Why Identifiability Matters
A fitted compartment model can produce a concentration-time curve that follows observations closely. That alone does not show that its parameters have been determined by the data. Different parameter values may produce exactly the same curve, or they may produce curves so similar that finite, noisy observations cannot reliably distinguish them. These are different problems: the first is structural identifiability; the second is practical identifiability.
The distinction matters when interpreting parameter estimates. If several parameter values are observationally equivalent, a precise-looking estimate from an optimizer does not make the underlying value uniquely recoverable. If a value is structurally recoverable in principle but weakly determined by the available observations, its estimate may still have substantial uncertainty. In either case, a good-looking fitted curve need not imply that every reported parameter is well supported.
Identifiability is always considered for a specified model, parameter domain, observation scheme, and set of known inputs. It is not a general property of a parameter name by itself. For example, changing which quantities are treated as known, or which outputs are observed, can change what the data can determine.
Structural Identifiability: Uniqueness in Ideal Conditions
“Exact noiseless observations” describes a theoretical question about the model’s predictions, not the usual circumstances of a pharmacokinetic study. Structural identifiability asks whether two allowed parameter values can produce exactly the same observable output under ideal conditions. If they can, collecting more precise measurements of that same output under the same design cannot resolve that exact ambiguity. If they cannot, structural identifiability does not guarantee that a real data set will estimate the parameter precisely.
The parameter domain matters. A model may be ambiguous if two parameters can take any real values but unambiguous after a scientifically justified ordering or positivity constraint. Such constraints must be stated and defensible. A convention that merely assigns labels does not create new information about underlying physiology.
Worked Example: An Output That Determines Only a Sum
Consider a hypothetical observation model with two parameters \(\theta_1\) and \(\theta_2\), each measured in milligrams, and one exact reported quantity \(y\), also in milligrams:
Suppose the ideal observation is \(y=10\) mg and the allowed domain requires both parameters to be positive. The pair \((\theta_1,\theta_2)=(4,6)\) mg gives \(y=4+6=10\) mg. But \((5,5)\) mg also gives \(y=5+5=10\) mg, as does \((2,8)\) mg. In fact, every pair \((a,10-a)\) mg with \(0<a<10\) produces the same observation.
Therefore, the sum is uniquely determined, but the two individual parameters are not structurally identifiable from this observation. Better measurement precision for \(y\) would improve knowledge of the sum; it would not separate the two components. This simple example illustrates the logic of structural non-identifiability without asserting that a particular compartment model has this exact observation equation.
A related ambiguity can arise from parameter labels. A biexponential concentration profile may be written \(C(t)=A_1e^{-\lambda_1t}+A_2e^{-\lambda_2t}\). Exchanging the paired labels \((A_1,\lambda_1)\) and \((A_2,\lambda_2)\) leaves the curve unchanged. If the amplitudes and rates are distinct, an ordering convention such as \(\lambda_1>\lambda_2\) can remove this label-switching ambiguity. It does not, by itself, establish that any underlying physiologic parameters are uniquely determined by the curve.
Practical Identifiability: Reliable Estimation With Real Data
In practice, observations are finite and affected by measurement error. A parameter may affect predictions only slightly over the observed time range, or its effect may resemble that of another parameter. The data can then support a narrow range of predicted profiles while leaving a broad range of parameter values plausible. Practical identifiability therefore depends on the design and the data as well as on the model.
A useful local justification begins with the model prediction vector \(f(\theta)\), whose components are predictions at the chosen observation times, and a small parameter change \(\Delta\theta\). Near a parameter value \(\theta\), a first-order Taylor approximation gives:
where \(J(\theta)\) is the sensitivity matrix, with entries \(J_{ij}=\partial f_i/\partial\theta_j\). If a nonzero change \(\Delta\theta\) lies in, or near, a direction where \(J(\theta)\Delta\theta\) is zero, or small relative to measurement error, then predictions change little in that direction. Parameters may be difficult to estimate separately. Under independent normal measurement errors with known standard deviations collected in a diagonal matrix \(\Sigma\), the local information matrix is proportional to \(J^\mathsf{T}\Sigma^{-1}J\). A small eigenvalue indicates a direction in parameter space with little local information under that design.
This is a local diagnostic, not a proof of global identifiability. A full-rank sensitivity matrix at one parameter value does not rule out a separate, distant parameter value with exactly the same predictions. Conversely, a nearly singular matrix can flag weak practical information even when exact global uniqueness holds. The approximation is most informative when changes are small and the model is sufficiently smooth near the parameter value being considered.
Worked Example: Two Distinct Profiles Nearly Hidden by Measurement Error
Suppose two candidate parameter sets in a hypothetical compartment-model analysis predict concentrations at three sampled times. Their predictions are listed below; each concentration is in milligrams per litre. Assume independent normal measurement errors with standard deviation \(\sigma=0.5\) mg/L at each time. The purpose is to compare the candidates’ predictions, not to claim that either set is the true one.
| Sample time | Candidate A | Candidate B | Difference, A minus B |
|---|---|---|---|
| 1 hour | 8.2 mg/L | 8.1 mg/L | 0.1 mg/L |
| 4 hours | 6.7 mg/L | 6.5 mg/L | 0.2 mg/L |
| 12 hours | 4.5 mg/L | 4.2 mg/L | 0.3 mg/L |
A standardized squared separation for these predictions is the sum of squared differences divided by the error variance:
The candidate predictions differ by less than one error standard deviation at every listed time, and their combined standardized squared separation is 0.56. Under the stated error scale, these observations would provide limited separation between the two candidates. This calculation is not a p-value, a confidence interval, or a formal model-selection test. It illustrates why a finite noisy profile can leave parameter values practically difficult to distinguish, even when their predicted concentrations are not exactly equal.
How to Assess Identifiability
Structural and practical questions require related but distinct checks. The structural question concerns whether the model’s exact predictions can be identical at different parameter values. The practical question concerns how much the observed data and its error model constrain the parameters. Convergence of an estimation algorithm answers neither question by itself.
State which parameters are unknown, which are fixed or known, the permitted parameter values, the observed outputs, and the observation conditions. Identifiability claims are conditional on these choices.
Compare model predictions under distinct allowed parameter values. An exact match for all specified ideal observations shows structural non-identifiability for the affected parameters; an arbitrary relabeling ambiguity may instead be removed by a defensible ordering convention.
Examine prediction sensitivity, uncertainty summaries, and whether substantially different parameter values give similarly plausible fits under the stated measurement-error model. Broad intervals, flat likelihood or objective-function profiles, and strong parameter dependence can signal weak practical information.
Diagnostics should be interpreted together. A small estimated standard error from one fit does not establish identifiability if the result depends on a boundary, a fixed parameter, or a local approximation. Large changes in estimates across starting values, broad profile intervals, or very strong correlations among estimates are reasons to investigate. Multiple optimizer solutions may reveal an ambiguity, but failing to find multiple solutions does not prove uniqueness.
Worked Example: More Informative Observations Can Improve Practical Separation
Return to the two hypothetical candidates and assume the same independent normal measurement error with \(\sigma=0.5\) mg/L. Suppose additional observations are considered at two times for which their predicted concentrations differ by 0.8 mg/L and 1.0 mg/L. For this illustration, assume these additional measurements have the same error standard deviation and are independent of the first three.
The two new contributions to standardized squared separation are \(0.8^2/0.5^2=0.64/0.25=2.56\) and \(1.0^2/0.5^2=1/0.25=4.00\). Including the original value \(D^2=0.56\), the total is:
For these assumed predictions and errors, the enlarged observation set separates the two candidates more than the original three observations did. This is a design illustration, not a guarantee that all parameters will become well estimated: another parameter direction could remain weakly informed, and actual observations may not follow either candidate. The practical lesson is to ask whether the available observations distinguish plausible parameter changes at the relevant noise level, rather than simply counting measurements.
What Identifiability Does—and Does Not—Tell You
Identifiability is not the same as goodness of fit. Two parameter sets can fit the observed concentrations well and still be difficult to distinguish. Nor is it the same as precision: an estimate can be reported with many digits while being weakly supported. Precision reflects information under a model and an analysis; numerical formatting does not create information.
When practical identifiability is weak, possible responses include obtaining observations that better distinguish competing predictions, improving measurement precision, or simplifying the parameterization. Fixing a parameter or imposing a strong constraint can also yield a more stable fit, but the resulting estimate is conditional on that choice. It should not be described as though the data independently identified the fixed or constrained quantity. If structural non-identifiability is present, more precise observations of the same outputs under the same conditions do not resolve the exact ambiguity; additional information or a changed observation scheme may be needed.
The main alternatives are therefore not interchangeable remedies. More observations can improve practical information when they add useful distinctions, but repeated observations that do not distinguish parameter effects may add little. Reparameterization can clarify which combinations are supported, but it does not create information. A simpler model can reduce the number of unknowns, but it must still be scientifically adequate. In all cases, conclusions should state which parameter values are estimated, which are fixed or constrained, and what observations support the result.
Common Mistakes
- Equating a close fit with identified parameters. A curve can fit well while multiple parameter sets give the same or nearly the same predictions. Check parameter uncertainty and alternative parameter values, not just residual fit.
- Calling a parameter structurally unidentified because its standard error is large. A large standard error suggests limited practical information in that analysis; structural non-identifiability is an exact property of the model’s input-output mapping.
- Treating optimizer convergence as proof of uniqueness. Convergence shows that an algorithm met its stopping conditions. It does not establish that no other parameter values produce equivalent predictions.
- Assuming that more samples always solve the problem. Additional data help only insofar as they provide information that distinguishes relevant parameter changes, given their error and the observation design.
- Imposing a constraint and then crediting the data with the resulting precision. Report estimates as conditional on fixed values or constraints; do not present those choices as information supplied by the observations.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A model reports \(y=\theta_1+\theta_2\), and an exact observation gives \(y=12\). Can the observation uniquely determine both positive parameters? What quantity is determined?
- Explain why structural identifiability does not guarantee a precise estimate from a real concentration-time data set.
- Two labeled exponential terms can be exchanged without changing their sum. What ambiguity does this create, and what can an ordering convention accomplish?
- A sensitivity analysis shows a direction in which predictions change by much less than the measurement standard deviation. What does that suggest, and what does it not prove?
- An optimizer converges from one starting value and reports a small standard error. Is this sufficient evidence of global structural identifiability? Explain.
Solutions. (1) No. Any positive pair \((a,12-a)\) with \(0<a<12\) gives the same observation; the sum \(\theta_1+\theta_2=12\) is determined, but the two components are not. (2) Structural identifiability concerns uniqueness with ideal observations. Real data are finite and noisy, and predictions for different parameter values may be too similar relative to measurement error for precise estimation. (3) It creates a label-switching ambiguity: exchanging the paired amplitudes and rates leaves the curve unchanged. A defensible ordering can assign consistent labels when the rates are distinct, but it does not prove that underlying physiologic parameters are identified. (4) It suggests weak practical information in that local direction under the examined design and error scale. It does not prove global structural non-identifiability, because a local sensitivity calculation is not a complete global analysis. (5) No. One converged fit and a small local standard error do not rule out distant equivalent parameter values, dependence on constraints, or other features that make the estimate unstable. Global structural identifiability requires examining uniqueness of the ideal model predictions over the allowed domain.