Common Errors in PTA Simulation

Why Small Simulation Errors Can Change a Dose Decision

Probability of target attainment (PTA) is a probability calculated for a specified drug regimen, target, and population. A simulation can implement the chosen model correctly in a computational sense and still answer the wrong question. A unit conversion error can reverse pass/fail classifications; a dosing-interval error can distort the exposure metric; and a sampling scheme that confuses population variability with parameter uncertainty can produce a PTA for a population that was never intended.

The introductory material on dose selection by simulation establishes the basic role of PTA in comparing regimens. Here the emphasis is on auditing the simulation that produces it: precisely what is sampled, which exposure is evaluated, and what the resulting probability means. The examples use hypothetical data and targets.

The event \(I_i(d)=1\) must be written down before running the simulation. For an AUC/MIC target, it might be \(AUC_{24,i}(d)/MIC_i \geq T\); for a time-above-MIC target, it might be \(T_{>MIC,i}(d)/\tau \geq q\), where \(\tau\) is the dosing interval and \(q\) is a target fraction. These are not interchangeable definitions. Each requires the correct exposure period, units, threshold, and regimen.

Audit the Target and Exposure Units

A ratio can be dimensionless or can carry units, depending on its numerator and denominator. For example, AUC in mg·h/L divided by MIC in mg/L has units of hours. If the target is expressed in those units, compare quantities after converting both to a common unit system. Do not compare the numerical values alone when the units differ. Likewise, a fraction of an interval is dimensionless, while time above the threshold is measured in time units.

A reliable implementation records units alongside every simulated quantity and checks the event definition against a hand calculation. This is especially important when data sources report MIC in different concentration units, or when a target has been copied from a source using a different exposure convention. The AUC period must also match: \(AUC_{24}\) is not automatically the same quantity as AUC over one dosing interval when \(\tau\) is not 24 hours.

Use the Actual Dosing Interval and Exposure Profile

Time-above-target metrics depend on both the concentration-time profile and the interval over which it is evaluated. For a steady-state regimen with interval \(\tau\), define \(T_{>MIC}\) as the total time within that interval for which the concentration exceeds MIC. Then the fraction above MIC is \(T_{>MIC}/\tau\). The denominator is the actual interval for the regimen being assessed, not a default interval or the duration of a convenient simulation grid.

In practice, the interval must agree across the dose schedule, simulated concentration-time profile, exposure calculation, and target definition. For a grid-based profile, check that the grid covers the full interval and resolves crossings adequately. A coarse grid can miss a short period below MIC or above a toxicity threshold. If the simulation concerns a treatment period rather than steady state, state which doses and time window define the metric; do not silently substitute a steady-state exposure.

Sample the Intended Population, Not a Convenient Distribution

A population PTA averages over the population that the dose recommendation is meant to address. Individual parameter draws should therefore reflect the model’s population distribution, including relevant covariates and the estimated random-effects covariance structure. Parameters that are correlated in the model should not be sampled independently unless independence is a justified feature of the model. Sampling each parameter from a marginal distribution while discarding correlation can create combinations of clearance, volume, or other parameters that are implausible under the fitted model.

It is also necessary to distinguish population variability from uncertainty in estimated model parameters. Population variability describes differences among individuals and belongs in the virtual population. Parameter uncertainty describes uncertainty about quantities such as fixed effects, random-effects variances, or covariate effects estimated from finite data. Treating a fresh draw of a fixed-effect estimate as if it were a new individual’s parameter can mix these sources and change the question. A common approach is to simulate individuals conditional on a fitted parameter set, then assess parameter uncertainty separately, for example by repeating the population simulation across parameter draws or bootstrap fits.

A parameter-uncertainty analysis answers a different question from the preceding Monte Carlo calculation. For example, if a bootstrap produces several plausible fitted parameter sets, each can yield its own PTA estimate. Variation across those estimates reflects sensitivity to parameter estimation as well as any simulation noise that remains. It should not be reported as though it were simply the spread of individual outcomes.

A Short Justification for the Monte Carlo Error Check

The standard error formula provides a useful check on whether an apparent difference between candidate doses is larger than simulation noise. Let \(I_i\) be independent indicators with \(P(I_i=1)=p\). Since \(I_i\) takes values 0 and 1, its expectation is \(p\) and its variance is \(p(1-p)\). For \(\widehat p=N^{-1}\sum I_i\), independence gives the variance of the sum as the sum of the variances:

Replacing the unknown \(p\) with \(\widehat p\) gives the estimated Monte Carlo standard error \(\sqrt{\widehat p(1-\widehat p)/N}\). This justification relies on independent, identically distributed simulation draws for the specified population. It does not correct a misspecified model, biased sampling, incorrect units, or an incorrect target event. More simulations reduce Monte Carlo error; they cannot repair a simulation that answers the wrong question.

Common Errors, Diagnostics, and Alternatives

Several checks can reveal errors before the results are used for dose selection. Recompute a few simulated subjects’ exposure metrics by hand; print the units and interval in the output; inspect whether target attainment changes plausibly with dose; and verify that the generated covariates and random effects match the intended population. A suspiciously abrupt PTA change, an impossible exposure scale, or a result that changes greatly when a modest number of draws is added warrants investigation. These checks complement, rather than replace, the diagnostics for implausible PTA curves.

The main alternatives are useful for different reasons, but none removes the need to define the target and sampling population correctly. Direct Monte Carlo simulation is flexible for nonlinear models and complicated regimens, but has finite simulation error. Deterministic numerical integration can reduce random simulation noise when the parameter dimension and integration structure make it practical, but can become difficult in high dimensions. Bootstrap or parameter-draw analyses address uncertainty from fitting the model, but do not by themselves define the intended patient population or fix a faulty exposure calculation. In every approach, the target event and units must be audited.

The central pitfall is to treat a polished percentage as self-validating. A PTA of 90% is meaningful only with its target, population, regimen, time window, model assumptions, and uncertainty made clear. It is a population-level result under those specifications, not an individual prediction or a promise of efficacy. It may also be limited by model misspecification, uncertain targets, unrepresented covariates, or a simulated population that differs from the population in which the dose will be used.