PTA Across Covariate-Stratified Populations

Why Calculate PTA by Covariate Stratum?

A dose can have different chances of meeting the same clinical target in different patient groups. For example, renal function may be associated with drug clearance, so a given regimen may produce different exposure distributions across renal-function strata. A pooled probability of target attainment (PTA) can obscure such differences: a satisfactory average need not mean that each prespecified group has satisfactory attainment.

The question here is conditional: for a candidate dose or regimen, what proportion of patients in each defined covariate stratum is predicted to meet the target? The clinical target is held fixed across strata. What changes is the distribution of patient characteristics, individual random effects, and consequent outcomes represented within each stratum.

Definition: Let $S$ denote a prespecified covariate stratum, $d$ a candidate dose or regimen, and $Y(d)$ the model-predicted clinical index under that regimen for a patient drawn from the stratum-specific population. If $\mathcal{T}$ is the clinical target set, then the stratum-specific probability of target attainment is $$ \mathrm{PTA}_s(d)=P\{Y(d)\in\mathcal{T}\mid S=s\}. $$ For a minimum efficacy target $Y(d)\geq T$, this becomes $P\{Y(d)\geq T\mid S=s\}$. The target $T$ and its clinical interpretation are the same in every stratum; the conditional distribution of $Y(d)$ may differ.

What Is Conditional on the Stratum?

A nonlinear mixed-effects model represents outcomes using population parameters, covariates, and individual-level random effects. For stratum-specific PTA, the simulated patients should have covariate values and random effects drawn from distributions appropriate to that stratum. The model then predicts the clinical index under the candidate regimen, and the target indicator is evaluated for each simulated patient.

The stratum definitions should be specified before examining the dose comparisons. A renal-function stratum, for instance, needs explicit cut points and a clear rule for assigning patients at a boundary. The model must also support predictions in the covariate range represented by each group. A stratum-specific calculation is not made valid merely by labelling simulations with subgroup names; the covariate and random-effect distributions used for that group must actually reflect its intended population.

Parameter uncertainty and patient variability have different meanings. Patient variability describes the distribution of outcomes among patients under a fixed model. Parameter uncertainty describes uncertainty about the model parameters themselves. When uncertainty propagation is part of the estimand, parameter draws can be incorporated according to the established simulation procedure, but uncertainty must not be counted twice. A PTA calculation conditional on a fitted parameter vector is not, by itself, a full account of parameter uncertainty.

Working assumption: Within each stratum, the model and the distributions used to simulate covariates and individual effects adequately represent the patients to whom the estimate is intended to apply. The same regimen definition, endpoint, follow-up period, and clinical target must be used across the strata being compared.

From a Conditional Probability to a Simulation Estimate

The probability can be written as an average of an indicator over the stratum-specific distribution. Let $Z$ collect a patient's covariates and individual random effects, and let $Y(d;Z)$ be the model prediction under dose $d$. For the minimum-target case, define $I\{Y(d;Z)\geq T\}$ to equal 1 when the target is met and 0 otherwise. Then

\mathrm{PTA}_s(d) = E\!\left[I\{Y(d;Z)\geq T\}\mid S=s\right] = \int I\{Y(d;z)\geq T\}\,dF_s(z),

where $F_s$ is the joint distribution of simulated patient characteristics and random effects in stratum $s$. This expression explains why the covariate distribution matters: changing $F_s$ can change the probability even though the target and dose are unchanged.

For $M_s$ simulated patients in stratum $s$, let $I_{is}$ be the target indicator for patient $i$. The Monte Carlo estimate and its simulation standard error are

\widehat{\mathrm{PTA}}_s(d) = \frac{1}{M_s}\sum_{i=1}^{M_s} I_{is}, \qquad \widehat{SE}_{MC,s} = \sqrt{\frac{\widehat{\mathrm{PTA}}_s(d)\{1-\widehat{\mathrm{PTA}}_s(d)\}}{M_s}}.

If the simulated patients are independent draws from the specified stratum distribution, each indicator is Bernoulli with success probability $\mathrm{PTA}_s(d)$. Thus the sample proportion is unbiased for that conditional probability, and its variance is $\mathrm{PTA}_s(d)\{1-\mathrm{PTA}_s(d)\}/M_s$. Replacing the unknown probability by its estimate gives the displayed Monte Carlo standard error. This measures numerical simulation error, not uncertainty from model misspecification, parameter estimation, or limited clinical data.

Worked Example: Analytic PTA in Two Clearance Strata

Suppose, hypothetically, that a fixed regimen produces steady-state exposure $X$ that is lognormally distributed in each of two prespecified renal-function strata. The clinical target is the same in both: $X\geq 8$ exposure units. For lower clearance, assume $\ln X\sim N(\ln 10,0.30^2)$; for higher clearance, assume $\ln X\sim N(\ln 9,0.35^2)$. The log-scale standard deviations and medians are stipulated model outputs for this illustration.

For a lognormal variable with $\ln X\sim N(\mu,\sigma^2)$, $P(X\geq T)=P(\ln X\geq\ln T)=\Phi\{(\mu-\ln T)/\sigma\}$, where $\Phi$ is the standard normal cumulative distribution function. For the lower-clearance stratum:

z_{\mathrm{lower}} = \frac{\ln(10)-\ln(8)}{0.30} = \frac{\ln(1.25)}{0.30} = \frac{0.22314}{0.30} = 0.7438, \qquad \mathrm{PTA}_{\mathrm{lower}}=\Phi(0.7438)\approx0.7715.

The estimated analytic PTA is therefore about 77.15%. For the higher-clearance stratum:

z_{\mathrm{higher}} = \frac{\ln(9)-\ln(8)}{0.35} = \frac{\ln(1.125)}{0.35} = \frac{0.11778}{0.35} = 0.3365, \qquad \mathrm{PTA}_{\mathrm{higher}}=\Phi(0.3365)\approx0.6318.

This is about 63.18%. The lower-clearance stratum has an estimated PTA higher by $0.7715-0.6318=0.1397$, or 13.97 percentage points. This is a comparison of conditional model predictions under the stipulated distributions, not evidence that clearance alone causes the difference. The exposure distributions, including any other modeled covariate effects, determine the result.

Monte Carlo Comparison Within Strata

Analytic calculations are useful when the outcome distribution has a tractable form. In a typical NLME application, however, the predicted endpoint may depend on nonlinear concentration-time profiles, several covariates, and random effects. Simulating patients in each stratum and counting target attainment provides a direct estimate without requiring a closed-form distribution for the endpoint.

Worked Example: Conditional Simulation Estimates and Precision

Consider a hypothetical regimen and the same clinical target in three prespecified renal-function strata. The model generates 1,000 independent virtual patients in each stratum. Suppose target attainment is observed for 846 patients in the preserved-function stratum, 713 in the moderate-impairment stratum, and 482 in the severe-impairment stratum.

The respective conditional estimates are $846/1000=0.846$, $713/1000=0.713$, and $482/1000=0.482$. Their approximate Monte Carlo standard errors are:

\widehat{SE}_{MC,\mathrm{preserved}} = \sqrt{\frac{0.846(1-0.846)}{1000}} = \sqrt{0.0001303} \approx0.01141,

\widehat{SE}_{MC,\mathrm{moderate}} = \sqrt{\frac{0.713(1-0.713)}{1000}} = \sqrt{0.000204631} \approx0.014305,

\widehat{SE}_{MC,\mathrm{severe}} = \sqrt{\frac{0.482(1-0.482)}{1000}} = \sqrt{0.0002497} \approx0.01580.

The simulated probabilities are 84.6%, 71.3%, and 48.2%. The reported standard errors correspond to approximately 1.14, 1.43, and 1.58 percentage points of Monte Carlo error, respectively. They indicate the precision of these simulation estimates under the assumed model and simulation design; they are not clinical confidence intervals and do not quantify uncertainty about whether the model is correct.

A large difference, such as 84.6% versus 48.2%, is unlikely to be explained by this amount of Monte Carlo noise alone. It still requires evaluation of the model and the stratum-specific inputs. If these simulation estimates are too imprecise for a decision near a threshold, increase the number of simulated patients and check that estimates are stable across independent simulation runs.

Using Stratum-Specific PTA for a Dose Decision

A dose decision should make the acceptance rule explicit. If a decision threshold is $\tau$, a regimen satisfies that rule in stratum $s$ when its PTA meets the prespecified criterion, commonly $\mathrm{PTA}_s(d)\geq\tau$. The same target and threshold can be applied across strata when that is the clinical decision being asked. The resulting comparison reveals whether the regimen meets the rule in each group; it does not automatically determine the best overall dose.

Worked Example: Applying a Common PTA Threshold

Suppose, hypothetically, 500 virtual patients are simulated per stratum for each of two regimens. The target is identical for both strata and regimens, and the decision rule requires PTA of at least 90% in each of two prespecified renal-function strata. For regimen A, 455 of 500 patients attain the target in stratum 1, and 390 of 500 do so in stratum 2. For regimen B, the corresponding counts are 470 and 440.

For regimen A, the stratum-specific estimates are $455/500=0.910$ and $390/500=0.780$. It meets the 0.90 rule in stratum 1 but not stratum 2. For regimen B, the estimates are $470/500=0.940$ and $440/500=0.880$. It meets the rule in stratum 1 but falls short in stratum 2. Therefore, neither regimen meets the stated criterion in both strata. Regimen B has higher estimated PTA in each stratum, but that fact alone does not establish that it is an acceptable dose: the second-stratum estimate remains below the specified threshold.

This conclusion is conditional on the target, threshold, model, and stratum definitions in the example. It does not imply that every patient in a group will or will not meet the target, and it does not guarantee a clinical outcome for an individual.

Diagnostics, Alternatives, and Limitations

Before interpreting a stratum comparison, check that the strata are prespecified, the assignment rules are unambiguous, and each simulated group has adequate representation in the data supporting the model. Examine whether simulated covariate values and individual effects are plausible within each stratum. Predictions at covariate extremes or outside the range represented in model development may be extrapolations and can be unreliable. For nonlinear models, compare simulated outcome distributions—not just their averages—with relevant observed data where possible.

Monte Carlo diagnostics are distinct from model diagnostics. A small simulation standard error indicates that the estimated probability is numerically stable for the chosen model and distributions; it cannot repair biased parameter estimates, an incorrect structural model, or poor representation of a stratum. Conversely, an adequate model may still produce a noisy PTA estimate if too few virtual patients were simulated. Increasing simulation size addresses the latter problem, not the former.

Stratified PTA is often more informative than a pooled PTA when subgroup performance matters. A pooled probability answers a different question: it averages across a population distribution and can conceal low attainment in a particular stratum. Another alternative is to model covariates continuously and report predictions over a range of values. That can avoid arbitrary cut points and reveal gradients, but it requires careful interpretation and does not replace a requested probability for a prespecified clinical group. Neither approach removes the need for credible covariate effects and a clearly defined estimand.

Stratum-specific estimates can also be unstable when a group is small or poorly represented in the source data. Very narrow strata may have little support, while broad strata can conceal clinically meaningful variation within the group. Report the group definition, simulated sample size, target, candidate regimen, and model assumptions alongside each PTA. If uncertainty in fitted parameters is relevant to the decision, report it separately or propagate it through the appropriate simulation procedure rather than presenting Monte Carlo precision as total uncertainty.

Common Mistakes

Changing the clinical target across groups without saying so. Then the probabilities do not answer a like-for-like comparison. Use the same target when the question is how one regimen performs across prespecified strata.
Using one pooled covariate distribution for every stratum. This removes the conditional distinction the analysis is intended to describe. Simulate each group using its appropriate covariate and individual-effect distribution.
Calling Monte Carlo error total uncertainty. The simulation standard error describes finite simulation variability under the model. It does not include model misspecification or automatically include parameter uncertainty.
Concluding that a dose is adequate because the pooled PTA is high. A pooled result can hide inadequate attainment in one group. Apply the stated decision criterion to each relevant stratum.
Interpreting PTA as an individual guarantee. A stratum PTA is a model-based population probability, not a certainty about any one patient.

Key takeaway: Stratum-specific PTA is the probability of meeting one stated clinical target, conditional on membership in a prespecified covariate group. Its meaning depends on using a credible stratum-specific population distribution, evaluating numerical precision separately from model adequacy, and interpreting each group's estimate against the decision rule.

Check Your Understanding

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

A target is $X\geq T$. What does $\mathrm{PTA}_s(d)$ condition on, and what must remain the same when comparing two strata?
In a stratum, 720 of 900 simulated patients meet the target. Calculate the estimated PTA and approximate Monte Carlo standard error.
Why can a pooled PTA above a decision threshold fail to show that the threshold is met in every stratum?
A stratum's Monte Carlo standard error is very small. Name one important source of uncertainty it does not necessarily measure.

Solutions. (1) It conditions on membership in stratum $s$ and evaluates the regimen $d$; for a like-for-like comparison, the clinical endpoint and target should be the same across strata. (2) The estimate is $720/900=0.800$. Its Monte Carlo standard error is $\sqrt{0.800(0.200)/900}=\sqrt{0.0001778}\approx0.01333$, or about 1.33 percentage points. (3) The pooled value averages across groups and can be raised by high attainment in one group despite lower attainment in another; inspect each conditional estimate against the rule. (4) It does not necessarily measure parameter uncertainty or model misspecification; it measures finite simulation error under the specified model and simulation design.