Why a PTA Curve Deserves a Diagnostic Check
A probability-of-target-attainment (PTA) curve plots the proportion of a simulated population that meets a specified pharmacokinetic/pharmacodynamic (PK/PD) target against dose. Its shape can affect dose comparisons, but a plausible-looking curve is not proof that the simulation is correct. A unit mismatch, a misplaced target threshold, finite simulation noise, or an incorrectly coded regimen can all alter the curve.
A downward step may be a simulation fluctuation, a coding error, or a real consequence of the model and target definition. A flat curve may indicate that nearly everyone already passes or fails, but it can also result from dose not reaching the exposure calculation. The diagnostic task is to identify which explanation fits the simulation and its assumptions before interpreting the plotted pattern.
The target must be specified before classifying individuals as passing or failing. For example, a time-above-MIC target might require the fraction of a dosing interval with concentration above the MIC to be at least a stated proportion. An AUC/MIC target instead compares a ratio with a threshold. Each target has its own units and comparison; a threshold from one definition cannot be substituted into another.
When Should PTA Increase With Dose?
A useful diagnostic benchmark is monotonicity. Consider the same virtual population at two doses \(d_1<d_2\), with the same regimen structure, covariates, MIC values, parameter draws, target definition, and simulation settings. Suppose each individual’s efficacy metric \(E_i(d)\) is nondecreasing with dose, and that individual \(i\) passes when \(E_i(d)\geq T_i\), where the target \(T_i\) is fixed across the two doses. Then an individual who passes at \(d_1\) must also pass at \(d_2\).
Summing these indicators over individuals preserves the ordering, so the estimated PTA cannot decrease in this paired simulation. Taking population probabilities gives the same conclusion for population PTA under the stated conditions. This is a diagnostic result, not a universal law: it relies on a nondecreasing efficacy metric, an unchanged target and regimen structure, and a like-for-like population. Nonlinear PK does not by itself invalidate the argument; what matters is whether the specific exposure metric is nondecreasing with dose under the model.
A further distinction matters. A monotone population PTA can be estimated by separate finite simulations at each dose, and those estimates need not be monotone. If separate random samples or parameter draws are used, simulation noise can produce a small downward step. Conversely, changes in frequency, infusion duration, or another regimen feature mean that the plotted points may no longer represent dose changes alone.
Audit the Target and Exposure Calculation
Start with the individual-level pass/fail calculation, not the graph. For every target, write down the metric, its units, the threshold, and the exact comparison. Check that the exposure is calculated over the intended interval and that MIC or other target-defining values are in compatible units. Trace several simulated individuals from dose through exposure to the pass/fail indicator; a final percentage cannot reveal which part of this chain is wrong.
Worked Example: A Time-Above-MIC Unit Mismatch
Suppose a hypothetical regimen has a 12-hour dosing interval. The target is for concentration to exceed the MIC for at least 60% of that interval. For one virtual individual, the simulated time above MIC is 5.5 hours. Determine whether this individual passes, and show how a faulty comparison can misclassify the result.
Correct calculation. Convert the time to a fraction of the interval: \(5.5\text{ h}/12\text{ h}=0.4583\), or 45.83%. Since \(45.83\%<60\%\), this individual fails.
Faulty calculation. Comparing the number 5.5 directly with 0.60 treats hours as though they were a unitless fraction. It would give \(5.5>0.60\) and incorrectly label the individual as a pass. The comparison is both dimensionally and logically invalid, and here it changes the classification.
The implementation should calculate the fraction consistently, for example as time above MIC divided by the 12-hour interval, and compare that fraction with 0.60. Alternatively, it may compare time above MIC with \(0.60\times12=7.2\) hours. Both correct forms give a failure at 5.5 hours.
Other useful checks include verifying whether the simulation uses the intended MIC value and whether the target applies to the interval, day, or another time window. Confirm that a ratio such as AUC/MIC divides quantities expressed on compatible scales. Inspect the unrounded exposure and target values for individuals close to the threshold: rounding before classification can move individuals across the pass/fail boundary.
Separate Monte Carlo Variation From a Real Change
At a fixed dose, each simulated pass/fail outcome is a Bernoulli observation with pass probability \(p\). If \(N\) individuals are simulated independently from the intended population, the estimated PTA has approximate standard error \(\sqrt{\hat p(1-\hat p)/N}\). This standard error describes Monte Carlo precision for the simulated population and model; it does not include uncertainty in the model, parameter estimates, or target choice.
For independently simulated estimates at two doses, an approximate standard error for their difference is the square root of the sum of the two estimated variances. This calculation helps judge whether a small observed change is readily explained by simulation variation. If simulations use common random numbers—for example, the same virtual individuals and parameter draws at every dose—the estimates are paired, and the variance of the difference should account for that pairing rather than treating the estimates as independent.
Worked Example: Checking a Small Downward Step
At one hypothetical dose, two independent simulation runs each use \(N=1{,}000\) individuals. The first estimates PTA as 0.720 and the second as 0.750. Assess whether this 3.0-percentage-point difference is large relative to the approximate Monte Carlo uncertainty.
For the first run, \(SE_1=\sqrt{0.720(1-0.720)/1{,}000} =\sqrt{0.0002016}=0.01420\), or 1.420 percentage points.
For the second run, \(SE_2=\sqrt{0.750(1-0.750)/1{,}000} =\sqrt{0.0001875}=0.01369\), or 1.369 percentage points.
Because these runs are independent, the approximate standard error of their difference is \(\sqrt{0.01420^2+0.01369^2}=0.01972\), or 1.972 percentage points. The observed difference is \(0.750-0.720=0.030\), or 3.0 percentage points; it is about \(0.030/0.01972=1.52\) standard errors. That difference is not compelling evidence of a change in the underlying PTA. It is compatible with ordinary Monte Carlo variation, although it does not prove that the implementation is correct.
If a curve has a small irregular step, increase the simulation size or repeat the simulation and compare results. For a more precise dose-to-dose comparison, simulate the same virtual individuals at both doses and examine the paired pass/fail changes. Persistent or large irregularities warrant an implementation audit, not just a larger simulation.
Verify What Each Dose Label Means
A dose label is not enough to identify a regimen. Confirm the amount administered at each administration, the dosing interval, the number of administrations represented, and the duration over which the target is evaluated. Check the conversion from a dose table into the model input, including units and any weight-based calculation. A curve can be orderly yet wrong if every regimen is consistently coded at the wrong amount.
Worked Example: A Daily-Dose Versus Per-Administration Error
Consider a hypothetical linear-PK check with clearance \(CL=5\text{ L/h}\). The intended regimen is 500 mg every 12 hours, and the target is based on AUC over 24 hours. At periodic steady state under linear PK, assuming the doses are systemic doses, \(AUC_{24}=\text{total dose in 24 h}/CL\). Calculate the intended AUC, then assess a coding error that treats 500 mg as the total daily amount.
The intended regimen gives two administrations in 24 hours, so the daily amount is \(2\times500=1{,}000\) mg. Thus \(AUC_{24}=1{,}000\text{ mg}/(5\text{ L/h})=200\text{ mg·h/L}\). The units follow because \(\text{mg}/(\text{L/h})=\text{mg·h/L}\).
If the code instead treats 500 mg as the total 24-hour amount, it computes \(500/5=100\text{ mg·h/L}\), half the intended AUC. A PTA curve based on that input could appear unexpectedly low or shifted to higher labeled doses. The check depends on the stated linear-PK and 24-hour-AUC assumptions; it is an arithmetic audit, not a general formula for nonlinear PK.
Also verify that a regimen change has not silently altered the endpoint. Equal daily amounts administered at different intervals can produce different concentration-time profiles and therefore different time-above-MIC results, even when a linear-PK AUC calculation gives the same total AUC. The simulated regimen and the target calculation must use the same interval definition.
Read the Shape Without Assuming It Is an Error
| Observed pattern | Checks to prioritize | Possible non-error explanation |
|---|---|---|
| Small downward steps | Simulation size, independent versus paired draws, dose mapping | Finite Monte Carlo variation |
| Large or repeated reversals | Exposure calculations, target indicator, covariates, regimen code | Nonmonotone exposure metric or a changing regimen |
| Sudden jump | Dose units, rounding, threshold logic, formulation or schedule changes | A genuinely discrete regimen change or a concentrated distribution near the target |
| Nearly flat curve | Does dose enter the model? Are units and simulation inputs correct? | Most individuals already pass or fail; the dose range may have little effect |
A sudden transition is not automatically a coding error. If many individuals have exposures near the target boundary, a modest dose increase can move many indicators from 0 to 1. Discrete available formulations or dose rounding can also create steps. Likewise, a flat curve can be legitimate when the simulated doses lie entirely in a range where nearly everyone fails or nearly everyone passes. Report the estimated PTA and its simulation precision rather than inferring a smooth dose response from a small number of simulated dose points.
Useful diagnostics include plotting individual exposure metrics against dose for a sample of virtual individuals, tabulating pass counts and denominator at each dose, and comparing a few hand-checked individuals with the simulation output. Confirm that the same population and target are used at every point. If a curve changes only when the regimen code changes, investigate the transition explicitly. These checks can locate errors but cannot establish that the structural PK model or target is clinically appropriate; those require separate scientific evaluation.
Common Pitfall and Practical Sequence
Pitfall: “PTA should rise with dose, so every downward step proves the simulation is wrong.” This reasoning ignores finite simulation precision and the conditions required for monotonicity. A small decrease from independent finite runs may be noise; a regimen change or a nonmonotone individual exposure metric may also break the simple benchmark. The correction is to verify the benchmark’s assumptions, quantify Monte Carlo uncertainty, and then trace the calculation before changing the model or code.
Write the target metric, threshold, time window, population, and regimen in explicit units.
Check dose input, exposure output, target comparison, and pass/fail indicator for selected individuals, including values near the threshold.
Use the binomial standard error or paired comparisons as appropriate; increase simulation size if Monte Carlo uncertainty is material.
Verify dose labels, administration frequency, interval, rounding, and that the same target and population definition are applied across points.
No single curve diagnostic can validate every component of an NLME-based simulation. A smooth curve may still use the wrong units, while a jagged curve may simply be imprecise. The most informative conclusion combines the curve’s shape with individual-level checks, simulation uncertainty, and an explicit account of how each regimen was encoded.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A 24-hour interval has a target of at least 50% time above MIC. An individual is above MIC for 10 hours. Does the individual pass? What unit error could produce the wrong answer?
- State the assumptions that support a nondecreasing PTA estimate across two doses when the same virtual individuals are used.
- At a dose, \(\hat p=0.40\) from \(N=400\) independent virtual individuals. Calculate the approximate Monte Carlo standard error.
- Why can equal total daily doses on different schedules have different time-above-MIC attainment?
- Give two plausible explanations for a flat PTA curve and one check that helps distinguish a coding problem from a genuine flat response.
Solutions. (1) The fraction is \(10/24=0.4167\), or 41.67%, so the individual fails the 50% target. Comparing 10 hours directly with 0.50 would be dimensionally invalid and would incorrectly suggest a pass. (2) The individual exposure metric must be nondecreasing with dose; the target and regimen structure must be unchanged; and the same virtual individuals, parameters, covariates, and target values must be used. Then each individual’s pass indicator cannot decrease, so neither can the estimated PTA. (3) \(SE\approx\sqrt{0.40(0.60)/400}=\sqrt{0.0006}=0.02449\), or 2.449 percentage points. (4) Different schedules create different concentration-time profiles, which can change time above MIC even if total daily dose is equal. (5) The curve may be flat because nearly everyone passes or fails over the selected dose range, or because dose is not being transmitted to the exposure calculation. Inspecting dose and individual exposure outputs across dose levels helps distinguish these explanations.