Why Precision Summaries Need to Be Read Together
A parameter estimate can look precise when viewed alone and still belong to a poorly determined combination of parameters. In a population PK model, two parameters may compensate for one another: moving both along a particular direction changes the fitted concentrations very little. The data then contain limited information about that direction, even if each reported estimate is finite and an optimizer has returned a solution.
This tutorial focuses on recognizing that pattern from standard errors, parameter correlations, and the stability of estimates across valid fits. These summaries are clues, not stand-alone tests of model adequacy. They describe uncertainty conditional on the model and data being analyzed; they do not establish that the model is correct or that every parameter is identifiable in all possible designs.
A standard error (SE) summarizes local uncertainty in an estimate. A relative standard error (RSE) expresses that uncertainty relative to the estimate, commonly as \(100\,SE/|\hat{\theta}|\) percent for an estimate \(\hat{\theta}\) on its reported scale. Parameter correlations describe how estimation errors in pairs of parameters tend to move together, as represented by the estimated covariance matrix. These quantities are most directly interpretable when the likelihood is smooth and approximately quadratic near an interior optimum and the covariance estimate is reliable.
When those conditions hold, the covariance matrix approximates the sampling covariance of the parameter estimates. A large SE can indicate limited information about one parameter. A correlation near \(+1\) or \(-1\) can indicate that two estimates move together or in opposite directions, leaving a linear combination much less certain than another. Instability across fits to the same data can provide another warning. No one pattern, by itself, identifies the cause.
What the Covariance Summary Can Reveal
For two estimates \(\hat{\theta}_1\) and \(\hat{\theta}_2\), let their standard errors be \(SE_1\) and \(SE_2\), and let their estimated correlation be \(r\). Their estimated covariance is \(rSE_1SE_2\). For a linear combination, the covariance determines how uncertainty adds or cancels:
If \(r\) is strongly positive, the sum tends to be less precise and the difference more precise. If \(r\) is strongly negative, the reverse tends to hold. This is one reason that individual SEs do not fully describe uncertainty in a multi-parameter model. The relevant combination depends on the correlation pattern and the model.
Worked Example: Finding the Less Certain Combination
Suppose a hypothetical PK analysis reports standard errors of \(0.20\) for \(\theta_1=\log(CL/(5\ \mathrm{L/h}))\) and \(0.30\) for \(\theta_2=\log(V/(50\ \mathrm{L}))\), with estimated correlation \(r=0.90\). The dimensionless log parameters are used here so that the comparison concerns relative changes in clearance and volume.
First calculate the covariance:
The variance of the sum is \(0.20^2+0.30^2+2(0.054)=0.04+0.09+0.108=0.238\). Its standard error is \(\sqrt{0.238}\approx0.488\). The variance of the difference is \(0.20^2+0.30^2-2(0.054)=0.04+0.09-0.108=0.022\), giving a standard error of \(\sqrt{0.022}\approx0.148\).
In this local approximation, the sum is much less precisely estimated than the difference. This is consistent with a direction in which the two log parameters increase together and the likelihood changes relatively little. It does not prove that the model is weakly identified: the approximation depends on the covariance estimate, and the full likelihood and fitting behavior should also be examined.
A correlation close to either \(+1\) or \(-1\) is a warning of strong dependence in the estimates, not a verdict. Its sign matters, and its interpretation depends on which parameters and scales were used. A high correlation may arise because two parameters genuinely trade off in the observed model, but it can also be affected by scaling, the parameterization, or an unreliable local covariance estimate. Inspect the SEs, the scientific meaning of the parameter combination, and the likelihood or fit stability before drawing a conclusion.
Look for Instability Across Equivalent Fits
A practical check is to fit the same model to the same data from multiple dispersed starting values. If several fits meet the same convergence criteria and have nearly identical objective values, yet return noticeably different parameter estimates, the data may not distinguish those estimates well. This evidence is stronger when the shifts follow a coherent pattern, such as two estimates changing together while the objective barely changes.
This check requires care. Different estimates do not indicate weak identification if some runs failed to converge, stopped at different local optima, used different data or model settings, or violated parameter constraints. Record the objective value, convergence diagnostics, and parameter estimates for each run. Similar objective values are not proof that the fits are equivalent in every respect, but they help distinguish a shallow range of solutions from a clearly inferior fit.
Worked Example: Estimates Shift While the Fit Barely Changes
Consider a hypothetical model fitted to the same sparse concentration data from three starting values. All three runs satisfy the chosen convergence checks. Their objective values are \(240.00\), \(240.06\), and \(240.04\), respectively; lower values indicate better fit. The estimated clearance and volume are:
| Fit | CL (L/h) | V (L) | Objective value |
|---|---|---|---|
| A | 5.0 | 50 | 240.00 |
| B | 6.4 | 64 | 240.06 |
| C | 4.2 | 42 | 240.04 |
The clearance range is \(6.4-4.2=2.2\ \mathrm{L/h}\), and the volume range is \(64-42=22\ \mathrm{L}\). Relative to the respective means, these ranges are \(2.2/5.2\approx42.3\%\) for clearance and \(22/52\approx42.3\%\) for volume. Meanwhile, the objective values span only \(240.06-240.00=0.06\). Suppose the estimated correlation between clearance and volume is \(+0.98\).
The estimates move together, the correlation is strongly positive, and the fits have very similar objective values. Taken together, these findings suggest that this data set may constrain the clearance-to-volume relationship more strongly than the two individual values. They are evidence consistent with weak identification of a parameter combination, not proof: the modeler should verify the fits, inspect the likelihood behavior, and assess whether the differences matter for the intended interpretation or predictions.
A Large RSE Is a Flag, Not a Diagnosis
RSE is useful for comparing uncertainty with the magnitude of an estimate, but it compresses a multivariate problem into one number. A high RSE can flag a parameter that is poorly estimated relative to its value. On its own, however, it cannot show whether that uncertainty comes from a trade-off with another parameter, a boundary or skewed likelihood, a limited design, or an unsuitable local approximation. Conversely, moderate individual RSEs do not rule out a poorly determined combination.
Consider an estimated interindividual variability SD \(\omega=0.12\) with \(SE=0.09\). Its RSE is \(100(0.09/0.12)=75\%\). Suppose clearance and volume have RSEs of 9% and 10%, and the estimated correlations of \(\omega\) with each are \(-0.12\) and \(0.08\). The large RSE for \(\omega\) deserves attention, but this summary does not show a strong pairwise trade-off involving \(\omega\). Because variability SDs are constrained to be nonnegative and can have asymmetric uncertainty, a symmetric local SE-based summary may also be a poor description near the boundary. The next example shows why the RSE should prompt follow-up rather than serve as the whole diagnosis.
Worked Example: Interpreting One High RSE in Context
For the hypothetical estimates above, \(RSE_\omega=100(0.09/0.12)=75\%\). The estimates of clearance and volume are \(5.0\ \mathrm{L/h}\) with \(SE=0.45\ \mathrm{L/h}\), and \(50\ \mathrm{L}\) with \(SE=5.0\ \mathrm{L}\). Their RSEs are \(100(0.45/5.0)=9.0\%\) and \(100(5.0/50)=10.0\%\).
The \(\omega\) estimate has substantially greater relative uncertainty than the two typical PK parameter estimates. That is a finding about its precision, not by itself evidence of a multi-parameter ridge. Suppose further investigation finds that fits from multiple starting values give \(\omega\) estimates near \(0.12\), while a likelihood-based interval is asymmetric, extending from \(0.02\) to \(0.24\). This interval indicates substantial uncertainty and asymmetry; the repeated estimates and modest reported correlations do not point to the same strong pairwise trade-off as in the preceding example.
The appropriate conclusion is not that \(\omega\) is known precisely, nor that it is definitively well identified. Rather, its large RSE is a warning that warrants methods suited to its boundary-constrained, potentially asymmetric uncertainty. A single RSE cannot reveal the shape of that uncertainty or establish its cause.
Diagnostic Workflow and Common Failure Modes
Use the summaries as complementary checks. The estimated covariance matrix describes local pairwise uncertainty; fit stability asks whether the optimizer returns similar solutions; and likelihood-based or resampling checks can assess uncertainty beyond a local quadratic approximation. Each has limitations. For example, repeated fits from similar starting values may all miss a broad region of solutions, while a covariance matrix based on a poorly behaved Hessian may give misleading SEs and correlations.
Confirm that runs use the same model and data, meet convergence criteria, and have comparable objective values. Exclude failed or clearly inferior fits from an identification argument.
Review SEs or RSEs together with the parameter correlation matrix. Ask whether large SEs or strong correlations suggest a meaningful trade-off, and identify which combinations could be affected.
Repeat fits from dispersed starting values. Where local approximations are questionable, use an appropriate likelihood-based or resampling assessment and examine whether estimates or conclusions change materially.
Separate evidence of relative imprecision, evidence of a parameter trade-off, and evidence that the overall model or numerical fit is unreliable. Do not infer more than the checks establish.
The condition number and RSE answer different questions. RSE is parameter-specific and relative to the estimate; a condition number summarizes the spread of information across directions and can be sensitive to parameter scaling. A high RSE for one parameter does not automatically imply a problematic condition number, and a problematic condition number does not identify which parameter interpretation is most uncertain without examining the corresponding parameter combinations. Neither replaces fit diagnostics or a scientifically meaningful assessment of uncertainty.
Common failures include treating a correlation near \(\pm1\) as proof of non-identifiability, interpreting a high RSE as a universal cutoff, and trusting SEs when the likelihood is markedly asymmetric or a parameter is near a constraint. Another error is to call estimates unstable when the compared fits have not converged to comparable solutions. In each case, the correction is to check the conditions behind the summary, examine multiple sources of evidence, and report what remains uncertain.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Two log-scale estimates have SEs of \(0.10\) and \(0.20\), with correlation \(-0.95\). Which of their sum or difference is expected to be less precise locally?
- Several converged fits to the same data have nearly identical objective values but substantially different estimates of two positively correlated parameters. What pattern does this suggest, and what should be checked before concluding weak identification?
- A parameter has an RSE of 90%, but its estimated correlations with other parameters are small. Does the RSE alone establish a multi-parameter identification problem?
- Why can a covariance-based SE be misleading for a parameter near a boundary?
- How does an RSE differ from a condition number as a precision diagnostic?
Solutions. (1) With negative correlation, the covariance term reduces the variance of the sum and increases the variance of the difference, so the difference is expected to be less precise. (2) The pattern is consistent with a weakly determined combination or trade-off. First confirm that the runs used identical data and settings, met convergence criteria, and reached comparable fits; then assess whether the shifts are meaningful. (3) No. The RSE flags substantial relative uncertainty in that parameter, but alone does not establish a multi-parameter trade-off or identify the cause. (4) Near a boundary, the likelihood may be asymmetric or truncated, whereas a covariance-based SE relies on a local approximation that is often most interpretable for a smooth, approximately quadratic likelihood at an interior estimate. (5) RSE compares one parameter's SE with its estimate; a condition number summarizes information across directions and can depend on parameter scaling. Neither diagnostic alone gives a complete account of identification.