Randomizing Conditions Across Donors

Why Randomize Conditions Across Donors?

In a single-cell study, treatment or condition is often assigned to a donor, while many cells are measured from that donor. The donor is therefore the relevant unit for random assignment when the condition is applied to the donor. Assigning thousands of cells to different conditions does not replace assigning donors: cells from one donor share biological history and do not constitute thousands of independently randomized donors.

Random assignment is a design method for making treatment groups comparable, on average, with respect to both measured and unmeasured donor characteristics. It does not promise that a particular small study will have similar groups. By contrast, whether the enrolled donors represent a wider population depends on recruitment and sampling, not on random assignment. These are separate design questions.

A causal interpretation requires a clearly defined intervention and condition, valid assignment at the donor level, and outcomes that correspond to the assigned conditions. In potential-outcomes terms, let \(Y_i(1)\) and \(Y_i(0)\) denote the outcome that enrolled donor \(i\) would have under A and B, respectively. Assume consistency: a donor assigned to a condition has the outcome defined for that condition. Also assume no interference between donors: one donor’s assignment does not change another donor’s outcome. The assignment mechanism must actually be random as specified, and outcomes must be observed or missingness handled without undermining the intended comparison. Under these conditions, randomization supports a causal comparison for the enrolled donors. It does not, by itself, make those donors representative of a broader population.

What Randomization Guarantees—and What It Does Not

Write \(Z_i=1\) if donor \(i\) is assigned to A and \(Z_i=0\) if assigned to B. With complete randomization, every subset of \(m\) donors is equally likely to form group A. The probability that any specified donor is assigned to A is \(m/n\). The crucial guarantee is about the assignment process: before the draw, donor characteristics cannot systematically determine the assigned condition if the process is implemented correctly. In expectation over all possible assignments, group averages of fixed baseline characteristics are equal. A realized allocation may nevertheless have a noticeable imbalance, especially when \(n\) is small.

The same reasoning applies to potential outcomes. The difference in observed group means is an unbiased estimator, over the randomization distribution, of the average treatment effect among the enrolled donors. This result describes repeated assignment of the same enrolled donors under the specified randomization; it does not mean every realized difference is unbiased in a particular sample or that chance imbalance is impossible.

A Short Justification

Under complete randomization, each donor has probability \(m/n\) of entering A. For fixed donor outcomes, the observed A mean can be written as \(\sum_i Z_iY_i(1)/m\). Taking its expectation over assignments gives the mean of the donors’ A potential outcomes:

Likewise, because each donor has probability \((n-m)/n\) of entering B, the expected B mean is \(n^{-1}\sum_iY_i(0)\). Subtracting these expectations shows that the expected difference in observed means is \(n^{-1}\sum_i[Y_i(1)-Y_i(0)]\), the average treatment effect for these enrolled donors. This is a design-based result: it relies on the assignment probabilities, not on a model asserting that donors are identical or that their outcomes follow a particular distribution.

Chance Imbalance in a Small Study

An important feature of randomization is that it prevents systematic assignment based on donor characteristics, not that it forces exact balance. To assess a design, distinguish a baseline difference that occurred by chance from a difference created by a compromised allocation process. A baseline imbalance alone does not show that randomization failed. Conversely, a process can be technically random yet still produce an unlucky imbalance that reduces precision or makes interpretation harder.

Concealing the Allocation

Random sequence generation is not enough if the person enrolling donors can see or predict the next assignment. Suppose a recruiter knows the upcoming condition and can delay enrollment, alter eligibility judgments, or choose which eligible donor enters next. Enrollment decisions may then become associated with treatment. In that case, the assignment sequence may have been generated randomly, but the enrolled donors’ allocation is no longer protected from selection bias.

A sound sequence is generated in advance or through a secure central system, and upcoming assignments remain inaccessible to those making enrollment decisions. Eligibility and consent should be established, and enrollment recorded, before the assignment is revealed. The sequence and any deviations should be retained for audit. Concealment does not hide the condition from laboratory staff after assignment; that is the separate aim of blinding, which may or may not be feasible.

Practical Checks, Alternatives, and Limits

Before assigning donors, specify the eligible population, the condition being assigned, the assignment unit, the number allocated to each condition, and the randomization procedure. Keep a record of sequence generation, enrollment order, assignment times, and deviations. Afterward, inspect group characteristics descriptively to understand what was realized, but do not treat a nonsignificant baseline test as proof that groups are equivalent. In a small study, such tests have little ability to detect meaningful imbalances; design records provide more direct evidence about whether assignment was implemented correctly.

Random assignment is preferable to predictable rules such as alternation, assignment by enrollment date, or assignment by recruiter judgment. Those methods can be anticipated and can align donor characteristics with condition. Random assignment also differs from statistical adjustment: adjustment models observed covariates, while randomization uses a chance mechanism to prevent systematic allocation from depending on donor attributes, including attributes not measured. Adjustment can sometimes improve precision or address observed imbalances, but it cannot repair every problem created by selection into treatment.

Other design strategies can improve comparability, but they answer additional design questions. For example, blocking on known characteristics assigns donors within pre-defined groups; its specific construction is a distinct approach. No assignment method repairs a narrow or biased recruitment frame, guarantees adherence, prevents differential outcome measurement, or eliminates loss of donors after assignment. In single-cell experiments, condition assignment should also be distinguished from downstream processing choices: if condition and processing batch align, an effect may be difficult or impossible to separate. Randomization must be planned for each relevant stage rather than assumed to solve every source of bias.