Why the Allocation Rule Matters
Minimization uses the covariates of the incoming participant and the assignments already made to compare the imbalance that would result from placing that participant in each treatment arm. The comparison produces a lower-score arm: the assignment that would yield the smaller value of the prespecified imbalance score. The next decision is how firmly to favor that arm. A deterministic rule always assigns to it; a randomized rule favors it with a specified probability but retains a chance of assigning to the other arm.
That difference affects more than the next assignment. Deterministic minimization can make the allocation highly predictable once the allocation history and the next participant’s covariates are known. A biased-coin rule adds a controlled element of chance. It does not remove the score-based preference, and it does not guarantee a particular balance after each assignment. It changes the probability with which the preferred assignment is made.
Scores and Conditional Assignment Probabilities
Consider a two-arm trial with arms A and B. Before assigning a new participant, calculate the imbalance score that would result under each possible assignment. Call these scores \(S_A\) and \(S_B\). A smaller score indicates the assignment that is preferred by the chosen minimization score. The score is calculated using information available before randomization; it does not use the participant’s future outcome.
The probability \(p\) attaches to whichever arm has the lower score on that particular assignment. It does not automatically mean “assign to A with probability \(p\).” In particular, if B has the lower score, then the probability of assigning to A is \(1-p\). The allocation probability is conditional on the observed allocation history and the incoming participant’s covariates, because these determine the two scores.
When \(p=1\), the biased-coin rule reduces to deterministic minimization whenever the scores are unequal. When \(p=1/2\), the assignment is equally likely to go to either arm, regardless of which score is lower. Values strictly between \(1/2\) and 1 express an intermediate degree of preference for the lower-score arm.
Assumptions for Applying the Rule
The comparison is meaningful only if the allocation procedure is defined before it is used. The imbalance score, its covariates and weights, the value of \(p\), and the handling of tied scores should be specified in advance. The needed covariate values must be available before the assignment, and the score calculation must use the same prespecified rule for both possible assignments. If there is a tie, the lower-score rule alone does not determine an assignment; the tie-breaking procedure must supply that decision.
The random component also depends on correct implementation. At each decision, the random-number mechanism must produce an assignment according to the stated conditional probabilities. The allocation record should preserve the score comparison, the probability used, and the resulting assignment. Staff should not be able to change the assignment after seeing the random result. These are procedural conditions, not assumptions about how treatment affects outcomes.
The examples below use a standard illustrative score: sum the absolute differences in arm counts across the observed levels of each covariate. They use equal weights and two treatment arms. Other prespecified scores are possible; the probability rule applies once the two hypothetical scores have been calculated.
Worked Example: Finding the Favored Arm From Marginal Counts
Suppose a hypothetical trial has two binary covariates, X and Y. Before the next assignment, the counts are as follows. The incoming participant has X1 and Y0.
| Covariate level | Arm A | Arm B |
|---|---|---|
| X0 | 4 | 3 |
| X1 | 2 | 4 |
| Y0 | 2 | 3 |
| Y1 | 4 | 4 |
For each possible assignment, calculate the sum of the absolute arm-count differences. If the participant is assigned to A, the X1 counts become 3 in A and 4 in B, while X0 remains 4 versus 3. The X contribution is therefore \(|3-4|+|4-3|=2\). For Y, Y0 becomes 3 versus 3 and Y1 remains 4 versus 4, giving contribution \(|3-3|+|4-4|=0\). Thus \(S_A=2+0=2\).
If the participant is assigned to B, X1 becomes 2 versus 5 and X0 remains 4 versus 3, so the X contribution is \(|2-5|+|4-3|=4\). Y0 becomes 2 versus 4 and Y1 remains 4 versus 4, so the Y contribution is \(|2-4|+|4-4|=2\). Thus \(S_B=4+2=6\).
Since \(S_A=2<S_B=6\), A is the lower-score arm. Deterministic minimization assigns the participant to A. With a biased-coin rule using \(p=0.80\), the conditional probabilities are \(P(A)=0.80\) and \(P(B)=1-0.80=0.20\). The score comparison favors A, but a 20% chance of assigning to B remains.
How a Biased Coin Is Applied
A direct implementation uses a fresh random draw for the assignment. For example, draw \(U\) uniformly from 0 to 1. If A has the lower score, assign to A when \(U<p\), and assign to B otherwise. If B has the lower score, reverse the labels: assign to B when \(U<p\), and to A otherwise. This procedure makes the probability attach to the lower-score arm regardless of its name.
For a single decision with unequal scores, let \(I\) equal 1 if the lower-score arm is selected and 0 otherwise. Under the biased-coin rule, \(I\) is a Bernoulli random variable with success probability \(p\). Consequently, \(E(I)=p\), since \(E(I)=1\cdot p+0\cdot(1-p)\). This establishes exactly what the bias parameter controls: the probability of choosing the currently favored arm at that decision. It does not specify the final treatment proportions or guarantee a particular amount of imbalance at the end of the trial.
Worked Example: The Lower Score Is for Arm B
For a hypothetical incoming participant, suppose the score calculations give \(S_A=9\) and \(S_B=5\). With a biased-coin rule using \(p=0.80\), B is favored because \(5<9\). Therefore
The probability of assigning to A is 0.20, not 0.80: \(p\) is the probability of the lower-score arm, which here is B. Under deterministic minimization, the same comparison would assign to B with probability 1.
This is a conditional statement about this decision, given the current history and covariates. It is not a claim that 80% of all participants in the completed trial will be assigned to B. The favored arm can differ from one arrival to the next as the observed history and covariates change.
Worked Example: Turning a Random Draw Into an Assignment
Suppose a hypothetical trial uses \(p=0.75\). At one decision, A has the lower score, and the random-number generator returns \(U=0.82\). Since \(0.82\) is not less than \(0.75\), assign to B. This assignment has probability \(1-p=0.25\) under the rule, even though B has the higher score.
At a later decision, suppose B has the lower score and the draw is \(U=0.55\). Because \(0.55<0.75\), assign to B. In both cases, the draw is compared with the same threshold, but the favored arm differs. That is why software should identify the lower-score arm before applying the probability.
The calculations are consistent with the rule: in the first decision, \(P(A)=0.75\) and \(P(B)=0.25\); in the second, \(P(B)=0.75\) and \(P(A)=0.25\). A realized assignment to the less-favored arm is not, by itself, evidence of an error.
What the Choice Changes
The main practical contrast is between guaranteed preference for the lower score and controlled uncertainty about the next assignment. Randomized minimization sacrifices some immediate certainty of choosing the lower-score arm in exchange for making the assignment less predictable. The table summarizes the distinction for unequal scores.
| Feature | Deterministic minimization | Biased-coin minimization |
|---|---|---|
| Probability of lower-score arm | 1 | \(p\), with \(1/2 \leq p \leq 1\) |
| Probability of other arm | 0 | \(1-p\) |
| Predictability from scores and history | Assignment is determined when scores differ | Assignment retains a random component when \(p<1\) |
| Role of the score | Always determines the assignment when scores differ | Determines which arm receives probability \(p\) |
Neither rule should be described as guaranteeing balance. A lower score is a preference according to a chosen score at the current decision, not a promise that every covariate will be balanced at the end. Nor does better covariate balance by itself establish equality of treatment outcomes. Allocation is one part of trial design; outcome analysis must respect the design actually used.
Deterministic allocation can be especially predictable when the score and covariates are known to people involved in enrollment. In practice, predictability can create opportunities for selection if assignments are not properly concealed. A biased coin reduces that certainty when \(p<1\), but it does not ensure concealment: staff may still know the probability, and a large value of \(p\) can leave the assignment substantially predictable. Centralized implementation and appropriate access controls remain important.
Diagnostics and Common Pitfalls
Implementation should be checked against the rule, not judged only by whether the final arm counts look balanced. For each assignment, retain the two scores, which arm was lower, the value of \(p\), and the assigned arm. Confirm that deterministic assignments always follow the lower score when scores differ. For a biased coin, confirm that the programmed probability is \(p\) for the lower-score arm and \(1-p\) for the other arm. Review software logic and test both possible lower-score arms before enrollment begins.
A run of assignments to the lower-score arm does not establish that a biased coin was implemented incorrectly; chance can produce such a run. Conversely, an assignment to the higher-score arm can be correct under the biased-coin rule. Implementation problems are suggested by a reversed probability, a probability applied to a named arm instead of the lower-score arm, or recorded assignments that do not match the stated randomization procedure.
- Wrong reasoning: “The rule uses \(p=0.80\), so assign to A with probability 0.80.” Correction: Assign to the lower-score arm with probability 0.80. If B has the lower score, B receives that probability.
- Wrong reasoning: “The less-favored arm was selected, so the algorithm failed.” Correction: Under a biased coin, that assignment is an intentional possibility with probability \(1-p\).
- Wrong reasoning: “A high \(p\) guarantees better final balance.” Correction: \(p\) controls the conditional chance of selecting the lower-score arm at each unequal-score decision; it does not guarantee a final balance level.
- Wrong reasoning: “The deterministic and biased-coin rules are identical because they use the same scores.” Correction: The scores identify the favored arm, but the allocation probabilities differ whenever \(p<1\).
Limitations and Practical Choice
Deterministic minimization removes assignment uncertainty when the scores differ. That feature can be undesirable when it makes the next assignment foreseeable. A biased coin retains uncertainty, but an assignment to the higher-score arm may temporarily increase the chosen imbalance score. The size of that effect depends on the score and the current data; the probability rule alone does not quantify it.
The value of \(p\) is a design choice that should be documented and fixed in advance. Values nearer 1 place stronger preference on the lower-score arm and leave less uncertainty; values nearer \(1/2\) weaken that preference. Choosing \(p\) does not solve other design questions, such as which covariates belong in the score, how they are weighted, or how ties are handled. Those choices must also be prespecified. This comparison concerns the allocation decision after scores have been calculated.
Check Your Understanding
Answer each question using the allocation rule for unequal scores. Short solutions follow the questions.
- If A has the lower score and \(p=0.70\), what are the assignment probabilities under a biased-coin rule?
- If B has the lower score and \(p=0.80\), what is the probability of assigning to A?
- What does the biased-coin rule become when \(p=1\)? What happens to score-directed preference when \(p=1/2\)?
- If the scores tie, does the lower-score rule determine the assignment? Explain.
- With \(p=0.75\), A has the lower score and the random draw is \(U=0.90\). Is assigning to B consistent with the rule?
Solutions. 1. A receives probability \(0.70\), and B receives \(1-0.70=0.30\). 2. B receives probability \(0.80\), so A receives \(1-0.80=0.20\). 3. At \(p=1\), the rule is deterministic for unequal scores. At \(p=1/2\), each arm has probability \(1/2\), so the scores do not create a probability preference. 4. No. Equal scores mean neither arm is lower; a separately specified tie rule is needed. 5. Yes. Since \(0.90\) is not less than \(0.75\), the rule assigns to the other arm, B, with probability \(1-0.75=0.25\).