Modeling Efficacy Plateaus And Toxicity Tradeoffs

Why an Efficacy Plateau Changes the Dose Comparison

A higher dose can increase the probability of benefit while also increasing the probability of toxicity. If efficacy continues to rise only slightly at the upper doses, the extra benefit may be small relative to the added harm. Dose selection therefore depends not only on which dose has the largest estimated efficacy, but also on where efficacy gains become practically small, how toxicity changes across that region, and how uncertain those estimates are.

A dose-response curve may approach a mathematical ceiling without ever becoming exactly flat at a finite dose. A clinically useful plateau is instead defined relative to a margin: for example, the efficacy increase from doubling a dose may be called negligible if it is less than three percentage points. The margin is a clinical judgment, not a property that the data determine by themselves.

Models for Saturating Efficacy and Increasing Toxicity

Let \(d\) denote dose, \(p_E(d)\) the probability of the prespecified efficacy outcome by its assessment time, and \(p_T(d)\) the probability of the prespecified toxicity outcome by its assessment time. These are marginal probabilities at each dose; a joint outcome model is not required simply to compare their expected contributions to a utility.

A common saturating model for a binary efficacy outcome is the Emax curve. In its simplest form, with Hill coefficient fixed at one, it is:

Here \(p_0\) is the efficacy probability at zero dose, \(E_{\max}\) is the maximum increase above \(p_0\), and \(ED_{50}\) is the dose that achieves half of that increase. Parameters must keep predicted probabilities between zero and one; one sufficient restriction is \(0\leq p_0\) and \(p_0+E_{\max}\leq 1\). More flexible models can estimate a Hill coefficient, but extra flexibility requires enough dose levels and data to identify it.

A simple increasing model for toxicity is logistic:

The monotonicity restriction encodes the working assumption that toxicity risk does not decrease as dose increases. These models are useful only if their shapes are reasonably compatible with the endpoint definitions, dose range, and observed data. They do not establish a biological mechanism.

The dose increment matters. A curve may satisfy a three-percentage-point margin for doubling doses but not for an increase of only 25%, or the reverse for a different comparison. A derivative-based definition is also possible: call the response practically flat where the increase in efficacy probability per unit dose is below a specified slope \(c\). Such a definition is sensitive to the units in which dose is measured, so the units and threshold must be reported.

Worked Example: Locating a Practical Plateau

Why the Plateau Criterion Has This Form

For the simple Emax curve, the marginal increase in efficacy probability per mg is the derivative:

Both parameters in the numerator are positive under the usual model constraints, so the derivative is positive. As \(d\) increases, the denominator increases, and the derivative decreases toward zero. This justifies describing the curve as saturating. It does not provide a unique finite plateau dose: the slope is positive at every finite dose in this model.

A slope threshold \(c>0\) gives an operational dose boundary. Solving the inequality \(d p_E(d)/d d\leq c\) yields:

If the right-hand side is negative, the slope is already at or below \(c\) for all nonnegative doses. This calculation is a derivation of the boundary for this particular model and criterion, not proof that the clinical margin \(c\) is appropriate. In practice, a finite dose grid may make a specified increment, such as doubling, easier to communicate than a per-mg slope.

Combining Plateau and Toxicity Information

A plateau does not, by itself, rank doses. A transparent comparison can use the dose utility established by the clinical priorities. For illustration, suppose a unit of efficacy probability receives weight \(1\), a unit of toxicity probability receives cost \(\lambda>0\), and expected utility is \(U(d)=p_E(d)-\lambda p_T(d)\). This is a deliberately simple linear summary: it assumes the stated weights adequately represent the decision and does not imply that efficacy and toxicity are independent. The expected value of an additive benefit-minus-harm score is the difference of the two marginal expectations.

Quantifying Uncertainty in the Preferred Dose

A dose-specific fitted curve is estimated, not known. Uncertainty in \(p_0\), \(E_{\max}\), and \(ED_{50}\) changes predicted efficacy gains and the location of a practical plateau. Uncertainty in the toxicity model and in the utility weights can also change the ranking. Therefore, a point estimate of the preferred dose should be accompanied by uncertainty summaries that address the actual comparison.

One approach is to calculate a confidence interval for a utility contrast. If \(\widehat{\Delta}\) estimates \(U(d_2)-U(d_1)\), and \(SE(\widehat{\Delta})\) is its estimated standard error, an approximate 95% confidence interval is \(\widehat{\Delta}\pm1.96SE(\widehat{\Delta})\) when the estimator is approximately normal. The interval describes uncertainty in the contrast, not the probability that either dose is truly best.

A second approach is a bootstrap: repeatedly resample participants in a way that respects the randomization and design, refit the efficacy and toxicity models, recalculate the plateau criterion and utility at each dose, and record which dose has the largest estimated utility in each replicate. With adaptive allocation, the resampling or simulation scheme must reflect the allocation process; an ordinary resample that ignores the design may not represent the relevant sampling variability. Frequencies across bootstrap replicates describe how often the fitted procedure selects each dose under the resampling scheme. They are not posterior probabilities that doses are optimal.

Diagnostics, Alternatives, and Limitations

Before interpreting a plateau, check whether the data support the curve that produced it. Plot dose-specific observed response proportions with uncertainty intervals against fitted probabilities. Look for systematic departures, such as a high-dose efficacy decrease that a monotone saturating curve cannot represent, or toxicity estimates that do not increase across doses despite the monotonicity restriction. Examine whether the estimated asymptote and \(ED_{50}\) are weakly identified; with sparse observations near the upper dose range, many combinations of these parameters can fit similarly while implying different plateau locations. Very wide uncertainty intervals for predictions or a boundary estimate for a parameter are reasons for caution, not evidence that the fitted curve is precise.

The Emax model is not the only option. A logistic efficacy curve can describe a sigmoid response, and a monotone spline can allow more flexible shapes. A flexible model can reduce shape bias but may be unstable with few dose levels; a simple parametric curve is easier to summarize but can give misleading plateau estimates if its shape is wrong. Directly comparing observed proportions by dose avoids imposing a smooth curve, but is noisy and cannot reliably interpolate between sparse dose levels. The choice is a balance between credible shape assumptions and the information available.

A practical plateau depends on the chosen margin and dose increment, so report both and, when conclusions are sensitive, show results for more than one clinically defensible margin. Model uncertainty is not the only limitation: utility weights may be contested, the assessment times may not capture delayed outcomes, and population-level probabilities do not describe every patient's benefit-harm balance. A plateau identified from one endpoint also need not apply to another endpoint or patient subgroup.