Why Pair Conditions Within Donors?
Donors can differ substantially in baseline cell composition, gene expression, and specimen quality. When the scientific question concerns a contrast between two conditions, measuring both conditions for each donor can make that contrast less sensitive to those baseline differences. In a single-cell study, the paired observations are typically specimens or aliquots from the same donor, not individual cells treated as independent replicates.
The design must begin with the intended comparison. For example, the question may concern the average change in a donor-level response after an ex vivo treatment, or the average difference between two tissue compartments sampled from the same donors. Those are different scientific targets. Pairing describes how observations are linked; it does not by itself establish that the conditions are interchangeable, that treatment was randomized, or that a contrast is causal.
A donor-level outcome must be defined in advance. It might be a summary of expression for a specified cell type, a cell-type proportion, or another prespecified response. The design should specify how that outcome will be obtained from each specimen and what happens if a specimen lacks the cell type or fails quality control. A large number of captured cells cannot compensate for having very few independent donors.
Designing Collection and Allocation
The strongest practical opportunity for pairing often occurs when material from one donor can be divided before the experimental conditions are applied. For example, a fresh specimen may be homogenized or mixed according to a validated protocol, then divided into aliquots. The aliquots can be assigned to conditions so that each donor supplies both conditions while the allocation within donor is randomized. This makes the condition contrast less likely to be entangled with which portion happened to receive a condition.
That plan is only suitable if splitting the material preserves the biological quantity of interest. If cells are unevenly distributed, fragile, or affected by the processing needed to divide them, aliquots may not be comparable. Pilot work can assess yield and variability. Collection protocols should also specify minimum material per aliquot, labeling and chain of identity, handling times, storage, and how a failed or contaminated specimen will be recorded.
When the two conditions are distinct tissues or naturally occurring states, a donor may contribute one specimen of each type, but the samples may not be exchangeable apart from the named condition. For example, tissue compartments differ in their local environment as well as in cell composition. The estimand should then be described as the within-donor difference between those defined specimens—not as the effect of changing one tissue into the other.
Where treatment can be assigned, randomize condition allocation within each donor when feasible. Keep the pair identifiable through collection, processing, library preparation, and analysis. If paired samples are processed together, use a consistent protocol and avoid introducing systematic differences in handling time. If they must be processed separately, plan how processing order and batches will be distributed across conditions rather than allowing one condition to be processed only in one way.
Worked Example: Splitting a Specimen Across Two Conditions
A hypothetical study will compare untreated and stimulated immune-cell aliquots from each of 10 donors. The target is the average within-donor change in a prespecified response score after stimulation. Each donor provides enough fresh material for two aliquots, and the same assay protocol will be used for both.
Collection and allocation. For each donor, the team mixes the prepared cell suspension using a validated procedure, divides it into two equal aliquots, and randomly assigns one aliquot to stimulation and the other to the untreated condition. Both aliquots receive the same planned culture duration and downstream handling, except for the intended stimulus.
Pairing and unit. Each donor contributes one stimulated and one untreated specimen, hence one paired contrast. The study has 10 donor pairs, not 20 independent donors and not a number of replicates equal to its cell count.
Feasibility checks. The plan is credible only if the available yield is sufficient for both aliquots and pilot work supports reasonably comparable aliquots. If stimulation can spread to the untreated aliquot through shared media or handling, the pair is exposed to interference and the allocation protocol must prevent it. The result targets the average within-donor response among donors represented by the study; it does not automatically generalize to every donor population.
Why Within-Donor Contrasts Can Be More Precise
Pairing can improve precision when donors with high outcomes under one condition also tend to have high outcomes under the other. The reason follows directly from the variance of a difference. Suppose the donor-level outcomes under the two conditions have variances \(\sigma_A^2\) and \(\sigma_B^2\), and their within-donor correlation is \(\rho\). Then
The covariance term is the key. With positive within-donor correlation, shared donor characteristics tend to move both outcomes in the same direction and partly cancel in the difference. Pairing does not guarantee a precision gain: if the correlation is small, the gain may be slight; if it is negative, the variance of differences can be larger than a comparison that ignores the matching. The actual outcome and specimen process determine the correlation.
Worked Example: Calculating Paired Differences
Suppose four hypothetical donors have the following values for a prespecified response score, measured in score units:
| Donor | Condition A | Condition B | \(D_i=A-B\) |
|---|---|---|---|
| 1 | 12 | 10 | 2 |
| 2 | 15 | 12 | 3 |
| 3 | 13 | 12 | 1 |
| 4 | 16 | 13 | 3 |
The four paired differences are \(2,3,1,3\) score units. Their mean is
The sample variance of the differences is
Thus the standard error of the mean paired difference, using these four pairs, is
These calculations summarize the matched contrasts; they do not show that the design is representative or that the conditions were assigned without bias. Four hypothetical pairs are also too few to establish that the observed pattern will hold in a broader donor population.
The derivation above assumes that each donor contributes a meaningful outcome under both conditions and that contrasts from different donors can be treated as independent for the intended analysis. A causal interpretation additionally requires a defensible treatment allocation and no relevant interference or carryover between conditions. For naturally occurring states, the paired difference is a within-donor association unless further design and assumptions support a causal claim.
When Pairing Is Infeasible or Vulnerable
A paired plan is not always possible or preferable. It may be infeasible if each donor can supply only one condition, if the required tissue is available in only one state, or if the sample is too small to divide without compromising both aliquots. In those cases, attempting to create nominal pairs by matching different donors does not make the observations genuinely paired; it creates a different design that needs its own justification.
Pairing can also fail in less obvious ways. A donor may have one specimen that fails quality control, leaving no usable contrast. If failures are related to condition or donor characteristics, the complete pairs that remain may not represent the intended target population. A plan should state whether a pair is excluded when either member fails, how partial data will be documented, and what population the resulting complete-pair analysis describes.
Order effects are another risk. If the first specimen collected or processed systematically receives one condition, a difference attributed to condition may instead reflect collection time, delay, operator learning, or instrument drift. Randomizing which condition is handled first within donor, balancing order across donors, and recording timestamps can help expose and reduce this risk. If the conditions themselves are administered sequentially to the same material, residual effects from the first condition may carry into the second. Separate, appropriately controlled aliquots may be needed; random order alone cannot remove biological carryover.
Worked Example: Checking an Order-Effect Plan
A hypothetical laboratory has six donor pairs and can process only one aliquot at a time. The team plans to process condition A first for every donor and condition B second. This makes condition and processing order perfectly aligned: any systematic difference due to later processing is indistinguishable from a condition difference.
A simple improvement is to assign the order within each pair and balance the two orders across donors. With six donors, three can be assigned A then B and three B then A. The assignment should be randomized among donors, and the schedule should be recorded before processing.
This allocation prevents a consistent order from being attached to only one condition, but it does not guarantee that all order effects vanish. The two orders could still have different handling delays or other protocol deviations. The team should record actual timing and deviations, and should reconsider the design if the biological treatment has carryover that cannot be controlled by using separate aliquots. If each donor cannot provide two suitable aliquots, the proposed paired comparison may not be feasible.
Choosing Between Paired and Independent Sampling
A paired design is most useful when both conditions can be measured meaningfully for the same donor, the pair can be collected and processed without introducing a systematic condition difference, and within-donor outcomes are expected to share donor-level variation. An independent-group design is more appropriate when only one condition can be observed per donor or when applying both conditions to one donor would create unacceptable interference. It avoids requiring two specimens per donor but does not remove baseline donor variation through within-donor subtraction.
Blocking on donor characteristics can improve comparability when the same donor cannot contribute both conditions, but it is not the same as a paired design. In a pair, each donor supplies both condition observations and a direct within-donor contrast. In blocking, different donors in the same block supply different conditions; the design controls a recorded grouping variable but does not create a within-donor match. The choice depends on what specimens can be collected and on the scientific comparison, not merely on which design appears more efficient.
Before collection, document the target contrast, the donor as the independent biological unit, the required specimens per donor, the allocation mechanism, the processing-order plan, and the rules for incomplete pairs. Confirm that the planned outcome is measurable under both conditions. These choices help distinguish a genuinely paired study from a study that only happens to contain multiple samples from the same person.
Check Your Understanding
Answer each question, then compare with the solutions.
- A study splits each donor’s cell suspension into two aliquots and randomly assigns one to each of two conditions. What is the paired unit, and how many pairs are there if 14 donors contribute usable aliquots?
- Why can positive within-donor correlation reduce the variance of a paired difference?
- All condition A specimens are processed first and all condition B specimens second. What problem does this create, and name one design remedy.
- Each donor contributes a specimen from two different tissue compartments. Can their difference automatically be interpreted as the causal effect of changing tissue type? Explain.
Solutions. (1) The donor contributes one matched pair of aliquots, so there are 14 pairs. (2) The variance of the difference includes \(-2\operatorname{Cov}(Y_{iA},Y_{iB})\); positive covariance reduces it because shared donor-level variation tends to cancel. (3) Condition is confounded with processing order, so an order or handling effect could appear to be a condition effect. Randomize order within donor and balance the two sequences across donors, while recording actual processing times. (4) No. The contrast is between those two defined tissue specimens, which may differ in more than the named condition; it does not by itself identify a causal effect.