Why the Unit of Inference Matters
A single-cell experiment can contain thousands of measured cells and still have only a handful of independent treatment replicates. The key question is not simply how many rows appear in the expression matrix. It is what unit received the treatment independently, and what units provide independent evidence about its effect? When treatment is assigned to donors, cells from the same donor share that assignment and many biological characteristics. Counting each cell as a separate treatment replicate can therefore suggest far more independent evidence than the design provides.
The introductory material on cells nested within donors establishes that cells from one donor form a cluster rather than a collection of unrelated biological replicates. Here the practical task is to use the study design to choose the unit for inference. That choice precedes selecting a statistical procedure: a sophisticated model cannot make a treatment independently assigned to three donors into a treatment independently assigned to thousands of cells.
Use the Assignment Mechanism to Choose
Suppose donor \(i\) is assigned condition \(T_i\), and the measured outcome for cell \(j\) from that donor is \(Y_{ij}\). If all cells from a donor receive the same donor-level assignment, then the treatment labels vary across donors, not independently across cells. The cells provide information about outcomes within each donor, but they do not create additional independent treatment assignments.
The donor is the natural independent unit when the study is designed so that donors are independently sampled or assigned, each donor contributes to one treatment condition, and the analysis concerns differences between those conditions. This reasoning assumes that donors do not influence one another's outcomes, that treatment is applied as specified, and that the observed donors and their cells are suitable for the intended scientific comparison. For a randomized experiment, the assignment procedure must be followed as designed; for an observational comparison, differences between donor groups may reflect factors other than treatment.
A useful justification comes directly from the randomization. If a study assigns three of six donors to treatment and the remaining three to control, the possible treatment allocations are choices of three donors from six. There are \(\binom{6}{3}=20\) such allocations. Recording 100 cells per donor does not change the allocation: each donor's cells carry that donor's one assignment. Thus the number of independently randomized units is six, not 600. The cell measurements may help characterize each donor's response, but they do not enlarge the set of treatment assignments.
Worked Example: Counting Independent Treatment Replicates
In a hypothetical experiment, investigators randomly assign six donors, three to a stimulation condition and three to control. They sequence 80 cells from every donor. Identify the treatment-assignment unit, measurement unit, and independent replication for the treatment comparison.
The treatment is assigned to each donor, so the donor is the treatment-assignment unit. Expression is recorded for each cell, so the cell is the measurement unit. There are \(6\times80=480\) measured cells, but only six independently assigned donors: three in each condition.
Under complete random assignment of three of six donors to stimulation, the number of allowable assignments is \( \binom{6}{3}=6!/(3!3!)=720/36=20 \). The assignment mechanism has 20 possible donor-level allocations; it does not have 480 independently assigned cell labels. For a treatment comparison across donors, the independent replication is therefore three donors per condition. The 80 cells per donor can improve the description of each donor's measured cell population, subject to representative cell sampling, but they do not turn the design into 240 treatment replicates per group.
This conclusion assumes the stated donor-level randomization was actually carried out, that there is no interference between donors, and that the sequencing and outcome measurements are comparable across conditions. Randomization supports a causal treatment comparison under these design conditions; it does not by itself establish that the six donors represent every population to which investigators might wish to generalize.
From Cell Measurements to Donor-Level Evidence
For a donor-level treatment question, a common conceptual route is to summarize the cells within each donor and compare those donor-level outcomes across conditions. The summary might be an average expression value, a cell-type proportion, or another outcome defined by the scientific question. The relevant comparison then has one outcome per donor, rather than treating all cells as independent replicates. The exact choice of outcome summary and the population quantity it targets require additional specification; the central point here is the identity of the independent units.
Before interpreting such a comparison, check that each donor's outcome was constructed comparably, the treatment labels are assigned at the donor level as claimed, and there are multiple independent donors in the conditions being compared. If a donor contributes an unusually small or selective set of cells, its summary may be less representative. That is a measurement and sampling concern; it does not change the assignment unit.
Worked Example: Comparing Donor Summaries Without Rounding Error
Suppose a hypothetical experiment measures a gene in six donors, with three donors per condition. After calculating a mean log-expression value across the sampled cells for each donor, the control donor values are 0.9, 1.2, and 1.0 log-expression units. The treated donor values are 2.1, 1.7, and 2.4 log-expression units. Find the difference in the two group means and identify the replication for inference.
The donor-level control mean is \( (0.9+1.2+1.0)/3=3.1/3=1.033333\ldots \) log-expression units. The treated mean is \( (2.1+1.7+2.4)/3=6.2/3=2.066666\ldots \) log-expression units. Calculating the difference before rounding gives \( (6.2/3)-(3.1/3)=3.1/3=1.033333\ldots \), or approximately 1.033 log-expression units.
If the displayed group means are rounded first to 2.067 and 1.033, their subtraction is 1.034. That is a subtraction of rounded displays, not the correctly rounded difference calculated from the underlying donor values. The difference should therefore be reported as approximately 1.033 log-expression units (or with sufficient digits to make the rounding explicit).
The six donor summaries, not the individual cells used to form them, provide the independent donor-level replication: three donors per condition. This arithmetic describes a difference in observed group means; by itself it does not establish statistical significance or rule out alternative explanations. Inference would also depend on the study design and assumptions, including independent donors and appropriate treatment assignment.
When the Answer Is Not Simply “Donor”
The unit of inference depends on the question and the assignment structure; it is not a rule that every single-cell analysis must use donors in exactly the same way. If the scientific goal is to describe the distribution of cell states within one donor, cells are the measurement units of interest. They can inform that donor's observed cell population, but they do not provide independent replication for a claim about a treatment effect across donors.
Likewise, if each donor is split into samples that receive different conditions, treatment may be assigned to samples or aliquots within donor rather than to whole donors. The analysis then needs to respect the within-donor comparison, while donors still supply independent replication across biological individuals. The assignment structure should be described precisely instead of applying a blanket “one cell, one replicate” or “one donor, one replicate” rule.
A hierarchical or mixed-effects model can represent both donor-level and cell-level variation, and may use cell observations without pretending that cells from the same donor are independent. It does not create additional independent donors. A donor-level summary is often easier to interpret for a donor-level question; a model using cell-level data may be useful for other goals or data structures. Neither choice repairs a design in which treatment and donor identity are inseparable.
| Scientific task | Relevant unit or structure | Key caution |
|---|---|---|
| Compare conditions assigned to different donors | Donors provide independent treatment replication; cells are measured within donors | Do not count cells as independent assignments |
| Describe the cells observed in one donor | Cells are the measurement units for the within-donor description | Do not claim donor-population treatment replication from one donor |
| Compare conditions applied to aliquots within donors | Condition assignment occurs at the aliquot level; donors provide independent biological blocks | Preserve the within-donor pairing in the analysis |
| Fit a model to cell-level outcomes | Represent the donor grouping and the assignment structure | A model cannot manufacture independent donor replication |
Diagnostics and Common Failure Modes
Several design checks help reveal whether the stated inference unit is plausible. First, make a donor-by-condition table: if condition varies only between donors, the treatment contrast is donor-level. Second, check how many distinct donors occur in each condition, not just how many cells. Third, inspect whether batch, processing date, or another factor is aligned with condition. If all treated donors were processed in one batch and all control donors in another, treatment and batch may be confounded; many cells per donor do not resolve that problem.
Also check whether the sampled cells plausibly represent the cell population to which the outcome refers. Strongly selective cell recovery, low cell counts for some donors, or condition-dependent measurement quality can affect donor summaries. These issues can weaken measurement or change what the observed data represent. They do not make cells independent treatment replicates.
- Counting cells as treatment replicates. “We have 4,000 cells, so we have 4,000 independent replicates” is wrong when treatment was assigned to donors. The correction is to identify the assignment mechanism first; for donor-level assignment, donors provide the independent treatment replication.
- Assuming more cells fix too few donors. More cells may better characterize a donor, but they do not replace additional independent donors for a donor-level treatment comparison.
- Ignoring treatment–batch alignment. If condition and batch always coincide, a difference cannot be cleanly attributed to treatment from that design alone.
- Calling a donor summary a population result. A summary for observed donors does not automatically represent all donors. Generalization depends on how donors were sampled and on the target population.
The principal limitation is that identifying the correct unit does not itself select an estimand, estimator, or uncertainty procedure. Those depend on the exact scientific question and design. The decision here is more basic but essential: do not base treatment-level replication on observations that did not receive independent treatment assignments.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A study assigns four donors to treatment and four to control, then sequences 50 cells per donor. How many independent donor-level treatment replicates are in each condition?
- Why can a study with 2,000 cells still have only six independent units for a treatment comparison?
- In a hypothetical dataset, the control donor summaries are 1.0 and 1.4, and the treated summaries are 2.0 and 2.6. Calculate the difference in group means using unrounded values.
- All treated donors are processed in batch A and all control donors in batch B. What concern should be raised, and would adding cells to each donor solve it?
- A donor is split into two aliquots, one assigned to each condition. Is this the same assignment structure as assigning whole donors to conditions?
Solutions. (1) There are four independent donor-level replicates in each condition. The 50 cells per donor yield 400 measured cells overall, but do not alter the four-per-condition donor replication. (2) If treatment was assigned to six donors, the treatment labels vary across six independent assignment units; the 2,000 cells inherit those donor labels and are measurements nested within them. (3) The control mean is \((1.0+1.4)/2=1.2\), and the treated mean is \((2.0+2.6)/2=2.3\). The treated-minus-control difference is \(2.3-1.2=1.1\) in the units of the donor summaries. There are two donors per group, regardless of how many cells contributed to each summary. (4) Treatment is confounded with processing batch, so the observed difference could reflect batch, treatment, or both. More cells within these same donors do not separate the two explanations; the design needs treatment and batch to be distinguishable. (5) No. Treatment is assigned to aliquots within donors, creating a within-donor comparison. The analysis should retain that pairing, with donors supplying independent biological replication across individuals.