Why Constrain Escalation by Posterior Risk?
In a combination trial, each candidate dose is a pair, and toxicity risk can depend on both components and their interaction. A model can estimate risk at each pair, but an estimate alone does not say how much uncertainty is acceptable before escalation. A posterior safety constraint makes that policy explicit: a candidate is eligible only when the posterior probability that its toxicity risk exceeds a chosen target is sufficiently small.
The constraint is a screening rule, not a complete dose-selection algorithm. It identifies which possible moves meet a specified safety standard; trial rules and any prespecified criteria for choosing among eligible moves still matter. This distinction is particularly important when two dose pairs have similar estimated risk.
This rule concerns uncertainty about the pair’s underlying DLT probability. It is not the same as requiring the posterior mean \(E(p_{jk}\mid D)\) to be below \(\phi\). A pair can have a mean below the target yet retain substantial posterior probability above it. Conversely, passing the constraint does not mean that toxicity cannot exceed the target; it means that, under the fitted model and current data, the posterior probability of exceeding it is below the chosen cutoff.
From Posterior Draws to an Admissible Move
Suppose the combination toxicity model provides posterior draws of the risk at each dose pair. The posterior safety probability is the posterior mass for which the draw exceeds the target. With \(S\) independent draws \(p_{jk}^{(1)},\ldots,p_{jk}^{(S)}\), estimate it by counting exceedances:
Here \(I(\cdot)\) equals 1 when its condition is true and 0 otherwise. For correlated Markov chain Monte Carlo (MCMC) draws, the same fraction estimates posterior mass, but its simulation precision depends on effective sample size rather than simply on the number of stored draws.
For an escalation decision, first identify which pairs are permitted next steps from the current pair. Partial ordering of combinations and the trial protocol determine this set; a safety calculation does not make a clinically or operationally prohibited move permissible. Then apply the constraint to each candidate separately. A useful formalization is:
where \((j_0,k_0)\) is the current pair and \(\mathcal{N}(j_0,k_0)\) is the set of protocol-permitted next moves. The set \(\mathcal{A}\) contains the admissible moves. If it is empty, the rule supplies no escalation move. Whether to stay, de-escalate, or stop must be determined by the trial’s other prespecified rules.
Worked Decisions
Worked Example: One Candidate Passes the Constraint
Consider a hypothetical trial with target \(\phi=0.30\) and posterior cutoff \(c=0.25\). The participant is currently at pair \((1,1)\), and the protocol permits either \((2,1)\) or \((1,2)\) as the next move. From 2,000 posterior draws, risk at \((2,1)\) exceeds 0.30 in 420 draws; risk at \((1,2)\) exceeds 0.30 in 560 draws. Assume the posterior model and its draws have passed the trial’s planned computational checks.
For \((2,1)\), \( \widehat q_{21}=420/2000=0.210 \), or 21.0%. Since \(0.210<0.25\), this pair passes the constraint.
For \((1,2)\), \( \widehat q_{12}=560/2000=0.280 \), or 28.0%. Since \(0.280\) is not less than \(0.25\), this pair does not pass.
Thus the admissible set contains \((2,1)\) but not \((1,2)\). The result is conditional on the stated target, cutoff, model, and allowed-move set; it does not establish that \((2,1)\) is free of excessive toxicity.
Worked Example: Two Similarly Risky Moves Pass
In another hypothetical decision, the current pair is \((2,1)\). The protocol permits \((3,1)\) and \((2,2)\), and no other next steps. With \(\phi=0.30\) and \(c=0.25\), 460 of 2,000 posterior draws exceed the target at \((3,1)\), while 480 of 2,000 exceed it at \((2,2)\).
The estimates are \( \widehat q_{31}=460/2000=0.230 \) and \( \widehat q_{22}=480/2000=0.240 \). Both are below 0.25, so both pairs are admissible. Their estimated posterior overdose probabilities differ by only \(0.010\), or one percentage point.
Suppose the protocol prespecified choosing the admissible pair with the smaller posterior overdose probability when two candidates are otherwise eligible. Under that rule, \((3,1)\) is selected. The calculation does not show that this is a uniquely superior dose: the difference is small, and a different prespecified rule, such as a clinical preference among admissible combinations, could select differently. If the cutoff had instead been \(c=0.20\), neither move would pass because \(0.230\) and \(0.240\) both exceed 0.20. The cutoff therefore has a direct operational effect and must not be chosen after inspecting the candidates merely to permit escalation.
Worked Example: No Candidate Meets the Safety Rule
Suppose a participant is at \((3,2)\), and the only permitted escalation moves are \((4,2)\) and \((3,3)\). In a hypothetical analysis with \(\phi=0.25\) and \(c=0.20\), 330 of 1,000 posterior draws exceed the target at \((4,2)\); 310 of 1,000 exceed it at \((3,3)\).
The estimated probabilities are \( \widehat q_{42}=330/1000=0.330 \) and \( \widehat q_{33}=310/1000=0.310 \). Since neither is below 0.20, the admissible escalation set is empty. Although \((3,3)\) has the smaller estimate, it still fails the constraint. The rule therefore does not justify escalating to the “less risky” option. The protocol must determine the appropriate response, such as remaining at the current pair, de-escalating, or stopping.
Why the Posterior Constraint Works—and What It Does Not Prove
The calculation follows directly from the definition of posterior probability. For a candidate pair, the event of concern is \(p_{jk}>\phi\). Its posterior probability is the posterior expectation of the indicator of that event:
A posterior sample approximates this expectation by averaging the indicator over draws, yielding the exceedance fraction above. This justifies using the fraction to estimate posterior mass; it does not remove uncertainty due to finite simulation or uncertainty due to model assumptions. In particular, \(q_{jk}<c\) is a rule about the posterior distribution conditional on the model and data. It is not a frequentist guarantee that the true toxicity probability is below \(\phi\) with probability \(1-c\).
For independent posterior draws, the indicators are Bernoulli with success probability \(q_{jk}\), so the approximate Monte Carlo standard error of the estimated fraction is:
For MCMC output, autocorrelation reduces effective information; use an effective sample size or another suitable MCMC error estimate rather than treating every stored draw as independent. If the estimated probability is close to \(c\), simulation error can affect which side of the operational boundary the estimate falls on.
Worked Example: Checking Monte Carlo Precision Near the Cutoff
For the estimate \(\widehat q_{31}=0.230\) from 2,000 approximately independent draws in the second example, the approximate Monte Carlo standard error is \( \sqrt{0.230(1-0.230)/2000}=\sqrt{0.00008855}=0.00941 \). This is about 0.94 percentage points.
The estimate is 0.020 below the cutoff of 0.25, roughly 2.13 such standard errors. This calculation describes simulation precision, not the posterior uncertainty in the underlying toxicity probability: that uncertainty is already represented by the posterior draws used to calculate \(q_{31}\). If the chains were autocorrelated, the effective sample size could be smaller and the Monte Carlo standard error larger. A decision very close to the cutoff warrants checking simulation precision and following any prespecified boundary convention; it does not justify silently changing the cutoff.
Assumptions, Diagnostics, and Alternatives
A posterior constraint is interpretable only in relation to a defined endpoint and a credible model. At minimum, specify the DLT definition and assessment window, the target \(\phi\), the cutoff \(c\), and the allowable transition set before applying the rule. The posterior distribution must reflect the trial data and the combination toxicity model’s assumptions, including how dose-pair risks and any interaction are represented. The Bayesian copula model supplies one way to obtain joint-combination risk; the constraint is applied to its posterior risk draws rather than replacing that model.
Before acting on the calculated probabilities, check that the posterior computation is reliable and that the model’s risk estimates are credible for the decision at hand. Relevant checks include chain mixing and convergence, effective sample sizes for risk summaries, and Monte Carlo error near \(c\). Review whether the fitted model reproduces important toxicity patterns in the observed data and whether posterior risk behaves plausibly across the combination grid. Sparse observations at a pair, a strong dependence on prior assumptions, or unexpected non-monotone risk patterns can make a precise-looking probability misleading. If model fit or computation is inadequate, a threshold comparison cannot repair it.
Posterior constraints differ from several related decision rules:
| Approach | What it screens or summarizes | Important distinction |
|---|---|---|
| Posterior overdose constraint | Posterior probability that risk exceeds \(\phi\) | Explicitly limits posterior mass above a risk target. |
| Posterior mean-risk rule | Posterior mean \(E(p_{jk}\mid D)\) | A mean below \(\phi\) does not ensure that \(P(p_{jk}>\phi\mid D)\) is small. |
| Point-estimate threshold | A fitted or empirical risk estimate | Does not directly account for uncertainty around that estimate. |
| EWOC feasibility rule | A feasibility boundary based on a specified posterior overdose criterion | Its feasibility definition and decision structure should be followed as specified, not assumed identical to every candidate-screening rule. |
The posterior constraint has limitations. Its conclusion depends on the model, prior, data, target, and cutoff; it is not a model-free safety guarantee. A cutoff that is too permissive may admit moves with material posterior overdose probability, while an overly restrictive cutoff may leave no admissible move. Nor does the rule alone define the best move when several candidates pass. In that case, use a prospectively stated tie-breaker or decision criterion; do not present a small difference in estimated \(q_{jk}\) as proof of a meaningful safety advantage.
Common Pitfalls
- Confusing posterior overdose probability with mean risk. The condition is \(P(p_{jk}>\phi\mid D)<c\), not merely \(E(p_{jk}\mid D)<\phi\).
- Treating a passing pair as guaranteed safe. Passing means the posterior mass above the target is below the chosen cutoff, conditional on the model and data.
- Choosing among candidates by an unplanned numerical tie-break. When several candidates pass and their risks are similar, use a prespecified rule and acknowledge that the safety screen may not distinguish them meaningfully.
- Ignoring the allowed transition set. A pair that passes the safety constraint is not automatically a permitted next move.
- Trusting a boundary result without checking computation. If \(\widehat q_{jk}\) is close to \(c\), evaluate MCMC precision and apply the previously specified rule for equality and numerical uncertainty.
Check Your Understanding
Answer each question using the stated target and cutoff. Solutions follow the questions.
- At a candidate pair, 150 of 1,000 posterior draws have DLT risk above \(\phi\). Does the pair pass when \(c=0.15\), using the strict rule \(q<c\)?
- Two allowed moves have estimated posterior overdose probabilities 0.18 and 0.19, with cutoff \(c=0.20\). What can be concluded, and what cannot be concluded without another decision rule?
- Explain why a posterior mean risk of 0.24 does not by itself establish passage when \(\phi=0.30\) and \(c=0.20\).
- For 1,000 independent posterior draws and estimated overdose probability 0.10, calculate the approximate Monte Carlo standard error.
Solutions. 1. The estimate is \(150/1000=0.15\). Because the rule requires \(q<c\), equality does not pass. 2. Both moves pass because \(0.18<0.20\) and \(0.19<0.20\). The screen alone does not determine which to select; a prespecified tie-breaker or further decision criterion is needed, and the 0.01 difference should not automatically be treated as decisive. 3. A mean summarizes the posterior risk values but does not give the posterior fraction above 0.30. The required quantity is \(P(p_{jk}>0.30\mid D)\), which must be calculated from the posterior distribution and compared with 0.20. 4. Using the independent-draw approximation, \(SE_{\mathrm{MC}}\approx\sqrt{0.10(0.90)/1000}=\sqrt{0.00009}=0.00949\), or about 0.95 percentage points.