Worked Example: PTA for an AUC Threshold

From Simulated Exposure to a Dose Decision

A population pharmacokinetic model can predict exposure for virtual individuals at several candidate doses. For an AUC target, the dose question can be made concrete: what proportion of the target population is expected to have an AUC at or above the specified threshold? That proportion is the probability of target attainment (PTA). A dose-grid simulation estimates PTA at each dose, while Monte Carlo error describes the random variation in the estimate caused by using a finite number of simulated individuals.

This tutorial focuses on a single, fixed lower AUC threshold. It does not treat a joint endpoint or a therapeutic window. Those choices require different attainment definitions. A dose recommendation also needs a rule stated in advance: for example, choose the lowest dose whose estimated PTA is at least 90%. The simulation estimates how the modelled population performs under each candidate dose; it does not by itself establish that the target or the selection rule is clinically appropriate.

Definition: For dose \(D\), an AUC threshold \(T\), and a target population represented by the model, define the individual attainment indicator as \(I_i(D)=1\) when individual \(i\)'s simulated AUC is at least \(T\), and \(I_i(D)=0\) otherwise. If \(N\) virtual individuals are simulated, the estimated PTA is \(\widehat{\mathrm{PTA}}(D)=\frac{1}{N}\sum_{i=1}^{N} I_i(D)\).

Specify the Simulation Before Calculating PTA

For each simulated individual, the model supplies the pharmacokinetic parameters needed to calculate exposure. In a linear, time-invariant model at steady state, the AUC over a dosing interval is determined by dose, bioavailability, clearance, and dosing interval. For an intravenous drug given once daily, with daily dose \(D\) and individual clearance \(CL_i\), the steady-state 24-hour AUC is \(AUC_{24,i}(D)=D/CL_i\). With \(D\) in mg and \(CL_i\) in L/h, the resulting units are mg·h/L. For an extravascular regimen, the corresponding linear-model expression includes bioavailability \(F_i\): \(AUC_{\tau,i}=F_iD/CL_i\), where \(D\) is the amount given during that interval.

The formula is a useful check, not a substitute for the model when its assumptions do not hold. Nonlinear elimination, time-varying clearance, or a more complex exposure definition may require calculating AUC from the simulated concentration-time profile. The model, dosing regimen, steady-state or other time point, target population, AUC definition, and threshold must all match the dose question.

A typical simulation draws a set of virtual individuals from the specified target population, including between-subject variation and relevant covariates. For each candidate dose, it calculates one AUC per individual and applies the same attainment rule. If the objective is between-subject PTA for an individual's model-predicted exposure, measurement error or residual observation noise should not be added as though it were an extra source of biological variation. Conversely, if the target is explicitly defined for noisy observed measurements, the simulation must represent that target appropriately. The earlier material on avoiding double counting of variability is directly relevant to this choice.

The basic Monte Carlo standard error treats the simulated attainment indicators as independent Bernoulli outcomes with a common model-implied probability. Under that setup, if the true PTA is \(p_D\), then \(\widehat{\mathrm{PTA}}(D)\) has approximate standard error \(\sqrt{p_D(1-p_D)/N}\); in practice, estimate it by substituting \(\widehat{\mathrm{PTA}}(D)\) for \(p_D\). The standard error measures simulation error, not uncertainty about whether the pharmacokinetic model, target, or population is correct.

SE_{\mathrm{MC}}\{\widehat{\mathrm{PTA}}(D)\} \approx \sqrt{\frac{\widehat{\mathrm{PTA}}(D)\left[1-\widehat{\mathrm{PTA}}(D)\right]}{N}}

Worked Examples

Worked Example: Calculating PTA From Individual AUC Values

Consider a hypothetical intravenous drug at steady state, given once daily. The fixed target is \(AUC_{24}\geq 50\) mg·h/L, and the candidate daily dose is 400 mg. Suppose 20 hypothetical virtual individuals have simulated clearances, in L/h, of 5, 6, 7, 8, 8.5, 9, 10, 11, 6.5, 7.5, 8.2, 9.5, 5.5, 7.8, 10.5, 6.8, 8.8, 9.2, 7.2, and 8.4. Assume the linear steady-state relationship \(AUC_{24,i}=D/CL_i\) applies.

At 400 mg, an individual attains the target when \(400/CL_i\geq 50\), which is equivalent to \(CL_i\leq 8\) L/h. Ten of the 20 clearances meet that condition: 5, 6, 7, 8, 6.5, 7.5, 5.5, 7.8, 6.8, and 7.2 L/h. Thus,

\widehat{\mathrm{PTA}}(400) = \frac{10}{20} = 0.500

The estimated PTA is 50.0%. Its estimated Monte Carlo standard error is \(\sqrt{0.500(1-0.500)/20}=\sqrt{0.0125}=0.1118\), or 11.18 percentage points. For illustration, a 95% Wilson interval for the simulated proportion is approximately 0.299 to 0.701. This wide interval reflects the small simulation size. It is an interval for the Bernoulli probability represented by this simulation, under its assumptions; it does not include uncertainty in the model or target definition.

Worked Example: Selecting the Lowest Dose That Meets a PTA Criterion

In a separate hypothetical simulation, \(N=2{,}000\) virtual individuals are evaluated at each candidate daily dose for the same kind of fixed AUC threshold. The prespecified rule is to choose the lowest dose with estimated PTA at least 0.90. The simulated attainment counts are:

Dose (mg/day)	Attainers out of 2,000	Estimated PTA
300	1,250	0.625
400	1,640	0.820
500	1,840	0.920
600	1,930	0.965

For 500 mg/day, \(\widehat{\mathrm{PTA}}=1{,}840/2{,}000=0.920\). Its Monte Carlo standard error is \(\sqrt{0.920(0.080)/2{,}000}=\sqrt{0.0000368}=0.006066\), or about 0.607 percentage points. A normal-approximation 95% interval is \(0.920\pm1.96(0.006066)\), approximately 0.908 to 0.932. The interval describes simulation precision, conditional on the model and simulation setup.

The estimated PTAs at 300 and 400 mg/day fall below 0.90, while the estimate at 500 mg/day is 0.920. Therefore, under the stated point-estimate rule, select 500 mg/day: it is the lowest candidate dose meeting the criterion. The result does not mean that every individual attains the target, nor does it establish that a 90% PTA criterion is clinically justified.

Worked Example: Planning Simulation Size for a Precise PTA Estimate

Suppose the aim is for a 95% normal-approximation interval to have a half-width no greater than 0.010 when the PTA is near 0.90. Using \(p=0.90\) for planning, the approximate half-width is \(1.96\sqrt{p(1-p)/N}\). Requiring this to be at most 0.010 gives:

1.96\sqrt{\frac{0.90(0.10)}{N}}\leq 0.010 \quad\Longrightarrow\quad N\geq \frac{1.96^2(0.90)(0.10)}{0.010^2} = 3457.44

Round up to \(N=3{,}458\) virtual individuals. This planning calculation uses an anticipated PTA; if that value is uncertain, a conservative choice is \(p=0.50\), which maximizes \(p(1-p)\) and requires more simulations for the same half-width. For example, at \(p=0.50\), the same calculation gives \(N\geq 9{,}604\). These values plan Monte Carlo precision only; a larger \(N\) cannot correct a misspecified model.

Why the Proportion and Its Error Formula Work

For a fixed dose, let \(I_i(D)\) equal one if simulated individual \(i\) attains the threshold and zero otherwise. Under independent, identically distributed simulation draws, each indicator is Bernoulli with probability \(p_D\). Its mean is \(p_D\), and its variance is \(p_D(1-p_D)\). Since the estimated PTA is the average of \(N\) such indicators, its expected value is \(p_D\), and its variance is the indicator variance divided by \(N\):

E\{\widehat{\mathrm{PTA}}(D)\}=p_D, \qquad \operatorname{Var}\{\widehat{\mathrm{PTA}}(D)\} = \operatorname{Var}\left\{\frac{1}{N}\sum_{i=1}^{N}I_i(D)\right\} = \frac{p_D(1-p_D)}{N}

The last equality follows because independent indicators have zero covariance, so the variance of their sum is the sum of their variances. Taking the square root gives the standard error; replacing the unknown \(p_D\) by the observed proportion gives its plug-in estimate. This justification depends on independent simulated individuals. If simulation draws are correlated, or the design uses clustered sampling that changes the variance, the simple binomial standard error may be inappropriate.

Diagnostics, Alternatives, and Limitations

A dose grid is easier to interpret when its assumptions and outputs are checked before applying the decision rule. Useful checks include:

Verify the exposure calculation. Check dose units, clearance units, interval, bioavailability where relevant, and whether the AUC is at steady state or over another specified period.
Inspect the simulated population. Confirm that covariates and random effects represent the intended target population, and that simulated parameter values are plausible under the fitted model.
Check the threshold classification. Apply the same inequality, AUC definition, and units to every dose. Decide in advance how values exactly equal to the threshold are classified.
Look for unexplained dose-grid patterns. Under a linear model with fixed individual parameters and a lower AUC threshold, increasing dose cannot turn an attaining individual's AUC into a nonattaining AUC. A substantial decrease in PTA across increasing doses may signal coding, regimen, or simulation problems. Sampling variation can still create small irregularities if separate random populations are used at each dose.
Check Monte Carlo stability. Increase \(N\), or repeat the simulation with independent random seeds, and see whether PTA and the dose selected by the rule are stable. Instability near the decision threshold signals that more simulation may be useful.

One alternative is to derive PTA analytically when the relationship between AUC and model parameters is simple and the parameter distribution permits a tractable probability calculation. For example, under the intravenous linear relationship, attainment is equivalent to \(CL_i\leq D/T\); if the clearance distribution function is known, PTA is that distribution's cumulative probability at \(D/T\). This avoids Monte Carlo error but still depends on the distributional assumptions and may not be convenient for complex covariate structures or nonlinear models. Simulation is flexible for those models, but its estimate has finite-sample error.

A second practical choice is to simulate independently at each dose or reuse the same virtual individuals across doses. Reusing individuals makes dose comparisons easier to interpret because each individual's exposure changes across the grid under the same underlying covariates and random effects. It can also make the attainment pattern monotone for a fixed threshold under a dose-proportional model. Independent simulations at each dose are not inherently invalid, but may add avoidable Monte Carlo variation to comparisons. In either design, estimate the PTA at each dose from the simulated attainment indicators, and do not mistake the resulting Monte Carlo interval for a confidence interval that accounts for all sources of scientific uncertainty.

Finally, the selected dose is conditional on the candidate grid, model, target population, AUC threshold, and decision rule. A dose between grid points could also meet the criterion, so a coarse grid may only identify the lowest passing candidate, not the exact boundary dose. Model uncertainty, parameter uncertainty, adherence, clinical consequences of target failure, and safety constraints require separate consideration. The dose-grid result is a model-based input to dose justification, not a guarantee of individual response.

Common pitfall: A simulated PTA of 0.92 is sometimes described as a 92% probability that the selected dose is correct. It instead estimates that 92% of individuals in the modelled population attain the specified AUC threshold at that dose, subject to Monte Carlo error and the model assumptions. It does not quantify the probability that the model or dose recommendation is correct.

Check Your Understanding

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

At a dose of 300 mg, 720 of 1,000 simulated individuals attain the AUC threshold. What is the estimated PTA?
For the estimate in question 1, calculate the plug-in Monte Carlo standard error.
A prespecified rule selects the lowest dose with estimated PTA at least 0.85. Dose A has PTA 0.82 and Dose B has PTA 0.87. Which dose is selected if these are the only candidates?
Why does increasing the number of virtual individuals reduce Monte Carlo error but not correct a wrong clearance model?

Solutions. (1) The estimate is \(720/1{,}000=0.720\), or 72.0%. (2) \(SE_{\mathrm{MC}}=\sqrt{0.720(0.280)/1{,}000}=\sqrt{0.0002016}=0.01420\), about 1.42 percentage points. (3) Dose B is selected because it is the only candidate meeting the 0.85 criterion. (4) A larger simulation reduces the sampling variation of the simulated proportion around the probability implied by the model. It does not change that model-implied probability or make incorrect assumptions about clearance, covariates, or the target population correct.