Reading a Precision Table as a Whole
A population PK precision table compresses several kinds of information into a few columns. The estimate describes the fitted parameter value; its standard error (SE) summarizes estimated sampling uncertainty on the parameter’s scale; a relative standard error (RSE) expresses that uncertainty relative to the estimate. Correlations between parameter estimates describe how their estimation errors vary together. A condition number summarizes a feature of a matrix used in estimation and can flag numerical sensitivity or weakly distinguished parameter directions.
These entries answer related but different questions. A small RSE may indicate relatively precise estimation of one parameter, while a large absolute SE may simply reflect that the parameter is measured on a large scale. Correlations describe joint uncertainty, not whether parameters are biologically associated in the population. A condition number is not an uncertainty measure for any single parameter. A defensible interpretation keeps these roles separate before bringing them together.
A standard-error table usually relies on a local approximation to the sampling uncertainty around the fitted values, commonly based on an estimated covariance matrix. Under that approximation, the diagonal elements of the covariance matrix are estimated variances and their square roots are SEs. Correlations are obtained by dividing a covariance by the product of the corresponding SEs. These summaries are most straightforward when the model is adequately fitted, the covariance calculation is reliable, and the local approximation is informative. A table alone cannot verify those conditions.
Check the Scale and the Arithmetic
RSE makes the SE interpretable relative to the estimated magnitude. For a positive parameter such as clearance, volume, or an absorption rate constant, its calculation is:
For a parameter that may be negative, the absolute value in the denominator gives a nonnegative relative magnitude, but the result may not be a useful summary near zero. RSE is also not a percentage confidence interval and does not by itself give a probability that a parameter lies within a stated range. Always retain the parameter’s units when discussing its estimate and SE; RSE is unitless.
Worked Example: Checking RSEs for PK Parameters
Suppose a hypothetical model reports clearance \(CL=5.20\ \mathrm{L/h}\) with \(SE=0.78\ \mathrm{L/h}\), volume \(V=42.0\ \mathrm{L}\) with \(SE=10.5\ \mathrm{L}\), and an absorption rate constant \(k_a=1.10\ \mathrm{h}^{-1}\) with \(SE=0.44\ \mathrm{h}^{-1}\). Check the RSEs and compare relative precision.
For clearance, \(100(0.78/5.20)=100(0.15)=15.0\%\). For volume, \(100(10.5/42.0)=100(0.25)=25.0\%\). For the absorption rate constant, \(100(0.44/1.10)=100(0.40)=40.0\%\).
The table’s relative uncertainty is smallest for clearance and largest for \(k_a\). That comparison is about relative, not absolute, uncertainty. In their own units, the SEs are \(0.78\ \mathrm{L/h}\), \(10.5\ \mathrm{L}\), and \(0.44\ \mathrm{h}^{-1}\); comparing those three numbers directly would not rank precision meaningfully. The RSE calculation is internally consistent with the supplied values.
A quick arithmetic check is useful when reviewing a table: recompute a few RSEs and confirm that the reported values agree after rounding. If an estimate is close to zero, a large or unstable RSE may result even when its SE is not especially large. If the parameter is constrained, such as being restricted to positive values, the estimate and its uncertainty may also be more naturally considered on a transformed scale. The table’s scale and the model’s parameterization therefore matter when interpreting a percentage.
Correlations Describe Joint Estimation Uncertainty
The estimated correlation between two parameter estimators ranges from \(-1\) to \(1\). A positive value indicates that, under the estimated joint uncertainty, upward deviations in one estimate tend to accompany upward deviations in the other. A negative value indicates a tendency for deviations in opposite directions. Values near either endpoint suggest strong dependence in the estimated uncertainty; values near zero suggest little linear dependence in that summary.
For estimates \(a\) and \(b\), the covariance and correlation are related by:
This relationship explains why a correlation matrix can make joint uncertainty easier to compare across parameters with different units. The covariance retains units equal to the product of the parameters’ units; the correlation is unitless. Neither quantity says that the biological values of \(a\) and \(b\) are correlated across individuals. That is a different model feature.
Worked Example: Interpreting a Clearance–Volume Correlation
In a hypothetical fit, \(SE(CL)=0.78\ \mathrm{L/h}\), \(SE(V)=10.5\ \mathrm{L}\), and the reported correlation between their estimates is \(-0.72\). Calculate the covariance and interpret the correlation.
Substitute the table entries into the covariance formula:
The negative correlation indicates substantial inverse dependence between the estimated clearance and volume errors in this approximation: fits with clearance estimates above their fitted values tend to have volume estimates below their fitted values, and conversely. It does not establish a negative biological relationship between clearance and volume among patients. The correlation is also not, by itself, proof that either parameter is poorly estimated; interpret it alongside the individual SEs and RSEs and other model diagnostics.
What a Condition Number Adds
A condition number is a property of a matrix, often an information matrix or a suitably scaled version of it. It summarizes how sensitive the matrix-based calculation may be to perturbations, or how different its strongest and weakest directions are. Its interpretation depends on the matrix definition and scaling: changing parameter units can change the condition number of an unscaled matrix. A reported value is therefore incomplete without knowing what matrix was used and how it was scaled.
A large condition number can flag that some parameter directions are difficult to distinguish numerically or are especially sensitive to perturbations. It does not identify which parameter is responsible, give an SE for a particular parameter, or supply a universal pass/fail threshold. Conversely, a moderate condition number does not establish that every parameter is precisely estimated. Read it alongside the table’s SEs, RSEs, and correlations, and alongside the fitting diagnostics.
The practical distinction is important: a condition number concerns matrix conditioning, while SEs and RSEs summarize estimated uncertainty. They can point in the same direction, but they are not interchangeable. A numerical difficulty can make a covariance calculation unreliable; a parameter may also be statistically uncertain even when the numerical calculation is stable. The earlier material on numerical ill-conditioning versus statistical uncertainty makes this distinction explicit.
Worked Example: Combining All Entries in a Precision Table
Consider this hypothetical summary from a fitted population PK model. The condition number is calculated from a scaled information matrix; the particular scaling convention is not specified in the table.
| Parameter or pair | Estimate / summary | SE | RSE or correlation |
|---|---|---|---|
| CL | 5.20 L/h | 0.78 L/h | 15.0% |
| V | 42.0 L | 10.5 L | 25.0% |
| \(k_a\) | 1.10 h\(^{-1}\) | 0.44 h\(^{-1}\) | 40.0% |
| CL and V estimates | — | — | −0.72 |
| CL and \(k_a\) estimates | — | — | 0.18 |
| V and \(k_a\) estimates | — | — | −0.31 |
| Scaled information matrix | Condition number 480 | — | — |
The reported RSEs match the estimates and SEs: \(100(0.78/5.20)=15.0\%\), \(100(10.5/42.0)=25.0\%\), and \(100(0.44/1.10)=40.0\%\). On this relative scale, clearance is estimated most precisely and \(k_a\) least precisely. The \(k_a\) RSE of 40.0% is a reason for caution, not a standalone decision rule about whether the estimate is usable.
The CL–V correlation of \(-0.72\) shows substantial inverse dependence in their estimated uncertainty. The other listed correlations are weaker in magnitude. The scaled condition number of 480 also merits attention as a potential sign of sensitivity, but it cannot be labeled acceptable or unacceptable without the matrix and scaling convention, the model context, and relevant diagnostics. It does not contradict or replace the parameter-specific RSEs.
A cautious summary would be: “In this hypothetical fit, clearance has the lowest relative uncertainty among the three listed parameters, whereas \(k_a\) has the highest RSE. The clearance and volume estimates show substantial negative correlation. The reported scaled information-matrix condition number warrants checking the scaling and numerical diagnostics; these summaries alone do not establish that the fit is invalid or that any parameter is unidentified.” This statement describes what the table supports and avoids claiming more than it can show.
Diagnostics, Alternatives, and Limits
Before drawing a conclusion from a precision table, check whether its entries are interpretable and coherent. Confirm that parameter units and scales are stated, that RSEs agree with the displayed estimates and SEs, that correlations lie between \(-1\) and \(1\), and that the condition number refers to a defined, appropriately scaled matrix. Review convergence and covariance-estimation diagnostics as well. A table can look complete even when the underlying approximation is unstable or does not represent the uncertainty well.
Warning signs include a very large RSE, an estimate close to a parameter boundary, a correlation close to \(1\) or \(-1\), or sensitivity of the condition number to scaling. None is an automatic verdict. For example, a high RSE for one parameter does not prove that all predictions are imprecise, and a high parameter correlation does not specify which parameter combination is well determined. Likewise, a modest RSE should not be treated as a guarantee that the model or estimate is correct.
The standard-error summary is compact and useful when its local covariance approximation is credible. Approximate confidence intervals can express uncertainty on the parameter scale, while profile-likelihood or bootstrap approaches can provide alternative assessments when local, symmetric uncertainty summaries are questionable. These alternatives have their own assumptions and computational limitations; they do not turn a questionable fit into a sound one. The table is best treated as one part of a precision assessment, not as a substitute for model evaluation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A parameter estimate is \(8.0\ \mathrm{L/h}\) and its SE is \(1.2\ \mathrm{L/h}\). What is its RSE, and what does that percentage describe?
- A pair of estimates has correlation \(0.85\). What does this say about their estimated uncertainty, and what does it not say?
- Why is a condition number of 480 insufficient, on its own, to classify a fit as acceptable or unacceptable?
- A table reports an estimate of \(0.10\) and an SE of \(0.08\) for a positive parameter. Calculate the RSE and give one reason to be cautious about interpreting it alone.
Solutions. (1) \(100(1.2/8.0)=15.0\%\). It describes the SE relative to the estimate, not a confidence level or a probability that the estimate is correct. (2) The positive correlation indicates that upward deviations in one estimate tend to accompany upward deviations in the other in the estimated joint uncertainty. It does not establish a positive biological correlation between individual-level parameter values. (3) Its meaning depends on the matrix and scaling convention, and there is no universal threshold that makes the number a verdict. It must be considered with other uncertainty and fitting diagnostics. (4) \(100(0.08/0.10)=80.0\%\). Because the estimate is relatively close to zero, the RSE can be large and may not by itself convey the parameter’s uncertainty adequately; the parameter scale, constraints, and other diagnostics matter.