Why Exact Covariate Balance Is Not Guaranteed

Balancing Tendency Is Not a Balance Guarantee

Minimization uses incoming participants’ baseline covariates to favor the treatment assignment that would leave the trial more balanced. But “favor” is not the same as “always choose.” In randomized minimization, the less favorable assignment can still occur by design. Even when it does not occur, the available assignments are discrete: assigning one participant cannot always make every margin exactly equal.

The key distinction is between a balancing tendency—a preference, in expectation, for the assignment with the lower imbalance score—and a guarantee that the final score will be zero. This tutorial uses a two-arm trial to show why the first does not imply the second.

Definition: Let \(S_A\) and \(S_B\) be the imbalance scores that would result if the next participant were assigned to arms A and B, respectively. A randomized minimization rule assigns the participant to the arm with the lower score with probability \(q\), where \(1/2 < q \leq 1\), and to the arm with the higher score with probability \(1-q\). If \(S_A=S_B\), a common rule assigns with probability \(1/2\) to each arm. The score is a nonnegative summary of imbalance, often formed from weighted differences in treatment counts across covariate levels and possibly an overall treatment-count difference. A score of zero means balance on the margins included in that score.

This setup assumes that the participant’s baseline covariates are recorded before assignment, the score and tie rule are specified in advance, and assignment follows the stated conditional probabilities given the allocation history and the incoming participant’s covariates. The randomization draw must be implemented as specified. The discussion concerns the allocation mechanism; it does not require outcome data or assume that the covariates predict the outcome correctly.

Why a Positive-Probability Path Can End Imbalanced

Suppose the trial starts with no participants and enrolls three participants who all have the same covariate profile. For this simple case, let the score be the absolute difference in the numbers assigned to the two arms. If the current arm counts are \(n_A\) and \(n_B\), the score is \(|n_A-n_B|\). Let \(q=0.8\), so the assignment that yields the lower score is chosen with probability 0.8 whenever the candidate scores differ.

Worked Example: A Three-Participant Imbalanced Path

All data here are hypothetical. For the first participant, assigning to A or B gives the same score, so the tie rule assigns to A with probability \(1/2\). Consider the path in which the first participant goes to A.

Second participant. Assigning to A would produce counts \((2,0)\), with score \(2\). Assigning to B would produce counts \((1,1)\), with score \(0\). Therefore, assigning to A is the less favorable choice and occurs with probability \(1-q=0.2\).

Third participant, given the first two went to A. Assigning to A would produce counts \((3,0)\), with score \(3\); assigning to B would produce counts \((2,1)\), with score \(1\). Again, assigning to A is the less favorable choice, with probability \(0.2\).

The probability of this complete path is \( (1/2)(0.2)(0.2)=0.02 \). At its end, the counts are \((3,0)\), and the score is \(3\), not zero. Thus there is a positive-probability allocation path that leaves residual imbalance. The path is allowed by the rule; its occurrence does not show that the rule was implemented incorrectly.

More generally, if a specified path has successive conditional assignment probabilities \(r_1,\ldots,r_m\), its probability is the product \(r_1\cdots r_m\). This follows by applying the multiplication rule for conditional probabilities to the sequential assignments. If every step on the path has positive conditional probability, the path itself has positive probability. A path need not be common to matter for this argument: its existence is enough to show that the design does not guarantee an exact final score.

Key distinction: A rule guarantees exact balance only if every allocation path it permits ends with score zero. Randomized minimization with a positive probability of choosing the higher-score assignment generally permits paths that do not. A balancing preference changes the probabilities of paths; it does not remove every imbalanced path.

The Balancing Tendency in Expected Score

The path argument establishes possibility, not typical performance. To describe the local balancing tendency precisely, compare the two candidate scores for one assignment decision. Let \(S_{\min}=\min(S_A,S_B)\), \(S_{\max}=\max(S_A,S_B)\), and \(d=S_{\max}-S_{\min}\). When \(d>0\), randomized minimization selects the lower score with probability \(q\). Its conditional expected score immediately after the assignment is therefore:

E(S_{\text{chosen}}\mid\text{current history and covariates}) = qS_{\min}+(1-q)S_{\max} = S_{\min}+(1-q)d.

If the two candidate assignments instead had been chosen by a fair coin, the corresponding expected score would be \(S_{\min}+d/2\). Subtracting the randomized-minimization expectation gives:

\left(S_{\min}+\frac{d}{2}\right) -\left(S_{\min}+(1-q)d\right) =\left(q-\frac{1}{2}\right)d > 0.

Thus, conditional on the same history and covariates, the expected immediate score is lower under the biased coin than under a fair coin whenever the candidate scores differ. If \(d=0\), the assignment choices have the same score and there is no immediate score advantage to either arm. This is a local statement about the next assignment; by itself, it is not a theorem that the score decreases at every step or that the final score is zero.

Worked Example: Quantifying the Local Preference

Suppose a hypothetical allocation history gives candidate scores \(S_A=2.4\) and \(S_B=5.4\), and the design uses \(q=0.8\). The score is measured in score points, not in a physical unit. Arm A is the preferred assignment, and \(d=5.4-2.4=3.0\).

The conditional expected score after randomized minimization is \( (0.8)(2.4)+(0.2)(5.4)=1.92+1.08=3.00 \) score points.

Under a fair coin between these same candidate assignments, the expected score would be \( (0.5)(2.4)+(0.5)(5.4)=1.20+2.70=3.90 \) score points. The difference, \(3.90-3.00=0.90\), agrees with \( (q-1/2)d=(0.8-0.5)(3.0)=0.90 \). The rule improves the immediate expected score relative to a fair coin, but there remains a \(0.2\) probability of selecting the score of \(5.4\).

Discrete Constraints Can Leave Residual Imbalance Too

Randomness is not the only reason exact balance may be absent. Scores are based on counts, and counts change in whole-person increments. An exact target may be unreachable at a particular enrollment size, even if the algorithm always chooses a lower-score assignment whenever the candidate scores differ.

Worked Example: A Tie at an Unavoidable Odd Total

Consider three hypothetical participants with the same covariate profile. Suppose the score includes the overall arm-count difference, and use deterministic minimization for this example: choose a lower-score assignment whenever one exists, with a fair coin for ties. The first participant is assigned by a tie-breaking coin; suppose the result is A.

For the second participant, assigning to A gives counts \((2,0)\) and score \(2\), while assigning to B gives \((1,1)\) and score \(0\). The lower-score assignment is B, so the counts become \((1,1)\).

For the third participant, either assignment gives an absolute difference of \(1\): counts \((2,1)\) if assigned to A, or \((1,2)\) if assigned to B. This is a tie, so either arm may be selected. The final score is \(1\) whichever assignment occurs.

Here, no participant was assigned to an arm with a higher candidate score. Exact overall equality is nevertheless impossible with three participants, because two integer arm counts that sum to \(3\) cannot be equal. The tie rule cannot overcome that arithmetic constraint.

What to Check When Imbalance Remains

A nonzero final score is compatible with a correctly functioning design. It is a reason to understand the realized allocation, not by itself evidence that the algorithm failed. In a practical review, first establish which score the design actually used: the included covariate margins, any weights, whether an overall arm-count term was included, the value of \(q\), and how ties were resolved. Then check that the assignment probabilities and randomization draws followed that specification at each enrollment.

Potential warning signs concern discrepancies between the declared rule and its operation. Examples include using a covariate value before it was available, calculating a candidate score from stale counts, applying the wrong tie rule, or assigning with a probability different from the prespecified one. Reviewing the allocation history can distinguish these implementation problems from a valid but unlucky randomization path. A single observed imbalance cannot, on its own, reveal which explanation applies.

How the Guarantee Differs Across Approaches

The relevant alternatives differ in what they prioritize. A deterministic rule that always takes the lower-score arm removes the deliberate probability of choosing the higher score, but it still cannot always attain score zero: ties, incompatible margins, and integer constraints can leave residual imbalance. Stratified permuted blocks instead use prespecified strata and block structure to control treatment counts within those strata, subject to the chosen block and stratum definitions. They do not automatically ensure balance on every covariate margin, especially when many profiles are sparse or participants fall into different strata.

A fair-coin allocation does not favor the lower candidate score at an assignment decision. Randomized minimization with \(q>1/2\) lies between that approach and deterministic minimization: it retains a positive chance of the higher-score assignment while increasing the chance of the lower-score assignment. These comparisons concern allocation behavior, not a claim that one method is universally best; the appropriate design depends on the trial’s covariates, operational constraints, and objectives.

Approach	Assignment preference	Does it guarantee zero score?
Fair coin between arms	No preference based on candidate scores	No
Randomized minimization, \(1/2<q<1\)	Favors the lower candidate score but allows the higher one	No
Deterministic minimization, \(q=1\)	Chooses a lower score when scores differ; ties still need a rule	No, not in general
Stratified permuted blocks	Controls allocation counts within defined strata and blocks	Only for the specified structure and conditions, not every possible margin

Common Pitfall and Limitations

Wrong reasoning: “The algorithm minimizes imbalance, so a final nonzero score proves it did not work.” Correction: Randomized minimization gives the lower-score assignment a greater probability, not certainty. A permitted higher-score assignment can leave residual imbalance.
Wrong reasoning: “A lower expected score means the score must fall after each assignment.” Correction: The expectation comparison is conditional and local. A realized score can rise, and later assignments can change the score again.
Wrong reasoning: “Choosing the lower score whenever possible guarantees exact balance.” Correction: Discrete counts and tied candidate scores can make zero unattainable, as the odd-total example shows.

The local expectation result compares the two candidate assignments at one decision. It does not establish a global probability of ending balanced, and it does not say how large a realized imbalance should be considered acceptable. Those questions require attention to the complete allocation and the trial’s balance criteria. Minimization also only acts on the quantities included in its score: it cannot guarantee balance on omitted covariates or correct a covariate measured inaccurately.

Key takeaway: Randomized minimization is designed to favor assignments that reduce a prespecified imbalance score. Because it may select the less favorable assignment and because count constraints can make exact equality unattainable, residual imbalance is possible without contradicting the method’s balancing tendency.

Check Your Understanding

Try each problem, then compare your reasoning with the solutions.

Two candidate scores are \(1.5\) and \(4.5\), and \(q=0.7\). What is the conditional expected score under randomized minimization?
In a two-arm trial, three participants all share one profile. If the score includes the overall count difference, can exact equality be achieved after all three are assigned? Explain.
Why does a positive-probability path that ends with a nonzero score disprove a guarantee of exact balance?
If both candidate assignments have the same score, what does the local expected-score comparison tell us about preferring one arm?
A trial ends with a nonzero score. Give one explanation consistent with correct implementation and one implementation problem worth checking.

Solutions. (1) The lower score is \(1.5\) and \(d=4.5-1.5=3.0\), so the expectation is \( (0.7)(1.5)+(0.3)(4.5)=1.05+1.35=2.40 \) score points. (2) No. Two integer counts summing to \(3\) cannot be equal, so the absolute overall count difference is at least \(1\). (3) A guarantee requires every allocation path allowed by the design to end at score zero; one allowed path with positive probability and a nonzero final score is enough to contradict that requirement. (4) If the candidate scores are equal, then \(d=0\), so the immediate expected-score advantage is zero; the assignment rule may break the tie randomly. (5) A permitted randomized choice of the higher-score assignment is consistent with correct implementation. A discrepancy between the prespecified and computed candidate scores, or use of the wrong assignment probability, should be checked.