Distinguishing the MTD from a Dose Recommendation

Why the MTD and the Recommended Dose Can Differ

A model can identify the dose at which its estimated toxicity probability reaches a chosen target. That dose is a model-based maximum tolerated dose (MTD). A clinical trial protocol, however, must choose from available doses while managing uncertainty and applying safety constraints. Its recommended dose is therefore a decision produced by a rule, not necessarily the dose whose estimated toxicity probability exactly equals the target.

The distinction matters because the dose-toxicity curve is uncertain, doses are usually administered from a finite candidate set, and a protocol may prohibit doses with too much posterior probability of exceeding a safety limit. The MTD describes a target on a toxicity curve; a recommendation says what action the protocol permits or selects given the evidence and its rules.

Definition: Fix a model-based toxicity probability \(p(d)\) at dose \(d\), and a target boundary \(\gamma\), with \(0<\gamma<1\). When the curve is increasing and reaches \(\gamma\), the boundary MTD is the dose \(d_{\mathrm{MTD}}\) satisfying \(p(d_{\mathrm{MTD}})=\gamma\). A dose recommendation is the dose selected by a prespecified decision rule from the candidate doses, using the available evidence and any protocol constraints. These are different objects: the first is a target-crossing dose; the second is an action under a rule.

In a Bayesian analysis, the toxicity curve depends on unknown parameters. Consequently, the MTD is uncertain too: for each posterior draw of the curve parameters, there may be a corresponding draw of \(d_{\mathrm{MTD}}\). A posterior median or mean MTD summarizes that distribution, but it does not, by itself, define what dose the protocol should recommend.

Finding the MTD on a BLRM Curve

The earlier material on the two-parameter Bayesian logistic regression model (BLRM) uses a dose reference \(d_{\mathrm{ref}}\) and a log-dose predictor. In the parameterization used here, the dose-specific toxicity probability is

\operatorname{logit}\{p(d)\} = \alpha+\beta\log\left(\frac{d}{d_{\mathrm{ref}}}\right), \qquad \operatorname{logit}(q)=\log\left(\frac{q}{1-q}\right).

For a fixed pair \((\alpha,\beta)\), the curve is increasing in dose when \(\beta>0\). If the target \(\gamma\) lies on the curve over the dose range being considered, the MTD solves \(\operatorname{logit}\{p(d_{\mathrm{MTD}})\}=\operatorname{logit}(\gamma)\). Rearranging gives

d_{\mathrm{MTD}} = d_{\mathrm{ref}} \exp\left\{ \frac{\operatorname{logit}(\gamma)-\alpha}{\beta} \right\}.

This equation gives a continuous-valued dose for a fixed parameter pair. It is not automatically one of the trial's candidate doses. In Bayesian analysis, applying the equation to each posterior draw \((\alpha^{(m)},\beta^{(m)})\) gives a posterior distribution of MTDs. A draw with \(\beta^{(m)}\leq 0\), or a target crossing far outside the dose range, calls for particular care: the usual increasing-curve interpretation may fail, or the computed crossing may depend on extrapolation.

Worked Example: Solving for a Continuous MTD

Suppose a hypothetical BLRM parameter pair is \(\alpha=-1.3863\) and \(\beta=1.20\), with reference dose \(d_{\mathrm{ref}}=20\) mg. Let the target boundary be \(\gamma=0.25\). Find the model-based MTD for this parameter pair.

The assumptions for this calculation are that the stated logistic curve applies at the doses considered, \(\beta>0\), and the target boundary is the protocol's chosen definition of the MTD. Here the positive slope gives an increasing curve. The target logit is \(\operatorname{logit}(0.25)=\log(0.25/0.75)=\log(1/3)\approx-1.0986\).

Substitute into the MTD formula:

d_{\mathrm{MTD}} = 20\exp\left\{ \frac{-1.0986-(-1.3863)}{1.20} \right\} = 20\exp(0.2398) \approx 25.42\ \text{mg}.

Thus this curve reaches a 25% toxicity probability at approximately 25.42 mg. If the trial only permits doses of 20 mg and 30 mg, neither is exactly the continuous MTD. Choosing between them requires a decision rule; rounding 25.42 mg to a candidate dose without reference to that rule would not be justified.

Posterior MTDs Are Not Recommendations

Let the posterior distribution of the BLRM parameters represent uncertainty after observing trial data. For each posterior draw, the MTD formula gives a draw of the boundary-crossing dose, when the formula is defined and meaningful. The resulting posterior distribution can be summarized by a median, a credible interval, or another prespecified summary. Those summaries describe uncertainty about the model-based boundary dose.

A recommendation requires an additional decision rule. It might use candidate doses, posterior toxicity probabilities, and constraints set out in the protocol. For example, a protocol might require that the posterior probability that a candidate dose exceeds a specified toxicity limit be no greater than a chosen cutoff. The limit, event definition, and cutoff must be specified by the protocol; there is no universal rule that converts a posterior MTD summary into a recommendation. The following example uses a deliberately simple illustrative rule, not a general BLRM definition of overdose.

Worked Example: A Candidate Dose Can Fail a Safety Rule

Consider four equally weighted hypothetical posterior draws for a positive quantity \(D\), where the dose-toxicity curve in each draw is \(p(d)=d/(3D+d)\). This simplified curve has \(p(D)=0.25\), so \(D\) is that draw's MTD for a 25% target. Suppose the candidate doses are 20, 30, and 40 mg, the target interval is 20% to 30%, and the illustrative protocol rule permits a dose only if the posterior probability of \(p(d)>0.30\) is at most 0.25. Take the four draws of \(D\) to be 18, 26, 35, and 45 mg.

At 40 mg, the probabilities in the four draws are \(40/(3\cdot18+40)=40/94\approx0.4255\), \(40/(3\cdot26+40)=40/118\approx0.3390\), \(40/(3\cdot35+40)=40/145\approx0.2759\), and \(40/(3\cdot45+40)=40/175\approx0.2286\). Two of four probabilities exceed 0.30, so the estimated posterior probability of exceeding the limit is \(2/4=0.50\). The 40 mg dose fails the stated rule.

At 30 mg, the four probabilities are \(30/84\approx0.3571\), \(30/108\approx0.2778\), \(30/135\approx0.2222\), and \(30/165\approx0.1818\). One of four exceeds 0.30, giving \(1/4=0.25\), which meets the rule. At 20 mg, the probabilities are \(20/74\approx0.2703\), \(20/98\approx0.2041\), \(20/125=0.1600\), and \(20/155\approx0.1290\); none exceeds 0.30. If the protocol selects the highest candidate dose passing this illustrative rule, it recommends 30 mg.

The posterior MTD draws are still 18, 26, 35, and 45 mg; their median is \((26+35)/2=30.5\) mg. The recommendation is 30 mg because it is the highest candidate passing the rule, not because the posterior median MTD is 30 mg. A different candidate set or constraint could produce a different recommendation from the same posterior draws.

Uncertainty Can Change the Decision

A useful way to see why a point summary is insufficient is to compare posterior distributions with the same median but different spread. Under a probability-based safety rule, a wide posterior distribution can make a candidate dose less acceptable even if its median MTD matches that of a more concentrated distribution. The rule responds to tail probability, whereas the median does not.

Worked Example: Same Median MTD, Different Recommendations

Use the same simplified curve \(p(d)=d/(3D+d)\), candidate doses 20, 30, and 40 mg, and illustrative rule \(P\{p(d)>0.30\mid\text{data}\}\leq0.20\). Compare two sets of four equally weighted hypothetical posterior draws for \(D\): a concentrated set \(26,28,28,30\) mg and a more dispersed set \(16,24,32,40\) mg. Both have median MTD 28 mg.

For this curve, \(p(d)>0.30\) exactly when \(d/(3D+d)>0.30\). Since \(3D+d\) is positive, rearranging gives \(d>0.90D+0.30d\), hence \(0.70d>0.90D\), or \(D<(7/9)d\). At 30 mg, the threshold is \(D<23.33\) mg. None of the concentrated draws is below this threshold, so its estimated exceedance probability is \(0/4=0\). In the dispersed set, only 16 mg is below the threshold, so the probability is \(1/4=0.25\), which exceeds the permitted 0.20.

At 20 mg, the threshold is \(D<15.56\) mg. Neither set contains a draw below that threshold, so 20 mg passes for both. At 40 mg, the threshold is \(D<31.11\) mg: all four concentrated draws and two of four dispersed draws fall below it, giving exceedance probabilities of 1.00 and 0.50. Thus 40 mg fails for both. Under a rule that selects the highest passing candidate, the concentrated set recommends 30 mg and the dispersed set recommends 20 mg, even though their median MTDs are both 28 mg.

The arithmetic illustrates a general decision principle: a posterior summary that ignores uncertainty can conceal relevant tail risk. The numerical rule here is only an example; an actual trial must use its prespecified constraint and its defined toxicity event.

Assumptions, Diagnostics, and Alternatives

The continuous MTD calculation relies on a meaningful dose-toxicity model, a specified target boundary, and a curve that reaches that boundary. For the displayed BLRM formula, a positive slope gives a unique crossing on an unrestricted positive dose scale. In a trial, however, the useful domain is usually the studied or permitted range. If a calculated crossing lies outside that range, it is an extrapolated model quantity rather than evidence that the dose can be administered or is acceptable.

Several checks help identify when the distinction is being mishandled:

Check the target convention. A target interval has a lower and an upper boundary. State which boundary defines the MTD; do not silently switch between the interval's center and its upper edge.
Inspect posterior uncertainty. A broad or asymmetric posterior distribution of MTDs means that one point summary cannot represent the uncertainty. Check the full posterior summary and any tail probabilities used by the protocol.
Check the candidate grid and rule. The continuous boundary dose may not be available. Verify which candidate doses pass the protocol's constraints before describing any dose as recommended.
Check model support. A crossing beyond the observed dose range, or posterior draws with problematic slopes, can make the MTD summary sensitive to model assumptions and extrapolation.

A target-based approach can report the continuous MTD or summarize its posterior distribution. It directly answers where the model places the target boundary, but does not by itself account for available doses or a protocol's safety rule. A decision-rule approach evaluates permitted candidate doses under prespecified criteria and returns an action. It answers what the protocol recommends, but its result depends on the chosen candidates and constraints. Both can be reported, provided their meanings are kept separate. Neither removes the need to assess whether the dose-toxicity model is credible.

The MTD is also not automatically the dose that optimizes overall clinical benefit. This tutorial concerns the distinction between a toxicity-boundary dose and a protocol dose recommendation; a recommendation based on toxicity constraints should not be described as a demonstrated optimum for efficacy, tolerability, or patient benefit.

Common pitfall: “The posterior median MTD is 30 mg, so the recommended dose must be 30 mg.” This is incorrect unless the protocol explicitly defines its recommendation to be that summary and all other constraints are satisfied. The correction is to report the posterior MTD summary as an estimate of a model-based target crossing, then apply the prespecified decision rule to the candidate doses to determine the recommendation.

Key takeaway: The MTD is defined by a toxicity boundary on a model curve. A dose recommendation is a protocol decision that can depend on posterior uncertainty, the candidate-dose set, and safety constraints; it need not equal a point estimate of the MTD.

Check Your Understanding

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

For a fixed increasing toxicity curve, what equation defines the MTD when the target boundary is \(\gamma\)?
Why does a posterior median MTD not, on its own, determine the recommended dose?
Under the illustrative curve \(p(d)=d/(3D+d)\), what is \(p(30)\) when \(D=20\) mg?
Give one reason a continuous MTD may differ from a recommended dose even if the model is correctly specified.

Solutions. (1) It is the dose \(d_{\mathrm{MTD}}\) satisfying \(p(d_{\mathrm{MTD}})=\gamma\). (2) The median summarizes the posterior distribution of boundary-crossing doses, while a recommendation applies a protocol decision rule to candidate doses and may impose uncertainty-based safety constraints. (3) \(p(30)=30/(3\cdot20+30)=30/90=0.3333\), or about 33.33%. (4) For example, the continuous crossing may not be an available candidate dose, or the candidate may fail a protocol constraint based on posterior uncertainty.