Why the Distinction Matters
An estimation run may return precise-looking parameter values even when the observations cannot distinguish them. Conversely, parameters that are distinguishable in principle may be estimated imprecisely because the available data are noisy or poorly timed. These are different problems: the first is a property of the model and observation design in an ideal setting; the second is a property of the information available in a particular finite data set.
The distinction guides the response to a failure. More optimization effort cannot resolve an exact equivalence between parameter values. A model transformation or new kind of observation may be needed. When the model is identifiable in principle but information is weak, improved sampling, more precise measurements, or a justified simplification may help.
Definitions in Terms of Model Predictions
Let \(\theta\) denote the model parameters, \(u\) the known inputs such as dose and covariates, and \(\mu(t;u,\theta)\) the model prediction at time \(t\). In a stochastic model, the corresponding object is the distribution of the observations conditional on the design and parameters. Identifiability concerns whether distinct parameter values imply distinguishable predictions or observation distributions under an ideal, sufficiently informative design.
A globally identifiable parameter has one compatible value when the ideal outputs are known. A locally identifiable parameter may have a finite set of compatible values, one of which is the true value locally, but may still have a distinct global alternative. The examples here emphasize a stronger and especially transparent failure: a continuous set of different parameter values produces exactly the same outputs.
These definitions assume that the model form, inputs, and observation process have been specified correctly. Structural identifiability is not a claim that the model is biologically true. It asks whether its parameters could, in principle, be distinguished if the model and ideal outputs were available. Practical identifiability also depends on the actual sample times, measurement error, and observed data.
An Exact Symmetry in Oral Pharmacokinetics
Consider a one-compartment model with first-order absorption after an oral dose \(D\). Let \(F\) be bioavailability, \(V\) apparent volume, \(CL\) clearance, and \(k_a\) the absorption rate constant. Assume positive parameters, \(k_a\ne CL/V\), linear elimination, and observations of concentration with dose and sampling times known. The predicted concentration is
The prediction depends on \(F\) and \(V\) through \(F/V\), and on \(CL\) and \(V\) through \(CL/V\). For any positive scaling factor \(a\) that keeps the transformed parameters within their allowed ranges, set \(F'=aF\), \(V'=aV\), and \(CL'=aCL\). Then \(F'/V'=F/V\) and \(CL'/V'=CL/V\), so every predicted concentration remains unchanged. The three individual parameters therefore cannot all be recovered from oral concentrations alone under this model.
Worked Example: Exposing an Oral-Model Symmetry
Suppose a hypothetical oral study uses \(D=100\) mg and has parameters \(F=0.20\), \(V=20\) L, \(CL=4\) L/h, and \(k_a=1.0\) h\(^{-1}\). Compare these with \(F'=0.40\), \(V'=40\) L, \(CL'=8\) L/h, and the same \(k_a\).
For the first set, \(F/V=0.20/20=0.0100\) L\(^{-1}\), and \(CL/V=4/20=0.200\) h\(^{-1}\). For the second, \(F'/V'=0.40/40=0.0100\) L\(^{-1}\), and \(CL'/V'=8/40=0.200\) h\(^{-1}\). Thus the coefficients and exponential rates in the concentration equation match at every time.
At \(t=2.0\) h, for either parameter set,
The equality is not an accident at \(2.0\) h: the transformed parameters give the same prediction for every \(t\). More precise measurements or a larger sample of the same kind cannot separate \(F\), \(V\), and \(CL\) individually. The data may inform \(F/V\), \(CL/V\), and \(k_a\), subject to adequate sampling, but the exact scaling symmetry prevents identifying all three original parameters.
Why the Symmetry Implies Structural Nonidentifiability
The argument can be stated directly. Write the prediction as a function of the two combinations \(\phi_1=F/V\) and \(\phi_2=CL/V\), along with \(k_a\). Any two parameter vectors that have the same values of these combinations and the same \(k_a\) yield the same \(C(t)\) for every observation time. For example, the transformation in the worked example changes the original parameter vector but leaves \((\phi_1,\phi_2,k_a)\) unchanged. Hence the mapping from \((F,V,CL,k_a)\) to ideal predictions is not one-to-one.
This also explains why a numerical optimizer may report one particular answer from a whole set of equivalent answers. Starting values, bounds, parameter scaling, or numerical details can determine where it stops along that set. That output does not show that the data selected those individual values. Rewriting the model in terms of \(F/V\) and \(CL/V\) makes the estimable combinations explicit; it does not create information about \(F\), \(V\), and \(CL\) separately.
A Broad Likelihood Is Not Necessarily an Exact Symmetry
Practical nonidentifiability is assessed from the likelihood under the chosen observation model and data. With independent normal errors of known standard deviation \(\sigma\), a useful scaled discrepancy for predictions \(\mu_i(\theta)\) and observations \(y_i\) is
For fixed \(\sigma\), the likelihood is proportional to \(\exp[-Q(\theta)/2]\). If changing a parameter over a substantial range changes \(Q\) very little, the likelihood changes little too: the data provide weak practical information in that direction. Unlike the exact oral-model symmetry, there is no requirement that the predictions or likelihood be identical at the alternative values.
Worked Example: Weak Information from a Narrow Sampling Window
Suppose a hypothetical analysis models log concentration \(z\) as \(z(t)=\alpha-kt\), where \(t\) is in hours, \(k\) is in h\(^{-1}\), and independent residual errors are normal with known standard deviation \(\sigma=0.050\) log-concentration units. The observations are \(z=2.00,\ 1.98,\ 1.97\) at \(t=1.0,\ 1.1,\ 1.2\) h. Consider two candidate parameter values.
For \(\alpha=2.03\) and \(k=0.050\) h\(^{-1}\), the predictions are \(1.980,\ 1.975,\ 1.970\). The residuals (observation minus prediction) are \(0.020,\ 0.005,\ 0.000\). Therefore,
For \(\alpha=2.10\) and \(k=0.120\) h\(^{-1}\), the predictions are \(1.980,\ 1.968,\ 1.956\), giving residuals \(0.020,\ 0.012,\ 0.014\). Thus,
The difference is \(Q_2-Q_1=0.126\). Under the stated normal-error model, the likelihood ratio for the second candidate relative to the first is \(\exp[-(Q_2-Q_1)/2]=\exp(-0.063)\approx0.939\). These two substantially different elimination rates therefore have very similar likelihoods for these data. This comparison alone does not establish the complete likelihood shape or the maximum-likelihood estimate, but it demonstrates weak discrimination over this range.
The model is structurally identifiable here: with exact observations at two distinct times, \(k\) follows from the slope, \(k=(z(t_1)-z(t_2))/(t_2-t_1)\), and then \(\alpha=z(t_1)+kt_1\). In the example, however, the observations span only \(0.2\) h and contain residual error. The small change in predicted log concentration across that window makes different slopes difficult to tell apart.
Distinguishing the Diagnoses and Choosing a Remedy
Both structural and practical nonidentifiability can produce small Hessian eigenvalues, high parameter correlations, unstable standard errors, or sensitivity to starting values. Those diagnostics signal weak curvature or numerical difficulty, but do not by themselves reveal its cause. A structural diagnosis requires examining the model map: seek parameter transformations that leave predictions or distributions exactly unchanged, or express predictions in terms of combinations that are actually observed. A practical diagnosis asks how much the likelihood changes over plausible parameter values under the actual design and error model.
For the oral model, the exact transformation proves structural nonidentifiability even before any data are collected. For the narrow-window log-linear example, distinct rates give distinct predictions at distinct times, so the model is identifiable in principle; the observed likelihood is simply insufficiently discriminating. This is the key contrast: exact equality under a parameter transformation versus small but nonzero differences relative to data uncertainty.
Different remedies address these different mechanisms:
| Approach | What it can address | What it cannot do by itself |
|---|---|---|
| Reparameterize using identifiable combinations | Makes the estimable quantities explicit and may improve numerical behavior | Does not recover separate values along an exact symmetry |
| Fix a parameter to a justified known value | Anchors a model when external information genuinely supplies that value | Does not make the fixed value data-informed |
| Add a different type of observation or input | Can break a structural symmetry if the new information responds differently to its parameters | May still leave practical uncertainty if the added data are sparse or noisy |
| Improve sampling times or measurement precision | Can strengthen practical information when predictions differ across the design | Cannot break an exact symmetry that applies to all observations of that type |
Worked Example: Adding Intravenous Information
Suppose oral data support apparent parameters \(CL/F=20\) L/h and \(V/F=100\) L, but cannot separately determine \(F\), \(CL\), and \(V\). A hypothetical intravenous study under the same linear one-compartment model supports \(CL=5\) L/h and \(V=25\) L.
The intravenous results imply \(F=CL/(CL/F)=5/20=0.250\). Independently, \(F=V/(V/F)=25/100=0.250\). The two calculations agree, so these values are internally consistent with the assumed model: \(F=0.250\), \(CL=5\) L/h, and \(V=25\) L. The intravenous information supplies quantities that oral concentrations alone could not separate, thereby breaking the oral-only scaling symmetry.
This conclusion depends on the assumptions that the studies share the relevant disposition parameters, the model is linear, and the apparent oral parameters have been estimated under the stated model. A reparameterization alone would not have supplied the absolute values; the additional route of administration provides new information.
Diagnostics, Limitations, and a Common Mistake
A practical diagnostic sequence is to write down the observation model, inputs, and parameters; inspect the equations for combinations and transformations; then compare likelihood behavior under the actual design. Numerical checks can support this reasoning: evaluate predictions at transformed parameter values, inspect whether the objective is exactly invariant to working precision, and examine whether nearby but distinct predictions are small relative to residual variability. Exact invariance should be justified algebraically, not inferred merely from a failed covariance step or a rounded objective-function trace.
Sampling design matters. Measurements taken at times when two candidate parameter settings predict nearly the same values will contribute little discrimination, even if the parameters are structurally identifiable. Conversely, collecting more observations of the same kind can reduce practical uncertainty but does not help when every such observation preserves an exact symmetry. The analysis also depends on the assumed error model: a likelihood comparison is meaningful only when its observation distribution and variance assumptions are appropriate.
Common mistake: “The standard error is enormous, so the parameter is structurally nonidentifiable.” A large standard error or flat-looking likelihood is evidence of weak information, not proof of an exact structural equivalence. The correction is to test the model equations for an exact transformation that preserves all ideal outputs. If none exists under the stated design, the problem may be practical; if one does exist, optimizing longer or reporting more digits cannot identify the individual parameters.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In the oral model, why does multiplying \(F\), \(V\), and \(CL\) by the same positive factor preserve every concentration prediction?
- Does rewriting the oral model in terms of \(F/V\) and \(CL/V\) identify \(F\), \(V\), and \(CL\) separately?
- In the log-linear example, what feature of the sampling design makes the elimination rate difficult to estimate precisely?
- Can a small Hessian eigenvalue alone establish structural nonidentifiability? What additional check is needed?
- Why can an intravenous study help distinguish parameters that oral observations alone cannot separate?
Solutions. (1) The transformation leaves both \(F/V\) and \(CL/V\) unchanged, as well as \(k_a\), so every term in \(C(t)\) is unchanged. (2) No. It exposes the combinations informed by the predictions but does not distinguish parameter values that share those combinations. (3) The observations span only \(0.2\) h, so candidate rates produce similar changes over the observed window relative to the specified measurement error. (4) No. It indicates weak curvature or numerical difficulty; inspect the model for an exact parameter transformation that preserves ideal outputs. (5) Intravenous data can inform absolute \(CL\) and \(V\), while oral data inform apparent combinations \(CL/F\) and \(V/F\); the additional information can determine \(F\) and break the oral-only scaling symmetry under the stated assumptions.