Two Ways to Balance Baseline Factors
Both minimization and stratified permuted-block randomization use baseline factors to reduce treatment-group imbalance. Their central difference is the unit at which balance is pursued. Minimization compares treatment counts across levels of selected factors as a participant arrives. Stratified permuted-block randomization instead forms strata from the joint combinations of the factors and randomizes within each stratum. Thus, two factors with several levels can produce many strata, even when each factor considered separately has only a few levels.
The comparison matters when factors are crossed. If a trial uses severity and site, balancing severity overall and site overall does not guarantee equal numbers in every severity-by-site combination. Conversely, creating a separate stratum for each combination may leave very few participants in some strata. The choice concerns the allocation constraints the design can realistically sustain, not just which method sounds more comprehensive.
These descriptions assume the factors are measured before allocation, their levels are defined in advance, and the allocation procedure is implemented as specified. For stratified randomization, the random sequence within each stratum must be generated and concealed appropriately. For randomized minimization, the score and the rule for converting scores to assignment probabilities must be prespecified; the simple score examples below illustrate comparison of candidate assignments, not a universal probability rule. Neither method uses future outcomes to make the allocation decision.
What Counts as a Stratum?
A stratum is defined by the joint profile of the stratification factors. Suppose severity has levels mild and severe, and site has levels North, Central, and South. Their combination defines \(2\times3=6\) possible strata. With three factors having \(L_1,L_2,L_3\) levels, the number of possible combinations is \(L_1L_2L_3\), assuming every combination is allowed. The actual number of participants per stratum depends on how often those profiles occur.
In stratified permuted-block randomization, each stratum has its own allocation sequence. With two treatments and a fixed block size of four, each completed block has two assignments to A and two to B. This is a constraint within that stratum. It does not require equal counts at every moment: an incomplete block may temporarily be imbalanced. Nor does a completed block guarantee equal overall treatment counts if some participants are in incomplete blocks or if strata have different enrollment totals.
Minimization does not ordinarily require a separate sequence for every joint profile. Instead, it uses the marginal counts at each factor level. For example, a new participant who is severe and at North contributes to the severe level and the North level when candidate assignments are compared. A chosen minimization score may sum absolute treatment-count differences across levels. This can target balance across each factor while avoiding a separate allocation constraint for every combination.
Worked Example: Comparing Candidate Minimization Scores
In a hypothetical trial, the current counts are as follows. There are two factors: severity (mild or severe) and site (North or South). A new participant is severe and at South.
| Factor level | A | B |
|---|---|---|
| Mild | 4 | 3 |
| Severe | 2 | 3 |
| North | 3 | 4 |
| South | 3 | 2 |
For illustration, define the score as the sum of absolute A-versus-B count differences at the levels of the factors that the new participant occupies. A lower score indicates less marginal imbalance for those levels.
If assigned to A, the severe counts become A \(=3\), B \(=3\), while the South counts become A \(=4\), B \(=2\). The candidate score is
If assigned to B, the severe counts become A \(=2\), B \(=4\), while the South counts become A \(=3\), B \(=3\). The score is
The scores tie. Under a randomized tie rule, either assignment may be selected according to the prespecified rule. The calculation shows that improving balance at one factor level can be offset by worsening it at another. It does not establish that the severe-by-South joint profile is balanced: the score here compares marginal levels, not every joint cell.
Why Sparse Strata Change the Comparison
As the number of joint combinations increases, enrollment is divided among more allocation sequences. Some strata may enroll only once or not at all; others may not complete a block. In those cells, a design’s within-stratum balance constraint has little opportunity to operate. This is not a defect in the randomization procedure: it is a consequence of trying to maintain local constraints with limited observations.
A simple occupancy calculation makes the issue visible. Assume, only for illustration, that participants independently have equal probability of entering any of 12 possible profiles. For each profile, the probability it is unoccupied after 30 enrollments is \((11/12)^{30}\). Summing the probabilities of occupancy over the 12 profiles gives the expected number of occupied profiles:
Although almost all profiles are expected to appear, that does not mean they will each have enough participants for completed blocks. Under the same assumptions, a particular profile has exactly one participant with probability \(30(1/12)(11/12)^{29}\). Across 12 profiles, the expected number of singleton profiles is
These are expected counts under a deliberately simple equal-probability model, not predictions for a particular trial. Unequal profile frequencies can make sparsity more severe. The calculation illustrates why counting possible combinations alone is not enough: the expected enrollment in each combination and the distribution of actual profiles matter.
Worked Example: A Sparse Joint Profile Under Permuted Blocks
Suppose a hypothetical trial uses severity and site to form six strata and a block size of four, with two assignments to each treatment in every completed block. At one site-severity combination, only three participants enroll before recruitment ends. If those three are the first three assignments in a fresh block, the block contains two assignments to one treatment and one to the other; the fourth assignment, if it occurs, must complete the block with the treatment assigned only once so far.
For example, if the first three assignments are A, B, A, then the fourth must be B. The completed block is balanced at two per treatment. But if no fourth participant with that profile enrolls, the observed stratum counts remain A \(=2\), B \(=1\). The block rule has not failed; its exact balance condition applies to a completed block, not to an incomplete one.
A minimization procedure would instead evaluate each of those participants using the current counts at their factor levels, including counts from other joint profiles that share those levels. It can therefore respond to broader marginal imbalance even when this particular joint profile is rare. The trade-off is that it does not generally force a precise A-versus-B count within that profile.
How to Compare the Allocation Constraints
| Feature | Minimization | Stratified permuted blocks |
|---|---|---|
| Unit used for balancing | Usually levels of factors considered across participants | Each joint combination of stratification factors |
| Allocation constraint | Favors an assignment with a lower score; the exact behavior depends on the prespecified rule | Completed blocks have specified treatment counts within each stratum |
| When combinations are sparse | Can use information from other profiles sharing factor levels | Some strata may have incomplete blocks or very few assignments |
| Primary practical concern | Define and implement the score and randomization rule consistently | Avoid excessive numbers of strata relative to expected enrollment |
The comparison is not a claim that one method is always more balanced. Stratified permuted blocks give a direct within-stratum constraint when enrollment supports it. Minimization can address marginal imbalances across several factors without requiring a complete block in every joint profile, but the achieved joint-profile counts are not guaranteed to match. Both methods require careful implementation and concealment; a highly predictable allocation can undermine protection against selection bias.
Diagnostics and Practical Limits
Before choosing between these approaches, tabulate the anticipated or observed counts for the joint profiles. Useful checks include the number of possible profiles, the expected enrollment in common and rare profiles, and the number of strata likely to end with incomplete blocks. During a trial, summaries should distinguish completed-block balance from the current totals in incomplete strata. A profile with unequal counts at an interim look is not, by itself, evidence that the randomization algorithm malfunctioned.
For minimization, verify that the recorded factor levels match the prespecified categories and that the allocation system applies the same scoring and tie rules each time. Compare counts at each factor level, not just the total arm sizes; acceptable overall totals can conceal imbalance in a prognostic level. For stratified randomization, verify that participants were assigned using the correct joint stratum and that the intended block procedure was applied. Unexpectedly empty or extremely small strata may indicate that the planned stratification is too fine for enrollment, but they may also reflect genuinely rare profiles.
Neither method guarantees balance on unrecorded factors, prevents chance imbalance in outcomes, or substitutes for appropriate analysis. Stratified permuted blocks can be awkward when many combinations are possible, and fixed block sizes may make upcoming assignments more predictable when prior assignments within a stratum are known. Minimization avoids a separate block sequence for every combination, but its score and randomization rule must be specified, and its balance target may not be intuitive from the final joint-profile table. These are design trade-offs to assess before enrollment, rather than reasons to select a method after inspecting outcomes.
Common Mistake
Wrong reasoning: “Minimization balances all combinations because it uses all the factors, so it guarantees the same within-cell balance as stratified blocks.” Including multiple factors in a score does not make every joint profile a separately constrained stratum. A score based on factor levels can improve marginal balance while leaving a particular combination uneven.
Correction: State exactly what the procedure balances. Stratified permuted blocks constrain treatment counts within each joint stratum when blocks are completed. Minimization compares assignments using its specified score, commonly built from factor-level counts, and does not generally impose a fixed allocation ratio in every joint cell. Sparse enrollment limits the practical benefit of either method’s intended balance.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A trial stratifies on a factor with three levels and a factor with four levels. How many joint strata are possible, and what practical concern follows?
- With two treatments and block size four, what treatment counts occur in a completed block if allocation is equal? Does that guarantee equal counts in an incomplete block?
- Why can minimization improve balance across factor levels without guaranteeing balance in every joint profile?
- In a hypothetical minimization calculation, assignment to A has score 5 and assignment to B has score 3. Which candidate has the lower score, and what does that comparison alone tell you?
- Give one diagnostic relevant to sparse strata and one limitation shared by both approaches.
Solutions. (1) There are \(3\times4=12\) possible joint strata. The trial must assess whether expected enrollment is sufficient to populate them; otherwise many may be empty or have incomplete blocks. (2) A completed block contains two assignments to each treatment. An incomplete block can be unequal because its remaining assignments have not yet occurred. (3) Minimization can aggregate counts at the levels of each factor, so participants in different joint profiles may contribute to the same level totals; its score does not necessarily impose a separate constraint on each combination. (4) Assignment to B has the lower score, 3 rather than 5. This identifies the preferred candidate under that score comparison but does not, without the prespecified randomization rule, determine the assignment probability or guarantee balance. (5) A relevant diagnostic is the number of expected or observed participants per joint profile, including the frequency of incomplete blocks. Both approaches can leave imbalance on unrecorded factors and neither guarantees balance in outcomes.