Why an Interval Target Needs an Explicit Rule
Population targets are not always reported as a single number. A source may give a clinically plausible range, a rounded interval, or an estimate accompanied by uncertainty. For example, a report might describe a biomarker median as 20 to 30 units, or give an estimated mean of 28 units with a standard error. These statements do not mean the same thing. The first may define a range of values considered acceptable; the second describes uncertainty about an estimated quantity. A virtual population can be judged differently depending on which meaning applies.
The distinction matters when a predicted summary is close to a range boundary. A strict rule can reject it even when the discrepancy is small. A permissive rule can allow a prespecified practical deviation. An uncertainty-aware rule additionally asks whether the estimate is precise enough to support the decision. None of these rules is automatically correct: the target's meaning, the scientific purpose, and the decision costs must determine the rule before results are inspected.
Three Acceptance Rules
The strict rule accepts a virtual population only when its summary lies within the reported interval:
This rule is appropriate when the endpoints are genuine hard bounds or when the analysis requires literal agreement with the reported range. It does not account for rounding, simulation error, or any practical allowance outside the range. A summary one small unit below \(L\) fails just as surely as a much more discrepant value.
The permissive rule expands the range by the tolerance margin and accepts when
The margin should be set before assessing candidate populations. It should reflect a meaningful scale, such as measurement precision or an explicitly accepted discrepancy, not be chosen afterward to make a preferred population pass. A margin is not a confidence level and does not quantify uncertainty in the target.
An uncertainty-aware rule accounts for uncertainty in the model summary as well as the target's acceptable range. Suppose \(m\) is estimated with standard error \(s_m\), and a suitable confidence interval is \(m \pm c s_m\). For example, \(c\) may be 1.96 for an approximate 95% normal interval when the normal approximation is justified. A conservative rule accepts only if this whole interval lies inside the tolerance-expanded target:
This is not a claim that there is a 95% probability that the fixed model quantity is in this particular interval. It is a confidence-interval-based decision rule: under the method's assumptions, the interval procedure has its stated repeated-sampling coverage. Requiring the entire interval to fit makes the rule more demanding than checking the point estimate alone.
Worked Example: Strict Versus Permissive Acceptance
Suppose a hypothetical QSP analysis targets a virtual-population median clearance between 20 and 30 L/h. A candidate population has a median of 19.2 L/h. Compare the strict rule with a prespecified tolerance margin of 2 L/h.
Under strict acceptance, the required condition is \(20 \leq 19.2 \leq 30\), which is false. The candidate does not pass.
The distance to the interval is
Under the permissive rule, the expanded interval is \([20-2,30+2]=[18,32]\) L/h. Since \(18 \leq 19.2 \leq 32\), the candidate passes. This conclusion depends on treating 2 L/h as a defensible allowance; it does not mean the original target range was met.
Using Uncertainty Without Hiding Mismatch
An uncertainty-aware assessment starts by specifying what varies. If the uncertainty is Monte Carlo error in estimating a fixed virtual-population summary, \(s_m\) should measure that estimation error. If the uncertainty represents variation across parameter sets or plausible models, the resulting interval answers a different question. Such a spread should not be labeled a confidence interval for a single fixed summary unless its construction supports that interpretation.
For an interval target \([L,U]\), the conservative confidence-interval rule can be checked by two endpoint comparisons. The lower endpoint of the model-summary interval must be no lower than \(L-\delta\), and its upper endpoint must be no higher than \(U+\delta\). If either condition fails, the point estimate may still pass the permissive rule, but the available precision or uncertainty does not support this stricter decision.
Worked Example: An Interval Target With Model Uncertainty
Suppose a hypothetical target for mean biomarker concentration is 25 to 30 mg/L. The model estimates a mean of 27.0 mg/L with standard error 1.0 mg/L. Use \(\delta=0\) and an approximate 95% normal interval, with \(c=1.96\). Assume the estimate is approximately normal and the standard error is appropriate for the uncertainty being assessed.
The point estimate passes the strict rule because \(25 \leq 27.0 \leq 30\). Its approximate 95% interval is
Both endpoints lie inside \([25,30]\) mg/L: \(25.04 \geq 25\) and \(28.96 \leq 30\). Thus the uncertainty-aware rule also passes. Now suppose the standard error were 1.5 mg/L, with the same point estimate and target. The interval would be
The point estimate still meets the strict rule, but the interval extends below 25 mg/L. The uncertainty-aware rule fails. The difference is not a change in the estimated mean; it is a change in how much precision supports the conclusion.
Some reports give a point estimate with uncertainty rather than an acceptable range. In that case, treating the uncertainty interval as if every value in it were an acceptable target can be misleading. A useful alternative is to compare the model estimate with the target estimate using a prespecified practical equivalence margin \(\delta\).
Why the Combined-Error Rule Works
The difference between two independent, approximately normal estimates has mean equal to the difference of their means and variance equal to the sum of their variances. In particular, if their errors have variances \(s_m^2\) and \(s_t^2\), independence gives a covariance of zero, so the variance of their difference is \(s_m^2+s_t^2\). The standard error is therefore \(s_d=\sqrt{s_m^2+s_t^2}\).
An approximate confidence interval for the difference is \([d-cs_d,d+cs_d]\). To require that the entire interval be inside \([-\delta,\delta]\), both inequalities must hold:
The first inequality is equivalent to \(d\geq-\delta+cs_d\), and the second to \(d\leq\delta-cs_d\). Together these say \(|d|+cs_d\leq\delta\). Thus the rule requires both a sufficiently small estimated discrepancy and enough precision to keep the whole interval within the practical margin. If the estimates are dependent, the covariance must be included; simply adding the variances is then not justified.
Worked Example: Comparing Two Uncertain Estimates
A hypothetical external study reports a target mean concentration of 28 mg/L with standard error 1.2 mg/L. A virtual population gives an estimated mean of 29 mg/L with standard error 0.8 mg/L. Treat the estimates as independent and approximately normal, and set the practical equivalence margin to \(\delta=4\) mg/L. Use \(c=1.96\).
The estimated difference, model minus target, is \(d=29-28=1\) mg/L. The combined standard error is
The approximate 95% interval for the difference is
The entire interval is within \([-4,4]\) mg/L, so the uncertainty-aware equivalence rule passes. The point discrepancy alone, 1 mg/L, would also be smaller than 4 mg/L, but it would not show whether uncertainty was small enough to support that conclusion. If the margin had been 3 mg/L, the interval would not fit, and the rule would fail.
Choosing and Checking a Rule
Before applying any rule, confirm that the model summary and target describe the same estimand: for example, both are means, both concern the same subgroup and time point, and both use compatible units and definitions. An interval for a median cannot be substituted for one for a mean. Nor should a marginal target be treated as a target for a joint distribution. Earlier material on matching means, variances, and quantiles and on comparing population distributions provides tools for defining the summary to be compared; interval handling does not repair a mismatch in its definition.
The interval's provenance also matters. A clinically specified acceptable range, a rounded published range, and a confidence interval around an estimate require different interpretations. A confidence interval is not automatically a tolerance margin: it describes uncertainty under a specified estimation procedure, whereas \(\delta\) represents a practical allowance. If both target and model estimates are uncertain, include both sources in the comparison when justified, rather than treating one as exact.
Useful diagnostics include recording the distance to the nearest endpoint, reporting the margin and the rule used, and checking whether the decision changes under plausible alternatives for \(\delta\) or the uncertainty estimate. For an uncertainty-aware interval rule, inspect both endpoints of the model interval. For a comparison of two estimates, inspect the assumptions behind independence and the normal approximation. If a decision changes with a tiny adjustment to a rounded endpoint or margin, report that sensitivity rather than presenting the pass/fail result as robust.
The main alternatives answer different questions. Strict acceptance asks whether the point estimate is inside the reported range. Permissive acceptance asks whether it is within a practical distance of that range. The uncertainty-aware rule asks whether the available precision supports the comparison at a stated confidence level. A probability-based rule, such as requiring a specified fraction of uncertainty draws to fall inside the range, is another possible approach, but its interpretation depends on how those draws were generated. It is not interchangeable with a confidence-interval rule.
| Rule | Decision criterion | What it does not establish |
|---|---|---|
| Strict | Point summary lies in \([L,U]\) | That the summary is precise or robust to uncertainty |
| Permissive | Point summary lies in \([L-\delta,U+\delta]\) | That the original interval was met or that \(\delta\) is statistically derived |
| Uncertainty-aware | Model uncertainty interval lies within the permitted region, or a difference interval lies within \([-\delta,\delta]\) | That assumptions, estimands, or the chosen margin are appropriate |
Common Pitfalls and Limitations
- Calling every reported interval a target range. A confidence interval represents estimation uncertainty; it is not automatically a band of acceptable model values. Identify what the source's endpoints mean.
- Choosing the margin after seeing the result. This makes acceptance easier to obtain and weakens the decision. Define \(\delta\) using scientific or decision-based reasoning in advance.
- Ignoring units or estimands. A numerical comparison is not meaningful when the model and target use different units, populations, time points, or summary definitions.
- Treating uncertainty as a cure for poor fit. A large standard error can make a model result inconclusive; it does not make the point estimate more concordant with the target.
- Assuming independence without checking. If the target and model estimates share data or a common calibration, their covariance affects the standard error of the difference.
All three rules reduce a complex population comparison to a decision about a specified summary. Passing one interval criterion does not establish that the entire distribution, its tails, or its dependence structure matches the target. These rules also depend on the validity of the virtual population and on the target's relevance to the intended application. They should therefore be reported alongside the exact summary, interval, margin, uncertainty method, and acceptance result.
Check Your Understanding
Answer each question, then compare with the solutions.
- A target interval is 40 to 50 units, and a candidate summary is 38.5 units. What is the distance to the interval, and does it pass a permissive rule with \(\delta=2\) units?
- A model estimate is 12 with standard error 0.5. Its target interval is 11 to 13, and \(\delta=0\). Does its approximate 95% interval fit entirely inside the target interval? Use \(c=1.96\).
- A model estimate and an independent target estimate differ by 0.5 mg/L. Their standard errors are 0.3 and 0.4 mg/L. Find the combined standard error.
- Why is a confidence interval around a target estimate not automatically a practical tolerance margin?
- Name two assumptions needed for the combined-error equivalence calculation in the worked example.
Solutions. 1. The distance is \(40-38.5=1.5\) units. The expanded interval is \([38,52]\), so 38.5 passes. 2. The interval is \(12\pm1.96(0.5)=[11.02,12.98]\), which lies inside \([11,13]\); it passes. 3. \(s_d=\sqrt{0.3^2+0.4^2}=\sqrt{0.25}=0.5\) mg/L. 4. A confidence interval describes uncertainty from an estimation procedure, while a tolerance margin specifies a practically acceptable discrepancy; they have different meanings. 5. The estimates must be independent (or their covariance must be included), and the normal approximation and standard-error estimates must be appropriate.