marginal variance \(\sigma_u^2+\sigma_e^2\), the resulting variance calculation would be \((\sigma_u^2+\sigma_e^2)/(Dm)\), which divides the shared donor variation by \(Dm\). That is too small when \(\sigma_u^2>0\) and \(m>1\). This argument relies on its stated variance model; real studies can have additional dependence, unequal cell counts, or specimen-level variation. Its central lesson remains: the unit that supplies independent replication must be identified before interpreting precision.
Worked Example: Quantifying the Cost of Ignoring Donor Clustering
For a hypothetical standardized cell measurement, suppose there are \(D=4\) independent donors, \(m=100\) cells per donor, donor variance \(\sigma_u^2=0.16\), and cell-level variance \(\sigma_e^2=0.84\). Under the variance model above, calculate the variance and standard error of the average across donors and cells.
The variance that respects donor clustering is:
The corresponding standard error is \(\sqrt{0.0421}=0.2052\) standardized units. If the 400 cells were incorrectly treated as independent, the calculated variance would be:
The smaller value is not evidence that the study has 400 independent biological replicates; it is the consequence of omitting shared donor variation. The calculation assumes equal cell counts and the simple variance decomposition stated above. It illustrates the direction and possible size of the error, not a universal correction for every single-cell analysis.
Building and Checking a Multistage Frame
A useful design description records each transition from the target population to the analyzed data. At each stage, ask which units are eligible, how they are identified, how selection occurs, and what causes a unit to be lost. A frame need not be a literal list of every cell: it may be an operational procedure, such as collecting a specified tissue and retaining cells that meet predefined quality criteria. What matters is stating what that procedure covers and what it excludes.
Name whether the target concerns people, specimens, or cells, and state the defining population characteristics and scope.
Describe the source of eligible donors, such as clinics, registries, or a defined community, and note who may be absent from that source.
State which tissue or specimen types are collected, whether multiple specimens can come from one donor, and how unavailable or unusable material is handled.
Record the capture and quality-control steps that determine which cells become observations, and consider whether recovery may vary across cell types or samples.
Keep donor and specimen identifiers linked to each cell so analyses and reports do not mistake nested observations for independent units.
Diagnostics begin with counts at every level, not just a total cell count. Tabulate eligible, recruited, sampled, and analyzed donors; specimens per donor; and cells retained per specimen. Look for exclusions concentrated in one condition, donor group, tissue source, or cell type. Large differences can indicate that the analyzed frame is narrower than intended or that the composition of observed data depends on the sampling and processing pathway.
Worked Example: Measuring Frame Coverage and Attrition
A hypothetical study aims to describe adults with a specified condition in a region. A clinic roster contains 120 eligible people. Investigators can contact 90, 54 consent, 45 provide the required tissue, and 36 specimens yield data that pass the stated quality criteria. Calculate the successive proportions and explain what they describe.
The contact proportion from the roster is \(90/120=0.750\), or 75.0%. Consent among those contacted is \(54/90=0.600\), or 60.0%. Tissue provision among those who consent is \(45/54=0.8333\), or 83.3%. Data yield among tissue providers is \(36/45=0.800\), or 80.0%. Overall, \(36/120=0.300\), or 30.0% of the roster has analyzable data.
These are stage-specific coverage or yield summaries for this hypothetical process; they do not establish that the 36 analyzed people are representative of all adults with the condition in the region. The roster itself may omit people who do not attend that clinic, and consent or tissue availability may be related to participant characteristics. The target is regional adults, while the feasible frame described here is a clinic roster, so generalization beyond the roster requires caution.
Choosing Among Donor-, Specimen-, and Cell-Level Views
The levels are complementary rather than competing definitions of “the sample.” A donor-level view is appropriate for describing how many independent people contribute biological replication. A specimen-level view is needed when material differs within a donor, such as across tissue sites. A cell-level view is essential for describing heterogeneity among measured cells. The mistake is not using cell-level data; it is treating cell-level observations as if they were independent selections from the donor population.
| Approach to describing the data | Useful for | Does not by itself solve |
|---|---|---|
| Count donors | Summarizing independent biological participation and group coverage | Variation across specimens or cells within donors |
| Count specimens | Showing tissue or site coverage and within-donor sampling | Dependence among specimens from the same donor |
| Count cells | Showing cellular resolution and measurement yield | Donor-level replication, selection bias, or cell-recovery bias |
A sampling frame also differs from a statistical model. A model may account for clustering among cells, but it cannot make an unrepresented population appear in the frame. Conversely, a broad donor frame does not ensure valid inference if within-donor dependence is ignored. Design description and analysis must agree about which units were sampled and how they are related.
Common Mistakes and Limitations
- Equating cell count with sample size. “We had 50,000 cells” describes measurement volume, not necessarily independent biological replication. Report donor and specimen counts as well.
- Calling every cell an experimental unit. A cell is not independently assigned to a donor’s disease status or treatment when that status is shared at donor level. State the assignment or sampling level explicitly.
- Calling the recruited cohort the target population. The recruited cohort is observed; the target population is the population the study aims to inform. Explain how the recruitment frame relates to that target.
- Ignoring losses after recruitment. Tissue availability, cell recovery, and quality control can change which donors or cells are represented. Report these transitions rather than only the final cell count.
- Assuming equal representation from equal capture counts. Equal numbers of captured cells per donor do not guarantee equal representation of the underlying tissues or cell types, because recovery may be selective.
Sampling-frame descriptions have limits: they can reveal gaps and support careful interpretation, but they cannot remove selection bias by themselves. A clinic-based frame may not cover people outside the clinic system; a tissue-based frame excludes other tissues; and a captured-cell frame reflects both biology and processing. Conclusions should be bounded by the units and populations the design actually reaches.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A study recruits 8 donors, collects 2 biopsies per donor, and captures 150 cells per biopsy. How many donors, specimens, and cells are represented? Which count indicates independent donor-level replication?
- In a randomized study, a treatment is assigned to each donor, and 500 cells are measured per donor. What is the experimental unit, and why are the cells not 500 independent treatment assignments?
- A study aims to represent residents of a state but recruits only from one specialist hospital. What is the difference between its target population and its feasible donor frame?
- Why can the number of cells passing quality control differ from the number of cells in the tissue before processing?
- Suppose 40 eligible donors are on a frame, 30 are contacted, and 18 enroll. Calculate the contact proportion from the frame and the enrollment proportion among those contacted.
Solutions. (1) There are 8 donors, \(8 \times 2=16\) specimens, and \(8 \times 2 \times 150=2{,}400\) captured cells. The donor-level replication count is 8. (2) The donor is the experimental unit because treatment is assigned to donors. Cells within a donor share the same assignment and biological context, so 500 measured cells do not constitute 500 independent assignments. (3) The target is state residents meeting the stated eligibility criteria; the feasible frame is eligible people reachable through that hospital. The frame may omit residents who do not use that hospital. (4) Dissociation, capture, sequencing, and quality-control filters can remove cells, and recovery can vary across cell types or specimens. (5) Contact proportion is \(30/40=0.750\), or 75.0%; enrollment among contacted people is \(18/30=0.600\), or 60.0%.