Why Diagnose the Distribution Before Using It?
Simulated exposure distributions are inputs to dose-justification decisions, including probability of target attainment (PTA). A simulation can produce a PTA estimate without errors or warnings and still give an implausible exposure distribution because a parameter was sampled on the wrong scale, a covariate was misapplied, or an exposure calculation used inconsistent units. Inspecting the distribution is therefore a model-checking step: it asks whether the simulated values are consistent with the intended model and calculation, not merely whether the code ran.
The goal is not to make every distribution symmetric or narrow. A real population model can imply substantial right skew, long tails, or clusters reflecting distinct covariate groups. Instead, ask whether the observed range, shape, and boundary values follow from the stated model, the simulated population, and the exposure metric. An unusual feature is a prompt to trace the simulation, not a reason by itself to delete observations or alter the model.
Define the Exposure and Its Expected Support
Let \(E_i\) denote the exposure metric calculated for simulated individual \(i\), such as AUC over a dosing interval at steady state, \(AUC_{\tau,ss}\), or a minimum concentration. Before inspecting a histogram, state exactly which metric was calculated, its units, the dose and regimen, and the population and model assumptions used. A distribution cannot be judged sensibly if its variable or units are unclear.
For example, under linear pharmacokinetics for an intravenously administered dose \(D\) over a dosing interval \(\tau\) at steady state, \(AUC_{\tau,ss}=D/CL\), where \(CL\) is clearance. If \(D>0\) and \(CL>0\), this AUC must be positive. A negative value is not an unusual tail; it contradicts the calculation’s assumptions or implementation. Other metrics have different supports: a fraction of time above a target must lie between zero and one, whereas a positive AUC is not bounded above by one.
Population parameters and individual random effects jointly determine the distribution. For instance, a common clearance model is \(CL_i=CL_{\mathrm{typ}}\exp(\eta_{CL,i})\), where \(CL_{\mathrm{typ}}>0\) is typical clearance and \(\eta_{CL,i}\) is a normally distributed random effect with mean zero. This construction keeps clearance positive. The distribution of a derived exposure can consequently be skewed even when the random effect itself is symmetric.
Shape, Skew, and a Useful Distributional Check
When \(AUC_{\tau,ss}=D/CL\) and \(CL=CL_{\mathrm{typ}}\exp(\eta_{CL})\), then \(AUC_{\tau,ss}=(D/CL_{\mathrm{typ}})\exp(-\eta_{CL})\). If \(\eta_{CL}\sim N(0,\omega^2)\), the log AUC is normally distributed, so AUC is lognormal under these assumptions. Its right skew is expected, not evidence of a faulty simulation.
This conclusion depends on the stated simplified model: a fixed positive dose, linear pharmacokinetics, positive clearance, and a normal random effect on log clearance. Covariates, parameter uncertainty, correlated random effects, or a different exposure calculation can change the resulting distribution. The check is a reference for tracing the simulated values, not a requirement that every exposure distribution be lognormal.
A short derivation clarifies the expected skew. If \(Y=\exp(Z)\) and \(Z\sim N(m,v)\), then \(E[Y]=\exp(m+v/2)\) and \(\operatorname{Var}(Y)=(\exp(v)-1)\exp(2m+v)\). The coefficient of variation is \(\sqrt{\exp(v)-1}\), and the skewness is \((\exp(v)+2)\sqrt{\exp(v)-1}\). Thus, for \(v>0\), the lognormal distribution has positive skew. In the AUC example, \(m=\log(D/CL_{\mathrm{typ}})\) and \(v=\omega^2\). Increasing random-effect variability increases the right tail, even though the typical value, \(D/CL_{\mathrm{typ}}\), stays fixed.
Useful first-pass summaries include the minimum, maximum, median, selected quantiles, and a histogram or empirical cumulative distribution. Compare quantiles and extremes with the model’s expected behavior and the intended clinical population. A plot can reveal a cluster, a pile-up at a boundary, or a long tail that a mean and standard deviation alone may obscure. These checks describe the simulated output; they do not establish that the population model is clinically correct.
Trace an Unusual Value Back to Its Inputs
For each suspicious exposure, work backward from the final metric. Check the exposure formula and units first, then the individual’s parameter values and covariates, and finally the random-effect draws and their transformations. Compare the observed values with the distribution specified in the simulation code. If random effects are correlated, verify that the simulated joint distribution uses the intended covariance structure; checking each marginal distribution alone will not detect every joint-sampling error.
Write down the exposure definition, units, regimen, and conditions under which it was calculated. Identify impossible values before interpreting tails.
Use summaries and plots to find negative values, zeros, extreme quantiles, isolated outliers, or multiple clusters. Check whether the feature appears in the raw parameter draws or only after calculating exposure.
Inspect the covariates, parameter values, random effects, and units for individuals in the tail or at a boundary. Recalculate their exposure independently from the stated formula.
Check transformation, variability scale, covariate coding, dependence among random effects, and any rounding or boundary-handling operation. Correct a demonstrated cause, then rerun and recheck.
Diagnostics should be reproducible. Preserve the random seed or the suspect simulated records, and report the values used in a hand-check. If an extreme value follows from a plausible covariate or random-effect draw, it may belong in the distribution. If it arises from an incorrect unit conversion or an invalid parameter value, changing the simulated population is not a defensible remedy; correct the implementation and rerun it.
Worked Example: An Extreme AUC and an Incorrect Random-Effect Scale
Consider a hypothetical intravenous regimen with \(D=600\) mg per interval and \(CL_{\mathrm{typ}}=6\) L/h. Under the assumed linear steady-state calculation, \(AUC_{\tau,ss}=D/CL\), in mg·h/L. A simulated record has \(CL=0.30\) L/h, giving an AUC far above the typical value. Determine what to check.
Calculate the exposure. The substitution gives \(AUC_{\tau,ss}=600/0.30=2000\) mg·h/L. The typical-clearance value is \(600/6=100\) mg·h/L. Both calculations use mg and L/h, so the AUC units are mg·h/L; the extreme value is not explained by a units mismatch in this formula.
Trace the parameter. Under \(CL=6\exp(\eta_{CL})\), the random effect implied by this clearance is \(\eta_{CL}=\log(0.30/6)=\log(0.05)\approx-2.996\). Suppose the intended model specifies \(\eta_{CL}\sim N(0,0.30^2)\). The implied draw is about \(2.996/0.30=9.99\) standard deviations below zero. Such a value is extraordinarily inconsistent with that specified distribution, so first inspect how the random effect was generated and whether this record’s parameter was overwritten or transformed incorrectly.
Check the implementation. Suppose investigation shows that the code used a standard deviation of 1.2 rather than 0.30. Then a value near \(-3.0\) is only \(2.50\) standard deviations below the mean, so occasional very low clearances and very high AUCs become much more likely. The key finding is not simply that AUC equals 2000; it is that the simulated random-effect scale does not match the intended model. Correct the scale, regenerate the population, and verify the resulting random-effect distribution before using the exposures.
Boundary Values and Mixtures Need Context
A boundary value can be either valid or a warning. An exact zero AUC is incompatible with a positive administered dose and finite positive clearance under \(AUC_{\tau,ss}=D/CL\). It may indicate a failed calculation, missing values converted to zero, rounding, or an exposure definition that genuinely permits zero. By contrast, a computed fraction of time above a threshold can legitimately equal zero or one. Always interpret a boundary relative to the metric’s definition.
Skew and multiple clusters also require different questions. A long upper tail might follow from a log-scale random effect; two clusters might reflect two covariate groups or a specified mixture population. A cluster that appears only after a transformation or calculation may instead signal an implementation issue. Compare the exposure plot with the corresponding input distributions and covariate strata. Do not infer a biological subgroup from a histogram alone.
Worked Example: A Bimodal AUC Distribution From Two Groups
A hypothetical population contains 80 individuals with clearance 4 L/h and 20 with clearance 12 L/h. Every individual receives an intravenous dose of 600 mg per interval, and AUC is calculated as \(D/CL\). Explain the two exposure clusters and calculate their weighted mean.
For the first group, \(AUC=600/4=150\) mg·h/L, giving 80 values at that level. For the second, \(AUC=600/12=50\) mg·h/L, giving 20 values. The two clusters follow directly from the two specified clearance groups; neither requires a random-effect sampling error to explain it.
The mean across all 100 individuals is \((80(150)+20(50))/100=(12000+1000)/100=130\) mg·h/L. The median is 150 mg·h/L because the first 80 ordered values are 150, followed by 20 values at 50 only if ordered incorrectly; ordering from smallest to largest gives 20 values at 50 and then 80 at 150, so observations 50 and 51 are both 150 and the median is 150 mg·h/L. The distribution is bimodal and the mean is below the larger cluster, but those features are consistent with the stated group proportions and clearances. This check supports a model-based explanation; it does not establish that the assumed groups or proportions represent a real target population.
Common Failure Modes and Alternative Responses
A diagnostic is most useful when it narrows the cause without imposing an unsupported repair. The table contrasts several observations with checks that address their likely sources.
| Observed feature | What to check | Response to avoid |
|---|---|---|
| Negative value for a metric that must be positive | Formula, parameter bounds, missing-value handling, and units | Keep it as an extreme tail or silently replace it |
| Many exact zeros in a positive AUC metric | Rounding, failed calculations, and whether zeros were used to encode missingness | Assume zero exposure is clinically meaningful without checking |
| Very long right tail | Random-effect scale, low-clearance draws, covariate range, and transformation | Delete high values solely because they increase the mean |
| Two or more clusters | Population mixture, covariate strata, parameter sampling, and coding | Force a single symmetric distribution without a model-based reason |
Possible responses include fixing a proven coding or unit error, clarifying the target population, or revisiting an assumption when the observed behavior is incompatible with it. Truncating, winsorizing, or discarding simulated values changes the distribution and can change PTA. Such operations are not routine repairs: they require a justified model or population definition and transparent reporting. A larger Monte Carlo sample may describe a tail more precisely, but it does not correct a wrongly specified random-effect distribution or exposure formula.
These diagnostics have limits. A plausible range and shape cannot prove that the structural model, parameter estimates, covariate distribution, or target population are appropriate. Conversely, a rare extreme can be compatible with a correctly specified stochastic model. The relevant question is whether the value is possible and whether its probability and cause are consistent with the assumptions used to generate it.
Worked Example: Deciding Whether Zero AUC Is a Valid Boundary
In a hypothetical simulation of 1,000 individuals, the exposure output is \(AUC_{\tau,ss}\) in mg·h/L after a positive intravenous dose. The output contains 997 positive values and three exact zeros. Under \(AUC_{\tau,ss}=D/CL\), with positive \(D\) and finite positive \(CL\), determine whether the zeros can be accepted as legitimate simulated exposures.
The model implies \(D/CL>0\), so an exact zero cannot result from that formula with valid inputs. The proportion of zero outputs is \(3/1000=0.003\), or 0.3%, but this proportion does not make the values valid. Inspect those three records for missing or infinite clearance, failed calculations, rounding to zero, or a step that substitutes zero for an invalid result. Recalculate each value from its dose and clearance. If, for example, all three were missing values converted to zero, repair that handling and rerun or appropriately resolve the invalid records. Do not simply count the zeros as zero exposure in PTA calculations unless zero is genuinely part of the metric’s defined support.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For \(AUC_{\tau,ss}=D/CL\), what feature of the inputs guarantees a positive AUC, and what would an exact zero suggest?
- If \(\eta_{CL}\sim N(0,\omega^2)\) and \(CL=CL_{\mathrm{typ}}\exp(\eta_{CL})\), is right skew in AUC automatically evidence of a simulation error? Explain.
- A simulated population has two AUC clusters that align exactly with two prespecified covariate groups. What should you check before deciding that the clusters are erroneous?
- Why is increasing the number of Monte Carlo replicates not a repair for an incorrectly scaled random effect?
Solutions. (1) Positive dose and finite positive clearance imply \(D/CL>0\); an exact zero suggests a calculation, missing-value, rounding, or definition issue to investigate. (2) No. Under the stated assumptions, \(\log(AUC_{\tau,ss})\) is normal and AUC is lognormal, which is right-skewed. (3) Verify group definitions, proportions, covariate application, and the exposure calculation; if these match the intended model, the clusters may be expected. (4) More replicates reduce Monte Carlo sampling error for the distribution generated by the code, but they do not make that distribution match the intended random-effect model.