Selecting a Phase II Dose From Efficacy and Exposure Data

From Dose-Ranging Data to a Phase II Dose Choice

A phase II dose-ranging trial should inform more than whether a drug has activity. It should help identify a dose, or a plausible range, whose expected benefit is large enough to matter without choosing a dose solely because it produced the largest observed mean. A model can connect the randomized dose groups to exposure and response, interpolate between studied groups, and quantify uncertainty. The model does not replace the observed group results: a useful dose decision checks whether its predictions are consistent with those results and makes the target and uncertainty criteria explicit.

The case below is entirely hypothetical. The endpoint is change from baseline on a symptom scale at week 12, with larger values representing greater improvement. Effects are reported as placebo-adjusted mean improvements, in scale points. A clinically relevant target has been set at an improvement of at least 8 points. The analysis is about selecting a dose based on efficacy and exposure evidence; it does not establish that a dose is acceptably safe or appropriate as a final regimen.

Hypothetical Trial Data and Exposure-Response Model

Suppose participants were randomized equally to placebo or one of four active doses. The endpoint is measured at baseline and week 12; the table shows the placebo-adjusted week-12 mean improvement. Each group has 40 participants. The exposure column gives the group's mean steady-state average concentration, \(C_{\mathrm{avg}}\), in mg/L. The dose-to-exposure relationship in this hypothetical trial is approximately \(C_{\mathrm{avg}}=0.20 \times \text{dose}\), with dose in mg.

For illustration, an exposure-response analysis fit a saturable \(E_{\max}\) model to the individual-level efficacy and exposure data. The fitted parameter estimates are \(E_{\max}=12\) points and \(EC_{50}=2\) mg/L. This model describes an increasing response that approaches a plateau as exposure increases. It is being used here as a compact description of the hypothetical data, not as a universal biological law.

Here, \(E(C)\) is the model-predicted placebo-adjusted mean improvement at exposure \(C\). \(E_{\max}\) is the limiting response as exposure becomes very large, and \(EC_{50}\) is the exposure at which the model predicts half of \(E_{\max}\). For this model to be appropriate, response must be adequately represented by a monotone saturating curve over the observed range, and the exposure and endpoint measurements must be suitable for the stated interpretation. A fitted curve also assumes that important differences between participants or groups have not made the common exposure-response relationship misleading.

Finding the Exposure Associated With the Target

A model can translate a response target into an exposure target. To find the exposure at which the fitted curve reaches 8 points, solve \(E(C)=8\). For a target \(T\) below \(E_{\max}\), rearranging the model gives \(C=T EC_{50}/(E_{\max}-T)\). A finite positive solution requires \(0<T<E_{\max}\); a target at or above the plateau cannot be reached at finite exposure under this model.

Given the hypothetical mean exposure relationship \(C_{\mathrm{avg}}=0.20 \times \text{dose}\), a mean exposure of 4 mg/L corresponds to \(4/0.20=20\) mg. This is a model-based mapping, not a guarantee that every participant assigned 20 mg will have an exposure of 4 mg/L or an 8-point response. Exposure varies among participants, and the dose-to-exposure relationship is estimated rather than exact.

Using Uncertainty to Screen Candidate Doses

A dose should not be declared adequate just because its point prediction reaches the target. The model-based uncertainty for each predicted mean should also be considered. Suppose the fitted model's covariance calculation gives prediction SEs of 0.55, 0.62, and 0.78 points at 10, 20, and 40 mg, respectively. These are SEs for predicted mean responses, including parameter uncertainty; they are not the observed-group SEs in the data table. For a prediction \(E(C)\), the delta-method approximation uses the gradient of the model with respect to its fitted parameters and their covariance matrix \(V\):

The gradient for parameters \((E_{\max},EC_{50})\) is \(\left(C/(EC_{50}+C),\ -E_{\max}C/(EC_{50}+C)^2\right)\). In practice, the covariance calculation must reflect the fitted model and the data; an SE should not be inferred from the curve's shape alone.

If three candidate doses are screened and the aim is to make a one-sided 90% family-wise lower-bound statement, a Bonferroni adjustment allocates a tail probability of \(0.10/3\) to each dose. Under a normal approximation, the one-sided critical quantile is \(z_{1-0.10/3}\approx1.834\). For each candidate, calculate the lower bound as prediction minus 1.834 times its prediction SE. Bonferroni's union-bound argument does not require the three predictions to be independent: the probability that any of the three bounds fails to cover its corresponding mean is no greater than the sum of the three individual failure probabilities, which is at most \(3(0.10/3)=0.10\).

Assumptions, Diagnostics, and Alternative Summaries

The dose recommendation is only as credible as the analysis behind it. Randomization supports comparisons of assigned-dose groups, provided allocation was implemented appropriately and endpoint data are handled without bias. Exposure is measured after randomization, however, and participants with different exposures may also differ in characteristics related to response. An exposure-response association alone is therefore not automatically a causal effect of changing exposure. The randomized dose-group results remain an important check on the model-based exposure interpretation.

The \(E_{\max}\) fit assumes an increasing saturating relationship with no downturn in the studied range. It also assumes that the chosen week-12 endpoint adequately captures the relevant response and that the model accounts for the exposure variation relevant to prediction. If efficacy changes over time, a week-12 summary can conceal an earlier benefit, delayed effect, or loss of response. The analysis should use a consistent endpoint definition and account for missing observations and departures from the planned treatment or measurement schedule.

Diagnostics should ask whether the model is failing in ways that matter to the dose decision. Plot observed group means and uncertainty alongside the predictions; inspect individual-level residual patterns against exposure and time; and look for curvature, systematic overprediction at high exposure, influential observations, or subgroup patterns. Compare alternative plausible curve shapes when justified by the data. A pattern of increasingly negative observed-minus-predicted differences at high exposure, for example, could suggest that the assumed curve or its exposure distribution is inadequate. A visually close fit to five group means is not by itself sufficient evidence, especially when the model was fitted to those same data.

The strongest practical interpretation uses these approaches together rather than treating them as competitors. The observed randomized means show what happened at the tested doses; the exposure-response model helps explain the pattern and estimate response at relevant exposures; and uncertainty bounds make the decision rule visible. If these sources disagree materially, investigate the disagreement rather than selecting whichever result supports a preferred dose.

Common Pitfall and Decision Limits

A common mistake is to select the dose with the largest observed mean and call it optimal. The highest mean can arise partly from sampling variability, and a modest additional response at the highest dose may not justify increased exposure. The correction is to define a clinically meaningful target and an uncertainty criterion in advance, compare observed results with model predictions, and state clearly what the evidence does and does not establish. Here, 40 mg is the only candidate whose simultaneous lower bound clears the target; this supports an efficacy-based candidate under the stated rule, not a complete clinical recommendation.

The target itself is a decision input, not something the fitted curve determines. A different target, an individual-responder criterion, or a different uncertainty requirement could change the selected dose. In addition, efficacy alone cannot establish benefit-risk balance, and a mean response target does not guarantee that an individual patient benefits. These limitations matter when translating a phase II result into later development decisions.