Monte Carlo Estimation of PTA

From a Target Event to a Simulated Proportion

Probability of target attainment (PTA) is the probability that an individual drawn from a specified population achieves a prespecified exposure target at a specified dose. In a nonlinear mixed-effects model, individual parameter vectors vary across the population, so the exposure metric can vary even when the dose is fixed. Monte Carlo estimation approximates the population probability by simulating individual vectors, calculating the metric for each vector, and counting target successes.

The target, dose, exposure metric, and population distribution define what probability is being estimated. The calculation does not itself decide whether the target is clinically appropriate or what PTA is sufficient for a dose decision; those are distinct questions.

For a lower-bound target, \(\mathcal{T}\) may be the set of metric values at or above a specified target \(M^*\). The population distribution must represent the intended population: this includes appropriate covariate values and the model's random-effects distribution. Individual parameter vectors must be generated from that distribution, not treated as if empirical Bayes estimates for observed individuals were population draws. These distinctions are important in the earlier material on empirical Bayes estimates and simulation uncertainty.

Why the Success Count Estimates PTA

Assume for the calculation that the simulated individual vectors are independent draws from the specified population distribution, the fitted model and dose are held fixed, and the exposure calculation and target rule are applied consistently to every draw. Then the success indicators \(I_i(d)\) are independent Bernoulli random variables with success probability \(\pi(d;\psi)\). Their sum \(S=\sum_{i=1}^{N}I_i(d)\) has a binomial distribution with parameters \(N\) and \(\pi(d;\psi)\).

The first equality follows from linearity of expectation: each indicator has expectation \(\pi(d;\psi)\), so the average of \(N\) indicators has that same expectation. For the variance, independence makes the variance of the sum equal to the sum of the variances. A Bernoulli variable with success probability \(\pi\) has variance \(\pi(1-\pi)\); dividing the sum by \(N\) therefore gives variance \(\pi(1-\pi)/N\).

Because the true PTA is not known, estimate its Monte Carlo standard error (MC SE) by substituting \(\widehat{\pi}\) for \(\pi\):

For sufficiently large \(N\), a normal-approximation interval is \(\widehat{\pi}\pm1.96SE_{\mathrm{MC}}\), with endpoints restricted to the probability range if necessary. This interval describes Monte Carlo error conditional on the fitted model and the stated simulation design. It does not include uncertainty in the fitted parameters, uncertainty in the target, or uncertainty about whether the model represents the real population. For a small number of successes or failures, a Wilson binomial interval is generally preferable to the simple normal approximation.

Comparing Candidate Doses With Shared Simulated Individuals

When comparing two doses, using the same simulated individual parameter vectors at both doses often improves precision. The comparison then respects that the same underlying individual may attain the target at one dose but not the other. Define the per-individual difference \(D_i=I_i(d_2)-I_i(d_1)\). It takes values \(-1,0,\) or \(1\), and its average estimates the PTA difference \(\pi(d_2;\psi)-\pi(d_1;\psi)\). Because each simulated individual contributes a paired outcome, the standard error must be calculated from the variability of the \(D_i\), not by treating the two dose estimates as independent.

Planning Replicates and Checking Precision

The MC SE decreases at approximately the rate \(1/\sqrt{N}\): multiplying the replicate count by four approximately halves the MC SE. For a planned normal-approximation half-width \(h\), the variance formula gives the planning relation \(N\approx z^2\pi(1-\pi)/h^2\), where \(z\) is the chosen normal critical value. A pilot estimate can provide a planning value for \(\pi\). If a conservative count is needed without a useful pilot, the bound \(\pi(1-\pi)\leq0.25\) gives a count that covers the largest possible Bernoulli variance.

Diagnostics, Alternatives, and Limits

A credible Monte Carlo estimate requires more than a large replicate count. Check that generated vectors are valid, that the covariance and dependence structure of individual parameters are preserved, and that each vector is evaluated with the intended dose, metric, and target rule. Inspect the number and fraction of failed simulations. Excluding failed parameter draws without justification changes the simulated distribution and can bias PTA; investigate the failure mechanism and follow a prespecified handling approach.

Useful checks include comparing success counts across independent batches, plotting the running estimate against replicate count, and confirming that reruns with independent random-number streams give estimates compatible with their MC SEs. Large batch-to-batch variation, a running estimate that continues to drift, or a substantial fraction of failed calculations calls for investigation. A fixed random seed supports reproducibility but does not by itself ensure independent draws or adequate precision.

When the target event can be integrated analytically or with reliable numerical quadrature, those approaches can avoid random simulation error. Monte Carlo is often practical when individual parameter vectors are high-dimensional or the exposure calculation is nonlinear and cumbersome to integrate. Its trade-off is computational cost and finite-replicate error. Increasing \(N\) reduces that numerical error, but does not correct model misspecification, an inappropriate population distribution, or an incorrectly defined target.

For comparing doses, independently simulating two unrelated populations can waste information and make the difference less precise. Shared simulated individuals create paired outcomes and allow the standard error to reflect the within-individual relationship. Conversely, calculating an independent-proportions standard error for paired results ignores that relationship. Neither the paired nor unpaired calculation should be interpreted as total uncertainty unless the other uncertainty sources have been addressed separately.