Dose Adaptation From Posterior Predictive Probabilities

From an Exposure Forecast to a Dose Rule

A posterior predictive forecast describes what exposure might occur under a proposed regimen, given the patient’s observed data and the model. A dose-adaptation rule turns that forecast into an action by specifying in advance which probabilities are acceptable. For example, a rule might require a sufficiently high probability of reaching a target exposure while keeping the probability of overexposure below a safety limit. The thresholds are part of the decision policy; the forecast alone does not determine them.

The probabilities must refer to the outcome that matters clinically: for example, steady-state trough concentration, AUC over a dosing interval, or peak concentration. They also refer to a particular candidate regimen and a particular prediction horizon. A probability for one exposure measure or regimen cannot be substituted for another.

Definition: Let \(Y(d)\) be a future exposure measure under candidate dose regimen \(d\), and let \(y\) denote the individual’s observed data. The posterior predictive distribution is the distribution of \(Y(d)\) conditional on \(y\), obtained by averaging the future-outcome model over the posterior distribution of the individual’s unknown parameters. For lower and upper target limits \(L\) and \(U\), define underexposure probability \(p_{\mathrm{under}}(d)=P(Y(d)<L\mid y)\), target-attainment probability \(p_{\mathrm{target}}(d)=P(L\leq Y(d)\leq U\mid y)\), and overexposure probability \(p_{\mathrm{over}}(d)=P(Y(d)>U\mid y)\).

When the exposure is continuous, the events \(Y(d)<L\), \(L\leq Y(d)\leq U\), and \(Y(d)>U\) partition all possible outcomes. Their probabilities therefore sum to one. The choice of including the boundaries in the target interval is a convention; for a continuous distribution, the probability of equalling a boundary is zero.

p_{\mathrm{under}}(d)+p_{\mathrm{target}}(d)+p_{\mathrm{over}}(d)=1.

A common policy is to identify the lowest regimen in a prespecified candidate set that satisfies a target-attainment requirement and a safety constraint. Other policies may prioritize limiting underexposure or may specify a maximum acceptable change from the current regimen. These are different rules, not competing calculations of the same probability. A probability threshold should be justified clinically and prespecified; it is not automatically supplied by Bayesian inference.

Definition: For a candidate set \(\mathcal{D}\), target-attainment threshold \(q\), and maximum acceptable overexposure probability \(r\), a feasible dose is any \(d\in\mathcal{D}\) satisfying \(p_{\mathrm{target}}(d)\geq q\) and \(p_{\mathrm{over}}(d)\leq r\). A lowest-feasible-dose rule selects the smallest feasible regimen, if one exists. If none exists, the rule returns “no feasible candidate” rather than silently selecting a dose that violates a constraint.

What the Predictive Probability Includes

Let \(\theta\) represent the individual’s unknown pharmacokinetic parameters. Conditional on the observed data \(y\), the posterior distribution \(p(\theta\mid y)\) expresses uncertainty about those parameters. A future exposure can also vary even if \(\theta\) were known, because of residual variability, occasion effects, or other stochastic components included in the model. If \(p(Y\mid\theta,d)\) is the future-outcome distribution under regimen \(d\), then the posterior predictive distribution is

p(Y(d)\mid y)=\int p(Y(d)\mid\theta,d)\,p(\theta\mid y)\,d\theta.

Thus, predictive probabilities generally account for both uncertainty about the individual parameters and future variability represented in the model. Using only a point estimate of \(\theta\) and calculating an exposure at that estimate omits parameter uncertainty and may also omit future variability. That calculation is not, by itself, a posterior predictive probability.

For example, if the future exposure is conditionally normal with mean \(m(\theta,d)\) and variance \(v(\theta,d)\), the predictive distribution is generally a mixture of normal distributions over posterior draws of \(\theta\), not necessarily a single normal distribution. A normal approximation can be useful, but it should be identified as an approximation and checked for suitability.

Justification: The law of total variance separates predictive uncertainty into uncertainty in the conditional mean and average conditional variability:

\operatorname{Var}(Y(d)\mid y) = \operatorname{Var}\!\left(E[Y(d)\mid\theta,d,y]\mid y\right) + E\!\left[\operatorname{Var}(Y(d)\mid\theta,d,y)\mid y\right].

The first term is variation in the expected exposure across the posterior distribution of \(\theta\); the second is average future-outcome variance conditional on \(\theta\). Both contribute to the predictive spread. This identity follows by writing \(Y-E(Y\mid y)=[Y-E(Y\mid\theta,d,y)]+[E(Y\mid\theta,d,y)-E(Y\mid y)]\), squaring, and taking expectations conditional on \(y\). The cross-product has expectation zero because the conditional expectation of \(Y-E(Y\mid\theta,d,y)\), given \(\theta,d,y\), is zero.

Worked Dose Rules

Worked Example: Requiring Target Attainment and a Safety Limit

Suppose, hypothetically, the target exposure is \(8\) to \(12\) units. The prespecified policy requires \(p_{\mathrm{target}}(d)\geq 0.65\) and \(p_{\mathrm{over}}(d)\leq 0.20\), and selects the lowest qualifying dose among 80, 100, and 120 mg. Assume the dose-specific predictive distributions are normal with the means and standard deviations shown in the calculations below. These are hypothetical forecasts, not observations from a real study.

At 80 mg, suppose \(Y(80)\mid y\sim N(8,3^2)\). The target probability is

p_{\mathrm{target}}(80) = \Phi\!\left(\frac{12-8}{3}\right) - \Phi\!\left(\frac{8-8}{3}\right) = \Phi(4/3)-\Phi(0) \approx 0.9088-0.5000 =0.4088.

The underexposure probability is \(\Phi(0)=0.5000\), and the overexposure probability is \(1-\Phi(4/3)\approx0.0912\). Although the overexposure constraint is met, the target-attainment requirement is not.

At 100 mg, suppose \(Y(100)\mid y\sim N(10,2^2)\). Then

p_{\mathrm{target}}(100) = \Phi\!\left(\frac{12-10}{2}\right) - \Phi\!\left(\frac{8-10}{2}\right) = \Phi(1)-\Phi(-1) \approx 0.8413-0.1587 =0.6826.

Each tail probability is approximately \(0.1587\), so \(p_{\mathrm{over}}(100)\approx0.1587\). Both requirements are met. At 120 mg, suppose \(Y(120)\mid y\sim N(12,2.4^2)\). Its target probability is \(\Phi(0)-\Phi(-1.6667)\approx0.5000-0.0478=0.4522\), and its overexposure probability is \(0.5000\). This dose fails both requirements. The lowest feasible candidate is therefore 100 mg.

Worked Example: Limiting Both Underexposure and Overexposure

Suppose a hypothetical policy selects the lowest dose for which underexposure probability is at most \(0.10\) and overexposure probability is at most \(0.05\), with target limits \(8\) and \(12\) units. Candidate doses are 80, 90, and 100 mg. Assume their predictive distributions are normal with standard deviation \(1.2\) units and respective means \(8.5\), \(9.5\), and \(10\) units.

For 80 mg, the underexposure probability is \(\Phi((8-8.5)/1.2)=\Phi(-0.4167)\approx0.3385\). Its overexposure probability is \(1-\Phi((12-8.5)/1.2)=1-\Phi(2.9167)\approx0.0018\). It fails the underexposure constraint.

For 90 mg, the two probabilities are

p_{\mathrm{under}}(90)=\Phi\!\left(\frac{8-9.5}{1.2}\right) =\Phi(-1.25)\approx0.1056,

p_{\mathrm{over}}(90)=1-\Phi\!\left(\frac{12-9.5}{1.2}\right) =1-\Phi(2.0833)\approx0.0186.

The safety constraint is met, but the underexposure probability is just above \(0.10\), so 90 mg does not qualify. For 100 mg, \(p_{\mathrm{under}}(100)=\Phi(-1.6667)\approx0.0478\) and \(p_{\mathrm{over}}(100)=1-\Phi(1.6667)\approx0.0478\). Both limits are met, so 100 mg is the only qualifying candidate and is selected by this rule.

Worked Example: Applying the Rule to Predictive Simulations

Normality may be an inadequate approximation when the predictive distribution is skewed or otherwise irregular. Suppose 4,000 hypothetical predictive draws are generated for each candidate dose, using posterior parameter draws and the model’s future-outcome variability. The target is 8 to 12 units; the rule requires target probability of at least \(0.80\) and overexposure probability no greater than \(0.05\).

At 60 mg, suppose 1,360 draws are below target, 2,560 are in the target interval, and 80 are above target. The estimated probabilities are \(1360/4000=0.3400\), \(2560/4000=0.6400\), and \(80/4000=0.0200\), respectively. This dose fails the target-attainment requirement.

At 80 mg, suppose the counts are 520 below target, 3,310 in the target interval, and 170 above target. The estimates are \(0.1300\), \(0.8275\), and \(0.0425\). Both requirements are met. At 100 mg, suppose the corresponding counts are 160, 3,240, and 600, giving target probability \(0.8100\) and overexposure probability \(0.1500\); this dose fails the safety limit. The lowest qualifying dose is 80 mg.

Simulation estimates have Monte Carlo error. For the 80 mg target probability, the estimated Monte Carlo standard error is \(\sqrt{0.8275(1-0.8275)/4000}\approx0.0060\). For the overexposure probability it is \(\sqrt{0.0425(1-0.0425)/4000}\approx0.0032\). The estimated overexposure probability is below \(0.05\), but fairly close; more draws can help determine whether numerical uncertainty could change the classification. Monte Carlo precision does not establish that the model itself is adequate.

Assumptions, Diagnostics, and Limits

A probability-based rule is meaningful only if its forecast and its inputs match the intended decision. Before applying it, check the following:

Outcome and horizon: the forecast is for the exposure measure, time period, and regimen that the target limits describe.
Regimen representation: the proposed dose, dosing interval, route, and relevant assumptions about adherence are represented in the predictive model.
Predictive uncertainty: the calculation integrates over the individual’s posterior uncertainty and includes the future variability required by the clinical question.
Thresholds and candidates: limits, probability thresholds, and allowable doses are specified before choosing among candidates, with units and any dose-rounding rules made explicit.
Numerical stability: analytic calculations use the stated distribution and parameterization correctly; simulation calculations use enough draws to estimate probabilities near decision thresholds with useful precision.

Warning signs include a dose recommendation that changes substantially under small reasonable changes in the candidate grid, predicted probabilities that are highly sensitive to model assumptions, or simulated probabilities that remain noisy around a policy threshold. A probability rule may also return no feasible dose: for example, every candidate may have too much underexposure risk, or doses that meet the target requirement may violate the safety constraint. That outcome is informative. It calls for review of the candidate regimens, the forecast, or the policy rather than an unreported relaxation of a constraint.

The main alternative is a point-forecast rule, such as selecting a dose whose predicted mean exposure lies inside the target interval. It is simpler, but it does not say how likely an individual future exposure is to fall outside the interval. A posterior predictive probability rule makes that uncertainty explicit, at the cost of relying on the predictive model and chosen probability thresholds. The dose with the highest target probability is another possible criterion, but it need not satisfy a separate safety limit and need not be the lowest acceptable dose.

Common Pitfall

Common mistake: “The predicted mean is inside the target interval, so the dose has a high probability of target attainment.” A mean inside the interval does not determine the probability of being inside it; the predictive spread and shape matter. Calculate the probability from the full predictive distribution. For example, a forecast centered at the lower target boundary can still assign substantial probability to underexposure, as the 80 mg calculation above shows.

A second error is to treat a threshold such as \(0.80\) as a universal standard. It is a policy choice, not a mathematical consequence of the posterior. Probability thresholds and safety constraints should have an explicit clinical rationale, and a rule that satisfies them should not be described as proving that the dose is safe or effective for every future outcome.

Key takeaway: For each candidate regimen, calculate the posterior predictive probabilities of the clinically defined exposure regions. Apply a prespecified rule to those probabilities, report its constraints and units, and state clearly when no candidate qualifies.

Check Your Understanding

Try each question, then compare your reasoning with the solutions.

If a predictive distribution assigns probabilities \(0.18\) to underexposure and \(0.07\) to overexposure, what is the target-attainment probability, assuming the three regions partition the outcomes?
A rule requires target probability at least \(0.75\) and overexposure probability at most \(0.05\). A candidate has probabilities \(0.80\) and \(0.06\), respectively. Does it qualify? Why?
Why can a dose with predicted mean exposure inside the target range still have a low target-attainment probability?
What two sources of uncertainty are represented by the predictive variance decomposition?

Solutions. (1) The probabilities sum to one, so \(p_{\mathrm{target}}=1-0.18-0.07=0.75\). (2) No. Although \(0.80\geq0.75\), the overexposure probability \(0.06\) exceeds its maximum \(0.05\). Both constraints must be met. (3) The predictive distribution may be wide or have substantial mass in one or both tails; the mean alone does not determine the probability inside the interval. (4) The decomposition includes variation in conditional mean exposure across the posterior parameter distribution and average future-outcome variability conditional on those parameters.