Why Pair Donors?
When the same donors contribute observations under two conditions, the comparison can be made within each donor. This removes differences between donors that affect both condition summaries, such as a donor’s generally high or low expression level. The quantity of interest is not a comparison of two unrelated sets of cells: it is the set of donor-specific differences and their average.
For a chosen gene or other outcome, let \(Y_{iA}\) and \(Y_{iB}\) be comparable donor-level summaries for donor \(i\) in conditions A and B. The summaries might, for example, be computed from cells of the same cell type in each condition. Define the contrast in the stated direction, B minus A. Each donor with both summaries supplies one contrast. The biological sample size for this analysis is the number of donors with complete pairs, not the number of cells.
Pairing describes the matching of the two observations by donor. It does not mean that a cell measured in condition A is paired with a corresponding cell in condition B. Cells from the same donor are clustered observations; they do not become independent biological replicates merely because they are numerous. As covered in the earlier material on donor-level summaries, the summaries must also be defined consistently across donors and conditions.
Estimating the Mean Contrast and Its Uncertainty
Let \(d_1,\ldots,d_n\) denote the donor contrasts. Their sample standard deviation, \(s_d\), measures how much the contrasts vary across donors. The estimated standard error of their mean is \(SE(\bar{d})=s_d/\sqrt{n}\). For a two-sided test of no average contrast, the hypotheses are \(H_0:\mu_d=0\) and \(H_a:\mu_d\ne0\), where \(\mu_d\) is the population mean of the donor-level B-minus-A contrasts.
The paired t procedure uses \(t=\bar{d}/(s_d/\sqrt{n})\), with \(df=n-1\). A confidence interval for \(\mu_d\) is \(\bar{d}\pm t^*_{n-1,\,1-\alpha/2}s_d/\sqrt{n}\), where \(t^*_{n-1,\,1-\alpha/2}\) is the corresponding t-distribution critical value. The interval and test concern the mean of the donor contrasts, not the mean of all individual cells.
Normality is an assumption about the differences, not a requirement that the two condition-specific outcomes separately be normally distributed. With few donors, a histogram or dot plot of the contrasts can reveal an extreme value or marked asymmetry, but diagnostics have limited power. A normality test that does not reject is not proof that the assumption holds. If one or two donors dominate the result, report that sensitivity and consider whether the t-based summary is adequate.
Worked Example: Mean Contrast, Test, and Confidence Interval
Suppose a hypothetical study measures the same gene in the same cell type for six donors under conditions A and B. The outcome is a donor-level summary on a log2 expression scale. The six B-minus-A contrasts, in log2 units, are \(1.2,\ 0.8,\ 1.5,\ -0.1,\ 1.0,\ 0.6\). Assume for this calculation that donors are independent and that the population distribution of contrasts is reasonably normal.
The sum of the contrasts is \(5.0\), so \(\bar{d}=5.0/6=0.8333\) log2 units. The squared deviations from the mean sum to approximately \(1.5333\), giving \(s_d=\sqrt{1.5333/(6-1)}=0.5538\) log2 units. Thus \(SE(\bar{d})=0.5538/\sqrt{6}=0.2261\) log2 units.
For \(H_0:\mu_d=0\) against \(H_a:\mu_d\ne0\), the test statistic is \(t=0.8333/0.2261=3.685\), with \(df=5\). The two-sided p-value is approximately \(0.014\). At significance level \(0.05\), this is evidence against a zero mean donor contrast, subject to the assumptions and study design.
The 95% confidence interval uses \(t^*_{5,0.975}\approx2.571\). Its margin of error is \(2.571(0.2261)=0.5813\) log2 units, so the interval is \(0.8333\pm0.5813=(0.252,\ 1.415)\) log2 units. In this hypothetical sample, the estimated mean B-minus-A contrast is positive. The interval gives values compatible with the data under the model; it does not mean there is a 95% probability that this fixed population mean lies in this particular interval.
Why Pairing Can Improve Precision
Suppose the two condition outcomes have variances \(\sigma_A^2\) and \(\sigma_B^2\), and their within-donor covariance is \(\operatorname{Cov}(Y_A,Y_B)\). Since \(d=Y_B-Y_A\), the variance of a donor contrast is
This identity follows by expanding the variance of a difference: \(\operatorname{Var}(Y_B-Y_A)=\operatorname{Var}(Y_B)+\operatorname{Var}(Y_A)-2\operatorname{Cov}(Y_A,Y_B)\). If donors who tend to have high summaries in A also tend to have high summaries in B, the covariance is positive and subtracting it can make the variance of the differences smaller. The paired standard error then reflects the observed variability of those differences, rather than variability between unrelated donor groups.
Pairing is not guaranteed to reduce uncertainty: if the within-donor covariance is small or negative, the differences may remain variable. Nor does matching create additional independent donors. The precision calculation still uses \(n\) donor contrasts, with \(n-1\) degrees of freedom for the paired t procedure.
Worked Example: Donor Pairing Is Not Cell Pairing
Suppose three hypothetical donors have the following donor-level mean expression summaries. The cell counts differ by donor and condition; the summaries, not the individual cells, are the paired observations.
| Donor | Condition A: mean, cells | Condition B: mean, cells | B minus A |
|---|---|---|---|
| 1 | 2.0, 20 | 3.0, 40 | 1.0 |
| 2 | 4.0, 100 | 3.5, 10 | -0.5 |
| 3 | 1.0, 5 | 2.0, 5 | 1.0 |
The paired donor estimate is the equally weighted mean of the three contrasts: \(\bar{d}=(1.0-0.5+1.0)/3=0.500\) expression units per donor. There are three donor pairs, not 180 paired cells. The cell counts do not change the number of independent biological replicates.
For comparison, pooling the cell-level means with cell-count weights gives an A mean of \((20(2.0)+100(4.0)+5(1.0))/125=445/125=3.56\), and a B mean of \((40(3.0)+10(3.5)+5(2.0))/55=165/55=3.00\). Their difference is \(-0.56\), opposite in sign to the equally weighted donor contrast. These are different summaries with different weighting: the pooled calculation gives donors influence according to their cell counts. It is not a substitute for the paired donor estimand when the target is the average donor-level contrast.
Choosing the Comparison and Checking the Analysis
A paired analysis is appropriate when the same donors contribute both condition summaries and the scientific target is a within-donor contrast. If different donors contribute to A and B, the observations are not paired; an unpaired donor-level comparison is the relevant alternative. Treating truly matched donor summaries as unrelated discards their matching and may lose precision. Treating unrelated donors as paired by row order invents a match and gives an invalid analysis.
A donor-restricted permutation procedure is another possible approach when its exchangeability conditions are justified. It uses the within-donor structure rather than permuting cells across donors. Such a procedure and the paired t procedure need not target the same assumptions: the t procedure relies on the distribution of donor differences for its small-sample reference distribution, whereas a permutation procedure requires a valid randomization or exchangeability argument. The choice should follow the design and target, not which method gives a smaller p-value.
Use a consistent direction for the contrast and state it explicitly. Reversing B-minus-A to A-minus-B changes the sign of every contrast, the mean, and both interval endpoints, but not the two-sided p-value. Pairing also requires both condition summaries for each included donor. A donor missing one condition cannot contribute a within-donor difference; report how many complete pairs remain and consider whether missingness could select an unrepresentative subset. Do not fill in a missing value by pairing that donor with another donor.
Before interpreting the result, check the donor identifiers, condition labels, cell-type definition, summary scale, and contrast direction. Plot the individual donor contrasts so that their spread and any influential donors are visible. Confirm that a donor has not been counted more than once as an independent pair and that the number of pairs used in the calculation is stated. In single-cell studies, these checks help prevent a large cell count from disguising a small donor count.
Worked Example: Reversing the Contrast and Interpreting a Wide Interval
In a separate hypothetical analysis, four independent donors have B-minus-A contrasts of \(0.2,\ 0.4,\ -0.1,\ 0.3\) expression units. Assume the differences are drawn from a normal population for this illustration. Their mean is \(\bar{d}=(0.2+0.4-0.1+0.3)/4=0.200\). The squared deviations sum to \(0.140\), so \(s_d=\sqrt{0.140/3}=0.2160\) and \(SE=0.2160/\sqrt{4}=0.1080\) units.
The test statistic for a zero mean is \(t=0.200/0.1080=1.852\), with \(df=3\), giving a two-sided p-value of approximately \(0.161\). The 95% critical value is \(t^*_{3,0.975}=3.182\), so the interval is \(0.200\pm3.182(0.1080)=(-0.144,\ 0.544)\) units. This interval includes zero, and the test does not provide evidence against a zero mean at the 0.05 level. It does not establish that the mean contrast is zero; with only four donors, the interval is wide.
If the contrast had instead been defined as A minus B, the mean would be \(-0.200\) and the interval would be \((-0.544,\ 0.144)\) units. The two-sided p-value would be unchanged. The direction should therefore be specified before reporting results, and the uncertainty should be interpreted in the stated units.
Common Mistakes and Limitations
- Pairing cells instead of donors. The incorrect reasoning is that many cells in both conditions create many pairs. The correction is to form one defined contrast for each matched donor; cell-level pairing requires a genuine, meaningful cell match and does not follow from sharing a donor.
- Using cell count as the t-test sample size. This treats clustered cells as independent replicates and understates uncertainty. The paired t calculation uses the number of independent donor contrasts.
- Ignoring the difference distribution. A t procedure is not justified merely because both condition means look plausible. Inspect the donor contrasts, particularly when \(n\) is small, and acknowledge that severe skew or influential donors can undermine the approximation.
- Interpreting a nonsignificant result as no effect. A large p-value is not evidence that the null hypothesis is true. Report the estimate and interval; they show which contrast values are compatible with the data and the model.
- Assuming pairing solves every design problem. Matching controls shared donor differences, but it cannot remove confounding that varies with condition, correct inconsistent summaries, or make a nonrepresentative donor sample representative.
A paired donor contrast is a focused analysis of an average within-donor difference. Its strength is that it respects the biological replication and the matching in the design. Its main limitations are equally important: the inference is supported by the number and quality of donors, the donor differences must be adequately modeled, and conclusions apply to the chosen summaries and target population—not automatically to every cell or every biological context.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For donor \(i\), condition A has summary 2.4 and condition B has summary 3.1. What is the B-minus-A contrast, and what does its sign mean?
- Eight donors each contribute both condition summaries, with 70 cells per donor in A and 120 in B. How many paired observations enter a donor-level paired t analysis?
- State the standard error and degrees of freedom used for inference from \(n\) complete donor contrasts.
- Why can positive within-donor correlation reduce the variance of the paired differences?
- A 95% confidence interval for the mean contrast includes zero. Does that prove the mean contrast is zero? Explain.
Solutions. (1) \(d_i=3.1-2.4=0.7\) summary units; B is higher for this donor. (2) There are eight paired observations, one contrast per donor; the cell counts do not increase the donor-level sample size. (3) \(SE=s_d/\sqrt{n}\) and \(df=n-1\), under the paired t assumptions. (4) The difference variance is \(\sigma_A^2+\sigma_B^2-2\operatorname{Cov}(Y_A,Y_B)\); positive covariance lowers it relative to the same marginal variances with zero covariance. (5) No. The interval indicates that zero is compatible with the data at the stated confidence level, but it may also include nonzero contrasts; it does not prove the null value.