Dose Expansion Decisions After Escalation

From Escalation to Expansion

Escalation and expansion answer related but distinct questions. During escalation, investigators learn about tolerability across doses and identify doses worth considering. Expansion then gathers more information at a selected dose, often in a population or setting where activity can be assessed more precisely. The decision to expand should therefore consider both whether the dose appears acceptably safe and whether there is enough evidence of activity to justify further study.

A dose is not automatically suitable for expansion because it was reached safely, nor because it produced responses in a small escalation cohort. Both observations are uncertain, and the expansion decision depends on prespecified clinical boundaries, the amount and relevance of the available evidence, and whether outcomes have been adequately observed. The earlier material on Bayesian dose selection and overdose control provides context for how escalation choices are made; here the focus is the transition from a selected dose to expansion.

Definition: At dose \(d\), let \(p_d\) denote the probability that a treated patient experiences the protocol-defined dose-limiting toxicity (DLT) within its assessment window, and let \(q_d\) denote the probability of the protocol-defined response within its assessment window. A dose may proceed to expansion under a prespecified rule requiring adequate posterior evidence that \(p_d\) is below a clinically chosen toxicity boundary and that \(q_d\) is above a clinically chosen activity boundary.

These probabilities refer to clearly defined populations, treatment schedules, and outcome windows. A DLT observed during a longer escalation follow-up window is not automatically comparable with one observed during a shorter expansion window. Likewise, a response rate in a biomarker-selected expansion cohort need not estimate the response probability in the broader escalation population.

A Simple Posterior Decision Framework

For a transparent dose-specific analysis, suppose that at dose \(d\), \(x_d\) of \(n_d\) patients have a DLT and \(y_d\) of \(m_d\) evaluable patients respond. Conditional on the dose-specific probabilities, model the outcomes as binomial. A convenient starting point is independent beta priors for the toxicity and response probabilities. For illustration, use uniform \(\operatorname{Beta}(1,1)\) priors; a real trial should justify its priors and assess reasonable alternatives.

X_d\mid p_d \sim \operatorname{Binomial}(n_d,p_d), \qquad Y_d\mid q_d \sim \operatorname{Binomial}(m_d,q_d), \qquad p_d,q_d\sim\operatorname{Beta}(1,1).

This model assumes that, within each outcome analysis, patients have the same underlying probability at the dose being assessed; conditional outcomes are independent; the enrolled patients represent the target population; and outcomes are classified consistently. It also assumes that patients included in each denominator have completed the relevant assessment window, or that an appropriate method handles incomplete follow-up. The independent priors and likelihoods imply separate posterior distributions for safety and response. This is a simplifying assumption, not a claim that toxicity and response are biologically unrelated.

Beta-binomial conjugacy gives direct posterior updates. With \(x_d\) DLTs among \(n_d\) patients, the posterior for toxicity is \(\operatorname{Beta}(1+x_d,1+n_d-x_d)\). With \(y_d\) responses among \(m_d\) patients, the response posterior is \(\operatorname{Beta}(1+y_d,1+m_d-y_d)\). These distributions quantify uncertainty in the dose-specific probabilities, conditional on the model and data.

p_d\mid \text{DLT data}\sim\operatorname{Beta}(1+x_d,1+n_d-x_d), \qquad q_d\mid \text{response data}\sim\operatorname{Beta}(1+y_d,1+m_d-y_d).

To see why, the binomial likelihood for toxicity is proportional to \(p_d^{x_d}(1-p_d)^{n_d-x_d}\). Multiplying by the \(\operatorname{Beta}(a,b)\) prior density, proportional to \(p_d^{a-1}(1-p_d)^{b-1}\), gives a posterior density proportional to \(p_d^{a+x_d-1}(1-p_d)^{b+n_d-x_d-1}\). This is the kernel of a \(\operatorname{Beta}(a+x_d,b+n_d-x_d)\) distribution. The response update follows by the same multiplication using its response count.

Let \(\phi\) be the maximum acceptable DLT probability, \(q_{\min}\) the minimum activity probability that would make expansion worthwhile, and let \(\gamma_S\) and \(\gamma_E\) be chosen posterior probability thresholds. A two-gate decision rule is:

P(p_d<\phi\mid\text{data})\geq\gamma_S \quad\text{and}\quad P(q_d>q_{\min}\mid\text{data})\geq\gamma_E.

The thresholds and boundaries are trial-specific choices, not universal constants. The rule says to proceed only if both conditions are met. It does not say that the dose is risk-free, effective, or proven superior to another dose. When the two posterior probabilities are calculated separately, their joint probability can be obtained as their product only under the stated posterior independence assumption. A decision rule requiring a joint probability should specify that requirement explicitly rather than silently substituting two separate gates.

Worked Decisions

Worked Example: Choosing a Dose After Escalation

Hypothetical data. Three doses have completed the DLT assessment window. At 2 mg, 0 of 6 patients had a DLT and 1 of 6 responded. At 4 mg, 1 of 6 had a DLT and 3 of 6 responded. At 6 mg, 3 of 6 had a DLT and 4 of 6 responded. Assume comparable patients and outcome definitions across doses, and use independent \(\operatorname{Beta}(1,1)\) priors. Suppose the protocol requires \(P(p_d<0.35\mid\text{data})\geq0.70\) and \(P(q_d>0.20\mid\text{data})\geq0.80\).

At 4 mg, the toxicity posterior is \(\operatorname{Beta}(2,6)\). For a \(\operatorname{Beta}(2,6)\) random variable, the cumulative probability at \(z\) is \(1-(1-z)^6(1+6z)\). Thus:

P(p_{4\text{ mg}}<0.35\mid\text{data}) =1-(0.65)^6(1+6(0.35)) =1-(0.07542)(3.10) =0.7662.

The safety gate passes because \(0.7662\geq0.70\). The response posterior is \(\operatorname{Beta}(4,4)\). Its cumulative probability at \(z\) is \(35z^4-84z^5+70z^6-20z^7\), so:

P(q_{4\text{ mg}}>0.20\mid\text{data}) =1-\left[35(0.20)^4-84(0.20)^5+70(0.20)^6-20(0.20)^7\right] =1-0.03334 =0.9667.

The activity gate also passes. At 2 mg, the response posterior is \(\operatorname{Beta}(2,6)\), giving \(P(q_{2\text{ mg}}>0.20\mid\text{data})=0.5767\), below the activity threshold. At 6 mg, the toxicity posterior is \(\operatorname{Beta}(4,4)\), giving \(P(p_{6\text{ mg}}<0.35\mid\text{data})\approx0.200\), below the safety threshold. Under these illustrative rules, 4 mg is eligible to proceed to expansion. This is a decision under uncertainty, not proof that 4 mg is the best dose or that its true DLT probability is below 0.35.

Worked Example: Updating With Expansion Data

Suppose the hypothetical expansion cohort at 4 mg enrolls 12 additional patients from the same target population, on the same schedule, with complete and comparable assessment. There are 0 DLTs and 6 responses in these 12 patients. Combining them with the escalation data gives 1 DLT among 18 patients and 9 responses among 18 patients. Under the exchangeability assumptions, the updated posteriors are \(\operatorname{Beta}(2,18)\) for toxicity and \(\operatorname{Beta}(10,10)\) for response.

For toxicity, the cumulative probability for a \(\operatorname{Beta}(2,18)\) variable at \(z\) is \(1-(1-z)^{18}(1+18z)\). Therefore:

P(p_{4\text{ mg}}<0.35\mid\text{all data}) =1-(0.65)^{18}(1+18(0.35)) \approx1-(0.000429)(7.30) =0.9969.

For response, the \(\operatorname{Beta}(10,10)\) posterior gives \(P(q_{4\text{ mg}}>0.20\mid\text{all data})\approx0.9984\). Both probabilities exceed the example thresholds. The additional cohort has strengthened the evidence at this dose, conditional on the model and on combining the cohorts being defensible. The result supports proceeding under the stated rule; it does not establish that the treatment causes responses, since this single-arm rate does not by itself provide a randomized comparison.

Worked Example: Expansion Patients Do Not Match the Escalation Population

Consider a different hypothetical scenario. The escalation cohort at 4 mg had 3 responses among 6 patients with a broad tumor-type mix. The proposed expansion is restricted to a biomarker-defined subgroup, and 1 of its first 12 patients responds. The expansion patients also have 1 DLT among 12. Suppose the decision is whether to continue enrolling this subgroup, using the same illustrative boundaries \(\phi=0.35\), \(q_{\min}=0.20\), \(\gamma_S=0.70\), and \(\gamma_E=0.80\).

Because the expansion targets a different population, use its outcomes for this subgroup-specific decision rather than pooling the escalation response count as if the patients were exchangeable. With the uniform prior, the subgroup response posterior is \(\operatorname{Beta}(2,12)\), and:

P(q_{4\text{ mg}}>0.20\mid\text{subgroup data}) =1-\left[1-(1-0.20)^{12}(1+12(0.20))\right] =0.2336.

The subgroup safety posterior is also \(\operatorname{Beta}(2,12)\), so:

P(p_{4\text{ mg}}<0.35\mid\text{subgroup data}) =1-(0.65)^{12}(1+12(0.35)) =0.9704.

Safety passes, but activity does not. Under the specified rule, the evidence does not support continuing expansion in this subgroup. The result does not prove that the dose is inactive; it indicates that the observed subgroup data do not meet the prespecified posterior activity requirement. Pooling the earlier responses without accounting for the population change could produce a more encouraging number while answering the wrong clinical question.

Operational Checks and Alternatives

A posterior gate is only as reliable as the data and assumptions that feed it. Before making the transition decision, confirm that DLT follow-up is complete or appropriately handled, response assessments use the planned criteria and time point, and each denominator is clear. If responses are delayed, treating patients whose response windows are incomplete as nonresponders can bias the apparent activity downward. Conversely, excluding patients with incomplete assessments without a defensible rule can distort the rate. The analysis plan should specify how such patients enter the likelihood.

Check whether the escalation and expansion cohorts can reasonably share a probability parameter. Differences in eligibility, disease subtype, prior treatment, dose schedule, supportive care, or calendar time may make direct pooling inappropriate. If clinically meaningful differences are expected, analyze the target expansion population separately or use a model that represents the differences explicitly. Do not add counts simply because both cohorts received the same nominal dose.

The beta-binomial calculation is a dose-by-dose alternative to a model that borrows information across doses, such as a logistic dose-toxicity model. The simple model is easy to audit but does not impose a monotone increase in toxicity with dose and may be unstable with very small cohorts. A model that shares information across doses can be more efficient when its structural assumptions are credible, but its conclusions can be sensitive to model specification. For response, a fixed posterior gate is distinct from choosing a dose by an overall benefit-risk utility: a utility approach can trade a higher response rate against greater toxicity, whereas the two-gate rule requires each condition to pass independently.

Useful diagnostics include examining the posterior distributions, not just their means; comparing results under plausible alternative priors; checking whether conclusions change when non-comparable cohorts are excluded; and reviewing the outcome counts and follow-up by cohort. A posterior mean below the toxicity boundary does not guarantee that the safety gate passes: the probability mass above the boundary also matters. Likewise, a promising response proportion in a small sample may still leave substantial posterior probability below the activity boundary.

No single posterior threshold removes clinical judgment. The boundary \(\phi\) should reflect the toxicity considered unacceptable at the dose and in the disease context; \(q_{\min}\) should represent an activity level worth pursuing, not simply the observed rate. Thresholds chosen after seeing the data invite optimistic decisions. Finally, an expansion decision is a design action, not a final efficacy conclusion. It remains conditional on the target population, endpoint definitions, priors, model, and decision rules.

Common Pitfall

Common mistake: “The escalation dose had a response rate above the activity target, so we should pool every expansion patient with it and proceed.”

Correction: A pooled posterior is appropriate only when the cohorts estimate the same dose-specific probabilities under comparable conditions, or when the model explicitly accounts for their differences. If expansion changes the target population, evaluate that population with relevant data and state the resulting uncertainty. Also require the prespecified safety evidence; activity alone does not satisfy a two-gate rule.

Check Your Understanding

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

With a uniform prior, 2 DLTs among 10 patients give which posterior distribution for \(p_d\)?
For a candidate dose, the posterior safety probability exceeds its threshold but the posterior activity probability does not. What does the two-gate rule recommend?
Why might a response observed in a broad escalation cohort not be pooled with responses from a biomarker-defined expansion cohort?
With a uniform prior, 0 responses among 8 patients give what posterior distribution for \(q_d\), and what is the posterior probability that \(q_d>0.20\)?

Solutions. (1) The posterior is \(\operatorname{Beta}(1+2,1+10-2)=\operatorname{Beta}(3,9)\). (2) The dose does not pass both required gates, so the rule does not recommend proceeding on the basis of these data; this is not proof of no activity. (3) The populations may have different response probabilities, so unadjusted pooling would treat non-exchangeable patients as if they shared one rate and could mislead the subgroup decision. (4) The posterior is \(\operatorname{Beta}(1,9)\). Since its cumulative probability at 0.20 is \(1-(0.80)^9=0.8658\), \(P(q_d>0.20\mid\text{data})=(0.80)^9=0.1342\).