Why Adjust Allocation for Covariates?
Patients entering a clinical trial can differ in baseline characteristics that predict response. If treatment effects also vary with those characteristics, a single allocation probability based on each arm’s overall response rate can hide clinically relevant differences. A covariate-adjusted rule instead asks: for this incoming patient, which arm has the greatest posterior probability of producing the best response?
The rule developed here uses the posterior distribution of each arm’s predicted response at the incoming patient’s covariate values. It is a patient-specific version of posterior probability matching. It does not assign an arm deterministically: it randomizes among arms, with probabilities determined by how often each arm is predicted to be best across posterior draws. The introductory material on Response Adaptation Within Covariate Strata considers adaptation by groups; the method here evaluates the incoming patient’s covariate profile directly.
Model and Allocation Rule
Suppose there are \(K\) available treatment arms. Let \(A_i\) denote the assigned arm for patient \(i\), let \(X_i=x\) denote that patient’s baseline covariates, and let \(Y_i\) be a binary response, with 1 representing response and 0 no response. Let \(D_t\) be the trial information available when patient \(i\) is randomized, including the outcomes observed by that time. For arm \(a\), define
where \(\theta\) contains the model parameters. The posterior distribution \(p(\theta\mid D_t)\) expresses uncertainty about those parameters after the available trial data are considered. The posterior predicted response for arm \(a\) at covariate value \(x\) is the random quantity \(p_a(x,\theta)\) as \(\theta\) varies over this posterior.
This rule compares arms at the same covariate values. For example, if higher baseline disease severity predicts a different relative response to treatment, the predictions for a high-severity patient need not have the same ranking as predictions for a low-severity patient. The probabilities sum to one, provided every posterior draw assigns its full allocation share among the arms tied for the maximum.
A common model for binary outcomes is logistic regression with treatment-by-covariate interactions. For example, with one covariate \(x\), one possible specification is \(\operatorname{logit}\{p_a(x,\theta)\}=\beta_0+\beta_a+(\gamma+\delta_a)x\), with a reference arm’s treatment terms set to zero. The interaction terms permit relative response to vary with \(x\). This is a model choice, not a requirement of the allocation rule; whatever model is used must produce coherent arm-specific predictions and account for uncertainty in them.
Conditions and Calculation
The interpretation of \(p_a(x,\theta)\) as an arm-specific response probability relies on a suitable outcome model and trial conduct. Covariates used in allocation must be measured before assignment. The randomized assignment mechanism must be implemented as specified, and the response definition and follow-up must be comparable across arms. A causal interpretation of arm contrasts additionally relies on standard randomized-trial conditions, including consistency of treatment assignment, no interference between patients, and adequate handling of missing outcomes. The model should represent relevant treatment-covariate variation; a covariate-adjusted rule cannot repair a poorly specified prediction model.
In practice, posterior simulation provides a direct calculation. For each posterior draw \(\theta^{(m)}\), compute all \(p_a(x,\theta^{(m)})\), identify the arm or arms with the largest prediction, and give each tied maximum a predeclared share. The average share over draws estimates \(q_a(x\mid D_t)\). This is a Monte Carlo approximation to a posterior probability, so its precision depends on the number and quality of posterior draws.
Record \(x\) before randomization, using the same definitions and coding used to fit the response model.
For each posterior draw, evaluate \(p_a(x,\theta)\) for all currently available arms.
Give the maximal arm a share of 1 when unique; divide the share according to the prespecified tie rule when the maximum is shared.
Average each arm’s shares across draws to obtain allocation probabilities, then randomize the patient using those probabilities, subject to any trial-wide constraints.
Worked Calculations
Worked Example: Three Arms for One Patient
Suppose a hypothetical trial has three arms, A, B, and C, and the incoming patient has covariates \(x\). Four equally weighted posterior draws give these predicted response probabilities:
| Posterior draw | Arm A | Arm B | Arm C | Predicted best |
|---|---|---|---|---|
| 1 | 0.72 | 0.66 | 0.61 | A |
| 2 | 0.58 | 0.70 | 0.64 | B |
| 3 | 0.63 | 0.67 | 0.71 | C |
| 4 | 0.68 | 0.69 | 0.62 | B |
There are no ties in these four draws. The estimated allocation probability is the number of draws in which an arm is best divided by four:
Thus this rule randomizes the patient to A, B, or C with probabilities 0.25, 0.50, and 0.25, respectively. The probabilities summarize uncertainty about which arm has the highest predicted response for this patient; they are not the posterior mean response probabilities themselves.
Worked Example: Conditional and Marginal Rankings Differ
Consider two hypothetical arms, A and B, and a binary baseline covariate with high- and low-severity groups. Assume the enrolled population is equally divided between the groups. Four equally weighted posterior draws give the following predictions:
| Draw | A, high | B, high | A, low | B, low |
|---|---|---|---|---|
| 1 | 0.70 | 0.80 | 0.90 | 0.60 |
| 2 | 0.65 | 0.75 | 0.85 | 0.55 |
| 3 | 0.60 | 0.70 | 0.80 | 0.50 |
| 4 | 0.55 | 0.65 | 0.75 | 0.45 |
For a high-severity patient, B is best in all four draws, so the conditional posterior matching probabilities are \(q_A(\text{high})=0/4=0\) and \(q_B(\text{high})=4/4=1\). For a low-severity patient, A is best in all four draws, giving \(q_A(\text{low})=1\) and \(q_B(\text{low})=0\).
Now average each arm’s predicted response over the two equally common groups in each draw. In draw 1, for example, A’s marginal prediction is \((0.70+0.90)/2=0.80\), while B’s is \((0.80+0.60)/2=0.70\). Applying that calculation to all draws gives:
A marginal probability-matching rule would select A as best in all four draws, yielding allocation probabilities 1 for A and 0 for B, regardless of the incoming patient’s severity. The covariate-adjusted rule instead selects B for a high-severity patient. This is not a contradiction: the two rules answer different questions. The marginal calculation averages over a specified covariate distribution; the adjusted calculation compares predictions at this patient’s covariates.
Worked Example: A Positive Probability of Ties
Suppose two arms have posterior predicted responses for a particular patient as follows across three equally weighted draws: draw 1 gives A = 0.70 and B = 0.70; draw 2 gives A = 0.80 and B = 0.72; draw 3 gives A = 0.60 and B = 0.65. The first draw is an exact tie. Prespecify that a tied maximum is split equally between the tied arms.
The allocation share for A is 0.5, 1, and 0 across the three draws. For B it is 0.5, 0, and 1. Therefore:
The tie rule matters: assigning the whole tied draw to A by a fixed priority rule would instead give A probability \(2/3\) and B probability \(1/3\). Continuous posterior parameters alone do not rule out ties in predicted responses. For example, if two arms have identical prediction functions, their predictions tie for every parameter draw. A zero-tie calculation is justified only when, for each relevant pair of arms, the posterior probability that their predicted-response difference is exactly zero is zero. Otherwise, state and apply a tie rule.
Justification and Practical Checks
For a unique maximum, define \(I_a(\theta,x)\) to equal 1 when arm \(a\) has the largest predicted response at \(x\), and 0 otherwise. With no posterior mass on pairwise ties, exactly one arm has indicator 1 for every draw. Consequently, \(\sum_{a=1}^K I_a(\theta,x)=1\). Taking the posterior expectation gives:
When ties occur, replace the indicator with the arm’s share under the tie rule. If those shares sum to 1 for every draw, the same expectation argument proves that the allocation probabilities sum to 1. This also explains why tie handling is part of the allocation procedure, not a cosmetic detail.
Before using the rule, inspect whether the incoming patient’s covariates are represented adequately in the data and model. Predictions for sparse or unsupported covariate profiles may be dominated by model assumptions rather than trial evidence. Check posterior computation and predictive behavior, including whether predictions are plausible and whether treatment-by-covariate patterns are too weakly identified to support the intended distinctions. If the model does not allow treatment effects to vary with a covariate, it cannot discover such variation through this rule.
Allocation probabilities are operational probabilities, not guarantees of benefit. In a trial, minimum allocation requirements, caps, or a protected control probability can modify the raw probabilities; the implemented probabilities must be documented and used in the randomization system. Record covariates, the information available at each update, and the probabilities actually used. Randomization probabilities do not replace an appropriate analysis of the trial outcomes.
Alternatives, Limitations, and a Common Pitfall
| Approach | What is compared | Key distinction |
|---|---|---|
| Covariate-adjusted posterior matching | Posterior probability each arm is best at the incoming patient’s \(x\) | Responds to patient-specific predicted rankings |
| Marginal response matching | Posterior probability each arm is best after averaging over a covariate distribution | Uses a population-level summary, not the individual profile |
| Posterior expected utility | Posterior mean of a specified utility for each arm at \(x\) | Ranks posterior averages rather than counting which arm is best in each draw |
| Softmax allocation | A transformation of arm scores into probabilities | Probability depends on the chosen score and temperature, rather than only the probability of being best |
Posterior probability matching treats each posterior draw’s best arm equally, whether its advantage over the alternatives is large or small. By contrast, expected utility can favor an arm with a modest but consistent advantage over one that is best often but performs poorly in some posterior draws. Neither criterion is universally preferable; the allocation objective should be prespecified and clinically defensible. The patient-specific rule also depends on covariates being available before randomization and may be unstable when subgroup information is limited.
Common mistake: “Arm A has the largest overall response rate, so every incoming patient should receive A with the largest probability.” This substitutes a marginal comparison for a conditional one. If response rankings vary across covariate profiles, the overall rate can conceal that variation. The correction is to calculate the posterior predicted response for every arm at the incoming patient’s covariates, apply the prespecified tie rule, and average the resulting best-arm shares over posterior draws.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- Across five equally weighted draws, an arm is uniquely best twice and shares a tied maximum equally with one other arm in two draws; it is not best in the fifth. What is its estimated allocation probability?
- Why can a continuous posterior distribution for model parameters still produce a positive probability of tied predicted responses?
- A marginal rule averages predictions over a population with 70% low-severity and 30% high-severity patients. What additional information is needed to use a covariate-adjusted rule for a new patient?
- In one sentence, distinguish posterior probability matching from posterior expected utility.
Solutions. (1) Its total share is \(2+2(0.5)=3\), so the estimated probability is \(3/5=0.60\). (2) Two prediction functions can be identical, or otherwise equal on a set with positive posterior probability, so parameter continuity alone does not ensure a zero probability of a prediction tie. (3) The patient’s pre-randomization covariates and posterior arm-specific predictions evaluated at those values are needed; the population mixture alone does not identify that patient-specific comparison. (4) Probability matching counts how often each arm is best across posterior draws, whereas expected utility ranks arms by their posterior average utility.