Exploration in Covariate-Adjusted Adaptive Designs

Why Preserve Exploration?

A covariate-adjusted adaptive design may use accumulating response information to change the probability of assigning the next participant to a treatment. This can make the probability very small for a treatment that the current model regards as less favorable. Such adaptation may be intentional, but extreme probabilities have a practical cost: within a covariate profile, very few participants may receive one treatment. That can leave little direct information about its outcomes for that profile and make subsequent model-based decisions more dependent on the assumptions of the outcome model.

The introductory material on updating model-based allocation probabilities establishes how a model can produce a probability for the next assignment. Here the focus is what to do when that probability becomes too close to zero or one for a design to retain meaningful randomization. A probability bound does not determine which treatment is better. It limits how far the assignment rule may move toward either treatment.

Definition: Consider two treatments, A and B. For participant $i$, let $X_i$ denote the participant's covariate profile and let $H_{i-1}$ denote the trial history available before assignment. Let $q_i$ be the model-based target probability of assigning treatment A. Given a prespecified exploration limit $\epsilon$, where $0 < \epsilon < 0.5$, the bounded assignment probability is $$ \pi_i=\min\{1-\epsilon,\max\{\epsilon,q_i\}\}. $$ Thus $\epsilon \le \pi_i \le 1-\epsilon$, and the probability of assigning B is $1-\pi_i$, also between $\epsilon$ and $1-\epsilon$.

The quantities $q_i$ and $\pi_i$ have different roles. The first is the probability suggested by the adaptive model; the second is the probability actually used by the randomization procedure. If $q_i$ is inside the permitted interval, bounding leaves it unchanged. If it lies below the lower bound or above the upper bound, the assigned probability is brought back to the nearest endpoint. This operation is called clipping.

For example, with $\epsilon=0.10$, the permitted range is from 0.10 to 0.90. A target of 0.68 remains 0.68, but targets of 0.03 and 0.97 become 0.10 and 0.90. The procedure still adapts within the allowed range; it does not require equal allocation.

Worked Example: Clipping a Model-Based Target

In a hypothetical two-treatment trial, the current model gives the next participant a target probability $q_i=0.96$ of receiving treatment A. The design uses $\epsilon=0.10$. Find the probability used for assignment and interpret it.

The lower bound is 0.10 and the upper bound is $1-0.10=0.90$. Apply the clipping rule:

\pi_i=\min\{0.90,\max\{0.10,0.96\}\} =\min\{0.90,0.96\}=0.90.

The randomization assigns A with probability 0.90 and B with probability $1-0.90=0.10$. The model's target is not used unchanged because it exceeds the permitted upper bound. If the target had instead been $q_i=0.74$, it would be inside the interval and the assignment probability would remain 0.74.

What the Bounds Guarantee

A lower bound on $\pi_i$ means that A retains at least probability $\epsilon$ for each eligible assignment. The upper bound means the same for B, because $1-\pi_i \ge \epsilon$. In particular, neither treatment is made impossible by the allocation rule for a participant whose assignment is governed by these bounds.

The guarantee applies conditionally at each assignment, even when the model target changes with trial history. Suppose $m$ participants with the same covariate profile arrive, and the assignment rule uses probabilities satisfying $\epsilon \le \pi_j \le 1-\epsilon$ for each of those $m$ assignments. Conditional on the history before each assignment, the chance of assigning B is at most $1-\epsilon$. Repeated conditioning therefore bounds the chance that all $m$ receive B:

P(\text{no A among the }m\text{ participants}) \le (1-\epsilon)^m.

The corresponding bound applies to observing no B. The argument does not require assignment probabilities to be constant; it requires the conditional probability bounds to hold at each assignment. For expected exposure, let $I_j$ equal 1 if participant $j$ receives A and 0 otherwise. Since $E(I_j\mid H_{j-1},X_j)=\pi_j$, the lower bound implies $E(I_j)\ge\epsilon$. Summing over the $m$ participants gives:

E\left(\sum_{j=1}^{m} I_j\right) =\sum_{j=1}^{m}E(I_j) \ge m\epsilon.

These results describe a minimum chance of exposure, not a promised count. Actual random assignments can still produce an uneven split or even no assignments to one treatment in a small profile. The probability bound can make such an event less likely than an unrestricted rule would, but it cannot rule it out.

Worked Example: Exposure Within a Repeated Profile

Suppose, hypothetically, 12 participants with a particular covariate profile enter a trial. At each of these assignments, the probability of A is at least 0.15, conditional on the trial history. Find a lower bound for the expected number assigned A and an upper bound for the probability that none receive A.

Let $I_j=1$ when participant $j$ receives A. For each of the 12 assignments, $E(I_j)\ge 0.15$. Therefore:

E\left(\sum_{j=1}^{12}I_j\right) =\sum_{j=1}^{12}E(I_j) \ge 12(0.15)=1.80.

The expected number assigned A is at least 1.80 participants. This is an expectation, not a claim that at least two participants must receive A in the realized trial.

For the probability of no A, at each step the conditional probability of assigning B is at most $1-0.15=0.85$. Hence:

P(\text{no A among the 12}) \le (0.85)^{12}=0.1422.

Thus the chance that all 12 receive B is at most about 14.22% under the stated conditional bounds. It is not necessarily exactly 14.22%, because the assignment probabilities may be larger than 0.15 and may vary with history.

Choosing and Applying a Bound

The value of $\epsilon$ is a design choice, not a universal constant. A larger value preserves more opportunity to assign the less-favored treatment, but it also restricts how strongly the adaptive model can influence allocation. A smaller value permits more extreme adaptation while offering a weaker exposure guarantee. The value should be specified before enrollment and considered alongside trial size, the expected frequency of covariate profiles, and the scientific and ethical constraints of the study.

A bound is particularly relevant when a profile is expected to occur only a few times. Even if the rule guarantees that each treatment has positive assignment probability, a rare profile may not have enough participants to produce useful treatment-specific information. For common profiles, a modest bound may still allow substantial exposure to both treatments. Thus, the same $\epsilon$ has different practical consequences depending on how many participants are expected in a profile.

Worked Example: How the Bound Changes a Small-Profile Risk

In a hypothetical trial, the model target for A in a particular profile is $q_i=0.02$. Consider 20 participants from that profile and compare bounds $\epsilon=0.05$ and $\epsilon=0.10$. For each choice, calculate the actual assignment probability, the expected number assigned A if the probability stays at its clipped value for these assignments, and the probability that none receive A.

With $\epsilon=0.05$, the permitted range is 0.05 to 0.95, so the target is clipped to $\pi_i=0.05$. If this probability is used for all 20 assignments, the expected number receiving A is:

$$ 20(0.05)=1.00. $$

The probability that none receive A is:

(1-0.05)^{20}=(0.95)^{20}=0.3585.

With $\epsilon=0.10$, the target is instead clipped to $\pi_i=0.10$. The expected number assigned A is $20(0.10)=2.00$, and the probability of no A is:

(1-0.10)^{20}=(0.90)^{20}=0.1216.

For these hypothetical fixed probabilities, the larger bound increases expected exposure to A and reduces the probability of observing no A. It also moves the allocation further from the model's target of 0.02. In an adaptive trial, probabilities may change after each outcome; the corresponding exposure and no-assignment statements then use the conditional bounds rather than assuming a fixed probability.

Clipping and Other Allocation Choices

Unrestricted model-based allocation uses $q_i$ directly. It can follow extreme targets, including probabilities close to zero or one, so it does not provide a positive minimum assignment chance for both treatments. Clipping is one way to retain a minimum chance while otherwise preserving the model target whenever it lies within the permitted range.

A different approach is an exploration mixture. For a two-treatment trial, one version is:

\pi_i^{\mathrm{mix}}=(1-\gamma)q_i+\gamma\frac{1}{2}, \qquad 0<\gamma\le 1.

This combines the model target with equal allocation. Since $0\le q_i\le1$, the mixture lies between $\gamma/2$ and $1-\gamma/2$. Unlike clipping, which leaves every target inside its interval unchanged, the mixture pulls all targets toward 0.50. For example, if $q_i=0.80$ and $\gamma=0.20$, then $\pi_i^{\mathrm{mix}}=0.80(0.80)+0.20(0.50)=0.74$. A clipping rule with bounds 0.10 and 0.90 would leave the same target at 0.80.

Neither approach is automatically preferable. Clipping directly enforces chosen endpoints and does not alter targets already inside them. A mixture gives a smooth pull toward equal allocation, with its minimum probability determined by $\gamma$. A conventional fixed-ratio randomization rule is another alternative, but it does not adapt allocation probabilities to the model target. The choice concerns the design's intended balance between adaptation and continued exposure, not a mathematical guarantee of a particular realized allocation.

Diagnostics and Limitations

A useful implementation check is to record, for each assignment, the model target $q_i$, the applied probability $\pi_i$, the covariate profile, and the treatment actually assigned. The records should confirm that the applied probability follows the stated rule and remains within the permitted range. They also show how often targets were clipped at either endpoint. Frequent clipping indicates that the bounds are materially shaping the design; it is not, by itself, evidence that the model is wrong.

Other warning signs include an applied probability of zero or one despite a stated positive $\epsilon$, a randomization system that uses $q_i$ rather than the clipped $\pi_i$, or a probability recorded after assignment rather than the probability used to make the assignment. A bounds check can verify implementation, but cannot establish that the outcome model is appropriate or that a profile will accrue enough participants.

Positive probability is not adequate sample size. A treatment can remain possible while still receiving very few participants in a rare profile.
Expected exposure is not guaranteed exposure. A lower bound on expected allocation does not impose a minimum realized count.
Bounds do not fix model problems. They restrict assignment probabilities but do not ensure that model-based targets are well estimated.
Bounds do not by themselves justify an analysis. The design and its assignment probabilities must still be represented appropriately in the analysis plan.

Key takeaway: Probability bounds preserve a specified minimum conditional chance of assigning either treatment at each eligible assignment. They limit extreme adaptive probabilities, but they do not guarantee exact balance, adequate sample size within every profile, or a correct outcome model.

Common Mistake: Treating a Bound as a Quota

A common mistake is to say, “With a lower bound of 0.10, at least 10% of the participants in each profile will receive A.” That conclusion confuses a conditional assignment probability with a realized proportion. The bound says that each eligible participant has probability at least 0.10 of receiving A, given the information available before that assignment. Random variation can still produce a realized proportion below 10%, including zero in a small profile. The correction is to describe the guarantee in terms of conditional probabilities and expected exposure, not as a guaranteed quota.

Check Your Understanding

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

A model target for A is 0.07 and the design uses $\epsilon=0.12$. What probability is used for A and what is the probability of B?
For 8 assignments, suppose the conditional probability of A is at least 0.20 at every step. Give a lower bound for the expected number assigned A and an upper bound for the probability that none receive A.
Explain why a lower bound of 0.10 does not guarantee that 10% of participants in a profile receive A.
For $q_i=0.80$, $\gamma=0.20$, and clipping bounds from 0.10 to 0.90, calculate the mixture probability and the clipped probability. State how they differ.

Solutions. (1) The target is below the lower bound, so $\pi_i=\max(0.12,0.07)=0.12$; the probability of B is $1-0.12=0.88$. (2) The expected number assigned A is at least $8(0.20)=1.60$. The probability that none receive A is at most $(1-0.20)^8=0.80^8=0.1678$. (3) The bound applies to each conditional assignment probability, not to the realized fraction. Random variation can yield fewer than 10% assigned A, especially when the profile is small. (4) The mixture probability is $(1-0.20)(0.80)+0.20(0.50)=0.74$. Since 0.80 is already within the clipping interval, the clipped probability is 0.80. The mixture moves the target toward equal allocation, whereas clipping leaves an in-range target unchanged.