Deriving Probability of Target Attainment

From an Exposure Target to a Population Probability

A dose-selection question is often framed as: what fraction of the intended population is expected to meet a specified exposure target at this dose? That fraction is not the exposure of a typical individual, nor is it the probability that the population model is correct. It is a probability calculated under a specified population model for a specified dose, regimen, exposure metric, and target.

For example, a target may require an exposure measure to be at least a prespecified value. Individuals differ in their pharmacokinetic parameters, so the same dose can produce different exposure values. Probability of target attainment (PTA) summarizes the proportion of the modeled population whose individual exposure satisfies the target. The target and exposure metric must be defined before calculating this probability; the result is conditional on those choices.

Definition: Fix a candidate dose or regimen $D$, an exposure metric $M(D)$ for an individual, and a prespecified target set $\mathcal{T}$. The probability of target attainment is $$ \operatorname{PTA}(D)=P\{M(D)\in\mathcal{T}\}, $$ where the probability is taken over the modeled distribution of individuals in the target population. For a lower-bound target $M(D)\geq T$, this becomes $\operatorname{PTA}(D)=P\{M(D)\geq T\}$.

In an NLME model, individual parameters are generated from a population distribution, often conditional on covariates. Let $\theta$ denote the population-model parameters, $X$ an individual's covariates, and $b$ the individual's random effects. Write the exposure metric as $M(D;\theta,b,X)$. With $\theta$ held fixed, the PTA for a lower-bound target is

\operatorname{PTA}(D\mid\theta) = \int \int \mathbf{1}\{M(D;\theta,b,x)\geq T\} \,f_{b\mid X}(b\mid x,\theta)\,db\,dP_X(x).

Here, $\mathbf{1}\{\cdot\}$ equals 1 when its condition is true and 0 otherwise. The density $f_{b\mid X}$ describes individual random effects conditional on covariates, and $P_X$ is the covariate distribution in the population for which the dose is being considered. Writing the covariate average with respect to $dP_X(x)$ allows for continuous, discrete, or mixed covariates; for discrete groups, it is a sum weighted by their population proportions. If covariates are fixed by the estimand, the integral over $x$ is replaced by that specified covariate value or stratum. The integrals average the indicator of attainment across the modeled population.

Why the Integral Gives PTA

The derivation follows from conditioning on the variables that make individuals differ. Given covariates $X=x$ and random effects $b$, the exposure metric is determined by the model (or has a conditional distribution if the metric itself includes additional stochastic variation). The indicator records whether that individual meets the target. Averaging this indicator over random effects gives the conditional probability of attainment for covariate value $x$; averaging again over covariates gives the population probability.

\begin{aligned} \operatorname{PTA}(D\mid\theta) &=E\!\left[\mathbf{1}\{M(D;\theta,b,X)\geq T\}\mid\theta\right]\\ &=E_X\!\left[ E_{b\mid X}\!\left[ \mathbf{1}\{M(D;\theta,b,X)\geq T\}\mid X,\theta \right]\right]\\ &=\int P\{M(D;\theta,b,x)\geq T\mid X=x,\theta\}\,dP_X(x). \end{aligned}

The first equality uses the fact that the expectation of an indicator equals the probability of its event. The second and third lines apply the law of total expectation, first averaging over random effects conditional on covariates and then over the covariate distribution. If exposure is a deterministic function of $b$ and $x$, the conditional probability in the last line is itself the integral of the indicator over $f_{b\mid X}$, yielding the double integral above.

In practice, the population parameters are usually estimated, and the primary plug-in calculation uses $\hat{\theta}$ in place of $\theta$. This gives PTA conditional on the fitted population model. It is distinct from integrating over uncertainty in $\theta$: doing that requires an explicitly defined parameter-uncertainty distribution and a corresponding estimand. Population variability in $b$ is not a substitute for parameter uncertainty, and adding uncertainty should not count the same source twice.

Worked Example: AUC Target Under Lognormal Clearance

Suppose a hypothetical population receives a single 100 mg dose under linear pharmacokinetics. Assume individual clearance $CL$, in L/h, satisfies $\ln(CL)\sim N(\ln 5,[\ln(1.5)]^2)$. Thus the median clearance is 5 L/h and the geometric standard deviation is 1.5. For this model, individual AUC is $M(D)=D/CL$, in mg·h/L. Calculate the PTA for the lower-bound target $M(D)\geq 40/3$ mg·h/L.

Because clearance is positive, the target event can be rearranged without changing its direction:

\frac{100}{CL}\geq\frac{40}{3} \quad\Longleftrightarrow\quad CL\leq\frac{100}{40/3}=7.5\ \mathrm{L/h}.

Standardizing the normal variable $\ln(CL)$ gives

z = \frac{\ln(7.5)-\ln(5)}{\ln(1.5)} = \frac{\ln(1.5)}{\ln(1.5)} =1.

Therefore,

\operatorname{PTA}(100) = P\{CL\leq 7.5\} = \Phi(1) \approx 0.8413.

Under the stated hypothetical model, about 84.1% of individuals attain the AUC target at 100 mg. This is a population-model probability, not a probability that a particular patient's AUC is above target after observing that patient's data. The calculation assumes the lognormal clearance distribution and the stated linear AUC relationship are appropriate.

Closed-Form Calculations and Covariate Averaging

An analytic expression is available when the exposure metric has a tractable distribution. For example, if $\ln M(D)\sim N(\mu_D,\sigma_D^2)$ and the target is $M(D)\geq T$, then the logarithm is monotone increasing, so

\operatorname{PTA}(D) = P\{\ln M(D)\geq\ln T\} = 1-\Phi\!\left(\frac{\ln T-\mu_D}{\sigma_D}\right) = \Phi\!\left(\frac{\mu_D-\ln T}{\sigma_D}\right).

This result requires $T>0$, $\sigma_D>0$, and the stated normal distribution for $\ln M(D)$. It is not a general formula for every exposure metric: nonlinear parameter transformations, covariate mixtures, and more complex PK models may require integration by other means.

Worked Example: PTA From a Lognormal Exposure Distribution

For a hypothetical candidate regimen, suppose the individual exposure metric $M$, in mg·h/L, has $\ln M\sim N(\ln 8,0.4^2)$. The target is $M\geq 6$ mg·h/L. Calculate PTA.

The log-scale mean is $\ln 8$, and the log-scale standard deviation is 0.4. Standardizing the target gives

z = \frac{\ln(6)-\ln(8)}{0.4} = \frac{\ln(0.75)}{0.4} = \frac{-0.2876821\ldots}{0.4} = -0.7192052\ldots.

Since the event is an upper tail on the log scale,

\operatorname{PTA} = P\{M\geq6\} = P\{Z\geq-0.7192052\ldots\} = \Phi(0.7192052\ldots) \approx0.7640,

where $Z\sim N(0,1)$. The modeled probability of meeting the target is about 76.4%. The result depends on the assumed lognormal distribution; it is not justified merely because the metric is positive.

When the target population contains covariate groups with different exposure distributions, overall PTA is a weighted average of group-specific PTAs. The weights must represent the intended population, not the relative number of individuals in an accidental or selectively sampled dataset. This follows directly from conditioning on group membership.

Worked Example: Averaging PTA Across Two Covariate Groups

Suppose a hypothetical target population is 60% Group A and 40% Group B. At a candidate regimen, assume $\ln M$ has standard deviation 0.3 in both groups. The group-specific medians of exposure are 12 and 8 mg·h/L, respectively. Find overall PTA for $M\geq10$ mg·h/L.

For Group A, the log-scale mean is $\ln(12)$. Its PTA is

\operatorname{PTA}_A = \Phi\!\left(\frac{\ln(12)-\ln(10)}{0.3}\right) = \Phi(0.6077385\ldots) \approx0.7283.

For Group B,

\operatorname{PTA}_B = \Phi\!\left(\frac{\ln(8)-\ln(10)}{0.3}\right) = \Phi(-0.7438\ldots) \approx0.2285.

Weighting by the stated population proportions gives

\operatorname{PTA} = 0.60(0.72832\ldots)+0.40(0.2285\ldots) = 0.52839\ldots \approx0.5284.

Thus about 52.8% of this modeled population meets the target. Taking an unweighted average of the two group PTAs would give a different result and would represent a 50:50 population, not the one specified here.

Assumptions, Diagnostics, and Alternatives

A PTA calculation is only as meaningful as its estimand and model. Before interpreting the result, check that the candidate regimen and target refer to the same dosing conditions and time period, and that the exposure metric is defined consistently. Verify the random-effects distribution, covariance structure, covariate relationships, and covariate distribution used for the target population. The model's parameterization must also produce physically meaningful individual parameters and exposure values over the range being integrated.

Useful checks include comparing simulated or integrated exposure distributions with observed data and model-based predictions, inspecting whether the relevant parameter tails are supported by data, and confirming that covariate weights match the intended population. For a derived metric, check that the metric is computed from the correct individual profile and regimen. If a nonattaining subpopulation or implausible exposure tail dominates the result, investigate model fit and assumptions rather than treating the single PTA value as self-validating. Numerical integration or simulation can approximate the population integral, but approximation does not repair a misspecified population model.

Be explicit about which random quantities are included. For a latent PK-derived exposure measure, residual assay or observation error is generally not part of the individual exposure distribution unless the estimand is specifically defined in terms of a noisy measured quantity. Likewise, random effects represent population heterogeneity conditional on the model; uncertainty in estimated fixed effects or variance parameters is a separate layer. If that additional uncertainty is part of the question, it must be introduced deliberately.

PTA is different from several nearby summaries. The exposure of a typical individual (for example, setting random effects to zero) does not average over the population and cannot generally stand in for PTA. A percentile of the exposure distribution describes a point in that distribution, whereas PTA is the probability mass on the target side of a threshold. Direct integration or simulation targets the population probability; a closed-form calculation is more transparent when its distributional assumptions hold, while numerical methods are more flexible for complicated NLME models. Each approach still requires the same clear target, population, and stochastic model.

Common pitfall: A learner may compute exposure for the typical parameter values, see that it exceeds the target, and conclude that PTA is high. That reasoning ignores the distribution of individual parameters. The typical value does not reveal how much population probability lies above or below the target. Calculate or estimate the probability over the full modeled population distribution instead.

Check Your Understanding

Answer each question using the definition of PTA and the assumptions stated.

If $M$ is lognormal with $\ln M\sim N(\ln 10,0.5^2)$, what is PTA for $M\geq10$?
Why must group-specific PTAs be weighted by the target population proportions?
Does a PTA calculated using fitted population parameters automatically include uncertainty in those parameter estimates?
If the target is $M\leq T$, what changes in the probability definition?
Why is exposure at zero random effect not generally equal to population PTA?

Solutions. (1) The target is the median, so $z=(\ln10-\ln10)/0.5=0$, and PTA $=\Phi(0)=0.5000$. (2) The weighted average must represent the covariate composition of the population in the estimand; different group proportions define a different overall probability. (3) No. Plugging in fitted population parameters conditions on them; integrating parameter uncertainty is a separate step. (4) Use $\operatorname{PTA}(D)=P\{M(D)\leq T\}$, integrating the indicator of the lower-tail event. (5) Zero random effect describes one point in the random-effects distribution, while PTA averages attainment indicators over that distribution and, as appropriate, over covariates.

Deriving Probability of Target Attainment

What You'll Learn