What a Predictive Interval Report Must Make Clear
A predictive interval is interpretable only when its target is explicit. A range for AUC in one virtual subject, a range for population AUC values, and a range for a population mean AUC answer different questions, even if their numerical limits happen to be similar. Reporting only two numbers hides which prediction was made and which sources of uncertainty were included.
Let \(Y\) denote the prediction quantity for the target specified by the analysis, such as an individual subject’s AUC after a stated dose. A PBPK model maps model inputs \(X\) to a prediction \(Y=g(X)\). The distribution assigned to \(X\), and whether it includes between-subject variability, uncertain parameter values, or both, determines the predictive distribution of \(Y\). A reported interval summarizes that distribution; it does not by itself establish that the model or its inputs are valid for a particular application.
The report should name the outcome and its units, the population or subgroup represented, and any relevant conditions such as dose, route, formulation, and time point. It should say whether the target is a future individual, a typical individual, or a population summary. These are not interchangeable targets. For example, an interval for individual concentrations ordinarily includes subject-to-subject variation if that is part of the specified prediction; an interval for a population mean does not describe the spread of individual concentrations.
Inputs, Simulated Population, and Sampling
For every input treated as uncertain, report what it represents and how its distribution was specified. Depending on the analysis, these inputs may include physiological quantities, drug-specific parameters, or subject characteristics. State distribution families and parameter values where appropriate, units, bounds or truncation, and dependencies among inputs. Also name important inputs held fixed. A statement such as “parameters were sampled” is insufficient if readers cannot tell which parameters, from what distributions, or with what dependence structure.
Describe the virtual population as well as the input distributions. State eligibility characteristics, covariate distributions, subgroup definitions, and how subjects were generated or selected. If scenarios are combined, report the scenario weights and whether the interval summarizes the weighted mixture or a separate distribution for each scenario. Do not imply that simulated subjects are a representative sample of a real population unless that interpretation is justified by the population construction.
Sampling details affect both reproducibility and the interpretation of simulation error. Report the number of model evaluations, the sampling method (for example, independent random draws, stratified sampling, or Latin hypercube sampling), any random seed if reproducibility requires it, and how failed or invalid simulations were handled. Explain whether the reported quantiles were calculated from all valid predictions or from a defined subset. If the sampling design induces dependence among predictions, an independent-draw Monte Carlo precision formula should not be presented as though it applied unchanged.
Worked Example: Reporting a Central Interval for AUC
Hypothetical analysis. A PBPK analysis simulates AUC from zero to infinity after a single 100 mg oral dose in virtual adults meeting a stated age and renal-function range. AUC is reported in mg·h/L. Uncertain clearance and fraction absorbed are sampled from specified distributions, with their dependence structure stated; body weight varies across virtual adults. Other inputs are held fixed at documented values. Independent random draws produce \(M=20{,}000\) model predictions.
The analysis defines a central 90% predictive interval by assigning 5% probability to each tail. Thus its population limits are \(q_{0.05}\) and \(q_{0.95}\), where \(q_p\) is the \(p\)th quantile of the specified predictive distribution. Suppose the empirical limits are 12.0 and 48.0 mg·h/L. The report should call this a central 90% predictive interval, not merely “the 90% interval,” because that wording does not specify tail allocation.
The interval can be reported as: “For the stated virtual-adult population and input distributions, the simulated central 90% predictive interval for individual AUC after a single 100 mg oral dose was 12.0 to 48.0 mg·h/L (20,000 independent simulations; empirical 5th and 95th percentiles).” The surrounding methods must specify the sampling distributions, dependence assumptions, quantile algorithm, and Monte Carlo precision. The limits describe the modeled predictive distribution under those choices, not a guarantee about every future individual.
Define the Interval and Its Monte Carlo Precision
For a continuous predictive distribution of \(Y\), a central \(100(1-\alpha)\%\) interval uses the quantiles \(q_{\alpha/2}\) and \(q_{1-\alpha/2}\). A one-sided upper \(100(1-\alpha)\%\) prediction limit is \(q_{1-\alpha}\); a one-sided lower limit is \(q_{\alpha}\). State which form is used, its nominal coverage, and the tail probability or probabilities. For finite simulations, also identify the empirical quantile convention, such as the interpolation rule used between ordered predictions. Different conventions can produce slightly different limits, especially with small simulation samples or discrete outcomes.
Monte Carlo precision concerns how accurately the simulation estimates a feature of the predictive distribution. It is not the predictive interval itself and does not measure uncertainty about whether the model is scientifically adequate. For independent draws from a continuous distribution with density \(f\) that is positive and smooth near \(q_p\), the large-sample standard error of the estimated \(p\)th quantile is approximately:
This approximation follows from the empirical cumulative distribution function. At the true quantile, each draw is below \(q_p\) with probability \(p\), so the empirical proportion below it has variance \(p(1-p)/M\). Near \(q_p\), a small change \(\Delta y\) in the value changes the cumulative probability by approximately \(f(q_p)\Delta y\). Dividing the standard deviation of the empirical proportion by this local rate of change gives the displayed approximation. It requires independent draws and a well-behaved, nonzero density near the quantile; it can be unreliable in sparse tails, at discontinuities, or for dependent sampling designs.
Worked Example: Quantifying Monte Carlo Error in a Limit
Continue the hypothetical AUC analysis with \(M=20{,}000\) independent draws. Suppose the estimated density at the 5th percentile, 12.0 mg·h/L, is \(0.0040\) per mg·h/L. The approximate Monte Carlo standard error of that percentile is:
If the estimated density at the 95th percentile is \(0.0030\) per mg·h/L, then:
A concise report can therefore give the estimated limits, 12.0 and 48.0 mg·h/L, together with approximate Monte Carlo standard errors of 0.385 and 0.514 mg·h/L, respectively, and state how the density estimates were obtained. These standard errors quantify simulation precision under the stated approximation; they are not additional predictive limits for AUC. If the density estimates are unstable, increase the simulation size or use an appropriate simulation-based precision assessment rather than reporting unjustified precision.
Match the Report to the Prediction Target
The interval definition must be consistent with the prediction target. If the target is an individual in a population, the predictive distribution should include the intended individual variability as well as any specified input uncertainty. If the target is a population mean, it should instead be the distribution of that mean under the stated model and sampling scheme. The interval’s label alone cannot resolve this distinction.
Worked Example: A One-Sided Prediction Limit
Hypothetical analysis. A model predicts steady-state trough concentration \(C_{\min,ss}\), in ng/mL, for individual virtual adults in a defined population receiving a stated regimen. The analysis uses 10,000 independent draws from its specified population and input distributions. It reports an upper one-sided 95% prediction limit, defined as the empirical 95th percentile. Suppose this limit is 2.70 ng/mL.
The tail probability is \(\alpha=0.05\), so the limit is \(q_{1-\alpha}=q_{0.95}\). It should be reported as: “The simulated upper one-sided 95% prediction limit for individual steady-state trough concentration was 2.70 ng/mL in the specified virtual-adult population (10,000 independent draws; empirical 95th percentile).” The report should also give the input distributions, sampling and quantile methods, and Monte Carlo precision. Calling 2.70 ng/mL a “95% central interval” would be incorrect: no lower endpoint or equal-tail allocation was calculated.
When multiple plausible scenarios are combined, distinguish scenario-specific intervals from an interval for their mixture. A mixture interval is calculated from the combined predictive distribution using the stated scenario weights; it is generally not obtained by averaging the scenario-specific interval limits. If the aim is to show scenario differences, report the scenario-specific results rather than relying only on a pooled range.
The same care applies to concentration-time results. A pointwise interval at each time describes the distribution at that time separately. It does not, by itself, describe an interval that contains an entire simulated concentration-time curve with a stated probability. Name the interval type and time grid explicitly so the reader does not mistake pointwise coverage for simultaneous curve coverage.
Diagnostics, Alternatives, and Limitations
Useful diagnostics include checking that the number of valid runs is reported, that simulated predictions are finite and physiologically plausible, and that the tails are adequately represented for the chosen quantiles. Inspect the prediction distribution for skewness, multimodality, or a point mass, since these features can make a normal approximation or a smooth-density Monte Carlo error estimate inappropriate. Compare quantile estimates across independent batches or larger simulation runs when feasible. For stratified, quasi-random, or otherwise dependent designs, use a precision method suited to that design rather than assuming independent draws.
The empirical quantile interval is a direct summary of simulated predictions and can reflect skewness or multimodality. A normal-approximation interval may be simpler, but it relies on an adequate approximation to the predictive distribution and can misrepresent tail probabilities when that approximation fails. A central interval and a one-sided limit answer different questions, so neither is a universal substitute for the other. A pointwise interval and a simultaneous band also have different coverage targets. State the chosen approach and why it matches the prediction being communicated.
No reporting checklist can compensate for misspecified input distributions, omitted dependencies, an unsuitable virtual population, or model limitations. The reported interval is conditional on the model structure and assumptions used to generate it. Numerical convergence, Monte Carlo precision, and biological credibility are separate considerations: a highly precise simulation can still produce a misleading prediction if its inputs or model structure do not support the intended target.
Common Reporting Pitfalls
- Reporting limits without a target. “The interval was 12 to 48” omits the outcome, units, population, and conditions. Name them.
- Calling Monte Carlo error predictive uncertainty. A quantile standard error describes numerical precision in estimating a limit, not the spread of individual outcomes.
- Leaving tail allocation implicit. State whether the interval is central or one-sided and identify the quantiles used.
- Pooling scenarios without saying so. A mixture interval depends on scenario weights and is not generally the average of scenario limits.
- Reporting excessive digits. Round limits to a precision supported by the simulation and the underlying inputs; include units.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A report gives an interval for individual \(C_{\max}\) but does not state the population or whether between-subject variability was simulated. What essential interpretation is missing?
- For a central 90% predictive interval, which quantiles define the limits?
- In one sentence, distinguish a Monte Carlo standard error for an estimated quantile from predictive uncertainty.
- Why should a pooled interval for two weighted scenarios not be described as the average of their scenario-specific limits?
- What additional caution is needed when quantile estimates come from dependent or quasi-random draws?
Solutions. (1) The reader cannot tell which individuals the prediction represents or whether the interval includes their between-subject spread, so the prediction target and included variability must be stated. (2) The 5th and 95th percentiles, \(q_{0.05}\) and \(q_{0.95}\). (3) Monte Carlo standard error describes numerical precision in estimating the quantile, whereas predictive uncertainty describes variation in the modeled outcomes. (4) A mixture distribution is formed by combining scenario-specific distributions with their weights; its quantiles are calculated from that combined distribution and generally are not averages of separate quantiles. (5) Independent-draw formulas may not apply; report a precision assessment appropriate to the sampling design.