Why Identifiable Parameters Can Still Be Hard to Estimate
A compartment model may have a unique set of parameters in principle, yet concentration measurements from a particular study may not determine those parameters precisely. The distinction matters: structural identifiability asks whether ideal, exact observations could distinguish parameter values, while practical identifiability asks whether the actual sampling schedule and measurement precision provide enough information to estimate them reliably. Structural identifiability is not a guarantee that a parameter will be estimated precisely from finite, noisy data.
This tutorial focuses on how the observation times, assay noise, and observed concentration range affect practical identifiability in one-, two-, and three-compartment concentration profiles. The introductory material on structural and practical identifiability provides the general distinction; here the emphasis is on the information carried by a particular set of samples. A useful starting point is that each sample informs a parameter only to the extent that changing that parameter changes the predicted concentration at that time by an amount distinguishable from measurement noise.
Sampling Coverage and Profile Components
A one-compartment IV bolus profile has one exponential component. A two-compartment profile can have a fast and a slow exponential component, and a three-compartment profile can have three. For illustration, write a multi-exponential concentration profile as
where \(C(t)\) is concentration at time \(t\), \(A_j\) is the concentration amplitude of component \(j\), and \(\lambda_j\) is its positive rate constant. In this expression, \(A_j\) has concentration units and \(\lambda_j\) has inverse-time units. The equation describes a profile; it does not mean that the amplitudes and rates are necessarily independent physiological parameters.
A sampling schedule has broad coverage when it contains informative observations across the portions of the profile that contribute to the parameters of interest. If observations are confined to a limited time window, one or more components may have already become very small, or may not yet be distinguishable from the other components. Having many samples does not automatically solve this problem: numerous observations at nearly the same times can provide less information about a changing profile than a smaller number spread across informative times.
Worked Example: A Fast Component Becomes Small
Consider the hypothetical two-component profile
with \(t\) measured in hours. The fast rate is \(1.0\ \text{h}^{-1}\) and the slow rate is \(0.10\ \text{h}^{-1}\). Compare the component sizes at 1 h and 5 h.
At 1 h, the fast contribution is \(80e^{-1}=29.430\) mg/L, and the slow contribution is \(20e^{-0.1}=18.097\) mg/L. Thus \(C(1)=29.430+18.097=47.527\) mg/L. Both components contribute substantially at this time.
At 5 h, the fast contribution is \(80e^{-5}=0.539\) mg/L, while the slow contribution is \(20e^{-0.5}=12.131\) mg/L. Therefore \(C(5)=0.539+12.131=12.670\) mg/L. The fast component accounts for \(100(0.539036/12.669649)=4.25\%\) of the total, using unrounded model values in the calculation.
A late observation at 5 h may therefore carry little information about the fast component: that component contributes only a small fraction of the predicted concentration. This calculation does not by itself prove that the fast rate cannot be estimated; it shows why observations from other parts of the profile may be needed to estimate it reliably.
Noise, Concentration Range, and Local Sensitivity
The relevance of a component depends not only on its size but also on the assay error at the corresponding concentration. For example, a component contributing \(0.5\) mg/L could be distinguishable with a highly precise assay but effectively obscured by a measurement error of several mg/L. The error model matters as well: with additive error of roughly constant standard deviation, small concentrations can be difficult to resolve; with proportional error, the absolute error typically changes with concentration. A design assessment should use the error model appropriate to the assay and analysis.
A formal way to describe this idea is through local sensitivity. Let \(f(t_i,\theta)\) be the model-predicted concentration at sample time \(t_i\), where \(\theta\) is the vector of parameters. Suppose, for this local calculation, observations have independent additive errors with mean zero and known standard deviations \(\sigma_i\). The sensitivity of the prediction to parameter \(\theta_j\) is \(\partial f(t_i,\theta)/\partial\theta_j\). Dividing by \(\sigma_i\) expresses the change relative to the measurement noise.
A parameter is weakly informed when its column has small entries: changing it barely changes predictions relative to assay noise. Two parameters can also be difficult to estimate separately if their columns are nearly proportional. In that case, a change in one parameter can be offset by a change in the other, leaving predictions nearly unchanged. These are local diagnostics: they describe behavior near a specified parameter value and do not replace examination of the fitted model and data.
A Local Information Justification
The link between sensitivity and precision can be seen by linearizing the model near a parameter value. With independent Gaussian additive errors, the expected local information matrix is \(J^\mathsf{T}J\), using the noise-scaled sensitivities above. The following steps show why small or redundant sensitivity patterns imply weak information.
For a small parameter change \(\Delta\theta\), the vector of predicted concentration changes is approximately \(X\Delta\theta\), where \(X_{ij}=\partial f(t_i,\theta)/\partial\theta_j\).
Dividing each prediction change by its standard deviation gives the noise-scaled change \(J\Delta\theta\). If this vector is small, the proposed parameter change produces little change relative to the expected noise.
The squared size of the scaled change is \((J\Delta\theta)^\mathsf{T}(J\Delta\theta)=\Delta\theta^\mathsf{T}J^\mathsf{T}J\Delta\theta\). If a nonzero \(\Delta\theta\) makes this quantity small, the data weakly distinguish that direction of parameter change.
This justification assumes a locally accurate linear approximation and the stated independent Gaussian error model. It explains why adding observations helps only when they add informative sensitivity, and why reducing assay error can help by increasing noise-scaled changes. In nonlinear mixed-effects analyses, the population model, between-subject variability, and individual sampling schedules also affect the information available; local sensitivity remains a useful conceptual diagnostic, not a complete substitute for evaluating the fitted NLME model.
Worked Example: Is the Fast Contribution Distinguishable From Assay Noise?
Use the hypothetical profile from the previous example and suppose the assay has additive measurement error with standard deviation 5% of the predicted concentration at each time. At 5 h, the predicted concentration is 12.670 mg/L, so the error standard deviation is \(0.05(12.669649)=0.6335\) mg/L.
The fast contribution at 5 h is 0.539 mg/L. Its size relative to the error standard deviation is \(0.539036/0.633482=0.851\). Thus, at this time, the fast contribution is smaller than one error standard deviation under the stated assumption. A single observation there would not sharply distinguish the full prediction from a prediction with that contribution removed.
At 1 h, the predicted concentration is 47.527 mg/L, giving error standard deviation \(0.05(47.5270)=2.376\) mg/L. The fast contribution is 29.430 mg/L, or \(29.430/2.376=12.39\) times that standard deviation. The contrast shows why the same assay can provide very different information about a component at different times. These ratios are diagnostic comparisons, not hypothesis tests or guarantees of parameter precision; estimation uses the full set of observations and the joint model.
Coverage in a Three-Component Profile
In a three-compartment profile, three exponential components may contribute over different parts of the observed time range. A schedule can span early, intermediate, and late times yet still fail to separate all three components if one component is negligible relative to the others or to assay noise throughout the observed range. Conversely, observations across a broad range can help distinguish changing contributions, provided the assay and error model support that distinction.
Worked Example: A Component Falls Below the Assay Scale
Consider the hypothetical three-component profile
where time is in hours. Suppose observations are planned at 0, 2, 8, and 24 h, and the assay has an additive error standard deviation of 0.20 mg/L. At 2 h, the component contributions are \(60e^{-2.4}=5.443\), \(30e^{-0.5}=18.196\), and \(10e^{-0.1}=9.048\) mg/L, giving \(C(2)=32.687\) mg/L.
At 8 h, the contributions are \(60e^{-9.6}=0.004\), \(30e^{-2}=4.060\), and \(10e^{-0.4}=6.703\) mg/L, giving \(C(8)=10.767\) mg/L. By 24 h, they are approximately \(0.000\), \(30e^{-6}=0.074\), and \(10e^{-1.2}=3.012\) mg/L, so \(C(24)=3.086\) mg/L.
At 24 h, the middle component contributes only 0.074 mg/L, less than the assumed 0.20 mg/L error standard deviation. This makes that component difficult to distinguish from noise at this time. The 2 h and 8 h observations have larger contributions from multiple components, but whether those observations identify all three amplitudes and rates depends on the full sensitivity pattern and the other observations. The example illustrates a design warning, not proof that a particular parameter is unidentifiable.
Design Choices, Diagnostics, and Limitations
When practical identifiability looks weak, the design response should target the source of weak information. A broader time range may help if observations omit a part of the profile where a parameter has appreciable influence. Better assay precision may help if a relevant component is present but small relative to noise. Increasing the number of subjects or obtaining repeated profiles can improve population-level information, but only if the resulting observations contribute useful information; repeatedly measuring the same uninformative time range may not resolve a weakly distinguished parameter.
| Approach | What it can improve | Key limitation |
|---|---|---|
| Spread observations across informative times | Captures changes in relative component contributions | Does not guarantee separation if components remain too similar or too small |
| Improve assay precision | Makes smaller prediction changes more distinguishable from noise | Cannot add information about a parameter that barely affects predictions at sampled times |
| Add subjects or replicate observations | Can improve population or average information when observations are informative | Does not automatically fix inadequate time coverage or strong parameter trade-offs |
| Simplify the parameterization or model | Can reduce unsupported complexity when justified scientifically | May impose assumptions and should not be used merely to force a precise estimate |
Useful warning signs include very wide parameter uncertainty intervals, estimates that change substantially with small changes to the data or starting values, strong parameter correlations, and fitted predictions that remain similar across markedly different parameter values. In an NLME analysis, also examine whether individual-level information is weak even when population estimates appear stable. A precise-looking population estimate does not establish that each subject's parameters are well determined.
A common mistake is to conclude that every model parameter is well estimated because the model is structurally identifiable, or because the data set contains many observations. Structural identifiability concerns the idealized model; practical identifiability depends on where and how precisely observations were collected. Likewise, many closely timed measurements may add little new information if their sensitivity patterns are nearly redundant.
No single sampling schedule is optimal for every purpose. The relevant times depend on the model, parameter values, assay error, and the parameters that matter for the scientific question. Sensitivity calculations also depend on the parameter values at which they are evaluated, and local diagnostics may miss nonlinear or global ambiguities. Consequently, use coverage and sensitivity to guide diagnosis, then assess uncertainty and model behavior with the actual estimation procedure.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In a profile with a fast and a slow component, why might observations taken only at late times provide little information about the fast rate?
- If two parameter-sensitivity columns are nearly proportional, what does that imply about estimating the parameters separately?
- For an additive-error assay with standard deviation 0.15 mg/L, a component contributes 0.30 mg/L at one sampled time. What is its contribution in error-standard-deviation units?
- Why might adding ten observations at nearly the same time be less useful than adding a few observations at distinct informative times?
- Does structural identifiability guarantee precise estimation from a finite study? Explain.
Solutions. (1) The fast contribution may have decayed to a small fraction of the total concentration, so its effect can be comparable to or smaller than assay noise. (2) The parameters can produce similar changes in the predictions, so different combinations may fit nearly equally well. (3) \(0.30/0.15=2.00\) error standard deviations; this is a scale comparison, not a test result. (4) Closely timed observations can have redundant sensitivity patterns, while distinct times may capture different component contributions. (5) No. Structural identifiability concerns ideal observations; finite, noisy data with limited coverage may still yield imprecise estimates.