Computing PTA Within a Therapeutic Window

Why a Therapeutic Window Requires Two Limits

A dose can produce exposures that are high enough to meet an efficacy target yet also high enough to exceed a safety limit. Conversely, a low exposure may avoid the safety concern but fail the efficacy target. When the clinical objective is to meet both constraints, neither the efficacy threshold nor the safety threshold alone describes success. The relevant quantity is the fraction of simulated subjects whose exposure is between the two limits.

The exposure measure and its definition must be fixed before counting. For example, the measure might be steady-state AUC over a dosing interval, with units of mg·h/L. The efficacy and safety limits must refer to that same measure, time frame, and unit. This tutorial concerns whether a subject's chosen exposure summary falls within a window; it does not calculate time spent above or below a limit.

The estimate is the proportion of the simulated cohort inside the window. It estimates a model-based population probability under the simulation design and assumptions; it is not a guarantee about any particular patient. Its interpretation depends on which sources of variability were represented, how the virtual cohort was generated, and what exposure summary was used. Those choices should be consistent with the intended population and dosing question.

Count Subjects That Satisfy Both Conditions

The direct calculation is deliberately simple: apply both inequalities to each subject, count the subjects satisfying both, and divide by the number eligible for the analysis. The limits are often inclusive: exposure equal to \(L\) meets the efficacy condition, and exposure equal to \(U\) does not exceed the safety limit. If a study or decision rule specifies strict inequalities instead, use that rule consistently. With continuous simulated exposures, exact equality is usually rare, but boundary handling still matters when exposures are rounded or discretized.

A useful accounting check partitions the cohort into three mutually exclusive groups: below the window, within the window, and above the window. For \(L<U\), these events are \(E<L\), \(L\leq E\leq U\), and \(E>U\). Their counts must sum to \(N\), and their proportions must sum to 1. In the example, the counts are 3, 7, and 2, respectively, giving \(3/12+7/12+2/12=1\).

Window Attainment Is Not Either Marginal Attainment

Two single-limit summaries are sometimes reported alongside window PTA. Efficacy-only attainment is the proportion with \(E_i \geq L\). Safety-limit attainment is the proportion with \(E_i \leq U\). Neither of these proportions requires subjects to satisfy the other condition, so neither is the proportion in the therapeutic window.

In the 12-subject example, 9 subjects meet the efficacy lower limit and 10 meet the safety upper limit. Those proportions are \(9/12=75.0\%\) and \(10/12=83.3\%\), respectively, whereas only \(7/12=58.3\%\) meet both. The difference arises because the two marginal counts include subjects failing the other constraint.

For \(L<U\), the two failure groups do not overlap: a value cannot be both below \(L\) and above \(U\). Consequently, the window proportion can also be calculated by subtracting both failure proportions from 1: $$ \widehat{\mathrm{PTA}}_{\mathrm{window}} = 1-\frac{\#\{i:E_i<L\}}{N}-\frac{\#\{i:E_i>U\}}{N}. $$ This identity is an accounting result, not an assumption that the two constraints are statistically independent. It follows because the below-window, in-window, and above-window groups form a complete partition. It also provides a simple check on a direct indicator calculation.

Comparing Doses When the Window Creates a Trade-Off

A dose increase may move more subjects above the efficacy threshold while also pushing more subjects past the safety limit. As a result, window PTA need not increase with dose. This is a central difference from a single lower-threshold attainment calculation: once exposure can exceed an upper bound, raising exposure can improve one part of the criterion and worsen the other. Compute the fraction inside the window separately at every candidate dose rather than inferring it from an efficacy PTA curve alone.

Assumptions, Checks, and Common Failure Modes

The arithmetic is valid only if the simulated exposures and limits refer to a coherent target. Before counting, confirm that every included subject has one appropriately defined exposure summary at the dose under evaluation, that the exposure and both limits use the same units and time basis, and that \(L<U\). If exposures are missing or invalid, define the analysis population and denominator explicitly; silently dropping subjects can change the estimated proportion and make comparisons misleading. A missing exposure should not automatically be classified as either a success or a failure.

For an unweighted virtual cohort generated to represent the target population, each subject contributes one equally weighted indicator. If the simulation deliberately uses unequal sampling weights to represent a population, the corresponding weighted fraction is instead \(\sum_i w_i I\{L\leq E_i\leq U\}/\sum_i w_i\), where \(w_i\) is subject \(i\)'s nonnegative weight. Do not switch between weighted and unweighted calculations without stating the target and the weighting rationale.

Several diagnostics are especially useful. First, verify that the below-window, in-window, and above-window counts sum to the denominator. Second, inspect the exposure distribution and the three group counts: an unexpected zero count or an abrupt change across similar doses may indicate a coding, unit, or simulation problem. Third, check boundary handling on values exactly equal to \(L\) or \(U\). Fourth, ensure that the same definition of exposure is used across doses and that a subject is not counted more than once. Monte Carlo error also means a simulated proportion can vary across runs; the replicate-count methods covered in the earlier material on quantifying Monte Carlo error help assess whether the simulation is precise enough for the intended comparison.

A common mistake is to calculate the proportion with \(E\geq L\) and call it window PTA, reasoning that reaching efficacy is the main objective. The correction is to test both inequalities for every subject: efficacy attainment alone includes subjects with \(E>U\), who fail the safety constraint. The converse mistake—using only \(E\leq U\)—counts subjects below the efficacy limit. A third error is to multiply the two marginal probabilities as if the events were independent. They are conditions on the same exposure, so that multiplication is not the direct window calculation and has no general justification.

Direct simulation is usually the clearest approach when a population PK model has complex covariate or random-effect structure: it preserves the simulated exposure distribution and counts the target event directly. An analytic calculation from a fitted exposure distribution can be efficient when its distributional assumptions are appropriate and its tail probabilities are tractable, but a misspecified distribution may distort the fraction near either limit. Reporting a median exposure or selected percentiles can help describe the distribution, but these summaries do not generally determine the fraction between two limits. The direct count remains specific to the chosen population, model, dose, exposure summary, and target window; it is not a substitute for checking those choices or for evaluating distinct endpoints.