Interpreting Differences Between PTA and Median Exposure

Why the Median Is Not a PTA

A simulated regimen can produce a median exposure close to a target and still have a low, moderate, or high probability of meeting that target. The reason is that a median locates the centre of a distribution, whereas probability of target attainment (PTA) depends on the fraction of the distribution on the target-attaining side of a threshold. How much exposure varies, whether the distribution is asymmetric, and where the threshold lies all matter.

This distinction is important when interpreting simulated results for dose justification. The median describes a typical exposure under the simulated conditions; PTA describes the proportion of the defined population whose exposure satisfies a criterion. Neither quantity can generally be inferred from the other. As in the earlier material on PTA curves, the target and the attainment rule must be specified before interpreting the result.

Definition: Let \(E\) denote an exposure measure for a randomly selected individual from a specified simulated population under a specified regimen. For a lower-bound target \(T\), \(PTA=P(E\geq T)\). For \(N\) equally weighted simulated individuals with exposures \(E_1,\ldots,E_N\), the empirical estimate is \(\widehat{PTA}=N^{-1}\sum_{i=1}^{N} I(E_i\geq T)\), where \(I(\cdot)\) equals 1 when its condition is true and 0 otherwise. The exposure measure, target, population, regimen, and use of \(\geq\) rather than \(>\) must be defined.

A median \(m\) is a value that divides a distribution so that at least half of its probability lies at or below \(m\), and at least half lies at or above \(m\). For a continuous distribution with no point mass at the median, exactly half the probability lies above it. Thus, if \(T=m\), then \(P(E\geq T)=0.5\) under that continuous-distribution condition. But if \(T\) is above or below the median, the median alone does not say how much probability lies beyond \(T\). For a discrete or rounded exposure distribution, ties can also make \(P(E\geq m)\) differ from 0.5.

Spread and the Target’s Location

A useful way to see the role of spread is to consider a lognormal exposure distribution. This is a possible model for positive, right-skewed exposures, not a universal assumption for simulated exposure. Suppose \(\ln(E)\sim N(\mu,\sigma^2)\), where \(\sigma\) is the standard deviation on the log scale. The median exposure is \(\exp(\mu)\). The probability of exceeding a positive target \(T\) is:

P(E\geq T) = P(\ln(E)\geq \ln(T)) = 1-\Phi\left(\frac{\ln(T)-\mu}{\sigma}\right)

Here, \(\Phi\) is the standard normal cumulative distribution function. The expression depends on both the target’s position relative to \(\mu\) and the log-scale spread \(\sigma\). This also shows why the median does not determine PTA: knowing the median determines \(\mu\), but not \(\sigma\).

Justification: Because the logarithm is increasing for positive values, \(E\geq T\) is equivalent to \(\ln(E)\geq\ln(T)\). Standardizing \(\ln(E)\sim N(\mu,\sigma^2)\) gives \(P(E\geq T)=P\{Z\geq[\ln(T)-\mu]/\sigma\}\), where \(Z\sim N(0,1)\). Taking the upper-tail probability gives the stated result.

When the target is above the median, increasing spread can increase the upper-tail probability: a broader distribution puts more probability far above the median, even as it also puts more below it. When the target is below the median, increasing spread can instead reduce the probability of exceeding the target, because more probability moves into the lower tail. This direction is a feature of the specified distributional comparison; it is not a general rule that greater variability always raises or always lowers PTA.

Worked Example: Same Median, Target Above It

Suppose two hypothetical simulated exposure distributions are lognormal and both have median exposure 100 mg·h/L. Their log-scale standard deviations are 0.20 and 0.50. For a target of 110 mg·h/L, compare the model-implied PTA. The lognormal assumption is stipulated for this example; in practice, the simulated distribution should be checked before relying on it.

Since the median is \(\exp(\mu)=100\), \(\mu=\ln(100)\). For the distribution with \(\sigma=0.20\):

z=\frac{\ln(110)-\ln(100)}{0.20} =\frac{\ln(1.10)}{0.20} =\frac{0.09531}{0.20} =0.4766

Therefore, \(PTA=1-\Phi(0.4766)\approx 1-0.6832=0.3168\), or 31.68%. For the distribution with \(\sigma=0.50\):

z=\frac{\ln(1.10)}{0.50} =\frac{0.09531}{0.50} =0.1906

Thus, \(PTA=1-\Phi(0.1906)\approx 1-0.5756=0.4244\), or 42.44%. Both distributions have the same median, but the broader one has the higher probability of exceeding this target above the median. This result follows from the stipulated lognormal model, not from the median alone.

Worked Example: Same Median, Target Below It

Keep the two hypothetical lognormal distributions from the previous calculation, with median 100 mg·h/L, but now set the target to 90 mg·h/L. The target is below both medians. For \(\sigma=0.20\):

z=\frac{\ln(90)-\ln(100)}{0.20} =\frac{\ln(0.90)}{0.20} =\frac{-0.10536}{0.20} =-0.5268

The probability of exceeding the target is \(1-\Phi(-0.5268)=\Phi(0.5268)\approx0.7009\), or 70.09%. For \(\sigma=0.50\), \(z=-0.10536/0.50=-0.2107\), giving \(PTA=1-\Phi(-0.2107)=\Phi(0.2107)\approx0.5835\), or 58.35%.

Here the broader distribution has lower PTA. More probability is spread into the lower tail, and the target lies below the common median. Together, the two examples show why the target’s location must be considered before interpreting a change in spread.

Skewness and Tail Weight

Exposure distributions are often asymmetric. In a right-skewed distribution, a relatively small fraction of individuals may have much higher exposure than the centre. The median can remain near the same value while the upper tail changes substantially. If the target is placed in that tail, even a modest change in the number of simulated individuals above the target can change PTA appreciably.

Skewness is not a substitute for examining the target-relevant tail. Two distributions can have similar medians and both appear right-skewed, yet have different proportions above a particular threshold. Conversely, a difference in a skewness summary does not by itself tell how many individuals meet the target. The direct quantity for a threshold PTA is the probability mass on the attaining side of that threshold.

Worked Example: Same Median, Different Upper-Tail Attainment

Consider two hypothetical empirical simulation distributions, each containing 1,001 equally weighted exposure values in mg·h/L. In both, 500 individuals have exposure 80 and one has exposure 100. In distribution A, 400 have exposure 108 and 100 have exposure 160. In distribution B, 100 have exposure 108 and 400 have exposure 160. In each distribution, the 501st ordered value is 100, so the median is 100 mg·h/L. Both have a right-hand tail, but the allocation of values in that tail differs.

For a target of 120 mg·h/L, only values of 160 meet the criterion. Distribution A has 100 attaining individuals:

\widehat{PTA}_A=\frac{100}{1001}=0.09990\approx 9.99\%

Distribution B has 400 attaining individuals:

\widehat{PTA}_B=\frac{400}{1001}=0.3996\approx 39.96\%

The medians are identical, but estimated PTA differs by approximately 30.0 percentage points because the upper-tail weight above 120 mg·h/L differs. These are deliberately simplified empirical distributions; they illustrate why the median and a general description such as “right-skewed” cannot replace counting or estimating the probability beyond the actual target.

Interpreting Results and Checking the Distribution

A sound interpretation pairs PTA with enough information to make the probability understandable. At minimum, identify the exposure metric and its units, the target and attainment rule, the regimen and simulated population, the median, and a description of spread and shape. When useful, report selected quantiles or a plot of the simulated distribution. A single spread measure may not describe a long or irregular tail well, so summaries near the target can be especially informative.

The empirical PTA is the fraction of simulated values that meet the criterion. It should be checked against the underlying exposure values or a plot, rather than inferred from the median, mean, or a spread summary. Also verify that the target and simulated exposure use the same units and definition, that the intended inequality was applied, and that the simulated individuals represent the population to which the interpretation refers. Rounding can create ties at the threshold, making the distinction between \(E\geq T\) and \(E>T\) consequential.

If a parametric distribution is used to calculate PTA, its assumptions need separate support. For the lognormal calculation above, exposure must be positive and the logarithms must be adequately represented by a normal distribution. A poor fit near the target can produce a misleading tail probability even when the fitted median looks plausible. Compare the fitted and simulated distributions around the threshold; do not judge fit only by agreement at the centre.

The empirical estimate and a parametric estimate are alternative ways to summarize the same target question. The empirical estimate counts simulated attainment directly and avoids choosing a distributional form, but it can be unstable in a tail when few simulated individuals lie near or beyond the target. A parametric estimate can provide a smooth probability and use the assumed distribution to describe a tail, but its validity depends on that assumption. Reporting the median alone is not an alternative PTA method: it answers a different question.

Simulation size also matters. With a finite set of simulated individuals, the empirical PTA is an estimate of the probability represented by the simulation, not an exact population truth. The tutorial on quantifying Monte Carlo error treats precision of that estimate. Here the key interpretive point is to separate uncertainty in the estimated PTA from the structural fact that a median does not specify the distribution’s shape or target-tail probability.

Common Mistakes and Limitations

Wrong reasoning: “The median equals the target, so PTA is high.” For a continuous distribution with no mass at the target, equality of median and target implies PTA of 50% under the \(E\geq T\) rule, not a high probability. With discrete or rounded values, ties can alter that result.
Wrong reasoning: “A higher spread always lowers PTA.” The direction depends on target location and distributional form. In the lognormal examples, greater spread raised PTA for a target above the median and lowered it for a target below the median.
Wrong reasoning: “The median and skewness tell me the PTA.” Neither generally gives the probability beyond a particular threshold. Inspect or estimate the fraction on the attaining side of that threshold.
Wrong reasoning: “A model-based PTA is reliable because the median fits.” Tail fit matters when the target is in a tail. Check fit near the target and state the assumptions behind any parametric calculation.

Finally, PTA is a model-based probability for a defined exposure criterion and simulated population. It is not, by itself, a probability of clinical response or safety, nor does a high PTA establish that the target is clinically appropriate. Those conclusions require the target’s clinical justification and the broader evidence supporting the model and regimen.

Key takeaway: The median describes the centre, while PTA depends on the probability beyond a specified threshold. Interpret their difference by locating the target relative to the distribution and examining spread, asymmetry, and tail weight at that target.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

A continuous exposure distribution has median 50 mg·h/L and no point mass at 50. What is \(P(E\geq50)\), and what assumption is needed for that answer?
A regimen’s median exposure is 100 mg·h/L. Explain why that fact alone does not determine PTA for a target of 115 mg·h/L.
For a right-skewed exposure distribution, what distribution feature should be examined when the target lies well above the median?
In a simulation of 2,000 equally weighted individuals, 530 meet the threshold. Calculate empirical PTA.
Why can a parametric PTA estimate be misleading even if its fitted median agrees with the simulated median?

Solutions. (1) \(P(E\geq50)=0.5\), provided the distribution is continuous and has no probability mass at the median. (2) The median gives the centre but not the spread or upper-tail probability; different distributions with median 100 can have different fractions above 115. (3) Examine the probability mass or fraction of simulated exposures above the target, and assess how well any fitted distribution represents that upper tail. (4) \(\widehat{PTA}=530/2000=0.265\), or 26.5%. (5) Agreement at the median does not establish an adequate fit near the target; a misspecified distribution can have an inaccurate tail probability and therefore an inaccurate PTA.