Joint Model-Based Escalation Decisions

Why Joint Posterior Decisions Need a Safety Gate

At an interim in dose finding, the decision is not simply whether the next dose has a higher estimated benefit rate. A higher dose may also raise toxicity, and the evidence about both outcomes is uncertain. A joint model represents benefit and toxicity together; a decision rule then uses that model to determine whether to escalate, de-escalate, or retain the current dose. The essential discipline is to screen every dose that could be assigned against a prespecified safety requirement before comparing its benefit-risk value.

This tutorial develops one transparent, local decision framework. It is not a single universally prescribed escalation algorithm: its safety threshold, utility values, and decision margin must be justified and specified before the trial. The posterior is updated using the joint outcome likelihood; the earlier material on likelihood construction for joint dose models gives the relevant likelihood context.

After interim data \(D\), the model yields a posterior distribution for \(\theta\). Choose a maximum acceptable toxicity probability \(\tau\) and a posterior confidence requirement \(\gamma\), with \(0<\gamma<1\). A dose passes the safety screen when the posterior probability that its toxicity probability is no greater than \(\tau\) reaches \(\gamma\). This is a statement about uncertainty in the dose-level toxicity rate, not the probability that a particular future participant will experience toxicity.

The rule is applied at every decision interim using the available data and the prespecified model. For example, a protocol might set \(\tau=0.30\) and \(\gamma=0.90\). A dose with \(q_d=0.89\) fails that rule, even if its posterior mean toxicity probability is below 0.30. These cutoffs are design choices, not quantities that the data alone determine.

Turning Joint Outcomes into a Dose Value

A safety screen alone does not rank doses that pass it. One way to compare them is to assign a prespecified utility to each joint outcome. The utility encodes the trial’s benefit-risk priorities: benefit without toxicity may be desirable, toxicity without benefit may be undesirable, and the joint occurrence of benefit and toxicity requires an explicit value rather than an assumption that it is equivalent to either outcome alone.

For a simple illustrative utility scale, suppose the values are \(U(0,0)=0\), \(U(1,0)=1\), \(U(0,1)=-1\), and \(U(1,1)=0.2\). The numbers are hypothetical and have no physical units. Under this scale, a participant with both benefit and toxicity has positive but substantially lower utility than a participant with benefit and no toxicity. A protocol may use a different scale, but it should define the scale before outcomes are examined.

Here \(\bar U(d;\theta)\) is the model-implied mean utility at dose \(d\) for a fixed parameter value, and \(\widetilde U_d\) averages that quantity over posterior uncertainty. Because utility is a fixed function of the outcome state, the posterior expected utility can also be calculated by applying the utility values to the posterior mean joint probabilities.

This equality follows by linearity of expectation: the utility values are constants, so \(E\{\sum U(b,t)p_{bt}\mid D\}=\sum U(b,t)E\{p_{bt}\mid D\}\). It is useful computationally, but it does not remove the need to account for posterior uncertainty in the safety screen. A favorable posterior mean utility cannot make an unsafe dose admissible.

A Local Rule for Escalation, De-escalation, and Retention

Suppose doses are ordered \(d_1<d_2<\cdots<d_K\), and the current dose is \(d_j\). For a simple local rule, consider the current dose and its immediately adjacent doses, when they exist. Let \(\epsilon\geq0\) be a prespecified utility margin. It prevents a very small estimated advantage from automatically triggering a dose change. The decision proceeds in this order:

This rule deliberately separates safety from preference: safety determines which doses are eligible, and utility ranks eligible options. A local comparison limits the size of a routine dose change; it does not establish that the chosen dose is globally optimal. If the model permits nonmonotone dose-response patterns, the safety screen must be calculated dose by dose rather than inferred from a neighboring dose. If a model imposes monotonicity, that assumption should be clinically and statistically defensible.

Assumptions, Diagnostics, and Alternatives

These decisions are meaningful only if the model and interim data support the quantities being used. The benefit and toxicity criteria, assessment window, and coding of joint outcomes must be prespecified and applied consistently. At the interim, outcomes used in the likelihood must be sufficiently ascertained under the trial’s analysis plan; treating an unresolved outcome as if it were a known absence can distort both the joint probabilities and the decision. Participants’ outcomes are ordinarily modeled as conditionally independent given their dose and the model parameters, unless the model explicitly accounts for dependence between participants. Dependence between benefit and toxicity within a participant must be represented appropriately by the joint model rather than removed by assumption without justification.

Several checks help reveal when the procedure is becoming unreliable. Inspect posterior predictive fit for the joint outcome categories, not only the marginal benefit and toxicity rates. Check whether the posterior is dominated by prior information at sparsely observed doses, whether estimates at adjacent doses change sharply without a plausible explanation, and whether the posterior safety probability is close to its cutoff. Recalculate decisions under clinically plausible alternatives for \(\tau\), \(\gamma\), utility values, and \(\epsilon\). A dose recommendation that changes under small reasonable changes deserves caution. Such checks diagnose sensitivity; they do not replace a prespecified decision rule.

A count-based escalation scheme is simpler and can be easier to implement, but it may not use the joint benefit-toxicity pattern or model information across doses. A rule based only on posterior expected utility uses both outcomes but can select a dose with an unacceptably high posterior probability of exceeding the toxicity limit. The safety-constrained approach avoids that particular failure by screening first, at the cost of relying on a chosen posterior cutoff and utility scale. It also depends on the model: misspecification, sparse data, or unjustified dose-response assumptions can lead to poor decisions even when the arithmetic is correct.