Mixture Distributions for Multiple Plausible Scenarios

Why Combine Predictions Across Scenarios?

A PBPK prediction can depend on a discrete physiological assumption: for example, whether a clearance pathway is present, which maturation category applies, or which of two defensible organ-flow representations is used. If the alternatives remain plausible and the prediction must represent uncertainty across them, the output should account for both uncertainty within each scenario and uncertainty about which scenario applies. A mixture distribution does this by assigning an explicit weight to each alternative and combining its scenario-specific predictive distribution.

The weights are part of the analysis, not a cosmetic choice. They may represent specified plausibility weights or probabilities supported by evidence and the intended predictive question. Equal weights are not automatically justified merely because two scenarios have been considered. If no defensible probabilities can be assigned, report the scenario-specific predictions or a clearly labelled range instead of presenting an arbitrary mixture as a probability distribution.

Defining the Mixture Precisely

Write \(F_k(y)=P(Y\leq y\mid S=k)\) for the predictive cumulative distribution function (CDF) under scenario \(k\). The mixture CDF is the weighted sum of these conditional CDFs:

When the conditional distributions have densities \(f_k(y)\), the mixture density is \(f(y)=\sum_{k=1}^{K} w_k f_k(y)\). PBPK predictions are often obtained by propagating uncertain inputs through the model, so \(F_k\) may instead be represented by Monte Carlo draws or a weighted empirical distribution. The same mixture definition applies: combine conditional probabilities or empirical masses using the scenario weights.

A mixture represents a two-stage predictive statement: first a scenario is selected according to its weights, then a prediction is drawn from that scenario's conditional distribution. This is a statement about the predictive distribution, not a claim that a particular simulated individual literally changes physiological scenario during prediction. The scenarios must be defined consistently, and each \(F_k\) must condition on the scenario-specific assumptions while propagating the other uncertainties intended for the analysis.

Means and the Two Sources of Variance

Let \(\mu_k=E(Y\mid S=k)\) and \(\sigma_k^2=\operatorname{Var}(Y\mid S=k)\). The mixture mean is the weighted mean of the conditional means. Its variance has two components: average within-scenario variance and variance among scenario means.

The first term describes predictive spread that remains when the scenario is fixed. The second describes additional spread because the conditional means differ across scenarios. A mixture can therefore be more variable than any one scenario, although the exact comparison depends on the conditional variances and means. These variance summaries do not fully describe the distribution: separated scenario predictions can produce a shape that is not well summarized by a single mean and standard deviation.

Why the Variance Decomposition Holds

For a complete justification, write \(\mu=E(Y)\). Conditional on \(S=k\), add and subtract \(\mu_k\) to express the deviation from the overall mean as \(Y-\mu=(Y-\mu_k)+(\mu_k-\mu)\). Squaring and taking the conditional expectation gives:

The cross-product term is zero because \(E(Y-\mu_k\mid S=k)=0\). Averaging over scenarios, with probabilities \(w_k\), yields \(\operatorname{Var}(Y)=\sum_k w_k[\sigma_k^2+(\mu_k-\mu)^2]\), which is the stated decomposition. This identity requires finite conditional means and variances; the CDF mixture itself does not require finite variance.

Quantiles Must Come From the Combined CDF

For a probability \(q\) between 0 and 1, define the mixture quantile by \(Q(q)=\inf\{y:F(y)\geq q\}\). To obtain it, first combine the scenario CDFs at each candidate \(y\), then find where the combined CDF reaches \(q\). In general, \(Q(q)\) is not the weighted average of the scenario-specific quantiles. Quantiles are nonlinear functionals of a distribution, so averaging them does not usually preserve the probability statement that defines a quantile.

Estimating Mixture Probabilities From Simulations

When scenario \(k\) has \(N_k\) simulated outcomes \(Y_{ki}\), an empirical estimate of the probability of exceeding a threshold \(c\) is:

where \(I(Y_{ki}>c)\) equals 1 if draw \(i\) in scenario \(k\) exceeds \(c\), and 0 otherwise. This estimator first estimates each conditional exceedance probability, then weights those estimates according to scenario plausibility. The scenario weights must not be replaced by the fraction of simulations allocated to each scenario unless that allocation was deliberately chosen to equal the target weights.

Diagnostics, Alternatives, and Limitations

Before interpreting a mixture, check that scenario weights are nonnegative, sum to one, and correspond to the predictive question. Confirm that the scenario definitions are mutually exclusive and collectively cover the alternatives represented by the weights. Within each scenario, inspect simulation convergence and verify that all outcomes use consistent units, endpoints, dosing conditions, and target populations. A numerically precise mixture cannot repair an implausible scenario set or unsupported weights.

A frequent implementation error is to pool all simulation draws and treat each draw as equally probable. That is appropriate only when the simulation allocation across scenarios matches the intended weights, or when a valid correction is applied. For example, allocating 1,000 draws to each scenario does not imply equal scenario probabilities. Another error is to average scenario quantiles and label the result a mixture interval; the worked example shows why the resulting endpoints need not have the stated coverage.

The main alternatives answer different questions. Reporting scenario-specific distributions keeps each conditional prediction visible and avoids assigning unsupported scenario probabilities. A range spanning scenario-specific intervals can show the envelope of outcomes, but it is not generally a prediction interval with a specified coverage probability. Reporting only a weighted mean summarizes one feature of the mixture while discarding its spread and distributional shape. A weighted mixture is useful when the weights are defensible and a single predictive distribution is needed; its interpretation remains conditional on the included scenarios and their weights.