Why the Maximum Tolerated Dose Is Not Automatically the Best Dose
In early-phase trials, dose escalation often focuses on toxicity: the goal is to identify a dose near a prespecified toxicity target without exceeding an acceptable risk. This is an important safety task, but it answers a narrower question than “Which dose offers the best clinical benefit relative to harm?” A lower dose can be safer yet have little efficacy; a higher dose can improve efficacy while also increasing toxicity. Neither the safest dose nor the most effective dose is necessarily the best choice for patients.
The distinction matters when dose selection is intended to support further clinical study. A dose decision should reflect the trial’s clinical objective, the toxicity constraint, the evidence at each dose, and how benefit and harm are to be compared. The calculations below use hypothetical probabilities to make these differences explicit; they are not results from a real trial.
An MTD rule is not a general measure of clinical value. It uses toxicity to locate a dose, but it does not by itself compare efficacy across doses. Likewise, the “safest dose” means the dose with the smallest toxicity probability among the options, not the dose with the best benefit-risk balance. An efficacy-toxicity objective explicitly brings both dimensions into the selection decision.
Conditions for a Fair Dose Comparison
The probabilities compared across doses must refer to clearly defined outcomes and comparable assessment periods. For example, toxicity risk over a short observation window cannot be fairly compared with toxicity risk over a longer window if events can occur later. Efficacy probabilities also need a common definition and follow-up period across doses. These quantities describe probabilities at the dose level; they do not by themselves describe whether efficacy and toxicity occur in the same person.
A simple working model treats each outcome separately as binary. For each dose \(d\), suppose \(n_d\) independent patients receive that dose, each with the same probability \(p_T(d)\) of the specified toxicity and the same probability \(p_E(d)\) of the specified efficacy outcome. If outcomes are completely observed and assessment is consistent, the counts of patients with each outcome can be summarized by dose-specific proportions. These assumptions support the calculations, but do not guarantee that estimated probabilities are precise. In particular, small dose groups can produce substantial sampling uncertainty.
Dose-response patterns are often expected to increase with dose, especially for toxicity, but that expectation should not be treated as an observed fact. Estimates may be irregular because of sampling variation, and efficacy may plateau or decline. A decision rule should therefore be applied to the evidence and assumptions of the trial, not to an assumed perfectly increasing curve.
Worked Example: The Safest Dose Versus the Highest Acceptable Dose
Suppose a hypothetical trial has four doses and uses a toxicity target of \(p_T^*=0.25\). The values below are hypothetical planning estimates. The design selects the highest dose with an estimated toxicity probability no greater than 0.25.
| Dose | Estimated efficacy probability | Estimated toxicity probability |
|---|---|---|
| 1 | 0.10 | 0.04 |
| 2 | 0.35 | 0.12 |
| 3 | 0.58 | 0.23 |
| 4 | 0.66 | 0.39 |
Dose 1 is the safest because its estimated toxicity probability, 0.04, is the smallest. But its estimated efficacy probability is only 0.10. Doses 1, 2, and 3 satisfy the safety boundary because their toxicity estimates are at most 0.25; dose 4 does not. Under the stated highest-acceptable-dose rule, dose 3 is selected. Its estimated toxicity probability is 0.23, only 0.02 below the target, and its efficacy estimate is 0.58.
Thus the safest dose and the MTD under this rule are different: dose 1 minimizes estimated toxicity, while dose 3 is the highest dose within the specified boundary. This comparison does not prove that dose 3 is best for patients. It shows only that a safety-target rule and a minimum-toxicity rule answer different questions, and that neither directly weighs the size of the efficacy gain against the toxicity increase.
Adding Efficacy to the Decision
One transparent approach is to treat toxicity as a constraint and then select the dose with the greatest efficacy among doses that satisfy it. This is a different objective from targeting a particular toxicity probability. Another approach summarizes efficacy and toxicity in a common decision score. Such a score requires clinical priorities to be specified; the numerical weights are not supplied by statistical calculation alone.
For illustration, define a simple score \(U(d)=p_E(d)-w p_T(d)\), where \(w\) is a prespecified weight expressing how much the decision penalizes toxicity relative to efficacy. This is only an illustrative comparison rule. Its numerical value depends on the chosen scale and weight, and it should not be interpreted as a universally valid measure of patient benefit.
Worked Example: A Benefit-Risk Score Can Prefer a Dose Below the MTD
Consider a second hypothetical comparison with a toxicity target of 0.25. Suppose the selection rule first excludes doses above that boundary and then compares the remaining doses using \(U(d)=p_E(d)-1.5p_T(d)\). The weight 1.5 is an illustrative value specified in advance for this example, not a general recommendation.
| Dose | \(p_E(d)\) | \(p_T(d)\) | \(U(d)=p_E(d)-1.5p_T(d)\) |
|---|---|---|---|
| 1 | 0.12 | 0.04 | 0.060 |
| 2 | 0.34 | 0.12 | 0.160 |
| 3 | 0.50 | 0.24 | 0.140 |
| 4 | 0.63 | 0.38 | 0.060 |
For dose 2, the calculation is \(0.34-1.5(0.12)=0.34-0.18=0.16\). For dose 3, it is \(0.50-1.5(0.24)=0.50-0.36=0.14\). Dose 4 is outside the boundary because \(0.38>0.25\), so it is excluded before comparing scores. Among the eligible doses, dose 2 has the largest score, 0.160, even though dose 3 is nearer the toxicity target and has higher efficacy.
This result depends on the constraint and the prespecified weight. A different clinically justified weight could change the ranking. The example’s point is not that one score is always correct, but that an MTD target does not itself determine how much efficacy gain is worth a toxicity increase.
Sampling Uncertainty and Decision Diagnostics
Dose-level probabilities are generally estimated, not known. A dose with an estimated toxicity probability below a boundary may still have considerable uncertainty, especially when few patients have been treated or toxicity follow-up is incomplete. A point estimate just below the boundary should not be treated as proof that the true probability is acceptable. The earlier material on posterior overdose control addresses decisions that explicitly account for uncertainty in toxicity risk; regardless of the method used, the decision rule and its uncertainty assessment should be reported together.
The following checks help identify when a dose comparison may be misleading:
- Unequal outcome assessment: if patients at one dose have shorter follow-up, fewer observed events may reflect less opportunity to observe them rather than lower risk.
- Sparse dose groups: one or two events can shift a proportion substantially when the denominator is small. Report denominators alongside estimated probabilities.
- Irregular dose patterns: a surprising decrease in estimated toxicity at a higher dose may be sampling noise, but it can also challenge a monotonicity assumption. Do not silently force the data into an expected pattern.
- Outcome mismatch: a toxicity target and efficacy estimate are not directly comparable if they use different populations, definitions, or assessment periods without a defensible rationale.
- Unstated priorities: a weighted score can create a precise ranking while concealing a subjective or clinically unjustified choice of weight.
An important limitation is that separate marginal probabilities do not reveal the patient-level relationship between efficacy and toxicity. Two doses can have the same efficacy and toxicity probabilities while differing in how often both outcomes occur in the same patient. A full joint-outcome analysis requires additional definitions and assumptions; it cannot be recovered from the two marginal probabilities alone.
Worked Example: Small Samples Can Make a Boundary Decision Fragile
In a third hypothetical trial, 20 patients are assessed at each of three doses. Assume complete follow-up over the same toxicity and efficacy assessment periods, independent patients within each dose, and binary outcomes. The observed counts are:
| Dose | Patients with efficacy outcome | Patients with toxicity | Total patients |
|---|---|---|---|
| 1 | 4 | 1 | 20 |
| 2 | 9 | 4 | 20 |
| 3 | 13 | 8 | 20 |
The estimated efficacy probabilities are \(4/20=0.20\), \(9/20=0.45\), and \(13/20=0.65\). The estimated toxicity probabilities are \(1/20=0.05\), \(4/20=0.20\), and \(8/20=0.40\). With a toxicity boundary of 0.25 and the highest-acceptable-dose rule, dose 2 is selected: its estimate is below the boundary, while dose 3’s estimate exceeds it. Dose 1 is safer by the point estimate but has a lower estimated efficacy probability.
The dose-2 toxicity estimate is based on only four events. Under the binomial working model, its estimated standard error is $$ SE(\hat{p}_T(2)) = \sqrt{\frac{\hat{p}_T(2)\{1-\hat{p}_T(2)\}}{n}} = \sqrt{\frac{0.20(0.80)}{20}} = 0.0894. $$ This standard error is large relative to the 0.05 distance between the estimate, 0.20, and the boundary, 0.25. The rule still selects dose 2 as written, but the point estimate alone gives limited reassurance about its true toxicity probability. A careful interpretation distinguishes the mechanical selection from the uncertainty surrounding it.
Choosing and Reporting the Objective
The main alternatives can be summarized as follows. Each answers a different question, so the method should follow the clinical objective rather than being chosen solely because it produces a familiar dose label.
| Approach | Question answered | Main limitation |
|---|---|---|
| Minimum estimated toxicity | Which dose appears safest? | Does not account for efficacy gains at higher doses. |
| MTD or toxicity-target rule | Which dose is near a toxicity target, or highest below a safety boundary? | Does not, by itself, rank doses by overall clinical benefit. |
| Maximum efficacy under a toxicity constraint | Which acceptable dose has the greatest estimated efficacy? | Depends on the constraint and may not distinguish doses with similar efficacy. |
| Prespecified weighted score | Which dose ranks highest under a stated efficacy-to-toxicity trade-off? | Depends on clinically justified weights and the score’s scale. |
A sound report states the outcome definitions and follow-up periods, the toxicity target or boundary, the rule used to select a dose, and the uncertainty in the estimates. If efficacy and toxicity are combined, it also states how that combination was specified. This makes clear whether a selected dose is the safest, toxicity-targeted, most effective among acceptable doses, or preferred under a broader benefit-risk criterion.
Common Mistake
Incorrect reasoning: “Dose 1 has the lowest toxicity probability, so it is the best dose.” This concludes that minimum harm guarantees the best clinical choice, without considering whether efficacy is adequate or how much efficacy could be gained at another dose.
Correction: First state the decision objective. If the goal is minimum toxicity, dose 1 may be the answer. If the goal is to select a dose near a toxicity target, the MTD rule applies. If the goal is to maximize efficacy subject to a safety boundary, compare efficacy among doses that meet the boundary. If the goal is a benefit-risk ranking, specify and justify how efficacy and harm are compared. These are distinct decisions, and the data alone do not determine which clinical objective is appropriate.
Check Your Understanding
Answer each question using the stated objective and hypothetical values.
- A trial uses a toxicity boundary of 0.20. Estimated toxicity probabilities at three doses are 0.06, 0.18, and 0.27. Under the highest-acceptable-dose rule, which dose is selected? Which is safest?
- At two doses, the lower dose has \(p_E=0.30\) and \(p_T=0.10\), while the higher dose has \(p_E=0.45\) and \(p_T=0.20\). Under \(U(d)=p_E(d)-p_T(d)\), which dose has the higher score?
- Why does an estimated toxicity probability of 0.19 not establish that a dose’s true toxicity probability is below a boundary of 0.20?
- Two doses have the same estimated efficacy and toxicity probabilities. Can those marginal probabilities establish that efficacy and toxicity occur together in the same patients? Explain.
Solutions. (1) The second dose is the highest dose at or below 0.20, so it is selected; the first dose is safest because 0.06 is smallest. (2) The lower-dose score is \(0.30-0.10=0.20\); the higher-dose score is \(0.45-0.20=0.25\). The higher dose has the larger score under this stated rule. (3) The value 0.19 is an estimate subject to sampling uncertainty. Its proximity to 0.20 does not prove that the unknown probability is below the boundary. (4) No. Marginal probabilities describe efficacy and toxicity separately; they do not determine their within-patient association or the probability of both outcomes occurring in one patient.