Why Block Donors Before Assigning Conditions?
Donors can differ in characteristics that are related to a single-cell study outcome. Age, for example, may be associated with baseline expression or cell-type abundance. If the donors assigned to one condition happen to be older than those assigned to another, an observed outcome difference may partly reflect this imbalance. Random assignment protects against systematic allocation bias, but in a modest sample it does not guarantee close balance on every characteristic.
Blocking addresses this design problem by grouping donors according to a small number of known, pre-condition characteristics and randomizing conditions separately within each group. It is useful when a characteristic is both measured before assignment and plausibly prognostic for the donor-level outcome. Blocking is not a substitute for randomization: it organizes the randomization so that important characteristics are represented across conditions.
A characteristic used to form blocks must be known before assignment; defining blocks using a response measured after treatment can distort the comparison. The characteristic should also be chosen for a scientific reason, not simply because it is available. If age is expected to predict the outcome, age groups may be useful. If sex is also relevant, combining age and sex categories can create cross-classified blocks, but each added division makes blocks smaller.
Planning the Allocation
Suppose there are two conditions, A and B, and \(n_b\) donors in block \(b\). The allocation plan should specify how many donors in that block receive each condition. To estimate a condition contrast within every block, each condition must receive at least one donor in each block. More donors per condition are preferable when feasible, because a comparison based on one donor per condition is highly sensitive to individual outcomes.
Within a block, assignment should be random among the allocations allowed by the plan. For example, if four donors in a block are to be split two per condition, each of the six possible choices of two donors for A can be given equal probability. Randomization within blocks means condition assignment is not decided by donor characteristics that were used to define the blocks.
Worked Example: Blocking on Age Before Randomization
A hypothetical study will compare conditions A and B using 12 donors. Before assignment, investigators place donors into three prespecified age bands, with four donors in each band. They expect age to be related to the donor-level expression outcome.
Within each age band, the investigators randomly select two of the four donors for condition A and assign the remaining two to condition B. There are \(\binom{4}{2}=6\) equally likely assignments in each band. If choices are made independently across the three bands, there are \(6^3=216\) possible overall assignments.
This plan guarantees two donors per condition in every age band. It does not guarantee that donors are identical within an age band, nor does it balance characteristics that were not used in the design. Its benefit is narrower: it prevents the realized allocation from placing all four donors in a particular age band into one condition.
The allocation is appropriate if the age bands were chosen before seeing outcomes, both conditions can be assigned to donors in every band, and the planned outcome can be measured comparably across conditions. If those conditions do not hold, the intended within-band comparison may not be available.
What Contrast Does Blocking Estimate?
Let \(\bar{Y}_{Ab}\) and \(\bar{Y}_{Bb}\) denote the mean donor-level outcomes in conditions A and B, respectively, within block \(b\). A block-specific contrast is \(\widehat{\tau}_b=\bar{Y}_{Ab}-\bar{Y}_{Bb}\). The overall blocked contrast combines these block-specific comparisons with prespecified weights \(w_b\), where the weights sum to one.
The weights determine which average difference is being summarized. Weights \(w_b=n_b/N\), where \(N\) is the total number of donors, target an average across the enrolled donor distribution when block-specific effects differ. Equal weights instead give each block equal influence, whether or not blocks contain the same number of donors. The weighting rule should match the target comparison rather than being chosen after looking at results.
A simple model helps show how blocking can affect precision. Let \(T_i=1\) for condition A and \(T_i=0\) for condition B for donor \(i\) in block \(b\), and write
Here \(\mu_b\) is the block-specific baseline mean and \(\tau\) is a common condition effect. For the variance calculation below, assume condition assignment is randomized within each block, each donor contributes one outcome, errors from different donors are independent with mean zero and common variance \(\sigma^2\), and the additive model has no condition-by-block interaction. The common-effect assumption means the treatment contrast is the same in every block; if effects vary, the weighted contrast still summarizes block-specific contrasts, but it targets their weighted average.
Under these assumptions, each block difference has variance \(\sigma^2(1/n_{Ab}+1/n_{Bb})\). Differences from separate blocks use different donors and have zero covariance under the model, so the variance of the weighted contrast is the sum of the weighted block variances.
This expression also explains the design trade-off. When a prognostic block characteristic accounts for substantial outcome variation, comparing donors within blocks can reduce the noise contributed by differences between blocks. But dividing a fixed number of donors into more blocks can leave fewer donors per condition within each block, increasing the variance terms \(1/n_{Ab}+1/n_{Bb}\). Blocking is not automatically more precise simply because it uses more categories.
Worked Calculation of a Blocked Contrast
Worked Example: Estimating a Donor-Level Condition Difference
Consider a hypothetical study with three age blocks and four donors per block: two assigned to A and two to B. The outcome is a prespecified donor-level log-expression summary, in log-expression units. Suppose the observed outcomes are as follows.
| Block | Condition B outcomes | Condition A outcomes | Within-block contrast, A minus B |
|---|---|---|---|
| 1 | 0, 2 | 1, 3 | 2 − 1 = 1 |
| 2 | 2, 4 | 5, 7 | 6 − 3 = 3 |
| 3 | 1, 3 | 3, 5 | 4 − 2 = 2 |
All blocks contain the same number of donors, so equal block weights are appropriate for the enrolled-donor average. The estimate is
For a model-based standard error, the six block-by-condition groups each contain two donors. Within every group, the two outcomes are one unit below and one unit above that group’s mean, so the sum of squared deviations is \(1^2+1^2=2\) per group. Thus the residual sum of squares is \(6(2)=12\). There are 12 observations and 6 group means, giving \(df=12-6=6\), and the pooled residual variance estimate is \(s^2=12/6=2\) square log-expression units.
Using \(w_b=1/3\) and \(n_{Ab}=n_{Bb}=2\), the estimated variance and standard error are
For illustration, if the independent, common-variance normal-error model is appropriate, a 95% \(t\) interval with 6 degrees of freedom is \(2.000\pm 2.447(0.8165)\), or approximately \([0.002,3.998]\) log-expression units. This interval describes uncertainty in the average condition contrast under the stated model; it is not a 95% probability statement about the fixed effect. The small residual degrees of freedom also make clear that this interval depends strongly on the model assumptions.
When Blocks Become Too Small
Cross-classifying characteristics can create sparse blocks quickly. If age has three categories and sex has two, the combined scheme can have six blocks. Some may contain only one or two donors. A one-donor block cannot contribute a within-block comparison between two conditions. A two-donor block can place one donor in each condition, but it provides no within-condition replication in that block. Very small blocks therefore limit allocation options and make block-specific estimates unstable.
Worked Example: A Sparse Age-by-Sex Blocking Scheme
A hypothetical roster has 16 donors in four age-by-sex blocks, with counts 2, 3, 3, and 8. Investigators wish to keep the two-condition allocation as balanced as possible overall while ensuring both conditions appear in every block.
One valid allocation gives condition A counts of 1, 1, 2, and 4, respectively; condition B then receives 1, 2, 1, and 4 donors. Each block contains both conditions, and the overall totals are 8 donors per condition. However, in the blocks of size two, the within-block contrast compares just one donor in A with one in B. Those comparisons are especially sensitive to donor-to-donor variability.
If any cross-classified block had only one donor, it would be impossible to assign both conditions within that block. The investigators should not pretend that a within-block contrast exists there. A more defensible plan could use fewer, scientifically prioritized blocks, or use unrestricted randomization with a prespecified procedure to seek reasonable balance on the known characteristics. Any coarsening should be specified before outcomes are examined; changing categories after seeing results risks making the analysis data-dependent.
Useful diagnostics include checking the number of donors assigned to each condition in every planned block, confirming that no block is empty for a condition, and reviewing whether the chosen categories actually represent meaningful differences in the prognostic characteristic. Unexpectedly sparse cells, extensive donor exclusions, or outcome measurement failures concentrated in one condition can undermine the intended comparison. Blocking balances only the characteristics used in the allocation; it does not ensure balance on all other variables.
Blocking Compared With Other Approaches
| Approach | Strength | Limitation |
|---|---|---|
| Unrestricted randomization | Simple allocation with many possible assignments | May produce chance imbalance in a small sample on a prognostic characteristic |
| Blocking before randomization | Guarantees representation of each condition within chosen blocks | Too many or too small blocks restrict allocation and can weaken within-block information |
| Adjustment for a measured characteristic in analysis | Can account for a prognostic variable without creating many allocation strata | Does not itself guarantee balanced assignment; results depend on the prespecified model |
In a randomized study, blocking is a design choice that can improve balance on selected known characteristics while retaining random assignment within blocks. In a nonrandomized comparison, merely grouping donors by age or sex does not remove confounding from other causes of condition assignment and outcome. Blocking also does not mean that each donor receives both conditions: conditions are allocated among distinct donors in this design.
Common Mistakes and Practical Limits
- Using too many characteristics. Creating a separate block for every combination can produce empty or singleton blocks. Prioritize a small number of prespecified prognostic characteristics.
- Assuming blocking guarantees precision gains. It may reduce variation between blocks, but small within-block sample sizes can increase uncertainty. The allocation counts and expected variability matter.
- Ignoring the target weights. Equal weighting of blocks and weighting by donor counts answer different questions when block sizes differ. State the intended average before interpreting the contrast.
- Treating cells as independent experimental units. The condition is assigned to donors, so the relevant independent units for this comparison are donors, not the many cells measured from each donor.
- Using outcome-informed blocks. Blocks should be based on pre-assignment characteristics, not chosen or redefined after inspecting condition outcomes.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Ten donors are divided into age blocks of four and six. What condition counts would ensure both conditions appear in both blocks, and what must be checked before randomization?
- Why can creating separate blocks for every age-by-sex combination be a poor plan?
- Three equally weighted blocks have estimated A-minus-B contrasts of \(-1\), \(2\), and \(5\) outcome units. What is the overall blocked contrast?
- Does blocking on age in a randomized study guarantee balance on all donor characteristics? Explain.
- Why is the donor, rather than each captured cell, the experimental unit in this allocation?
Solutions. (1) For example, allocate two donors to each condition in the block of four, and three to each condition in the block of six. Check that the blocks were defined before assignment, both conditions are feasible in each block, and the allocation is randomized within blocks. (2) Cross-classification can create tiny or empty blocks; a singleton cannot receive both conditions, and a two-donor block has only one donor per condition. (3) The estimate is \(\frac{1}{3}(-1+2+5)=2\) outcome units. (4) No. Blocking guarantees representation across conditions for age blocks under the allocation plan, but other characteristics can still be imbalanced by chance. (5) Condition is assigned to donors. Cells are measurements nested within those donors, not independently randomized units.