From a Fitted Model to a Development Prediction
A population PK model is useful for a development decision only when its parameters are translated into quantities relevant to that decision. A covariate effect on clearance, for example, is not itself a statement about how much exposure differs between patient groups. The predicted exposure also depends on the dose, the rest of the PK model, the distribution of individual variability, and uncertainty in the fitted parameters.
The goal here is to predict and compare exposure distributions for clinically relevant subgroups. These are model-based predictions for populations under specified conditions, not summaries of observed subgroup averages and not evidence that subgroup membership causes an exposure difference. The earlier material on structural models, covariate models, model diagnostics, and parameter uncertainty supplies the ingredients; this tutorial shows how to combine them for subgroup prediction.
For linear PK at steady state, the area under the concentration-time curve over a dosing interval, \(AUC_{\tau}\), is related to dose and clearance by \(AUC_{\tau}=F\cdot Dose/CL\), where \(F\) is bioavailability. For an intravenous dose, \(F=1\). This relation is a useful illustration, but predictions for other exposure metrics, routes, dosing schedules, or nonlinear models should be generated from the corresponding fitted model rather than inferred from this formula.
Typical Exposure Is Not the Exposure Distribution
Suppose a model predicts that subgroup \(g\) has a typical exposure \(M_g\), and that individual exposure is lognormally distributed:
Here \(M_g\) is the median exposure for the subgroup, and \(\omega_g\) is the standard deviation on the log scale. The geometric standard deviation is \(\exp(\omega_g)\). This representation assumes that log exposure is normally distributed within the subgroup under the specified regimen; it does not assert that all individuals have exposure \(M_g\).
A central 95% individual prediction interval, when the parameters are treated as fixed, is approximately
This interval describes the model-predicted spread for individuals, not uncertainty about the subgroup median. Parameter uncertainty is a separate source of uncertainty and should be propagated when the aim is to assess how reliably the model supports a development decision.
Worked Example: Translating Typical Clearance Into Typical AUC
Consider a hypothetical linear intravenous PK model for a 100 mg dose. At the reference covariate values, fitted typical clearance is 5.00 L/h. For a subgroup, the fitted covariate model predicts typical clearance of 4.00 L/h. Assume bioavailability is 1 and use \(AUC_{\tau}=Dose/CL\) for the relevant steady-state interval.
Reference group. The typical AUC is \(100\ \mathrm{mg}/(5.00\ \mathrm{L/h})=20.0\ \mathrm{mg\cdot h/L}\).
Subgroup. The typical AUC is \(100\ \mathrm{mg}/(4.00\ \mathrm{L/h})=25.0\ \mathrm{mg\cdot h/L}\). The ratio is \(25.0/20.0=1.25\), so the model predicts a 25.0% higher typical AUC for this subgroup at the same dose.
This is a comparison of typical values only. It does not show that every subgroup member has higher exposure than every reference-group member. To assess individual-level overlap, the model’s between-individual variability and any relevant uncertainty in the fitted parameters are also needed.
Comparing Distributions and Their Overlap
A useful comparison asks about two individuals sampled independently, one from each subgroup, under the specified regimen. If their log AUCs are normally distributed and independent, then their difference is also normal. In particular, if subgroup A and B have medians \(M_A\) and \(M_B\), and log-scale standard deviations \(\omega_A\) and \(\omega_B\), then
The variance adds because the two sampled individuals are independent. If the comparison instead involves paired individuals or correlated predictions, this independence assumption is not appropriate; the covariance must be included. The expression below gives the probability that the randomly selected A individual has higher exposure than the randomly selected B individual:
This probability is a compact description of distributional separation. It is not the percentage of individuals in subgroup A who exceed a fixed clinical threshold, and it does not imply that subgroup membership causes the difference.
Worked Example: Probability That an A Individual Has Higher AUC Than a B Individual
For a hypothetical comparison at the same dose, suppose subgroup A has median AUC \(M_A=20.0\ \mathrm{mg\cdot h/L}\), subgroup B has median AUC \(M_B=16.0\ \mathrm{mg\cdot h/L}\), and both have log-scale standard deviation \(\omega=0.300\). Assume independent draws from the two subgroup distributions.
The mean of the log AUC difference is \(\ln(20.0/16.0)=\ln(1.25)=0.223144\). Its standard deviation is \(\sqrt{0.300^2+0.300^2}=\sqrt{0.180}=0.424264\). Therefore,
The model therefore predicts about a 70.1% chance that a randomly selected A individual has higher AUC than an independently selected B individual. There is still substantial overlap: this probability is not close to 1, even though the typical AUC in subgroup A is 25% higher. Reporting both the typical ratio and an overlap-oriented measure gives a more informative comparison than reporting the ratio alone.
Threshold Probabilities and Parameter Uncertainty
A development question may concern the fraction of individuals above a specified exposure threshold, rather than the comparison between two randomly selected individuals. Under the lognormal model with fixed parameters, the probability of exceeding threshold \(T\) is
The threshold must have the same units and refer to the same exposure metric as the model prediction. The resulting proportion is conditional on the model, regimen, subgroup definition, and chosen threshold. It is not automatically a clinical risk: connecting exposure to an outcome requires evidence and methods beyond the exposure-distribution prediction.
Parameter uncertainty matters because \(M_g\), \(\omega_g\), and covariate effects are estimated. A practical approach is to draw parameter sets from an appropriate uncertainty distribution, such as a parameter bootstrap or a justified approximation based on the estimated parameter covariance, and simulate individuals for each draw. The spread of subgroup summaries across draws describes uncertainty in those summaries. When each draw also simulates individual random effects, the resulting predictive distribution reflects both parameter uncertainty and population variability; those components should be reported distinctly where possible.
Worked Example: Fraction Predicted Above an AUC Threshold
For hypothetical subgroup A, suppose \(M_A=20.0\ \mathrm{mg\cdot h/L}\), \(\omega_A=0.300\), and the exposure threshold is \(25.0\ \mathrm{mg\cdot h/L}\). First treat these model parameters as fixed. The standardized log threshold is \(\ln(25.0/20.0)/0.300=0.223144/0.300=0.7438\). Thus,
The fixed-parameter model predicts that about 22.85% of subgroup A would exceed this threshold under the specified regimen. This estimate describes individual variability conditional on the fitted parameters; it does not by itself express uncertainty about the model’s estimated median or variability.
Now suppose the estimated log median, \(\widehat{\mu}_A=\ln(20.0)\), has approximate standard error \(0.0500\), and for illustration assume this uncertainty is independent of individual log variability, with \(\omega_A=0.300\) treated as known. Under a normal approximation for uncertainty in the true log median, the marginal predictive log variance is \(0.0500^2+0.300^2=0.0925\), with standard deviation \(\sqrt{0.0925}=0.3041\). The threshold probability accounting for this uncertainty is approximately \(1-\Phi(0.223144/0.3041)=1-\Phi(0.7337)\approx0.2316\), or 23.16%.
This simplified calculation illustrates how uncertainty in the estimated median can broaden a predictive distribution. In a real analysis, uncertainty in multiple correlated parameters, covariate effects, and variability terms should be propagated using the fitted model’s joint uncertainty, not by assuming independence without justification.
Practical Prediction Workflow and Diagnostics
For complex models, simulation is generally more flexible than attempting a closed-form calculation. A defensible workflow makes the target population and uncertainty sources explicit before producing subgroup summaries.
Specify the subgroup, dose and schedule, exposure metric, time horizon, and any clinically relevant threshold. State which covariate values or distributions define the target subgroup.
Use the fitted parameter estimates for a conditional prediction, or draw joint parameter sets from a justified uncertainty procedure when parameter uncertainty is required.
For each parameter set, generate covariates appropriate to the subgroup and individual random effects from the model assumptions. Calculate the requested exposure metric using the fitted structural model.
Report subgroup medians or other prespecified summaries, individual prediction intervals, threshold proportions, and distributional overlap. Separate variability across people from variation across parameter sets.
Assess whether the covariate ranges, dosing conditions, and concentration-time patterns relevant to the prediction are supported by the data and model diagnostics. Examine sensitivity to plausible model and uncertainty assumptions.
A common failure is to simulate a subgroup using covariate values that do not represent the intended population. For example, assigning every member the subgroup’s typical covariate value suppresses covariate variation and can understate the exposure spread. Conversely, comparing groups with different age or weight distributions when the intended question is a subgroup effect can confound subgroup contrast with population composition. Define whether the goal is prediction in each group as observed, or a standardized comparison under a common covariate distribution.
Predictions also become unreliable when the target subgroup lies outside the covariate range or dosing conditions represented in the development data. Sparse sampling, influential observations, unresolved model misspecification, or weakly estimated variability can produce apparently precise summaries that are not well supported. Examine simulation-based checks and external or temporal evaluation where available, and investigate implausible tails or subgroup patterns rather than treating simulation output as validation.
Choosing a Summary and Recognizing Its Limits
| Approach | What it communicates | Main limitation |
|---|---|---|
| Typical exposure ratio | Relative difference between subgroup typical values | Hides individual variability and distributional overlap |
| Individual prediction intervals | Model-predicted spread of exposure within each subgroup | May omit parameter uncertainty unless it is propagated |
| Probability of one subgroup individual exceeding another | Distributional separation for two independently sampled individuals | Depends on distributional assumptions and is not a threshold probability |
| Threshold exceedance probability | Predicted proportion above a prespecified exposure value | Does not establish clinical consequence or validate the threshold |
Empirical Bayes estimates are not a substitute for these population predictions. They are individual-level estimates informed by each subject’s data and can be shrunk toward the population mean, especially with sparse information. Comparing their raw subgroup distributions may therefore reflect sampling design and shrinkage as well as underlying population differences. Population simulation instead uses the fitted model to generate the target population under explicit assumptions.
No single summary answers every development question. A typical ratio is easy to communicate but can conceal substantial overlap. Prediction intervals display spread but can be difficult to compare when they are wide or asymmetric. Threshold probabilities are directly relevant only when the threshold is meaningful for the decision. All are conditional on the exposure metric, regimen, model structure, covariate representation, and uncertainty procedure.
Common Pitfall
Correction: The 25% statement compares typical values only. Individual exposures vary, and subgroup distributions may overlap substantially. Report the typical ratio alongside individual variability and, when relevant, an overlap or threshold measure. State whether parameter uncertainty is included.
A related mistake is to interpret a simulated percentage above a threshold as an observed clinical event rate. The simulation estimates a proportion of exposures under the model; it does not establish the probability of toxicity or benefit. That interpretation requires an appropriate exposure-outcome analysis and its assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A subgroup has median AUC \(30.0\ \mathrm{mg\cdot h/L}\) and log-scale standard deviation \(0.200\). What is its approximate central 95% individual prediction interval when parameters are treated as fixed?
- Why is the probability that one A individual exceeds one B individual not the same as the probability that an A individual exceeds a fixed threshold?
- What happens if a subgroup simulation assigns every individual the same covariate value, despite meaningful variation in that covariate in the target population?
- In a simulated subgroup prediction, identify two distinct sources of variation that should not be conflated.
Solutions. (1) The interval is \([30.0\exp(-1.96\times0.200),30.0\exp(1.96\times0.200)]\) \(=[30.0\exp(-0.392),30.0\exp(0.392)]\approx[20.3,44.4]\ \mathrm{mg\cdot h/L}\). (2) The first probability compares two independently sampled individual exposures; the second compares one individual’s exposure with a specified value. They use different events and, generally, different calculations. (3) The simulation removes covariate variation and may understate the target population’s exposure spread; it may also misrepresent its median if the chosen value is not representative. (4) Between-individual variability describes differences among people under fixed parameters. Parameter uncertainty describes uncertainty in the fitted model’s population parameters. A prediction can include either or both, but should say which.