Confirming a Phase III Dose With Phase II Evidence

From a Phase II Dose Rationale to a Phase III Prediction

A Phase II exposure–response relationship can help answer a specific development question: if the proposed regimen is given to the intended Phase III population, what benefit should the model predict, and how uncertain is that prediction? The answer is not simply the Phase II treatment effect copied forward. Phase III patients may have different characteristics, and the proposed dose may produce a different exposure distribution. Both can change the predicted average benefit.

The calculation is a projection conditional on the model and its assumptions. It supports a dose decision; it does not establish that the Phase III result will match the prediction. The starting point is to specify the efficacy measure, the population to which the prediction applies, the proposed regimen, and what magnitude of benefit would be considered adequate.

Definition: Let $m(c,z;\theta)$ be the model-predicted treatment benefit for a patient with exposure $c$ and characteristics $z$, under model parameters $\theta$. The target-population average benefit at the proposed regimen is $$ \psi(\theta)=\int m(c,z;\theta)f_{\mathrm{III}}(c,z)\,dc\,dz, $$ where $f_{\mathrm{III}}$ is the expected joint distribution of exposure and patient characteristics in the Phase III population under that regimen. For a clinically chosen minimum acceptable benefit $\delta$, the risk of inadequate benefit can be expressed as $P(\psi(\theta)<\delta\mid\text{available evidence})$, when uncertainty about $\theta$ and the target population is represented by a probability distribution.

Here, benefit is defined so that larger positive values are better. For example, if an endpoint is a symptom score for which decreases are beneficial, the model can express benefit as the placebo-adjusted reduction rather than the raw score change. The units and direction must be stated before interpreting any threshold.

Assumptions Behind the Projection

Before using the calculation, specify what must hold for Phase II evidence to inform Phase III. The exposure–response model must describe the relevant treatment effect over the exposure range expected at the proposed regimen. The Phase II and Phase III endpoints, assessment times, background care, and treatment-effect definition must be sufficiently comparable, or their differences must be represented in the model. The Phase III exposure distribution must be estimated for the intended regimen and population, not borrowed uncritically from a different dose or patient group.

Transport also requires that important modifiers of treatment benefit are measured and represented. If disease severity or another characteristic changes benefit even at the same exposure, then $m$ needs to include that characteristic, or the projection may be biased. The Phase III population should have adequate overlap with the Phase II data: a prediction for exposures far beyond those observed is extrapolation, even when the fitted curve produces a precise-looking number. Finally, adherence, dose interruptions, concomitant treatment, and missing-data assumptions should be compatible with the estimand being projected.

The target exposure distribution may come from a population PK model combined with the planned dose and Phase III covariates. The integration should preserve relevant relationships: if exposure depends on body size or renal function, independently averaging exposure and those covariates can create a fictitious population. When the population is represented by strata or simulated patients, the average is a weighted average over those groups or patients.

Deriving the Target-Population Average

The basic justification is the law of total expectation: an average benefit in the target population is obtained by averaging the conditional model prediction over the distribution of patients and exposures in that population. For a finite set of $J$ strata, with proportion $p_j$ in stratum $j$, this becomes:

\psi(\theta)=\sum_{j=1}^{J}p_jm(c_j,z_j;\theta), \qquad \sum_{j=1}^{J}p_j=1.

The weights describe the Phase III target population; the model predictions describe benefit conditional on the stratum’s exposure and characteristics. Replacing the target weights with Phase II weights answers a different question. If the model parameters are uncertain, then $\psi$ is uncertain too. In a Bayesian analysis, posterior draws of $\theta$ can be passed through the calculation to obtain a distribution for $\psi$. Uncertainty in the Phase III covariate or exposure distribution can also be represented by drawing or resampling that distribution for each calculation.

For a fixed target distribution, the calculation is repeated for each parameter draw. The proportion of draws with $\psi<\delta$ estimates the posterior probability of inadequate average benefit, conditional on the model, prior, and data. It is not the probability that an individual patient will fail to benefit. Individual outcomes also vary around their conditional mean, and that residual variability is a separate component.

Worked Example: Predicting Benefit at the Proposed Exposure

Hypothetical case. A Phase II model estimates the placebo-adjusted mean improvement on a clinical scale, in points, as $m(c)=E_{\max}c/(EC_{50}+c)$. The fitted values are $E_{\max}=12.0$ points and $EC_{50}=4.0$ mg·h/L. Suppose the proposed Phase III regimen is expected to produce a representative exposure of $c=6.0$ mg·h/L. Calculate the model-predicted benefit at that exposure.

Substitute the fitted values:

m(6.0)=\frac{12.0\times 6.0}{4.0+6.0} =\frac{72.0}{10.0} =7.20\text{ points}.

Under this model, the predicted placebo-adjusted mean improvement at 6.0 mg·h/L is 7.20 points. This is a conditional prediction at one exposure, not yet a Phase III population average. It assumes the exposure–response relationship is applicable at 6.0 mg·h/L and that the stated model adequately describes benefit for the relevant patients. A single representative exposure would be insufficient if the target population has substantial exposure variability or important effect modifiers.

Accounting for the Phase III Population

A dose does not generally produce one exposure in every patient. Population characteristics can shift the exposure distribution, and that shift can change expected benefit even when the exposure–response parameters remain the same. The next calculation replaces the single exposure with a population-weighted average.

Worked Example: Averaging Across Phase III Exposure Groups

Hypothetical case. Use the same Phase II model and parameter estimates. A Phase III population projection predicts that 70% of patients will have exposure $6.0$ mg·h/L and 30% will have exposure $3.0$ mg·h/L. Assume no additional effect modification by patient characteristics once exposure is specified. Calculate the target-population average benefit.

First calculate the conditional benefits. At $6.0$ mg·h/L, the benefit is $7.20$ points, as calculated above. At $3.0$ mg·h/L:

m(3.0)=\frac{12.0\times 3.0}{4.0+3.0} =\frac{36.0}{7.0} =5.143\text{ points}.

Weight each prediction by its Phase III population proportion:

\psi=0.70(7.20)+0.30(5.143) =5.040+1.543 =6.583\text{ points}.

The projected average is 6.583 points, lower than the 7.20-point prediction at the representative higher exposure. The difference arises from the lower-exposure group’s contribution. It would be incorrect to use the 70% group’s prediction alone or to substitute an unweighted average of the two stratum predictions, because the groups do not have equal proportions. This calculation also assumes the two exposures and their proportions adequately represent the planned Phase III population.

Quantifying the Risk of Inadequate Average Benefit

A point prediction does not show whether the dose decision is robust to uncertainty in the Phase II model. Set a clinically justified threshold $\delta$ before examining the resulting risk. The threshold should use the same endpoint scale and benefit direction as the model prediction. For example, if at least 5.0 points of average placebo-adjusted improvement are required, inadequate average benefit is the event $\psi<5.0$ points.

Worked Example: Posterior Probability Below a Minimum Benefit

Hypothetical case. Suppose a Bayesian analysis combines Phase II data with prespecified prior information and generates 1,000 posterior draws of the exposure–response parameters. For every draw, the target-population average is recalculated using the Phase III exposure proportions from the preceding example. Of the 1,000 resulting predictions, 182 are below the minimum acceptable benefit $\delta=5.0$ points.

The estimated posterior probability of inadequate average benefit is:

P(\psi<5.0\text{ points}\mid\text{evidence}) \approx\frac{182}{1000}=0.182=18.2\%.

Thus, under the specified model, prior, data, and target-population assumptions, 18.2% of posterior probability lies below the 5.0-point threshold. This is a measure of uncertainty about the target-population average benefit, not a statement that 18.2% of patients will receive inadequate benefit. It is also not a data-independent property of the drug: a different prior, model, threshold, or target population can change the result.

If the analysis is frequentist, a bootstrap distribution can instead show how often resampled estimates fall below the threshold and can support an interval for $\psi$. That fraction describes sampling uncertainty under the bootstrap procedure; it is not automatically a posterior probability that the true benefit is inadequate. The probability interpretation depends on the uncertainty framework used.

Diagnostics, Alternatives, and Limitations

A projection should be examined, not merely reported. Plot Phase II observations and fitted exposure–response predictions, marking the range of exposures expected in Phase III. Check that the predicted range is supported by data, that residual patterns do not show systematic model misfit, and that uncertainty is not dominated by sparse observations near the proposed exposure. Compare observed and model-predicted outcomes across relevant covariate groups. If the relationship changes by a clinically plausible modifier, assess whether the model includes that modifier and whether Phase II has enough information to estimate the difference.

Useful sensitivity checks include varying the Phase III exposure distribution within credible bounds, comparing plausible exposure–response functional forms, and checking alternative assumptions about effect modifiers or missing data. A prediction that crosses the adequacy threshold under small, credible changes is fragile even if its central estimate is favorable. Conversely, a narrow interval is not reassuring if it results from unsupported extrapolation or omits important sources of uncertainty.

Approach	What it contributes	Main caution
Phase II observed treatment contrast	A direct estimate for the Phase II population and dose.	May not represent the Phase III exposure distribution or population.
Exposure–response projection	Translates Phase II exposure–response evidence to a proposed regimen and target population.	Depends on model fit, exposure overlap, and transportability assumptions.
Simple exposure or dose comparison	Provides a transparent reference when exposure-response modeling is unavailable.	Does not fully account for exposure variability or its relationship to patient characteristics.

The projection estimates a model-based population average, not the full distribution of clinical outcomes, the probability of trial success, or a causal effect in a population that differs in unmodeled ways. Those are distinct questions. The main limitation is structural: uncertainty calculations can quantify uncertainty within a specified model, but they cannot guarantee that the model captures every relevant biological or clinical mechanism.

Common pitfall. “The mean predicted benefit at the median Phase III exposure is the Phase III population mean.” This is generally wrong: for a nonlinear exposure–response curve, the prediction at median exposure need not equal the average of predictions across the exposure distribution. The correction is to average the individual or stratum-specific model predictions using the target population’s distribution, as in $\psi=\int m(c,z;\theta)f_{\mathrm{III}}(c,z)\,dc\,dz$.

Key takeaway: Confirming a proposed Phase III dose with Phase II evidence requires transporting the exposure–response model to the planned regimen and target population, averaging predicted benefit over that population, and evaluating uncertainty against a prespecified minimum. The conclusion is only as credible as the model, exposure projection, and transport assumptions.

Check Your Understanding

Answer each question using the stated model and assumptions. Solutions follow the questions.

A model predicts benefit $m(c)$ in points. Why is prediction at median Phase III exposure not necessarily the target-population average?
In a two-stratum population, 40% have predicted benefit 8 points and 60% have predicted benefit 4 points. What is the weighted average benefit?
What does a posterior probability of 12% that $\psi<\delta$ mean, and what does it not mean?
Name two checks that would help identify an unreliable Phase III projection.

Solutions. (1) If $m$ is nonlinear, averaging $m(c)$ over exposures generally differs from applying $m$ to the median exposure; the predictions must be averaged over the population distribution. (2) $0.40(8)+0.60(4)=3.2+2.4=5.6$ points. (3) Under the specified posterior and assumptions, 12% of posterior probability for the target-population average is below the threshold; it is not the proportion of patients with inadequate individual outcomes. (4) Examples include checking Phase II exposure overlap with the Phase III range, assessing residual and covariate-stratified fit, and testing sensitivity to plausible exposure distributions or model forms.