Posterior Probability of Meeting an Exposure Target

Why a Target Probability Uses a Predictive Distribution

An exposure target is often stated as a range: exposure should be high enough to support efficacy and low enough to avoid unacceptable toxicity. For an individual, however, the relevant question is not only what exposure is expected under a proposed regimen. It is also how uncertain that expectation is, and how much future variability remains. A posterior predictive probability combines these sources of uncertainty to answer: given the individual’s observed data, how likely is a future exposure to lie inside the specified range?

Let \(D\) denote the individual’s observed data, \(\theta\) the unknown individual-specific parameters, and \(Y^*\) a future exposure of interest under a specified regimen and prediction occasion. Let \(L\) and \(U\) be the lower efficacy limit and upper safety limit, with \(L\leq U\). The probability is conditional on the observed data and on the model, regimen, and prediction conditions used to define \(Y^*\). It is not a probability that the limits themselves are correct, nor does it by itself prescribe a dose change.

This differs from checking whether a point summary, such as the posterior mean exposure, lies in the range. A posterior mean inside the range does not imply a high probability of being inside it: substantial uncertainty or residual variability could put much of the predictive distribution outside. Conversely, a point summary outside the range does not determine the probability of meeting the target.

Definition by Posterior Predictive Integration

Suppose a model specifies the distribution of future exposure given the individual parameters and observed data, \(p(y^*\mid\theta,D)\), and the individual posterior distribution \(p(\theta\mid D)\). The posterior predictive distribution averages the future-exposure distribution over that posterior. This is the key step: plausible parameter values contribute in proportion to their posterior probability, and future variability conditional on those values is retained.

For a closed interval, the corresponding target probability is

If the predictive distribution has cumulative distribution function \(F_{Y^*\mid D}\), the exact closed-interval expression is

where \(F_{Y^*\mid D}(L^-)=P(Y^*<L\mid D)\) is the left limit of the CDF at \(L\). For a continuous predictive distribution there is no probability mass at \(L\), so \(F(L^-)=F(L)\) and the familiar difference \(F(U)-F(L)\) is valid. For a discrete or mixed distribution, the left limit matters: using \(F(L)\) would omit any mass exactly at the lower boundary.

Why Parameter and Future Variability Both Matter

A short derivation clarifies the role of the two uncertainty sources. Suppose, specifically, that conditional on \(D\), \(\theta\sim N(m,V)\), and that \(Y^*=\theta+\varepsilon\), where \(\varepsilon\mid D\sim N(0,R)\) and \(\varepsilon\) is independent of \(\theta\) conditional on \(D\). These are conditional assumptions: marginal independence of \(\varepsilon\) and \(\theta\) alone is not sufficient, because observing \(D\) can induce dependence or provide information about \(\varepsilon\).

Under these assumptions, the conditional moment-generating function of \(Y^*\) is the product of the two normal moment-generating functions. For any real \(t\),

This is the moment-generating function of a normal variable with mean \(m\) and variance \(V+R\). Therefore, in this specific model, \(Y^*\mid D\sim N(m,V+R)\). The variance \(V\) represents posterior uncertainty about the individual parameter, while \(R\) represents future variability conditional on that parameter. This normal result is not automatic in nonlinear PK models or when the conditional predictive error depends on \(\theta\); in those settings, use the model’s full posterior predictive distribution.

Computing the Probability From Posterior Predictive Draws

For complex models, the integral defining the target probability may not have a closed form. A direct simulation approach is to draw parameter values from \(p(\theta\mid D)\), then draw a future exposure from the model conditional on each parameter draw and the prediction conditions. Each simulated future exposure is classified as inside or outside the target. The fraction inside estimates the posterior predictive probability.

If \(M\) independent predictive draws are generated, write them as \(Y^{*(1)},\ldots,Y^{*(M)}\). The Monte Carlo estimate is

where \(I\{\cdot\}\) equals 1 when the condition holds and 0 otherwise. When the draws are independent, the estimated Monte Carlo standard error is \(\sqrt{\widehat{P}(1-\widehat{P})/M}\). This quantifies simulation error in estimating the probability, not uncertainty about whether the model is adequate. If draws are dependent, as can occur with posterior simulation algorithms, the independent-draw standard error is not generally appropriate; dependence must be accounted for when assessing Monte Carlo precision.

Atoms at Target Boundaries

A future exposure distribution need not be continuous. For example, rounding, a discrete endpoint, or a model with a probability of zero exposure can create point masses. In that case, boundary inclusion is part of the target definition, and the closed interval formula must include mass at both \(L\) and \(U\).

Assumptions, Diagnostics, and Alternative Summaries

A target probability is only as meaningful as the predictive distribution used to calculate it. Before interpreting the result, check that the forecast matches the intended clinical quantity and time: for example, an AUC over the relevant interval is not interchangeable with a trough at one time point. Verify that the regimen, dosing history, adherence assumptions, and any prediction conditions used to construct the forecast are appropriate for the individual. The probability is conditional on these choices.

Also check how the calculation represents uncertainty. A plug-in calculation that fixes \(\theta\) at a point estimate and includes only residual variability omits posterior uncertainty in \(\theta\). Conversely, using a posterior distribution for a mean exposure while omitting meaningful future variability can make the predictive distribution too narrow. In either case the resulting target probability may be poorly calibrated to the intended future outcome. For nonlinear relationships, transforming a posterior mean parameter is generally not equivalent to averaging the transformed exposure over the posterior; prediction should use the model’s exposure function for each draw.

Useful practical checks include confirming that predictive draws have plausible units and ranges, that the target event is coded with the correct strict or inclusive boundaries, and that repeated simulation runs give sufficiently stable estimates. For simulation estimates, report \(M\), the estimated probability, and an appropriate measure of Monte Carlo precision. A narrow Monte Carlo error does not establish that the structural model, parameter assumptions, or target limits are clinically suitable.

A common mistake is to see that the posterior mean lies between \(L\) and \(U\) and conclude that the individual is likely to meet the target. The mean alone does not determine the probability: distributions with the same mean can have different variances, shapes, and probabilities outside the interval. The correction is to calculate or simulate the posterior predictive probability for the actual target event. A second mistake is to use \(F(U)-F(L)\) for a closed interval without checking continuity. The correction is to use \(F(U)-F(L^-)\), which includes any predictive mass exactly at the lower limit.