A Worked Minimization Design for a Two-Arm Trial

Why Specify Minimization as a Full Randomized Rule?

A minimization design assigns each new participant by comparing how balanced the trial would be under each possible assignment. That comparison is only as clear as its score: a protocol must say which covariates enter, how each is measured, how discrepancies are scaled and weighted, and what happens when candidate assignments tie. It must also specify the randomization step. A deterministic rule always chooses the arm with the better score; a randomized rule can favor that arm while preserving a positive probability of assignment to either arm.

The design below uses two arms, A and B, categorical baseline severity, and continuous age. It uses no response outcomes: assignment depends on the covariates of the incoming participant and the assignments already made. The numerical choices are illustrative and would need clinical and operational justification in an actual protocol.

Definition: For a candidate participant, a minimization score is a prespecified numerical measure of the covariate imbalance that would result under a possible assignment. A randomized minimization procedure calculates a score for assigning the participant to A and one for assigning the participant to B, then randomizes with probabilities that favor the lower-scoring assignment.

Specify the Covariate Scores Before Enrollment

Let severity have three mutually exclusive, exhaustive categories: mild, moderate, and severe. For category \(k\), let \(n_{Ak}\) and \(n_{Bk}\) be the numbers already assigned to A and B. Define the categorical severity discrepancy as the sum of the absolute differences across categories. For age, compare the two arm means after the candidate assignment. In this example, express the age discrepancy in units of 10 years so that it is dimensionless.

C=\sum_{k\in\{\text{mild, moderate, severe}\}} |n_{Ak}-n_{Bk}|, \qquad G=\frac{|\bar{x}_A-\bar{x}_B|}{10\ \text{years}}, \qquad S=C+2G.

Here \(C\) is the categorical score, \(G\) is the scaled age score, and \(S\) is the total score. The weight of 2 means that a 10-year difference between arm mean ages contributes 2 score units. A one-person discrepancy in one severity category contributes 1 unit. These weights are design choices, not universal constants: they should be chosen in advance, justified for the trial, and kept fixed during enrollment.

For each candidate assignment, update the relevant category count and age mean as though the participant had been assigned to that arm, then calculate \(S\). The arm means are defined only when each arm has at least one participant. One simple startup rule is to randomize the first two participants in a random order, with one assigned to each arm; apply the minimization rule from participant three onward. This rule is part of the design and should be implemented as written.

Worked Candidate Scores

Worked Example: Categorical and Age Components Can Favor Different Arms

Suppose, after the startup assignments, arm A has severity counts (mild 2, moderate 1, severe 1) and ages 40, 50, 60, and 70 years. Arm B has counts (mild 1, moderate 2, severe 1) and ages 45, 55, 65, and 75 years. A new participant has moderate severity and age 50 years. Use the specified score \(S=C+2G\).

Candidate assignment to A. The updated severity counts are A (2, 2, 1) and B (1, 2, 1). The category differences are (1, 0, 0), so \(C_A=|1|+|0|+|0|=1\). The current age means are \(\bar{x}_A=220/4=55\) years and \(\bar{x}_B=240/4=60\) years. After assigning the participant to A, \(\bar{x}_A=(220+50)/5=54\) years, while \(\bar{x}_B=60\) years. Thus \(G_A=|54-60|/10=0.6\), and \(S_A=1+2(0.6)=2.2\).

Candidate assignment to B. The updated counts are A (2, 1, 1) and B (1, 3, 1). The differences are (1, -2, 0), so \(C_B=|1|+|-2|+|0|=3\). The means would be 55 years in A and \((240+50)/5=58\) years in B. Therefore \(G_B=|55-58|/10=0.3\), and \(S_B=3+2(0.3)=3.6\).

The age component favors B, because its candidate age-mean gap is smaller, but the severity component favors A. The total score is lower for A, so the rule favors A. A lower score does not, by itself, determine the assignment: the allocation probability must also be specified.

Design rule: If one candidate score is lower, assign the participant to the lower-scoring arm with probability \(q=0.8\), and to the other arm with probability \(1-q=0.2\). If the scores are equal, assign with probability \(0.5\) to each arm. Use an independent random draw with these probabilities at each step, and record the probabilities used.

Worked Example: Applying the Allocation Probability

Continue the preceding example, where \(S_A=2.2\) and \(S_B=3.6\). Because A has the lower score, the specified biased-coin rule assigns the participant to A with probability 0.8 and to B with probability 0.2. A uniform random draw on the interval from 0 to 1 can implement this: assign to A if the draw is below 0.8, and otherwise assign to B.

The probability statement describes the randomization mechanism for this participant given the current trial history and covariates. It does not say that A is certain to be selected, nor that A is clinically superior. The nonzero 0.2 probability of assignment to B is an intentional part of the design.

Worked Example: Handling a Tie

Suppose each arm currently has one participant with mild severity and age 50 years, and one participant with severe severity and age 60 years. The new participant has moderate severity and age 55 years. If assigned to A, the moderate-category difference becomes 1 and the other category differences remain 0, so \(C_A=1\). The age means after assignment are still 55 years in both arms, so \(G_A=0\) and \(S_A=1+2(0)=1\).

If assigned to B, the severity differences are (0, -1, 0), giving \(C_B=1\), and the arm age means again both equal 55 years, giving \(G_B=0\). Thus \(S_B=1\) as well. The tie rule assigns with probability 0.5 to A and 0.5 to B. Choosing an arm arbitrarily in this case would not implement the stated randomized procedure.

Why the Categorical Score Uses Absolute Differences

The categorical component adds the magnitude of the discrepancy in every category. This avoids cancellation: a surplus in mild severity cannot offset a deficit in severe severity merely because their signed differences have opposite signs. For example, differences \((1,-1,0)\) sum to zero, but they indicate discrepancies in two categories, and the score is \(|1|+|-1|+|0|=2\).

When the two arm totals are equal, the sum of the signed category differences is zero. In that case, the positive differences sum to the same amount as the absolute values of the negative differences. Therefore, half the sum of absolute differences equals the one-sided excess count when the two arm totals are equal. This design uses the unhalved sum. The unhalved score is not itself the one-sided excess count, even when arm totals are equal.

The age component also uses an absolute difference, so the direction of imbalance does not change the score: an age-mean gap of 6 years contributes the same amount whether A or B has the higher mean. The score is designed to measure imbalance, not to favor a particular direction.

Worked Example: A Further Candidate Comparison

Suppose the current severity counts are A (2, 1, 1) and B (1, 2, 1), and the current ages are 40, 50, 60, and 70 years in A and 45, 55, 65, and 75 years in B. A new participant has severe severity and age 70 years.

If assigned to A, the category counts become A (2, 1, 2) and B (1, 2, 1), with differences (1, -1, 1). Thus \(C_A=1+1+1=3\). The updated mean age in A is \((220+70)/5=58\) years, compared with 60 years in B. Hence \(G_A=|58-60|/10=0.2\), and \(S_A=3+2(0.2)=3.4\).

If assigned to B, the counts become A (2, 1, 1) and B (1, 2, 2), with differences (1, -1, -1), so \(C_B=3\). The mean age in A remains 55 years; in B it becomes \((240+70)/5=62\) years. Therefore \(G_B=|55-62|/10=0.7\), and \(S_B=3+2(0.7)=4.4\). A has the lower score, so the allocation probabilities are 0.8 to A and 0.2 to B.

Assumptions, Diagnostics, and Practical Limits

This rule requires severity to be classified consistently into the stated categories and age to be measured in years before assignment. The relevant covariates must be available for the incoming participant, and the current allocation history must be accurate. The scoring rule, startup rule, weights, tie handling, and randomization probabilities should be fixed before enrollment; changing them in response to observed imbalances makes the actual design harder to characterize. Assignment should be concealed until the required baseline covariates are collected and the allocation is generated.

Operational checks can identify implementation failures early. For a sample of allocation records, independently recompute both candidate scores from the recorded history and confirm that the lower-score arm received the higher probability, or that tied scores received equal probabilities. Verify that age values are in years, severity labels are valid, both candidate scores are recalculated after every assignment, and the recorded assignment probability matches the randomization rule. A missing age or severity classification is not a reason to silently substitute a value or omit a component: the protocol needs a prespecified missing-covariate procedure.

The age scale matters. Dividing by 10 years makes a 10-year difference contribute one unit to \(G\); dividing by 5 years would double that component for the same data. The weights and scale therefore determine the trade-off between balancing age and severity. Review the score on realistic trial profiles before enrollment, and confirm that no coding or unit choice gives one component an unintended dominance.

Minimization does not guarantee exact balance, and a randomized rule can assign a participant to the arm with the higher score. It also does not ensure balance on covariates omitted from the score. Deterministic minimization would always choose the lower-score arm, but can make assignments predictable and removes the positive assignment probability to the other arm at non-ties. Simple randomization does not use these covariates when assigning participants, while stratified or blocked randomization can balance prespecified factors through separate allocation sequences. Those approaches may be easier to implement in some settings, but continuous age requires categorization or another specified handling if used as a stratification factor. The choice depends on the trial’s operational needs and the covariates requiring control.

Common Mistakes

Letting signed discrepancies cancel. Adding \(1+(-1)+0=0\) incorrectly suggests perfect category balance. The defined score adds absolute values, giving 2.
Changing the score scale mid-trial. A different age divisor changes the contribution of age. Use the prespecified units and weights throughout.
Treating the lower score as a certain assignment. This is a biased-coin design, not a deterministic rule: the lower-scoring arm has probability 0.8, not 1.
Calling a balanced allocation a treatment effect. The score describes baseline covariate balance under candidate assignments; it does not estimate or establish a difference in outcomes.
Ignoring implementation details. Tie handling, initial assignments, missing covariates, and allocation-probability records are components of the design, not optional software choices.

Key takeaway: A reproducible minimization design specifies the candidate scores, units and weights, startup and tie rules, and randomization probabilities. Recalculate both candidate scores from the current trial history, then use the prespecified random rule rather than assigning deterministically by default.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

For severity differences (1, -1, 0), calculate the unhalved absolute-difference score. If arm totals are equal, what is half this score?
If candidate scores are \(S_A=2.4\) and \(S_B=1.8\), what assignment probabilities follow from the stated rule?
If the age divisor changes from 10 years to 5 years, what happens to the age component for a fixed 10-year mean difference?
Why should a tie receive a prespecified probability rather than an arbitrary choice by the allocator?
Name two checks that could reveal an implementation error in this minimization procedure.

Solutions. 1. The score is \(|1|+|-1|+|0|=2\); half is 1, the one-sided excess count when arm totals are equal. 2. B has the lower score, so assign to B with probability 0.8 and to A with probability 0.2. 3. The scaled age discrepancy doubles: it is \(10/10=1\) with the original divisor and \(10/5=2\) with the new divisor. 4. A prespecified tie rule makes the randomization reproducible and avoids allocator discretion; here each arm receives probability 0.5. 5. For example, independently recompute candidate scores from recorded covariates and assignments, or check that the stored assignment probability agrees with the score comparison and tie rule.