Why Reporting Requires More Than a Point Estimate
A population pharmacokinetic estimate is most useful when a reviewer can identify what it represents, its scale and units, and how precisely it was estimated. A clearance estimate of 5.2 is incomplete without units and without saying whether it is a typical value, a population mean, or an individual estimate. A variability estimate is also incomplete unless its parameterization is stated. Finally, a precise-looking point estimate does not show whether the data support it reliably.
Reporting should separate two kinds of variation. Between-individual variability describes how individual parameters vary around the population model. Parameter uncertainty describes how much the estimated population parameters might vary across repeated data sets under the assumed model and estimation procedure. These quantities answer different questions and must not be presented as if they were interchangeable.
What to State Alongside Each Estimate
For each reported parameter, give its name, meaning, estimate, and units where applicable. State the parameterization: for example, whether a typical value is defined on the original scale or log scale, and whether a variability term is a variance, standard deviation, or coefficient of variation (CV). Include a measure of precision and, when appropriate, an uncertainty interval. A table should also identify the estimation method and the data or analysis set to which the estimates apply; those details allow the numbers to be interpreted in the context of the analysis.
In a common log-normal individual-parameter model, a positive parameter \(P_i\) for individual \(i\) is represented as \(P_i = \theta \exp(\eta_i)\), where \(\theta\) is the typical value and \(\eta_i \sim N(0,\omega^2)\). Here, \(\omega^2\) is the variance of the random effect on the log scale; it is not a percentage. The corresponding between-individual CV, expressed as a percentage, is:
The parameterization should be explicit because some reports give \(\omega^2\), some give \(\omega\), and some give the derived CV. These are not numerically interchangeable. When random effects for two parameters are correlated, report the estimated correlation or covariance as appropriate, and say which one is shown. A covariance depends on the scales of both random effects, whereas a correlation is dimensionless.
Worked Example: Reporting Log-Normal Between-Individual Variability
Suppose a hypothetical final model estimates the typical clearance as 5.2 L/h and the log-scale random-effect variance for clearance as \(\omega^2=0.16\). Calculate the random-effect standard deviation and the corresponding CV.
The log-scale standard deviation is \(\omega=\sqrt{0.16}=0.400\). The CV is calculated from the log-normal parameterization:
A clear report could state: “Typical clearance was 5.2 L/h, with an estimated log-scale random-effect variance of 0.16 (standard deviation 0.400), corresponding to a between-individual CV of 41.7% under the log-normal model.” The CV describes modeled individual variation; it is not a 41.7% uncertainty interval for the estimated typical clearance.
Residual Error Needs Its Own Scale
Residual error describes remaining differences between observations and model predictions under the specified observation model. For example, a combined additive and proportional model can be written as \(C_{\mathrm{obs}}=C_{\mathrm{pred}}(1+\epsilon_{\mathrm{prop}})+\epsilon_{\mathrm{add}}\), with independent, mean-zero error terms. In this parameterization, the proportional error is dimensionless, while the additive error has the units of concentration. A report should say whether it presents error variances or standard deviations and give units for the additive component.
This distinction matters because “residual error = 20%” could refer to a proportional standard deviation, a variance transformed to a percentage, or another convention. State the model and the reported scale rather than relying on a label alone. If the model estimates both components, give both and retain their different units.
Worked Example: Presenting a Combined Residual Error Model
In a hypothetical concentration model, suppose the proportional residual-error variance is estimated as \(0.0225\), and the additive residual-error standard deviation is estimated as 0.08 mg/L. Report the terms on interpretable scales.
The proportional standard deviation is \(\sqrt{0.0225}=0.150\), or 15.0% on the proportional scale. The additive term is already a standard deviation and remains 0.08 mg/L. A suitable description is: “The combined residual-error model had a proportional error standard deviation of 0.150 (15.0%) and an additive error standard deviation of 0.08 mg/L.” The units differ because the proportional term scales with predicted concentration and the additive term is on the concentration scale.
Precision Measures and Uncertainty Intervals
For an estimator \(\hat{\theta}\), a model-based SE is commonly obtained from an estimated covariance matrix of the parameter estimates. The SE for a parameter is the square root of its estimated variance, given by the corresponding diagonal element of that matrix. The calculation relies on the chosen estimation method and its approximation to the estimator’s sampling distribution; it is conditional on the model and data analysis. Report the method used to obtain uncertainty, rather than presenting an SE as method-free.
For a positive parameter whose estimate is meaningfully separated from zero, the relative standard error (RSE) is often reported as \(100\times SE/\hat{\theta}\). RSE expresses the SE relative to the point estimate; it is a precision summary, not an estimate of between-individual variability. It can be unhelpful when the estimate is near zero, when the parameter can take negative values, or when its sampling distribution is strongly asymmetric.
An approximate two-sided 95% interval based on a normal approximation is \(\hat{\theta}\pm1.96SE\). This interval is justified when the estimator’s sampling distribution is approximately normal and the approximation is adequate for the parameter scale. It can be inappropriate for a positive parameter if the lower endpoint is negative, or for a variance parameter at a boundary. For a positive parameter estimated on the log scale, an interval can instead be formed on that scale and transformed back. The interval’s coverage interpretation concerns repeated use of the procedure under its assumptions; it does not mean there is a 95% probability that a fixed parameter lies in the particular computed interval.
Worked Example: Calculating an SE, RSE, and Approximate Interval
Suppose a hypothetical model estimates typical clearance as 5.20 L/h with model-based SE 0.520 L/h. Calculate the RSE and a normal-approximation 95% interval, assuming the estimator is approximately normal on the reported scale.
The RSE is \(100(0.520/5.20)=10.0\%\). The interval is:
Rounded to three significant figures, the approximate interval is 4.18 to 6.22 L/h. It gives a range of values compatible with the estimate under the stated normal approximation and uncertainty calculation. It is not a range containing 95% of individual patients’ clearances; that would concern between-individual variability.
Worked Example: Reporting an Asymmetric Interval for a Positive Estimate
Suppose a hypothetical positive parameter has an estimate of 5.20 L/h and an SE of 0.080 on the log scale of the parameter. Assume a normal approximation is reasonable on that scale. Calculate the interval on the log scale and transform it to L/h.
The estimate on the log scale is \(\log(5.20)\). Form the interval there, then exponentiate both endpoints:
The transformed interval is asymmetric around 5.20 L/h, as expected when a symmetric interval on the log scale is returned to the original scale. The report should identify that the interval was constructed on the log scale. This approach respects positivity, but it is justified only if the log-scale approximation and the reported SE are appropriate.
Choosing and Presenting an Uncertainty Method
A normal-approximation interval from the estimated covariance matrix is compact and useful when the likelihood or estimator is well behaved and the local approximation is credible. A bootstrap interval uses repeated resampling and refitting to approximate sampling variability; the resampling scheme, number of successful fits, and interval construction should be stated. Bootstrap results can be informative when the sampling distribution is asymmetric, but they are not automatically reliable: resamples may fail to fit, and the resampling design must reflect the independent sampling units.
For either method, report uncertainty for parameters that matter to interpretation, including variability and residual-error terms where feasible. Do not imply that an SE or interval fully captures model uncertainty: it usually reflects uncertainty conditional on the chosen model, analysis data, and estimation procedure. A concise table can make the reporting consistent.
| Quantity | Example report | What the reader should be able to identify |
|---|---|---|
| Typical clearance | 5.20 L/h; SE 0.520 L/h; RSE 10.0%; approximate 95% interval 4.18–6.22 L/h | Parameter meaning, units, estimate, and precision |
| Clearance variability | \(\omega^2=0.160\) on the log scale; CV 41.7% | Variance parameterization and derived variability scale |
| Residual error | Proportional SD 0.150; additive SD 0.08 mg/L | Error model, scale, and units for each component |
Diagnostics, Failure Modes, and Limitations
Check that the reported values agree with the final model output and with each other. Verify units, parameter labels, transformations, and whether each variability term is a variance, standard deviation, or CV. If a table gives an estimate, SE, RSE, and interval, recompute the RSE and check that the interval is consistent with the stated method. Small discrepancies can result from rounding, but material inconsistencies make review difficult.
A very large SE or RSE, a very wide interval, an interval extending outside the parameter’s meaningful range, or a variability estimate near a boundary may indicate weak information or a poor local approximation. These are prompts to qualify the precision and inspect the estimation results; they are not, by themselves, proof that the model is wrong. Likewise, a small SE does not demonstrate that the model structure is correct. Uncertainty summaries cannot replace transparent reporting of the parameterization and estimation procedure.
The main alternatives are complementary rather than interchangeable: an SE or RSE is a compact precision summary; a model-based interval provides a range on a specified scale; and a bootstrap interval can reflect asymmetry that a normal approximation misses. Each remains conditional on its assumptions and analysis setup. Clear reporting names the chosen method and avoids suggesting that any one statistic summarizes every source of uncertainty.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A log-normal random-effect variance is 0.09. Calculate \(\omega\) and the corresponding CV percentage.
- A typical volume estimate is 30 L with SE 3 L. Calculate its RSE. Does that RSE describe between-individual variability?
- Why should a report identify whether a residual-error quantity is a variance or a standard deviation?
- A normal-approximation interval for a positive parameter extends below zero. What should the analyst check before reporting it as a meaningful interval?
Solutions.
1. \(\omega=\sqrt{0.09}=0.300\). The CV is \(100\sqrt{\exp(0.09)-1}=100\sqrt{1.0942-1}\approx30.7\%\). This is modeled between-individual variation under the log-normal parameterization.
2. The RSE is \(100(3/30)=10.0\%\). No. It summarizes the SE relative to the estimated typical volume, not variation among individuals.
3. A standard deviation is the square root of a variance, so the values differ numerically and their units differ: a variance has squared units, while a standard deviation has the original scale’s units. The report must also clarify whether a proportional term is dimensionless and whether an additive term has concentration units.
4. Check whether the estimator is approximately normal on the reported scale and whether the parameter is constrained to be positive. If positivity and skewness make the normal approximation unsuitable, consider a justified transformed-scale or bootstrap interval, and state how it was constructed. Do not silently truncate the interval at zero.