Why a Predictive Interval Can Change the Meaning of a Threshold Comparison
A predicted exposure ratio can lie on one side of a labeling threshold while plausible values lie on both sides. Reporting only the central prediction then hides a decision-relevant feature: uncertainty about whether the ratio crosses the boundary. A predictive interval makes the range of model-predicted values visible, while a probability such as \(P(R>T)\) quantifies how much predictive probability lies beyond a specified threshold \(T\).
These summaries answer related but different questions. An interval indicates the spread of predicted values over a stated coverage level. The exceedance probability gives the probability mass on one side of the threshold. Neither, by itself, determines a label recommendation. The interpretation depends on the endpoint, the target population and scenario, the uncertainty represented, and the clinical and regulatory context.
In PBPK applications, predictive draws may propagate parameter uncertainty, between-individual variability, trial variability, or combinations of these. The analyst must state what the draws represent. An interval for uncertainty in a population-mean ratio is not interchangeable with an interval for ratios across individual patients, or with a prediction interval for a future clinical trial estimate. The endpoint must also be explicit: an AUC ratio and a Cmax ratio can have different predictive distributions and may face different decision considerations.
The earlier material on confidence intervals distinguishes uncertainty about an estimated quantity from prediction of possible values. That distinction remains essential here: do not call a predictive interval a confidence interval, and do not infer its meaning from its label alone. A model-based predictive interval is not automatically a frequentist coverage guarantee or a probability statement about a fixed unknown parameter. Its interpretation follows from the predictive distribution that was constructed.
Thresholds, Intervals, and Probabilities
For an upper threshold, the central question is often whether the ratio is greater than \(T\). If a predictive interval lies entirely above \(T\), all values within that interval exceed the boundary; if it lies entirely below, all values within it do not. When the interval straddles \(T\), the interval alone does not say how much probability lies on either side. A threshold near one endpoint of an interval is different from a threshold near its center, even though both cases involve overlap.
A common PBPK approximation is that the logarithm of a positive exposure ratio is normally distributed. Write \(Y=\ln(R)\), with \(Y\sim N(m,s^2)\), where \(m\) and \(s\) are the mean and standard deviation on the log scale. Then \(R\) has a lognormal predictive distribution. Its central \(100(1-\alpha)\%\) predictive interval is obtained by exponentiating the log-scale quantiles:
Here \(z_q\) is the \(q\) quantile of the standard normal distribution. The probability of exceeding an upper threshold \(T\) is
where \(\Phi\) is the standard normal cumulative distribution function. The calculation uses the same predictive distribution as the interval; it does not require the threshold to be an interval endpoint. If the predictive distribution is skewed, multimodal, truncated, or otherwise poorly represented by a lognormal model, estimate the interval and exceedance probability from the full predictive draws or an appropriate fitted distribution instead.
Worked Interpretations at a Decision Boundary
Worked Example: An Interval Crosses the Threshold but Most Probability Does Not
Suppose a hypothetical PBPK analysis predicts an AUC ratio \(R\), with \(\ln(R)\sim N(\ln(1.80),0.25^2)\). For illustration, take an upper labeling boundary of \(T=2.00\). Calculate the central 90% predictive interval and the probability of exceeding the boundary.
For a central 90% interval, \(z_{0.95}=1.645\). The log-scale limits are \(\ln(1.80)\pm1.645(0.25)=0.5878\pm0.4113\), or \(0.1765\) and \(0.9991\). Exponentiating gives
The interval includes values below and above 2.00. The exceedance calculation is
Thus, about 33.7% of the predictive probability is above the hypothetical boundary, despite the interval extending well above it. The result supports describing the prediction as uncertain with a meaningful but not majority probability of exceeding 2.00; interval overlap alone would not justify saying that exceedance is more likely than not.
Worked Example: A Narrower Interval With a High Exceedance Probability
A second hypothetical PBPK analysis has \(\ln(R)\sim N(\ln(2.40),0.15^2)\), using the same upper boundary \(T=2.00\). Calculate its central 90% predictive interval and exceedance probability.
The log-scale interval is \(\ln(2.40)\pm1.645(0.15)=0.8755\pm0.2468\), giving limits \(0.6287\) and \(1.1223\). Therefore, the ratio-scale interval is
The lower limit is just below 2.00, so the interval still crosses the boundary. But
About 88.8% of the predictive probability exceeds the threshold. Compared with the first example, the interval is narrower on the log scale and centered further above the boundary. The combination supports a stronger statement that exceedance is likely under this model and scenario, while the crossing interval still makes clear that values below the threshold remain plausible.
Worked Example: A Wide Interval Signals Substantial Predictive Spread
Consider a third hypothetical result: \(\ln(R)\sim N(\ln(1.70),0.50^2)\), again with \(T=2.00\). Find the central 90% interval and probability above the threshold.
The log-scale limits are \(\ln(1.70)\pm1.645(0.50)=0.5306\pm0.8225\), or \(-0.2919\) and \(1.3531\). Exponentiation gives
For the exceedance probability,
The interval is much wider than in the second example, and the probability of exceeding 2.00 is about 37.3%. The result does not imply a low or high probability simply because the interval is wide; width describes dispersion, while the probability depends on where the predictive distribution lies relative to the threshold. Here, the broad spread indicates substantial uncertainty, and the exceedance probability quantifies the direction of that uncertainty.
Why the Exceedance Calculation Works
The lognormal calculation follows directly from the monotonicity of the logarithm. Because \(R>0\), taking logs preserves the ordering: \(R>T\) if and only if \(\ln(R)>\ln(T)\). Standardizing \(Y=\ln(R)\) gives \(Z=(Y-m)/s\sim N(0,1)\). Therefore,
The same transformation explains why the interval is asymmetric on the ratio scale: equal distances above and below \(m\) on the log scale become multiplicative, rather than additive, distances around the ratio-scale median \(\exp(m)\). In practice, the lognormal assumption should be checked against predictive draws. If it is unsuitable, the probability can be calculated as the fraction of draws above the threshold, and interval endpoints can be estimated as the corresponding empirical quantiles.
How to Describe the Strength of a Recommendation
A useful interpretation separates three pieces of information: the location of the predictive distribution relative to the boundary, its spread, and the consequences of being on either side. The exceedance probability summarizes the first two for a particular threshold, but does not encode the clinical cost of a false reassurance or an overly conservative action. Recommendation strength should therefore be expressed in calibrated language rather than inferred from a universal probability cutoff.
| Predictive finding | What it supports saying | What it does not establish by itself |
|---|---|---|
| Most predictive probability is below an upper boundary, with a narrow interval below it | The modeled scenario provides relatively consistent support for remaining below that boundary | That the ratio can never exceed the boundary, or that no clinical action is needed |
| The interval crosses the boundary and exceedance probability is near one half | The model leaves substantial uncertainty about which side of the boundary applies | A decisive classification based only on the point prediction |
| The interval crosses the boundary but exceedance probability is high | Exceedance is likely under the stated predictive distribution, though lower values remain plausible | Certainty, or a recommendation independent of clinical context |
| A very wide interval spans values far below and above the boundary | Predictive uncertainty is large and deserves explicit attention | That the central prediction alone adequately represents the decision evidence |
For a lower boundary, such as a ratio threshold below which a concern is triggered, use the corresponding probability \(P(R<T)\), or transform the ratio and boundary consistently. Always say which side is decision-relevant. The numerical threshold and the action associated with it are scenario-specific; the interval calculation does not establish a universal regulatory action rule.
Diagnostics, Alternatives, and Limitations
Before interpreting an interval that crosses a boundary, confirm that the predictive distribution targets the decision quantity. Verify the ratio definition, whether the endpoint is AUC or Cmax, the victim or perpetrator context, and the population and dosing scenario. Check that all intended uncertainty sources were propagated and that the simulated draws preserve parameter dependence. A narrow interval can reflect a narrowly defined target or omitted uncertainty rather than strong evidence about all relevant patients or trials.
For simulation-based predictions, inspect the number of draws, quantile stability, and the estimated fraction beyond the threshold. If that fraction changes materially across repeated simulation batches or the interval endpoints are unstable, increase simulation precision before making fine distinctions. Inspect the distribution itself for skewness, multiple modes, extreme tails, truncation, and influential parameter combinations. A single mean and standard deviation on the log scale can conceal these features. Sensitivity analyses across plausible model assumptions and input values help reveal whether the threshold conclusion is robust or scenario-dependent.
The main alternatives answer different questions. A point prediction is compact but does not show uncertainty. A predictive interval shows a central range but does not, without the distribution or probability calculation, reveal the fraction on either side of a boundary. A confidence interval for an estimated mean or parameter describes uncertainty about that estimand, not the spread of future individual or trial outcomes. A categorical label classification may be necessary for communication, but it discards information about distance from the threshold and predictive uncertainty. Reporting the estimate, interval, and exceedance probability together is often more informative than choosing only one summary.
A predictive interval also has limitations. Its coverage is conditional on the model and the uncertainty sources represented; model misspecification, biased inputs, or omitted variability can make it poorly calibrated. A 90% interval does not mean that the probability of being above the threshold is 10%, 50%, or any other fixed value. Nor does a threshold-exceedance probability represent the probability that a regulatory label is correct. It is a probability under the specified predictive distribution, not a substitute for clinical judgment or evidence about consequences.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A central 90% predictive interval crosses an upper boundary. Does that establish a 50% probability of exceeding the boundary? Explain.
- Suppose \(\ln(R)\sim N(\ln(1.50),0.20^2)\) and the upper boundary is \(T=2.00\). Is exceedance likely under this predictive distribution? Calculate \(P(R>T)\) approximately.
- Why can two predictive intervals that both cross the same threshold support different strengths of statement?
- What does a predictive interval mean if its simulations include between-individual variability but not uncertainty in model parameters?
Solutions. 1. No. The interval specifies quantiles; the probability above the threshold depends on the threshold’s location within the predictive distribution. Calculate the tail probability directly. 2. The standardized threshold is \(\{\ln(2.00)-\ln(1.50)\}/0.20=0.2877/0.20=1.439\). Thus \(P(R>2.00)=1-\Phi(1.439)\approx 0.075\), or about 7.5%; exceedance is not likely under this distribution. 3. The intervals can differ in width and in how much predictive probability lies beyond the threshold. Their exceedance probabilities and uncertainty should be compared, not just the fact of overlap. 4. It describes the modeled spread across individuals conditional on the parameter values used, but it does not include parameter uncertainty; it should not be presented as covering that omitted uncertainty.