Diagnosing Imbalance Across Donors

Why Diagnose Donor Imbalance?

A single-cell data set may contain many cells but very different numbers of cells from different donors. That imbalance matters because a donor with few cells may have a less precisely measured outcome, while a donor with many cells may contribute a large share of a cell-weighted summary. Neither fact means that the donor is less biologically important or that the donor should automatically be excluded.

The aim here is a disciplined diagnostic: describe how many cells contribute for each donor, summarize the donor-level outcomes, and check whether the count distribution is associated with the outcome summaries or the conclusions one might draw from them. Counts are not independent biological replicates. For comparisons about donors or populations, the donor remains the biological replication unit.

Counts must refer to a comparable set of observations. For example, if the question concerns one cell type, use counts for that cell type rather than each donor's total across all cell types. For a condition comparison, record counts separately by donor and condition. A count can be small because of biology, sampling, filtering, or technical quality; the count alone does not identify the cause.

Summarizing the Count Distribution

Start with a table or plot of \(n_j\) by donor. Useful descriptive summaries include the minimum and maximum count, the maximum-to-minimum ratio, each donor's share \(w_j=n_j/\sum_{k=1}^{J}n_k\), and the coefficient of variation \(s_n/\bar{n}\), where \(\bar{n}\) and \(s_n\) are the sample mean and sample standard deviation of the donor counts. These summaries describe different features: a ratio highlights the range, shares show how concentrated the total is, and the coefficient of variation describes spread relative to the mean.

A concentration summary can make the count distribution easier to compare across data sets. Define \(N_{\mathrm{conc}}=1/\sum_j w_j^2\). This is the number of equally contributing donors that would produce the same sum of squared shares. It is a descriptive equivalent number, not the number of biological replicates and not a replacement for \(J\).

No single threshold defines unacceptable imbalance in every study. A maximum-to-minimum ratio can look dramatic when the minimum is very small, while a moderate ratio can conceal one donor contributing a large share when there are many donors. Report the distribution and relevant summaries rather than relying on one cutoff.

Compare Donor Summaries With Cell-Weighted Summaries

An equally weighted donor summary gives each donor the same contribution: \(\bar{y}_{D}=J^{-1}\sum_{j=1}^{J}\bar{y}_j\). A cell-count-weighted summary instead gives donor \(j\) weight \(w_j=n_j/\sum_k n_k\): \(\bar{y}_{C}=\sum_j w_j\bar{y}_j\). These answer different descriptive questions. The first is the average of the observed donor summaries, giving each donor equal weight. The second is the average over the observed cells, so donors with more eligible cells contribute more. The appropriate target depends on the scientific question; neither summary is automatically the correct inferential estimator.

To see this identity, write \(\bar{y}_{C}=\sum_j w_j(\bar{y}_j-\bar{y}_{D})+\bar{y}_{D}\sum_jw_j\). Since the weights sum to one, subtracting \(\bar{y}_{D}\) gives the displayed expression. Thus the difference between the summaries depends on whether donors with above-average counts tend to have above- or below-average outcomes. If all counts are equal, every \(w_j=1/J\), and the summaries agree. A large difference is a useful warning that the weighting choice is consequential; by itself, it does not establish bias or causation.

Counts and the Precision of Donor Outcomes

A low count can make a donor's cell-based outcome summary noisy, but the amount of additional information from more cells depends on dependence among cells from the same donor. For a simple diagnostic, suppose the cell outcomes within a donor have common marginal variance \(\sigma^2\) and common pairwise correlation \(\rho\), with \(0\leq\rho\leq1\). Under this exchangeable-correlation model, the variance of a donor's cell mean based on \(n_j\) cells is \(\sigma^2[\rho+(1-\rho)/n_j]\). This calculation assumes the cells are sampled comparably within donor and that the variance and correlation adequately describe their dependence. It is a precision model for a donor summary, not a model for between-donor variation.

The expression also explains why raw cell counts should not be read as independent sample sizes. When \(\rho>0\), the variance approaches \(\sigma^2\rho\), rather than zero, as \(n_j\) grows. More cells can improve measurement of a donor, but they do not create more donors. The earlier material on intraclass correlation and cell sampling uncertainty develops these dependence and precision ideas further.

A Practical Diagnostic Sequence

Use a consistent set of summaries before fitting or interpreting donor-level comparisons. The following sequence keeps count imbalance, outcome variation, and inferential choices distinct.

A table can be particularly useful when outcomes are condition-specific: include one row per donor-condition combination, with its cell count and outcome summary. This reveals whether an apparently balanced total count hides a sparse donor in one condition. In paired designs, both conditions need usable information for the donor-level contrast; a large total across conditions does not guarantee adequate measurement in each one.

Comparisons, Failure Modes, and Limits

A common mistake is to see one donor with many more cells and conclude that the analysis has many more independent replicates for that donor. The correction is to distinguish cell-level measurement from donor-level replication: extra cells may refine a donor summary, but they do not add independent donors. Another mistake is to equal-weight donors in every setting without asking what the target is. Equal weighting is informative for a donor-average description, while cell weighting describes the observed-cell average; the scientific estimand and analysis model must guide the choice.

Other warning signs include comparing counts from different cell populations, hiding a few low-count donors behind a large total, or interpreting a count-outcome association as evidence that count imbalance caused the outcome pattern. The diagnostic summaries are descriptive. They do not resolve whether count differences reflect biological composition, technical sampling, or design, and they do not replace donor-aware uncertainty estimation. Nor should a large count or an extreme outcome alone be treated as a formal finding of donor influence; that requires a separate influence assessment.

The practical conclusion should be calibrated: state how uneven the counts are, whether the donor summaries appear to depend on that unevenness, and which assumptions limit interpretation. When imbalance is substantial, report it transparently and show how the chosen donor-aware analysis handles unequal measurement. Do not claim that a larger count automatically deserves greater biological weight.