Response Adaptation Within Covariate Strata

Why Adapt Within Strata?

A treatment can have different response rates in different patient groups. If a prognostic characteristic is known before randomization, such as baseline risk category, the trial can maintain response-adaptive randomization separately within each prespecified stratum. A participant’s allocation probability then depends on the outcomes observed for participants in that same stratum, rather than on an undifferentiated comparison across all participants.

This approach targets the treatment comparison relevant to the participant’s stratum. It also avoids allowing a large stratum to determine allocation for a smaller one. The trade-off is that each stratum has less data available for updating its probabilities, so estimates can be noisy. The method is an allocation rule, not a guarantee that the treatment is effective or that the trial will estimate every stratum-specific effect precisely.

For a two-arm trial, denote control by \(C\) and treatment by \(T\). Let \(k\) index the strata, and let \(D_k\) denote the accrued, usable outcome data in stratum \(k\) at the time of an allocation update. The probability assigned to treatment is \(\pi_{T,k}(D_k)\); the control probability is \(1-\pi_{T,k}(D_k)\). The participant’s stratum is determined before the random assignment, and the current rule is applied to that stratum.

A valid implementation specifies the strata, response definition, update schedule, data eligible for updates, and probability rule in advance. It also requires reliable stratum classification and consistent outcome measurement. For a binary outcome, a simple model assumes that, within each arm–stratum cell, responses are independent Bernoulli observations with a constant success probability \(\theta_{a,k}\). The probability may differ across arms and strata. This model does not assume that treatment has the same effect in every stratum.

Posterior Probability Matching Within Each Stratum

One transparent rule for two arms is posterior probability matching. Within stratum \(k\), calculate the posterior probability that the treatment success probability exceeds the control success probability. Assign that probability to treatment, with the remainder assigned to control. This is a defined allocation rule; it is not the only possible response-adaptive rule.

For a binary endpoint coded as success or failure, suppose the prior probabilities are independent, with \(\theta_{a,k}\sim\operatorname{Beta}(1,1)\) for each arm and stratum. If there are \(s_{a,k}\) successes among \(n_{a,k}\) participants with observed outcomes, the binomial likelihood and beta prior give

The posterior update is made separately in every cell. Data from another stratum do not enter this calculation under the independent, stratum-specific model. If no outcomes are yet available in a stratum, its two posteriors remain the priors, so the probability-matching rule assigns one-half to each arm there.

How the Overall Allocation Mix Is Determined

Stratum-specific probabilities do not usually equal the overall fraction of participants assigned to treatment. The overall fraction also depends on how many participants enter each stratum. If the probability of entering stratum \(k\) is \(w_k\), where the stratum weights sum to one, the expected overall treatment allocation probability is

This is an enrollment-weighted average, not generally a simple average across strata. The weights can change during a trial as the participant mix changes. Consequently, a change in the overall treatment allocation proportion can occur even when none of the stratum-specific rules has changed.

Sequential Updating and Checks

At each scheduled update, the trial uses the current eligible responses within each stratum to recalculate that stratum’s allocation probability. The randomization system then uses the probability for the incoming participant’s stratum. An outcome in stratum A changes the information state for A, not the posterior calculation for B, under the independent cell model described above. The update schedule and treatment of unavailable outcomes must be specified so that each assignment uses information available at that time.

Several checks help reveal when the design is not behaving as intended:

• Check cell sizes and response counts. A stratum with few participants can produce unstable, highly data-sensitive probabilities. A stratum with no completed outcomes remains at its prior allocation under probability matching.
• Check that every arm remains possible. If a probability reaches zero or one, later participants in that stratum cannot be assigned to the other arm. Prespecified allocation constraints can prevent this loss of positivity.
• Check classification and outcome coding. Misclassified strata or inconsistent success definitions undermine the meaning of the cell-specific comparisons.
• Track stratum-specific and overall assignments separately. A change in the overall treatment share may reflect changing stratum composition rather than a change in any within-stratum rule.
• Assess the design before use. Simulation under plausible response rates and enrollment mixes can expose excessive allocation imbalance, poor operating characteristics, or instability in small strata.

Alternatives, Limitations, and a Common Mistake

A fixed-ratio stratified design uses the same chosen allocation ratio within each stratum, regardless of observed responses. It avoids response-driven changes in allocation but does not direct more participants toward the arm currently favored by the response data. A pooled response-adaptive rule instead uses information across strata without separate cell-specific updates; it may be more stable when cells are small, but can obscure genuine differences between strata. A covariate-adjusted allocation approach uses a different modeling strategy; the present rule makes no joint covariate model or cross-stratum borrowing assumption.

The main limitation of separate adaptation is data fragmentation. Each stratum has its own arm-specific response information, so the total trial sample may still leave some cell posteriors imprecise. Separate rules can also result in sharply different assignment probabilities. These features should be considered in design-stage operating-characteristic assessment, alongside the clinical rationale for the strata and the practical ability to classify participants consistently.

Common mistake: “Treatment has a 95% allocation probability in the lower-risk stratum and 5% in the higher-risk stratum, so half of all participants will be assigned to treatment.” This does not follow. The overall expected treatment fraction depends on the proportions of participants in the strata, as well as the stratum-specific probabilities. Weight those probabilities by the enrollment mix; do not average them equally unless the strata contribute equally.

Response-adaptive probabilities describe how the trial assigns participants, not the probability that a treatment is effective in a clinical or causal sense. Posterior probabilities depend on the specified outcome model and prior, and adaptive assignment does not remove the need for an appropriate analysis of the treatment effects.