From Relative Standard Error to an Interval
A parameter estimate is a point on a scale; its relative standard error (RSE) expresses its estimated sampling uncertainty relative to that point. An approximate confidence interval adds a range of plausible values, under a specified sampling approximation. In population pharmacokinetics, this can help summarize uncertainty in estimates such as clearance or volume, but the calculation is only as dependable as the approximation behind it.
For a positive estimate \(\hat{\theta}\), RSE is commonly reported as a percentage:
The standard error here describes sampling uncertainty in the estimated parameter. It is not the standard deviation of individual pharmacokinetic values, interindividual variability, or residual error. Converting RSE to a standard error assumes that the reported RSE is calculated on the same parameter scale as \(\hat{\theta}\).
This interval relies on the estimator being approximately normally distributed around the true parameter, with a standard error that estimates its sampling variability adequately. The approximation is most defensible when the model is identifiable, the estimate is not near a parameter boundary, and the likelihood and estimator behave regularly enough for a normal approximation to be reasonable. An RSE value alone cannot verify these conditions.
Why the Wald Formula Works
If the estimator is approximately normal, then the standardized estimation error is approximately standard normal:
A standard normal variable lies between \(-z_{1-\alpha/2}\) and \(z_{1-\alpha/2}\) with probability approximately \(1-\alpha\). Rearranging that event gives the interval formula:
The confidence level describes the long-run coverage of the procedure under its assumptions: across repeated comparable data sets, approximately 95% of intervals constructed this way would contain the fixed parameter. It does not mean that, after observing one interval, there is a 95% probability that the fixed parameter lies inside it.
Worked Example: A 95% Interval for Clearance
Suppose a hypothetical population analysis estimates clearance as \(\hat{\theta}=5.0\) L/h, with an RSE of 12%. Construct a raw-scale approximate 95% interval.
First convert the percentage to a standard error:
Using \(z_{0.975}=1.96\), the margin of error is \(1.96\times0.60=1.176\) L/h. Therefore:
The approximate 95% interval is 3.824 to 6.176 L/h. Its endpoints are positive, and the RSE is not especially large, so the result is not immediately implausible on the parameter scale. That does not establish normality or coverage; those remain model- and data-dependent assumptions.
Choosing the Confidence Level
The critical value changes with the confidence level. For a two-sided interval, a 90% interval uses approximately 1.645, a 95% interval uses approximately 1.96, and a 99% interval uses approximately 2.576. Higher confidence gives a wider interval when the estimate and standard error are unchanged. State the confidence level and units whenever reporting an interval.
Worked Example: A 90% Interval for Volume
Suppose a hypothetical estimate of volume is 40 L with RSE 25%. Construct a two-sided approximate 90% raw-scale interval.
The standard error is \(40\times0.25=10\) L. With \(z_{0.95}=1.645\), the margin of error is \(1.645\times10=16.45\) L. Thus:
Under the normal-approximation assumptions, this procedure has approximately 90% coverage in repeated comparable analyses. The 90% interval is narrower than a 95% interval would be because it uses a smaller critical value; it is not evidence that the estimate is more precise.
When a Raw-Scale Interval Is Unsuitable
A symmetric Wald interval can extend below zero even when the parameter must be positive. That is a warning that a normal approximation on the raw scale may be poor, especially when the estimate is close to a boundary or its uncertainty is large relative to its value. A negative endpoint is not a meaningful negative clearance or volume, and it should not be interpreted as one.
For example, if an estimate is \(0.08\ \mathrm{h}^{-1}\) with RSE 80%, its standard error is \(0.08\times0.80=0.064\ \mathrm{h}^{-1}\). A raw-scale 95% interval would be:
The lower endpoint conflicts with the parameter's positive domain. Simply replacing it with zero does not preserve the original interval's stated coverage. The calculation signals that a symmetric interval on this scale should not be relied on without further justification.
Worked Example: Diagnosing a High-RSE Interval
Consider the hypothetical rate estimate above, \(0.08\ \mathrm{h}^{-1}\), with RSE 80%. The arithmetic produces a lower endpoint below zero. What should be concluded, and can the RSE alone be converted into a dependable interval on a transformed scale?
Calculation and diagnosis. The standard error is \(0.064\ \mathrm{h}^{-1}\), and the approximate raw-scale 95% interval is \([-0.04544,\ 0.20544]\ \mathrm{h}^{-1}\). Since a negative rate is outside the parameter's domain, the result exposes a possible mismatch between the normal approximation and the parameter scale.
What the calculation does not establish. It does not show that the true rate is negative, nor does it provide a valid positive interval after truncation. Nor does it establish the standard error of the log-rate. For small relative uncertainty, the first-order approximation \(\mathrm{SE}(\log\hat{\theta})\approx\mathrm{SE}(\hat{\theta})/\hat{\theta}\) may be useful, but at 80% RSE that approximation should not be treated as a dependable correction.
Next step. Inspect the fitted model and its uncertainty estimates. If the parameter is estimated on a log scale, an interval based on the estimated log-scale uncertainty can be transformed back by exponentiation, provided the log-scale approximation is justified. Otherwise, a profile-likelihood or suitable bootstrap interval may better reflect asymmetry. Those approaches still require checking their own assumptions and performance.
Alternatives, Diagnostics, and Limitations
The raw Wald interval is convenient because it needs only an estimate and its standard error. Alternatives are useful when the parameter scale or likelihood makes symmetric raw-scale uncertainty doubtful. They do not automatically solve problems such as weak identification or inadequate data.
| Approach | What it does | Important limitation |
|---|---|---|
| Raw-scale Wald interval | Adds and subtracts a normal critical value times the raw-scale standard error. | May be asymmetric uncertainty poorly represented, or may extend beyond a parameter boundary. |
| Transformed-scale interval | Forms an interval on a scale such as the log scale, then transforms its endpoints back. | Requires appropriate uncertainty on the transformed scale and a defensible approximation there. |
| Profile-likelihood interval | Uses the likelihood over candidate values of the parameter to identify a supported range. | Depends on a suitable likelihood and reliable optimization; it can still be uninformative with weak identification. |
| Bootstrap interval | Uses estimates from repeated resampling or simulation under a specified scheme. | Can perform poorly if the resampling scheme is inappropriate or the model repeatedly fails to fit. |
Before reporting an RSE-based interval, check that the RSE is finite and corresponds to the reported estimate and scale. Then examine whether the interval respects the parameter's domain, whether the estimate is near a boundary, and whether the model fit and uncertainty calculations are trustworthy. A non-finite or unavailable RSE cannot be turned into this interval. A very wide interval, boundary endpoint, or unstable fit calls for investigation rather than a more confident interpretation of the Wald calculation.
The main pitfall is treating an interval formula as a repair for every uncertainty problem. For example, “the parameter must be positive, so I will set a negative lower endpoint to zero” changes the procedure without establishing its coverage. The correction is to recognize that the raw-scale normal approximation may be unsuitable, and use a scale or interval method supported by the model and data.
Check Your Understanding
Work through these questions, then compare your reasoning with the solutions.
- A hypothetical estimate of clearance is 8 L/h with RSE 10%. Calculate its standard error and approximate 95% raw-scale interval.
- A hypothetical parameter estimate is 12 units with standard error 3 units. Calculate an approximate two-sided 90% interval using \(z=1.645\), and state its units.
- A positive parameter has estimate 0.05 units and RSE 70%. Calculate the approximate 95% raw-scale interval and explain the diagnostic significance of its lower endpoint.
- What does the 95% confidence level mean for an RSE-based interval, and what does it not mean?
Solutions. (1) The standard error is \(8\times0.10=0.80\) L/h. The interval is \(8\pm1.96(0.80)=8\pm1.568=[6.432,\ 9.568]\) L/h. (2) The margin is \(1.645\times3=4.935\) units, so the interval is \(12\pm4.935=[7.065,\ 16.935]\) units. (3) The standard error is \(0.05\times0.70=0.035\) units. The interval is \(0.05\pm1.96(0.035)=0.05\pm0.0686=[-0.0186,\ 0.1186]\) units. The negative endpoint is outside the parameter's positive domain and warns that the raw-scale normal approximation may be unsuitable; setting it to zero does not justify the original coverage. (4) If the assumptions hold, the interval-producing procedure would cover the fixed parameter in approximately 95% of repeated comparable analyses. It does not mean there is a 95% probability that the fixed parameter lies in this particular observed interval.