Parameter Identifiability and Precision Review

Why Parameter Precision Matters

A fitted model can converge successfully and still leave important parameters poorly informed by the available data. This review asks a different question from whether the numerical algorithm found a solution: how much information in the data supports each parameter estimate, and what claims depend on that information? A parameter can be estimated imprecisely because the study provides little information about it, because the model structure allows parameters to compensate for one another, or because the likelihood is not well approximated by a symmetric curve near the estimate.

These distinctions matter in MIDD because a report may describe individual parameters, combinations of parameters, or predictions that depend on them. A parameter estimate should not be called “precise” merely because it is printed with several decimal places. Precision is about uncertainty, not display resolution.

Definition: A parameter is structurally identifiable if, under the specified model and ideal, noise-free observations over the design, distinct parameter values cannot produce exactly the same observable predictions. It is practically identifiable if the actual design and data provide enough information to estimate it with useful precision. A structurally identifiable parameter can be practically weakly identified when the data are sparse, noisy, or uninformative over the relevant range.

Identifiability is a property of the model together with what is observed; precision also reflects the data and estimation uncertainty. For example, if observations only determine a ratio of two parameters, the individual parameters may be difficult to separate even when the ratio is well estimated. Conversely, a model can be structurally identifiable in principle but have a nearly flat likelihood for one parameter in a particular study.

Standard Errors and Confidence Intervals

For an estimate \(\hat{\theta}\), the standard error (SE) summarizes the estimated sampling uncertainty of that estimate under the fitted model and the assumptions used to calculate it. When an estimate is approximately normally distributed and not constrained near a boundary, a Wald confidence interval can be formed as the estimate plus or minus a normal quantile times its SE. This approximation requires a locally symmetric likelihood and a reliable covariance estimate; it is not automatically appropriate for every parameter.

\hat{\theta} \pm z_{1-\alpha/2}SE(\hat{\theta})

For a 95% interval, \(z_{0.975}\) is approximately 1.96. Under the approximation, repeated intervals constructed in this way would contain the fixed parameter in about 95% of repetitions under the model and sampling assumptions. A particular interval is not a statement that there is a 95% probability the fixed parameter lies inside it. In a model report, interpret the interval in context and with the parameter’s units.

A useful descriptive quantity is the relative standard error, often reported as a percentage:

RSE(\hat{\theta}) = 100\frac{SE(\hat{\theta})}{|\hat{\theta}|}\%

Relative SE can help compare the scale of uncertainty across positive parameters, but it becomes misleading or undefined when the estimate is close to zero. It also does not replace an interval: two estimates with the same relative SE can have different practical implications because their units, constraints, and roles in the model differ.

Worked Example: Interpreting a Wald Interval for Clearance

In a hypothetical population pharmacokinetic analysis, clearance is estimated as \(5.00\) L/h with \(SE=0.80\) L/h. Suppose the estimate is interior to its permitted range, the covariance calculation is stable, and the likelihood near the estimate is approximately symmetric and quadratic. These assumptions make a Wald interval a reasonable local summary.

The 95% interval is:

5.00 \pm 1.96(0.80) =5.00 \pm 1.568 =[3.432,\ 6.568]\text{ L/h}.

The relative SE is \(100(0.80/5.00)=16.0\%\). The interval describes the uncertainty in clearance under the specified model and approximation. It does not establish that every value in the interval is equally plausible, nor does it show whether a decision based on clearance would change across that range; that depends on how the parameter enters the decision-relevant model output.

Parameter Correlations and Trade-Offs

The estimated covariance matrix describes uncertainty jointly across parameters. Its diagonal entries are estimated variances; off-diagonal entries describe how estimation errors tend to vary together. A correlation is the covariance divided by the product of the two standard errors, and ranges from \(-1\) to \(1\). A correlation near either extreme signals that the estimates are strongly linked locally: changing one parameter can be partly offset by changing another while preserving similar model predictions.

High correlation is a warning to investigate, not a proof that either parameter is unidentifiable. It is also important when calculating uncertainty for a derived quantity. Ignoring covariance can materially overstate or understate that uncertainty, depending on the signs of the derivatives and covariance.

Worked Example: Correlated Clearance and Volume Estimates

Suppose a hypothetical analysis estimates clearance \(CL=5.0\) L/h and volume \(V=50\) L, with \(SE(CL)=0.80\) L/h, \(SE(V)=12\) L, and estimated correlation \(-0.96\). The covariance is:

\operatorname{Cov}(CL,V) =(-0.96)(0.80)(12) =-9.216\text{ L}^2/\text{h}.

The estimates are strongly negatively correlated. Consider the derived elimination rate \(k=CL/V\), in h\(^{-1}\). Its estimate is \(5.0/50=0.100\) h\(^{-1}\). A first-order uncertainty approximation uses the gradient \((1/V,\ -CL/V^2)\) and the estimated covariance matrix:

\begin{aligned} \operatorname{Var}(\hat{k}) \approx {}& \left(\frac{1}{V}\right)^2\operatorname{Var}(CL) +\left(\frac{-CL}{V^2}\right)^2\operatorname{Var}(V)\\ &+2\left(\frac{1}{V}\right)\left(\frac{-CL}{V^2}\right) \operatorname{Cov}(CL,V). \end{aligned}

Substituting the estimates gives:

\begin{aligned} \operatorname{Var}(\hat{k}) \approx {}& (0.02)^2(0.64)+(-0.002)^2(144) +2(0.02)(-0.002)(-9.216)\\ ={}&0.000256+0.000576+0.00073728\\ ={}&0.00156928\text{ h}^{-2}. \end{aligned}

Thus \(SE(\hat{k})\approx\sqrt{0.00156928}=0.03961\) h\(^{-1}\). If covariance were omitted, the estimated variance would be \(0.000832\) h\(^{-2}\), giving an SE of \(0.02884\) h\(^{-1}\). Here omission understates uncertainty because the negative covariance combines with derivatives of opposite signs to give a positive covariance contribution. The example shows why a review should inspect joint uncertainty, especially when a claim concerns a function of correlated parameters.

Profile Likelihood and Non-Symmetric Uncertainty

A standard error usually summarizes local curvature at the optimum. It can be unreliable when the likelihood is asymmetric, flat in one direction, or affected by parameter constraints. A profile likelihood examines one parameter at a time while allowing the other parameters to take their best-fitting values for each fixed value of the parameter of interest.

Definition: The profile log-likelihood for parameter \(\theta\) is \(\ell_p(\theta)=\max_{\boldsymbol{\eta}}\ell(\theta,\boldsymbol{\eta})\), where \(\boldsymbol{\eta}\) denotes the other model parameters and \(\ell\) is the log-likelihood. Under regular likelihood assumptions, an approximate 95% profile-likelihood interval for one parameter contains the values satisfying \(2[\ell_p(\hat{\theta})-\ell_p(\theta)]\leq 3.84\), using the 95th percentile of a chi-square distribution with 1 degree of freedom.

The interval is found by locating where the profile falls 1.92 log-likelihood units below its maximum, because \(3.84/2=1.92\). This result relies on regularity conditions, including an interior parameter value and a suitable large-sample likelihood approximation. Profile shape is itself informative: a broad, shallow profile indicates weak information; a strongly asymmetric profile warns that a symmetric Wald interval may be misleading; and a profile that does not cross the chosen cutoff on one side may yield only a one-sided bound over the explored range.

Worked Example: An Asymmetric Profile Interval

In a hypothetical analysis, a positive parameter \(\theta\) has estimate \(\hat{\theta}=1.00\), and the maximized profile log-likelihood is \(-120.00\). Assume \(\theta\) is interior to its allowed range and the regular likelihood approximation is considered suitable. The 95% profile cutoff is:

-120.00-\frac{3.84}{2}=-121.92.

Suppose the profile crosses \(-121.92\) at \(\theta=0.55\) and \(\theta=2.40\). The approximate profile interval is therefore \([0.55,\ 2.40]\), which is notably asymmetric around the estimate. If a local standard error of \(0.35\) were used for a Wald interval, it would give \(1.00\pm1.96(0.35)=[0.314,\ 1.686]\). The two methods differ because the profile is not well represented by a symmetric quadratic over the interval. The report should show and explain the profile-based uncertainty rather than presenting the Wald interval as definitive.

Boundary Estimates and Review Diagnostics

Some parameters are constrained, such as a variance that cannot be negative or a fraction restricted to \([0,1]\). If an estimate lies at a boundary, a symmetric Wald interval can extend into impossible values and the usual normal approximation may fail. At a boundary, regular likelihood-ratio reference distributions may also change. A reviewer should therefore ask how uncertainty was obtained and whether the method accounts for the constraint; a standard error alone may not answer that question.

Worked Example: A Variance Estimate at Zero

Suppose a hypothetical population model estimates an interindividual variance \(\omega^2\) as \(0\), its permitted lower boundary. A symmetric Wald interval based on an ordinary SE would be centered at zero and extend below zero, which is not a valid range for a variance. It should not be interpreted as a valid confidence interval.

Instead, suppose a parametric bootstrap calibrated a one-sided 95% profile cutoff of \(2\Delta\ell=2.71\) for this particular analysis. This is a hypothetical, analysis-specific calibration, not a universal cutoff. The profile reaches that cutoff at \(\omega^2=0.16\) (variance units). The corresponding upper bound on the standard deviation is \(\sqrt{0.16}=0.40\) in the standard-deviation units. The result supports reporting an estimate at the boundary with an upper uncertainty bound; it does not prove that the true variance is exactly zero. The calibration method and boundary constraint should be disclosed.

A practical review combines numerical summaries with the shape and behavior of the uncertainty calculation. Warning signs include very large relative SEs, confidence limits that are implausible or incompatible with parameter constraints, correlations close to \(\pm1\), profiles that remain nearly flat, and large differences between Wald and profile intervals. These findings call for checking the estimation output, parameterization, and information in the design. They are not, by themselves, evidence of a particular model defect.

Parameter precision should also be linked to the claim. A weakly informed parameter does not automatically make every prediction imprecise: a prediction may depend mainly on a well-estimated combination of parameters. Conversely, a prediction or dosing claim that changes substantially over plausible values of a weak parameter needs qualification. The review should identify which parameters are weakly informed, what model outputs depend on them, and whether the report’s conclusions remain supported by the uncertainty presented.

Choosing and Comparing Uncertainty Summaries

Approach	Useful when	Main caution
Wald SE and interval	The estimate is interior and the likelihood is locally symmetric and well approximated by a quadratic.	Can mislead with asymmetry, constraints, or weak information; depends on a reliable covariance estimate.
Profile likelihood	Uncertainty may be asymmetric, or trade-offs with other parameters need examination.	Requires repeated optimization and appropriate likelihood reference assumptions; numerical failures can obscure the profile.
Bootstrap-calibrated uncertainty	Regular approximations are questionable, including some boundary settings, and a defensible resampling design is available.	Depends on the bootstrap model and implementation; a calibration is not automatically valid for every parameter or study.

These approaches answer related but not identical questions. A profile can expose likelihood shape that a single SE hides, while a bootstrap can calibrate uncertainty under a specified repeated-sampling procedure. Neither repairs a model or design that contains little information about a parameter. Reviewers should ask whether the method’s assumptions match the estimate, parameter constraints, and analysis, rather than selecting a method solely because it gives a narrower interval.

Check the parameter scale and constraints.
Confirm units, permitted range, and whether the estimate is near a boundary before interpreting a symmetric interval.

Read uncertainty jointly.
Review SEs or intervals together with parameter correlations and, where relevant, uncertainty in derived quantities.

Inspect shape and assumptions.
Use profile or calibrated alternatives when local symmetry or regularity is doubtful, and disclose the method and limitations.

Trace uncertainty to the claim.
State whether the decision-relevant conclusion depends on weakly informed parameters and qualify the claim accordingly.

Common Pitfall

Wrong reasoning: “The model converged, and the parameter estimate has three decimal places, so it is precisely estimated.” Convergence indicates that the fitting procedure reached a solution under its numerical criteria; decimal places indicate formatting. Neither establishes that the data distinguish this parameter from alternatives.

Correction: Evaluate uncertainty using SEs and intervals only when their assumptions are credible, inspect correlations and profile shape for weak information or trade-offs, and check boundaries. Then determine whether the claimed interpretation or decision is robust to the uncertainty in the parameters it uses. Precision and convergence are separate review questions.

Key takeaway: A parameter estimate is informative only when its uncertainty and the model’s information structure support the intended interpretation. Review the estimate, its interval or profile, its relationships with other parameters, any constraints, and the claims that rely on it.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A parameter estimate is \(2.0\) units with \(SE=0.5\) units. Under a suitable Wald approximation, calculate its 95% interval and relative SE.
Why can a high correlation between two parameter estimates be important even if each estimate has a reported SE?
A profile log-likelihood maximum is \(-50.00\). For a regular one-parameter 95% profile interval, what log-likelihood cutoff is used?
A nonnegative variance parameter is estimated as zero. Why is an ordinary symmetric Wald interval problematic, and what should a reviewer look for instead?

Solutions. (1) The interval is \(2.0\pm1.96(0.5)=[1.02,\ 2.98]\) units; the relative SE is \(100(0.5/2.0)=25.0\%\). (2) The estimates may trade off, and their covariance affects uncertainty in combinations or predictions involving both; separate SEs do not describe the full joint uncertainty. (3) The cutoff is \(-50.00-3.84/2=-51.92\). The interval includes parameter values whose profile log-likelihood is at least \(-51.92\), subject to the regularity assumptions. (4) A symmetric interval extends below the allowed value zero, and boundary estimates may not have the usual normal or likelihood-ratio behavior. Review a constraint-aware interval or boundary-calibrated profile or bootstrap method, including how it was calibrated.