Why Dose Selection Must Consider the Distribution
A dose that produces a suitable average exposure in children may still leave a meaningful fraction below a therapeutic target or above a safety threshold. This is especially important when body size, maturation, and other covariates vary within an age or weight band. Dose selection therefore requires examining the distribution of predicted exposures, not just its mean or median.
The earlier material on interpreting pediatric exposure targets establishes the role of a prespecified target range. Here, the task is to evaluate candidate doses within pediatric age or weight bands against that range. The target and the acceptable risks should be defined before comparing doses; there is no universal acceptable percentage for every drug or decision.
For an age or weight band, simulate children whose characteristics represent the intended population in that band. For each candidate dose, predict exposure using the population PK model and the relevant variability. In schematic form, a simulated exposure is \(E_i=f(d,x_i,\theta,\eta_i)\), where \(d\) is the dose, \(x_i\) contains the child’s characteristics, \(\theta\) denotes model parameters, and \(\eta_i\) represents individual random effects. Any residual error should be included only if it belongs in the exposure quantity being predicted; measurement noise is not automatically the same as between-child variability in true exposure.
The simulation is only as appropriate as its assumptions. The PK model must be adequate for the ages, weights, doses, and regimen being considered; the simulated covariates must represent the intended children; and the target must be relevant to the population and exposure measure. If dose proportionality is assumed, that assumption must be supported over the dose range. Nonlinear PK, maturation, or correlations among covariates and random effects can make simple scaling or independent sampling misleading.
Calculating Risk From an Exposure Distribution
A useful calculation can be made when exposure is lognormally distributed. This is a modeling assumption, not a general property of pediatric exposure. Suppose \(E=G\exp(\sigma Z)\), where \(Z\sim N(0,1)\), \(G\) is the geometric mean exposure, and \(\sigma\) is the standard deviation of log exposure. Taking logarithms gives \(\ln(E/G)=\sigma Z\). Thus the target boundaries can be put on the standard normal scale:
Here \(\Phi\) is the standard normal cumulative distribution function. Since the three exposure categories partition all possible exposures, target attainment is one minus the two tail risks, or equivalently the probability between the standardized boundaries:
Worked Example: Risks Under a Lognormal Exposure Model
Suppose, hypothetically, that AUC at a candidate dose is lognormal with geometric mean \(G=49\) mg·h/L and log-scale standard deviation \(\sigma=0.25\). The prespecified target range is 40 to 80 mg·h/L. Calculate the three probabilities.
Check the model and inputs. The assumed distribution is positive and lognormal, and the geometric mean and log-scale standard deviation are specified for this dose and population. The AUC and target boundaries all use mg·h/L. The calculation describes model-predicted exposure variability; it does not by itself account for uncertainty in the model parameters.
For the lower boundary, the standard normal value is \(\ln(40/49)/0.25\approx-0.812\). The lower-tail probability is \(\Phi(-0.812)\approx0.2085\), or 20.85% underexposure risk.
For the upper boundary, \(\ln(80/49)/0.25\approx1.961\). Therefore the overexposure risk is \(1-\Phi(1.961)\approx1-0.9750=0.0250\), or about 2.50%.
Target attainment is \(0.9750-0.2085=0.7665\), or about 76.65%. The probabilities sum to 1, subject to rounding: \(0.2085+0.7665+0.0250=1.0000\). This candidate dose therefore leaves about one child in five below the lower target boundary, even though the overexposure risk is small.
For simulations from a population PK model, the probabilities are often estimated by counting simulated exposures in each category. With \(N\) simulated children and \(n_{\mathrm{low}}\), \(n_{\mathrm{target}}\), and \(n_{\mathrm{high}}\) exposures in the three categories, the estimates are the corresponding counts divided by \(N\). The categories must be mutually exclusive and exhaustive: \(n_{\mathrm{low}}+n_{\mathrm{target}}+n_{\mathrm{high}}=N\). Use enough simulations to make Monte Carlo variation small for the decision, and report the number simulated as well as the estimated risks.
Worked Example: Comparing Candidate Doses With Prespecified Risk Limits
Consider a hypothetical pediatric weight band. A population PK model is used to simulate \(N=2{,}000\) children at each of three candidate doses. The target range and model are held fixed across candidates. Suppose the prespecified selection criteria are an underexposure risk no greater than 10% and an overexposure risk no greater than 5%.
| Candidate dose | Below target | Within target | Above target |
|---|---|---|---|
| 10 mg | 280 | 1,660 | 60 |
| 11 mg | 160 | 1,760 | 80 |
| 12 mg | 100 | 1,700 | 200 |
At 11 mg, the estimated underexposure risk is \(160/2{,}000=0.080\), or 8.0%; target attainment is \(1{,}760/2{,}000=0.880\), or 88.0%; and overexposure risk is \(80/2{,}000=0.040\), or 4.0%. The counts add to 2,000.
At 10 mg, the estimated underexposure risk is \(280/2{,}000=14.0\%\), which exceeds its 10% limit, although the overexposure risk is \(60/2{,}000=3.0\%\). At 12 mg, underexposure is \(100/2{,}000=5.0\%\), but overexposure is \(200/2{,}000=10.0\%\), above its 5% limit. Of these candidates, only 11 mg meets both stated limits. This is a conditional decision based on the hypothetical model, its simulated population, and the chosen limits—not evidence that 11 mg is suitable for children outside the band or under different assumptions.
Simulation results also vary by chance. For an estimated risk \(\hat p\) from \(N\) independent simulated children, its Monte Carlo standard error is approximately \(\sqrt{\hat p(1-\hat p)/N}\). At 11 mg, the underexposure risk estimate has standard error \(\sqrt{0.08(0.92)/2{,}000}\approx0.00607\), or 0.607 percentage points. This describes simulation noise under the specified model; it is not a measure of uncertainty about whether the model itself is correct.
Checking the Whole Band, Not Just Its Center
A band-level summary can conceal an unacceptable subgroup. If exposure changes systematically with weight or age, simulating only a typical child may miss high underexposure near one end and high overexposure near the other. Evaluate the distribution across the band, including its boundaries and relevant interior values. If the simulated population is sampled from the whole band, report its composition and also inspect clinically meaningful subgroups when a pooled result could mask different risks.
Worked Example: Detecting Risk Near a Weight-Band Boundary
Suppose a hypothetical fixed dose of 8 mg is being considered for children weighing 10 to less than 20 kg. At each of two weights, the model simulates 1,000 children with the remaining covariates and individual variability represented. The same target interval and risk criteria apply throughout the band.
At 10 kg, the simulations give 180 exposures below target, 800 within target, and 20 above target. The estimated risks are \(180/1{,}000=18\%\) underexposure and \(20/1{,}000=2\%\) overexposure; target attainment is 80%.
Near the upper boundary, at 19.9 kg, the counts are 40 below target, 870 within target, and 90 above target. The corresponding estimates are 4% underexposure, 87% target attainment, and 9% overexposure. Each set of counts sums to 1,000.
If the prespecified maximum risks were 10% for both underexposure and overexposure, the dose would fail at both examined weights, but for different reasons: underexposure at 10 kg and overexposure near 20 kg. A pooled band average could make either problem look smaller. The appropriate next step is to compare revised candidate doses or a different banding scheme using the same model and risk criteria, then re-evaluate across the band. The results do not establish that a particular split or dose will work without that additional simulation.
Assumptions, Diagnostics, and Alternatives
Before interpreting simulated risks, check that the analysis answers the intended decision. The exposure endpoint, target limits, dose regimen, age or weight boundaries, and risk criteria should be explicit. Review whether the population PK model captures relevant size and maturation effects, whether the covariate distribution reflects the intended children, and whether the dose range is supported by the model. Compare simulated predictions with observed data where available, and examine prediction behavior by age and weight rather than relying only on an overall goodness-of-fit summary.
Warning signs include abrupt or implausible changes in predicted exposure across adjacent ages or weights, a high-risk tail driven by a few extreme simulated covariates, simulated values outside the model’s support, and risk estimates that change substantially with the random seed or number of simulations. Investigate these patterns rather than treating a large simulation as a remedy: more simulations reduce Monte Carlo error but do not repair model misspecification or unrepresentative covariates. Sensitivity analyses can vary uncertain model parameters, random-effect assumptions, or the target-population distribution to show how robust the dose choice is.
The main alternative to a fixed dose within a band is a dose adjusted continuously by body weight or another covariate. Continuous adjustment can reduce within-band exposure variation, but it may be impractical or produce doses that cannot be measured or administered accurately. A fixed band dose is simpler but can create sharp changes at a cutoff and may leave greater variability within each band. A mg/kg rule is another alternative, but it should not be presumed to equalize exposure: clearance may scale nonlinearly with size, maturation may matter independently of weight, and PK may be nonlinear. Compare implementable regimens in the model, including rounding or dose caps when relevant, rather than selecting a theoretical dose that cannot be delivered.
The method has important limits. Risk estimates depend on the exposure target, population, model structure, parameter values, and variability included in the simulations. They do not alone establish clinical benefit or safety, and a target-attainment percentage is not proof of an outcome. Parameter uncertainty and structural uncertainty require separate consideration from ordinary simulation variability. Finally, a band that performs adequately at selected simulated ages or weights may still perform poorly elsewhere; the chosen grid must be sufficiently informative for the decision.
Define the exposure endpoint, target interval, pediatric band, candidate regimens, and acceptable underexposure and overexposure risks.
Use the population PK model and covariates appropriate to the band, including relevant individual variability and any justified parameter uncertainty.
Count exposures below, within, and above target; check the counts, Monte Carlo precision, and variation across age or weight.
Select only among candidates meeting the prespecified criteria under justified assumptions, and examine sensitivity to important uncertainties.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A simulation produces 50 exposures below target, 420 within target, and 30 above target from 500 children. What are the three estimated probabilities?
- For a lognormal exposure model, explain why the standardized lower boundary is \(\ln(L/G)/\sigma\), rather than \((L-G)/\sigma\).
- A dose meets the underexposure limit but exceeds the prespecified overexposure limit. Is it acceptable under those criteria? Explain.
- Why can a pooled risk estimate for an age or weight band be inadequate even when its total number of simulations is large?
Solutions. (1) Underexposure is \(50/500=10\%\), target attainment is \(420/500=84\%\), and overexposure is \(30/500=6\%\); the counts sum to 500 and the probabilities sum to 100%. (2) Under the assumed model \(E=G\exp(\sigma Z)\), taking logarithms gives \(\ln(E/G)=\sigma Z\). Setting \(E=L\) therefore gives \(Z=\ln(L/G)/\sigma\). Subtracting \(G\) on the original scale does not transform a lognormal boundary to the normal scale. (3) No. Both criteria must be met if the decision rule requires risks below their respective limits; meeting one does not compensate for exceeding the other unless a different trade-off was specified in advance. (4) A pooled estimate can hide distinct risks at different ages or weights. A large total simulation reduces Monte Carlo error for the pooled estimate, but does not ensure that the pooled composition represents every subgroup or reveal risk peaks near a band boundary.